A data and mechanism coordinated energy storage battery abnormal attenuation hierarchical diagnosis method

Abstract

This invention belongs to the field of safety monitoring technology for electrochemical energy storage systems, and discloses a hierarchical diagnostic method for abnormal degradation of energy storage batteries that combines data and mechanisms. The specific steps are as follows: Step 1: Multi-dimensional dynamic feature extraction. Data preprocessing, voltage range feature identification based on a sliding window, multi-dimensional dynamic feature extraction, and feature optimization and filtering are performed on the original operating data of the energy storage battery. Through standardized formulas for voltage, capacity, and temperature, and related formulas for feature calculation, the optimal feature set characterizing the battery's aging state is obtained. This invention, based on a holographic dynamic characterization method that combines sliding window voltage range feature identification with incremental capacity-internal resistance multi-source fusion, overcomes the limitations of single parameters and achieves multi-dimensional and accurate characterization of the battery's aging state. Simultaneously, a hybrid reasoning framework combining a lightweight physical neural network and temporal fuzzy logic is constructed, ensuring interpretability while quantifying uncertainty.

Description

Technical Field

[0001] This invention belongs to the field of safety monitoring technology for electrochemical energy storage systems, specifically a hierarchical diagnostic method for abnormal degradation of energy storage batteries that combines data and mechanisms. Background Technology

[0002] In large-scale applications of energy storage batteries, accurate diagnosis of abnormal degradation is crucial for ensuring the safe and stable operation of the system and extending battery life. Although current technologies for diagnosing abnormal degradation in energy storage batteries have made some progress, significant bottlenecks remain in areas such as adaptability to complex operating conditions, diagnostic accuracy, and engineering practicality, making it difficult to meet the stringent requirements of real-world applications. 1. Existing feature extraction techniques mostly rely on single-dimensional or static feature parameters, lacking adaptation to the coupling characteristics of multiple physical fields. They are difficult to fully capture the weak dynamic features in the abnormal battery degradation process, are easily affected by noise interference and operating condition fluctuations, and have insufficient sensitivity in early degradation feature identification, failing to provide high-quality data support for subsequent diagnosis. 2. There is a disconnect between physical mechanisms and data-driven approaches in the modeling process. Pure physical models have high computational complexity and are difficult to deploy in engineering. Pure data-driven models lack physical constraint embedding, have weak generalization ability, and are prone to diagnostic bias in cross-operating conditions and cross-battery type scenarios. Furthermore, the challenge of balancing lightweight model design with diagnostic accuracy has not yet been effectively solved. 3. At the reasoning and decision-making level, existing methods mostly adopt a single reasoning mode, which does not fully consider the quantification of uncertainties in the diagnosis process, resulting in insufficient reliability of reasoning results. At the same time, the diagnosis and decision-making lacks hierarchical design, which cannot provide "observable, credible and controllable" decision support for on-site operation and maintenance. Furthermore, there is no closed-loop feedback mechanism, making it difficult to adapt to the dynamic evolution characteristics of battery degradation. To address the aforementioned technical challenges, this invention proposes a hierarchical diagnostic method for abnormal degradation of energy storage batteries that integrates data and mechanisms. This method constructs a full-chain diagnostic system, enabling holographic characterization of battery aging features, high-precision prediction of state of harmonic exhaust (SOH), accurate reasoning of degradation levels, and practical implementation of operation and maintenance decisions. This effectively overcomes the shortcomings of existing technologies. Summary of the Invention

[0003] The purpose of this invention is to provide a hierarchical diagnostic method for abnormal degradation of energy storage batteries that combines data and mechanisms, in order to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a hierarchical diagnostic method for abnormal degradation of energy storage batteries that combines data and mechanisms, the specific steps of which are as follows: Step 1: Multi-dimensional dynamic feature extraction. Perform data preprocessing, voltage range feature identification based on sliding window, multi-dimensional dynamic feature extraction and feature optimization and screening operations on the original operating data of the energy storage battery. Obtain the optimal feature set characterizing the battery aging state through voltage, capacity and temperature standardization formulas and feature calculation formulas. Step 2: Construction and training of a lightweight neural network model with embedded physical constraints. A four-layer lightweight neural network architecture integrating MLP and PINN is built, embedding relevant formulas for the continuity of charged state constraints and thermodynamic energy balance constraints. A combined optimization strategy of knowledge distillation, parameter pruning, and quantization is used to compress the model size. The model training is completed in three stages: pre-training, knowledge distillation training, and parameter pruning and quantization optimization, through loss function fusion formula, to obtain a high-precision and lightweight SOH prediction model. Step 3: Hybrid intelligent reasoning and uncertainty quantification. Construct a hybrid reasoning architecture that combines neural networks and temporal fuzzy logic. The neural network submodule outputs the initial battery decay probability, which is then converted into a semantic decay level by the temporal fuzzy logic submodule using probability change rate calculation, membership function, and defuzzification formula. Finally, uncertainty quantification is completed through confidence weighted calculation. Step 4: Hierarchical diagnostic decision-making. Based on the dual-dimensional results of attenuation level and confidence level, perform diagnostic status classification, dual-dimensional judgment, and hierarchical action execution to form standardized operation and maintenance decision instructions. Step 5: Dynamic updates and feedback. Based on the operation and maintenance execution results and subsequent monitoring data, complete the upgrade / downgrade determination of battery diagnostic status, evaluate the decision effect, feed back relevant data to the preceding modules to complete model iteration and optimization, and build an operation and maintenance knowledge base to realize the reuse of decision-making experience. Step Six: Full-chain closed-loop verification. Verify the output results of the above steps from multiple dimensions to ensure the effectiveness of feature extraction, the accuracy of model prediction, the reliability of reasoning conclusions, and the implementation of decision-making actions, forming a full-chain diagnostic closed loop of "feature extraction - modeling and reasoning - decision-making and maintenance - feedback optimization".

[0005] Preferably, the data preprocessing in step one includes raw data screening and cleaning, and data standardization. The raw data screening and cleaning removes cyclical data from the final stage of charging when the current drops sharply, when the charging time is below a set threshold, and when the SOH label fluctuates abnormally. The data standardization maps voltage, capacity, and temperature data to the [0,1] interval, respectively, to eliminate the influence of individual battery parameter differences and operating condition fluctuations. The data standardization formula is as follows: Voltage standardization:

[0006] Capacity standardization:

[0007] Temperature standardization:

[0008] In the formula, This is the original voltage data. The rated maximum charging voltage of the battery. This is the battery discharge cutoff voltage; Q represents the cumulative capacity on a single charge, and Q is the rated capacity of the battery. The raw temperature data, For the maximum allowable operating temperature of the battery, The minimum allowable operating temperature for the battery.

[0009] Preferably, the voltage range feature recognition based on the sliding window in step one adopts a strategy combining equal interval division and adaptive sliding window. First, the charging voltage range is divided into basic windows according to ΔV∈[0.005V,0.05V]. Then, the sliding window length L∈[3,10] and sliding step size S∈[1,3] are adaptively adjusted according to the SOH value of the previous round, and basic parameters such as cumulative charging capacity, charging time, and temperature change in each window are calculated. The adaptive adjustment formula is as follows:

[0010] The cumulative capacity within the window is calculated by integrating the current: During constant current charging In the formula, The window start time, The end time of the window. For charging current, The charging current is constant.

[0011] Preferably, the multi-dimensional dynamic feature extraction and incremental capacity analysis and internal resistance evaluation technology described in step one extracts three major categories of features from three dimensions: capacity, impedance, and temperature. These features include basic statistical features, incremental capacity features, and internal resistance temperature coefficient change rate features, thereby achieving a holographic characterization of the battery aging state. The core calculation formulas include: Cumulative average capacity within the window:

[0012] Cumulative capacity standard deviation within the window:

[0013] Mean rate of temperature change within the window:

[0014] Rate of voltage change within the window:

[0015] Incremental capacity:

[0016] Incremental capacity peak offset:

[0017] Area of ​​the incremental capacity curve:

[0018] Temperature coefficient of internal resistance:

[0019] Rate of change of internal resistance temperature coefficient:

[0020] Internal resistance temperature coefficient stability index:

[0021] In the formula, q is the cumulative capacity of the k-th basic window within the sliding window, L is the length of the sliding window, Δp is the temperature change of the k-th basic window within the sliding window, Δt is the charging time of the basic window, Vs is the starting voltage of the sliding window, and Ve is the ending voltage of the sliding window. This represents the cumulative change in capacity within the sliding window. This represents the voltage change within the sliding window. This represents the voltage corresponding to the peak value of the current window's incremental capacity. Let IC(V) be the peak voltage of the incremental capacity under healthy battery conditions, and let IC(V) be the function of incremental capacity as a function of voltage. Window start temperature The corresponding equivalent internal resistance, The end temperature of the window The corresponding equivalent internal resistance, This is the internal resistance temperature coefficient of the previous sliding window. The standard deviation of the temperature coefficient of internal resistance within the window. This represents the average value of the temperature coefficient of internal resistance within the window; The feature optimization screening adopts a three-step method of "correlation analysis - feature importance assessment - redundancy removal". First, Pearson correlation analysis is used to screen features that are strongly correlated with SOH. Then, the random forest algorithm is used to assess the importance of features. Finally, highly redundant features are removed to obtain the optimal feature set. The formula for calculating the Pearson correlation coefficient is as follows:

[0022] The formula for calculating the Gini coefficient is as follows:

[0023] In the formula, X is the feature variable, Y is the SOH value, Cov(X,Y) is the covariance of X and Y, Var(X) is the variance of X, Var(Y) is the variance of Y, p is the probability of the k-th value of feature X, and K is the number of values ​​of feature X.

[0024] Preferably, the four-layer lightweight neural network architecture in step two is a feature input layer - feature enhancement layer - physical constraint embedding layer - lightweight output layer. The feature input layer performs batch normalization on the optimal feature set. The feature enhancement layer uses a lightweight MLP structure to enhance features. The physical constraint embedding layer fuses the output of the feature enhancement layer with physical constraints. The lightweight output layer uses a single neuron + 1×1 convolutional kernel structure to output the SOH prediction value. The batch normalization formula for the feature input layer is:

[0025] The formula for calculating the feature enhancement layer is:

[0026] The formula for calculating the lightweight output layer is:

[0027] in, The mean of the batch data. Let γ and β be the variance of the batch data, β be learnable parameters, and ε be a small value to prevent the denominator from being zero (ε = 1e-5). , Here is the weight matrix for each hidden layer. , Here are the bias vectors for each hidden layer. , The output feature maps of each hidden layer. The weight vector of the output layer. For the bias term of the output layer, Feature maps for incorporating physical information.

[0028] Preferably, the physical constraint embedding in step two involves transforming the state of charge continuity constraint and the thermodynamic energy balance constraint into a constraint loss function, which is then weighted and fused with the model prediction loss function to form the total loss function. This is achieved through a backpropagation algorithm, realizing a deep fusion of physical constraints and data-driven approaches. The state of charge continuity constraint includes SOC temporal continuity constraints and SOC-voltage relationship constraints; the thermodynamic energy balance constraint includes energy conservation constraints and temperature change rate constraints. These constraints are fused and embedded through the loss function to form the model's total loss function. The formula for the SOC temporal continuity constraint is: In the formula, Let be the predicted SOC value at time t. The SOC prediction value at time t-1; the SOC-voltage relationship constraint formula is: g(u) = a3u3 + a2u2 + a1u1 + a0, where, Let SOC be the voltage mapping function, ε be the allowable error, and ε ≤ 0.02; the energy conservation constraint formula is: In the formula, The total electrical energy input during the charging process. The change in chemical energy stored in a battery. The generated heat energy; the temperature change rate constraint formula is: In the formula, k1 is the proportionality coefficient, and k2 is the environmental temperature influence coefficient. Let be the equivalent internal resistance at time t; the formula for the total loss function is:

[0029] In the formula, α is the weighting coefficient, and α∈[0.6, 0.8]. For the prediction loss function, Here, N is the constraint loss function, and N is the number of samples. Let SOH be the predicted value for the i-th sample. Let be the true SOH value of the i-th sample, and λ1, λ2, λ3, and λ4 be constraint weight coefficients that satisfy λ1 + λ2 + λ3 + λ4 = 1. For SOC temporal continuity constraint loss, The loss is constrained by the SOC-voltage relationship. Losses are due to energy conservation constraints. The loss is constrained by the rate of temperature change.

[0030] Preferably, in step two, the pre-training stage of model training divides the battery's entire lifecycle cyclic data into training, validation, and test sets in a 7:2:1 ratio, and uses the Adam optimizer and early stopping strategy to complete model initialization training. In the knowledge distillation training stage, a deep MLP teacher model and a lightweight student model are constructed, and a dual-label training method using "soft labels + hard labels" is employed to achieve knowledge transfer. In the parameter pruning and quantization optimization stage, redundant parameters are first removed and fine-tuned, and then the model weights and activation values ​​are quantized from 32-bit floating-point to 8-bit integers, ultimately obtaining a lightweight model that meets engineering deployment requirements. The model performance must satisfy RMSE ≤ 0.015, MAE ≤ 0.01, inference time ≤ 10ms, and model size ≤ 5MB. The distillation loss function formula for the knowledge distillation training is as follows:

[0031] In the formula, μ is the hard label weight coefficient, and μ∈[0.3, 0.5]. The MSE loss between the student model's predicted values ​​and the actual SOH values; The KL divergence loss is used to calculate the probability distributions of the student model and the teacher model. Let i be the output probability distribution of the teacher model for the i-th sample. Let i be the output probability distribution of the i-th sample in the student model; The quantization optimization formula is as follows: In the formula, It is a 32-bit floating-point number. It is the quantized 8-bit integer. The maximum value of the parameter. This is the minimum value of the parameter.

[0032] Preferably, in step three, the neural network submodule takes the optimal feature set of the current time step and the previous two time steps as input, adopts a lightweight structure of "input layer - temporal correlation layer - output layer", introduces a temporal attention mechanism to strengthen the weight of recent features, outputs the initial decay probability and forms a historical probability sequence; the temporal fuzzy logic submodule takes the historical decay probability sequence as input, defines the current decay probability and probability change rate as fuzzy input variables, and decay level as fuzzy output variable, designs triangular and trapezoidal membership functions and 12 core fuzzy rules, and uses the Mamdani inference method and centroid method to complete fuzzy inference and defuzzification, outputting a semantic decay level; the uncertainty quantification uses the membership value corresponding to the decay level as the basic weight and the consistency coefficient of the historical decay level sequence as the secondary weight, and calculates the confidence level by weighting at a ratio of 0.6:0.4; if the confidence level is lower than 50%, a re-inference mechanism is triggered, the initial decay probability is optimized and inference is performed again to ensure that the confidence level is ≥50%; the input of the neural network submodule is The core calculation formula is as follows:

[0033] In the formula, For batch normalization operations, This is the temporal attention weight vector. For element-wise product operation, , This is the weight matrix. , It is the bias vector; The formula for calculating the probability change rate in the temporal fuzzy logic submodule is as follows:

[0034] In the formula, V is the daily average rate of change of probability, V>0 indicates that the decay is intensifying, and V<0 indicates that the decay trend is easing. The centroid method for resolving fuzziness is as follows:

[0035] In the formula, To obtain the precise attenuation level value after deblurring, The attenuation levels are quantified as follows: 0 = Normal, 1 = Slight attenuation, 2 = Moderate attenuation, 3 = Severe attenuation. This represents the final membership degree of the corresponding level; The confidence level calculation formula for the uncertainty quantification is as follows:

[0036] In the formula, This represents the membership value corresponding to the final attenuation level. The sequence consistency coefficient, =7, This refers to the number of times in the historical sequence that the current decay level is consistent with the current level.

[0037] Preferably, the diagnostic status classification in step four combines SOH value, attenuation level, and core feature fluctuation amplitude to classify the battery into four states: normal SOH ≥ 0.9, slight attenuation 0.8 ≤ SOH < 0.9, moderate attenuation 0.6 ≤ SOH < 0.8, and severe attenuation SOH < 0.6. A triple verification mechanism of "SOH judgment + attenuation level judgment + core feature verification" is adopted. The dual-dimensional judgment constructs a two-dimensional decision matrix of "diagnostic status × confidence level". The confidence level is divided into high confidence Conf>90%, medium confidence 70% ≤ Conf ≤ 90%, and low confidence Conf<70%. Each intersection node corresponds to a unique operation and maintenance decision action. The hierarchical action execution clarifies the execution specifications, responsible parties, and execution time limits of each decision action according to the operation and maintenance priority, so as to realize the standardized implementation of decision instructions.

[0038] Preferably, the dynamic update and feedback in step five is based on the changes in diagnostic status and confidence level monitored for three consecutive times, and the status is upgraded / downgraded; the decision effect is evaluated using the fluctuation amplitude of core features, the rate of change of decay probability, and the rate of decrease of SOH as evaluation indicators, and strategy optimization is triggered if the standards are not met; the evaluation data and status update data are fed back to the preceding module as 10% of the supplementary training data, and the model fine-tuning training is completed every six months.

[0039] The beneficial effects of this invention are as follows: 1. This invention proposes a holographic dynamic characterization method based on sliding window voltage range feature recognition and incremental capacity-internal resistance multi-source fusion, which overcomes the limitations of a single parameter and achieves multi-dimensional and accurate characterization of battery aging state.

[0040] 2. This invention constructs a hybrid reasoning framework that combines lightweight physical neural networks with temporal fuzzy logic, which achieves uncertainty quantification while ensuring interpretability, and significantly improves diagnostic robustness under complex working conditions.

[0041] 3. This invention enables pre-dial diagnostics and early warning based on historical automatic assessment results, which can identify potential abnormal battery degradation risks 24 hours in advance and provide semantic risk level assessment and confidence quantification to support preventive maintenance decisions.

[0042] 4. This invention constructs a full-chain diagnostic system from feature extraction and hybrid reasoning to hierarchical decision-making, forming a systematic solution adapted to complex working conditions. Through lightweight design and computational optimization that takes into account both early identification and engineering practicality, it supports the engineering application of the algorithm at the edge. Attached Figure Description

[0043] Figure 1 This is a flowchart of the present invention; Figure 2 This is a system module architecture diagram of the present invention. Detailed Implementation

[0044] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] like Figures 1 to 2 As shown in the figure, this embodiment of the invention provides a hierarchical diagnostic method for abnormal degradation of energy storage batteries that combines data and mechanisms. The specific steps are as follows: Step 1: Multi-dimensional dynamic feature extraction. Perform data preprocessing, voltage range feature identification based on sliding window, multi-dimensional dynamic feature extraction and feature optimization and screening operations on the original operating data of the energy storage battery. Obtain the optimal feature set characterizing the battery aging state through voltage, capacity and temperature standardization formulas and feature calculation formulas. Step 2: Construction and training of a lightweight neural network model with embedded physical constraints. A four-layer lightweight neural network architecture integrating MLP and PINN is built, embedding relevant formulas for the continuity of charged state constraints and thermodynamic energy balance constraints. A combined optimization strategy of knowledge distillation, parameter pruning, and quantization is used to compress the model size. The model training is completed in three stages: pre-training, knowledge distillation training, and parameter pruning and quantization optimization, through loss function fusion formula, to obtain a high-precision and lightweight SOH prediction model. Step 3: Hybrid intelligent reasoning and uncertainty quantification. Construct a hybrid reasoning architecture that combines neural networks and temporal fuzzy logic. The neural network submodule outputs the initial battery decay probability, which is then converted into a semantic decay level by the temporal fuzzy logic submodule using probability change rate calculation, membership function, and defuzzification formula. Finally, uncertainty quantification is completed through confidence weighted calculation. Step 4: Hierarchical diagnostic decision-making. Based on the dual-dimensional results of attenuation level and confidence level, perform diagnostic status classification, dual-dimensional judgment, and hierarchical action execution to form standardized operation and maintenance decision instructions. Step 5: Dynamic updates and feedback. Based on the operation and maintenance execution results and subsequent monitoring data, complete the upgrade / downgrade determination of battery diagnostic status, evaluate the decision effect, feed back relevant data to the preceding modules to complete model iteration and optimization, and build an operation and maintenance knowledge base to realize the reuse of decision-making experience. Step Six: Full-chain closed-loop verification. Verify the output results of the above steps from multiple dimensions to ensure the effectiveness of feature extraction, the accuracy of model prediction, the reliability of reasoning conclusions, and the implementation of decision-making actions, forming a full-chain diagnostic closed loop of "feature extraction - modeling and reasoning - decision-making and maintenance - feedback optimization".

[0046] This design forms a closed-loop diagnostic process from feature extraction to feedback optimization. Each step is progressive and collaborative, ensuring diagnostic accuracy and generalization through mechanism and data fusion, while also improving engineering practicality through lightweight design and quantitative analysis. At the same time, hierarchical decision-making and dynamic feedback enable more precise implementation of maintenance instructions, and can continuously optimize the model to adapt to the dynamic characteristics of battery degradation, thus comprehensively solving the pain points of traditional diagnostics.

[0047] Step one, data preprocessing, includes raw data screening and cleaning, and data standardization. Raw data screening and cleaning removes cyclical data from the final charging stage with a sudden drop in current, charging time below a set threshold, and abnormal fluctuations in the SOH label. Data standardization maps voltage, capacity, and temperature data to the [0,1] interval to eliminate the influence of individual battery parameter differences and operating condition fluctuations. The data standardization formula is as follows: Voltage standardization:

[0048] Capacity standardization:

[0049] Temperature standardization:

[0050] In the formula, This is the original voltage data. The rated maximum charging voltage of the battery. This is the battery discharge cutoff voltage; Q represents the cumulative capacity on a single charge, and Q is the rated capacity of the battery. The raw temperature data, For the maximum allowable operating temperature of the battery, The minimum allowable operating temperature for the battery.

[0051] In step one, the voltage range feature recognition based on the sliding window adopts a strategy combining equal interval division and adaptive sliding window. First, the charging voltage range is divided into basic windows according to ΔV∈[0.005V,0.05V]. Then, the sliding window length L∈[3,10] and sliding step size S∈[1,3] are adaptively adjusted according to the SOH value of the previous round. Basic parameters such as cumulative charging capacity, charging time, and temperature change in each window are calculated. The adaptive adjustment formula is as follows:

[0052] The cumulative capacity within the window is calculated by integrating the current: During constant current charging In the formula, The window start time, The end time of the window. For charging current, The charging current is constant.

[0053] This design balances the accuracy and adaptability of feature extraction. Data preprocessing removes invalid data through cleaning and eliminates individual and operating condition interference through standardization, laying a high-quality data foundation for subsequent analysis. The equal interval division combined with the SOH-driven adaptive sliding window can accurately match the characteristics of different battery aging stages, flexibly capture dynamic changes, and the current integral calculation window capacity also ensures the accuracy of basic parameters, improving the targeting and reliability of feature recognition.

[0054] In step one, multi-dimensional dynamic feature extraction is simultaneously integrated with incremental capacity analysis and internal resistance assessment technology. This extracts three main categories of features from three dimensions: capacity, impedance, and temperature. These features include basic statistical characteristics, incremental capacity characteristics, and internal resistance temperature coefficient change rate characteristics. This achieves a holographic characterization of the battery's aging state. The core calculation formulas include: Cumulative average capacity within the window:

[0055] Cumulative capacity standard deviation within the window:

[0056] Mean rate of temperature change within the window:

[0057] Rate of voltage change within the window:

[0058] Incremental capacity:

[0059] Incremental capacity peak offset:

[0060] Area of ​​the incremental capacity curve:

[0061] Temperature coefficient of internal resistance:

[0062] Rate of change of internal resistance temperature coefficient:

[0063] Internal resistance temperature coefficient stability index:

[0064] In the formula, q is the cumulative capacity of the k-th basic window within the sliding window, L is the length of the sliding window, Δp is the temperature change of the k-th basic window within the sliding window, Δt is the charging time of the basic window, Vs is the starting voltage of the sliding window, and Ve is the ending voltage of the sliding window. This represents the cumulative change in capacity within the sliding window. This represents the voltage change within the sliding window. This represents the voltage corresponding to the peak value of the current window's incremental capacity. Let IC(V) be the peak voltage of the incremental capacity under healthy battery conditions, and let IC(V) be the function of incremental capacity as a function of voltage. Window start temperature The corresponding equivalent internal resistance, The end temperature of the window The corresponding equivalent internal resistance, This is the internal resistance temperature coefficient of the previous sliding window. The standard deviation of the temperature coefficient of internal resistance within the window. This represents the average value of the temperature coefficient of internal resistance within the window; Feature optimization screening employs a three-step method: "correlation analysis - feature importance assessment - redundancy removal". First, Pearson correlation analysis is used to screen features strongly correlated with SOH. Then, the random forest algorithm is used to assess feature importance. Finally, highly redundant features are removed to obtain the optimal feature set. The Pearson correlation coefficient calculation formula is as follows:

[0065] The formula for calculating the Gini coefficient is as follows:

[0066] In the formula, X is the feature variable, Y is the SOH value, Cov(X,Y) is the covariance of X and Y, Var(X) is the variance of X, Var(Y) is the variance of Y, p is the probability of the k-th value of feature X, and K is the number of values ​​of feature X.

[0067] This design integrates multiple technologies and multi-dimensional feature extraction to achieve a holographic characterization of battery aging from three dimensions: capacity, impedance, and temperature, accurately capturing dynamic aging characteristics. It quantifies core features through a series of formulas to ensure the scientific nature and accuracy of feature extraction. The three-step feature screening method progressively eliminates invalid and redundant features, retains core features strongly correlated with SOH, and provides high-quality and highly targeted feature input for subsequent modeling, improving the accuracy and efficiency of model diagnosis.

[0068] In step two, the four-layer lightweight neural network architecture consists of a feature input layer, a feature enhancement layer, a physical constraint embedding layer, and a lightweight output layer. The feature input layer performs batch normalization on the optimal feature set. The feature enhancement layer uses a lightweight MLP structure to enhance features. The physical constraint embedding layer fuses the output of the feature enhancement layer with physical constraints. The lightweight output layer uses a single neuron + 1×1 convolutional kernel structure to output the SOH prediction value. The batch normalization formula for the feature input layer is:

[0069] The formula for calculating the feature enhancement layer is:

[0070] The formula for calculating the lightweight output layer is:

[0071] in, The mean of the batch data. Let γ and β be the variance of the batch data, β be learnable parameters, and ε be a small value to prevent the denominator from being zero (ε = 1e-5). , Here is the weight matrix for each hidden layer. , Here are the bias vectors for each hidden layer. , The output feature maps of each hidden layer. The weight vector of the output layer. For the bias term of the output layer, Feature maps for incorporating physical information.

[0072] This architecture features a layered design and formulaic quantification of parameters. Batch normalization eliminates feature scale differences, and lightweight MLP enhances feature representation capabilities. The physical constraint embedding layer integrates mechanism rules to improve model interpretability and generalization. The lightweight output layer simplifies the structure, balancing computational efficiency and prediction accuracy. While ensuring the accuracy of SOH prediction, the overall architecture achieves model lightweighting through formulaic parameter control, adapting to engineering deployment requirements and solving the problems of weak generalization and difficult deployment of pure data-driven models.

[0073] In step two, the physical constraint embedding transforms the state of charge (SOC) continuity constraints and thermodynamic energy balance constraints into constraint loss functions, which are then weighted and fused with the model prediction loss function to form the total loss function. This is achieved through a backpropagation algorithm, enabling deep integration of physical constraints and data-driven processes. The SOC continuity constraints include SOC temporal continuity constraints and SOC-voltage relationship constraints; the thermodynamic energy balance constraints include energy conservation constraints and temperature change rate constraints. These constraints are fused and embedded using loss functions to form the model's total loss function. The formula for the SOC temporal continuity constraint is: In the formula, Let be the predicted SOC value at time t. The SOC prediction value at time t-1; the SOC-voltage relationship constraint formula is: g(u) = a3u3 + a2u2 + a1u1 + a0, where, Let SOC be the voltage mapping function, ε be the allowable error, and ε ≤ 0.02; the energy conservation constraint formula is: In the formula, The total electrical energy input during the charging process. The change in chemical energy stored in a battery. The heat energy generated; the temperature change rate constraint formula is: In the formula, k1 is the proportionality coefficient, and k2 is the environmental temperature influence coefficient. Let be the equivalent internal resistance at time t; the formula for the total loss function is:

[0074] In the formula, α is the weighting coefficient, and α∈[0.6, 0.8]. For the prediction loss function, Here, N is the constraint loss function, and N is the number of samples. Let SOH be the predicted value for the i-th sample. Let be the true SOH value of the i-th sample, and λ1, λ2, λ3, and λ4 be constraint weight coefficients that satisfy λ1 + λ2 + λ3 + λ4 = 1. For SOC temporal continuity constraint loss, The loss is constrained by the SOC-voltage relationship. Losses are due to energy conservation constraints. The loss is constrained by the rate of temperature change.

[0075] This design transforms electrochemical and thermodynamic physical constraints into loss functions and integrates them with weighted predicted losses. Through backpropagation, it achieves a deep integration of mechanism and data-driven approaches, allowing the model to learn to conform to the actual operating laws of the battery, significantly improving generalization and interpretability. It solves the problems of pure data models lacking physical constraints and being prone to deviations across operating conditions. The hierarchical loss weights can be flexibly adjusted, ensuring both the accuracy of SOH prediction and that the model conforms to physical mechanisms, laying a reliable modeling foundation for subsequent diagnostics.

[0076] In step two, the pre-training phase of model training divides the battery's entire lifecycle cyclic data into training, validation, and test sets in a 7:2:1 ratio, and uses the Adam optimizer and early stopping strategy to complete model initialization training. The knowledge distillation training phase constructs a deep MLP teacher model and a lightweight student model, employing a dual-label training method ("soft label + hard label") to achieve knowledge transfer. The parameter pruning and quantization optimization phase first removes redundant parameters and fine-tunes them, then quantizes the model weights and activation values ​​from 32-bit floating-point to 8-bit integers, ultimately obtaining a lightweight model suitable for engineering deployment. The model performance must meet the following requirements: RMSE ≤ 0.015, MAE ≤ 0.01, inference time ≤ 10ms, and model size ≤ 5MB. The distillation loss function formula for knowledge distillation training is as follows:

[0077] In the formula, μ is the hard label weight coefficient, and μ∈[0.3, 0.5]. The MSE loss between the student model's predicted values ​​and the actual SOH values; The KL divergence loss is used to calculate the probability distributions of the student model and the teacher model. Let i be the output probability distribution of the teacher model for the i-th sample. Let i be the output probability distribution of the i-th sample in the student model; The quantization optimization formula is: In the formula, It is a 32-bit floating-point number. It is the quantized 8-bit integer. The maximum value of the parameter. This is the minimum value of the parameter.

[0078] This design employs phased training and optimizes the process through formula quantization. Pre-training ensures the basic accuracy of the model, while knowledge distillation leverages dual-label transfer of deep model knowledge, balancing accuracy and lightweight design. Parameter pruning and quantization further compress the model size, reducing inference time and storage usage. The entire strategy meets stringent performance metrics such as RMSE and MAE while achieving lightweight model design, adapting to engineering deployment requirements, and addressing the pain point of high-precision models being difficult to deploy.

[0079] In step three, the neural network submodule takes the optimal feature sets of the current time step and the previous two time steps as input, adopts a lightweight structure of "input layer - temporal correlation layer - output layer", introduces a temporal attention mechanism to strengthen the weight of recent features, outputs the initial decay probability and forms a historical probability sequence; the temporal fuzzy logic submodule takes the historical decay probability sequence as input, defines the current decay probability and probability change rate as fuzzy input variables, and decay level as fuzzy output variable, designs triangular and trapezoidal membership functions and 12 core fuzzy rules, and uses the Mamdani inference method and centroid method to complete fuzzy inference and defuzzification, outputting a semantic decay level; uncertainty quantification uses the membership value corresponding to the decay level as the basic weight and the consistency coefficient of the historical decay level sequence as the secondary weight, and calculates the confidence level by weighting them in a ratio of 0.6:0.4; if the confidence level is lower than 50%, a re-inference mechanism is triggered, the initial decay probability is optimized and the inference is executed again to ensure that the confidence level is ≥50%; the input of the neural network submodule is The core calculation formula is as follows:

[0080] In the formula, For batch normalization operations, This is the temporal attention weight vector. For element-wise product operation, , This is the weight matrix. , It is the bias vector; The formula for calculating the probability change rate in the temporal fuzzy logic submodule is as follows:

[0081] In the formula, V is the daily average rate of change of probability, V>0 indicates that the decay is intensifying, and V<0 indicates that the decay trend is easing. The formula for resolving fuzzy issues using the center-of-gravity method is:

[0082] In the formula, To obtain the precise attenuation level value after deblurring, The attenuation levels are quantified as follows: 0 = Normal, 1 = Slight attenuation, 2 = Moderate attenuation, 3 = Severe attenuation. This represents the final membership degree of the corresponding level; The formula for calculating the confidence level of uncertainty quantification is as follows:

[0083] In the formula, This represents the membership value corresponding to the final attenuation level. The sequence consistency coefficient, =7, This refers to the number of times in the historical sequence that the current decay level is consistent with the current level.

[0084] This design integrates the advantages of neural networks and fuzzy logic. The lightweight network combined with the temporal attention mechanism accurately fits the decay probability and strengthens the influence of recent features. Fuzzy inference transforms abstract probability into semantic decay levels, which is more in line with actual operation and maintenance. Confidence weighted calculation quantifies the uncertainty of inference, and the low-confidence heavy inference mechanism ensures the reliability of the conclusion. The whole architecture takes into account diagnostic accuracy, interpretability and practicality, and improves the robustness of inference under complex working conditions.

[0085] In step four, the diagnostic status classification combines SOH value, attenuation level, and core feature fluctuation amplitude to divide the battery into four states: normal SOH ≥ 0.9, slight attenuation 0.8 ≤ SOH < 0.9, moderate attenuation 0.6 ≤ SOH < 0.8, and severe attenuation SOH < 0.6. A triple verification mechanism of "SOH judgment + attenuation level judgment + core feature verification" is adopted. A two-dimensional decision matrix of "diagnostic status × confidence level" is constructed for the two-dimensional judgment. The confidence level is divided into high confidence Conf>90%, medium confidence 70% ≤ Conf ≤ 90%, and low confidence Conf<70%. Each intersection node corresponds to a unique operation and maintenance decision action. The hierarchical action execution clarifies the execution specifications, responsible parties, and execution time limits of each decision action according to the operation and maintenance priority, so as to realize the standardized implementation of decision instructions.

[0086] This design classifies battery status through triple verification of SOH, degradation level, and core characteristics to ensure accurate and unbiased classification; it constructs a two-dimensional decision matrix of diagnostic status and confidence level, so that each operating condition has a dedicated maintenance action, avoiding blind decision-making; it clarifies the execution specifications, subjects, and time limits according to priority, so as to achieve standardized implementation of decision instructions, taking into account both maintenance efficiency and pertinence, while the tiered handling can accurately match the degree of battery degradation, providing clear and implementable practical guidance for on-site maintenance.

[0087] In step five, dynamic updates and feedback are based on changes in diagnostic status and confidence level from three consecutive monitoring sessions, and status upgrades / downgrades are performed. The decision-making effect is evaluated using the fluctuation amplitude of core features, the rate of change of decay probability, and the rate of decrease of SOH as evaluation indicators. If the criteria are not met, strategy optimization is triggered. The evaluation data and status update data are fed back to the preceding modules as 10% of the supplementary training data, and model fine-tuning training is completed every six months.

[0088] This design relies on multi-indicator monitoring to dynamically upgrade / downgrade battery status, accurately tracking real-time changes in battery aging; it evaluates decision-making effectiveness using core characteristics and degradation probability as indicators, and optimizes strategies immediately if standards are not met, ensuring the effectiveness of operation and maintenance; it feeds operation and maintenance data as a supplementary training set back to the preceding modules, periodically fine-tuning the model so that the diagnostic system continuously adapts to battery degradation characteristics, forming a data-driven closed-loop optimization, and significantly improving the long-term adaptability and accuracy of the entire process diagnosis.

[0089] Example: This embodiment uses an 18650 lithium-ion power battery pack (rated voltage 3.2V, capacity 200Ah) as the application object to describe in detail a hierarchical diagnostic method for abnormal degradation of energy storage batteries based on data and mechanism collaboration. This method is based on a full-chain diagnostic system of "feature extraction - modeling and reasoning - decision-making and maintenance - feedback optimization," constructing four core modules. Each module independently implements specific technical functions and forms a closed-loop collaboration through data interaction. The specific implementation steps are as follows: Step 1: The multi-dimensional dynamic feature extraction module performs feature extraction operations. This module is the basic data processing unit of the diagnostic system. It consists of a data preprocessing subunit, a sliding window voltage range identification subunit, a multi-dimensional feature extraction subunit, and a feature optimization and screening subunit connected in series. The processing flow conforms to the ISO15118-3 standard. The specific implementation steps and technical parameters are as follows: 1.1 Data Preprocessing Subunit: Time-series data (sampling frequency 1Hz) of the battery's 0.3C constant current charging process were collected, including voltage u(t), current i(t), temperature p(t), charging time t, and SOH tag. The charging cutoff voltage was 3.65V, and the discharging cutoff voltage was 2.5V. A three-step cleaning operation was performed to remove cyclic data with a sudden drop in current at the end of charging >0.05C, charging time <2h, and SOH tag deviation from the mean ±5%. Anomaly identification adopted the statistical standard 3σ criterion (confidence level 99.7%). Subsequently, data standardization was performed on the cleaned data, and voltage, capacity, and temperature were standardized respectively, all mapped to the [0,1] interval to eliminate scale differences. The standardization formula and specific parameters are as follows: Voltage standardization: =(u(t)-2.5) / (3.65-2.5) (voltage range 2.5V~3.65V); Capacity standardization: = / 200 (rated capacity 200Ah); Temperature standardization: =(p(t)-25) / (60-25) (Operating temperature range 25℃~60℃, which is the normal operating range of lithium-ion batteries). The input to this sub-unit is the original battery runtime sequence data; the output is standardized clean data with a data integrity of ≥98%, meeting the acceptable standards for engineering applications.

[0090] 1.2 Sliding Window Voltage Range Identification Subunit: First, the voltage range is divided into equal intervals, based on the standardized voltage curve. A basic window is divided using a voltage interval ΔV of 0.01V. This interval is a commonly used optimal value for ICA analysis of lithium-ion batteries, balancing feature extraction accuracy and data processing efficiency. The entire charging voltage range (2.5V~3.65V, standardized from 0~1V) is divided into 115 basic windows accordingly, with the voltage range of the i-th window being […]. , +0.01V]( =2.5V, =3.65V); then an adaptive sliding window design was implemented, which adaptively adjusted the window length L and sliding step size S based on the previous SOH value. The specific implementation rules have been verified in engineering to be adaptable to the feature capture requirements of different aging stages: When L > 0.8, L = 9, S = 2; when L ≤ 0.6, S = 2. When ≤0.8, L=6, S=1; When <0.6, L=4, S=1; finally, calculate the basic parameters within the window, that is, the cumulative charging capacity q within each sliding window (calculated by current integration, q=0.3×200×( ) in the constant current 0.3C scenario). ), charging time Δt, temperature change Δp; The input to this sub-unit is standardized cleaning data; the output is the set of partitioned sliding windows and the basic parameters of each window.

[0091] 1.3 Multi-dimensional Feature Extraction Subunit: Integrating incremental capacity analysis and internal resistance assessment techniques, dynamic features are extracted from three dimensions: capacity, impedance, and temperature, to achieve a comprehensive characterization of battery aging status. The specific implementation process involves extracting three main categories of features (a total of 12 core features). The extraction methods and specific parameters for each feature are as follows: 1) Basic statistical characteristics: mean cumulative capacity within the window =Σ / 9 (when L=9), cumulative capacity standard deviation =√[Σ( - [ )² / (9-1)], mean temperature change rate =ΣΔ / 9. Voltage change rate =0.01V / Δt (ΔV=0.01V); 2) Incremental capacity characteristics: Incremental capacity IC within the window = Δq / 0.01V (ΔV = 0.01V), IC peak value (identification threshold 0.5Ah / V), IC peak offset (difference between the current peak voltage and the peak voltage at healthy state SOH = 1.0, in V), IC curve area (integration range 2.8V~3.6V). 3) Characteristics of the temperature coefficient of internal resistance: Equivalent internal resistance R within the window (1kHz AC impedance test, test voltage 10mV, frequency accuracy ±0.1%), temperature coefficient of internal resistance α=ΔR / Δp (Δp=5℃, temperature range 25℃~30℃), rate of change of temperature coefficient of internal resistance β=( - ) / ×100%, internal resistance temperature coefficient stability index γ= / ( Let α be the standard deviation. (where α is the mean). The input to this sub-unit is a set of sliding windows and the basic parameters of each window, as well as AC impedance test data; the output is a multi-dimensional dynamic feature set.

[0092] 1.4 Feature Optimization and Screening Subunit: A three-step method of "correlation analysis - feature importance assessment - redundancy removal" is adopted. The parameters and thresholds of each step conform to the industry's conventional value standards, as follows: The first step is to perform Pearson correlation analysis, with the significance level set to the statistically standard 0.05. After calculating the correlation coefficient between each feature and the SOH value, strongly correlated features with an absolute value > 0.6 are retained (8 items are retained in this embodiment, which conforms to the conventional threshold for battery feature screening), and weakly correlated features < 0.3 are removed (2 items are removed); The second step is to perform feature importance assessment, using random... The machine forest algorithm (100 decision trees, maximum depth 10 layers, all standard hyperparameters) calculates the Gini coefficients of the features and selects the top 6 important features (in this embodiment, the peak IC, equivalent internal resistance R, internal resistance temperature coefficient α, etc. are selected). The third step is to remove redundancy by calculating the Pearson correlation coefficients between important features and removing redundant features with an absolute correlation coefficient > 0.8 and a small Gini coefficient (in this embodiment, 1 feature is removed, and the redundancy removal threshold conforms to industry standards). Finally, the optimal feature set (5 core features, dimension 5) is output. The input to this sub-unit is a multi-dimensional dynamic feature set and a sequence of SOH values; the output is a 5-dimensional optimal feature set.

[0093] Step 2: Construct and train the SOH prediction model by embedding physical constraints into a lightweight modeling module.

[0094] This module is the core modeling unit of the diagnostic system, consisting of a model architecture design subunit, a physical constraint embedding subunit, a lightweight optimization subunit, and a model training subunit. It can be deployed on an ARM Cortex-M4 processor. Specific implementation details are as follows: 2.1 Model Architecture Design Subunit: A four-layer architecture is constructed, consisting of a "feature input layer - feature enhancement layer - physical constraint embedding layer - lightweight output layer," combining MLP and PINN to achieve feature mapping and SOH prediction. The specific implementation process and parameters are as follows: The feature input layer takes a 5-dimensional optimal feature set as input, and a batch normalization operation is added. The batch normalization formula is: BN(x) = γ × (x - ... ) / (√( +ε))+β, where γ=1.0 and β=0.0 (initial values). This is the batch average. The batch variance is ε = 1e-5. The feature enhancement layer adopts a lightweight MLP structure with two hidden layers (10 neurons in the first layer and 7 neurons in the second layer). The activation function is ReLU (threshold 0) to avoid gradient vanishing. The physical constraint embedding layer receives the output of the feature enhancement layer and the physical constraint conditions, fuses physical information and data features, and outputs a fused feature map (7-dimensional). The lightweight output layer adopts a single neuron structure with the activation function Sigmoid (output mapped to the [0,1] interval, corresponding to the SOH value). A 1×1 convolution kernel (weight initialization uses He normal distribution) is used to implement feature mapping, reducing model parameters (total parameters ≤ 500). The mathematical description of the model for this sub-unit is as follows: ,in For a 5-dimensional optimal feature set, The physical constraints are θ, the learnable parameters of the model are f(·), and the model mapping function is f(·). The input of this sub-unit is the optimal feature set and the physical constraints. The output is the SOH prediction value (preliminary) and the fused feature map (7-dimensional).

[0095] 2.2 Physical Constraint Embedded Sub-unit: Two major constraints are designed and fused into the embedded model using a loss function. These constraints include SOC (State of Charge) continuity constraints and thermodynamic energy balance constraints. The SOC continuity constraints include temporal continuity constraints (absolute value of the difference between adjacent SOC times ≤ 0.02) and SOC-voltage relationship constraints. g(·) is the SOC-voltage mapping function fitted by a cubic polynomial (fitting coefficient R² ≥ 0.99). ≤0.02); thermodynamic energy balance constraints include energy conservation constraints and temperature change rate constraints ( =0.002× × +0.01); the constraint embedding method is: constructing a constraint loss function. , and the prediction loss function Weighted fusion into the total loss function =0.7× +0.3× ; Using MSE loss, The weighted sum of the constraint deviations (λ1=0.3, λ2=0.2, λ3=0.3, λ4=0.2, λ1+λ2+λ3+λ4=1). The inputs to this sub-unit are the SOC sequence, voltage sequence, current sequence, temperature sequence, and equivalent internal resistance sequence; the outputs are the total loss function, constraint loss components, and physical constraint fusion feature map (7-dimensional).

[0096] 2.3 Lightweight Optimization Subunit: A triple optimization approach of "knowledge distillation + parameter pruning + quantization" is adopted. The parameters for each step are as follows: First, knowledge distillation is performed to construct a teacher-student model architecture. The teacher model is a high-precision deep MLP (4 hidden layers, 5000 total parameters), and the student model is a physically constrained embedding model. Dual-label training using "soft labels (probability distribution of the teacher model) + hard labels (true SOH values)" is employed, and the loss function is distilled. =0.4× +0.6× , First, the model parameters were compressed from 500 to 175 using KL divergence loss (divergence ≤ 0.1). Second, parameter pruning was performed by statistically analyzing the absolute value distribution of weights, setting the 10th percentile (pruning threshold 0.001) as the pruning threshold, and performing structured pruning (removing redundant neurons). After pruning, fine-tuning was performed (using the 10% dataset, 40 training epochs), reducing the number of parameters from 175 to 122, and the model file size to 48KB (20% of the initial size of 240KB). Third, quantization optimization was performed by quantizing the weights and activation values ​​from FP32 to INT8, calibrating using the 5% dataset (setting the scaling factor to 128, offset to 0), and fine-tuning was performed after quantization (20 training epochs), reducing memory usage to 30KB (25% of the initial size of 120KB). The input to this sub-unit is a pre-trained model, a teacher model, and a training dataset; the output is a lightweight model (INT8 precision, 48KB size, 8ms inference time), which meets the requirements for embedded deployment.

[0097] 2.4 Model Training Subunit: The specific implementation process is divided into three stages. The parameter settings of each stage follow the industry standard for neural network training and embedded deployment, as follows: The first stage is the pre-training stage. The dataset is divided into a training set (1400 samples), a validation set (400 samples), and a test set (200 samples) in a conventional ratio of 7:2:1. The model parameters are initialized using He normal initialization (weight mean 0, standard deviation √(2 / 10)) and zero initialization (bias). The hyperparameters are set to lr=0.001, Epochs=200, Batch=32 (a conventional hyperparameter combination). The optimizer is Adam (β1=0.9, β2=0.999, default parameters), and an early stopping strategy is adopted (continuous 2... The first stage involves training the pre-trained model (validation set RMSE = 0.018) without any decrease in RMSE on the validation set. The second stage is the knowledge distillation training stage, where the teacher model is trained first (hyperparameters set to lr = 0.0005, Epochs = 300, Batch = 32), and then the pre-trained student model and teacher model are loaded. The distillation training hyperparameters are set to lr = 0.0001, Epochs = 150, Batch = 32, and the optimal student model is saved. The third stage is the pruning and quantization optimization stage, where parameter pruning and fine-tuning, quantization calibration and fine-tuning are performed. The model performance is evaluated using a test set, and the final result meets the common indicators for embedded deployment (error ≤ 3%, inference time ≤ 500ms, model size ≤ 50MB). The inputs to this sub-unit are the feature dataset (training set / validation set / test set), the pre-trained model, and the teacher model; the outputs are the final lightweight SOH prediction model and the model performance evaluation report.

[0098] Step 3: The hybrid intelligent reasoning and uncertainty quantification module performs reasoning and confidence calculation.

[0099] This module is the core inference unit of the diagnostic system, consisting of a neural network inference subunit, a temporal fuzzy logic inference subunit, and an uncertainty quantization subunit. It adopts a "parallel processing-serial fusion" architecture with a processing latency of ≤8ms. Specific implementation details are as follows: 3.1 Neural Network Inference Subunit: The input is designed to be the optimal feature set of the current time step and the previous two time steps ( (Time interval 1 hour), the output is the initial decay probability. ([0, 1] interval, multiplied by 100 to convert to percentage); The network structure adopts a lightweight structure of "input layer - temporal correlation layer - output layer". The input layer includes a batch normalization operation (parameters are the same as before). The temporal correlation layer includes 1 hidden layer (with 6 neurons), a Leaky ReLU activation function (slope 0.01), and a temporal attention mechanism (weight A = [0.6, 0.3, 0.1]). The output layer uses a Sigmoid activation function; The training process uses the training dataset of this embodiment (1400 groups of samples) for synchronous training, an Adam optimizer (lr = 0.0005, decay coefficient 1e - 5, β1 = 0.9, β2 = 0.999), an MSE loss function (label P = 1 - SOH), an early stopping strategy (the validation set loss does not decrease for 15 consecutive rounds), and the validation set loss of the trained model = 0.008; During the inference stage, while outputting the preliminary decay probability, record the probabilities at the previous 7 moments to form a sequence P = (time span 7h); The input of this subunit is the optimal feature set at the current and the previous 2 moments, and the historical decay probability sequence; The output is the preliminary decay probability and the historical decay probability sequence (length 7).

[0100] 3.2 Temporal fuzzy logic inference subunit: First, define fuzzy variables. The input variables are the current decay probability P ([0, 100%]) and the probability change rate V (daily average change rate, V = 1 / 6×Σ(P_t - i + 1 - P_t - i), i = 1~6, unit % / h), and the output variable is the decay level G (including 4 fuzzy subsets: normal, slight decay, moderate decay, severe decay); Subsequently, design membership functions. Use triangular / trapezoidal membership functions to perform fuzzy processing on P (low / medium / high), V (slow / medium / fast), and G (four types of levels) respectively. Among them, the low membership function of P: =max(0, (30 - P) / 30) (P ≤ 30%), the medium membership function: =max(0, min((P - 10) / 20, (50 - P) / 20)) (10% < P < 50%), the high membership function: =max(0, (P - 30) / 70) (P ≥ 30%), the slow membership function of V: =max(0, (0.5 - V) / 0.5) (V ≤ 0.5% / h), the medium membership function: =max(0, min((V - 0.2) / 0.3, (0.8 - V) / 0.3)) (0.2% / h < V < 0.8% / h), the fast membership function: =max(0, (V - 0.5) / 0.5) (V ≥ 0.5% / h); A fuzzy rule base is then constructed, and 12 core rules are designed based on expert experience and attenuation mechanisms (e.g., P≤30% and V≤0.5% / h → normal, P≥60% and V≥0.8% / h → severe attenuation, etc.). Finally, fuzzy inference and defuzzification are performed using the Mamdani inference method (max-min inference), traversing the rule base to calculate activation strength, synthesizing the output fuzzy set, and using the centroid method to defuzzify, transforming the fuzzy set into precise level values. The final attenuation level is determined based on the maximum membership degree. The input to this sub-unit is the initial decay probability. Historical decay probability sequence; output is the final decay level (normal / mild / moderate / severe), and the membership vector μ=[ , , , ].

[0101] 3.3 Uncertainty Quantification Subunit: First, the membership value corresponding to the final decay level output by the temporal fuzzy logic inference subunit is taken as the basic weight W1 (if the normal level is taken as...). Then, count the number of times the current decay level matches the previous level across the seven historical time points. Calculate the consistency coefficient W2= / 7; Then calculate the confidence level Conf=(0.6×W1+0.4×W2)×100% to strengthen the dominant role of the current membership; if Conf<50%, trigger the re-inference mechanism, use the feature set of the previous 5 time steps to re-input the neural network inference subunit, optimize and infer again (re-inference delay ≤10ms) to ensure that the confidence level Conf of the final output is ≥50%; The inputs to this sub-unit are the final attenuation level, the membership vector of each level, and the historical attenuation level sequence; the outputs are the diagnostic confidence Conf (%) and the uncertainty correction signal (0 when no correction is needed, and 1 when correction is required).

[0102] Step 4: Hierarchical Diagnosis Closed-Loop Optimization Decision Module Executes Hierarchical Decisions and Actions.

[0103] This module serves as the decision output unit of the diagnostic system. It comprises a diagnostic status classification subunit, a two-dimensional decision-making subunit, a hierarchical action execution subunit, and a dynamic update feedback subunit, forming a closed-loop decision-making mechanism with a decision latency of ≤15ms. Specific implementation details are as follows: 4.1 Diagnostic Status Classification Subunit: Four types of diagnostic status are defined, specifically including the normal status (SOH ≥ 0.9, attenuation level is "normal", the fluctuation range of core features (IC peak value, equivalent internal resistance) ≤ 5%, corresponding attenuation probability P ≤ 10%), the slight attenuation status (SOH ∈ [0.8, 0.9), attenuation level is "slight attenuation", the fluctuation range of core features is 5% - 15%, corresponding attenuation probability 10% < P ≤ 30%), the moderate attenuation status (SOH ∈ [0.6, 0.8), attenuation level is "moderate attenuation", the fluctuation range of core features is 15% - 30%, corresponding attenuation probability 30% < P ≤ 60%), and the severe attenuation status (SOH < 0.6, attenuation level is "severe attenuation", the fluctuation range of core features > 30%, corresponding attenuation probability P > 60%); A triple verification mechanism of "SOH determination + attenuation level determination + core feature verification" is adopted (the verification weights are 0.4, 0.3, and 0.3 respectively). When the determination results of the three are inconsistent, the core feature verification result shall prevail, and an exception review process shall be triggered (review delay ≤ 30ms); The input of this subunit is the SOH prediction value, attenuation level, and fluctuation data of core features (incremental capacity peak value, equivalent internal resistance); the output is the verified diagnostic status (normal / slight / moderate / severe attenuation).

[0104] 4.2 Two-Dimensional Decision-Making Subunit: First, confidence levels are divided into high confidence (Conf>90%), medium confidence (70%≤Conf≤90%), and low confidence (Conf<70%). Then, a two-dimensional decision matrix is ​​constructed, clarifying the decision-making actions, execution time limits, and responsible parties for each cross-node. The core priority principle is to prioritize severe degradation and avoid misjudgment in normal states. Specific decision rules are as follows: Under high confidence, normal state is judged as normal operation with no action; maintenance personnel record the status once a week. Slight degradation triggers a "mild warning," and the system automatically pushes maintenance prompts (optimizing charging strategy, reducing charging current from 0.3C to 0.27C, a 10% reduction in speed), and maintenance personnel complete the adjustment within one working day. Moderate degradation triggers a "moderate warning," immediately initiating targeted maintenance (single cell equalization voltage ±5mV, reducing discharge depth from 100% to 80%), and maintenance personnel arrive on-site within 2 hours. Severe degradation triggers a "severe warning + shutdown intervention," the system immediately disconnects the battery unit, and maintenance personnel arrive within 30 minutes. Upon arrival, replace the battery with a backup (same model, 200Ah) and complete safety testing; under medium confidence, the normal state is marked as "suspected anomaly pending verification," and maintenance personnel retest core characteristics twice a week; under slight degradation, initiate "expert verification + characteristic retesting," and issue a trend assessment report within 3 working days; under moderate degradation, initiate "emergency expert verification + full-dimensional detection," and issue a decision report within 24 hours; under severe degradation, immediately initiate "expert consultation + comprehensive detection," and complete risk assessment and emergency response within 6 hours; under low confidence, the normal state is marked as "key tracking," and the system performs full-characteristic monitoring + historical trend analysis once a day; under slight degradation, mark as "continuous tracking," monitor once every 3 days, and upgrade to medium confidence handling if the confidence level rises to ≥70% twice consecutively; under moderate degradation, mark as "emergency tracking," monitor once every 12 hours, and upgrade to medium confidence handling if the data confirms no change in degradation level for 3 consecutive times; under severe degradation, mark as "special tracking," monitor once every hour, and simultaneously initiate backup battery switching (switching delay ≤1min). The output of this sub-unit is standardized decision instructions (including action type, execution time limit, responsible party, and monitoring requirements).

[0105] 4.3 Hierarchical Action Execution Subunit: Based on decision-making instructions, a standardized execution system is constructed, clarifying the execution specifications and acceptance standards for actions at each level; for example, a maintenance prompt is pushed for a high-confidence, slight degradation status, reducing the charging current from 0.3C to 0.27C, to be completed within one working day, with the acceptance standard being a charging efficiency ≥90%; after the action is executed, the maintenance personnel enter the execution results, forming an execution closed loop (closed loop completion time ≤24 hours), and outputting the execution results, feature retest data, and effect evaluation report.

[0106] 4.4 Dynamic Update Feedback Subunit: Based on three consecutive monitoring data (1-hour intervals), status updates are performed according to upgrade / downgrade thresholds (confidence change ≥5%). The decision-making effectiveness is evaluated weekly based on indicators such as core feature fluctuation ≤10%, decay probability rate ≤0.2% / h, and SOH decrease rate ≤0.01 / month. If these indicators are not met, the strategy is optimized within 24 hours. 200 sets / half-year evaluation and status data are used as supplementary training data and fed back to the preceding module. The model is fine-tuned every six months (50 rounds). Decision-making cases are collected weekly to build an operation and maintenance knowledge base, outputting status update instructions, model optimization data, and knowledge base update content.

[0107] Step 5: Closed-loop collaboration and optimization across the entire supply chain

[0108] The four core modules interact via CAN bus 2.0 (communication rate 500kbps). The optimal feature set output by the multi-dimensional dynamic feature extraction module (delay ≤20ms) is synchronously sent to the modeling and inference modules. The SOH prediction value output by the modeling module is fed back to the feature extraction and inference modules. The attenuation level and confidence level output by the inference module are sent to the decision module. The optimized data and execution results output by the decision module are fed back to all preceding modules. The entire chain of technical parameters and strategies is dynamically optimized every six months, forming a complete diagnostic closed loop. This enables accurate diagnosis and intelligent operation and maintenance of abnormal attenuation of 18650 lithium-ion power battery packs, effectively supporting the large-scale and safe operation of energy storage batteries.

[0109] In summary, through the specific design and coordinated operation of the above modules, this embodiment realizes the engineering implementation of the abnormal degradation diagnosis technology solution for energy storage batteries of the present invention. It can be directly applied to abnormal degradation diagnosis scenarios of lithium-ion power battery packs, effectively supporting the operation and maintenance needs of large-scale and safe operation of energy storage batteries, and has good engineering application value.

[0110] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0111] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A hierarchical diagnostic method for abnormal degradation of energy storage batteries that combines data and mechanisms, characterized in that, The specific steps are as follows: Step 1: Multi-dimensional dynamic feature extraction. Perform data preprocessing, voltage range feature identification based on sliding window, multi-dimensional dynamic feature extraction and feature optimization and screening operations on the original operating data of the energy storage battery. Obtain the optimal feature set characterizing the battery aging state through voltage, capacity and temperature standardization formulas and feature calculation formulas. Step 2: Construction and training of a lightweight neural network model with embedded physical constraints. A four-layer lightweight neural network architecture integrating MLP and PINN is built, embedding relevant formulas for the continuity of charged state constraints and thermodynamic energy balance constraints. A combined optimization strategy of knowledge distillation, parameter pruning, and quantization is used to compress the model size. The model training is completed in three stages: pre-training, knowledge distillation training, and parameter pruning and quantization optimization, through loss function fusion formula, to obtain a high-precision and lightweight SOH prediction model. Step 3: Hybrid intelligent reasoning and uncertainty quantification. Construct a hybrid reasoning architecture that combines neural networks and temporal fuzzy logic. The neural network submodule outputs the initial battery decay probability, which is then converted into a semantic decay level by the temporal fuzzy logic submodule using probability change rate calculation, membership function, and defuzzification formula. Finally, uncertainty quantification is completed through confidence weighted calculation. Step 4: Hierarchical diagnostic decision-making. Based on the dual-dimensional results of attenuation level and confidence level, perform diagnostic status classification, dual-dimensional judgment, and hierarchical action execution to form standardized operation and maintenance decision instructions. Step 5: Dynamic updates and feedback. Based on the operation and maintenance execution results and subsequent monitoring data, complete the upgrade / downgrade determination of battery diagnostic status, evaluate the decision effect, feed back relevant data to the preceding modules to complete model iteration and optimization, and build an operation and maintenance knowledge base to realize the reuse of decision-making experience. Step Six: Full-chain closed-loop verification. Verify the output results of the above steps from multiple dimensions to ensure the effectiveness of feature extraction, the accuracy of model prediction, the reliability of reasoning conclusions, and the implementation of decision-making actions, forming a full-chain diagnostic closed loop of "feature extraction - modeling and reasoning - decision-making and maintenance - feedback optimization".

2. The hierarchical diagnostic method for abnormal degradation of energy storage batteries based on data and mechanism synergy as described in claim 1, characterized in that: The data preprocessing in step one includes raw data screening and cleaning, and data standardization. Raw data screening and cleaning removes cyclical data from the final charging stage with a sudden drop in current, charging times below a set threshold, and abnormal fluctuations in the SOH label. Data standardization maps voltage, capacity, and temperature data to the [0,1] interval to eliminate the influence of individual battery parameter differences and operating condition fluctuations. The data standardization formula is as follows: Voltage standardization: Capacity standardization: Temperature standardization: In the formula, This is the original voltage data. The rated maximum charging voltage of the battery. This is the battery discharge cutoff voltage; Q represents the cumulative capacity on a single charge, and Q is the rated capacity of the battery. The raw temperature data, For the battery's maximum allowable operating temperature, The minimum allowable operating temperature for the battery.

3. The hierarchical diagnostic method for abnormal degradation of energy storage batteries based on data and mechanism synergy, as described in claim 1, is characterized in that: The voltage range feature recognition based on the sliding window described in step one adopts a strategy combining equal interval division and adaptive sliding window. First, the charging voltage range is divided into basic windows according to ΔV∈[0.005V,0.05V]. Then, the sliding window length L∈[3,10] and sliding step size S∈[1,3] are adaptively adjusted according to the SOH value of the previous round. Basic parameters such as cumulative charging capacity, charging time, and temperature change in each window are calculated. The adaptive adjustment formula is as follows: The cumulative capacity within the window is calculated by integrating the current: During constant current charging In the formula, The window start time, The end time of the window. This is the charging current. The charging current is constant.

4. The hierarchical diagnostic method for abnormal degradation of energy storage batteries based on data and mechanism synergy as described in claim 1, characterized in that: The multi-dimensional dynamic feature extraction and simultaneous fusion of incremental capacity analysis and internal resistance assessment technology described in Step 1 extracts three major categories of features from three dimensions: capacity, impedance, and temperature. These features include basic statistical features, incremental capacity features, and internal resistance temperature coefficient change rate features. This achieves a holographic characterization of the battery aging state. The core calculation formulas include: Cumulative average capacity within the window: Cumulative capacity standard deviation within the window: Mean rate of temperature change within the window: Rate of voltage change within the window: Incremental capacity: Incremental capacity peak offset: Area of ​​the incremental capacity curve: Temperature coefficient of internal resistance: Rate of change of internal resistance temperature coefficient: Internal resistance temperature coefficient stability index: In the formula, q is the cumulative capacity of the k-th basic window within the sliding window, L is the length of the sliding window, Δp is the temperature change of the k-th basic window within the sliding window, Δt is the charging time of the basic window, Vs is the starting voltage of the sliding window, and Ve is the ending voltage of the sliding window. This represents the cumulative change in capacity within the sliding window. This represents the voltage change within the sliding window. This represents the voltage corresponding to the peak value of the current window's incremental capacity. Let IC(V) be the peak voltage of the incremental capacity under healthy battery conditions, and let IC(V) be the function of incremental capacity as a function of voltage. Window start temperature The corresponding equivalent internal resistance, The end temperature of the window The corresponding equivalent internal resistance, This is the internal resistance temperature coefficient of the previous sliding window. The standard deviation of the temperature coefficient of internal resistance within the window. This represents the average value of the temperature coefficient of internal resistance within the window; The feature optimization screening adopts a three-step method: "correlation analysis - feature importance assessment - redundancy removal". First, Pearson correlation analysis is used to screen features that are strongly correlated with SOH. Then, the random forest algorithm is used to assess the importance of features. Finally, highly redundant features are removed to obtain the optimal feature set. The formula for calculating the Pearson correlation coefficient is as follows: The formula for calculating the Gini coefficient is as follows: In the formula, X is the feature variable, Y is the SOH value, Cov(X,Y) is the covariance of X and Y, Var(X) is the variance of X, Var(Y) is the variance of Y, p is the probability of the k-th value of feature X, and K is the number of values ​​of feature X.

5. The hierarchical diagnostic method for abnormal degradation of energy storage batteries based on data and mechanism synergy as described in claim 1, characterized in that: The four-layer lightweight neural network architecture described in step two consists of a feature input layer, a feature enhancement layer, a physical constraint embedding layer, and a lightweight output layer. The feature input layer performs batch normalization on the optimal feature set. The feature enhancement layer uses a lightweight MLP structure to enhance features. The physical constraint embedding layer fuses the output of the feature enhancement layer with physical constraints. The lightweight output layer uses a single neuron + 1×1 convolutional kernel structure to output the SOH prediction value. The batch normalization formula for the feature input layer is: The formula for calculating the feature enhancement layer is: The formula for calculating the lightweight output layer is: in, The mean of the batch data. Let γ and β be the variance of the batch data, β be learnable parameters, and ε be a small value to prevent the denominator from being zero (ε = 1e-5). , Here is the weight matrix for each hidden layer. , Here are the bias vectors for each hidden layer. , The output feature maps of each hidden layer. The weight vector of the output layer. For the bias term of the output layer, Feature maps for incorporating physical information.

6. The hierarchical diagnostic method for abnormal degradation of energy storage batteries based on data and mechanism synergy as described in claim 1, characterized in that: Step two involves embedding physical constraints by transforming the state of charge (SOC) continuity constraints and thermodynamic energy balance constraints into constraint loss functions. These are then weighted and fused with the model prediction loss function to form the total loss function. A backpropagation algorithm is used to achieve deep integration of physical constraints and data-driven processes. The SOC continuity constraints include SOC temporal continuity constraints and SOC-voltage relationship constraints; the thermodynamic energy balance constraints include energy conservation constraints and temperature change rate constraints. These constraints are fused and embedded through loss functions to form the model's total loss function. The formula for the SOC temporal continuity constraint is: In the formula, Let be the predicted SOC value at time t. The SOC prediction value at time t-1; the SOC-voltage relationship constraint formula is: g(u) = a3u3 + a2u2 + a1u1 + a0, where, Let SOC be the voltage mapping function, ε be the allowable error, and ε ≤ 0.02; the energy conservation constraint formula is: In the formula, The total electrical energy input during the charging process. The change in chemical energy stored in a battery. The generated heat energy; the temperature change rate constraint formula is: In the formula, k1 is the proportionality coefficient, and k2 is the environmental temperature influence coefficient. Let be the equivalent internal resistance at time t; the formula for the total loss function is: In the formula, α is the weighting coefficient, and α∈[0.6, 0.8]. For predicting the loss function, Here, N is the constraint loss function, and N is the number of samples. Let SOH be the predicted value for the i-th sample. Let be the true SOH value of the i-th sample, and λ1, λ2, λ3, and λ4 be constraint weight coefficients that satisfy λ1 + λ2 + λ3 + λ4 = 1. For SOC temporal continuity constraint loss, The loss is constrained by the SOC-voltage relationship. Losses are due to energy conservation constraints. The loss is constrained by the rate of temperature change.

7. The hierarchical diagnostic method for abnormal degradation of energy storage batteries based on data and mechanism synergy as described in claim 1, characterized in that: In step two, the pre-training phase of model training divides the battery's entire lifecycle cyclic data into training, validation, and test sets in a 7:2:1 ratio, and uses the Adam optimizer and early stopping strategy to complete model initialization training. In the knowledge distillation training phase, a deep MLP teacher model and a lightweight student model are constructed, employing a "soft label + hard label" dual-label training method to achieve knowledge transfer. In the parameter pruning and quantization optimization phase, redundant parameters are first removed and fine-tuned, then the model weights and activation values ​​are quantized from 32-bit floating-point to 8-bit integers, ultimately obtaining a lightweight model suitable for engineering deployment. The model performance must meet the following requirements: RMSE ≤ 0.015, MAE ≤ 0.01, inference time ≤ 10ms, and model size ≤ 5MB. The distillation loss function formula for the knowledge distillation training is as follows: In the formula, μ is the hard label weight coefficient, and μ∈[0.3, 0.5]. The MSE loss between the student model's predicted values ​​and the actual SOH values; The KL divergence loss is used to calculate the probability distributions of the student model and the teacher model. Let i be the output probability distribution of the teacher model for the i-th sample. Let i be the output probability distribution of the i-th sample in the student model; The quantization optimization formula is as follows: In the formula, It is a 32-bit floating-point number. It is the quantized 8-bit integer. The maximum value of the parameter. This is the minimum value of the parameter.

8. The hierarchical diagnostic method for abnormal degradation of energy storage batteries based on data and mechanism synergy as described in claim 1, characterized in that: In step three, the neural network submodule takes the optimal feature sets of the current time step and the previous two time steps as input, adopts a lightweight structure of "input layer - temporal correlation layer - output layer", introduces a temporal attention mechanism to strengthen the weights of recent features, outputs the initial decay probability and forms a historical probability sequence; the temporal fuzzy logic submodule takes the historical decay probability sequence as input, defines the current decay probability and probability change rate as fuzzy input variables, and decay level as fuzzy output variable, designs triangular and trapezoidal membership functions and 12 core fuzzy rules, and uses the Mamdani inference method and centroid method to complete fuzzy inference and defuzzification, outputting a semantic decay level; the uncertainty quantification uses the membership value corresponding to the decay level as the basic weight and the consistency coefficient of the historical decay level sequence as the secondary weight, and calculates the confidence level by weighting them in a ratio of 0.6:0.4; if the confidence level is lower than 50%, a re-inference mechanism is triggered, the initial decay probability is optimized and inference is performed again to ensure that the confidence level is ≥50%; the input of the neural network submodule is The core calculation formula is as follows: In the formula, For batch normalization operations, This is the temporal attention weight vector. For element-wise product operation, , This is the weight matrix. , It is the bias vector; The formula for calculating the probability change rate in the temporal fuzzy logic submodule is as follows: In the formula, V is the daily average rate of change of probability, V>0 indicates that the decay is intensifying, and V<0 indicates that the decay trend is easing. The centroid method for resolving fuzziness is as follows: In the formula, To obtain the accurate attenuation level value after deblurring, The attenuation levels are quantified as follows: 0 = Normal, 1 = Slight attenuation, 2 = Moderate attenuation, 3 = Severe attenuation. This represents the final membership degree of the corresponding level; The confidence level calculation formula for the uncertainty quantification is as follows: In the formula, This represents the membership value corresponding to the final attenuation level. The sequence consistency coefficient, =7, This refers to the number of times in the historical sequence that the current decay level is consistent with the current level.

9. The hierarchical diagnostic method for abnormal degradation of energy storage batteries based on data and mechanism synergy as described in claim 1, characterized in that: Step four, in its diagnostic status classification, combines SOH value, attenuation level, and core characteristic fluctuation amplitude to divide the battery into four categories: normal SOH ≥ 0.9, slight attenuation 0.8 ≤ SOH < 0.9, moderate attenuation 0.6 ≤ SOH < 0.8, and severe attenuation SOH < 0.

6. A triple verification mechanism of "SOH determination + attenuation level determination + core characteristic verification" is adopted. The dual-dimensional determination constructs a two-dimensional decision matrix of "diagnostic status × confidence level". The confidence level is divided into high confidence Conf>90%, medium confidence 70% ≤ Conf ≤ 90%, and low confidence Conf<70%. Each intersection node corresponds to a unique operation and maintenance decision action. The hierarchical action execution clarifies the execution specifications, responsible parties, and execution time limits of each decision action according to the operation and maintenance priority, realizing the standardized implementation of decision instructions.

10. The hierarchical diagnostic method for abnormal degradation of energy storage batteries based on data and mechanism synergy as described in claim 1, characterized in that: The dynamic update and feedback described in step five are based on the changes in diagnostic status and confidence level monitored for three consecutive times, and the status is upgraded / downgraded. The decision-making effect is evaluated using the fluctuation amplitude of core features, the rate of change of decay probability, and the rate of decrease of SOH as evaluation indicators. If the criteria are not met, strategy optimization is triggered. The evaluation data and status update data are fed back to the preceding module as 10% of the supplementary training data, and the model fine-tuning training is completed every six months.

Images