Health assessment method and system for retired battery

Abstract

The invention discloses a health assessment method and system for a decommissioned battery, and relates to the technical field of battery state assessment, and the method comprises the steps: recognizing a decommissioned battery identifier, automatically generating an assessment task number, and setting the battery state as "to-be-detected"; a battery charging mode is set, battery state data are collected in real time, battery features are extracted, and a feature vector G is constructed; and inputting the feature vector G into a parallel attention neural network to carry out static SOH estimation to obtain a preliminary SOH estimated value, carrying out SOH correction by using joint extended Kalman filtering, and obtaining the current internal resistance. According to the invention, the method can achieve the efficient and precise evaluation of the health state of the decommissioned battery, remarkably improves the comprehensiveness and deep modeling capability of feature extraction, improves the reliability and dynamic adaptability of an SOH estimation result, effectively reduces the error fluctuation, and improves the generalization capability of an evaluation system for a complex decommissioned state.

Description

Technical Field

[0001] The present invention relates to the technical field of battery status assessment, and in particular to a health assessment method and system for retired batteries. Background Art

[0002] With the rapid development of new energy vehicle industry and renewable energy system, lithium-ion batteries are widely used as core energy storage components in multiple scenarios such as power batteries, grid energy storage, and power tools. However, after a certain number of charge and discharge cycles, the battery will gradually reach its performance degradation threshold and be retired. According to existing industry data, a large-scale retired battery market will be formed in the next few years. In order to improve resource utilization, reduce environmental pressure and promote cascade utilization applications, it is urgent to conduct scientific and effective health status assessment of retired batteries. Existing assessment methods mostly rely on static capacity testing, internal resistance measurement or full life cycle experiments. These methods not only have long test cycles and high operating costs, but also have difficulty adapting to the reality of the complex sources and strong discrete performance of retired batteries. In recent years, with the integration and application of big data and artificial intelligence technologies in battery management systems, data-driven SOH estimation and life prediction methods have developed rapidly, but in the retired battery health assessment scenario, they still face outstanding problems such as poor modeling generalization, redundant feature dimensions, insufficient estimation accuracy, and lack of multi-source decision fusion.

[0003] Existing technologies have limitations in many key aspects. First, in terms of feature extraction, most methods rely only on simple statistical features, such as constant current duration and termination voltage, and fail to tap the deep health features hidden in time series data such as voltage and current, especially lacking comprehensive utilization of wavelet frequency domain, differential signal, and dynamic trend indicators. Second, in terms of SOH estimation methods, existing technologies mostly use a single model (such as neural network or Kalman filter) to process static data, lack a mechanism to integrate the advantages of multiple models for joint modeling and adaptive correction, and cannot take into account both prediction accuracy and dynamic tracking capabilities. Third, the feature selection process mostly relies on manual experience or simple filtering, lacks intelligent and optimization-driven feature screening strategies, and is prone to missing key features or introducing redundant information, affecting model performance. Finally, in terms of evaluation output, the evaluation results are often presented as a single indicator, lacking systematic report packaging and platform access capabilities, and it is difficult to meet the industry's needs for battery full life cycle management. Therefore, existing retired battery health assessment technologies generally have problems such as weak modeling capabilities, low prediction accuracy, poor feature utilization, and insufficient platform integration, making it difficult to support the engineering application of large-scale retired battery cascade utilization in the future. Summary of the invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides a method and system for health assessment of retired batteries, which solves the problems commonly existing in the existing health assessment technologies of retired batteries, such as weak modeling ability, low prediction accuracy, poor feature utilization rate, and insufficient platform integration, and it is difficult to support the engineering application of large-scale retired battery secondary utilization in the future.

[0006] To solve the above technical problems, the present invention provides the following technical solutions: In the first aspect, the present invention provides a method for health assessment of retired batteries, which includes: Identifying the retired battery identifier and automatically generating an assessment task number, and setting the battery status to "to be detected"; Setting the battery charging mode, collecting battery status data in real time, and extracting battery features to construct a feature vector G; The battery status data includes voltage, current, and time data; Inputting the feature vector G into a parallel attention neural network for static SOH estimation to obtain a preliminary SOH estimated value, using an extended Kalman filter for SOH correction, and obtaining the current internal resistance; Estimating the maximum available power of the battery based on the battery internal resistance and real-time voltage , and based on SOH and Performing battery health level classification and application scenario recommendation; Packaging all data into an assessment report and archiving it in the cloud for storage.

[0007] As a preferred solution of the method for health assessment of retired batteries according to the present invention, wherein: the setting of the battery charging mode, collecting battery status data in real time, and extracting battery features to construct a feature vector G means collecting the voltage, current, and time data of the battery and recording them as the original data set; After preprocessing the data in the original data set, extracting the primary features of the data, including the charging duration in the constant current stage , the charging duration in the constant voltage stage , the rising rate of the middle voltage section , the main peak value of the incremental capacity , the corresponding voltage position , and the peak time point of the differential voltage curve ; Integrating the extracted features to form a primary feature set ; Using Daubechies4 wavelet basis function to perform 3-layer wavelet decomposition on the voltage and current in the data set to obtain six groups of sub-signals, for each sub-signal, extracting features and integrating them to generate a wavelet feature set ; Combining the feature sets and Perform merging to construct the total feature set F; Fully discharge each battery to obtain the true capacity label a. Use the Pearson correlation evaluation method to calculate the correlation between each dimension of features and the true capacity label. Retain the features with a correlation greater than the preset threshold and integrate them into the filtered feature set F'. Construct a binary selection variable for each feature in F' and construct a binary vector of length k , where: indicates whether the k-th feature in the feature set is selected, and each B corresponds to a feature subset combination; Use the sparrow search algorithm to search for the optimal feature subset; Based on the feature data, use the XGBoost regressor to obtain the predicted capacity. Calculate the average error between the predicted capacity and the true capacity as the fitness value, and set the individual with the lowest fitness as the global optimal solution ; For the discoverer, execute the exponential decay perturbation update strategy; Globally guide and adjust the remaining individuals with reference to the current global optimal solution; Mark the individual with the worst current fitness as the vigilante, execute mutation perturbation, and jump out of the local optimum; Perform probability mapping on the continuous coding vectors of all individuals through the Sigmoid function, and then perform threshold binarization to obtain the updated binary coding individuals. After reaching the maximum number of iterations W, output the final global optimal coding solution A. Extract the selected feature dimensions according to the optimal coding A, form the optimal feature subset, perform normalization processing, and splice them into a numerical vector G.

[0008] As a preferred scheme of the health assessment method for retired batteries described in the present invention, wherein: the step of inputting the feature vector G into the parallel attention neural network for static SOH estimation to obtain the preliminary SOH estimation value includes: Divide the feature vector G into two sub-vectors according to the feature source, and use them for different model channel processing respectively; Extract from the main variables reflecting the evolution of the battery charge characteristics, denoted as the signal sequence x(t). Use the empirical mode decomposition algorithm to decompose the time series signal into several intrinsic mode functions and residuals; Select the first two order IMF components and for reconstructing the main sequence ; Combine the features of all time steps into an input tensor , where t is the time step and d is the number of input feature dimensions at each time step; Construct a time channel using an LSTM model, construct the first layer LSTM-1, initialize the input dimension, the number of hidden units, and the activation function, and input the Use a linear layer to increase the dimension to the same dimension to match the residual dimension, and normalize the output vector for each time step; Send the normalized output into the second layer LSTM-2 network, repeat the residual mechanism, perform a temporal average pooling operation on the output vector, send the pooled vector into a fully connected network, and output it by the output layer; Collect historical battery data as the training set and input it into the LSTM model for iterative training. Define the mean squared error as the training objective loss function and the Adam optimizer to iteratively optimize the model parameters. If the validation error does not decrease for P consecutive rounds, stop the iteration and obtain the optimal model parameters as the initial parameter set E; Expand E into a one-dimensional vector, use it as the initial position of the particle and initialize the APSO population, set the population size, and initialize a velocity vector for each particle at the same time; Start iterative optimization, set the maximum number of rounds, perform velocity and position updates for each particle, calculate the fitness value corresponding to each particle, and update the individual optimal and global optimal solutions; After completing the global search iteration of APSO, directly extract the current global optimal particle position as the final optimization solution and restore it to the LSTM model parameter structure, and load the optimized parameters into the original LSTM model to obtain the optimized model; Input the into the optimized LSTM model to obtain a preliminary SOH prediction value; Use Attention to construct a morphological channel, set the multi-head attention parameters, including the number of attention heads and the output dimension of each head, and construct three sets of mapping matrices for each attention head. For Perform linear transformation to obtain query Q, key K, and value V tensors, perform self-attention calculation for each attention head, and splice the output vectors of all heads along the dimension to form the final attention modeling output, representing the health state characterization vector modeled by the current morphological feature channel; Receive the outputs of the two model channels and splice them to form a fusion vector V. Construct an MLP network structure, including an input layer, a hidden layer, and an output layer, and input V into the MLP to obtain a preliminary SOH estimate .

[0009] As a preferred solution of the health assessment method for retired batteries described in the present invention, wherein: the use of the extended Kalman filter for SOH correction and obtaining the current internal resistance includes: Collect the rated capacity of the battery at the factory , and calculate the initial capacity estimate based on the SOH estimate and the rated capacity; Construct a combined state variable, define and initialize the filtering state vector, where the state vector includes the current state of charge, RC voltage term, equivalent internal resistance, and the current capacity of the battery; Meanwhile, initialize the covariance matrix, state noise, and observation noise based on historical data; Extract the dynamic observation data sequence from the standard charging test process, and use the battery RC equivalent model to predict the state variables, including predicting the SOC and RC polarization voltage; Calculate the current predicted voltage according to the state prediction value ; Calculate the error between the actual voltage and the predicted voltage; Update the state vector and state covariance using the Kalman gain formula of the standard EKF; After all sampling points are iterated, obtain the final state vector estimate, including the current state of charge, the estimated value of the RC polarization voltage component, the current estimated value of the equivalent internal resistance, and the current capacity estimate; Extract the capacity estimate from the state vector estimate, and calculate the final health estimate using the estimated capacity and the rated capacity 。

[0010] As a preferred solution of the method for evaluating the health of retired batteries according to the present invention, wherein: estimating the maximum available power of the battery based on the internal resistance and real-time voltage of the battery , and based on SOH and The battery health level classification and application scenario recommendation include: Extract the current estimated value of the equivalent internal resistance from the state vector estimate , and estimate the maximum available power of the battery based on the real-time voltage U ; Integrate the obtained final health estimate and the maximum available power of the battery to generate an input feature vector. Use a random forest as a classification model, define the health level labels as A level, B level, and C level, corresponding to high health, medium health, and low health respectively. Input the input feature vector into the random forest model to obtain the corresponding health level label, and automatically recommend the corresponding application scenario according to the classification result.

[0011] As a preferred solution of the method for evaluating the health of retired batteries according to the present invention, wherein: identifying the retired battery identifier and automatically generating an evaluation task number, and setting the battery state to "to be detected" means that at the battery warehousing link, the RFID tag information attached to the battery is read by an RFID scanning device and used as the unique number of the battery. After obtaining the battery number, synchronously generate the evaluation task number corresponding to the battery, bind the task number to the battery number one by one, and set the initial state of the task to "to be detected" after the information is bound.

[0012] As a preferred solution of the health assessment method for retired batteries according to the present invention, wherein: the encapsulation of all data into an assessment report and cloud archiving storage means integrating the basic battery information and the data generated by analysis to generate a complete battery health assessment report and uploading it to the cloud database platform.

[0013] In a second aspect, the present invention provides a health assessment system for retired batteries, including A task management module, used to identify the retired battery identifier and automatically generate an assessment task number, and set the battery status to "to be detected"; A feature extraction module, used to set the battery charging mode, collect battery status data in real time and extract battery features to construct a feature vector; An estimation module, used to input the feature vector into a parallel attention neural network for static SOH estimation to obtain a preliminary SOH estimated value; A correction module, used to perform SOH correction using an extended Kalman filter in combination and obtain the current internal resistance; A level classification module, used to estimate the maximum available power of the battery based on the battery internal resistance and the real-time voltage, and perform health level classification and application scenario recommendation based on the SOH; An archiving module, used to encapsulate all data into an assessment report and perform cloud archiving storage.

[0014] In a third aspect, the present invention provides a computer device, including a memory and a processor, wherein: when the computer program is executed by the processor, it implements any step of the health assessment method for retired batteries as described in the first aspect of the present invention.

[0015] In a fourth aspect, the present invention provides a computer-readable storage medium, on which a computer program is stored, wherein: when the computer program is executed by the processor, it implements any step of the health assessment method for retired batteries as described in the first aspect of the present invention.

[0016] The beneficial effects of the present invention are as follows: By introducing a multi-scale feature extraction mechanism based on wavelet decomposition and correlation assessment, combining sparrow search and XGBoost to achieve optimal feature subset screening, and constructing a parallel neural network integrating LSTM and multi-head attention structure for SOH estimation, and then using an extended Kalman filter in combination to achieve state correction and internal resistance estimation, the present invention can efficiently and accurately assess the health status of retired batteries, significantly improve the comprehensiveness of feature extraction and the deep modeling ability, enhance the reliability and dynamic adaptability of the SOH estimation result, effectively reduce the error fluctuation, and improve the generalization ability of the assessment system for complex retired states. Description of the Drawings

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0018] Figure 1 It is a flowchart of the health assessment method for retired batteries in Embodiment 1.

[0019] Figure 2 It is a structural diagram of the health assessment system for retired batteries in Embodiment 1.

[0020] Figure 3 It is a schematic flow diagram of feature optimization in Embodiment 1. Specific Embodiments

[0021] To make the above objects, features, and advantages of the present invention more obvious and understandable, the following will describe the specific embodiments of the present invention in detail with reference to the drawings of the specification.

[0022] In the following description, many specific details are set forth to fully understand the present invention. However, the present invention can also be implemented in other ways different from those described herein. Those skilled in the art can make similar generalizations without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0023] Secondly, the so-called "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that can be included in at least one implementation manner of the present invention. The "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment that excludes other embodiments.

[0024] Embodiment 1, referring to Figures 1 to 3 , is the first embodiment of the present invention. This embodiment provides a health assessment method for retired batteries, including the following steps: S1. Identify the retired battery identification and automatically generate an assessment task number, and set the battery status to "to be detected"; Specifically, identifying the retired battery identification and automatically generating an assessment task number, and setting the battery status to "to be detected" means that in the battery warehousing link, the RFID tag information attached to the battery is read by an RFID scanning device and used as the unique number of the battery. If the reading fails (such as no tag or damaged tag), a unique number based on the current date, test station, and task serial number is automatically generated to ensure that each battery has an independent identification; After obtaining the battery number, synchronously generate the evaluation task number corresponding to the battery. The task number is composed of the current date and the evaluation serial number, and is unique. Bind the task number to the battery number one by one, and set the initial state of the task to "to be detected" after the information is bound.

[0025] Through the RFID or the system's automatic number generation mechanism, ensure that each battery obtains a unique identifier when entering the evaluation system, avoiding data confusion problems caused by duplicate, lost numbers or manual entry errors. The automatic generation and binding of the evaluation task number enables the battery to have a unified indexing system throughout the process from identification to detection. Since some retired batteries may have problems such as missing RFID tags, damaged tags or fallen-off information, the traditional method relying on manual registration or barcode scanning input is inefficient and has a high error rate. The present invention designs a numbering system generated based on the combination of time and work station number, which can ensure the allocation of unique numbers even in the case of missing physical identifiers, enhancing the system's adaptability and fault tolerance to abnormal battery states.

[0026] S2. Set the battery charging mode, collect the battery status data in real time and extract the battery characteristics to construct the feature vector G; Specifically, setting the battery charging mode, collecting the battery status data in real time and extracting the battery characteristics to construct the feature vector G means placing the battery to be tested into the temperature control detection chamber, setting the test environment temperature, and starting the charging process (for example, charging at a constant current of 0.5C until the voltage reaches 4.20V, maintaining the voltage at 4.20V, and continuing to charge until the current drops to 0.05A). During the entire charging process, collect the voltage, current and time data of the battery and record them as the original data set; After preprocessing the data in the original data set, extract the first-level features of the data, including the charging duration in the constant current stage 、the charging duration in the constant voltage stage 、the rising rate of the middle section of the voltage (for example, the average rate of the voltage rising from 3.85V to 4.00V) 、the main peak value of the incremental capacity :

[0027] In the formula, Q is the electric charge and V is the voltage; The corresponding voltage position and the peak time point of the differential voltage curve : ; In the formula, I is the current in the constant current stage and T is the time; Integrate the extracted features to form the first-level feature set ; The voltage and current in the dataset are decomposed by 3-layer wavelet using Daubechies4 (db4) wavelet basis function to obtain six groups of sub-signals. For each sub-signal, 16 types of statistical features are extracted to obtain 128-dimensional DWT features. The extracted features are integrated to generate a wavelet feature set ; The feature sets and are merged to construct the total feature set F; Each battery is fully discharged to obtain the true capacity label a. The Pearson correlation evaluation method is used to calculate the correlation between each dimension of the feature and the true capacity label. The features with a correlation greater than the preset threshold (set by statistical method) are retained and integrated into the filtered feature set F'. A binary selection variable is constructed for each feature in F' to indicate whether it is selected into the final subset, and a binary vector with length o is constructed, where: : indicates that the o-th feature is retained; indicates whether the o-th feature in the feature set is selected; : indicates that the o-th feature is not retained; o is the remaining feature dimension after preliminary screening; Each B corresponds to a feature subset combination; The sparrow search algorithm is used to search for the optimal feature subset. The population size is initialized as N, the maximum number of iterations is W, and the individual roles are divided into the first The individual role is the discoverer, and the remaining individuals are followers. Each individual is initially encoded as a binary vector with length d, where p is the discoverer proportion coefficient, represents the number of individuals calculated by proportion, taking an integer (usually rounding down); Based on the feature data, the XGBoost regressor is used to obtain the predicted capacity. The average error between the predicted capacity and the true capacity is calculated as the fitness value, and the individual with the lowest fitness is set as the global optimal solution ; For the discoverer, an exponential decay perturbation update strategy is executed to promote the search accuracy:

[0028] In the formula, is the continuous position vector of the i-th individual in the t-th round, is the updated position of the i-th individual, is the step size factor controlling the exponential decay rate, W is the maximum number of iterations, is the local perturbation intensity factor, is a standard normal distribution random number, and Set by experimental parameter tuning; Globally guide and adjust the remaining individuals with reference to the current global optimal solution:

[0029] In the formula, is the current position of the j-th follower, is the continuous encoding of the optimal individual in the current population, is the guiding influence factor, set by the grid search method, is a uniformly distributed random number, is the updated position of the follower individual; Mark the individual with the worst current fitness as the vigilant, perform mutation perturbation, and jump out of the local optimum:

[0030] In the formula, is the position of the -th individual with the worst fitness, is the position of an individual randomly selected from the current population, is the jump perturbation amplitude factor, set by the differential optimization algorithm, sign() is the sign function, is the position of the vigilant individual after jumping; Perform probability mapping on the continuous encoding vectors of all individuals through the Sigmoid function, and then perform threshold binarization to obtain the updated binary encoding individuals. After reaching the maximum number of iterations W, output the final global optimal encoding solution A. Extract the selected feature dimensions according to the optimal encoding A, form the optimal feature subset, perform normalization processing, and splice them into a numerical vector G.

[0031] The primary feature set in the feature vector G can directly characterize the electrochemical behavior and performance response during the battery charging process; while the DWT features perform three-layer Daubechies4 wavelet decomposition on the voltage and current curve signals, obtain six groups of sub-signals, and extract statistical parameters such as energy, entropy, skewness, and kurtosis to further reveal the healthy signal features in different frequency bands. This time-frequency joint feature expression method can effectively capture the nonlinear and non-stationary characteristic evolution process of retired batteries. The SSA optimization objective function is the mean absolute error (MAE) between the predicted capacity and the true capacity of the XGBoost regressor, which makes the feature subset search not only reflect dimension optimization but also be deeply coupled with the specific task performance, thus enhancing the model adaptability. Finally, the output optimal feature subset forms a numerical feature vector G after normalization processing. This vector has the characteristics of compactness, high discrimination, and modeling friendliness, providing accurate input for the subsequent static SOH estimation model and dynamic state filter.

[0032] The proposed technical solution has the following beneficial effects: First, through the integration of wavelet decomposition and statistical features, the feature extraction is extended from the traditional single time domain to the time-frequency domain, enhancing the ability to capture latent health degradation information. Second, the correlation screening and intelligent optimization algorithm are used in cooperation to perform feature selection, significantly improving the effectiveness and stability of the selected feature set and reducing the risk of model overfitting. Third, the constructed feature vector G is structurally convenient for input into the deep model for processing, supporting the time modeling and morphological recognition of the subsequent neural network, and improving the accuracy and reliability of the entire health assessment process. Fourth, the overall process has a high degree of automation, is suitable for engineering deployment and batch evaluation requirements, and is particularly applicable to the rapid sorting scenario of retired batteries before cascade utilization.

[0033] S3. Input the feature vector G into the parallel attention neural network for static SOH estimation to obtain a preliminary SOH estimate value, use the extended Kalman filter for SOH correction, and obtain the current internal resistance. Specifically, inputting the feature vector G into the parallel attention neural network for static SOH estimation to obtain a preliminary SOH estimate value includes: Dividing the feature vector G into two sub-vectors according to the feature source, including the time series feature vector ( ), and the morphological feature vector (statistical features such as energy, mean, kurtosis, skewness, and entropy value extracted from six groups of sub-signals under wavelet decomposition (such as db4)), which are respectively used for processing by different model channels. The model channels include the time channel and the morphological channel; Extracting the main variables (such as voltage sequence, voltage rising section, dQ / dV interval) reflecting the evolution of the battery charge characteristics from ), denoted as the signal sequence x(t), and using the empirical mode decomposition algorithm to decompose the time series signal into several intrinsic mode functions and residuals:

[0034] In the formula, is the original signal sequence, is the jth intrinsic mode function, is the residual signal remaining after extracting the nth IMF, representing the trend component, n is the total number of IMFs, and t is the time step; Selecting the first two-order IMF components and for use as the reconstructed main sequence :

[0035] Taking as the new time-like sequence input and uniformly replacing the original Sequence features; Combine the features of all time steps into an input tensor , where t is the time step and d is the number of input feature dimensions at each time step; Use the LSTM model to construct a time channel, construct the first layer LSTM-1, initialize the input dimension, the number of hidden units, and the activation function, and input the Use a linear layer to increase the dimension to the same dimension to match the residual dimension, and normalize the output vector of each time step; Send the normalized output into the second layer LSTM-2 network, initialize the input dimension, the number of hidden units, and the output shape, and repeat the residual mechanism (input the Use a linear layer to increase the dimension to the same dimension to match the residual dimension, and normalize the output vector of each time step), perform a temporal average pooling operation on the output vector, send the pooled vector into a fully connected network, and output it by the output layer; Collect historical battery data as the training set and input it into the LSTM model for iterative training. Define the mean squared error as the training objective loss function and the Adam optimizer for iterative optimization of the model parameters. If the validation error does not decrease for consecutive P rounds, stop the iteration, obtain the optimal model parameters as the initial parameter set E, and retain the minimum validation RMSE during the training process as the fitness evaluation reference; Expand E into a one-dimensional vector, use it as the initial position of the particle, and initialize the APSO population. Set the population size and initialize a velocity vector for each particle simultaneously; Start iterative optimization, set the maximum number of rounds, perform velocity and position updates for each particle, calculate the fitness value corresponding to each particle, and update the individual optimal and global optimal solutions; After completing the global search iteration of APSO, directly extract the current global optimal particle position as the final optimization solution and restore it to the LSTM model parameter structure. Load the optimized parameters into the original LSTM model to obtain the optimized model; The Input it into the optimized LSTM model to obtain a preliminary SOH prediction value; Use Attention to construct a morphological channel, set the multi-head attention parameters, including the number of attention heads and the output dimension of each head. Construct three sets of mapping matrices for each attention head, and perform Linear transformation to obtain query Q, key K, and value V tensors. Perform self-attention calculation for each attention head, and concatenate the output vectors of all heads along the dimension to form the final attention modeling output, representing the health state characterization vector modeled by the current morphological feature channel; Receive the outputs of two model channels, splice them to form a fused vector V, construct an MLP network structure including an input layer, a hidden layer, and an output layer, and input V into the MLP to obtain a preliminary SOH estimation value .

[0036] Through the feature partitioning and modeling strategy, the time dynamics and frequency-domain morphological signals are fully decoupled, improving the modeling accuracy; the EMD and multi-scale feature structures are introduced to enhance the expression ability of non-stationary signals; the structure of LSTM is optimized by combining APSO to solve the problem of weak stability of traditional deep network structures; the internal dependence between high-order statistical features is modeled through the Attention mechanism, enhancing the adaptability of SOH estimation to complex aging features. This method effectively improves the accuracy and generalization ability of SOH static estimation, providing strong modeling support for the automated and intelligent screening of retired batteries.

[0037] Furthermore, the joint extended Kalman filter is used for SOH correction, and the current internal resistance is obtained as follows Collect the rated capacity of the battery at the factory , and calculate the initial capacity estimate based on the SOH estimate and the rated capacity :[[]]END]]

[0038] Construct a joint state variable, define and initialize the filtering state vector, where the state vector includes the current state of charge, the RC voltage term (electrochemical polarization voltage), the equivalent internal resistance, and the current capacity of the battery (using the capacity estimate as the initial value of the current capacity of the battery); At the same time, initialize the covariance matrix, state noise, and observation noise based on historical data; Extract the dynamic observation data sequence from the standard charging test process, including current, terminal voltage, and time interval, align and remove anomalies from the data, and use the battery RC equivalent model to predict the state variables, including predicting the SOC and RC polarization voltage:

[0039]

[0040] where is the state of charge at the k-th moment, is the capacity estimate at the k-th moment, is the current, is the sampling time interval, is the polarization resistance, is the polarization capacitance, is the state of charge at the next moment, is the RC polarization voltage at the next moment, is the RC polarization voltage; Among them, the capacity and internal resistance remain unchanged in state prediction (regarded as constants during the prediction phase); Calculate the current predicted voltage based on the state prediction value :

[0041] In the formula, is the open-circuit voltage value, obtained by experimental fitting or by looking up the OCV–SOC table, is the ohmic internal resistance; Calculate the error between the actual voltage and the predicted voltage :

[0042] In the formula, is the actual measured voltage value; Update the state vector and state covariance using the Kalman gain formula of the standard EKF:

[0043]

[0044]

[0045] In the formula, is the Kalman gain, O is the observation noise covariance matrix, is the Jacobian matrix of the observation function, is the updated state vector, is the predicted value of the state covariance, is the transpose matrix, is the predicted state at the current moment, is the updated state covariance matrix, is the identity matrix, used to keep the dimensions of the matrix unchanged; When all sampling points are iterated, the final state vector estimate is obtained, including the current state of charge, the estimate of the RC polarization voltage component, the current equivalent internal resistance estimate, and the current capacity estimate; Extract the capacity estimate from the state vector estimate, and calculate the final health estimate using the estimated capacity and the rated capacity :

[0046] In the formula, is the final estimated capacity.

[0047] In terms of state modeling, the construction of the joint state vector not only includes the real-time state of charge (SOC) and RC voltage term of the battery, but also introduces two core parameters, the current equivalent internal resistance and capacity, which upgrades the SOH estimation process from a single "voltage-capacity mapping" problem to a multi-parameter collaborative evolution modeling problem. This state-parameter joint filtering structure enhances the model's response ability to internal state perturbations of the system, making the SOH estimation more adaptable and real-time. In the prediction stage, the dynamic response of the battery is accurately modeled through the RC equivalent circuit model, the SOC and polarization voltage are calculated, and the predicted voltage is calculated in combination with the OCV-SOC mapping function, effectively restoring the terminal voltage behavior under the real operating state. The observation residual is introduced through the error calculation with the measured voltage, and the system dynamically adjusts the state estimation. State correction is performed through the Kalman gain calculation, which can maintain high estimation stability and accuracy in the face of uncertainties such as voltage noise and load disturbances. This feedback-regulated estimation mechanism is crucial for enhancing the robustness of the SOH output.

[0048] S4. Estimate the maximum available power of the battery based on the internal resistance and real-time voltage of the battery , based on SOH and conduct battery health level classification and application scenario recommendation; Specifically, estimate the maximum available power of the battery based on the internal resistance and real-time voltage of the battery , based on SOH and conduct battery health level classification and application scenario recommendation, including: Extract the current equivalent internal resistance estimation from the state vector estimation , and estimate the maximum available power of the battery based on the real-time voltage U :

[0049] Integrate the obtained final health estimation and the maximum available power of the battery to generate an input feature vector. Use random forest as the classification model, define the health level labels as A level, B level, and C level, corresponding to high health, medium health, and low health respectively, and input the input feature vector into the random forest model to obtain the corresponding health level label; Automatically recommend the corresponding application scenarios according to the classification results, including A level recommended for industrial energy storage systems, grid backup power supplies, microgrid main power supply systems, data center UPS, etc., B level recommended for low-speed electric vehicles, household small energy storage systems, emergency communication guarantee power supplies, etc., and C level recommended for material recycling, regeneration disassembly, cell echelon remanufacturing, etc.

[0050] By combining the current internal resistance with the real-time voltage for power capacity modeling, it breaks through the single-dimensional limitation of traditional independent estimation based only on open-circuit voltage or internal resistance. Introducing the SOP factor in the health level classification, constructing a joint feature vector, and considering the "remaining life" and "current performance" of the battery together improve the classification granularity and application matching accuracy. When dealing with multi-dimensional numerical features (such as SOH, SOP, mean of feature subsets, skewness, etc.), the random forest can automatically evaluate the feature importance and eliminate redundant features, making the classification results more stable and interpretable. Matching the classification results with the actual application scenarios and outputting directly breaks through the information barrier between the evaluation end and the application end. For example, recommending Class A batteries for industrial and commercial energy storage or microgrid systems, matching Class B batteries with light-load transportation tools, and importing Class C batteries into the recycling process, which greatly improves the implementation and transformation efficiency of the evaluation results and helps to promote the construction of the closed-loop industrial chain for battery secondary utilization.

[0051] S5. Package all data into an evaluation report and archive it in the cloud for storage; Specifically, packaging all data into an evaluation report and archiving it in the cloud for storage means integrating the basic battery information and the data generated by analysis (health level) according to the task number and the bound data, generating a complete battery health evaluation report, outputting the report in both PDF format and JSON structured data, uploading the report to the cloud database platform, docking with the battery full-life cycle management system (BMS), and marking the corresponding battery life cycle status as "evaluated".

[0052] It realizes the organic unity of visualization of evaluation results, improvement of data availability, and intelligent linkage of the system, significantly enhancing the industrial adaptability, data reusability, and management intelligence level of the retired battery health evaluation system.

[0053] This embodiment also provides a health evaluation system for retired batteries, including: A task management module, used to identify the retired battery identifier and automatically generate an evaluation task number, and set the battery status to "to be detected"; A feature extraction module, used to set the battery charging mode, collect battery status data in real time, and extract battery features to construct a feature vector; An estimation module, used to input the feature vector into a parallel attention neural network for static SOH estimation to obtain a preliminary SOH estimated value; A correction module, used to perform SOH correction using the extended Kalman filter and obtain the current internal resistance; A level classification module, used to estimate the maximum available power of the battery based on the battery internal resistance and real-time voltage, and perform health level classification and application scenario recommendation based on SOH; An archiving module, used to package all data into an evaluation report and archive it in the cloud for storage.

[0054] This embodiment also provides a computer device, which is applicable to the situation of the health assessment method for retired batteries, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the health assessment method for retired batteries as proposed in the above embodiment.

[0055] The computer device can be a terminal. The computer device includes a processor, a memory, a communication interface, a display screen, and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be implemented through WIFI, a carrier network, NFC (Near Field Communication), or other technologies. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer covered on the display screen, or a button, a trackball, or a touchpad set on the shell of the computer device, or an external keyboard, a touchpad, or a mouse, etc.

[0056] This embodiment also provides a storage medium, on which a computer program is stored. When the program is executed by a processor, it implements the health assessment method for retired batteries as proposed in the above embodiment; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM for short), Electrically Erasable Programmable Read-Only Memory (EEPROM for short), Erasable Programmable Read-Only Memory (EPROM for short), Programmable Read-Only Memory (PROM for short), Read-Only Memory (ROM for short), magnetic memory, flash memory, a magnetic disk, or an optical disc.

[0057] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered by the scope of the claims of the present invention.

Claims

1. A health assessment method for retired batteries, characterized in that: including identifying the retired battery identification and automatically generating an evaluation task number, and setting the battery status to "to be detected"; setting the battery charging mode, collecting battery status data in real time, and extracting battery features to construct a feature vector G; the battery status data includes voltage, current, and time data; inputting the feature vector G into a parallel attention neural network for static SOH estimation to obtain a preliminary SOH estimate value, using an extended Kalman filter for SOH correction, and obtaining the current internal resistance; Estimating the maximum available power of a battery based on the internal resistance and real-time voltage , based on SOH and conducting battery health level classification and application scenario recommendation; encapsulating all data into an evaluation report and archiving and storing it in the cloud.

2. The health assessment method for retired batteries according to claim 1, wherein: The setting of the battery charging mode, collecting battery status data in real time, and extracting battery features to construct a feature vector G means collecting the voltage, current, and time data of the battery and recording them as the original data set; After preprocessing the data in the original dataset, the primary features of the data are extracted, including the charging duration in the constant current stage , the charging duration in the constant voltage stage , the rising rate of the middle voltage segment , the main peak of the incremental capacity , the corresponding voltage position and the peak time point of the differential voltage curve ; Integrate the extracted features to form a first-level feature set ; The voltage and current in the dataset are decomposed by the Daubechies 4 wavelet basis function for three layers to obtain six groups of sub-signals. For each sub-signal, features are extracted and integrated to generate a wavelet feature set ; Merge the feature sets and to construct the total feature set F; Fully discharge each battery to obtain the true capacity label a. Use the Pearson correlation evaluation method to calculate the correlation between each dimension feature and the true capacity label. Retain the features with a correlation greater than the preset threshold and integrate them into the filtered feature set F'. Construct a binary selection variable for each feature in F', and construct a binary vector of length k , where: indicates whether the k-th feature in the feature set is selected, and each B corresponds to a feature subset combination; using the sparrow search algorithm to search for the optimal feature subset; Using the XGBoost regressor based on the feature data to obtain the predicted capacity, calculating the average error between the predicted capacity and the true capacity as the fitness value, and setting the individual with the lowest fitness as the global optimal solution ; executing an exponential decay perturbation update strategy for the discoverers; globally guiding and adjusting the remaining individuals with reference to the current global optimal solution; marking the individual with the worst current fitness as the vigilant, performing mutation perturbation, and jumping out of the local optimum; probability mapping all individuals' continuous coding vectors through the Sigmoid function, and then performing threshold binaryzation to obtain updated binary coding individuals. After reaching the maximum number of iterations W, output the final global optimal coding solution A. Extract the selected feature dimensions according to the optimal coding A, form the optimal feature subset, perform normalization processing, and splice them into a numerical vector G.

3. The health assessment method for retired batteries according to claim 2, wherein: The inputting the feature vector G into a parallel attention neural network for static SOH estimation to obtain a preliminary SOH estimate value includes: dividing the feature vector G into two sub-vectors according to the feature source for processing by different model channels; Extract the main variables reflecting the evolution of the battery charge characteristics from it, denoted as the signal sequence x(t), and use the empirical mode decomposition algorithm to decompose the time series signal into a number of intrinsic mode functions and residuals; Select the first two order IMF components and are used to reconstruct the main sequence ; Combine the features of all time steps into an input tensor , where t is the time step and d is the number of input feature dimensions at each time step; Construct a temporal channel using the LSTM model, construct the first layer LSTM-1, initialize the input dimension, the number of hidden units, and the activation function, and input the Use a linear layer to increase the dimension to the same dimension to match the residual dimension, and normalize the output vector for each time step; sending the normalized output to the second-layer LSTM-2 network, repeating the residual mechanism, performing a time average pooling operation on the output vector, sending the pooled vector to a fully connected network, and outputting by the output layer; collecting historical battery data as a training set and inputting it into the LSTM model for iterative training, defining the mean square error as the training objective loss function and the Adam optimizer for iterative optimization of model parameters. If the validation error does not decrease for consecutive P rounds, stop the iteration, and obtain the optimal model parameters as the initial parameter set E; expanding E into a one-dimensional vector, using it as the initial position of the particle and initializing the APSO population, setting the population size, and initializing a velocity vector for each particle at the same time; starting iterative optimization, setting the maximum number of rounds, performing velocity and position updates for each particle, calculating the fitness value corresponding to each particle, and updating the individual optimal and global optimal solutions; after completing the global search iteration of APSO, directly extract the current global optimal particle position as the final optimization solution and restore it to the LSTM model parameter structure, and load the optimized parameters into the original LSTM model to obtain the optimized model; Input into the optimized LSTM model to obtain a preliminary SOH prediction value; Construct a morphological channel using Attention, set the parameters of multi-head attention, including the number of attention heads and the output dimension of each head, and construct three sets of mapping matrices for each attention head respectively. For perform linear transformation to obtain query Q, key K, and value V tensors, perform self-attention calculation for each attention head, splice the output vectors of all heads along the dimension to form the final attention modeling output, which represents the health state representation vector modeled by the current morphological feature channel; Receive the outputs of two model channels and splice them to form a fused vector V. Construct an MLP network structure, including an input layer, a hidden layer, and an output layer, and input V into the MLP to obtain a preliminary SOH estimate value .

4. The health assessment method for retired batteries according to claim 3, characterized in that: The using an extended Kalman filter for SOH correction and obtaining the current internal resistance includes: Collect the rated capacity of the battery at the factory , calculate the initial estimated capacity based on the SOH estimated value and the rated capacity; constructing a joint state variable, defining a filter state vector and initializing it. The state vector includes the current state of charge, RC voltage term, equivalent internal resistance, and the current capacity of the battery; Initialize the covariance matrix, state noise, and observation noise based on historical data simultaneously; Extract the dynamic observation data sequence from the standard charging test process, and use the battery RC equivalent model to predict the state variables, including predicting the SOC and RC polarization voltage; Calculate the current predicted voltage based on the status prediction value ; Calculate the error between the actual voltage and the predicted voltage; Update the state vector and state covariance using the Kalman gain formula of the standard EKF; When all sampling points are iterated, obtain the final state vector estimation, including the current state of charge, the estimation of the RC polarization voltage component, the current equivalent internal resistance estimation, and the current capacity estimation; Extract the capacity estimate from the state vector estimate and calculate the final health estimate using the estimated capacity and the rated capacity .

5. The health assessment method for retired batteries according to claim 4, characterized in that: Estimating the maximum available power of the battery based on the internal resistance and real-time voltage , based on SOH and Performing battery health level classification and application scenario recommendation includes: Extract the current equivalent internal resistance estimate from the state vector estimate and estimate the maximum available power of the battery based on the real-time voltage U ; Integrate the obtained final health estimation and the maximum available power of the battery to generate an input feature vector. Use a random forest as a classification model, define the health level labels as A-level, B-level, and C-level, corresponding to high health, medium health, and low health respectively. Input the input feature vector into the random forest model to obtain the corresponding health level label, and automatically recommend the corresponding application scenario according to the classification result.

6. The health assessment method for retired batteries according to claim 5, wherein: The identification of the retired battery identifier and the automatic generation of the evaluation task number, and setting the battery status to "to be detected" means that in the battery warehousing link, the RFID tag information attached to the battery is read by the RFID scanning device and used as the unique number of the battery. After obtaining the battery number, the evaluation task number corresponding to the battery is generated synchronously, and the task number is bound to the battery number one by one. After the information is bound, the initial state of the task is set to "to be detected".

7. The health assessment method for retired batteries according to claim 6, characterized in that: The encapsulation of all data into an evaluation report and cloud archiving storage means integrating the basic information of the battery and the data generated by the analysis to generate a complete battery health evaluation report, and uploading it to the cloud database platform.

8. A health assessment system for retired batteries, based on the health assessment method for retired batteries according to any one of claims 1 to 7, characterized in that: Including, A task management module for identifying the retired battery identifier and automatically generating an evaluation task number, and setting the battery status to "to be detected"; A feature extraction module for setting the battery charging mode, collecting the battery status data in real time, and extracting the battery features to construct a feature vector; An estimation module for inputting the feature vector into a parallel attention neural network for static SOH estimation to obtain a preliminary SOH estimation value; A correction module for using the extended Kalman filter to correct the SOH and obtain the current internal resistance; A level classification module for estimating the maximum available power of the battery based on the battery internal resistance and the real-time voltage, and performing health level classification and application scenario recommendation based on the SOH; An archiving module for encapsulating all data into an evaluation report and performing cloud archiving storage.

9. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the health assessment method for retired batteries according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the health assessment method for retired batteries according to any one of claims 1 to 7.

Images