Multi-type fault diagnosis method for electrochemical energy storage

Abstract

The invention relates to the technical field of energy storage fault diagnosis, in particular to a multi-type fault diagnosis method for electrochemical energy storage. According to the method, parameters such as battery voltage, temperature distribution, internal resistance change and state of charge are collected, a multi-fault coupling feature space is constructed, and correlation mapping relations such as overcharge-temperature abnormity, connection fault-local high temperature and aging-internal resistance change are established; an improved salat swarm optimization algorithm is adopted, and uniform initialization, dynamic path search and bidirectional optimization strategies are introduced, so that the search precision and convergence performance are improved; constructing a multi-scale multi-layer neural network, fusing monomers, modules, clusters and system-level features, and performing connection initialization based on association mapping weights; and finally, fault classification, probability weighted fusion and causal chain traceability are realized, and primary and secondary faults and grades are output. According to the method, the coupling fault identification accuracy and the early warning advance are improved, the false alarm rate is reduced, and the method has engineering application value.

Description

Technical Field

[0001] This invention relates to the field of electrochemical energy storage fault diagnosis technology, specifically to a multi-type fault diagnosis method for electrochemical energy storage. Background Technology

[0002] With the large-scale integration of new energy sources, electrochemical energy storage has become an important means for power grid peak shaving, frequency regulation, and backup. Its safe and stable operation is directly related to the reliability of the power system.

[0003] In existing technologies, most methods still focus on monitoring and identifying single fault characteristics, such as relying solely on voltage or temperature thresholds. This fails to effectively address coupled faults like overcharging, temperature rise, aging, and connection anomalies. While some studies have introduced intelligent algorithms for optimization, these often suffer from problems such as blind initialization, rigid search paths, and susceptibility to local optima, leading to unstable diagnostic results. Furthermore, traditional neural networks neglect the constraints of electrochemical physical mechanisms during parameter training, often requiring long training times to converge, and lacking sufficient root cause localization capabilities. These shortcomings make it difficult for existing methods to simultaneously achieve high accuracy, rapid convergence, and interpretability, failing to meet the urgent needs of energy storage systems for early warning and precise operation and maintenance. Summary of the Invention

[0004] This invention provides a multi-type fault diagnosis method for electrochemical energy storage, which aims to improve the operational safety and maintenance efficiency of energy storage systems under complex operating conditions through collaborative diagnosis and source tracing analysis of multiple types of faults, and has important engineering significance.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A multi-type fault diagnosis method for electrochemical energy storage includes: S100: Collects battery module voltage, charging and discharging current, module temperature distribution, internal resistance change and state of charge characteristic data through the battery management system, constructs a multi-fault coupling characteristic space, and establishes the correlation mapping relationship between multiple types of faults; S200: Based on the aforementioned correlation mapping relationship, the sand cat swarm optimization algorithm is initialized using a uniform initialization sequence, and the sequence dimension is set to correspond to the number of battery modules; a dynamic path search strategy is introduced, and the spiral search parameters are set according to the overcharge-temperature anomaly mapping; a bidirectional optimization strategy is introduced, and the reverse learning triggering conditions are adjusted by combining the SEI film damage mechanism and the electrolyte decomposition mechanism. S300: Construct a multi-scale, multi-layer neural network, design network connection weights for coupled fault identification based on the aforementioned correlation mapping relationship, and integrate individual-level, module-level, cluster-level, and system-level features; S400: Utilizes an improved sand cat swarm optimization algorithm to optimize network parameters, inputs multi-fault coupling feature data into the neural network, and performs coupled fault tracing and correlation diagnosis based on fault correlation mapping.

[0006] As a preferred embodiment of the present invention, the step of constructing the multi-fault coupling feature space includes: The battery module voltage data is divided into charging, discharging and resting segments according to the time window, and the voltage variance and voltage gradient in each segment are calculated. Spatial difference processing is performed on the module temperature distribution data to obtain the temperature gradient matrix and the maximum temperature difference value; The first and second derivatives of the internal resistance change are extracted as dynamic characteristics of the internal resistance; By combining the state of charge with voltage, temperature, and internal resistance characteristics in multiple dimensions, a multi-fault coupling feature vector is formed.

[0007] As a preferred embodiment of the present invention, the establishment of the correlation mapping relationship between multiple types of faults includes: Set voltage and temperature anomaly thresholds, and establish a coupling weight matrix for the overcharge-temperature anomaly correlation mapping; Calculate the internal resistance difference between adjacent modules. When the difference exceeds a preset threshold, it is marked as a connection abnormality. Establish a linear mapping function that correlates connection faults with local high temperatures. Based on the ratio of the state of charge decay rate to the internal resistance growth rate, a weight matrix for the aging-internal resistance change correlation mapping is constructed.

[0008] As a preferred embodiment of the present invention, the step of initializing the sand cat swarm optimization algorithm with a uniform initialization sequence includes: The boundary range of the search space is determined based on the overcharge-temperature anomaly correlation mapping, the connection failure-local high temperature correlation mapping, and the aging-internal resistance change correlation mapping. A low-difference quasi-random sequence is generated as a uniform initialization sequence, with the sequence dimension matching the number of battery modules; The sampling points of the initialization sequence are preferentially distributed in the fault coupling sensitive region determined by the correlation mapping relationship; Adjust the initial density of individuals in each fault coupling space according to the weight distribution of the association mapping matrix.

[0009] As a preferred embodiment of the present invention, the dynamic path search strategy includes: Based on the coupling strength in the overcharge-temperature anomaly correlation mapping, the pitch parameter and search trajectory of the spiral search are dynamically adjusted. Based on the gradient direction of the correlation mapping between connection failure and local high temperature, the priority search path of the spiral search is set. When the aging-internal resistance change correlation mapping indicates a deepening of fault coupling, the spiral search radius is reduced and the search accuracy is increased.

[0010] As a preferred embodiment of the present invention, the bidirectional optimization strategy includes: The timing of SEI membrane failure and electrolyte decomposition is predicted by utilizing the aging-internal resistance change correlation mapping relationship, and a trigger threshold for reverse learning is set. The trigger probability of reverse learning is calculated based on the SEI film thickness growth rate and electrolyte concentration change rate. When the association mapping relationship indicates an increased risk of multiple fault coupling, the reverse learning frequency is increased and the search direction weights are adjusted.

[0011] As a preferred embodiment of the present invention, the step of constructing a multi-scale, multi-layer neural network includes: The input layer nodes are configured to correspond one-to-one with the parameters collected by the battery management system, including the battery voltage and internal resistance at the individual cell level, the maximum and minimum temperatures and temperature difference at the module level, the cluster voltage and cluster current at the cluster level, and the total output power and cycle number at the system level. Based on the coupling weight coefficients in the association mapping relationship, initialize the connection weights between the input layer and the hidden layer, and assign the overcharge-temperature anomaly mapping weights, connection failure-local high temperature mapping weights, and aging-internal resistance change mapping weights to the corresponding network connections respectively. The number of hidden layer nodes is set to be adaptively adjusted according to the capacity of the energy storage system, and a nonlinear transformation is performed using the Sigmoid activation function; The output layer nodes are configured to correspond to various fault modes of the energy storage system, including overcharge fault, over-discharge fault, battery aging fault, battery connection fault, and abnormal temperature fault.

[0012] As a preferred embodiment of the present invention, the step of optimizing network parameters using the improved sand cat swarm optimization algorithm includes: The weight matrix and threshold vector of the multilayer neural network are encoded into the position vector of the sand cat individual, and the encoding length is equal to the total number of connection weights and node thresholds in the network. The fitness function is set as a weighted combination of the fault classification error and prediction accuracy of the neural network output, and the fitness value of each sand cat individual under the corresponding network parameters is calculated. The sand cat determines whether to use a prey search strategy or a predator strategy based on the magnitude of the control parameter R. When the absolute value of R is less than or equal to 1, a local search is performed to update the network weights. When the absolute value of R is greater than 1, a global search is performed to explore a new parameter space. By iteratively updating the sand cat's position, the weights and threshold parameters of the neural network are gradually optimized until the preset number of iterations or fitness convergence condition is reached.

[0013] As a preferred embodiment of the present invention, the coupled fault tracing and correlation diagnosis based on fault correlation mapping includes: The normalized multi-fault coupling feature data is input into the optimized neural network, and the output probability of each fault type is calculated through forward propagation. Based on the weight matrix in the fault association mapping relationship, the fault probabilities output by the neural network are weighted and fused to identify the main fault type and the secondary fault type. Based on the causal chain in the correlation mapping relationship, trace back from the detected fault results to determine the root cause and propagation path of the fault; By combining the coupling strength coefficients in the correlation mapping matrix, the degree of mutual influence between different types of faults is evaluated, and the fault level and maintenance recommendations are output.

[0014] The beneficial effects of this invention are: 1. This invention establishes a correlation mapping relationship between multiple types of faults and introduces coupling features such as overcharging-temperature anomalies, connection faults-local high temperatures, and aging-internal resistance changes into the diagnostic model. This allows the neural network to possess physical prior constraints from the initial training stage, avoiding the shortcomings of traditional methods that rely on a single threshold or blind learning. This enables rapid identification and causal tracing of complex coupled faults. This mechanism significantly improves diagnostic accuracy and interpretability, solving the problem of existing technologies struggling to handle multi-fault coupling relationships.

[0015] 2. The improved sand cat swarm optimization algorithm of this invention integrates uniform initialization, dynamic path search, and a bidirectional optimization strategy based on electrochemical mechanisms. This ensures both precise searching of fault-sensitive regions and effective escape from local optima, achieving efficient optimization of neural network parameters. Compared with existing optimization methods, this invention demonstrates superior convergence speed, stability, and false alarm rate. Combined with a multi-scale, multi-layer network structure, it can achieve early warning and precise maintenance in engineering applications, improving the safety and reliability of energy storage systems. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart illustrating a multi-type fault diagnosis method for electrochemical energy storage according to the present invention. Figure 2 This is a schematic diagram of the improved sand cat swarm optimization algorithm of this invention. Detailed Implementation

[0017] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0018] Example 1: As Figure 1 As shown, the present invention provides a multi-type fault diagnosis method for electrochemical energy storage, comprising: S100: Collects battery module voltage, charging and discharging current, module temperature distribution, internal resistance change and state of charge characteristic data through the battery management system, constructs a multi-fault coupling characteristic space, and establishes the correlation mapping relationship between multiple types of faults; Furthermore, the step of constructing the multi-fault coupling feature space includes: The battery module voltage data is divided into charging, discharging and resting segments according to the time window, and the voltage variance and voltage gradient in each segment are calculated. Spatial difference processing is performed on the module temperature distribution data to obtain the temperature gradient matrix and the maximum temperature difference value; The first and second derivatives of the internal resistance change are extracted as dynamic characteristics of the internal resistance; By combining the state of charge with voltage, temperature, and internal resistance characteristics in multiple dimensions, a multi-fault coupling feature vector is formed.

[0019] Specifically, the collected battery module voltage data is first divided into charging, discharging, and resting segments according to the charging and discharging states. Within each segment, the voltage variance is calculated, which is the sum of the squares of the differences between the voltage value at each moment and the average voltage of that segment, divided by the total number of time points. Simultaneously, the voltage gradient is calculated, which is the difference between voltage values ​​at adjacent moments divided by the time interval, resulting in a voltage change rate sequence.

[0020] A mapping is established between the location of each temperature sensor and its temperature value. The temperature difference between adjacent locations is calculated to construct a temperature gradient matrix. The temperature value of a given sensor is subtracted from the temperature values ​​of the four adjacent sensors in each of the four directions to obtain the temperature gradient components in those four directions. Simultaneously, the highest and lowest temperature values ​​are recorded, and the maximum temperature difference is calculated.

[0021] The difference in internal resistance between adjacent time points is calculated and divided by the time interval to obtain the first derivative sequence of the internal resistance change. Further differentiation of the first derivative sequence yields the second derivative sequence, which serves as the dynamic characteristic of the internal resistance.

[0022] A four-dimensional feature matrix is ​​established: the first dimension is the state of charge value, the second dimension is the voltage gradient value at the corresponding time, the third dimension is the temperature gradient value at the corresponding location, and the fourth dimension is the first derivative value of the internal resistance at the corresponding time. Multiple four-dimensional feature vectors are formed by using a sliding time window, constituting a multi-fault coupling feature space.

[0023] By segmenting and combining multidimensional features, the accuracy of abnormal feature identification is improved, providing a more comprehensive data representation for complex coupled faults.

[0024] Furthermore, establishing the correlation mapping relationship between multiple types of faults includes: Set voltage and temperature anomaly thresholds, and establish a coupling weight matrix for the overcharge-temperature anomaly correlation mapping; Calculate the internal resistance difference between adjacent modules. When the difference exceeds a preset threshold, it is marked as a connection abnormality. Establish a linear mapping function that correlates connection faults with local high temperatures. Based on the ratio of the state of charge decay rate to the internal resistance growth rate, a weight matrix for the aging-internal resistance change correlation mapping is constructed.

[0025] Specifically, thresholds for voltage and temperature anomalies are set. A two-dimensional coordinate system is established, with the horizontal axis representing voltage deviation and the vertical axis representing temperature deviation. Data points in historical data that simultaneously exhibit both voltage and temperature anomalies are marked on the coordinate system. A voltage-temperature correlation curve is obtained by fitting the data using the least squares method, and the slope and intercept of the curve are used as weight parameters for the correlation mapping.

[0026] Calculate the internal resistance difference between adjacent modules, and take the average of the internal resistance differences between all adjacent modules as the benchmark value. When the internal resistance difference exceeds 2-3 times the benchmark value, it is marked as a connection abnormality. The specific multiple is determined according to the rated internal resistance range of the module: for modules with a rated internal resistance less than 1mΩ, the threshold is set to 3 times the benchmark value; for modules with a rated internal resistance greater than or equal to 1mΩ, the threshold is set to 2 times the benchmark value. At the same time, record the temperature rise at this location. Establish a linear mapping function with the internal resistance difference as the independent variable and the local temperature rise as the dependent variable, and determine the mapping coefficients through linear regression.

[0027] Calculate the state-of-charge decay rate and the internal resistance growth rate. Construct a correlation weight matrix, where rows correspond to the state-of-charge decay intervals and columns correspond to the internal resistance growth intervals. The matrix elements are weight coefficients, which are determined by statistically analyzing the proportion of samples in each interval combination to the total number of samples.

[0028] The three types of correlation mappings established by statistical regression quantify the coupling relationship between faults, enabling the system to predict fault propagation paths, effectively reducing the fault misjudgment rate, and providing prior knowledge guidance for subsequent intelligent optimization algorithms.

[0029] S200: Based on the aforementioned correlation mapping relationship, the sand cat swarm optimization algorithm is initialized using a uniform initialization sequence, and the sequence dimension is set to correspond to the number of battery modules; a dynamic path search strategy is introduced, and the spiral search parameters are set according to the overcharge-temperature anomaly mapping; a bidirectional optimization strategy is introduced, and the reverse learning triggering conditions are adjusted by combining the SEI film damage mechanism and the electrolyte decomposition mechanism. Furthermore, the initialization step of the sand cat swarm optimization algorithm using a uniform initialization sequence includes: The boundary range of the search space is determined based on the overcharge-temperature anomaly correlation mapping, the connection failure-local high temperature correlation mapping, and the aging-internal resistance change correlation mapping. A low-difference quasi-random sequence is generated as a uniform initialization sequence, with the sequence dimension matching the number of battery modules; The sampling points of the initialization sequence are preferentially distributed in the fault coupling sensitive region determined by the correlation mapping relationship; Adjust the initial density of individuals in each fault coupling space according to the weight distribution of the association mapping matrix.

[0030] Specifically, the boundary range of the search space is determined based on the overcharge-temperature anomaly correlation mapping, the connection fault-local high temperature correlation mapping, and the aging-internal resistance change correlation mapping. The weight distribution range in the coupling weight matrix of the overcharge-temperature anomaly correlation mapping is used as the search boundary of the voltage-temperature dimension; the parameter range of the linear mapping function of the connection fault-local high temperature correlation mapping is used as the boundary of the internal resistance-temperature dimension; and the weight matrix distribution range of the aging-internal resistance change correlation mapping is used as the boundary of the SOC-internal resistance dimension.

[0031] A low-difference quasi-random sequence is used as the uniform initialization sequence. The specific generation steps are as follows: based on the number of battery modules... Determine sequence dimensions and generate 3D sequence, sequence length is the population size A linear mapping transformation from [0,1] to the search space [L,U] is performed on each dimension of the sequence, matching the sequence dimension with the number of battery modules. The initial position vectors of the individual sand cats are: ; in, and Let be the lower and upper bound vectors of the search space, respectively. For the first The low-discrepancy quasi-random sequence vector corresponding to each individual. This indicates element-wise multiplication.

[0032] In terms of distribution, 70% of the initial individuals are placed in the "fault coupling sensitive region" determined by the above mapping relationship. For example, in this embodiment, When both voltage and temperature deviations are greater than 1.5 times their average values, the corresponding parameter range is marked as a sensitive region. When the internal resistance difference between adjacent modules exceeds the threshold, the corresponding temperature rise range will be marked as a sensitive area. When the ratio of SOC decay rate to internal resistance growth rate is in an abnormal range, it is marked as an aging-sensitive region.

[0033] The remaining 30% of individuals are evenly distributed throughout the global space to maintain diversity.

[0034] By adopting a strategy of prioritizing the distribution of low-discrepancy sequences and sensitive regions, the initial population of the optimization algorithm is made closer to the optimal solution region, which significantly improves the convergence speed and avoids the problem of low search efficiency caused by random initialization.

[0035] Furthermore, the dynamic path search strategy includes: Based on the coupling strength in the overcharge-temperature anomaly correlation mapping, the pitch parameter and search trajectory of the spiral search are dynamically adjusted. Based on the gradient direction of the correlation mapping between connection failure and local high temperature, the priority search path of the spiral search is set. When the aging-internal resistance change correlation mapping indicates a deepening of fault coupling, the spiral search radius is reduced and the search accuracy is increased.

[0036] Specifically, based on the coupling strength in the overcharge-temperature anomaly correlation mapping, the pitch parameter and search trajectory of the helical search are dynamically adjusted. When an increased risk of overcharge-temperature anomaly coupling is detected, the pitch parameter is reduced for a finer search. The pitch parameter b is calculated according to the formula... Adjustments, among which This is the initial pitch parameter (value 1.0). The coupling weight coefficient for the overcharge-temperature anomaly mapping (normalized value range [0,1], calculated using the correlation coefficient between voltage deviation and temperature deviation; when the correlation coefficient > 0.7) is used. >0.8 indicates strong coupling; when the coupling risk is low, the pitch parameter... To broaden the search scope.

[0037] After incorporating the dynamic path search strategy, the position update formula for the sand cat swarm optimization algorithm is: ; ; ; ; in: For the optimal position of the population, This indicates the current location of the individual Sand Cat. This is the helical shape constant, which is empirically set to 1.0; Represents a random number within the range [-1, 1]; This represents the random position of an individual in the current population. It means A random number between [a certain number] and [a certain number]; R is the search-expansion balance factor; For the hearing sensitivity of sand cats; This represents the current iteration number; This represents the maximum number of iterations.

[0038] In practice: if the slope of the connection fault-local high temperature mapping function is large, then the search is intensified along the gradient direction; if the aging-internal resistance mapping indicates that the coupling degree is deepening, then the helix radius is reduced to improve the search accuracy.

[0039] By adaptively adjusting the search parameters based on the fault coupling strength, the algorithm can maintain efficient search under different fault modes. Compared with the standard sandcat algorithm, it significantly improves the optimization accuracy and effectively avoids the local optimum trap.

[0040] Furthermore, the bidirectional optimization strategy includes: The timing of SEI membrane failure and electrolyte decomposition is predicted by utilizing the aging-internal resistance change correlation mapping relationship, and a trigger threshold for reverse learning is set. The trigger probability of reverse learning is calculated based on the SEI film thickness growth rate and electrolyte concentration change rate. When the association mapping relationship indicates an increased risk of multiple fault coupling, the reverse learning frequency is increased and the search direction weights are adjusted.

[0041] Specifically, by utilizing the aging-internal resistance change correlation mapping relationship, based on the SEI film thickness growth rate... and electrolyte concentration change rate Calculate the trigger probability of reverse learning.

[0042] Based on the SEI film thickness growth rate Unit: nm / cycle, calculated based on the internal resistance growth rate: The conversion factor Based on the calibration of electrochemical impedance spectroscopy, among which This represents the module's internal resistance at the current test moment (unit: mΩ). This is the internal resistance value of the module when it is initially put into operation (unit: mΩ).

[0043] Electrolyte concentration change rate Unit: mol / L·cycle, calculated based on capacity decay rate: The conversion factor The value was determined based on linear regression analysis of electrolyte composition and capacity testing, where... The initial rated capacity of the battery module (unit: Ah). (Current measured available capacity (unit: Ah), conversion factor).

[0044] It should be noted that and All are substituted in the form of dimensionless fractions; The unit should be rewritten as nm / (cycle·fraction). and The results were obtained by performing EIS and capacity decay regression calibration on more than 50 groups of samples at 25℃ and constant charge and discharge current (report regression R² and standard error). Introducing a random disturbance factor: ; in This represents the current iteration number. To determine the maximum number of iterations, during each iteration, [the value will be determined]. and Random number comparison, if If the size is larger, reverse learning is performed. This design aims to make the algorithm more exploratory in the early stages of iteration and more developmental in the later stages.

[0045] Based on the standard back-learned solution, a guiding term is introduced towards the current optimal individual, and its update formula is: ; in: To solve for the upper bound vector of the solution space, To solve for the lower bound vector of the space, For the first The optimal position of the generation, for express Random numbers within an interval.

[0046] Regarding the execution conditions: when the aging-internal resistance change mapping indicates that the SEI film thickness growth rate exceeds the threshold and the electrolyte concentration change rate increases, the trigger probability of back learning is increased; when all three types of mapping simultaneously indicate an increased risk of multi-fault coupling, the back learning frequency is increased, and the search direction weight is adjusted during the update process to ensure that the algorithm quickly jumps out of local optima and captures the characteristics of complex coupled faults.

[0047] Through the synergistic effect of the three improvement strategies mentioned above, the improved sand cat swarm optimization algorithm can better adapt to the complex characteristics of multi-type fault coupling in electrochemical energy storage systems, providing high-quality parameter configurations for subsequent neural network optimization.

[0048] S300: Construct a multi-scale, multi-layer neural network, design network connection weights for coupled fault identification based on the aforementioned correlation mapping relationship, and integrate individual-level, module-level, cluster-level, and system-level features; Furthermore, the step of constructing a multi-scale, multi-layer neural network includes: The input layer nodes are configured to correspond one-to-one with the parameters collected by the battery management system, including the battery voltage and internal resistance at the individual cell level, the maximum and minimum temperatures and temperature difference at the module level, the cluster voltage and cluster current at the cluster level, and the total output power and cycle number at the system level. Based on the coupling weight coefficients in the association mapping relationship, initialize the connection weights between the input layer and the hidden layer, and assign the overcharge-temperature anomaly mapping weights, connection failure-local high temperature mapping weights, and aging-internal resistance change mapping weights to the corresponding network connections respectively. The number of hidden layer nodes is set to be adaptively adjusted according to the capacity of the energy storage system, and a nonlinear transformation is performed using the Sigmoid activation function; The output layer nodes are configured to correspond to various fault modes of the energy storage system, including overcharge fault, over-discharge fault, battery aging fault, battery connection fault, and abnormal temperature fault.

[0049] Specifically, the multi-level operational characteristics collected by the battery management system are first used as input nodes: Individual cell characteristics: cell voltage and internal resistance; Module-level characteristics: module maximum temperature, minimum temperature, and internal temperature difference within the module; Cluster-level characteristics: cluster voltage, cluster current; System-level characteristics: total output power, number of cycles.

[0050] Let the total number of input layer nodes be Each input node corresponds to a monitoring parameter, denoted as a vector: ; in For the first Each feature data point is normalized before being fed into the network.

[0051] The number of hidden layer nodes is adaptively adjusted according to the capacity of the energy storage system. Specifically, the larger the system capacity, the more battery modules and clusters there are, and the more the number of hidden layer nodes increases accordingly to capture more complex fault coupling relationships. This embodiment uses two hidden layers: the number of nodes in the first hidden layer is set to... This is used to capture the coupling characteristics between individual units and modules; the number of nodes in the second hidden layer is set to... This is used to capture global features between the cluster level and the system level. The weighted sum from the input layer to the first hidden layer is: ; in, For the input layer Nodes and Hidden Layers Node connection weights For the first Node threshold.

[0052] The output of the first hidden layer is: ; Nonlinear features are introduced by using the Sigmoid activation function.

[0053] When initializing the connection weights from the input layer to the hidden layer, the aforementioned fault mapping relationships are considered: for the overcharge-temperature anomaly mapping, the corresponding weight values ​​in the coupling weight matrix obtained in S100 are directly assigned to the weights connecting the voltage input node and the temperature input node to the same hidden layer node; for the connection fault-local high temperature mapping, the slope coefficients of the linear mapping function are proportionally allocated to the connection weights from the internal resistance node and the temperature node to the hidden layer node; for the aging-internal resistance change mapping, the weight coefficients corresponding to the SOC decay rate and the internal resistance growth rate in the weight matrix are assigned to the connection weights from the SOC node and the internal resistance node to the hidden layer node. In this way, the network can highlight the impact of multi-fault coupling characteristics in the early stages of training.

[0054] The output of the second hidden layer is used as input, and after weighted calculation and threshold adjustment, the output layer result is obtained: ; in, The number of output layer nodes corresponds to various fault modes of the energy storage system: overcharge fault, over-discharge fault, battery aging fault, battery connection fault, and abnormal temperature fault.

[0055] The output layer nodes use the Sigmoid activation function to map the diagnostic results to... The interval, where the output value of each node represents the probability of the corresponding fault occurring.

[0056] Through this structural design, the network achieves the fusion of multi-level features at the individual, module, cluster, and system levels: individual and module-level features capture local coupling relationships through the first hidden layer; cluster and system-level features extract global correlation patterns through the second hidden layer; and the output layer combines weighted weights of multiple mapping relationships to achieve multi-type classification and diagnosis of complex coupling faults.

[0057] S400: Utilizes an improved sand cat swarm optimization algorithm to optimize network parameters, inputs multi-fault coupling feature data into the neural network, and performs coupled fault tracing and correlation diagnosis based on fault correlation mapping.

[0058] All trainable parameters in the neural network are encoded into vector form in a fixed order, which serves as the individual position vectors for the sand cat swarm optimization algorithm.

[0059] The iteration begins with the initial population output by S200 (uniform initialization, and the density of sensitive regions has been given by S200).

[0060] In each iteration, the fitness of each individual is calculated according to the fitness function, and the current best individual is determined based on the fitness. .

[0061] The individual position is updated according to the formula in the paper based on the control parameter R and the dynamic path strategy, and the reverse learning update is triggered when necessary (by the aging and internal resistance mapping trigger condition control).

[0062] Iterate until maximum algebraic convergence or fitness convergence, decode the optimal vector to obtain the final network parameters, and use this network for coupled fault prediction, weighted fusion, and source tracing. For example... Figure 2 As shown in the flowchart, the details of each step are explained below.

[0063] Furthermore, the step of optimizing network parameters using the improved sand cat swarm optimization algorithm includes: The weight matrix and threshold vector of the multilayer neural network are encoded into the position vector of the sand cat individual, and the encoding length is equal to the total number of connection weights and node thresholds in the network. The fitness function is set as a weighted combination of the fault classification error and prediction accuracy of the neural network output, and the fitness value of each sand cat individual under the corresponding network parameters is calculated. The sand cat determines whether to use a prey search strategy or a predator strategy based on the magnitude of the control parameter R. When the absolute value of R is less than or equal to 1, a local search is performed to update the network weights. When the absolute value of R is greater than 1, a global search is performed to explore a new parameter space. By iteratively updating the sand cat's position, the weights and threshold parameters of the neural network are gradually optimized until the preset number of iterations or fitness convergence condition is reached.

[0064] Specifically, the parameters of a multi-layer neural network are flattened into column vectors in a fixed order. First, collect the weight matrix from the input layer to the first hidden layer in layer order. Expand the vectors row-wise; then add the first hidden layer bias vector. Then, the weights of the first hidden layer to the second hidden layer are expanded sequentially. Bias Until the output layer weights and biases are determined. The encoded vector is denoted as... ,in This represents the total number of all weighted and biased elements in the network. This vector is the position vector for each individual sand cat. The encoding order remains fixed in the implementation to ensure accurate decoding back to the network parameters after optimization.

[0065] The fitness function is a weighted sum of mean squared error and parameter L2 regularization, defined as: ; in The number of training samples, For the network in parameters Next to the The output probability vector of each sample This is the true one-hot label vector for this sample. To prevent overfitting, an L2 regularization term is added: ; Combine the two items with fixed weights to form the final fitness (which needs to be minimized): ; The weight ratio (0.9:0.1) is based on the trade-off results of previous experiments, aiming to prioritize classification accuracy (high weight for error term) while introducing appropriate regularization to prevent overfitting (low weight for regularization term).

[0066] A uniformly initialized population generated using S200. As the initial individual position. For each individual Perform the following steps: Fitness assessment: Decoded into neural network parameters Perform forward propagation on the training set to compute the output. Substitute into the fitness function .

[0067] Determine the current optimal state: Find the individual with the lowest fitness among all individuals. .

[0068] Computational Search-Development Balance Factor R: In the dynamic path search strategy, the computational search-development balance factor determines whether to perform local predatory updates or global exploratory updates.

[0069] Location Update: In this step, the algorithm still runs based on the dynamic path search strategy and bidirectional optimization strategy introduced in S200, which will not be elaborated here.

[0070] Boundary control and validity verification: After location update, Any dimensional component exceeds the search boundary If so, then the component value is set to the boundary value closest to it (i.e., projection is performed using the truncation method).

[0071] Convergence criterion: If the maximum algebra is reached... (Typically set to 500-1000) or the optimal fitness change is less than the threshold over several generations. (usually set to) If the condition is met, then stop iterating and output the optimal individual. .

[0072] The improved sand cat swarm algorithm optimizes network parameters. Compared with the traditional gradient descent method, it avoids the gradient vanishing problem, improves the accuracy of parameter optimization, and enhances the network's generalization ability.

[0073] Furthermore, the coupled fault tracing and correlation diagnosis based on fault correlation mapping includes: The normalized multi-fault coupling feature data is input into the optimized neural network, and the output probability of each fault type is calculated through forward propagation. Based on the weight matrix in the fault association mapping relationship, the fault probabilities output by the neural network are weighted and fused to identify the main fault type and the secondary fault type. Based on the causal chain in the correlation mapping relationship, trace back from the detected fault results to determine the root cause and propagation path of the fault; By combining the coupling strength coefficients in the correlation mapping matrix, the degree of mutual influence between different types of faults is evaluated, and the fault level and maintenance recommendations are output.

[0074] Specifically, Decoding back to network parameters The parameters are written back to the network weights and biases directly in the reverse order of encoding (i.e., output layer first, input layer last to the first hidden layer) and row priority rule to form the final diagnostic model. Small-amplitude gradient descent fine-tuning can be performed on the same training set, but this implementation uses the output parameters of the sand cat swarm optimization algorithm as the final weights to maintain the optimization path description consistent with the previous description.

[0075] The raw feature vectors acquired in real time are normalized to the range [0,1] and then input into the optimized network. The normalization formula is: ; in , This represents the maximum and minimum values ​​of the feature within the training or historical window.

[0076] The neural network performs layer-by-layer computation on the input features and maps them to a probability interval at the output node using the Sigmoid function. Let the output of the optimized network for the current sample be denoted as . ,in This represents the number of fault categories.

[0077] Using the correlation mapping weight matrix established in S100 We perform weighted fusion of the network output probabilities to reflect the known fault coupling structure. The fusion score vector is defined as: ; in Indicates the first prediction from the network Class 1 faults The mapping weights for each fault class, the matrix, have been obtained and normalized from historical sample statistical regression in S100. After fusion, the scores for each fault class are normalized to ensure... .

[0078] Will Sort by value in descending order: the category with the highest score is the primary fault, and the category with the highest score is the secondary fault. To clearly define the output diagnostic level, based on statistical analysis of a large number of historical fault cases, the following fault severity judgment thresholds are set: If the main fault score This is classified as an emergency fault; like This is classified as a high-risk fault. like It was determined to be a low-to-medium risk fault.

[0079] The established association mapping includes not only the mapping weights between faults, but also the causal coefficient matrix from the input group to the fault. The rows correspond to input groups (such as individual cell voltage group, internal resistance group, temperature group, SOC group), and the columns correspond to fault categories. Each input group is defined. The root cause contributions are divided into: ; Pick The input set is used to infer the root cause. This method utilizes the causal chain of association mapping to trace back layer by layer to the most upstream input factor, thereby providing actionable inspection or maintenance targets. This calculation corresponds to the claim's requirement to "trace upwards from the detected fault results to determine the root cause and propagation path of the fault."

[0080] Define any two types of faults The coupling strength is: ; in The coupling coefficients are obtained statistically from the association mapping matrix. The total system coupling risk score is defined as: ; Based on engineering experience and safe operating procedures, the following handling recommendations are given according to fixed threshold values: If This was determined to be a high-coupling risk and immediate offline inspection was recommended; if It was determined to be of moderate coupling risk and local checks and balancing were recommended; if The model was classified as having low coupling risk and close monitoring was recommended. This quantification method combines the coupling strength of the mapping matrix with the model output probability to obtain an actionable risk scale.

[0081] It should be noted that It consists of the association frequency and regression coefficient of historical samples and is normalized row by row (the sum of each row is 1). Obtained by the structural causal discovery algorithm (describe the algorithm used). Values ​​are normalized to [0,1], reducing system coupling risk. Defined as normalized .

[0082] The optimized system will output diagnostic results according to the following structure: primary fault type and level (urgent / high risk / medium-low risk); secondary fault type; root cause input group (based on...). (Ranking given as the top two); Coupling risk score. The output includes corresponding handling suggestions (immediate offline inspection / partial inspection and balancing / continued monitoring); in the case of an emergency fault, preliminary actions will also be suggested, such as "immediately disconnect the charging circuit and initiate module-level isolation checks." The suggestion text is linked to the risk level and interacts with the BMS / EMS to trigger corresponding measures. This output format facilitates automated response and manual intervention by the operations and maintenance system.

[0083] By using weighted fusion and root cause analysis, we can not only identify the type of fault, but also locate the source of the fault and the propagation path, providing precise guidance for maintenance decisions, significantly improving maintenance efficiency and reducing the risk of incorrect or missed repairs.

[0084] Example 2: This example selects an energy storage power station in a certain area, with 8 clusters, each cluster having 12 modules, and each module having 16 individual units. The data acquisition cycle is 1 second, with continuous sampling for 30 days, resulting in 2.59 × 10^6 samples. Acquisition parameters include: voltage, module internal resistance, cluster current / voltage, module maximum / minimum temperature and temperature difference, SOC, cycle count, etc. (consistent with Example 1). Historical events are labeled in three categories: Event A (Overcharging with Temperature Rise): Abnormal charging power limitation is triggered, resulting in a coupling of increased voltage deviation and increased temperature deviation; Event B (Connection Failure): Poor terminal contact leads to an increase in the internal resistance difference between adjacent modules, accompanied by localized temperature rise; Event C (Aging-driven): After long-term cycling, the SOC decay rate and internal resistance growth rate increase, accompanied by a slight temperature rise.

[0085] Annotation set: 37 cases in category A, 22 cases in category B, 41 cases in category C, the rest are normal operating conditions. Training / verification / testing ratio is 6:2:2, randomly divided (segmented by event to avoid leakage).

[0086] The method of this invention is: multi-scale multi-layer network in Example 1 + fault association mapping weighting + improved sand cat swarm optimization algorithm (three improvements: uniform initialization + dynamic path search (spiral) + bidirectional optimization / backward learning, triggering association with SEI / electrolyte mechanism).

[0087] Baseline 1 (BP): Standard BP neural network, without associated mapping and without swarm optimization.

[0088] Baseline 2 (PSO+NN): A PSO-optimized NN with the same structure and no associated mappings.

[0089] Baseline 3 (Original Sand Cat Swarm Optimization Algorithm + NN): The three improved sand cat swarm optimization algorithms of this invention are not introduced, nor is there a physical triggering strategy.

[0090] All methods share the same input features and labels, and output 5 types of fault / normal probabilities; evaluation metrics: overall accuracy, macro F1, AUC, average warning lead time (h), false alarm rate (%), convergence algebra, and root cause accuracy.

[0091] Table 1 Overall Results

[0092] The results show that the present invention outperforms the baseline in terms of accuracy, F1, AUC, early warning lead time, and false alarm rate; it has a significantly smaller convergence algebra than the PSO / original sand cat swarm optimization algorithm; and it also has high source tracing accuracy (which can directly support maintenance suggestion linkage).

[0093] Review classic events that occurred during testing: Event A: Overcharge-Temperature Anomaly Coupling Phenomenon: During the charging phase, both the individual cell voltage deviation and the module temperature deviation increase simultaneously.

[0094] The invention diagnoses the following: 18 minutes after the abnormal start point, it determines "Main: Overcharge (0.87), Minor: Temperature Abnormal (0.62)" and provides an emergency level and power limitation suggestion.

[0095] In comparison, BP, PSO+NN, and the original sand cat swarm optimization algorithm took 72 minutes, 41 minutes, and 33 minutes respectively to recognize the target.

[0096] Advantages: Increases the lead time for early warning by 3-4 hours, avoiding over-temperature protection triggering.

[0097] Event B: Connection Failure—Localized High Temperature Phenomenon: The difference in internal resistance between adjacent modules suddenly increases, causing local temperature rise.

[0098] The invention diagnoses the following: within 11 minutes, the fault is identified as "Main: Connection Fault" and points to a specific module terminal.

[0099] Comparison: BP misjudged it as a temperature anomaly, and both PSO+NN and the original sand cat swarm optimization algorithm required more than 30 minutes.

[0100] Advantages: The accuracy of traceability is improved by about 27%, and the scope of maintenance is more focused.

[0101] Event C: Aging-driven: Phenomenon: Increased SOC decay rate, increased internal resistance, accompanied by a slight temperature rise.

[0102] The invention diagnoses "Main: Aging (0.83)" and provides recommendations for derating and balanced maintenance.

[0103] Comparison: BP and PSO+NN misjudged temperature anomalies in some periods, and the original sand cat swarm optimization algorithm got stuck in a local optimum.

[0104] Advantages: False alarm rate reduced to 2.8%, effectively distinguishing between aging and thermal anomalies.

[0105] The following section demonstrates the necessity of the three improvements through ablation testing. Using this invention as a baseline, the improvement strategies were disabled one by one, and the test set results are as follows: Table 2 Ablation Test Table

[0106] All three are indispensable: uniform initialization mainly affects the convergence speed and lead time; dynamic path search improves both F1 score and lead time; bidirectional optimization is particularly crucial for aging class identification.

[0107] This invention demonstrates significant effectiveness in practical power plant applications: through multi-fault coupling mapping and improved optimization strategies, the average early warning lead time is increased by more than 2 hours, enabling the EMS to limit power and conduct thermal management in a timely manner; fault root cause localization is more focused, reducing the average number of manual inspection points by 50%, significantly improving maintenance efficiency; in high-coupling-risk scenarios, the system can automatically provide clear suggestions of "immediately offline inspection or isolation," thereby reducing the risk of secondary accidents; at the same time, compared with traditional PSO and original sandcat swarm optimization algorithms, the convergence algebra of this method is reduced by approximately 55% and 57%, respectively, significantly shortening the model online and parameter tuning cycle, ensuring stable and reliable diagnostic results.

[0108] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-type fault diagnosis method for electrochemical energy storage, characterized in that, The method comprises the following steps: S100: Collecting battery module voltage, charging and discharging current, module temperature distribution, internal resistance change and state of charge characteristic data by a battery management system, constructing a multi-fault coupling feature space, and establishing a correlation mapping relationship between multiple types of faults; S200: Based on the correlation mapping relationship, the sand cat swarm optimization algorithm is initialized by using a uniform initialization sequence, the sequence dimension is set to correspond to the number of battery modules; a dynamic path search strategy is introduced, the spiral search parameters are set according to the overcharge-temperature anomaly mapping; a bidirectional optimization strategy is introduced, the reverse learning trigger condition is adjusted combined with the SEI membrane damage mechanism and the electrolyte decomposition mechanism; S300: Constructing a multi-scale multi-layer neural network, designing network connection weights for coupling fault identification based on the correlation mapping relationship, and fusing single-body, module, cluster and system level features; S400: Optimizing network parameters by using the improved sand cat swarm optimization algorithm, inputting multi-fault coupling characteristic data into the neural network, and performing coupling fault tracing and correlation diagnosis according to the fault correlation mapping.

2. The multi-type fault diagnosis method for electrochemical energy storage according to claim 1, characterized in that, The step of constructing a multi-fault coupling feature space comprises: Divide the battery module voltage data into charging section, discharging section and static section according to time window, calculate the voltage variance and voltage gradient in each section; Perform spatial difference processing on the module temperature distribution data to obtain a temperature gradient matrix and a maximum temperature difference value; Extract the first derivative and second derivative of the internal resistance change as the internal resistance dynamic characteristics; Cross-combine the state of charge with voltage, temperature and internal resistance characteristics in multiple dimensions to form a multi-fault coupling feature vector.

3. The multi-type fault diagnosis method for electrochemical energy storage according to claim 1, characterized in that, The step of establishing a correlation mapping relationship between multiple types of faults comprises: Set the voltage anomaly threshold and the temperature anomaly threshold, establish the coupling weight matrix of the overcharge-temperature anomaly correlation mapping; Calculate the internal resistance difference value between adjacent modules, mark as connection abnormality when the difference value exceeds the preset threshold, and establish the linear mapping function of the connection fault-local high temperature correlation mapping; Based on the ratio relationship between the state of charge attenuation rate and the internal resistance growth rate, construct the weight matrix of the aging-internal resistance change correlation mapping.

4. The multi-type fault diagnosis method for electrochemical energy storage of claim 1, wherein, The step of initializing the sand cat swarm optimization algorithm by using a uniform initialization sequence comprises: Determine the boundary range of the search space according to the overcharge-temperature anomaly correlation mapping, the connection fault-local high temperature correlation mapping and the aging-internal resistance change correlation mapping; Generate a low-difference quasi-random sequence as a uniform initialization sequence, the sequence dimension matches the number of battery modules; Distribute the sampling points of the initialization sequence preferentially in the fault coupling sensitive area determined by the correlation mapping relationship; Adjust the initial density of population individuals in each fault coupling space according to the weight distribution of the correlation mapping matrix.

5. The multi-type fault diagnosis method for electrochemical energy storage according to claim 1, characterized in that, The dynamic path search strategy comprises: Based on the coupling strength in the overcharge-temperature anomaly correlation mapping, dynamically adjust the pitch parameter and search trajectory of the spiral search; According to the gradient direction of the connection fault-local high temperature correlation mapping, set the priority search path of the spiral search; When the aging-internal resistance change correlation mapping indicates that the fault coupling degree is deepened, reduce the spiral search radius and increase the search accuracy.

6. The multi-type fault diagnosis method for electrochemical energy storage of claim 1, wherein, The bidirectional optimization strategy comprises: The occurrence timing of SEI film damage and electrolyte decomposition is predicted by using an aging-internal resistance change correlation mapping relationship, and a trigger threshold of reverse learning is set; Based on the SEI film thickness growth rate and the electrolyte concentration change rate, the trigger probability of reverse learning is calculated; When the correlation mapping relationship indicates that the risk of multiple fault coupling increases, the reverse learning frequency is increased and the search direction weight is adjusted.

7. The multi-type fault diagnosis method for electrochemical energy storage of claim 1, wherein, The step of constructing a multi-scale multi-layer neural network comprises: The input layer nodes are set to correspond one-to-one to the parameters collected by the battery management system, including the battery voltage and internal resistance at the single cell level, the highest and lowest temperature and temperature difference at the module level, the cluster voltage and cluster current at the cluster level, and the total output power and cycle number at the system level; According to the coupling weight coefficient in the correlation mapping relationship, the connection weight between the input layer and the hidden layer is initialized, and the overcharge-temperature anomaly mapping weight, the connection fault-local high temperature mapping weight, and the aging-internal resistance change mapping weight are respectively assigned to the corresponding network connection; The number of hidden layer nodes is adaptively adjusted according to the capacity scale of the energy storage system, and a Sigmoid activation function is used for nonlinear transformation; The output layer nodes are configured to correspond to multiple fault modes of the energy storage system, including overcharge fault, overdischarge fault, battery aging fault, battery connection fault and temperature anomaly fault.

8. The multi-type fault diagnosis method for electrochemical energy storage of claim 1, wherein, The step of optimizing network parameters by using the improved sand cat swarm optimization algorithm comprises: The weight matrix and threshold vector of the multi-layer neural network are encoded into the position vector of the sand cat individual, and the encoding length is equal to the total number of all connection weights and node thresholds in the network; The fitness function is set as the weighted combination of fault classification error and prediction accuracy of neural network output, and the fitness value of each sand cat individual under the corresponding network parameters is calculated; According to the size relationship of the control parameter R, it is judged whether the sand cat executes the search prey strategy or the hunting prey strategy, when the absolute value of R is less than or equal to 1, the local search is executed to update the network weight, and when the absolute value of R is greater than 1, the global search is executed to explore new parameter space; The position of the sand cat is updated by iteration, and the weight and threshold parameters of the neural network are gradually optimized until the preset iteration number or the fitness convergence condition is reached.

9. The multi-type fault diagnosis method for electrochemical energy storage of claim 1, wherein, The coupling fault tracing and correlation diagnosis according to the fault correlation mapping comprises: The normalized multi-fault coupling feature data is input into the optimized neural network, and the output probability of each fault type is calculated by forward propagation; Based on the weight matrix in the fault correlation mapping relationship, the fault probability output by the neural network is weighted and fused to identify the main fault type and the secondary fault type; According to the cause-and-effect chain in the correlation mapping relationship, the root cause and propagation path of the fault are determined by tracing upwards from the detected fault result; Combined with the coupling strength coefficient in the correlation mapping matrix, the mutual influence degree between each type of fault is evaluated, and the fault level and maintenance suggestion are output.

Images