A mine area battery product iteration optimization method and system based on fault big data

Abstract

The application provides a mine area battery product iterative optimization method and system based on fault big data, relates to the cross technical field of mine area battery health management and artificial intelligence, and solves the technical problems of fault data scarcity, poor physical consistency of diagnosis models, and prediction results unable to drive product iteration. The method comprises the following steps: constructing an electrochemical-thermal-aging coupled physical model and generating a fault evolution sequence, and then generating a fault simulation data set through a generative adversarial network; performing multi-modal feature extraction on real and synthetic data, and screening indirect health factors; embedding the physical model into a physical information neural network to obtain a diagnosis model fused with physical constraints through training, and outputting health state estimation and fault warning; determining key features and their contribution degrees through interpretable AI attribution analysis, and combining with working condition backtracking to determine the product optimization direction; converting the optimization scheme into an adjustment instruction, and feeding back the performance data of the optimized product to update the physical model, thereby forming a closed-loop iteration.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of battery health management in mining areas and artificial intelligence. Specifically, it is a method and system for iterative optimization of battery products in mining areas based on big data on faults. Background Technology

[0002] Electrified equipment in mining areas (such as electric mining trucks, underground transport vehicles, and loaders) places extremely high demands on the safety, reliability, and lifespan of power batteries. However, the harsh mining environment, with its steep slopes under heavy loads, alternating high and low temperatures, strong vibrations, and high dust levels, makes batteries prone to accelerated capacity decay, increased internal resistance, localized overheating, and even thermal runaway during use. To ensure operational safety in mining areas and reduce maintenance costs, battery state of health (SOH) estimation and early fault warning technologies have become a research hotspot. Existing technologies are mainly divided into three categories: First, there are methods based on electrochemical-thermal coupled physical models, which describe the internal aging mechanism of batteries by constructing a system of differential equations. This method has clear physical meaning, but it is computationally complex, difficult to calibrate parameters, and difficult to adapt to the complex and ever-changing working conditions in mining areas. Second, there are purely data-driven methods, such as support vector machines, recurrent neural networks, and Transformers, which train models with a large amount of historical data to achieve state estimation and fault prediction. However, these methods heavily rely on high-quality and diverse fault data. The probability of serious faults (such as thermal runaway and internal short circuits) occurring in mining batteries during actual operation is extremely low, resulting in a severe scarcity of real fault samples. Model training is prone to overfitting or underfitting, and the generalization ability drops sharply under unseen working conditions. The prediction results may violate basic electrochemical laws. Third, there are hybrid methods that combine physical models with data-driven approaches. However, existing hybrid methods mostly use simple series-parallel structures, and there is a lack of deep two-way information interaction between physical constraints and data models. This may lead to the model outputting physically unreasonable prediction results under extreme working conditions.

[0003] Furthermore, existing battery diagnostic models generally lack interpretability, outputting only health status values ​​or failure probabilities. They fail to reveal the key characteristics leading to failures and their contribution, and it is even more difficult to establish a correlation between diagnostic results and battery product design parameters, thus failing to effectively guide product iteration and optimization of batteries in mining areas. Therefore, how to solve the technical problems of scarce battery failure data in mining areas, poor physical consistency of diagnostic models, and uninterpretable prediction results that cannot drive closed-loop product iteration has become a bottleneck that urgently needs to be overcome in this field. Summary of the Invention

[0004] This application provides a method and system for iterative optimization of mining battery products based on big data of faults. It solves the technical problems in the prior art, such as the scarcity of real fault data of mining batteries leading to insufficient training samples for diagnostic models, the tendency of pure data-driven models to deviate from physical laws in prediction results under extreme and complex working conditions, resulting in poor generalization ability, and the lack of interpretability of fault diagnosis results, which makes it difficult to effectively guide product design improvement and closed-loop iteration.

[0005] To achieve the above objectives, this application adopts the following technical solution: Firstly, a method for iterative optimization of battery products in mining areas based on big data on faults is provided, including: Battery operation data is collected using multi-source sensors, and an aging coupled physical model is constructed based on electrochemical and thermal theories. Based on the aging coupled physical model, the numerical solution method is used to solve the problem step by step with a set time step by preset mining area working conditions and fault triggering conditions to obtain the evolution trajectory of the internal state variables of the battery. The time series of the state variables from the initial state to the point where the fault threshold is exceeded is recorded to obtain the fault evolution sequence. The fault evolution sequence is input into a generative adversarial network to generate a fault simulation dataset; Multimodal features are extracted from the battery operation data and the fault simulation dataset, and indirect health factors related to battery health status are screened from the multimodal features using the Pearson correlation coefficient. The aging coupled physical model is embedded into a physical information neural network, and the indirect health factor is used as input to train a battery state diagnosis model that integrates physical constraints, and outputs battery health state estimation and fault warning signal. The output of the battery state diagnostic model is analyzed using interpretable AI methods to identify key features that cause the failure and their contribution, and product optimization directions are determined by combining the backtesting of operating conditions. The optimization direction is transformed into product design parameter adjustment instructions, and the optimized product performance data is fed back to the aging coupled physical model to update the model parameters, forming a closed-loop iteration.

[0006] Based on the above technical solutions, this application provides an iterative optimization method for mining battery products based on fault big data. By constructing an aging coupled physical model and using generative adversarial networks to synthesize fault simulation datasets, the problem of scarce real fault data in mining areas is effectively solved, providing sufficient and diverse samples for model training. Multimodal feature extraction and Pearson correlation coefficient screening of indirect health factors are employed, enabling the accurate extraction of low-redundancy features highly correlated with battery health status from massive amounts of raw data, reducing the dimensionality of model input and improving diagnostic efficiency. Next, the aging coupled physical model is embedded into a physical information neural network, allowing the diagnostic model to strictly follow the electrochemical-thermal-aging physical laws while being data-driven, significantly enhancing the model's generalization ability and prediction reliability under extreme conditions. Then, interpretable AI methods are used to perform attribution analysis on the model output, and combined with operational condition backtracking to determine the product optimization direction, providing clear and quantifiable basis for product design improvement. Finally, by feeding optimization instructions back to the physical model and forming a closed-loop iteration, the battery product can continuously evolve with the accumulation of operational data, thereby systematically improving the safety, reliability, and economy of mining batteries.

[0007] Furthermore, the construction process of the aging coupled physical model includes: A system of differential-algebraic equations is constructed by coupling state variables from an electrochemical sub-model, a thermal sub-model, and an aging sub-model; among which, The electrochemical sub-model adopts a single-particle model or a pseudo-two-dimensional model to describe the diffusion process of lithium ions in the solid phase of positive and negative electrode materials and the ion transport process in the electrolyte phase. The thermal sub-model uses the lumped parameter method or the finite element method to describe the energy balance during the heat generation and dissipation process inside the battery. The aging sub-model includes a solid electrolyte interface film growth sub-model, a lithium dendrite formation sub-model, and an active material loss sub-model, which are used to describe the capacity decay and internal resistance increase of the battery due to side reactions, mechanical stress, and thermal effects during cycling. The solid phase concentration and overpotential output by the electrochemical sub-model and the temperature output by the thermal sub-model are used as inputs to the aging sub-model. At the same time, the SEI film thickness and internal resistance growth output by the aging sub-model are fed back to the electrochemical sub-model and the thermal sub-model to form a closed-loop coupling, thus obtaining the aging coupled physical model.

[0008] Furthermore, the generation of fault evolution sequences under different operating conditions based on the aging coupled physical model includes: A set of mining area operating condition parameters is set, including ambient temperature, discharge rate, charge rate, depth of discharge, vibration acceleration amplitude and frequency; For each combination of operating parameters, preset fault triggering conditions are injected into the aging coupled physical model. The fault triggering conditions include diaphragm puncture, electrolyte drying, local overheating, and current collector breakage. Starting with the initial value of battery health state SOH=100%, the Runge-Kutta method is used to numerically solve the differential algebraic equations, and the evolution trajectory of the battery's internal state variables is calculated step by step with a set time step. When any state variable exceeds the corresponding fault threshold, the state variable sequence of several time steps before the fault time is recorded, including voltage, current, temperature, solid phase concentration distribution, SEI film thickness and lithium dendrite length, forming a fault evolution sequence from healthy state to fault state. Iterate through all combinations of operating parameters and fault triggering conditions to generate an original fault evolution library containing several fault evolution sequences.

[0009] Furthermore, the specific process for generating the fault simulation dataset includes: A conditional generative adversarial network (GAN) is used as the fault data generator, and the GAN includes a generator G and a discriminator D; wherein... The generator G takes a random noise vector and a condition vector as input, and the condition vector includes a working condition type code, a fault type code, and a degradation stage code. The generator G consists of a fully connected layer, a batch normalization layer, a LeakyReLU activation layer, and an upsampled convolutional layer connected in sequence, and outputs a synthesized fault data sequence. The discriminator D consists of a convolutional layer, a dropout layer, a fully connected layer, and a sigmoid activation layer connected in sequence, and its output is the probability that the input data comes from a real evolutionary sequence. The generator G and discriminator D are trained alternately until convergence. The trained generator G is then sampled in batches in the conditional vector space to generate several synthetic fault data sequences, which are then merged with the original fault evolution library to form a fault simulation dataset.

[0010] Furthermore, the multimodal feature extraction of the battery operation data and the fault simulation dataset includes: For each time series in the battery operation data and the fault simulation dataset, perform feature extraction operations in the following three dimensions: First, calculate the time-domain characteristic statistics within each charge-discharge cycle, including mean, standard deviation, peak value, peak-to-peak value, skewness, kurtosis, root mean square value, waveform factor, impulse factor, and margin factor. Second, perform a fast Fourier transform on the voltage and current sequences to extract frequency domain features, which include the centroid frequency, mean square frequency, frequency variance, and amplitude ratio of the first 5 harmonics in the spectrum. Third, wavelet packet decomposition is performed on the temperature sequence to extract time-frequency domain features, which include the energy proportion and energy entropy of each frequency band node; The time-domain feature statistics, frequency-domain features, and time-frequency-domain features are concatenated to form the original multimodal feature vector, thus obtaining the multimodal features.

[0011] Furthermore, the step of using the Pearson correlation coefficient to screen out indirect health factors related to battery health status from the multimodal features includes: For each feature in the original multimodal feature vector, calculate the Person correlation coefficient between the feature and the battery health state label SOH, using the following formula: Where N is the total number of samples, Let i be the feature value of the j-th sample. Let be the sample mean of feature i. Let j be the battery health status label for the j-th sample. The sample mean of the health status label. Let i be the correlation coefficient between feature i and the battery health status label; Features with correlation coefficients greater than a preset threshold are selected to form a candidate health factor set; Calculate the variance inflation factor of the features in the candidate health factor set, and remove features whose variance inflation factor is greater than a preset value to obtain the final indirect health factor set.

[0012] Furthermore, the battery state diagnostic model integrating physical constraints adopts a parallel branch physical information neural network architecture, including an input layer, a convolutional neural network branch, a Transformer branch, a cross-branch sensing modulator, a physical constraint embedding layer, and an output layer; wherein, The input layer receives a time series matrix composed of the indirect health factors. The convolutional neural network branch is used to extract features from the time series matrix through multiple convolutions and output a first feature tensor. The Transformer branch is used to extract features from the time series matrix through a multi-head self-attention mechanism and output a second feature tensor. The cross-branch perceptual modulator includes a first modulation unit and a second modulation unit. The first modulation unit receives the first feature tensor, maps the first feature tensor to an attention bias vector through a fully connected layer, and adds the attention bias vector to the attention weight matrix of the multi-head attention layer in the Transformer branch. By performing softmax normalization on the attention weights after adding the bias and calculating the weighted sum, a modulated second feature tensor is obtained. The second modulation unit receives the second feature tensor, generates convolution kernel modulation coefficients through a gating mechanism, and multiplies the modulation coefficients with the weights of each convolution kernel in the convolutional neural network. By applying the modulated convolution kernels to the input data for convolution operations, a modulated first feature tensor is obtained. The physical constraint embedding layer is used to embed the aging coupled physical model as a physical constraint into the battery state diagnosis model. During the model training process, the physical consistency residual is added as a regularization term to the total loss function to achieve consistency constraint between the prediction results of the battery state diagnosis model and the physical evolution law. The output layer receives the modulated first feature tensor and the modulated second feature tensor, as well as the physical consistency residual output by the physical constraint embedding layer. The modulated first feature tensor and the modulated second feature tensor are concatenated along the channel dimension and then passed through the global average pooling layer and the fully connected layer in sequence to output the battery health status estimate and the fault warning signal.

[0013] Furthermore, the total training loss function of the battery state diagnostic model is: ;in, The mean squared error loss for health status estimation. For cross-entropy loss in fault warning, For physical consistency loss, This is a cross-branch consistency loss used to constrain the consistency of the distributions of the first and second feature tensors in the latent space. , , These are the corresponding weighting coefficients.

[0014] Furthermore, the workflow of the physical constraint embedding layer is as follows: The modulated first feature tensor and the modulated second feature tensor are concatenated along the channel dimension and then input into a decoder composed of a fully connected layer network. The decoder outputs the predicted internal state vector of the battery. The predicted battery internal state vector is input into the aging coupled physical model, and the mean square value of the difference between the output vector obtained by the physical model recursion and the predicted battery internal state vector is calculated to obtain the physical consistency residual for the next time step. The physical consistency residual, as part of the physical constraint loss term, is used in the training process through automatic differentiation and backpropagation to force the prediction results of the battery state diagnosis model to meet the dynamic evolution law of the aging coupled physical model.

[0015] Furthermore, the attribution analysis of the output of the battery state diagnostic model using interpretable AI methods to identify key features leading to the fault and their contribution includes: The SHAP value method is used for attribution analysis. The fault warning signal output by the battery state diagnosis model is used as the target variable. For each sample to be explained, the SHAP value of the i-th indirect health factor is calculated. Sort all samples by absolute value from largest to smallest, and select the top k features whose cumulative contribution rate reaches a preset percentage as key features. The contribution of each key feature is defined as the proportion of the SHAP value to the sum of the absolute values ​​of all feature SHAP values; where k is a positive integer greater than 1. The battery health status estimate output by the battery status diagnosis model is used to assist in verifying the rationality of the attribution results: when key features change, it is determined whether the trend of the battery health status estimate is consistent with the sign of the contribution of the key features; if they are consistent, the attribution is marked as reasonable.

[0016] Furthermore, the product optimization direction is determined based on the key characteristics of the fault, the contribution of the key characteristics, and operational condition backtracking, specifically including: Based on the type of the key features, a mapping table between features and product design parameters is established; the mapping table records the upstream design parameters corresponding to each indirect health factor. Combined with the working condition retrospective results, the working condition retrospective results refer to retrieving the mining area vehicle operation logs within the corresponding time period based on the timestamp of the time when the fault occurred, and extracting the ambient temperature, road slope, load capacity, charging pile power, and continuous operation duration as working condition tags. The contribution of the key features is used as a weight to perform weighted voting on each design parameter in the mapping table, and the design parameters whose total contribution exceeds a preset ratio are selected as priority optimization parameters. Based on the operating condition labels, the critical conditions for the occurrence of the fault are determined. If the frequency of a certain operating condition label in all fault samples is greater than a preset frequency threshold, the boundary conditions corresponding to the operating condition label are used as optimization constraints to generate a product optimization scheme that includes the priority optimization parameters, optimization parameter types, optimization directions, optimization magnitude suggestions, and verification operating conditions.

[0017] Furthermore, the step of converting the optimization direction into product design parameter adjustment instructions includes: The priority optimization parameters, parameter types, optimization directions, suggested optimization magnitudes, and verification conditions in the product optimization plan are encapsulated into structured adjustment instructions; the data format of the structured adjustment instructions is JSON, containing the following fields: The parameter identifier field is used to uniquely identify the design parameter to be adjusted. Adjust the type field, with possible values ​​including "increase", "decrease", and "switch"; Adjust the step size field to use relative percentages or absolute values; The constraint field is used to record the boundary conditions in the verification condition; The safety limit field is used to record the upper and lower boundaries that the adjusted parameters must not exceed; The structured adjustment instructions are sent to the edge side of the mining area battery management system via OTA control and download technology, and simultaneously synchronized to the parameter configuration file of the aging coupled physical model; Perform verification simulation in the aging coupled physical model: substitute the adjusted design parameters into the model, and rerun the fault evolution simulation under the verification conditions; if the simulation results show that no similar faults are triggered, mark the adjustment instruction as "verified" and submit it to the product design change system; if the fault is still triggered in the verification simulation, trigger the rollback mechanism, restore the design parameters to the values ​​before adjustment, and push them to the manual analysis queue. After receiving the verified adjustment instruction, the product design change system generates a product design parameter modification notification and triggers the remote parameter update process of the on-site battery management system in the mining area.

[0018] Secondly, this application provides an iterative optimization system for battery products in mining areas based on fault big data, including: a physical modeling module, a fault data synthesis module, a feature extraction module, a model diagnosis module, an attribution analysis module, and a closed-loop optimization module; wherein, The physical modeling module is used to collect battery operating data using multi-source sensors and to construct an aging coupled physical model based on electrochemical and thermal theories. The fault data synthesis module is used to calculate the evolution trajectory of the internal state variables of the battery by using a numerical solution method with a set time step based on the aging coupled physical model, through preset mining area working condition parameters and fault triggering conditions, and recording the time series of the state variables from the initial stage to when they exceed the fault threshold, thereby obtaining the fault evolution sequence. The fault evolution sequence is then input into a generative adversarial network to generate a fault simulation dataset. The feature extraction module is used to perform multimodal feature extraction on the battery operation data and the fault simulation dataset, and to use the Pearson correlation coefficient to screen out indirect health factors related to the battery health status from the multimodal features. The model diagnostic module is used to embed the aging coupled physical model into a physical information neural network, and train a battery state diagnostic model that integrates physical constraints with the indirect health factor as input, and output battery health state estimation and fault warning signal. The attribution analysis module is used to perform attribution analysis on the output of the battery state diagnostic model using interpretable AI methods, identify the key features that lead to the failure and the contribution of the key features, and determine the product optimization direction by combining the backtracking of operating conditions. The closed-loop optimization module is used to convert the optimization direction into product design parameter adjustment instructions, and to feed back the optimized product performance data to the aging coupled physical model to update the model parameters, forming a closed-loop iteration.

[0019] Compared with the prior art, the beneficial effects of this application are: First, this application fundamentally solves the problem of scarce real-world battery fault samples in mining areas by constructing a closed-loop coupled electrochemical-thermal-aging multiphysics model and using conditional generative adversarial networks to synthesize fault data. This method not only simulates the complete evolution of various faults such as separator puncture and lithium dendrite growth under different operating conditions, but also accurately identifies indirect health factors highly correlated with health status through multimodal feature extraction in the time, frequency, and time-frequency domains and Pearson correlation coefficient screening, significantly improving the training efficiency and generalization ability of subsequent diagnostic models. Simultaneously, by embedding the physical model into a parallel branched physical information neural network, convolutional neural networks and Transformer branches are used to capture local transient features and global aging trends, respectively. A cross-branch perceptual modulator enables bidirectional information interaction, allowing the model to strictly adhere to physical laws while being data-driven, thus maintaining high-precision health status estimation and early fault warning even under extreme mining conditions.

[0020] Secondly, this application introduces the interpretable SHAP value method to perform attribution analysis on indirect health factors, targeting fault warning signals. It combines operational condition backtracking to identify key characteristics leading to faults and their contribution, and establishes a mapping relationship between these characteristics and design parameters, achieving a leap from "black-box prediction" to "white-box attribution." Weighted voting is used to select priority optimization parameters, and operational condition labels are used to determine optimization constraints, generating a product optimization plan that includes specific adjustment directions, step sizes, and verification operational conditions. Furthermore, the optimization plan is encapsulated as structured instructions and distributed to the edge via OTA. Simulation verification is performed in the physical model. Only instructions that pass verification can be submitted to the product design change system and trigger remote updates, forming a closed-loop iterative mechanism that effectively avoids the safety risks associated with blind adjustments.

[0021] Finally, this application feeds back the optimized performance data to the aging coupled physical model, continuously updating the model parameters and enabling the system to continuously evolve as mining operation data accumulates. This closed-loop iterative approach not only significantly shortens the cycle from problem discovery to design improvement for battery products and reduces on-site testing costs, but also significantly improves the safety, reliability, and economy of mining batteries in complex and variable environments, providing complete technical support for the intelligent operation and maintenance and continuous product upgrades of electrified equipment in mining areas. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 A system architecture diagram of a mining area battery product iterative optimization system based on fault big data is provided for embodiments of this application; Figure 2 A flowchart illustrating an iterative optimization method for battery products in mining areas based on fault big data, provided as an embodiment of this application; Figure 3 A flowchart illustrating another method for iterative optimization of battery products in mining areas based on fault big data, provided as an embodiment of this application; Figure 4 A flowchart illustrating another method for iterative optimization of battery products in mining areas based on fault big data, provided as an embodiment of this application; Figure 5 This is a flowchart illustrating another method for iterative optimization of battery products in mining areas based on fault big data, provided as an embodiment of this application. Detailed Implementation

[0024] In the description of this application, unless otherwise stated, " / " means "or," for example, A / B can mean A or B. The "and / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone. Furthermore, "at least one" means one or more, and "multiple" means two or more. The terms "first," "second," etc., do not limit the quantity or order of execution, and "first," "second," etc., do not necessarily imply differences.

[0025] It should be noted that, in this application, the terms "exemplary" or "for example" are used to indicate that something is being described as an example, illustration, or illustration. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or design solutions. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0026] The iterative optimization method for battery products in mining areas based on fault big data provided in this application embodiment can be applied to, for example... Figure 1 The system shown is an iterative optimization system for battery products in mining areas based on big data of faults. Figure 1 As shown, the communication system includes: a physical modeling module, a fault data synthesis module, a feature extraction module, a model diagnosis module, an attribution analysis module, and a closed-loop optimization module; among which, The physical modeling module is used to collect battery operating data using multi-source sensors and to build an aging coupled physical model based on electrochemical and thermal theories. The fault data synthesis module is used to generate fault evolution sequences under different working conditions based on the aging coupled physical model, and input the fault evolution sequences into the generative adversarial network to generate a fault simulation dataset. The feature extraction module is used to extract multimodal features from battery operation data and fault simulation datasets, and uses Pearson correlation coefficient to screen out indirect health factors related to battery health status from multimodal features. The model diagnostic module is used to embed the aging coupled physical model into the physical information neural network, take indirect health factors as input, train the battery state diagnostic model that integrates physical constraints, and output battery health state estimation and fault warning signal. The attribution analysis module is used to perform attribution analysis on the output of the battery state diagnostic model using interpretable AI methods, identify the key features that lead to the failure and the contribution of the key features, and determine the direction of product optimization by combining the backtracking of operating conditions. The closed-loop optimization module is used to convert the optimization direction into product design parameter adjustment instructions, and to feed back the optimized product performance data to the aging coupled physical model to update the model parameters, forming a closed-loop iteration.

[0027] To address the technical problems of inaccurate fault diagnosis and low iterative optimization efficiency in existing mining battery products under complex operating conditions, such as... Figure 2 As shown in the figure, this application provides an iterative optimization method for battery products in mining areas based on fault big data. The method includes the following steps.

[0028] S1 uses multi-source sensors to collect battery operation data and constructs an aging coupled physical model based on electrochemical and thermal theories.

[0029] The aging coupled physical model refers to a mathematical and physical model that simultaneously describes the coupling relationship between the electrochemical performance degradation and thermal behavior evolution of a battery during charge-discharge cycles. This model aims to reveal the battery aging mechanism and quantify the positive feedback effect between internal side reactions such as lithium deposition, electrolyte decomposition, and temperature rise, thereby providing physical constraints and prior knowledge for the generation of subsequent fault evolution sequences. This model can be constructed using various techniques, such as: first, simplifying the electrochemical pseudo-two-dimensional model to obtain a single-particle model or equivalent circuit model to describe the solid-phase lithium concentration distribution and potential changes; second, coupling a lumped-parameter thermal model or a three-dimensional heat conduction model to calculate the internal temperature field of the battery; third, introducing the Arrhenius equation to describe the effect of temperature on the reaction rate and aging rate; and fourth, identifying and correcting the model parameters using experimental data such as capacity decay curves and temperature rise curves from cyclic aging tests. Furthermore, reduction-order modeling techniques can be used to transform complex partial differential equations into ordinary differential equations to improve computational efficiency. The construction of this model is the cornerstone for realizing the integration of physics and data, and can make up for the lack of physical interpretability in purely data-driven methods.

[0030] S2 generates fault evolution sequences under different operating conditions based on the aging coupled physical model, and inputs the fault evolution sequences into a generative adversarial network to generate a fault simulation dataset.

[0031] The fault evolution sequence refers to the time-varying trajectory of fault characteristic parameters obtained from aging-coupled physical model simulations under specific operating conditions such as high temperature, high-rate charge / discharge, and overcharge / over-discharge. Examples include internal resistance growth sequences, capacity decay sequences, and local hotspot temperature sequences. A Generative Adversarial Network (GAN) consists of a generator and a discriminator in its core structure. These are trained through a game-like process, enabling the generator to produce synthetic data highly similar to the distribution of real data. The fault evolution sequence is input into the GAN because actual battery fault data from mining areas is scarce and costly to obtain, while the evolution sequences generated by the physical model, although conforming to physical laws, lack diversity. The GAN can learn the potential distribution of the physical evolution sequence and, based on this, generate a large amount of statistically similar but detailed fault simulation data, thereby expanding the diversity and coverage of the dataset. This GAN requires training on two types of data: the first is the fault evolution sequence from the aging-coupled physical model as a substitute for real samples; the second is random noise vectors, which the generator maps into synthetic sequences. During training, the generator attempts to deceive the discriminator, while the discriminator strives to distinguish between real and synthetic sequences. This process iterates alternately until Nash equilibrium is reached, thereby expanding the fault dataset, addressing the small sample size problem, and improving the generalization ability and robustness of subsequent diagnostic models.

[0032] S3 extracts multimodal features from battery operation data and fault simulation datasets, and uses Pearson correlation coefficient to screen out indirect health factors related to battery health status from the multimodal features.

[0033] The simultaneous use of battery operating data and fault simulation datasets is chosen because battery operating data, collected from real sensors, is highly reliable but has a limited sample size and incomplete coverage of operating conditions. The fault simulation dataset, generated by a generative adversarial network, while exhibiting some model bias, covers various extreme and rare operating conditions deduced by the physical model. Combining these two datasets balances realism and comprehensiveness, reducing the risk of overfitting the model to a single data source. Indirect health factors refer to measurable features strongly correlated with battery health, extracted from observable multimodal characteristics such as voltage curve slope, temperature change rate, charging time interval, and current impulse response, rather than directly measuring battery health states like remaining capacity or internal resistance. Pearson correlation coefficients are used for factor selection to quantify the linear correlation strength between each candidate feature and battery health state, thereby eliminating redundant and noisy features and retaining the most strongly correlated subset. The Pearson correlation coefficient ranges from -1 to +1, with a value closer to one indicating a stronger linear correlation. This selection process reduces the input dimensionality of subsequent models, decreases computational overhead, and avoids model instability caused by multicollinearity.

[0034] S4 embeds the aging coupled physical model into the physical information neural network, takes indirect health factors as input, trains to obtain a battery state diagnosis model that integrates physical constraints, and outputs battery health state estimation and fault warning signals.

[0035] Among them, the Physical-Informed Neural Network (PIN) is based on the idea of ​​embedding physical laws into the loss function of the neural network in the form of partial differential equations or algebraic constraints. This ensures that the network's predictions not only fit the training data but also satisfy physical laws. A specific implementation of embedding the aging-coupled physical model into the PIN is as follows: First, the aging-coupled physical model can be described as a system of partial differential equations containing time and spatial derivatives, such as Fick's law describing lithium-ion concentration diffusion and Fourier's law describing heat conduction. Then, a physical residual term is added to the loss function of the neural network. This term is the sum of squared differences between the left and right sides of the partial differential equations after substituting the neural network output. Finally, multi-task learning is used to simultaneously minimize the data fitting loss and the physical residual loss. Using indirect health factors as input, the battery state diagnosis model can be a multi-layer fully connected network or a long short-term memory network. Its output layer contains two branches: the first branch outputs a battery health state estimate, such as a percentage of remaining capacity or health level; the second branch outputs a fault warning signal, which can be a binary classification result such as normal or abnormal, or a continuous risk probability value. Furthermore, the output battery health status estimate can provide a quantitative basis for battery replacement and maintenance in mining areas, and the fault early warning signal can allow the system to issue an alarm before a fault occurs, avoiding equipment downtime and safety accidents. After incorporating physical constraints, this diagnostic model can maintain physical consistency even in data-sparse regions, thereby improving extrapolation ability and reliability.

[0036] S5 utilizes interpretable AI methods to perform attribution analysis on the output of the battery state diagnostic model, identify the key features that lead to the failure and the contribution of the key features, and combine this with operational condition backtracking to determine the direction of product optimization.

[0037] Explainable Artificial Intelligence (EAI) aims to break down the black-box nature of deep learning models and output decision-making rationale that is understandable to humans. A typical implementation of EAI for attribution analysis is the Shapley Additive Explanations (SHAP) method. Based on Shapley values ​​in cooperative game theory, this method calculates the marginal contribution of each input feature to the model output relative to the baseline output and distributes the total contribution among the features. Specifically, one approach to identifying key features and obtaining their contributions is as follows: First, input a set of indirect health factor samples into a trained battery status diagnostic model to obtain health status estimates or failure probabilities. Then, generate a baseline sample for this baseline, such as averaging all features. Next, iterate through all possible feature subsets, calculate the change in model output after adding a certain feature, and perform a weighted average according to the definition of Shapley values ​​to obtain the contribution value of each feature. The larger the absolute value of the contribution, the more critical the feature's impact on the current output.

[0038] In one implementation, the method for determining product optimization directions by combining operating condition backtesting is as follows: Correlation analysis is performed between the key features with the highest contribution and their corresponding operating conditions, such as ambient temperature, discharge rate, and charge / discharge cutoff voltage. For example, if a significant increase in the contribution of temperature features is found under high-temperature conditions, indicating insufficient existing heat dissipation design, then product optimization directions could include improving the thermal management system or selecting high-temperature resistant materials. Similarly, if a prominent contribution of voltage polarization features is found during high-current charging, then optimization directions could include reducing electrode internal resistance or optimizing the charging strategy. Thus, attribution analysis pushes the model output back to physically actionable improvement points.

[0039] S6 transforms the optimization direction into product design parameter adjustment instructions and feeds back the optimized product performance data to the aging coupled physical model to update the model parameters, forming a closed-loop iteration.

[0040] In some implementations, translating optimization directions into product design parameter adjustment instructions involves establishing a mapping rule library from critical fault characteristics to design parameters. For example, if attribution analysis indicates that the contribution of thermal accumulation characteristics under high-temperature conditions exceeds a preset threshold, the corresponding design parameter adjustment instruction would be to reduce the electrode coating thickness or increase the thermal interface material; if it indicates that the contribution of internal resistance growth characteristics is high, the instruction would be to adjust the electrolyte additive formulation or optimize the tab welding process. These instructions can be directly input into a computer-aided design system or a manufacturing execution system to achieve automated design changes.

[0041] In some implementations, the optimized performance data is fed back to the aging coupled physical model (ACP) to update its parameters. During actual operation after product iteration, newly collected performance data, such as capacity decay curves and internal resistance change trajectories, are used to recalibrate key parameters in the ACP, including reaction rate constant, diffusion coefficient, heat capacity, and thermal resistance. Specifically, Bayesian parameter estimation or extended Kalman filtering can be used to fuse prior model parameters with new observation data, resulting in a posteriorly updated parameter distribution. The updated model parameters allow the ACP to more accurately reflect the aging behavior of the actual product, thus providing a more precise physical prior for the next iteration. This closed-loop iterative mechanism enables battery product optimization to become a continuously self-evolving process, with each optimization based on feedback from the previous actual effect, gradually approaching the optimal design.

[0042] Based on this, this application significantly improves the accuracy and interpretability of battery fault diagnosis in mining areas by integrating real data from multiple sensor sources with simulated data from generative adversarial networks, combining aging coupled physical models with physical information neural networks, and introducing interpretable AI attribution analysis and closed-loop feedback iteration. This reduces data acquisition costs, enables continuous optimization of product design, and effectively solves the technical problems in existing technologies such as fault diagnosis relying on a large amount of labeled data, lack of physical constraints in models, and unclear optimization directions.

[0043] In one possible implementation of the embodiments of this application, combined with Figure 2 ,like Figure 3 As shown, the above S1 can be implemented through the following S101, S102 and S103, which are explained in detail below: S101. Collect battery operation data through multi-source sensors deployed on the battery system in the mining area.

[0044] The multi-source sensors include a voltage sensor, a current sensor, a multi-point thermocouple temperature sensor, an acoustic emission sensor, and an electrochemical impedance spectroscopy (EIS) measurement module. These are used to collect battery terminal voltage, charge / discharge current, multi-point temperature, acoustic emission signal, and EIS data, respectively. The sampling frequency is set according to the sensor type: the sampling frequency for voltage and current is 10Hz–100Hz, the temperature sampling frequency is 1Hz, the acoustic emission signal sampling frequency is 1MHz, and the EIS measurement frequency range is 0.01Hz–10kHz.

[0045] In some implementations, preprocessing operations are performed on the collected raw data: median filtering is used to remove impulse noise from voltage and current signals; linear interpolation is used to fill in data loss caused by communication packet loss; wavelet thresholding is used to denoise the acoustic emission signals; and then the multi-source data are aligned with a unified timestamp to form a time series matrix.

[0046] It should be noted that the acoustic emission sensor is used to detect the elastic waves released during lithium dendrite growth and microcrack formation, and is a key data source for distinguishing between electrochemical aging and mechanical damage. The electrochemical impedance spectroscopy (EIS) module obtains the battery's impedance spectrum by applying a small-amplitude AC signal, which can reflect changes in SEI film growth and charge transfer resistance, and plays an irreplaceable role in the extraction of indirect health factors. The preprocessed data is segmented according to charge-discharge cycles. Each cycle's data segment contains a complete constant-current charging, constant-voltage charging, and discharging process, and is stored with the cycle number as a time stamp.

[0047] For example, on a 200Ah lithium iron phosphate mining battery, 16 voltage monitoring points (one for each cell), 2 current sensors (total current and branch current), 8 thermocouples (distributed at the top, bottom, left, right, and center of the battery module), 4 acoustic emission sensors (attached to the battery casing surface), and 1 electrochemical impedance spectroscopy measurement module were deployed. Data was continuously collected for 30 days at a sampling period of 1 second, yielding approximately 2.6 million raw data records. After preprocessing, a time-series dataset containing 320 complete charge-discharge cycles was formed.

[0048] It should be noted that the method provided in this application has been deployed and applied to lithium-ion battery vehicles in multiple mining areas, covering various types of mining electric vehicles including monorail cranes, trackless rubber-tired vehicles, and special vehicles. The compatible battery pack specifications mainly include the 73600 / 320 series and the 32000 / 320 series. The 73600 / 320 series indicates a single cell rated voltage of 3.2V, 100 cells in series, a single cell capacity of 320Ah, and a total voltage of 320V and a total capacity of 320Ah (i.e., 73600Wh). The 32000 / 320 series indicates a single cell voltage of 3.2V, 100 cells in series, a single cell capacity of 100Ah, and a total voltage of 320V and a total capacity of 100Ah (i.e., 32000Wh). Each battery pack has the following characteristics: a 1:1 voltage and temperature sampling ratio, ensuring independent voltage and temperature monitoring for each cell; compliance with coal mine-specific explosion-proof standards and coal mine safety certification; support for independent control button activation of the system, featuring an ultra-low power standby mode that automatically shuts down after a timeout, reducing power consumption to zero; each enclosure is equipped with an independent display screen for real-time viewing of operating information and historical fault records; and a built-in thermal runaway prediction algorithm and initial handling plan. The system can be configured in various ways, including single-pack independent use, two packs in series or parallel, four packs in parallel, and two in parallel and two in series, allowing for flexible configuration based on the actual power and range requirements of mining vehicles. Actual operating data for the aforementioned battery products is collected through multi-source sensors in the S101, providing a solid data foundation for subsequent steps; specific specifications (such as rated capacity, series / parallel connection method, and thermal runaway warning threshold) also serve as important input conditions for aging coupling physical model calibration and fault evolution sequence generation.

[0049] S102. Based on electrochemical and thermal theories, an electrochemical sub-model, a thermal sub-model, and an aging sub-model are constructed respectively, and the three are coupled through state variables to form a system of differential algebraic equations.

[0050] Among them, the electrochemical sub-model adopts a single-particle model to describe the diffusion process of lithium ions in the solid phase of positive and negative electrode materials and the ion transport process in the electrolyte phase; the thermal sub-model adopts the lumped parameter method to describe the energy balance in the internal heat generation and heat dissipation process of the battery; the aging sub-model includes a solid electrolyte interface film growth sub-model, a lithium dendrite formation sub-model and an active material loss sub-model to describe the capacity decay and internal resistance increase caused by side reactions, mechanical stress and thermal effects during battery cycling.

[0051] In some implementations, the state equation of the electrochemical sub-model is expressed as: , ;in, Solid-phase lithium ion concentration (unit: mol / m³). denoted as the solid-phase diffusion coefficient (unit: m² / s), and r as the radial coordinate of the particle (unit: m). Where is the particle radius (unit: m), I is the battery current (positive for discharging, negative for charging, unit: A), and F is the Faraday constant (96485 C / mol). Let be the active surface area per unit volume of the electrode (unit: m² / m³), and L be the electrode thickness (unit: m). This equation describes the diffusion behavior of lithium ions inside spherical particles. Its physical meaning is that the rate of change of lithium ion concentration with time is equal to the divergence of diffusion flux, and the boundary conditions give the relationship between lithium ion flux and current density on the particle surface.

[0052] The state equation of the thermal quantum model is expressed as: , ;in, The average density of a single battery cell (unit: kg / m³). is the specific heat capacity (unit: J / (kg·K)), T is the temperature (unit: K), and k is the thermal conductivity (unit: W / (m·K)). Heat production rate per unit volume (unit: W / m³). Internal resistance in ohms (unit: Ω). Open circuit potential (unit: V). The activation overpotential is expressed in V. The three terms of the heat production rate correspond to ohmic heat. Reversible entropy heat and activation polarization heat Among them, reversible entropy heat is related to the rate of temperature change and has a significant impact on battery temperature rise in low-temperature environments.

[0053] The expression for the solid electrolyte interface film growth sub-model in the aging sub-model is: , ;in, SEI film thickness (unit: m). is the reaction rate constant (unit: mol / (m²·s)). , where is the activation energy (unit: J / mol), and R is the gas constant (8.314 J / (mol·K)). This refers to the side reaction current (unit: A). This is the transmission coefficient (dimensionless, usually taken as 0.5). The value represents the overpotential of the side reaction (unit: V). This equation describes the kinetic process of the SEI film gradually thickening on the negative electrode surface due to the solvent reduction reaction. The thickening of the SEI film consumes active lithium and increases internal resistance.

[0054] The expression for the lithium dendrite formation model is: ;in, Lithium dendrite length (unit: m). Dendrite growth rate constant (unit: m·s) - ¹·(A / m²) -n ), denoted as the activation energy for dendrite growth (unit: J / mol), where J is the local current density (unit: A / m²). Let be the critical current density for dendrite growth (unit: A / m²), and n be the exponential factor (usually taken as 1 to 2). This model demonstrates that when the local current density exceeds the critical value, lithium metal will deposit in the form of dendrites, and its growth rate increases power-lawfully with the increase of the overcurrent density.

[0055] The expression for the active material loss sub-model is: ;in, This represents the amount of active material lost (unit: mol / m³). Loss rate constant (unit: mol·m) - ³·s - ¹·Pa -m ), The activation energy for energy loss (unit: J / mol). The stress (in Pa) is the internal stress of the particles, caused by the difference in volume expansion / contraction due to the lithium concentration gradient, and m is the stress exponent (usually taken as 1 to 2). This equation describes the phenomenon of particle breakage and detachment of positive and negative electrode materials caused by repeated charge and discharge, which is one of the important mechanisms of capacity decay.

[0056] It should be noted that the four sub-models mentioned above are coupled with each other through state variables: the solid-phase lithium-ion concentration distribution and overpotential output by the electrochemical sub-model, and the temperature output by the thermal sub-model, serve as input conditions for the aging sub-model (for example, the SEI film growth rate is highly sensitive to temperature, and the side reaction current is exponentially related to the overpotential); simultaneously, the SEI film thickness output by the aging sub-model... Update the relationship through internal resistance Feedback is provided to the electrochemical and thermal sub-models, forming a closed-loop coupling. This is the proportionality coefficient between the SEI film thickness and the increase in internal resistance (unit: Ω / m). The initial ohmic internal resistance is given. This two-way coupling mechanism allows the model to simulate the gradual degradation of battery performance as it ages.

[0057] S103. The differential algebraic equation system formed by coupling the electrochemical sub-model, the thermal sub-model, and the aging sub-model through state variables is used as the aging coupled physical model, and its parameters are calibrated and verified.

[0058] Specifically, the differential-algebraic equations established in S102 are encapsulated as an aging-coupled physical model. The inputs to this model are operating parameters (ambient temperature, current rate, depth of discharge, etc.) and initial states (initial temperature, initial SOC, initial SEI film thickness, etc.), and the outputs are the evolution trajectories of the battery's internal state variables over time, including temperature, voltage, solid phase concentration, SEI film thickness, lithium dendrite length, and active material loss.

[0059] In some implementations, a combination of hybrid pulse power characteristic testing and electrochemical impedance spectroscopy is used to calibrate the model parameters. The specific steps are as follows: First, HPPC testing was conducted under standard operating conditions (25℃ constant temperature, 0.5C charge / discharge) to obtain the ohmic internal resistance, polarization internal resistance, and open-circuit potential curves of the battery at different SOCs. Then, EIS testing was performed at different temperatures (-20℃, 0℃, 25℃, 45℃) to obtain the frequency response characteristics of the SEI film resistance, charge transfer resistance, and diffusion impedance. Finally, a genetic algorithm was used to calibrate the uncalibrated parameters in the aging sub-model. , , , , The optimization process involves minimizing the root mean square error between the model-predicted capacity decay curve and the actual cyclic aging experimental data. Through continuous genetic iterative optimization until convergence, the values ​​of each calibration parameter are obtained.

[0060] Based on the above technical solution, S1 obtained an aging coupling physical model that accurately reflects the electrochemical-thermal-aging coupling behavior of batteries in the mining area through multi-source sensor acquisition, coupling construction of three sub-models, and parameter calibration and verification. This model not only provides a physical basis for the subsequent generation of fault evolution sequences, but also provides a computable mathematical expression for the physical constraint embedding of physical information neural networks.

[0061] In one possible implementation of the embodiments of this application, combined with Figure 2 ,like Figure 4 As shown, the above S2 can be implemented through the following S201, S202 and S203, which are explained in detail below: S201. Set a set of mining area operating condition parameters and inject preset fault triggering conditions into the aging coupling physical model.

[0062] The operating parameters include ambient temperature, discharge rate, charge rate, depth of discharge, vibration acceleration amplitude, and frequency; the fault triggering conditions include diaphragm puncture, electrolyte drying, localized overheating, and current collector fracture. The combination of operating parameters was generated using orthogonal experimental design to cover various extreme scenarios that may occur in actual operation in the mining area.

[0063] In some implementation methods, the range of parameters for each operating condition is first determined based on historical operating data of the mining area: ambient temperature -40℃~55℃, discharge rate 0.2C~3C, charging rate 0.2C~1.5C, depth of discharge 20%~100%, vibration acceleration amplitude 0.1g~5g, and vibration frequency 5Hz~200Hz. An L27 orthogonal array is used to design 27 typical operating condition combinations, each representing a typical mining area operation scenario, such as slow charging at low temperatures in winter, heavy load at high temperatures in summer, and frequent start-stop cycles on ramps. For each operating condition combination, four fault triggering conditions are injected into the aging coupling physical model: diaphragm puncture is simulated by instantaneously reducing the diaphragm thickness parameter by 90% at a specific time step; electrolyte drying is simulated by linearly decreasing the liquid phase diffusion coefficient to 10% of its initial value over time; and localized overheating is simulated by applying an additional heat source (power density of 10) to the battery's geometric center for 10 seconds. 5 The current collector fracture is simulated by increasing the current collector resistance of a battery cell by 100 times instantaneously.

[0064] It should be noted that the timing of fault triggering is set according to the characteristics of the operating conditions: for faults related to the number of cycles (such as separator puncture and current collector breakage), triggering occurs at three nodes when the state of health (SOH) drops to 85%, 75%, and 65% respectively; for faults related to environmental stress (such as local overheating), triggering occurs when the battery temperature exceeds 55°C; for faults related to electrolyte decomposition (such as drying out), triggering occurs when the cumulative charge and discharge capacity exceeds 200 times the rated capacity. This design can simulate real-world scenarios where faults occur at different stages of degradation.

[0065] S202. The Runge-Kutta method is used to numerically solve the differential algebraic equations. The evolution trajectory of the internal state variables of the battery is calculated step by step with a set time step. When any state variable exceeds the corresponding fault threshold, the fault evolution sequence is recorded.

[0066] Starting with an initial battery health state of 100% (SOH=100%), the fourth-order Runge-Kutta method was used to numerically solve the differential-algebraic equations corresponding to the aging coupled physical model. The time step Δt was set to 1 to 10 seconds (dynamically adjusted according to operating conditions: 1 second for rapid transient conditions and 10 seconds for steady-state conditions). At each time step, the internal state variables of the battery were calculated: temperature T, terminal voltage V, and solid-phase lithium-ion concentration. SEI film thickness Lithium dendrite length Loss of active materials Etc. Fault thresholds are preset for each state variable: temperature exceeding 60℃, voltage below 2.0V or above 4.2V, SEI film thickness exceeding 1μm, lithium dendrite length reaching 80% of the separator thickness (25μm), and active material loss exceeding 30% of the initial active material. A fault is determined to have occurred when any state variable exceeds the corresponding threshold.

[0067] In some implementations, a sequence of state variables for 1000 time steps prior to the time of the fault is recorded, forming a complete evolution sequence from a healthy state to a fault state. The data recorded at each time step includes: timestamp, voltage, current, temperature, solid phase concentration distribution (simplified to surface concentration and average concentration), SEI film thickness, lithium dendrite length, active material loss, and internal resistance. All records are uniformly stored in CSV format, with each line representing a sample from one time step. The last column is labeled "Fault Flag," with the fault flag being 0 before the fault occurs and 1 at and after the fault occurs.

[0068] It should be noted that for progressive failures (such as excessive SEI film growth or slow loss of active material), even if the state variables do not exceed the hard threshold, when the capacity decay rate reaches an inflection point (i.e., a change in the sign of d²Q / dt²) or the internal resistance growth rate exceeds five times the initial value, it is also considered a "functional failure" and its evolution sequence is recorded. This allows the failure evolution library to include more types of degradation modes.

[0069] S203. Input the original fault evolution sequence into the conditional generative adversarial network, alternately train the generator G and the discriminator D, generate a synthetic fault data sequence and merge it with the original evolution library to form a fault simulation dataset.

[0070] In this study, a conditional generative adversarial network (GAN) is used as the fault data generator. The GAN consists of a generator G and a discriminator D. The input to the generator G is a random noise vector z ~ N(0,1) and a conditional vector c. The conditional vector c includes an operating condition type code (represented by a one-hot vector), a fault type code (represented by a one-hot vector for four types: diaphragm puncture, electrolyte desiccation, local overheating, and current collector breakage), and a degradation stage code (represented by a continuous value indicating the current state of equilibrium (SOH) or the normalized cycle number). The generator G consists of a fully connected layer, a batch normalization layer, a LeakyReLU activation layer, and an upsampling convolutional layer connected in sequence. The output is a synthesized fault data sequence. The discriminator D consists of a convolutional layer, a dropout layer, a fully connected layer, and a sigmoid activation layer connected in sequence. Its output is the probability D(x|c) that the input data comes from the real evolutionary sequence.

[0071] In some implementations, the training process of a generative adversarial network (GAN) is as follows: First, each sequence in the original fault evolution library is divided into segments of fixed length L = 200 time steps, each segment accompanied by a corresponding condition vector c. These real segments are used as positive samples. The generator G generates synthetic segments of the same length as negative samples based on random noise z and randomly selected condition vector c.

[0072] In this process, the discriminator D and the generator G are updated alternately: the loss function of the discriminator D is: The first term is the discrimination loss for real samples, the second term is the discrimination loss for generated samples, and the third term is the gradient penalty term used to stabilize training. =10, , These are random interpolation points on the line connecting the real sample and the generated sample.

[0073] The loss function of generator G is: That is, we hope that the generator can "deceive" the discriminator, so that the discriminator will judge the generated sample as a real sample.

[0074] The Adam optimizer was used for training, with a learning rate of 0.0002 and a batch size of 64. The discriminator was trained twice, followed by the generator once. The number of training iterations could be set according to the actual training situation.

[0075] It should be noted that, in order to increase the diversity of synthetic data and prevent pattern collapse, a feature matching loss is also introduced during training: the mean squared error of the feature vectors of generated samples and real samples in the intermediate layer of the discriminator is calculated and added as an auxiliary loss to the total loss of the generator, with a weight of 0.1. In addition, the empirical replay technique can be used: the 1000 most recently generated samples are cached and mixed into real samples with a certain probability, making the training more stable.

[0076] Based on the above technical solutions, S2 successfully constructed a large-scale and diverse fault simulation dataset through operating condition parameter design, fault injection simulation, and conditional generative adversarial network synthesis. This dataset not only covers typical operating conditions and various fault types in mining areas, but also includes the complete evolution process from healthy to faulty, effectively solving the problem of scarce real fault data and providing a sufficient data foundation for subsequent multimodal feature extraction and diagnostic model training.

[0077] In one possible implementation of the embodiments of this application, combined with Figure 2 The above-mentioned S3 can be implemented through the following S301, S302 and S303, which are explained in detail below: S301. For each time series in the battery operation data and fault simulation dataset, perform time-domain feature extraction, frequency-domain feature extraction and time-frequency-domain feature extraction respectively to form the original multimodal feature vector.

[0078] The battery operation data comes from actual mining area data collected by multi-source sensors in S101, while the fault simulation dataset comes from synthesized fault data sequences in S203. Each time series corresponds to a complete charge-discharge cycle or a fixed-length time window. For each sequence, feature extraction operations are performed in the following three dimensions.

[0079] In some implementations, time-domain feature extraction includes calculating the mean, standard deviation, peak value, peak-to-peak value, skewness, kurtosis, root mean square value, waveform factor, impulse factor, and margin factor for each charge-discharge cycle. These statistics can reflect the fluctuation characteristics of battery voltage, current, and temperature signals from different perspectives: skewness and kurtosis describe the distribution shape, waveform factor and impulse factor reflect the impulse characteristics of the signal, and margin factor is sensitive to weak impulses.

[0080] Frequency domain feature extraction includes performing Fast Fourier Transform on the voltage and current sequences respectively to extract the centroid frequency, mean square frequency, frequency variance, and amplitude ratio of the first five harmonics in the spectrum. These frequency domain features can reveal the periodic fluctuations and anomalous harmonic components of the internal electrochemical reactions of the battery. For example, the thickening of the SEI film leads to an increase in low-frequency impedance, which manifests as a shift in the centroid frequency.

[0081] Time-frequency domain feature extraction includes: performing wavelet packet decomposition on the temperature sequence, with a decomposition level set to 3, and extracting the energy proportion and energy entropy of each frequency band node. Wavelet packet decomposition simultaneously decomposes the signal into a low-frequency approximation part and a high-frequency detail part, which can capture the transient changes of the temperature signal during charging and discharging and the energy distribution of different frequency bands, and is sensitive to detecting faults such as local overheating.

[0082] The above-mentioned time-domain features (10 features), frequency-domain features (5 features extracted from voltage and 5 features extracted from current, for a total of 10 features), and time-frequency-domain features (3 layers of wavelet packets, 8 nodes per layer, for a total of 24 energy percentages plus 3 energy entropies, for a total of 27 features) are concatenated to form the original multimodal feature vector. ,in .

[0083] S302. For each feature in the original multimodal feature vector, calculate its Pearson correlation coefficient with the battery health status label SOH, and select features with correlation coefficients greater than a preset threshold to form a candidate health factor set.

[0084] The SOH label comes from the raw data: for battery operation data, SOH is obtained through capacity calibration experiments, i.e., a 0.5C constant current discharge is performed every 50 cycles, and the ratio of actual capacity to rated capacity is calculated; for fault simulation datasets, SOH is directly output by the aging coupled physical model. There are N samples in total (including real and synthetic data), each sample corresponding to an SOH value and a set of 47 feature values.

[0085] In some implementations, the formula for calculating the Pearson correlation coefficient is: ;in, Let i be the feature value of the j-th sample. This is the sample mean of this feature. Let j be the health status label of the j-th sample. This represents the label mean. The value range is [-1, 1], and the closer the absolute value is to 1, the higher the degree of linear correlation. A positive correlation indicates that the SOH increases when the eigenvalue increases (healthy), and a negative correlation indicates that the SOH decreases when the eigenvalue increases (deterioration).

[0086] Set correlation coefficient threshold (This can be adjusted based on actual data, generally taken as 0.5 to 0.7). Filter out those that meet the requirements. The characteristics constitute the candidate health factor set If the number of selected features is too small (e.g., less than 10), the threshold should be appropriately lowered to 0.5; if it is too large (e.g., more than 30), the threshold should be increased to 0.65 to ensure the conciseness of the candidate set.

[0087] It should be noted that the Pearson correlation coefficient only measures linear relationships and may be insensitive to nonlinear relationships. Therefore, before calculation, a visualization analysis (scatter plot) of the features and SOH can be performed. If a significant nonlinear monotonic relationship is found, such as exponential decay, the features or SOH should be appropriately transformed (e.g., by taking the logarithm) before calculating the correlation coefficient. Furthermore, since the sample size in the fault simulation dataset is very large (over 140,000 records), a block-based calculation method can be used when calculating the correlation coefficients of all 47 features to avoid memory overflow.

[0088] S303. Calculate the variance inflation factor of the features in the candidate health factor set, and remove features whose variance inflation factor is greater than the preset value to obtain the final indirect health factor set.

[0089] The variance inflation factor is used to detect multicollinearity, which is the high correlation between features. If two features are highly correlated, retaining them will lead to model instability and increased variance in parameter estimates.

[0090] In some implementations, for each of the m features in the candidate health factor set, one of them is treated as the dependent variable, and the remaining m-1 features are treated as independent variables. A linear regression model is then established, and the coefficient of determination for each feature is calculated. The variance inflation factor is defined as: ; The closer a feature's VIF is to 1, the greater the linear explanation it receives from other features, and the higher the VIF. A VIF > 10 is generally considered to indicate severe multicollinearity, requiring feature removal or feature merging. The specific removal strategy is as follows: starting with the feature with the highest VIF, if its VIF > 10, remove it from the candidate set. Then recalculate the VIF of the remaining features, repeating this process until the VIF of all features is ≤ 10. The features ultimately retained constitute the indirect health factor set. .

[0091] For example, the candidate health factor set obtained in S302 contains 15 features. Calculating the VIF revealed that the VIFs for constant current charging time and constant voltage charging time were 18.5 and 22.3, respectively, showing a strong negative correlation (because the total charging time is fixed, a longer constant current time results in a shorter constant voltage time). Therefore, the constant voltage charging time was removed, and the constant current charging time was retained. Additionally, the VIF for the median voltage of the discharge plateau and the peak height of the incremental capacity curve was 12.1, exceeding the threshold. Considering that the peak height better reflects the structural degradation of the cathode material, the median voltage of the discharge plateau was removed. After three rounds of iterative removal, nine features remained: constant current charging time, temperature rise rate, peak height of the incremental capacity curve, ohmic internal resistance, charge transfer internal resistance, charging curve inflection point voltage, discharge cutoff voltage drop rate, cycle number normalization, and wavelet packet energy entropy. The VIFs of these features are all less than 5, constituting the final indirect health factor set. The formula for calculating the peak height of the incremental capacity curve is: Where Q represents capacity and V represents voltage. The peak height is directly related to the phase transition degree of the cathode material and the amount of lithium ions available, and is a sensitive indicator reflecting the state of equilibrium (SOH).

[0092] Based on the above technical solution, S3 extracts a set of low-redundancy, high-correlation indirect health factors from the original signal through multimodal feature extraction, Pearson correlation coefficient screening, and VIF multicollinearity test. These factors include time-domain statistics reflecting electrochemical characteristics, spectral features revealing frequency-domain characteristics, and wavelet packet energy features capturing transient temperature changes. Compared to directly using the original voltage, current, and temperature sequences, using indirect health factors as input to the diagnostic model can significantly reduce model complexity and improve training efficiency. Furthermore, since the linear correlation between these factors and SOH has been verified, the model can more easily learn accurate mapping relationships. In addition, this screening process utilizes both real and synthetic data, ensuring that the selected factors are applicable not only to normal operating conditions but also to degradation modes during fault evolution, laying a solid feature foundation for subsequent training of the physical information neural network.

[0093] In one possible implementation of the embodiments of this application, combined with Figure 2 The above-mentioned S4 specifically includes the following S401 to S403: S401. Construct a parallel branch physical information neural network, including an input layer, a convolutional neural network branch, a Transformer branch, a cross-branch perceptual modulator, a physical constraint embedding layer, and an output layer, wherein the input layer receives a time series matrix composed of indirect health factors.

[0094] The indirect health factors were selected from the final set of indirect health factors screened in S303. For each sample, its nine indirect health factor values ​​within a time window were arranged into a matrix. , where d=9 is the factor dimension. The input layer receives this matrix.

[0095] In some implementations, the convolutional neural network branch, used to extract local temporal features, consists of three sequentially connected convolutional blocks. Each convolutional block includes a one-dimensional convolutional layer, a batch normalization layer, and a ReLU activation layer. The kernel size of the first convolutional block is... (That is, covering 3 time steps and all feature dimensions), with a stride of 1 and 32 output channels; the kernel size of the second convolutional block is... The stride is 1, and the number of output channels is 64; the kernel size of the third convolutional block is... The stride is 1, and the number of output channels is 128. The output of the convolutional neural network branch is a local feature tensor. .

[0096] The Transformer branch is used to extract global aging trajectory features, including a position encoding layer, a multi-head self-attention layer, and a feedforward network layer. The position encoding layer uses sine-cosine encoding to add temporal location information to the input sequence: for position pos and dimension i, the encoding is as follows: , The multi-head self-attention layer contains 8 attention heads, and each head is calculated as follows: Where Q, K, and V are the query, key, and value matrices, respectively. Let be the dimension of the key vector. The feedforward network layer consists of two fully connected layers with ReLU activation in between. The output of the Transformer branch is the global feature tensor. .

[0097] The cross-branch sensing modulator includes a first modulation unit and a second modulation unit. The first modulation unit receives a local feature tensor. It is mapped to an attention bias vector through a fully connected layer. And this bias vector is added to the attention weight matrix of the multi-head self-attention layer in the Transformer branch, i.e. Then, the attention weights after biasing are softmax normalized and a weighted sum is calculated to obtain the modulated global features. The second modulation unit receives the global feature tensor. The convolution kernel modulation coefficients are generated through a gating mechanism. And multiply this coefficient by the weights of each convolutional kernel in the convolutional neural network branch, i.e. Then, the modulated convolution kernel is applied to the input data to perform convolution operations, thereby obtaining the modulated local features. .

[0098] The physical constraint embedding layer is used to embed the aging coupled physical model (the model obtained in S103) as a physical constraint into the neural network. Its specific structure is as follows: it embeds the modulated local features... and modulated global features The data is concatenated along the channel dimension and input into a decoder consisting of two fully connected layers. The decoder outputs the predicted internal state vector of the battery. .

[0099] The output layer receives modulated local features. Modulated global features and the physical consistency residuals output by the physical constraint embedding layer .Will and After being concatenated along the channel dimension, the data is passed sequentially through a global average pooling layer and two fully connected layers to output battery health state estimates. and fault warning signals (0 indicates normal, 1 indicates fault).

[0100] It should be noted that the parallel branch design enables the model to simultaneously capture multi-scale features of battery aging: CNN branches are sensitive to local transient changes, such as voltage drops within a loop; Transformer branches are sensitive to long-term degradation trends, such as capacity decay across loops; and the cross-branch sensing modulator enables bidirectional information exchange between the two branches. The physically constrained embedding layer is a key innovation that distinguishes it from purely data-driven models. It embeds the electrochemical-thermal-aging coupled physical model into the neural network, subjecting the model's predictive behavior to physical constraints.

[0101] S402. Define the total training loss function, including the mean squared error loss for health status estimation, the cross-entropy loss for fault prediction, the physical consistency loss, and the cross-branch consistency loss, and set the weight coefficients for each loss.

[0102] In some implementations, the total training loss function is: ;in, , which is the mean squared error loss of the health status estimate, and measures the deviation between the predicted SOH and the actual SOH; , which is the binary cross-entropy loss, used for fault early warning.

[0103] The physical consistency loss is calculated as shown in S403; Cross-branch consistency loss, used to constrain the consistency of the distribution of local and global features in the latent space, is defined as the maximum mean difference (MMD) distance between two feature tensors: ;in This is a radial basis function kernel mapping. This loss forces the features extracted by the two branches to tend to be consistent in statistical distribution, avoiding the situation where local features contradict global features; , , These are weighting coefficients, all ranging from [0,1], and satisfying the following conditions: Normal settings =0.3, =0.3, =0.2, and the remaining 0.2 is allocated to SOH to estimate the loss.

[0104] It should be noted that physical consistency loss It is the core of the physical information neural network, which uses the aging coupled physical model as a regularization term, forcing the neural network's predicted state to satisfy the evolution law of the physical equation. The cross-branch consistency loss ensures that the features extracted from the two branches from different perspectives are aligned in the latent space, thereby improving the robustness of the model.

[0105] S403. Using a fault simulation dataset and real battery operation data, with indirect health factors as input, the total loss function is minimized through the backpropagation algorithm, and the network parameters are iteratively updated until the model converges, resulting in a trained battery state diagnosis model. In each forward propagation, the physical consistency residual is calculated through a physical constraint embedding layer. And it is added as a regularization term to the loss function.

[0106] In some implementations, the training process is as follows: The indirect health factor dataset obtained from S303 (containing real and synthetic data, with a total of over 140,000 fragments) is divided into a training set, a validation set, and a test set in an 8:1:1 ratio. The input for each sample is... The output labels include the SOH value of the last time step within the time window and a flag indicating whether a fault occurred within that window. The Adam optimizer is used, with an initial learning rate of 0.001, decaying to 0.9 times the original rate every 10 epochs. The batch size is set to 64, the maximum training epochs are 200, and an early stopping strategy is employed: if the total loss on the validation set no longer decreases for 20 consecutive epochs, training is stopped and the best model is saved.

[0107] In each forward propagation, the specific workflow of the physical constraint embedding layer is as follows: The predicted internal state vector output by the decoder is... ( The input is given to the aging coupled physical model, based on the discretized state transition function. Calculate the physical state at the next time step. Then calculate the predicted state. The mean square error between the physical state and the physical state is used as the physical consistency residual: Where K=T-1 is the number of recursive steps within the time window. This is the control input for step k. This represents the battery's internal state vector predicted at the (k+1)th step by the decoder output in the physical constraint embedding layer. For the discretized state transition function, specifically: ;in, This represents the internal state vector of the battery. For the predicted average battery temperature, For the predicted terminal voltage, The predicted average thickness of the SEI film. For the predicted ohmic internal resistance, The predicted solid-phase average lithium-ion concentration; Let k be the control input vector for step k. Battery current, Ambient temperature; The open-circuit potential function is based on the average lithium-ion concentration in the solid phase. Calculations show that This is the activation overpotential at step k+1. Let the current be the current at step k+1. Let be the internal resistance of the ohm at step k+1; This represents the discretization time step. The average density of a single battery cell. This refers to the specific heat capacity of a single battery cell. Let be the volume of the battery cell, h be the convective heat transfer coefficient between the battery surface and the environment, and A be the heat dissipation area of ​​the battery cell. Let be the reaction rate constant for SEI film growth. R is the activation energy for the SEI film growth reaction, and R is the ideal gas constant. Let F be the side reaction current at step k, and F be the Faraday constant. Let be the initial ohmic internal resistance of the battery, β be the proportionality coefficient between the SEI film thickness and the increase in internal resistance, and L be the electrode thickness. This residual is backpropagated through automatic differentiation, and the gradient simultaneously updates the parameters of the decoder as well as the parameters of the CNN and Transformer branches, forcing the entire network to learn a representation that satisfies physical laws.

[0108] It should be noted that in the early stages of training, the model's predictions may deviate significantly from the laws of physics, leading to... The physical constraints are significant; as training progresses, they gradually take effect, and the internal state sequence output by the model will approach the recursive result of the physical model. Furthermore, to prevent overly strict physical constraints from causing the model to fail to fit true anomalies in the data, [the following can be done]: Set to dynamic decay mode: take a small value (such as 0.1) in the early stage of training to let the model learn the main patterns in the data first; increase it to the target value (such as 0.3) in the later stage of training to strengthen physical constraints.

[0109] Based on the above technical solution, S4 constructs a parallel branch physical information neural network and embeds an aging coupled physical model to train a battery state diagnostic model that integrates physical constraints. This model utilizes the multi-scale feature extraction capabilities of convolutional neural networks and Transformers, and ensures that the predictions conform to the electrochemical-thermal-aging laws through physical consistency loss. Thus, it can maintain high-precision health state estimation and early fault warning even under extreme working conditions in mining areas, effectively solving the problems of poor generalization ability and physical inconsistency of pure data-driven models.

[0110] In one possible implementation of the embodiments of this application, combined with Figure 2 ,like Figure 5 As shown, the above S5 can be implemented through the following S501, S502 and S503, which are explained in detail below: S501. The SHAP value method is used to perform attribution analysis on the output of the battery state diagnosis model. The fault warning signal is used as the target variable to calculate the SHAP value of each indirect health factor, and the key features and their contribution are screened out accordingly.

[0111] The battery state diagnosis model is a model trained in S403 that incorporates physical constraints, and its output includes an estimate of the battery health state. and fault warning signals Attribution analysis targets fault warning signals, that is, explaining why the model classifies a particular sample as faulty. The input features are the set of indirect health factors selected from S303. For each sample to be explained, corresponding to the sequence of indirect health factors within a time window, the SHAP value of the i-th feature is calculated. .

[0112] In some implementations, the theoretical formula for calculating the SHAP value is: ; in, Let S be the set of all input features, and let S be the subset of features that does not contain feature i. This represents the probability that the model outputs a fault warning signal of 1 when using only features from subset S, i.e., the probability that the model determines a fault. This formula calculates the average marginal contribution of feature i across all possible feature combinations. Because accurate calculation requires iterating through... For each subset, the KernelSHAP approximation algorithm is used in practical applications: by fitting the relationship between the model output and the feature mask through weighted linear regression, and setting the number of samplings to 2000, the SHAP value of each feature can be estimated efficiently.

[0113] Sort the SHAP values ​​of all samples correctly predicted as faults in the test set from largest to smallest absolute value, and select the top k features whose cumulative contribution rate reaches a preset percentage (e.g., 80%) as key features. The contribution of each key feature is defined as the proportion of its absolute SHAP value to the sum of the absolute SHAP values ​​of all features: Where k is a positive integer greater than 1, and is usually taken as the smallest k value that makes the cumulative contribution rate exceed 80% for the first time.

[0114] It should be noted that SHAP values ​​are additive: the difference between the model output probability and the baseline value is equal to the sum of the SHAP values ​​of all features. Therefore, a positive SHAP value indicates that the feature makes a positive contribution to fault determination (i.e., increases the fault probability), while a negative SHAP value indicates a negative contribution. When selecting key features, both the absolute value and the sign are important: for batteries in mining areas, features that positively contribute to increasing the fault probability are typically prioritized. Simultaneously, the model output can be utilized... Perform auxiliary verification: When a key feature changes (e.g., a feature value increases), determine... If the trend of change is consistent with the sign of the feature contribution, that is, an increase in the positive contribution feature value should lead to a decrease in SOH, then the attribution is marked as reasonable; otherwise, it is necessary to check whether there are any anomalies in the model or data.

[0115] S502. Based on the backtracking results of operating conditions, establish a mapping table between key features and product design parameters, and extract the operating condition labels when the fault occurs.

[0116] Specifically, operational condition retrospection refers to retrieving the operating logs of mining vehicles within the corresponding time period based on the timestamp of the fault occurrence, and extracting data such as ambient temperature, road gradient, load capacity, charging pile power, and continuous operation duration as operational condition tags. This data is typically recorded by the mining vehicle's information collection system and can be obtained through onboard terminals or cloud databases.

[0117] In some implementations, a mapping table between features and product design parameters is first established. This table records the upstream design parameters corresponding to each indirect health factor, for example:

[0118] For each key feature selected in S501, the corresponding design parameters are looked up in the mapping table to form an initial set of candidate optimization parameters. Then, for each fault sample, its operating condition label is extracted. Specifically, based on the end time of the fault sample's time window, i.e., the time of the fault occurrence, the operating condition data within one hour before and after that time are queried from the mine vehicle operation log, and the average or most frequent value is taken as the operating condition label for that sample. If some operating condition data is missing (e.g., the charging pile power is only recorded during the charging period), only the valid data is recorded.

[0119] It should be noted that the accuracy of operational condition backtracking depends on the completeness of the mine vehicle operation logs and the accuracy of time synchronization. It is recommended to install high-precision GPS timing modules on the vehicles to ensure that the error between the timestamps of each sensor and the log timestamp is less than 1 second. For different fault types (such as thermal runaway and internal short circuit), the sensitive operating conditions may differ: thermal runaway is more concerned with ambient temperature and discharge rate, while internal short circuit is more concerned with vibration acceleration and the number of charge-discharge cycles. Therefore, when extracting operational condition labels, differentiated extraction can be performed by combining the fault type (real labels from samples in S501 or model-predicted labels).

[0120] S503. Using the contribution of key features as weights, perform weighted voting on each design parameter in the mapping table, and combine the operating condition labels to determine the critical conditions for the occurrence of faults, generating a product optimization scheme that includes priority optimization parameters, optimization direction, optimization range suggestions, and verification operating conditions.

[0121] In some implementations, the weighted voting process is as follows: Suppose there are K key features (K=4), and each key feature i corresponds to one or more design parameters. For each design parameter p, count which key features it is associated with, and calculate the weighted score for that design parameter: ;in This is the set of key feature indices associated with the design parameter p. The design parameters with the highest scores are selected as priority optimization parameters, typically the top few parameters with a total score exceeding 50%. The optimization direction is determined by the contribution sign and actual physical meaning of the key features: if the positive contribution of a key feature is that an increase in the feature value leads to an increase in the failure probability, then the optimization direction should be to reduce that feature value, which translates into the direction of adjusting the design parameters. For example, if an increase in the constant current charging time leads to an increase in the failure probability, then the optimization direction is to reduce the constant current charging time, and the corresponding design parameter adjustment direction is to increase the constant current charging rate or decrease the charging cutoff current.

[0122] Determining the recommended adjustment range: Based on the distribution of key features in the fault samples, calculate the deviation ratio between their values ​​at the time of fault occurrence and those under normal operating conditions (e.g., the average of the first 100 cycles), and take the median of each sample as the recommended adjustment range. Simultaneously, determine the critical conditions for fault occurrence based on operating condition labels: For each operating condition label, such as ambient temperature or road slope, statistically analyze its frequency of occurrence in all fault samples. If the frequency of a certain operating condition label's value range in the fault samples exceeds a preset frequency threshold (e.g., 70%), then the boundary condition corresponding to that operating condition label is used as an optimization constraint. For example, if 80% of the fault samples occur when the ambient temperature is above 40℃, the optimization plan should specify "applicable to high-temperature operating conditions with ambient temperatures ≥ 40℃".

[0123] Finally, the above information is packaged into a product optimization plan, including: priority optimization parameters (name, current value), optimization parameter type (such as "charging strategy", "thermal management parameter"), optimization direction ("increase" / "decrease" / "switch"), optimization range suggestion (relative percentage or absolute value) and verification conditions (conditions that need to be tested, such as high temperature heavy load uphill).

[0124] It should be noted that the results of weighted voting may be biased due to the distribution of the dataset. It is recommended to repeat the above process on multiple independent subsets of failure samples (e.g., divided by different mining areas or seasons) to obtain robust optimization parameters. In addition, the optimization magnitude should be considered in conjunction with engineering feasibility: the adjustment step size should not be too large, generally not exceeding 20% ​​of the current value, and iteratively implemented in subsequent verification simulations.

[0125] Based on the above technical solutions, S5 achieves a leap from "black-box diagnosis" to "white-box traceability" through interpretable SHAP value attribution analysis, operating condition backtracking and mapping relationship establishment, weighted voting and critical condition judgment. It provides a clear and quantifiable direction for product design improvement and effectively solves the problem that the diagnostic results in the existing technology are not interpretable and cannot guide product iteration.

[0126] In one possible implementation of the embodiments of this application, combined with Figure 2 The above-mentioned S6 can be implemented through the following S601, S602 and S603, which are explained in detail below: S601. Encapsulate the priority optimization parameters, optimization parameter types, optimization directions, optimization magnitude suggestions, and verification conditions in the product optimization plan into structured adjustment instructions, and send them to the edge side of the mining area battery management system via OTA, while simultaneously synchronizing them to the parameter configuration file of the aging coupled physical model.

[0127] The structured adjustment instructions are in JSON format and include the following fields: The parameter identifier field (param_id) is used to uniquely identify the design parameter to be adjusted, such as "charge_current_ratio" and "coolant_flow_rate". Adjust type field (adjust_type): Values ​​include "increase", "decrease", or "toggle"; Adjust the step field: Use a relative percentage (e.g., "+5%)" or an absolute value (e.g., "0.05C"); Constraints field: Records the boundary conditions in the verification working condition, such as "ambient temperature ≥ 40℃ and slope ≥ 5%"; The safety limit field (safe_limits) records the upper and lower boundaries that the adjusted parameters must not exceed, such as "maximum constant current ratio 1.2C, minimum 0.5C".

[0128] In some implementations, the optimized solution obtained from S503 is first converted into a JSON object. For example: { "param_id": "charge_current_ratio", "adjust_type": "increase", “step”: “+0.05C”, “constraints”: {“temp_min”: 40, “slope_min”: 5}, "safe_{limits": {"min": 0.5, "max": 1.2} } Then, the command is sent via OTA (Over-The-Air) server to the edge of the battery management system of all relevant vehicles in the mining area. After receiving the command, the edge device temporarily stores it in the local configuration cache, and it will take effect the next time the vehicle is powered on or charged. At the same time, the command is synchronized to the parameter configuration file of the aging coupled physical model in the cloud for subsequent verification simulation.

[0129] It is important to note that OTA (Over-The-Air) updates must ensure communication security and command integrity. TLS encryption can be used for transmission, and commands can be digitally signed to prevent tampering. For multiple vehicles, updates can be rolled out in batches to avoid network congestion. If a vehicle is offline for more than a preset time, the command should be retained and rolled out immediately upon regaining online status. Additionally, when synchronizing to the physical model configuration file, the parameter modification history should be preserved for rollback purposes.

[0130] S602. Perform verification simulation in the aging coupled physical model: Substitute the adjusted design parameters into the model and rerun the fault evolution simulation under verification conditions; if the simulation results show that no similar faults are triggered, mark the adjustment instruction as "verified" and submit it to the product design change system; if the fault is still triggered, trigger the rollback mechanism, restore the design parameters to the values ​​before adjustment, and push them to the manual analysis queue.

[0131] The verification simulation employs the same fourth-order Runge-Kutta method as S202, with a time step Δt = 2 seconds, and a total simulation duration covering at least 1000 charge-discharge cycles or until a fault occurs. The verification operating conditions use the critical operating conditions defined in S503. The adjusted design parameters are substituted into the aging coupled physical model of S103, and the simulation is run to monitor for the occurrence of similar faults.

[0132] In some implementations, the fault judgment criteria are set to be consistent with S202: temperature exceeding 60℃, voltage below 2.0V or above 4.2V, SEI film thickness exceeding 1μm, lithium dendrite length reaching 80% of the separator thickness, and active material loss exceeding 30%. If no fault threshold is triggered throughout the entire simulation cycle, and the slope of the battery health state (SOH) decay curve does not show an abnormal inflection point (i.e., the sign of d²SOH / dt² does not change abruptly), the adjustment is considered effective, and the instruction is marked as "verified". Otherwise, the adjustment is considered ineffective, triggering a rollback mechanism: restoring the parameters in the physical model to their values ​​before adjustment, generating a failure report, including the simulation time when the fault occurred, the state variable exceeding the limit, and suggested directions for manual troubleshooting, and pushing it to the manual analysis queue.

[0133] It should be noted that verification simulations can be accelerated using parallel computing: for multiple combinations of design parameters, multiple simulation instances can be run simultaneously. Furthermore, the rollback mechanism should not only restore the physical model parameters but also automatically send a "rollback command" to the edge side to undo previously issued adjustments, preventing invalid modifications from being applied to the actual vehicle.

[0134] For example, after increasing the constant current rate from 0.8C to 0.85C, the simulation was re-run under the verification conditions (45°C, 8% slope). The results showed that the battery temperature reached a maximum of 58°C at the 800th cycle (not exceeding 60°C), and the SEI film thickness ultimately reached 0.9μm (not exceeding 1μm), with no thermal runaway occurring. Compared to the unadjusted state (thermal runaway occurred at 420 cycles at 0.8C), the battery life was nearly doubled. Therefore, this adjustment instruction was marked as "verified" and submitted to the product design change system. Simultaneously, another parameter, "increasing the cooling flow rate by 10%", was also verified. The results showed that increasing the cooling flow rate further reduced the battery temperature rise; however, considering energy consumption and cost, adjusting only the constant current rate was chosen as the preferred solution.

[0135] S603. After receiving the verified adjustment instruction, the product design change system generates a product design parameter modification notification and triggers the remote parameter update process of the on-site battery management system in the mining area. At the same time, the optimized performance data of the product (including the new aging curve and capacity decay data in the verification simulation) is fed back to the aging coupled physical model to update the model parameters and form a closed-loop iteration.

[0136] The product design change system is an internal engineering change management tool for the company. Upon receiving a verified adjustment instruction, it automatically generates a modification notification form, which includes: parameter name, original value, new value, basis for modification (attribution analysis report), verification simulation results, and effective date. After the notification form is electronically approved, it triggers the remote parameter update process of the battery management system at the mining site: a confirmation instruction is reissued via OTA, ensuring that the vehicle BMS at the site actually adopts the new parameters.

[0137] In some implementations, the key to closed-loop iteration is feeding the optimized performance data back to the aging-coupled physical model. Specifically, this involves adding new aging evolution data obtained from verification simulations, such as the battery's capacity decay curve, internal resistance growth curve, and temperature evolution sequence under adjusted parameters, as additional training samples to the parameter calibration dataset of the aging sub-model. Incremental learning is then used to recalibrate some parameters in the aging sub-model, enabling the physical model to more accurately reflect the optimized battery behavior. The updated model parameters are saved and used for fault evolution sequence generation and diagnostic model training in the next iteration.

[0138] It should be noted that the feedback data includes both synthetic data from the verification simulation and real performance data collected during subsequent real-world vehicle operation. These two types of data work together to continuously bring the physical model closer to the actual degradation patterns.

[0139] Based on the above technical solutions, S6 achieves the executable, safe, and continuous evolution of product optimization through structured instruction encapsulation, OTA (Over-The-Air) updates, physical model verification and simulation, rollback mechanisms, and performance data feedback. This closed-loop iterative mechanism not only avoids the safety risks caused by blind adjustments but also enables battery products to continuously self-optimize as operational data accumulates, improving the long-term reliability and economy of battery systems in mining areas.

[0140] How this application works: First, based on the electrochemical-thermal-aging multiphysics coupling model, fault evolution sequences covering different operating conditions and fault types are generated through numerical simulation. A conditional generative adversarial network is then used to expand this into a large-scale fault simulation dataset, addressing the scarcity of real fault samples in mining areas. Second, time-domain, frequency-domain, and time-frequency-domain multimodal features are extracted from the original data. Pearson correlation coefficients and variance inflation factors are used to screen out indirect health factors that are highly correlated with battery health status and have low redundancy. Then, a parallel branch physical information neural network is constructed. Using the indirect health factors as input, the discretized state transition function of the aging coupling physical model is added as a physical consistency residual to the total loss function during training. Backpropagation forces the model output to follow the electrochemical-thermal-aging law, thereby outputting high-precision health status estimates and fault warning signals. Next, the SHAP value method is used to perform attribution analysis targeting the fault warning signals, quantifying the contribution of each indirect health factor to fault determination. Combined with operating condition backtracking, operating condition labels such as ambient temperature and road slope at the time of fault occurrence are extracted to establish a mapping relationship between features and product design parameters. Weighted voting is used to select priority optimization parameters and adjustment directions. Finally, the optimization scheme is encapsulated into structured instructions, which are then sent to the edge via OTA and synchronized to the physical model for verification simulation. If the verification is effective, a design change is submitted and the battery management system parameters are updated remotely. At the same time, the optimized performance data is fed back to the physical model to update its parameters, forming a closed-loop iteration of "data acquisition → fault synthesis → feature screening → diagnostic modeling → attribution and tracing → design optimization → verification and deployment → data feedback", enabling the battery product to continuously evolve itself.

[0141] Although this application has been described herein in conjunction with various embodiments, those skilled in the art, by reviewing the accompanying drawings, disclosure, and appended claims, will understand and implement other variations of the disclosed embodiments in carrying out the claimed application. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude multiple instances. A single processor or other unit can implement several functions listed in the claims. While different dependent claims may recite certain measures, this does not mean that these measures cannot be combined to produce good results.

[0142] Although this application has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made thereto without departing from the spirit and scope of this application. Accordingly, this specification and drawings are merely illustrative descriptions of the application as defined by the appended claims, and are considered to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from the spirit and scope of this application. Thus, if such modifications and variations of this application fall within the scope of the claims of this application and their equivalents, this application is also intended to include such modifications and variations.

Claims

1. A method for iterative optimization of battery products in mining areas based on fault big data, characterized in that, include: Battery operation data is collected using multi-source sensors, and an aging coupled physical model is constructed based on electrochemical and thermal theories. Based on the aging coupled physical model, the numerical solution method is used to solve the problem step by step with a set time step by preset mining area working conditions and fault triggering conditions to obtain the evolution trajectory of the internal state variables of the battery. The time series of the state variables from the initial state to the point where the fault threshold is exceeded is recorded to obtain the fault evolution sequence. The fault evolution sequence is input into a generative adversarial network to generate a fault simulation dataset; Multimodal features are extracted from the battery operation data and the fault simulation dataset, and indirect health factors related to battery health status are screened from the multimodal features using the Pearson correlation coefficient. The aging coupled physical model is embedded into a physical information neural network, and the indirect health factor is used as input to train a battery state diagnosis model that integrates physical constraints, and outputs battery health state estimation and fault warning signal. The output of the battery state diagnostic model is analyzed using interpretable AI methods to identify key features that cause the failure and their contribution, and product optimization directions are determined by combining the backtesting of operating conditions. The optimization direction is transformed into product design parameter adjustment instructions, and the optimized product performance data is fed back to the aging coupled physical model to update the model parameters, forming a closed-loop iteration.

2. The method for iterative optimization of battery products in mining areas based on fault big data as described in claim 1, characterized in that, The construction process of the aging coupled physical model includes: A system of differential-algebraic equations is constructed by coupling state variables from an electrochemical sub-model, a thermal sub-model, and an aging sub-model; among which, The electrochemical sub-model adopts a single-particle model or a pseudo-two-dimensional model to describe the diffusion process of lithium ions in the solid phase of positive and negative electrode materials and the ion transport process in the electrolyte phase. The thermal sub-model uses the lumped parameter method or the finite element method to describe the energy balance during the heat generation and dissipation process inside the battery. The aging sub-model includes a solid electrolyte interface film growth sub-model, a lithium dendrite formation sub-model, and an active material loss sub-model, which are used to describe the capacity decay and internal resistance increase of the battery due to side reactions, mechanical stress, and thermal effects during cycling. The solid phase concentration and overpotential output by the electrochemical sub-model and the temperature output by the thermal sub-model are used as inputs to the aging sub-model. At the same time, the SEI film thickness and internal resistance growth output by the aging sub-model are fed back to the electrochemical sub-model and the thermal sub-model to form a closed-loop coupling, thus obtaining the aging coupled physical model.

3. The method for iterative optimization of battery products in mining areas based on fault big data as described in claim 2, characterized in that, The generation of fault evolution sequences under different operating conditions based on the aging coupled physical model includes: A set of mining area operating condition parameters is set, including ambient temperature, discharge rate, charge rate, depth of discharge, vibration acceleration amplitude and frequency; For each combination of operating parameters, preset fault triggering conditions are injected into the aging coupled physical model. The fault triggering conditions include diaphragm puncture, electrolyte drying, local overheating, and current collector breakage. Starting with the initial value of battery health state SOH=100%, the Runge-Kutta method is used to numerically solve the differential algebraic equations, and the evolution trajectory of the battery's internal state variables is calculated step by step with a set time step. When any state variable exceeds the corresponding fault threshold, the state variable sequence of several time steps before the fault time is recorded, including voltage, current, temperature, solid phase concentration distribution, SEI film thickness and lithium dendrite length, forming a fault evolution sequence from healthy state to fault state. Iterate through all combinations of operating parameters and fault triggering conditions to generate an original fault evolution library containing several fault evolution sequences.

4. The method for iterative optimization of battery products in mining areas based on fault big data as described in claim 3, characterized in that, The specific process for generating the fault simulation dataset includes: A conditional generative adversarial network (GAN) is used as the fault data generator, and the GAN includes a generator G and a discriminator D; wherein... The generator G takes a random noise vector and a condition vector as input, and the condition vector includes a working condition type code, a fault type code, and a degradation stage code. The generator G consists of a fully connected layer, a batch normalization layer, a LeakyReLU activation layer, and an upsampled convolutional layer connected in sequence, and outputs a synthesized fault data sequence. The discriminator D consists of a convolutional layer, a dropout layer, a fully connected layer, and a sigmoid activation layer connected in sequence, and its output is the probability that the input data comes from a real evolutionary sequence. The generator G and discriminator D are trained alternately until convergence. The trained generator G is then sampled in batches in the conditional vector space to generate several synthetic fault data sequences, which are then merged with the original fault evolution library to form a fault simulation dataset.

5. The method for iterative optimization of battery products in mining areas based on fault big data as described in claim 1, characterized in that, The multimodal feature extraction of the battery operation data and the fault simulation dataset includes: For each time series in the battery operation data and the fault simulation dataset, perform feature extraction operations in the following three dimensions: First, calculate the time-domain characteristic statistics within each charge-discharge cycle, including mean, standard deviation, peak value, peak-to-peak value, skewness, kurtosis, root mean square value, waveform factor, impulse factor, and margin factor. Second, perform a fast Fourier transform on the voltage and current sequences to extract frequency domain features, which include the centroid frequency, mean square frequency, frequency variance, and amplitude ratio of the first 5 harmonics in the spectrum. Third, wavelet packet decomposition is performed on the temperature sequence to extract time-frequency domain features, which include the energy proportion and energy entropy of each frequency band node; The time-domain feature statistics, frequency-domain features, and time-frequency-domain features are concatenated to form the original multimodal feature vector, thus obtaining the multimodal features.

6. The method for iterative optimization of battery products in mining areas based on fault big data as described in claim 1, characterized in that, The battery state diagnostic model incorporating physical constraints adopts a parallel branch physical information neural network architecture, including an input layer, a convolutional neural network branch, a Transformer branch, a cross-branch sensing modulator, a physical constraint embedding layer, and an output layer; wherein... The input layer receives a time series matrix composed of the indirect health factors. The convolutional neural network branch is used to extract features from the time series matrix through multiple convolutions and output a first feature tensor. The Transformer branch is used to extract features from the time series matrix through a multi-head self-attention mechanism and output a second feature tensor. The cross-branch perceptual modulator includes a first modulation unit and a second modulation unit. The first modulation unit receives the first feature tensor, maps the first feature tensor to an attention bias vector through a fully connected layer, and adds the attention bias vector to the attention weight matrix of the multi-head attention layer in the Transformer branch. By performing softmax normalization on the attention weights after adding the bias and calculating the weighted sum, a modulated second feature tensor is obtained. The second modulation unit receives the second feature tensor, generates convolution kernel modulation coefficients through a gating mechanism, and multiplies the modulation coefficients with the weights of each convolution kernel in the convolutional neural network. By applying the modulated convolution kernels to the input data for convolution operations, a modulated first feature tensor is obtained. The physical constraint embedding layer is used to embed the aging coupled physical model as a physical constraint into the battery state diagnosis model. During the model training process, the physical consistency residual is added as a regularization term to the total loss function to achieve consistency constraint between the prediction results of the battery state diagnosis model and the physical evolution law. The output layer receives the modulated first feature tensor and the modulated second feature tensor, as well as the physical consistency residual output by the physical constraint embedding layer. The modulated first feature tensor and the modulated second feature tensor are concatenated along the channel dimension and then passed through the global average pooling layer and the fully connected layer in sequence to output the battery health status estimate and the fault warning signal.

7. The method for iterative optimization of battery products in mining areas based on fault big data as described in claim 1, characterized in that, The attribution analysis of the output of the battery state diagnostic model using interpretable AI methods to identify key features leading to the fault and their contribution includes: The SHAP value method is used for attribution analysis. The fault warning signal output by the battery state diagnosis model is used as the target variable. For each sample to be explained, the SHAP value of the i-th indirect health factor is calculated. Sort all samples by absolute value from largest to smallest, and select the top k features whose cumulative contribution rate reaches a preset percentage as key features. The contribution of each key feature is defined as the proportion of the SHAP value to the sum of the absolute values ​​of all feature SHAP values; where k is a positive integer greater than 1. The battery health status estimate output by the battery status diagnosis model is used to assist in verifying the rationality of the attribution results: when key features change, it is determined whether the trend of the battery health status estimate is consistent with the sign of the contribution of the key features; if they are consistent, the attribution is marked as reasonable.

8. The method for iterative optimization of battery products in mining areas based on fault big data as described in claim 1, characterized in that, The product optimization direction is determined based on the key characteristics of the fault, the contribution of the key characteristics, and backtesting of operating conditions, specifically including: Based on the type of the key features, a mapping table between features and product design parameters is established; the mapping table records the upstream design parameters corresponding to each indirect health factor. Combined with the working condition retrospective results, the working condition retrospective results refer to retrieving the mining area vehicle operation logs within the corresponding time period based on the timestamp of the time when the fault occurred, and extracting the ambient temperature, road slope, load capacity, charging pile power, and continuous operation duration as working condition tags. The contribution of the key features is used as a weight to perform weighted voting on each design parameter in the mapping table, and the design parameters whose total contribution exceeds a preset ratio are selected as priority optimization parameters. Based on the operating condition labels, the critical conditions for the occurrence of the fault are determined. If the frequency of a certain operating condition label in all fault samples is greater than a preset frequency threshold, the boundary conditions corresponding to the operating condition label are used as optimization constraints to generate a product optimization scheme that includes the priority optimization parameters, optimization parameter types, optimization directions, optimization magnitude suggestions, and verification operating conditions.

9. The method for iterative optimization of battery products in mining areas based on fault big data as described in claim 8, characterized in that, The step of converting the optimization direction into product design parameter adjustment instructions includes: The priority optimization parameters, parameter types, optimization directions, suggested optimization magnitudes, and verification conditions in the product optimization plan are encapsulated into structured adjustment instructions; the data format of the structured adjustment instructions is JSON, containing the following fields: The parameter identifier field is used to uniquely identify the design parameter to be adjusted. Adjust the type field, with possible values ​​including "increase", "decrease", and "switch"; Adjust the step size field to use relative percentages or absolute values; The constraint field is used to record the boundary conditions in the verification condition; The safety limit field is used to record the upper and lower boundaries that the adjusted parameters must not exceed; The structured adjustment instructions are sent to the edge side of the mining area battery management system via OTA control and download technology, and simultaneously synchronized to the parameter configuration file of the aging coupled physical model; Perform verification simulation in the aging coupled physical model: substitute the adjusted design parameters into the model, and rerun the fault evolution simulation under the verification conditions; if the simulation results show that no similar faults are triggered, mark the adjustment instruction as "verified" and submit it to the product design change system; if the fault is still triggered in the verification simulation, trigger the rollback mechanism, restore the design parameters to the values ​​before adjustment, and push them to the manual analysis queue. After receiving the verified structured adjustment instruction, the product design change system generates a product design parameter modification notification and triggers the remote parameter update process of the on-site battery management system in the mining area.

10. A mining area battery product iterative optimization system based on fault big data, characterized in that, include: The system includes a physical modeling module, a fault data synthesis module, a feature extraction module, a model diagnosis module, an attribution analysis module, and a closed-loop optimization module; among them, The physical modeling module is used to collect battery operating data using multi-source sensors and to construct an aging coupled physical model based on electrochemical and thermal theories. The fault data synthesis module is used to calculate the evolution trajectory of the internal state variables of the battery by using a numerical solution method with a set time step based on the aging coupled physical model, through preset mining area working condition parameters and fault triggering conditions, and recording the time series of the state variables from the initial stage to when they exceed the fault threshold, thereby obtaining the fault evolution sequence. The fault evolution sequence is then input into a generative adversarial network to generate a fault simulation dataset. The feature extraction module is used to perform multimodal feature extraction on the battery operation data and the fault simulation dataset, and to use the Pearson correlation coefficient to screen out indirect health factors related to the battery health status from the multimodal features. The model diagnostic module is used to embed the aging coupled physical model into a physical information neural network, and train a battery state diagnostic model that integrates physical constraints with the indirect health factor as input, and output battery health state estimation and fault warning signal. The attribution analysis module is used to perform attribution analysis on the output of the battery state diagnostic model using interpretable AI methods, identify the key features that lead to the failure and the contribution of the key features, and determine the product optimization direction by combining the backtracking of operating conditions. The closed-loop optimization module is used to convert the product optimization direction into product design parameter adjustment instructions, and to feed back the optimized product performance data to the aging coupled physical model to update the model parameters, forming a closed-loop iteration.

Images