A fault diagnosis method and system for an electric control system of a seawater hydrogen production device

Abstract

This invention discloses a fault diagnosis method and system for a seawater hydrogen production electronic control system, belonging to the field of hydrogen production fault diagnosis technology. By constructing a five-dimensional multimodal sensor array, comprehensive monitoring of key locations in the seawater hydrogen production electronic control system is achieved. Combining physical mechanism feature calculation and data-driven feature extraction, a hybrid fault fingerprint database containing multi-dimensional information is established, providing a rich feature foundation for subsequent fault diagnosis. The hybrid fault fingerprint database can more comprehensively reflect the characteristics of latent faults unique to the seawater environment. Through causal reasoning and reinforcement learning collaboration, environmental interference can be eliminated through causal graphs, and diagnostic strategies can be dynamically optimized through deep Q-network models. A dynamic threshold adjustment mechanism adapts to fluctuations in seawater temperature and salinity, avoiding misjudgments caused by environmental changes. For latent faults such as PCB heat dissipation copper foil crystallization and slight corrosion of connectors, early identification and warning can be provided, reducing the risk of system downtime.

Description

Technical Field

[0001] This invention relates to the field of hydrogen production fault diagnosis technology, specifically to a fault diagnosis method and system for a seawater hydrogen production electronic control system. Background Technology

[0002] Fault diagnosis of seawater hydrogen production electronic control systems is essentially an intelligent immune system that ensures the safe and efficient operation of the system. It is not just a simple fault alarm, but uses advanced sensing technology, control algorithms and data analysis methods to monitor and analyze the entire hydrogen production process in real time in order to identify abnormalities, locate malfunctions and provide solutions.

[0003] , , Crystals slowly deposit in the PCB heat dissipation copper foil, connector pins, and heat sink fins of control modules for weeks to months, forming insulating or conductive deposits. This leads to increased thermal resistance, causing false triggering of chip thermal protection and micro-short circuits between pins, resulting in signal crosstalk. Moisture condensation and salt crystallization accelerate electrochemical corrosion. Such non-electrical causes leading to electrical failures are almost non-existent in terrestrial industrial systems and represent a hidden fault mode unique to marine environments. Existing technical solutions cannot provide targeted multi-dimensional monitoring and treatment of the hidden nature of salt crystallization, leading to circuit failures. Summary of the Invention

[0004] The purpose of this invention is to provide a fault diagnosis method and system for a seawater hydrogen production electronic control system, which can solve the technical problems mentioned in the background art.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] A fault diagnosis method for a seawater hydrogen production electronic control system includes:

[0007] A five-dimensional multimodal sensor array integrating conductivity, acoustic emission, vibration, pH, and thermal imaging was constructed to collect dynamic operating data of PCB heat dissipation copper foil, connector pins, and heat dissipation fins in the seawater hydrogen production electronic control system. Based on the thermodynamic kinetic model of crystal deposition, several physical characteristics were calculated. Combined with the polarization curve analysis of electrochemical corrosion, several electrochemical characteristics were extracted, and a hybrid fault fingerprint database was established.

[0008] Based on the obtained hybrid fault fingerprint database, the direct causal relationship between faults and sensing parameters is identified through causal graph reasoning, eliminating the indirect correlation caused by environmental disturbances. A deep Q-network is used to optimize the fault classification strategy, and the diagnostic threshold is dynamically adjusted according to real-time sensing data to achieve accurate identification of latent faults in the marine environment.

[0009] Furthermore, the dynamic operating data specifically includes conductivity data, acoustic emission data, vibration data, pH data, and thermal imaging data.

[0010] Furthermore, the calculation of physical mechanism characteristics includes the calculation of crystal growth rate and crystal porosity; the calculation of electrochemical corrosion characteristics is carried out by polarization curve analysis and the Tafel extrapolation method to calculate the corrosion current density and corrosion potential of PCB heat dissipation copper foil and connector pins.

[0011] Furthermore, when establishing a hybrid fault fingerprint database, five-dimensional dynamic operation data of the seawater hydrogen production electronic control system are collected under normal state, thermal resistance increase fault state, signal crosstalk fault state, and electrochemical corrosion fault state.

[0012] The collected five-dimensional dynamic operation data is preprocessed, and the preprocessed five-dimensional dynamic operation data is nonlinearly mapped to extract 16-dimensional data-driven features.

[0013] By integrating data-driven features, physical mechanism features, and electrochemical features, a hybrid fault fingerprint database is constructed.

[0014] Furthermore, based on the hybrid fault fingerprint database, a causal graph containing fault nodes, sensor parameter nodes, and confusion variable nodes is constructed;

[0015] The direct causal correlation strength between faults and sensing parameters is calculated using causal entropy weighted correlation coefficient, and the correlation is dynamically used as the edge of direct causality based on the analysis results of the causal strength.

[0016] Furthermore, when using deep Q-networks to optimize the fault classification strategy, we define the state space and action space, and design a reward function based on the constructed causal graph.

[0017] Furthermore, the initial diagnostic threshold is set at the mean value of each state in the hybrid fault fingerprint database ± 3 times the standard deviation. The diagnostic threshold is then dynamically adjusted based on the Q-value changes in real-time sensor data.

[0018] Furthermore, when dynamically adjusting the diagnostic threshold, if the change in the Q value corresponding to the feature is greater than 0, the probability of failure predicted by the corresponding deep Q-network model increases, thus increasing the threshold sensitivity; if the change in the Q value corresponding to the feature is less than 0, the probability of failure predicted by the corresponding deep Q-network model decreases, thus decreasing the threshold sensitivity.

[0019] Furthermore, the diagnostic threshold is dynamically adjusted, and the corresponding expression is:

[0020] ;in, The new diagnostic threshold for the k-th feature; k is the feature index; The original diagnostic threshold for the k-th feature; The learning rate is adjusted to a threshold value, ranging from [0.01, 0.1]. The change in Q value corresponding to the k-th feature. , These are the current Q value and the historical average Q value, respectively.

[0021] A fault diagnosis system for a seawater hydrogen production electronic control system, comprising:

[0022] Hybrid Fault Fingerprint Database Data Processing Construction Module: Construct a five-dimensional multimodal sensor array integrating conductivity, acoustic emission, vibration, pH, and thermal imaging to collect dynamic operating data of PCB heat dissipation copper foil, connector pins, and heat dissipation fins in the seawater hydrogen production electronic control system. Based on the thermodynamic kinetic model of crystal deposition, calculate several physical characteristics. Combined with the polarization curve analysis of electrochemical corrosion, extract several electrochemical characteristics and establish a hybrid fault fingerprint database.

[0023] The fault diagnosis dynamic correlation identification and optimization module: Based on the processed and acquired hybrid fault fingerprint database, it identifies the direct causal relationship between faults and sensing parameters through causal graph reasoning, eliminates the indirect correlation influence of environmental disturbances, adopts a deep Q-network to optimize the fault classification strategy, and dynamically adjusts the diagnostic threshold according to real-time sensing data to achieve accurate identification of latent faults in the seawater environment.

[0024] Compared to existing solutions, the beneficial effects achieved by this invention are:

[0025] This invention achieves comprehensive monitoring of key locations in the seawater hydrogen production electronic control system by constructing a five-dimensional multimodal sensor array; it establishes a hybrid fault fingerprint database containing multi-dimensional information by combining physical mechanism feature calculation and data-driven feature extraction, providing a rich feature foundation for subsequent fault diagnosis; compared with traditional single feature databases, the hybrid fault fingerprint database can more comprehensively reflect the characteristics of hidden faults unique to the seawater environment, improving the accuracy and reliability of fault diagnosis.

[0026] This invention, through the synergy of causal reasoning and reinforcement learning, can both eliminate environmental interference through causal graphs and dynamically optimize diagnostic strategies through deep Q-network models, enabling accurate identification of latent faults. By adapting to fluctuations in seawater temperature and salinity through a dynamic threshold adjustment mechanism, it can avoid misjudgments caused by environmental changes. For latent faults such as PCB heat dissipation copper foil crystallization and slight corrosion of connectors, it can identify and warn in advance, reducing the risk of system downtime and enhancing early warning capabilities. Attached Figure Description

[0027] The invention will now be further described with reference to the accompanying drawings.

[0028] Figure 1 This is a flowchart illustrating the steps of implementing a fault diagnosis method for a seawater hydrogen production electronic control system according to the present invention.

[0029] Figure 2 This is a block diagram of a fault diagnosis system for a seawater hydrogen production electronic control system according to the present invention. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] Example 1, as Figure 1 As shown, this invention is a fault diagnosis method for a seawater hydrogen production electronic control system, comprising:

[0032] A five-dimensional multimodal sensor array integrating conductivity, acoustic emission, vibration, pH, and thermal imaging was constructed to collect dynamic operating data of PCB heat dissipation copper foil, connector pins, and heat dissipation fins in the seawater hydrogen production electronic control system. Based on the thermodynamic kinetic model of crystal deposition, several physical characteristics were calculated. Combined with polarization curve analysis of electrochemical corrosion, several electrochemical characteristics were extracted, and a hybrid fault fingerprint database was established. The specific steps include:

[0033] A five-dimensional multimodal sensor array is deployed at key locations such as the PCB heat dissipation copper foil, connector pins, and heat dissipation fins of the seawater hydrogen production electronic control system. Specifically, it includes conductivity sensors, acoustic emission sensors, vibration sensors, pH sensors, and thermal imaging sensors.

[0034] Among them, the conductivity sensor adopts a four-electrode method and is deployed near the copper foil for heat dissipation on the PCB to detect local conductivity anomalies caused by crystal deposition;

[0035] The acoustic emission sensor uses a wideband piezoelectric sensor, which is installed on the connector pin base to capture the micro-acoustic signals generated by crystal growth and corrosion propagation;

[0036] The vibration sensor is a triaxial accelerometer fixed to the surface of the heat sink fins to monitor changes in vibration characteristics caused by crystal deposition;

[0037] The pH sensor uses a stain-resistant glass electrode and is installed inside the electronic control system to detect local pH fluctuations caused by moisture condensation.

[0038] The thermal imaging sensor is a non-contact infrared thermal imager, deployed outside the electronic control system, used to monitor abnormal temperature distribution caused by increased thermal resistance due to crystal deposition;

[0039] Based on the deployment of a five-dimensional multimodal sensor array, dynamic operating data of PCB heat dissipation copper foil, connector pins, and heat dissipation fins in the seawater hydrogen production electronic control system are collected and preprocessed to obtain operating data.

[0040] The dynamic operation data specifically includes conductivity data, acoustic emission data, vibration data, pH data, and thermal imaging data; the preprocessing specifically includes removing high-frequency noise from the conductivity data using a sliding window averaging method, with a window size of 10 sampling points;

[0041] The acoustic emission data was denoised using wavelet transform, with a db4 wavelet basis selected and a decomposition layer of 5 layers used to reconstruct the high-frequency coefficients.

[0042] Vibration data are converted from time-domain signals to frequency-domain signals using Fast Fourier Transform, and frequency features in the range of 0~1kHz are extracted.

[0043] pH data were filtered using median filtering to remove outliers, with a window size of 5 sampling points.

[0044] The thermal imaging data was segmented using the Otsu's method to extract feature parameters of temperature anomaly regions; these are all existing conventional technical solutions, and the specific implementation steps are not described here.

[0045] When performing calculations of physical mechanism characteristics, it includes the calculation of crystal growth rate and crystal porosity;

[0046] Among them, based on , , The thermodynamic kinetic model for crystal deposition is used to calculate the crystal growth rate, and the corresponding expression is:

[0047] ;in, This represents the crystal growth rate; The reaction rate constant is determined experimentally, for example... for , for , for ; The ion concentration is obtained by converting the measurement value from the conductivity sensor. The conversion formula is as follows: , For electrical conductivity, molar conductivity, for example for , for , for ; Let be the reaction order. for , for , for ; For activation energy, for , for , for ; is the gas constant, with a value of 8.314; This is absolute temperature, obtained by converting measurements from a thermal imaging sensor. , Temperature in Celsius;

[0048] It should be noted that the calculations involved in the embodiments of the present invention all involve standardization of the calculation data before the calculation, including but not limited to extracting the numerical values ​​of the calculation data and normalization, which are all existing conventional technical solutions. The specific implementation steps are not described here.

[0049] And, calculate the crystal porosity: ;in, It is the crystalline porosity; The actual density of the crystalline deposit layer is calculated from the change in vibration frequency measured by a vibration sensor, using the following formula: , The proportionality constant is determined experimentally and its value is [value missing]. ; This represents the change in vibration frequency. The theoretical density of crystallization, for , for , for ;

[0050] When performing electrochemical corrosion characteristic calculations, the corrosion current density and corrosion potential of PCB heat dissipation copper foil and connector pins are calculated using polarization curve analysis and the Tafel extrapolation method. The corrosion current density and corrosion potential of the PCB heat dissipation copper foil correspond to the characteristic threshold of early crystallization of PCB micro-cracks in the mixed fault fingerprint database; the corrosion current density and corrosion potential of the connector pins correspond to the characteristic threshold of slight corrosion of connector pins in the mixed fault fingerprint database. Specifically:

[0051] The polarization curve was measured using the linear scanning voltammetry method, with a scanning range of -0.2V to +0.2V and a scanning rate of 1mV / s relative to the open circuit potential. Potential and current data of the PCB heat dissipation copper foil and connector pins were collected in real time.

[0052] Convert the collected current data into current density: ;in, Current density; The collected current value; The exposed area of ​​the working electrode;

[0053] And, calculate the logarithm of the absolute value of the current density. ;

[0054] When identifying and fitting the Tafel region, data is imported using Origin or MATLAB software, and polarization curves are plotted, with the horizontal axis representing the polarization potential E and the vertical axis representing the polarization potential E. ;

[0055] The potential range of the anodic Tafel region is , This is the open-circuit potential; the potential in this region is similar to... The relationship is linear and conforms to the Tafel equation for the anode: ;in, This is the polarization potential; This is the corrosion potential; The slope of the anodic polarization curve; This represents the corrosion current density.

[0056] The potential range of the cathode Tafel region is This region conforms to the cathode Tafel equation: ;in, The slope of the cathodic polarization curve;

[0057] In this process, linear fitting is performed on the data from the anodic Tafel region, and the slope of the fitted line is the anodic Tafel slope. Furthermore, a linear fit is performed on the data from the cathode Tafel region, and the absolute value of the slope of the fitted line is the cathode Tafel slope. Linear fitting is a conventional technical solution, and the specific implementation steps will not be elaborated here.

[0058] The corrosion current density is obtained by extending the fitted straight lines of the anode and cathode and finding the intersection point. Alternatively, it can be calculated using a formula, the corresponding expression of which is:

[0059] ;in, The anodic fitting intercept potential is the value of the anodic Tafel fitting line at... Potential at time; The cathode fitting intercept potential, i.e., the cathode Tafel fitting line at... Potential at time;

[0060] The potential value corresponding to the intersection of the extended lines of the fitted straight lines of the anode and cathode is the corrosion potential. It can also be calculated using a formula, the corresponding expression of which is:

[0061] ;

[0062] When establishing a hybrid fault fingerprint database, five-dimensional dynamic operation data of the seawater hydrogen production electronic control system are collected under normal state, thermal resistance increase fault state, signal crosstalk fault state, and electrochemical corrosion fault state.

[0063] In normal conditions, specifically: the system runs continuously for 100 hours in a standard seawater environment, collecting one set of data every second, for a total of 360,000 sets; the standard seawater environment is specifically a salinity of 35‰ and a temperature of 25℃.

[0064] The thermal resistance increase fault condition was specifically described as follows: a 1mm thick simulated crystalline layer was coated on the surface of the heat sink fins, and data was collected over 24 hours, totaling 86,400 sets; the crystalline layer contained... : =7:3 mixed powder;

[0065] Signal crosstalk fault status, specifically: simulated humidity of 90% was continuously introduced into the PCB heat dissipation copper foil, and data was collected for 48 hours, totaling 172,800 sets.

[0066] Electrochemical corrosion failure state, specifically: the connector pins are immersed in simulated seawater for 72 hours, then removed, dried, and installed in the system for 12 hours of operation, for a total of 43,200 sets;

[0067] The collected five-dimensional dynamic operation data is preprocessed and divided into training set, validation set and test set in a ratio of 7:2:1;

[0068] Furthermore, a variational autoencoder (VAE) is used to perform nonlinear mapping on the preprocessed five-dimensional dynamic running data to extract 16-dimensional data-driven features;

[0069] Among them, the pre-processed five-dimensional dynamic operation data are PCB crystal growth rate, PCB corrosion current density, PCB corrosion potential, connector corrosion potential, and connector contact resistance change rate.

[0070] The VAE encoder maps the 5-dimensional raw data to a 16-dimensional latent space through a nonlinear transformation of a fully connected layer, specifically in two steps:

[0071] A 5-dimensional to 32-dimensional mapping is achieved through the first fully connected layer:

[0072] The original 5-dimensional data is input into a fully connected layer containing 32 neurons, with each neuron corresponding to a linear transformation plus a ReLU activation function: ;in, This is the intermediate feature vector, which corresponds to the intermediate representation of the 5-dimensional original data after nonlinear transformation; The weight matrix is ​​5×32, with 5 rows corresponding to the 5-dimensional original data and 32 columns corresponding to the 32 intermediate feature neurons; is a 32-dimensional bias vector, with 32 elements corresponding to 32 intermediate feature neurons; x is a 5-dimensional original data vector;

[0073] Furthermore, a 32-dimensional to 16-dimensional mapping is achieved through a second fully connected layer:

[0074] The 32-dimensional intermediate features are input into a fully connected layer containing 16 neurons, and the mean of the 16-dimensional latent features is output. and variance :

[0075] ;

[0076] ;in, The mean of the 16-dimensional latent features describes the central location of the latent features; It is the logarithmic form of the variance of the 16-dimensional latent features, describing the degree of dispersion of the latent features; Both are 32×16 weight matrices. The 32 rows correspond to the 32-dimensional intermediate features, and the 16 columns correspond to the mean of the 16-dimensional latent features. The 32 rows correspond to the 32-dimensional intermediate features, and the 16 columns correspond to the variance of the 16-dimensional latent features; Both are 16-dimensional bias vectors. The 16 elements correspond to the mean of the 16-dimensional latent features. The 16 elements correspond to the variance of the 16-dimensional latent features;

[0077] In addition, the 16-dimensional latent features extracted by VAE are a nonlinear abstraction and fusion of the 5-dimensional original sensing data. Each feature corresponds to a specific physical mode of PCB or connector failure and can be divided into 4 core feature groups, specifically including the PCB micro-gap early crystallization failure feature group (dimensions 1-4), the connector pin slight corrosion failure feature group (dimensions 5-8), the system-level corrosion risk feature group (dimensions 9-12), and the failure uncertainty feature group (dimensions 13-16).

[0078] Among them, the early crystallization failure feature group of PCB micro-gap includes crystal growth rate-dominant feature, crystallization-corrosion synergistic feature, crystallization uniformity feature, and crystallization risk accumulation feature. The crystal growth rate-dominant feature takes the PCB crystal growth rate as the core and integrates the coupling information of corrosion current density. The crystallization-corrosion synergistic feature is a nonlinear combination of crystal growth rate and corrosion potential. The crystallization uniformity feature is the time series fluctuation feature of crystal growth rate. The crystallization risk accumulation feature is the integral feature of crystal growth rate.

[0079] The connector pin slight corrosion failure feature group includes corrosion potential-dominant feature, corrosion-contact failure synergistic feature, corrosion uniformity feature, and contact risk accumulation feature. The corrosion potential-dominant feature takes the connector corrosion potential as the core and integrates the coupling information of the contact resistance change rate. The corrosion-contact failure synergistic feature is a nonlinear combination of corrosion potential and contact resistance change rate. The corrosion uniformity feature is the time series fluctuation feature of corrosion potential. The contact risk accumulation feature is the integral feature of contact resistance change rate.

[0080] The system-level corrosion risk feature group includes multi-component corrosion synergy features, corrosion-crystallization coupling features, environmental interference suppression features, and comprehensive system health features. The multi-component corrosion synergy features are a nonlinear combination of the corrosion potentials of the PCB and connectors; the corrosion-crystallization coupling features are a fusion of the PCB crystallization features and the connector corrosion features; the environmental interference suppression features are the decoupling features between the raw data and environmental parameters (seawater temperature, salinity); and the comprehensive system health features are a nonlinear fusion of all the raw data.

[0081] The Fault Uncertainty Feature Group includes crystallization Fault Uncertainty Features, Corrosion Fault Uncertainty Features, System Fault Uncertainty Features, and Data Reliability Features. The crystallization fault uncertainty feature is the variance of PCB crystallization features, the corrosion fault uncertainty feature is the variance of connector corrosion features, the system fault uncertainty feature is the variance of system health features, and the data reliability feature is the reconstruction error of the original data.

[0082] The VAE model structure is as follows: the input layer contains 5-dimensional dynamic data; the encoder contains two fully connected layers with 32 and 16 neurons respectively; the decoder contains two fully connected layers with 32 and 5 neurons respectively.

[0083] The specific training parameters are: batch size of 32, learning rate of 0.001, number of training epochs of 1000, and loss function is the weighted sum of reconstruction loss and KL divergence, with the corresponding expression as follows:

[0084] ;in, This is the weighting coefficient, with a value range of [0, 1], and a default value of 0.9; The original input data is a standardized vector of the five-dimensional dynamic operating data. The reconstructed data vector output by the VAE decoder; Mean squared error; For the encoder posterior distribution; It is the prior distribution; KL divergence measures the difference between the encoder distribution and the standard normal distribution. The formula for its calculation is: ;in, The variance of the encoder output; The variance is logarithmic. This is the mean vector output by the encoder;

[0085] Load the trained VAE model and switch the VAE model to evaluation mode, that is, turn off dynamic parameters during training such as dropout and batch normalization to avoid introducing random noise during feature extraction.

[0086] The test set data was divided into batches of 32 samples each, and the batches were input into the model encoder. The first 16 dimensions of the encoder output were extracted as data-driven features, corresponding to the mean vector part of the model output.

[0087] When data-driven features, physical mechanism features, and electrochemical features are integrated to construct a hybrid fault fingerprint database, the normal state includes data-driven features, normal physical mechanism features, and normal electrochemical features. Among them, the normal physical mechanism features correspond to a crystal growth rate of <0.1 and a crystal porosity of <10%; the normal electrochemical features correspond to a corrosion current density of <1 and a corrosion potential of >-0.5.

[0088] The thermal resistance increase fault includes data-driven characteristics, first physical mechanism characteristics, and first electrochemical characteristics; among them, the first physical mechanism characteristics correspond to a crystal growth rate > 1 and a crystal porosity > 30%; the first electrochemical characteristics correspond to a corrosion current density < 5 and a corrosion potential > -0.4.

[0089] Signal crosstalk faults include data-driven characteristics, second physical mechanism characteristics, and second electrochemical characteristics; among them, the second physical mechanism characteristics correspond to a crystal growth rate > 0.5 and a crystal porosity > 20%; the second electrochemical characteristics correspond to a corrosion current density < 3 and a corrosion potential > -0.45.

[0090] Electrochemical corrosion faults include data-driven characteristics, third physical mechanism characteristics, and third electrochemical characteristics; among them, the third physical mechanism characteristics correspond to a crystal growth rate < 0.5 and a crystal porosity < 20%; the third electrochemical characteristics correspond to a corrosion current density > 10 and a corrosion potential < -0.6.

[0091] It should be noted that the multimodal sensing array can simultaneously monitor multiple fault mechanisms such as crystal deposition, moisture condensation, and electrochemical corrosion, covering the main causes of latent faults unique to the marine environment; the calculation of physical mechanism features is based on thermodynamic kinetic models and electrochemical theories, which can essentially reflect the development process of the fault and improve the interpretability of fault diagnosis; the data-driven feature extraction adopts a nonlinear dimensionality reduction method, which can effectively extract the potential information in the original data and improve the representativeness of the features; the hybrid fault fingerprint database integrates multi-dimensional features, which can more accurately describe the characteristics of different fault types and provide a more reliable basis for subsequent fault classification.

[0092] In this embodiment of the invention, a five-dimensional multimodal sensor array was constructed to achieve comprehensive monitoring of key locations in the seawater hydrogen production electronic control system. By combining physical mechanism feature calculation and data-driven feature extraction, a hybrid fault fingerprint database containing multi-dimensional information was established, providing a rich feature foundation for subsequent fault diagnosis. Compared with the traditional single feature database, the hybrid fault fingerprint database can more comprehensively reflect the characteristics of hidden faults unique to the seawater environment, improving the accuracy and reliability of fault diagnosis.

[0093] Based on the obtained hybrid fault fingerprint database, the direct causal relationship between faults and sensing parameters is identified through causal graph reasoning, eliminating the indirect correlation caused by environmental disturbances. A deep Q-network is used to optimize the fault classification strategy, and the diagnostic threshold is dynamically adjusted based on real-time sensing data to achieve accurate identification of latent faults in the marine environment. The specific steps include:

[0094] Based on a hybrid fault fingerprint database, a cause-effect graph is constructed that includes fault nodes, sensor parameter nodes, and confounding variable nodes; specifically:

[0095] The fault nodes specifically include normal state, thermal resistance increase fault, signal crosstalk fault, and electrochemical corrosion fault.

[0096] The sensing parameter nodes specifically include conductivity, acoustic emission signal, vibration acceleration, pH value, and thermal imaging temperature;

[0097] The obfuscated variable nodes specifically include seawater temperature and salinity;

[0098] The direct causal correlation strength between the fault and the sensing parameters is calculated using the causal entropy weighted correlation coefficient, and the corresponding expression is:

[0099] ;in, The direct causal strength between the i-th fault and the j-th sensing parameter is defined, with a value range of [0, 1]; i is the fault number, and j is the sensing parameter number. Let be the Pearson linear correlation coefficient between fault i and sensing parameter j, reflecting the degree of linear correlation, with a value range of [-1, 1]. , Let be the covariance between the sample set corresponding to fault i and the sensing parameter j. These are the standard deviations of the two values; The causal entropy, with a value range of [0, 1], is calculated by controlling for confounding variables such as seawater temperature and salinity using a backdoor adjustment method. , This is a set of environmental confounding variables, specifically including seawater temperature, seawater salinity, seawater current velocity, and environmental pH. Let be the entropy of sensing parameter j under fault condition i; To control the entropy of sensor parameter j under fault i after controlling environmental variables;

[0100] Preserving causal strength Edges that are considered direct causal relationships should be eliminated. The remaining cases are spurious correlations, such as indirect correlations caused by environmental disturbances; the remaining cases are weak correlations, which need to be analyzed in conjunction with other features, such as the correlation between signal crosstalk faults and pH values.

[0101] It should be noted that the above steps can clarify the direct causal link between the fault and the sensing parameters, effectively eliminate indirect interference from environmental disturbances such as seawater temperature and salinity fluctuations, and avoid the situation where changes in conductivity caused by temperature rise are misjudged as faults; it can also provide causal constraints for subsequent reinforcement learning models, ensuring that the model learns the essential correlation of the fault rather than a false correlation.

[0102] When using a Deep Q-Network (DQN) to optimize the fault classification strategy, the state space is defined as follows: the output hybrid fault fingerprint is used as the input state of the Deep Q-Network, with a dimension of 20, represented as a state vector. ;

[0103] Among them, the 20-dimensional hybrid fault fingerprint is obtained by splicing preprocessed standardized data in the order of 16-dimensional data-driven features, 2-dimensional physical mechanism features, and 2-dimensional electrochemical features.

[0104] The 2D physical mechanism features are specifically the micro-gap geometry and connector contact pressure features;

[0105] The geometric characteristics of micro-slits, namely the width-to-depth ratio of PCB micro-slits, reflect the spatial constraints on crystal deposition. When the width-to-depth ratio of PCB micro-slits is greater than 1.2, the flow of seawater within the micro-slits is obstructed, and the crystal deposition rate increases by 3-5 times, which is the core cause of PCB crystallization failure.

[0106] The contact pressure characteristic of a connector refers to the contact pressure between the pin and the socket, which reflects the sealing performance of the contact interface. When the contact pressure is <8N, seawater can easily seep into the contact interface, increasing the corrosion rate of the pin by 2-3 times, which is the core cause of connector corrosion failure.

[0107] Two-dimensional electrochemical characteristics, specifically polarization resistance and passivation film integrity; the polarization resistance characteristic is the polarization resistance of the PCB copper foil, reflecting the reciprocal of the corrosion rate. When the polarization resistance of the PCB copper foil is less than 50 kΩ·cm... 2 At that time, the corrosion rate exceeded the normal level by more than 2 times;

[0108] The passivation film integrity characteristics are the passivation film integrity of the gold-plated layer of the connector pin. It is measured by the capacitive arc radius of the electrochemical impedance spectroscopy (EIS). The capacitive arc radius is positively correlated with the passivation film integrity. When it is <200, microcracks appear in the gold-plated layer. Seawater directly contacts the base metal, and the corrosion rate increases by 4-6 times.

[0109] Define the motion space: Set 4 discrete actions These correspond to the normal state, the thermal resistance increase fault state, the signal crosstalk fault state, and the electrochemical corrosion fault state, respectively.

[0110] Based on the constructed causal graph, design the reward function:

[0111] ;in, Action reward value;

[0112] The structure of a deep Q-network model includes an input layer, hidden layer 1, hidden layer 2, and an output layer.

[0113] Hidden layer 1 contains 64 neurons and uses the ReLU activation function.

[0114] Hidden layer 2 contains 64 neurons and uses the ReLU activation function;

[0115] The output layer outputs a 4-dimensional result, corresponding to the Q value of 4 actions, which correspond to the normal state, thermal resistance increase fault state, signal crosstalk fault state, and electrochemical corrosion fault state, respectively.

[0116] When training the deep Q-network model, the specific training parameters are: batch size 64, learning rate 0.001, experience replay pool capacity 10000, target network update frequency 100 steps, and training epochs 500.

[0117] Furthermore, the historical data of the hybrid fault fingerprint database is divided into training and validation sets in a 7:3 ratio, and the correlation between data is broken through empirical replay to estimate the stable Q-value of the target network.

[0118] Training and validation of deep Q-network models are existing conventional technical solutions, and the specific implementation steps will not be elaborated here.

[0119] It should be noted that by designing a reward function based on causal constraints, the model is guided to prioritize learning the essential correlation between faults and sensing parameters, thus avoiding the spurious correlation and overfitting problem common in reinforcement learning.

[0120] The experience replay mechanism of deep Q-networks can effectively utilize historical fault data in the seawater environment to optimize the generalization ability of classification strategies.

[0121] Using the mean value ± 3 standard deviation of the features of each state in the mixed fault fingerprint database as the initial diagnostic threshold, the diagnostic threshold is dynamically adjusted based on the Q-value changes of real-time sensor data. The corresponding expression is:

[0122] ;in, The new diagnostic threshold for the k-th feature; k is the feature index; The original diagnostic threshold for the k-th feature; The learning rate is adjusted to the threshold to control the adjustment range, with a value range of [0.01, 0.1] and a default value of 0.05. The change in Q value corresponding to the k-th feature. , These are the current Q value and the historical average Q value, respectively.

[0123] Among them, the features k=1, 2, 3, 4 correspond to PCB crystal growth rate, PCB corrosion current density, connector corrosion potential, and connector contact resistance change rate, respectively.

[0124] PCB crystal growth rate, which is the growth rate of salt crystals in the micro-gaps of the PCB, is measured in μm / h. The normal range is 0~0.07, and the fault range is 0.08-0.12.

[0125] PCB corrosion current density, also known as corrosion current density of PCB heat dissipation copper foil, is measured in μA / cm². The normal range is <0.5, and the fault range is 0.5~3.

[0126] The connector corrosion potential, also known as the self-corrosion potential of the connector pins, is measured in volts (V). The normal range is -0.4 to 0.3 volts, and the fault range is -0.55 to -0.45 volts.

[0127] The connector contact resistance change rate, i.e. the relative change rate of the contact resistance between the pin and the socket, is expressed in % / h. The normal range is <2, and the fault range is ≥2.

[0128] When dynamically adjusting diagnostic thresholds, when When the probability of failure predicted by the corresponding deep Q-network model increases, the threshold sensitivity is increased. For example, the crystal growth rate threshold for thermal resistance increase failure is adjusted from >1μm / h to >0.8μm / h.

[0129] when When the probability of a fault predicted by the corresponding deep Q-network model decreases, the threshold sensitivity is reduced to avoid misjudgment. For example, the crystal growth rate threshold for the fault of increased thermal resistance is adjusted from >1μm / h to >1.2μm / h.

[0130] It should be noted that the above steps can adapt to dynamic changes in the seawater environment, such as the drift of sensing parameters caused by the rise in seawater temperature in summer, and maintain diagnostic accuracy by adjusting the threshold; for latent faults such as PCB heat dissipation copper foil crystallization and slight corrosion of connectors, the threshold sensitivity can be adjusted in advance to achieve early warning.

[0131] When implementing accurate identification and verification of latent faults, laboratory simulation samples and field test samples are constructed;

[0132] Among them, when constructing laboratory simulation samples, in the seawater environment simulation platform, the seawater temperature is controlled at 25±2℃ and the salinity at 30±2‰ to manufacture two types of latent faults: early crystallization of PCB heat dissipation copper foil and slight corrosion of connector pins.

[0133] Specifically, for the early crystallization of PCB heat dissipation copper foil, by controlling humidity and current density, 100 samples with a crystal growth rate of 0.08~0.12μm / h and a porosity of 8~12% were generated and collected.

[0134] The connector pins showed slight corrosion. Electrochemical corrosion tests were conducted to generate samples with corrosion current densities of 1~3 μA / cm² and corrosion potentials of -0.55~0.45V. 100 samples were collected.

[0135] Normal state samples, with the fault-free electronic control system operating under the same environment, 800 sets were collected;

[0136] When constructing the field test samples, sensor data during the latent fault early warning stage of the actual seawater hydrogen production power plant were collected to supplement 200 sets, including 50 sets of crystallization faults, 50 sets of corrosion faults, and 100 sets of normal samples.

[0137] Additionally, each sample is manually labeled based on the feature thresholds of the hybrid fault fingerprint database.

[0138] Specifically: crystal growth rate ≥ 0.1 μm / h and porosity ≥ 10% are marked as crystallization failure state; corrosion current density ≥ 2 μA / cm² and corrosion potential ≤ -0.5V are marked as corrosion failure state; the rest are marked as normal state;

[0139] All samples were divided into a test set and a validation set in an 8:2 ratio. The test set was used for model performance evaluation, and the validation set was used for subsequent model optimization.

[0140] Obtain the standardized feature vectors of the test dataset and input them into the trained deep Q-network model to output the diagnostic results of the corresponding samples;

[0141] Based on the diagnostic results of all samples, the diagnostic accuracy, recall, and F1 score of the deep Q-network model were calculated.

[0142] The diagnostic accuracy rate is the ratio of the number of correctly classified samples to the total number of samples.

[0143] Recall is the ratio of the number of correctly identified faulty samples to the actual number of faulty samples.

[0144] ;

[0145] Compare the calculated diagnostic indicators with the validation criteria: when the accuracy is ≥98%, the recall is ≥95%, and the F1 score is ≥96%, the requirements for accurate identification of latent faults are met.

[0146] If all three indicators meet the standards, the model is deemed to have passed the validation; if any one indicator fails to meet the standards, the model is deemed to have failed the validation, and the reasons will be analyzed.

[0147] In this embodiment of the invention, by combining causal reasoning and reinforcement learning, environmental interference can be eliminated through causal graphs, and diagnostic strategies can be dynamically optimized through deep Q-network models, enabling accurate identification of latent faults. By adapting to fluctuations in seawater temperature and salinity through a dynamic threshold adjustment mechanism, misjudgments caused by environmental changes can be avoided. For latent faults such as PCB heat dissipation copper foil crystallization and slight corrosion of connectors, early identification and warning can be provided, reducing the risk of system downtime and enhancing early warning capabilities.

[0148] Example 2, as Figure 2 As shown, a fault diagnosis system for a seawater hydrogen production electronic control system includes:

[0149] Hybrid Fault Fingerprint Database Data Processing Construction Module: Construct a five-dimensional multimodal sensor array integrating conductivity, acoustic emission, vibration, pH, and thermal imaging to collect dynamic operating data of PCB heat dissipation copper foil, connector pins, and heat dissipation fins in the seawater hydrogen production electronic control system. Based on the thermodynamic kinetic model of crystal deposition, calculate several physical characteristics. Combined with the polarization curve analysis of electrochemical corrosion, extract several electrochemical characteristics and establish a hybrid fault fingerprint database.

[0150] The fault diagnosis dynamic correlation identification and optimization module: Based on the processed and acquired hybrid fault fingerprint database, it identifies the direct causal relationship between faults and sensing parameters through causal graph reasoning, eliminates the indirect correlation influence of environmental disturbances, adopts a deep Q-network to optimize the fault classification strategy, and dynamically adjusts the diagnostic threshold according to real-time sensing data to achieve accurate identification of latent faults in the seawater environment.

[0151] In the several embodiments provided by this invention, it should be understood that the disclosed system can be implemented in other ways. For example, the embodiments of the invention described above are merely illustrative; for example, the division of modules is only a logical functional division, and there may be other division methods in actual implementation.

[0152] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0153] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated module can be implemented in hardware or in the form of hardware plus software functional modules.

[0154] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the essential characteristics of the present invention.

[0155] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A fault diagnosis method for a seawater hydrogen production electronic control system, characterized in that, include: A five-dimensional multimodal sensor array integrating conductivity, acoustic emission, vibration, pH, and thermal imaging was constructed to collect dynamic operating data of PCB heat dissipation copper foil, connector pins, and heat dissipation fins in the seawater hydrogen production electronic control system. Based on the thermodynamic kinetic model of crystal deposition, several physical characteristics were calculated. Combined with the polarization curve analysis of electrochemical corrosion, several electrochemical characteristics were extracted, and a hybrid fault fingerprint database was established. Based on the obtained hybrid fault fingerprint database, the direct causal relationship between faults and sensing parameters is identified through causal graph reasoning, eliminating the indirect correlation caused by environmental disturbances. A deep Q-network is used to optimize the fault classification strategy, and the diagnostic threshold is dynamically adjusted according to real-time sensing data to achieve accurate identification of latent faults in the marine environment.

2. The fault diagnosis method for a seawater hydrogen production electronic control system according to claim 1, characterized in that, The dynamic operating data specifically includes conductivity data, acoustic emission data, vibration data, pH data, and thermal imaging data.

3. The fault diagnosis method for a seawater hydrogen production electronic control system according to claim 1, characterized in that, When performing physical mechanism characteristic calculations, it includes the calculation of crystal growth rate and crystal porosity; when performing electrochemical corrosion characteristic calculations, it uses polarization curve analysis and the Tafel extrapolation method to calculate the corrosion current density and corrosion potential of PCB heat dissipation copper foil and connector pins.

4. The fault diagnosis method for a seawater hydrogen production electronic control system according to claim 3, characterized in that, When establishing a hybrid fault fingerprint database, five-dimensional dynamic operation data of the seawater hydrogen production electronic control system are collected under normal state, thermal resistance increase fault state, signal crosstalk fault state, and electrochemical corrosion fault state. The collected five-dimensional dynamic operation data is preprocessed, and the preprocessed five-dimensional dynamic operation data is nonlinearly mapped to extract 16-dimensional data-driven features. By integrating data-driven features, physical mechanism features, and electrochemical features, a hybrid fault fingerprint database is constructed.

5. The fault diagnosis method for a seawater hydrogen production electronic control system according to claim 1, characterized in that, Based on the hybrid fault fingerprint database, a causal graph containing fault nodes, sensor parameter nodes, and confusion variable nodes is constructed. The direct causal correlation strength between faults and sensing parameters is calculated using causal entropy weighted correlation coefficient, and the correlation is dynamically used as the edge of direct causality based on the analysis results of the causal strength.

6. The fault diagnosis method for a seawater hydrogen production electronic control system according to claim 5, characterized in that, When using deep Q-networks to optimize fault classification strategies, we define the state space and action space, and design a reward function by combining the constructed causal graph.

7. A fault diagnosis method for a seawater hydrogen production electronic control system according to claim 6, characterized in that, The initial diagnostic threshold is set at the mean value of each state in the mixed fault fingerprint database ± 3 times the standard deviation. The diagnostic threshold is then dynamically adjusted based on the Q-value changes in real-time sensor data.

8. A fault diagnosis method for a seawater hydrogen production electronic control system according to claim 7, characterized in that, When dynamically adjusting the diagnostic threshold, if the change in the Q value corresponding to the feature is greater than 0, the probability of failure predicted by the corresponding deep Q-network model increases, thus increasing the threshold sensitivity; if the change in the Q value corresponding to the feature is less than 0, the probability of failure predicted by the corresponding deep Q-network model decreases, thus decreasing the threshold sensitivity.

9. A fault diagnosis method for a seawater hydrogen production electronic control system according to claim 7, characterized in that, The expression for dynamically adjusting the diagnostic threshold is: ;in, The new diagnostic threshold for the k-th feature; k is the feature index; The original diagnostic threshold for the k-th feature; The learning rate is adjusted to a threshold value, ranging from [0.01, 0.1]. The change in Q value corresponding to the k-th feature. , These are the current Q value and the historical average Q value, respectively.

10. A fault diagnosis system for a seawater hydrogen production electronic control system, employing a fault diagnosis method for a seawater hydrogen production electronic control system as described in any one of claims 1-9, characterized in that, include: Hybrid Fault Fingerprint Database Data Processing Construction Module: Construct a five-dimensional multimodal sensor array integrating conductivity, acoustic emission, vibration, pH, and thermal imaging to collect dynamic operating data of PCB heat dissipation copper foil, connector pins, and heat dissipation fins in the seawater hydrogen production electronic control system. Based on the thermodynamic kinetic model of crystal deposition, calculate several physical characteristics. Combined with the polarization curve analysis of electrochemical corrosion, extract several electrochemical characteristics and establish a hybrid fault fingerprint database. The fault diagnosis dynamic correlation identification and optimization module: Based on the processed and acquired hybrid fault fingerprint database, it identifies the direct causal relationship between faults and sensing parameters through causal graph reasoning, eliminates the indirect correlation influence of environmental disturbances, adopts a deep Q-network to optimize the fault classification strategy, and dynamically adjusts the diagnostic threshold according to real-time sensing data to achieve accurate identification of latent faults in the seawater environment.

Images