© M. Xu et al., Published by EDP Sciences, 2025

1 Introduction

When warships, ships, and other vessels are sailing at sea, a large amount of water vapor is generated on the sea surface due to the transpiration of seawater [1]. The beating of waves can also cause an increase in water mist on the sea surface. These water vapors contain a large amount of high salt spray particles, including chlorine, sodium, potassium, iodine, etc. [2]. The presence of these particles can cause corrosion of the ship's hull. When the temperature of the ship itself is high, it will accelerate the corrosion rate [3]. This corrosion phenomenon will increase the maintenance cost and shorten the service life of warships, ships, etc. Therefore, it is necessary to closely monitor the concentration of these high salt spray particles in the ocean and take corresponding measures. Chlorine ions are a high content component in marine high salt spray particles and an important indicator for water quality monitoring. Chlorine ions are easily polarized. When exceeding the critical value, they will rapidly accelerate the corrosion rate of the ship [4]. The existing chloride ion detection only involves detection in materials such as steel bars and concrete. There is no in-depth analysis of chloride ions in ocean high salt spray particles yet. Therefore, with the increasing impact of Marine high salt spray particles on Marine navigation and Marine structures, it is particularly important to develop efficient and accurate monitoring systems to track the concentration of corrosive components such as chloride ions in real time. The rapid development of intelligent sensor technology provides a new solution for Marine corrosion monitoring, so it is of great practical significance to innovate the combination of intelligent algorithm and traditional ion selective electrode technology to improve the accuracy of chloride ion detection.

Intelligent sensing technology has achieved good practical results in different fields [5]. Scholars have designed intelligent wireless corrosion detection systems. Yu A et al. designed an intelligent wireless detection system for monitoring steel corrosion using wireless sensor networks. This system achieved real-time data query, effectively reflected the corrosion of steel bars, and facilitated the maintenance of staff [6]. However, there is a lack of relevant research on intelligent sensing technology for monitoring high salt spray particles in the ocean. To prolong the service life of warships, ships, etc., an intelligent system for detecting high salt spray particles in the ocean is designed. This intelligent system introduces a Neural Network (NN) for establishing a sensor temperature compensation model. The introduction of NN improves the detection accuracy of sensors in ocean high salt spray particles. Chlorine ions have a high content in ocean high salt spray particles and have a significant impact on the ship's hull. Therefore, the detection of chloride ions will be an experimental focus. On this basis, an intelligent detection system for detecting chloride ions is established. This is beneficial for helping staff monitor the corrosion of the ship at any time and take corresponding measures.

Jiaqi W et al. optimized traditional radial basis functions using NN. Based on this, they established a temperature compensation model. After model training, the improved temperature compensation model had advantages in compensation accuracy and other aspects [7]. Wang Q et al. proposed an adaptive control method based on NN for temperature compensation of motors. After field experiments, the established temperature compensation model effectively reduced temperature errors. This method effectively improved the phenomenon of slipping in traditional tillage methods [8]. Huang L et al. established a temperature compensation model using NN. They used NN to reduce errors in non-uniform temperature fields and improve the accuracy of aircraft assembly [9]. Yeo W J et al. utilized Long Short-term Memory (LSTM) for temperature compensation during diamond processing. LSTM was adopted for thermal error estimation at different positions. In the final result, the temperature compensation method based on LSTM reduced the impact of thermal errors [10]. Azamathulla H M et al. proposed a method based on artificial NNs and gene expression programming for the temperature prediction problem in the Tabuk region of Saudi Arabia. Among them, the prediction effect of the gene expression programming model was superior to that of the NN [11]. Soh K et al. proposed a lightweight encoding-decoder SAR image detection method based on optimized U-Net for the problem of Marine oil spill monitoring, achieving efficient detection performance with F1 of 91.65% and IoU of 84.59% on the sub-dataset containing polarization information [12]. Yang D et al. proposed a detection method for Marine objects in intelligent ships based on color model cameras for the problem of Marine object detection in intelligent ships, emphasizing the importance of deep learning technology in improving detection performance [13]. Sagadevan S et al. explored effective methods for the sampling and detection of radionuclides in the Marine environment and emphasized the importance of choosing appropriate sampling and detection techniques [14].

In summary, the temperature compensation method based on NN has high accuracy. Back Propagation Neural Network (BPNN) can quickly calculate the partial derivatives of parameters through error backpropagation, which is used to find the global minimum value [15]. The Sparrow Search Algorithm (SSA) has strong local development ability and high stability [16]. SSA can expand the search range and effectively avoid getting stuck in local optima too early. Therefore, the improved SSA is utilized in the optimization of BPNN to obtain an intelligent detection system for chloride ion detection. The key contribution of the research is to design a new funnel-shaped intelligent sensor, combining with the detection demand of chloride ions in NaCl solution, the ion selective electrode is used to detect high salt spray particles. Using BPNN and ISSA to optimize the temperature compensation model, the detection accuracy of the sensor at different temperatures is significantly improved. Compared with similar studies, the innovation points of this research mainly lie in several aspects. First, a new funnel-shaped intelligent sensor is designed. This structure not only optimizes the collection efficiency of chloride ions but also enhances the sensor's response sensitivity to high salt spray particles in the ocean. Secondly, the research combines traditional ion-selective electrode technology with modern intelligent algorithms, introducing BPNN and ISSA to achieve more precise temperature compensation and detection performance.

The main structure of the study is divided into four sections. The first section introduces the influence of Marine high salt spray particles on Marine navigation and structure and the research background. In the second section, the research methods are described in detail, including electrochemical chloride ion detection, improved intelligent sensor technology and the design of funnel chloride ion sensor. The third section reports the experimental results, including the algorithm performance comparison of the smart sensor, the sensitivity test of the sensor and the influence law analysis. In the fourth section, the practical significance of the experimental results and the limitations of the sensor are discussed, and the suggestions for future research are given.

2 Methods and materials

2.1 Electrochemical-based chloride ion detection

Sensor technology for Marine chloride detection plays a crucial role in protecting the Marine environment and extending the service life of Marine navigation facilities. These sensors can monitor the concentration of chloride ions in seawater in real time, promptly identify potential corrosive environments, and thereby provide effective risk assessment and protective measures for ships and Marine structures. Through precise detection of chloride concentration, sensor technology plays a pivotal role not only in preventing and mitigating equipment corrosion and damage, thereby reducing maintenance expenses and minimizing downtime, but also in supplying crucial data for scientific research. This, in turn, fosters the development and execution of marine environmental protection policies and accelerates the achievement of sustainable development objectives.

The existing detection methods for chloride ion concentration in solutions include spectrophotometry, ion chromatography, and modal analysis. However, these methods cannot meet the detection of chloride ions in the marine atmospheric environment. Electrochemical sensors are commonly used structures in ion detection. The ion selective electrode has a simple structure and low equipment requirements, which is conducive to achieving ion detection in a specific environment. Therefore, the detection of chloride ions in marine high salt spray particles is carried out based on ion selective electrodes. This method requires calculating the Nernst potential of the tested component and obtaining the concentration of the tested component through indirect calculation. Figure 1 is a classification diagram of ion selective electrodes.

The above structures are mainly divided into basic primary electrodes, sensitized electrodes, and solid-state electrodes. The sensitized electrode and solid-state electrode are used to modify the original electrode. The main components of an ion selective electrode include an electrode cavity, an internal reference electrode, an internal reference solution, and a sensitive membrane. The sensitive membrane is an important structure of ion selective electrodes. By determining whether it is crystalline, the basic primary electrode can be divided into crystalline and amorphous membrane electrodes. A sensitive electrode, namely Ag/AgCl electrode, is designed based on ion selective electrode as the working and reference electrode. The response potential is calculated using the Nernst equation. This paper calculates the differences between fixed potentials and ultimately obtained chloride ions values. Ion selective electrodes need to detect the membrane potential on the working electrode, which is the sum of interface potential and diffusion potential. A reference electrode is added to the sensor. The electrode potential of the sensor can be expressed using formula (1).

$$\phi ={\phi }_C+{\phi }_M=k^{\prime }\pm \frac{RT}{nF}{\alpha }_{C{l}^{-}}.$$(1)

In formula (1), φ refers to the electrode potential. k^' refers to the equipotential point potential. φ_C refers to the potential of the internal reference solution. φ_M refers to membrane potential. R refers to the molar gas constant. T refers to Kelvin temperature. n refers to the total charge carried by ions in electrode reactions. F refers to the Faraday constant. α refers to ion activity. A two-electrode detection system is selected, with Ag/AgCl as the sensor. Figure 2 shows the structure and schematic diagram of the ion selective electrode. This structure has the characteristics of simple operation, convenient use, and high sensitivity.

When the working electrode is in the test solution, ion exchange occurs on the surface of the electrode membrane. At this point, φ is represented by formula (2).

$$\phi ={\phi }_{A{g}^{+}}^o+\frac{RT}{nF}\ln {\alpha }_{A{g}^{+}}.$$(2)

In formula (2), ${\phi }_{A{g}^{+}}^o$ refers to the potential at the equipotential point. After deduction, the output electromotive force E of the sensor in formula (3) can be obtained.

$$E=E_o^{\prime }-25.69\ln C_{C{l}^{-}}/E=E_o^{\prime }-59.16\lg C_{C{l}^{-}}.$$(3)

In formula (3), $E_o^{\prime }$ refers to the standard potential. $C_{C{l}^{-}}$ refers to the chloride ion concentration. E and $C_{C{l}^{-}}$ have a linear relationship. The slope $-25.69mV/ln{C}_{C{l}^{-}}$ or $-59.16mV/\lg C_{C{l}^{-}}$ is the sensitivity of the ion selective electrode.

Fig. 1 Classification diagram of ion selective electrodes.

Fig. 2 Structure of ion selective electrode and working schematic diagram.

2.2 Improved intelligent sensor technology

In the calculation of the Nernst equation, the corresponding potential of ions is affected by temperature. This is because sensors are easily affected by temperature. This will increase the error of chloride ion detection. Therefore, temperature compensation method is adopted to solve this problem. If the sensor is calibrated at a temperature of T_o, the actual measured temperature is T_o + ΔT. Therefore, based on the derivation, the temperature compensation value of the $C_{C{l}^{-}}$ logarithm in formula (4) can be obtained.

$$\ln C_{C{l}^{-}(T_o)}=m_o+n_o{E}_{(T_o+\mathrm{\Delta }T)}.$$(4)

In formula (4), $E_{(T_o+\mathrm{\Delta }T)}$ refers to the corrected measured potential. After deduction, the temperature compensation formula for the $C_{C{l}^{-}}$ logarithm in formula (5) can be obtained.

$$\ln C_{C{l}^{-}(T_o+\mathrm{\Delta }T)}=(1+\frac{\mathrm{\Delta }T}{T_o})(m_o+n_o{E}_{(T_o+\mathrm{\Delta }T)})/2.303.$$(5)

The above temperature compensation method can to some extent reduce the detection error of the sensor. However, this method is susceptible to the influence of sensor hardware and software configuration. The common temperature compensation methods for sensors include hardware and software compensation methods. The hardware compensation method involves circuit design and cannot achieve full compensation. This method has low accuracy and poor universality, which is difficult to adjust. Software compensation methods can utilize artificial intelligence and numerical analysis methods for temperature compensation. This improves the universality and accuracy of the method, and has high value in practical applications. Therefore, to further reduce the chloride ion detection error, NN is introduced to improve the above method.

The main reason for choosing NN as the temperature compensation model in the research is its superiority in handling complex nonlinear relationships. NN can effectively capture the complex dynamic relationship between input variables (such as temperature and chloride ion concentration) and output results, and has strong learning and generalization capabilities, with strong adaptability. It is particularly suitable for data-driven monitoring tasks. While other artificial intelligence approaches, including explainable artificial intelligence (XAI) and genetic programming (GEP), boast their own distinct advantages, they might not match the direct effectiveness of NNS in terms of learning efficiency and model complexity when confronted with highly nonlinear and time-varying environments. Additionally, an improved SSA is adopted in the experiment to optimize BPNN, namely ISSA-BPNN. Sparrow search algorithm shows strong local development ability and global search ability in the process of finding the optimal solution, which makes the model work effectively under changing environmental conditions. BPNN is excellent at dealing with complex functional relationships and can capture nonlinear and complex relationships. The core basis for choosing the ISSA algorithm to combine with BPNN in the research lies in their collaborative optimization mechanism: Although BPNN has a strong nonlinear fitting ability, its gradient descent method is prone to fall into local optimum and is sensitive to the initial weights. The improved ISSA significantly enhances the global search efficiency by introducing quasi-reflection initialization, Levy flight strategy and dynamic step size factor, and can provide the optimal initial weight threshold combination for BPNN. In the scenario of dynamic changes in ocean temperature, ISSA's population diversity maintenance mechanism and ARO perturbation strategy enable the model to continuously track the time-varying characteristics of the temperature-potential relationship, breaking through the premature convergence limitation of traditional optimization algorithms in dynamic environments.

SSA can use random functions to initialize the population. However, this can result in insufficient individual diversity and uneven distribution. Therefore, Harris Hawks Optimization (HHO) is adopted to expand the search range of the population and improve the optimization ability of SSA [17]. If X_j = (x₁, x₂, ..., x_j) in a space with dimension j, its quasi reflective solution is ${\displaystyle \overline{X}}=({{\displaystyle \overline{x}}}_1,{{\displaystyle \overline{x}}}_2\mathrm{...},{{\displaystyle \overline{x}}}_j)$. Thus, the quasi reflective solution ${\displaystyle \overline{X}}$ in formula (6) is obtained.

$${{\displaystyle \stackrel{-}{X}}}_j=rand(l{b}_j+u{b}_j)/2,X_j).$$(6)

In formula (6), X refers to an individual sparrow. lb and ub refer to parameters' upper and lower bounds. SSA has issues such as premature convergence. To address this deficiency, the Chameleon Swarm Algorithm (CSA) is added to this paper [18]. This approach can improve information exchange between sparrow populations. Formula (7) refers to the updated explorer position obtained after improvement.

$$X_{i,j}^{t+1}=\{\begin{array}{c}\hfill X_{i,j}^t+\mu ((u{b}_j-l{b}_j)r_1+l{b}_j)\mathrm{sgn}(rand-0.5)\begin{array}{cc}\hfill \mathrm{}\hfill & \hfill \mathrm{}\hfill \end{array}{R}_2<ST\hfill \\ \hfill X_{i,j}^t+Q\text{g}L\begin{array}{cc}\hfill \mathrm{}\hfill & \hfill \mathrm{}\hfill \end{array}{R}_2\ge ST\hfill & .\end{array}$$(7)

In formula (7), μ refers to the convergence factor. r₁ and rand refer to a random number with a range of (0,1). R₂ refers to the danger warning value. Q refers to a random number that follows a normal distribution. ST refers to a safety threshold. In SSA, anti-predators can achieve optimal results with individual i. At this moment, formula (8) represents the position update when the anti-predator can sense the crisis and escape in a timely manner.

$$X_{i,j}^{t+1}=X_{i,j}^t+K\cdot (\frac{\left|X_{i,j}^t-X_{worst}^t\right|}{(f_i-f_w)+\epsilon }).$$(8)

In formula (8), ε refers to a constant. K refers to a random number with a range of [−1,1]. $X_{worst}^t$ refers to the current global worst position. f_i and f_w refer to the current individual and worst fitness results of sparrows. After the above steps, the position of the anti-predator belongs to the current optimal position. This situation will result in a narrowing of the individual's search range and lead to premature SSA. To avoid these situations mentioned above, this paper incorporates the Levy flight strategy to expand the search range of anti-predators and enhance global search capabilities. Formula (9) is an improved position update.

$$X_{i,j}^{t+1}=\{\begin{array}{c}\hfill X_{best}^t+\beta \cdot \left|X_{i,j}^t-X_{best}^t\right|\begin{array}{cc}\hfill \mathrm{}\hfill & \hfill \mathrm{}\hfill \end{array}{f}_i>f_g\hfill \\ \hfill X_{i,j}^t+\alpha \cdot \left|X_{i,j}^t-X_{worst}^t\right|\cdot Levy(\xi )\begin{array}{cc}\hfill \mathrm{}\hfill & \hfill \mathrm{}\hfill \end{array}{f}_i=f_g\hfill \end{array}.$$(9)

In formula (9), $X_{best}^t$ refers to the current optimal solution position. α and β refer to random step sizes. f_g refers to the current global optimal fitness value. ξ refers to the vector dimension. The random search path of Levy's flight strategy is represented by formula (10).

$$Levy(\xi )\sim \frac{u}{{\left|v\right|}^{1/2}}.$$(10)

In formula (10), u and v refer to two random numbers. The larger the ξ, the stronger the development capability of SSA. To accelerate the convergence speed and global optimization ability of SSA, the Artificial Rabbits Optimization (ARO) is utilized to update the positional perturbations of sparrow individuals. Formula (11) is the updated individual position of sparrows.

$$X_{i,j}^{t+1}=X_{best}^t+L\cdot c\cdot (X_{else}^t-X_{best}^t)+S\cdot n_1.$$(11)

In formula (11), L refers to the step size factor. $X_{else}^t$ refers to the position of other sparrow individuals. S refers to rounding. n₁ refers to a random number. Then, the sparrow individuals update their positions in the front and back directions. Formula (12) is the calculation method for L.

$$L=(e-\exp {(\frac{t-1}{ite{r}_{\max }})}^2)\cdot \sin (2\pi r_2).$$(12)

In formula (12), e refers to a natural constant. iter_max refers to the maximum iteration. sin refers to a sine function. r₂ refers to a random number with a range of (0,1). Figure 3 is a schematic diagram of L changing with iteration.

Figure 3 shows how the step size factor in ISSA is adjusted as the iterative process changes. This graph shows that in the initial stage of the algorithm, the step size factor is relatively large to enhance the global search ability and allow the algorithm to conduct extensive exploration in the search space. As the iteration progresses, the step size factor gradually decreases to enhance the convergence speed and optimization accuracy. This dynamic adjustment mechanism holds significant physical significance, reflecting the strategies of balancing exploration and development during the optimization process. It enables the algorithm to maintain efficient search while gradually focusing on the region of the optimal solution, thereby enhancing the overall performance and stability of the algorithm. After the above steps, an improved SSA is obtained. Incorporating randomness into the model enables it to perform multiple random samplings during the training phase, thereby effectively mitigating the risk of overfitting. The search space can be extended to avoid the algorithm falling into local optimal solution. Adding random noise can bring the model closer to the actual situation, so as to better evaluate the performance and stability of the sensor under the influence of different environmental variables. The management method of randomness introduction can first set an appropriate noise level, and conduct several independent experiments on the results after randomness introduction to generate different random seeds. The improved method is utilized to optimize BPNN, resulting in ISSA-BPNN. ISSA-BPNN is utilized for temperature compensation of sensors in the ion selective electrode mentioned above. This can improve the accuracy of sensor temperature compensation. Figure 4 shows the application structure of ISSA-BPNN in the sensor temperature compensation model.

In the above sensor temperature compensation model, the input layer data come from the ambient temperature and the chloride ion concentration before compensation. The output data are the predicted chloride ion concentration after compensation. The sensor temperature compensation model includes one hidden layer. The quantity of nodes l is determined by the empirical equation in formula (13).

$$l=\sqrt{m+n}+a.$$(13)

In formula (13), m and n refer to the quantity of input and output nodes. a refers to an adjustment constant within the range of (1,10). The model input layer consists of 2 nodes. The output layer consists of 1 node. The loss function employed by this temperature compensation model is the mean square error. The establishment of this sensor temperature compensation model is completed through the above steps. Figure 5 shows the process of establishing a sensor temperature compensation model based on ISSA-BPNN.

In the above temperature compensation model, first, it should distinguish between the training and testing samples. The sensor temperature and chloride ion concentration are input for training. Then, the BPNN structure is determined, relevant parameters are set, and weights and thresholds are initialized. Next, the parameters of the improved SSA are initialized. The fourth step is to use HHO for population initialization. Next, the fitness and position of individual sparrows are calculated. Then, CSA is utilized to update individual sparrow explorers' position. The seventh step involves updating the position of followers. Subsequently, the position of the anti-predator is updated based on Levy's flight strategy. The ninth step is to use ARO to update the optimal position of sparrow individuals. Then, by updating the fitness value, it is determined whether the maximum iteration or minimum error was met. If the requirements are met, the optimal weight and threshold of BPNN are determined. In the final operational steps, this paper conducts experiments under temperature and other conditions. After a series of data processing, the final chloride ion concentration is output. If the requirements are not met in step nine, the fifth step is returned after population update. In the above parameter selection, reasonable node Settings enable the network structure to effectively represent the relationship between input and output, prevent overfitting or underfitting, and improve the adaptability of the model under different conditions. Appropriate hidden layer and node number can effectively capture complex nonlinear features, and further improve the performance of the model and the generalization ability of new samples. The proper learning rate can prevent the shock and instability in the training process and make the model converge gradually. Appropriate population size can improve the global search ability. The number of iterations determines the depth of the model optimization search, and appropriate Settings ensure that the model is optimized in sufficient time.

Fig. 3 Step size factor changing with iteration.

Fig. 4 ISSA-BPNN in the sensor temperature compensation model.

Fig. 5 Operational workflow of ISSA-BPNN temperature compensation model with hybrid optimization mechanisms.

2.3 Design of funnel-shaped chloride ion sensors

Based on the above intelligent sensor design, a funnel-shaped chloride ion sensor was designed in the experiment. The funnel-shaped design facilitates efficient collection of seawater vapor and mist. Due to the complexity of sea surface climate, the structure of funnel can guide water vapor to gather in the center, which increases the accuracy and sensitivity of chloride ion detection. The funnel's sloped design facilitates the downward flow of liquid, enhancing the liquefaction process and minimizing the liquid's residence time on the electrode surface upon receiving the chloride ion sample. The design of this sensor includes the casing and electrodes. The shell design needs to meet basic requirements such as sealing, corrosion resistance, high temperature resistance, waterproofing, insulation, and light avoidance. Based on economic and feasibility requirements, after material comparison, Polytetrafluoroethylene (PTFE) is selected as the outer shell of the sensor. PTFE has the characteristics of high temperature resistance, corrosion resistance, low temperature resistance, insulation, non-toxicity, and non-adhesiveness, meeting practical requirements. The cover is the main structure utilized by sensors for sample collection. Due to the transpiration of seawater, chloride ions are present in water vapor during actual measurements. Therefore, a funnel-shaped cover is adopted. This is beneficial for the liquefaction of water vapor and facilitates the measurement of chloride ions. The sensor base is a cylindrical structure without a cover, utilized to store electrolytes and reference electrodes. Proton Exchange Membrane (PEM) can conduct protons and isolate electrolytes, which is beneficial for the preservation of liquid electrolytes. PEM needs to meet the requirements of good proton conductivity, electrochemical stability, and reliable mechanical strength. Meanwhile, the point permeation between water molecules and gas in PEM is smaller. Taking all factors into consideration, DuPont Nafion117 PEM is adopted, sourced from DuPont Company in the United States, with the model of Nafion117. Figure 6 is a schematic diagram of the designed funnel-shaped chloride ion sensor.

In the above sensor structure, electrode design is the core of the sensor. The Ag/AgCl electrode is adopted, whose performance determines the application of the sensor. Enhancing the specific surface area of the working electrode can significantly boost the sensor's sensitivity and expedite its response time. Therefore, multiple slit-shaped small holes are designed in the working electrode for flowing the solution to be tested. The working electrode needs to be thin to reduce pore depth. Due to the pre-tightening force of bolts during assembly, the working electrode is prone to deformation. Therefore, it should increase the outer ring's thickness of the electrode. The reference electrode's reaction needs to be single and reversible. The reference electrode adopted is Ag/AgCl, which has high stability, reproducibility, and anti-polarization ability, which is easy to prepare. The Ag/AgCl reference electrode's shape is a long cylindrical pin combined with a rectangular silver plate, which meets practical needs and is easy to process. The preparation of reference electrode and working electrode is carried out using direct chlorination method and chronopotentiometry method for coating [19,20]. Based on the above steps, the design of a funnel shaped chloride ion sensor is completed.

Fig. 6 Schematic diagram of funnel-shaped chloride ion sensor.

3 Results

3.1 Comparison of algorithmic performance in intelligent sensors

The ISA-BPNN model could be run in real time through an embedded system or microcontroller. The experiment met the following computing resource conditions: After optimization, the BPNN structure of this model was streamlined (with 2 nodes in the input layer, approximately 5–10 nodes in the hidden layer, and 1 node in the output layer), and the forward inference computation required about 102–103 floating-point operations per prediction. Millisecond-level response could be achieved on ARM Cortex-M4/M7 (main frequency ≥100 MHz) or RISC-V microcontrollers equipped with FPU. 20–50 KB of RAM should be reserved to store network parameters and intermediate variables, and 100–200 KB of Flash should be reserved to store the solidified model. Meanwhile, fixed-point quantization or pruning techniques should be adopted to further reduce the model size. The ISSA optimization during the training stage needed to be completed on the upper computer. During the deployment stage, only the inference function was retained. Through the pipeline design of sensor data acquisition + model prequel + temperature compensation, the real-time monitoring requirements in the Marine environment were met.

ISSA-BPNN was introduced in the intelligent sensors design to establish a temperature compensation model. ISSA-BPNN was compared with Particle Swarm Optimization-Radial Basis Function (PSO-RBF) [21], Improved Adaptive Genetic Algorithm-Back Propagation Neural Network (IAGA-BPNN) [22], and Whale Optimization Algorithm-Artificial Neural Network (WOA-ANN) [23]. The selected evaluation index was Relative Error (RE). The experiments were conducted under conditions ranging from 0°C to 40°C to simulate the temperature fluctuations that may occur in the Marine environment. A constant temperature control system was set up in the laboratory to ensure that the temperature was stable within the set value of ±1°C during the experiment. The ambient humidity was 60%–80% RH, which simulated Marine fog conditions. A fan was used to maintain uniform air flow in the laboratory to prevent local concentration errors and interference caused by uneven air flow, and the air speed was controlled at 0.1 m /s. A high-precision digital data acquisition system was used to record the potential changes of the sensor in real time. The data acquisition frequency was set to record a potential change every 0.1 s, and the data was stored in a computer for subsequent analysis after each set of experiments. Prior to the experiment, the sensor was rigorously calibrated using a standard chloride ion solution. Zero calibration and span calibration were carried out before each experiment to ensure the measurement accuracy of the sensor. The data used in the study was from March 1, 2023 to October 30, 2023, and 30 sample points were collected once a week. The ambient temperature was 15 °C–30 °C. The salinity of the collection environment was 30 g/L–40 g/L. The relative humidity of the collection environment was 60%–85%RH. The air velocity in the sample area was 0.5 m/ S–2.0 m /s. The depth of water samples collected was 0.5 m to 2.0 m. The sensor adopted an 8 mm diameter Ag/AgCl working electrode (with an effective surface area of 50.24 mm²). After calibration with 0.01–1 mol/L standard NaCl solution, it showed a Nernst slope of −55.78 mV/dec (at 25 °C), and the response time was less than 20 s. The total height of the funnel structure was 42 mm, the inlet diameter was 15 mm, the cone Angle was 60°, and the electrode spacing was 2.5 mm. The ion-selective membrane used Nafion117 PEM with a thickness of 178µm (ion exchange capacity 0.89 mmol/g, proton conductivity 0.083 S/cm). After pretreatment with 0.5 mol/L H₂SO₄, the chloride ion selectivity coefficient logK_(Cl-/Br-) reached −1.2. The overall weight of the sensor was 18 g, with an operating temperature range of −20 °C to 80 °C, and it met the IP68 protection grade. The recalibration frequency of sensors usually depended on their usage environment and application requirements. It was studied that sensors should be calibrated every four months to ensure the accuracy of measurements. The expected operating life of this signal before degradation was between 1 and 3 yr, depending on the material, design and environmental conditions in which it was located. The research aimed to extend the lifespan of sensors and maintain their operational stability by regularly monitoring their performance and promptly calibrating them.

To maintain the performance of the sensor in humidity, pressure and multi-ion environments (such as bromine and sulfate ions), the research adopted moisture-proof and corrosion-resistant materials to reduce the influence of moisture on the electrode and sensitive membrane and ensure stable electrochemical reactions. In addition, the selectivity of the ion-selective electrode could be optimized to enhance the sensitivity to the target ions and reduce the influence of interfering ions. In terms of design, temperature and concentration compensation mechanisms could be introduced to adjust the sensor output through real-time feedback, ensuring high precision under different environmental conditions.

Figure 7 shows the predicted RE of each method under different chloride ion concentrations. Data point refers to the RE data point between the predicted value and the standard value corresponding to each chloride ion concentration measured in the experiment. The prediction results of ISSA-BPNN were mainly concentrated near the standard concentration, with a maximum RE of 1.786%. Other temperature compensation models' RE was greater than 2%.

The chloride ion at a concentration of 0.5 mol/L was adopted as the standard concentration. Figure 8 further compares the RE of the predicted chloride ion concentration and the standard concentration of each method at different temperatures. Temperature indicates the temperature during the experiment, which was used to analyze the performance of the algorithm under different temperature conditions. The predicted concentration of ISSA-BPNN had the smallest RE compared to the standard value. The average RE of ISSA-BPNN was the smallest, at 0.972%. The RE of PSO-RBF, IAGA-BPNN, and WOA-ANN were 4.364%, 3.579%, and 3.268%, respectively. In addition, temperature could affect the sensitivity of the sensor. The error of ISSA-BPNN decreased with increasing temperature. Therefore, the temperature compensation model based on ISSA-BPNN had higher accuracy in practical measurements. The introduction of ISSA-BPNN effectively improved the detection reliability of sensors.

After comparing the above methods, the temperature compensation model based on ISSA-BPNN improved the detection accuracy of the sensor. This was because the improved SSA can enhance the generalization and learning ability of the temperature compensation method, ultimately improving the compensation effect. The introduction of ISSA-BPNN effectively improved the detection reliability of sensors. Meanwhile, ISSA-BPNN introduced different optimization strategies during the calculation process to avoid falling into global optima. The above results confirmed the practical application effect of ISSA-BPNN in predicting chloride ion concentration. Therefore, the introduction of intelligent algorithms was beneficial for the detection of high salt spray particles in the ocean. In the temperature drift verification experiment, the study adopted a standard constant temperature bath (with an accuracy of ±0.1 °C) and six groups of NaCl solutions of different concentrations (0.01–1 mol/L) for verification. Table 1 shows the drift compensation effect of the sensor when the temperature fluctuates within ±15 °C at a reference temperature of 25 °C.

The experimental data analysis in Table 1 showed that the ISSA-BPNN temperature compensation model improved the performance of the sensor. Within the concentration range of 0.01 mol/L to 1 mol/L, the fluctuation amplitude of the measured value after compensation was significantly reduced from the original temperature drift of 13.4 mV to within 0.5 mV. Among them, the maximum absolute error at 0.1 mol/L concentration was 0.4 mV, and the corresponding relative error at the concentration was only 0.44%. It is particularly worth noting that when the concentration increased to 1 mol/L, the relative error further decreased to 0.18%, verifying the stability advantage of the model in high-concentration detection scenarios. The average deviation between the measured value after compensation and the theoretical Nernst value was 0.25 mV, which was 89.3% lower than that before compensation. The error level shows a regular decreasing trend with the increase of concentration, gradually decreasing from 0.25% of 0.01 mol/L to 0.18% of 1 mol/ L. This feature confirmed that this model could effectively eliminate the nonlinear error caused by temperature and fully meet the detection accuracy requirements within the temperature fluctuation range of ±15 °C in the Marine environment.

Fig. 7 Prediction relative errors of various methods under different chloride ion concentrations.

Fig. 8 Relative error between chloride ion prediction and standard concentration for various methods at different temperatures.

3.2 Performance evaluation and impact analysis of funnel-shaped intelligent sensors

Sensitivity refers to the ratio of the output to input variation of a sensor under stable operating conditions. The sensitivity of the funnel shaped intelligent sensor was verified at 25 °C. During the process, sodium chloride solution needed to be added dropwise, with a concentration between 0.001 mol/L and 1 mol/L. Figures 9a and 9b show the sensitivity test potential and fitting curve of the funnel-shaped smart sensor, and the potential value recorded by the sensor was represented numerically to better observe the linear relationship of potential response when the sensitivity changed. After adding sodium chloride solution dropwise for 20 s, the potential instantly increased and then gradually stabilized. The corresponding potentials of sensors were well distinguished under different sodium chloride solutions. Two pairs' difference remained above 50 mV, consistent with the Nernst equation's theoretical value. This result indicated that the designed funnel-shaped intelligent sensor met the experimental requirements. In Figure 9b, the fitted curve obtained had a high linearity, with a correlation coefficient of R² = 0.998. The sensor's slope was −55.78$mv/\lg C_{C{l}^{-}}$ between 0.001 mol/L and 1 mol/L concentration, indicating a sensitivity value of −55.78. This result was close to the theoretical value of 59.16$mv/\lg C_{C{l}^{-}}$.

Stability refers to the ability of a sensor to maintain its own measuring ability and maintain a constant state. Drift can measure the stability of intelligent sensors. The stability of the sensor was tested using drift in Figure 10. After the lower limit detection of the sensor, the effective detection range ≥10⁻⁴ mol/L. In the range of 10⁻⁴ mol/L∼1 mol/L sodium chloride solution, the higher the concentration, the smaller the drift rate, and the more stable the potential.

After designing the funnel-shaped intelligent sensor, the potential and concentration curves were calibrated. This is the foundation that can be directly applied to concentration calculation in subsequent practical measurements. The true concentration of chloride ions could be directly calculated through calibration curves. The concentration range set was 10⁻⁴ mol/L∼1 mol/L, with a sampling time of 5 min and a sampling rate of 0.1 s. Two parallel measurements were conducted for each concentration, resulting in 10 sets of data results in Table 2. When the concentration was 10⁻⁴ mol/L, the calibration curve's average RE was the highest, at 13.95%, meeting the actual measurement requirements.

Due to the transpiration of seawater and the beating of seawater, significant water mist was formed. Water mist contains cations such as sodium, potassium, and magnesium, as well as anions such as chlorine, iodine, fluorine, and sulfate ions. These anions and cations could interfere with the detection of funnel-shaped intelligent chloride ion sensors. Due to the varying ion content in various ocean currents, the average values of Yangshan seawater and global seawater were adopted for calculation. Table 3 shows the calculated results. Sodium content was the ratio of other ions to sodium ions. The concentrations of potassium, calcium, magnesium, sulfate ions, and bromine ions were significantly higher. Therefore, the impact analysis was mainly conducted on these components. After calculating the interference potential curve, potassium, calcium, and magnesium ions had significant potential responses. The ion concentration was inversely proportional to the potential. The slope errors of the Nernst equations for potassium, calcium, magnesium, and sulfate ions were all below 7%. The measurement interference for chloride ions was relatively small. Therefore, in actual measurements, the interference of the aforementioned ions on chloride ion detection could be ignored. However, smart sensors were more sensitive to the content of bromine ions. When the bromine ion content increased, the sensor's potential drif was significant. Therefore, in actual measurement, it should exclude the interference of bromine ions.

The detection results of other Marine ions by the proposed sensor are shown in Table 4. The measured potentials of different ions at specific concentrations were 48.5 mV (Br^-), 49.0 mV (I^-), 50.8 mV (Na+) and 50.0 mV (K+). These potential values indicated that the sensor could successfully identify the presence of different ions in the mother liquor. The sensitivity of the sensor to each ion was good. The sensitivity of I^- was 54.0 mV /mol/L, and the sensitivity of Br^- was 52.0 mV /mol/L. The relative error for Br^- was 2.5%, for I^- 2.3%, for Na + 1.9% and for K + 1.7%. The relative errors of all ions were within the acceptable range, indicating that the measurement results of the sensor were relatively accurate.

The material choice of the sensor (e.g., teflon and Ag/AgCl electrodes) could significantly improve corrosion resistance and was expected to last for several years, but needed to be calibrated and cleaned regularly in practice to prevent salt crystallization or biological attachment. In addition, the effect of salinity changes on chloride ion measurements needed to be considered, as high salinity may enhance the signal, while low salinity might reduce sensitivity. Additionally, changes in ocean currents would also affect the stability of the sensor, so the layout should be reasonably designed during installation to optimize the measurement accuracy. To verify the advancement of the proposed sensor type, a comparative analysis was conducted using traditional electrochemical sensors and optical fiber sensors. The results are shown in Table 5.

Experimental comparison data showed that the sensor in this study demonstrated significant advantages in key performance indicators: the detection range was expanded to 1 × 10⁻⁴ to 1 mol/L, which was one and two orders of magnitude higher than the lower limit detection capability of traditional electrochemical sensors and optical fiber sensors, respectively. The relative error was controlled within the range of 0.18% to 0.44%, which was 82.4% to 93.0% lower than that of the control group, verifying the effectiveness of the ISSA-BPNN algorithm. The response time was shortened to 18 s, reaching 40% of that of traditional methods and 15% of that of optical fiber sensors, meeting the real-time monitoring requirements of Marine environments. The temperature sensitivity index of 0.02 mV/°C was 94.3% lower than that of traditional sensors, highlighting the practical value of the temperature compensation model. Although the anti-bromide ion interference coefficient of −1.2 was slightly inferior to −1.5 of the optical fiber sensor, considering the balance of detection accuracy, response speed and power consumption of 8.5 mW comprehensively, the sensor in this study showed better engineering applicability in the Marine corrosion detection scenario, and its 18-month service life effectively complemented the existing technology.

Fig. 9 Sensitivity test potential and fitting curve.

Fig. 10 Sensitivity test potential and potential drift.

4 Discussion

The above-mentioned study proposed a funnel-shaped intelligent sensor based on ISSA-BPNN for the precise detection of chloride ions in Marine high salt spray particles. In the experimental verification, ISSA-BPNN demonstrated significant advantages in the temperature compensation model, and its average relative error (0.972%) was much lower than that of PSO-RBF (4.364%), IAGA-BPNN (3.579%), and WOA-ANN (3.268%). The main reason for the research advantage was that by introducing HHO to initialize the population, the diversity of the sparrow population was enhanced. The explorer position was updated in combination with CSA, enhancing the local development capability. The addition of the Levy flight strategy has expanded the search range of anti-predators and avoided premature convergence. Not only did this approach excel in dynamic parameter adjustment, but it also dynamically fine-tuned the step size factor via the ARO strategy, thereby significantly bolstering the model's robustness and adaptability. In similar studies, Wang Y et al. proposed a simple solution treatment strategy at room temperature to synthesize perovskite CsPbBr3 quantum dots in response to the insufficient environmental adaptability and moisture stability in traditional synthesis methods, and enhanced their stability by modifying them with zinc-based organic ligands. The prepared perovskite quantum dots were uniformly dispersed in the trimethylbenzene solvent with a small particle size of approximately 9 nm. Subsequently, a sensor with a sensitivity of 0.025 to 400 ppm ethanol gas was developed by spin-coating perovskite quantum dots onto a finger electrode fabricated with response/recovery times of 3.9 s and 3.6 s, respectively [24]. Compared with the above-mentioned studies, the research method significantly reduced the compensation error through multi-strategy fusion and had higher sensitivity in the detection of low-concentration chloride ions. The results showed that the funnel-shaped chloride ion sensor structure design proposed in the study improved the detection performance, stability and sensitivity through efficient ion exchange and anti-interference optimization in the face of complex Marine environments.

The research content, such as “Responsible Consumption and Production” (SDG 12) and “Conservation and Sustainable Use of Marine Ecosystems” (SDG 14), is closely related to other goals and thus holds significant necessity within the Sustainable Development Goals (SDGS). As the problem of corrosion by high salt spray particles on Marine vessels and structures becomes increasingly serious, real-time monitoring and assessment of the concentration of corrosive chloride ions are of vital importance for maintaining the safety of the Marine environment and Marine transportation. By developing new intelligent sensor technologies, the research can effectively enhance monitoring accuracy, reduce the maintenance costs of Marine equipment, thereby promoting efficient resource utilization and minimizing negative impacts on the environment. This not only helps to extend the service life of Marine facilities, but also promotes the sustainability of Marine ecosystems, providing technical support and practical guarantees for achieving the global sustainable development goals.

5 Conclusion

At present, high salt spray particles in the ocean have a strong impact on ship corrosion, while the existing detection methods face problems such as insufficient sensitivity and temperature interference in the dynamic Marine environment. Therefore, a high-precision intelligent sensor was proposed in the research to achieve real-time monitoring of chloride ion concentration. This sensor combined the ISSA and BPNN intelligent algorithms and adopted a funnel-shaped structure to enhance ion exchange efficiency. It used a PTFE housing and Ag/AgCl electrodes to improve corrosion resistance and isolated interfering ions through a PEM. In the experimental results, the method proposed by the research achieved good results. Through algorithm optimization and structural innovation, high-precision detection of chloride ions was realized, providing a reliable tool for corrosion protection of ships. Although the experiment achieved certain results, there are still some shortcomings. For example, smart sensors were more sensitive to the content of bromine ions. When the bromine ion content increased, the sensor potential drift was significant. The interference of bromine ions was not thoroughly studied in the experiment. In addition, the study was not tested in a real Marine environment and did not take into account the complexity of the Marine environment and the reliability of the sensors under strong winds and wave dynamics. Therefore, it is necessary to further eliminate the influence of interfering ions in future studies, and selective membranes can be used to improve the selectivity of chloride ions. A thin modification layer was added on the electrode surface to achieve physical or chemical shielding of bromine ions. Consider designing a sensor with multiple working electrodes, where the response of each electrode to different ions was evaluated and compared. Subsequent research requires large-scale and long-term field tests and performance evaluations in real Marine or nearshore environments to ensure its effectiveness and reliability in actual conditions. In addition, a two-electrode system is designed in this experiment, and its service life needs to be further improved. Subsequent research can consider using a three-electrode system to improve its service life. Additionally, the study will explore the application of nonlinear models in sensor performance optimization, especially in complex Marine environmental conditions, to enhance the accuracy and reliability of chloride and other ion detection.

Funding

The research is supported by: Research on numerical simulation techniques for condensation behavior of electronic devices in marine environments (2023000205001).

Conflicts of interest

The authors declare that they have no conflicts of interest.

Data availability statement

The data used to support the findings of the research are available from the corresponding author upon reasonable request.

Author contribution statement

Mingxuan Xu: Data curation,Funding acquisition, Investigation, Methodology, Software,Validation, Visualization, Writing-original draft, Writing-review & editing. Zebo Gu: Conceptualization, Data curation, Formal analysis, Funding acquisition, Methodology, Project administration, Resources, Supervision, Validation, Visualization, Writing-original draft, Writing-review & editing. Xianlian Mu: Formal analysis, Methodology, Project administration, Resources, Visualization, Writing-original draft. Xingxue Dai: Formal analysis, Investigation, Methodology, Project administration, Resources, Software, Supervision, Writing-review & editing. Chen Zhu: Methodology, Project administration, Resources, Software, Supervision, Validation, Visualization, Writing-review & editing.

References

All Tables

All Figures

	Fig. 1 Classification diagram of ion selective electrodes.
In the text

	Fig. 2 Structure of ion selective electrode and working schematic diagram.
In the text

	Fig. 3 Step size factor changing with iteration.
In the text

	Fig. 4 ISSA-BPNN in the sensor temperature compensation model.
In the text

	Fig. 5 Operational workflow of ISSA-BPNN temperature compensation model with hybrid optimization mechanisms.
In the text

	Fig. 6 Schematic diagram of funnel-shaped chloride ion sensor.
In the text

	Fig. 7 Prediction relative errors of various methods under different chloride ion concentrations.
In the text

	Fig. 8 Relative error between chloride ion prediction and standard concentration for various methods at different temperatures.
In the text

	Fig. 9 Sensitivity test potential and fitting curve.
In the text

	Fig. 10 Sensitivity test potential and potential drift.
In the text