A battery state of health evaluation and fault diagnosis model construction method

Abstract

The present application relates to a battery health state evaluation and fault diagnosis model construction method, belonging to the technical field of battery management and health prediction. The method comprises the following steps: collecting multi-dimensional data including voltage, current, impedance parameter, AC internal resistance, DC internal resistance, phase angle, temperature, humidity, and cycle number, simultaneously applying an AC sweep signal of 0.1Hz to 10kHz to the battery to obtain electrochemical impedance spectrum data, and performing normalization processing on the electrochemical impedance spectrum data; after correcting the collected data, a high-dimensional feature vector is constructed; a support vector machine (SVM) model is constructed for SOC estimation, a neural network (NN) model is constructed for SOH estimation and RUL prediction, and the results are consistently integrated and interactively corrected; then, a CNN model is used for fault diagnosis; finally, the model is trained, and after quantization, the model is deployed to an embedded MCU to realize real-time diagnosis reasoning. The present application can realize accurate, real-time and comprehensive evaluation and early fault diagnosis of the battery health state.

Description

Technical Field

[0001] This invention belongs to the field of battery management and health prediction technology, specifically relating to a method for constructing a battery health status assessment and fault diagnosis model. Background Technology

[0002] Existing battery health (SOH) assessment technologies, such as patents CN117452259A and CN117970153A / B, primarily rely on statistical regularities or empirical correlations of macroscopic parameters to achieve indirect assessment. The former calculates the voltage coefficient of variation (CV) of batteries in the same batch within a fixed voltage range, establishing a correspondence between CV and SOH for rapid sorting; the latter analyzes the sensitivity of multiple assessment indicators (such as voltage and internal resistance) to changes within a specific target range, comprehensively calculating the health score.

[0003] However, these existing technologies have significant drawbacks. First, they rely on indirect inference from external time-domain parameters such as voltage and current, failing to grasp the essential electrochemical state inside the battery, resulting in inherently low accuracy and poor anti-interference capabilities. Second, the mathematical statistics or simple fitting models used are insufficient to characterize the complex nonlinear process of battery aging, leading to low accuracy limits and weak generalization. Third, the output is only a single SOH value, unable to specifically diagnose internal fault types (such as lithium dendrites or electrolyte drying), making its function relatively limited. Fourth, the application conditions are demanding, heavily reliant on batch consistency, or requiring extensive prior knowledge to set standard values ​​and target ranges, resulting in poor engineering applicability. Summary of the Invention

[0004] The purpose of this invention is to improve the accuracy of battery health assessment and fault diagnosis capabilities, address the shortcomings of existing technologies, and provide a method for constructing a battery health status assessment and fault diagnosis model. Through an automated, multi-parameter fusion artificial intelligence method, it achieves accurate, real-time, and comprehensive assessment of the health status of lead-acid batteries and lithium batteries, as well as early fault diagnosis.

[0005] To achieve the above objectives, the technical solution of the present invention is: a method for constructing a battery health status assessment and fault diagnosis model, comprising the following steps:

[0006] Step S1: Multi-source data acquisition and preprocessing: Acquire multi-dimensional data including voltage, current, impedance parameters, AC internal resistance, DC internal resistance, phase angle, temperature, humidity, and cycle number. At the same time, apply an AC sweep frequency signal from 0.1Hz to 10kHz to the battery to obtain electrochemical impedance spectroscopy data, and normalize the electrochemical impedance spectroscopy data.

[0007] Step S2, Multi-source feature correction and compensation: For the data collected in step S1, establish correction models based on environment or usage conditions to perform compensation and correction.

[0008] Step S3: Construct high-dimensional feature vectors: Based on the data corrected in step S2, construct high-dimensional feature vectors;

[0009] Step S4, Health Status Prediction Model Construction: A hierarchical fusion ensemble learning framework is adopted to construct a support vector machine (SVM) model for battery state of charge (SOC) estimation and a neural network (NN) model for battery health status (SOH) estimation and battery remaining life (RUL) prediction, respectively. The SVM model and the NN model share the high-dimensional feature vector extracted from step S2, and the SOC estimation results, SOH estimation results, and RUL prediction results are consistently integrated and interactively corrected through a result coordination layer.

[0010] Step S5: Correlation between state assessment results and fault diagnosis: The state assessment results obtained in step S4, including the SOC estimation results, SOH estimation results, and RUL prediction results and their anomaly indicators, are fused with the high-dimensional feature vector constructed in step S3 to form an enhanced feature vector;

[0011] Step S6, Fault diagnosis based on CNN model: The enhanced feature vector obtained in step S5 is used to classify internal battery faults through convolutional layer operations, pooling layer operations, output layer calculation, backpropagation and optimization of the CNN model.

[0012] Step S7, Model Training, Optimization and Embedded Deployment: Train and optimize the model using historical battery data, quantize the model and deploy it to the embedded MCU to achieve real-time diagnostic inference.

[0013] Furthermore, in step S1, the formula for normalizing the electrochemical impedance spectroscopy data is:

[0014]

[0015] in, and , , represent the mean and standard deviation of the corresponding batch of electrochemical impedance spectroscopy data, respectively, and Z represents the electrochemical impedance spectroscopy data before normalization. These are the normalized electrochemical impedance spectroscopy data.

[0016] Furthermore, in step S2, the data collected in step S1 are compensated and corrected using correction models based on environmental or usage conditions to eliminate or reduce the impact of unhealthy factors on the characteristic quantities. Specifically, this includes:

[0017] (1) Voltage and current correction model:

[0018] A dynamic compensation model based on the temperature-current-voltage coupling relationship is established to correct voltage drift and current fluctuations caused by temperature changes. The formula is as follows:

[0019]

[0020]

[0021] in, Here, T represents the reference temperature, H represents the ambient temperature, and H represents the ambient humidity. This is the temperature compensation coefficient. This is the temperature and humidity coupling coefficient; , V and I represent the voltage and current for ambient temperature detection, respectively; V and I represent the corrected voltage and current, respectively.

[0022] (2) Impedance parameter correction model:

[0023] A humidity compensation model is established for the impedance parameters, and the formula is as follows:

[0024]

[0025] in, This represents the real part of the impedance measured under ambient humidity. This is a temperature and humidity coupling compensation function. This is the real part of the corrected impedance;

[0026] (3) Temperature correction model for electrochemical impedance spectroscopy:

[0027] Using an Arrhenius-type temperature correction model, electrochemical impedance spectroscopy data measured at different temperatures are uniformly corrected to equivalent values ​​at a reference temperature (e.g., 25°C). The specific formula is as follows:

[0028]

[0029] The impedance value is the actual measured value at ambient temperature T (it can be the entire electrochemical impedance spectroscopy, i.e., the EIS spectrum, or the impedance at a certain frequency point). The reference temperature obtained for calibration Impedance value at (e.g., 25℃); The activation energy is an intrinsic parameter of the material, representing the sensitivity of the electrochemical process to temperature. This value needs to be determined through preliminary experiments. is the universal gas constant, a fixed value of 8.314 J·mol⁻¹·K⁻¹.

[0030] Furthermore, in step S3, the constructed high-dimensional feature vector is represented as follows:

[0031]

[0032] Where V is voltage, I is current, T is temperature, and H is humidity. The number of loops. Due to internal communication barriers, DC internal resistance, The rate of change of voltage. The rate of change of current, (@f1) represents the real part of the impedance at the predetermined frequency f1. The imaginary part of the impedance at a predetermined frequency f1, The phase angle at the predetermined frequency f3.

[0033] Furthermore, in step S4, the hierarchical fusion ensemble learning framework includes a feature preprocessing layer, a model parallel layer, and a result collaboration layer;

[0034] Feature preprocessing layer: Performs unified preprocessing on the high-dimensional feature vector F and generates a feature subset for SOC estimation. and feature subsets for SOH estimation / RUL prediction :

[0035]

[0036]

[0037] Where V is voltage, I is current, T is temperature, and H is humidity. The number of loops. Due to internal communication barriers, DC internal resistance, The rate of change of voltage. The rate of change of current, (@f1) represents the real part of the impedance at the predetermined frequency f1. The imaginary part of the impedance at a predetermined frequency f1, The phase angle at a predetermined frequency f3;

[0038] Parallel Model Layer: Support Vector Machine (SVM) Model Reception SOC estimation is performed by a neural network (NN) model. SOH estimation and RUL prediction were performed, with the two models running in parallel.

[0039] Results Collaboration Layer: Performs a consistency check on the SOC value output by the Support Vector Machine (SVM) model and the SOH value output by the Neural Network (NN) model. If the SOC value exceeds the SOH value, the SOC value is limited to the SOH value to ensure that the evaluation results are reasonable.

[0040] Furthermore, in step S4, the neural network NN model introduces a two-stage collaborative training strategy during the training phase:

[0041] Phase 1: Independent training, using individual methods and Train the Support Vector Machine (SVM) model and the Neural Network (NN) model until convergence;

[0042] The second stage: Collaborative fine-tuning. The parameters of the Support Vector Machine (SVM) model are fixed, and the SOC estimate output by the SVM model is added as a new feature. After being spliced ​​together, they are used as input to the neural network (NN) model to fine-tune the NN model so that its SOH estimation and RUL prediction can sense and adapt to the dynamic changes of SOC.

[0043] Furthermore, in step S4, to improve the accuracy of SOC estimation, balance model complexity and training error, and reduce the model's oversensitivity to training data, a Support Vector Machine (SVM) model and a Support Vector Regression (SVR) model are adopted, introducing... - Insensitive loss function optimization, the objective function is:

[0044]

[0045] The constraints are:

[0046]

[0047] Furthermore, a radial basis function (RBF) kernel is used, with the following expression:

[0048]

[0049] In the above formula, w is the weight vector; b is the bias term; and ||w||² is the margin maximization term. Here, is the error penalty term, representing the sum of the total costs for all samples violating the constraints; This is a slack variable, representing the degree of deviation when the deviation between the predicted and actual values ​​of the sample exceeds a preset insensitive interval; Let be the slack variable for the i-th sample; Let be the dual relaxation variable for the i-th sample; m be the total number of samples; and C be the penalty factor. This represents the i-th input sample; For reference sample; Represents the i-th input sample The corresponding target value; (·) is a high-dimensional space function; γ is the kernel function parameter; exp() is the natural exponential function, which converts the distance into a similarity value in the interval [0,1].

[0050] Furthermore, in step S4, to match the physical meaning of SOH and RUL, sharing low-level features allows the SOH and RUL tasks to mutually promote each other, utilizing common information to improve overall prediction accuracy. Specifically, to ensure that SOH remains within the range of 0 to 1, a sigmoid activation function is used to limit the final estimated value of SOH; while a linear function allows RUL to predict arbitrary values, adapting to the infinite range of remaining lifetime. Simultaneously, to avoid imbalanced results caused by data loss during training, a loss function is used to balance the final results. Specifically, the neural network (NN) model adopts a structure with a shared low-level feature extraction layer and two independent output layers in the upper branches; the input layer receives high-dimensional feature vectors; the shared hidden layer consists of 2-3 fully connected layers, with the activation function being... Activation function; Output layer 1 is used for SOH estimation, using the Sigmoid activation function, with an output range of (0,1), corresponding to 0%~100% SOH; Output layer 2 is used for RUL prediction, using a linear activation function, outputting the remaining battery cycle life or number of days;

[0051] The loss function used is a combined loss function:

[0052]

[0053] in Mean square error, The mean absolute error is given, and α and β are hyperparameters. It is expressed as follows:

[0054]

[0055] It is expressed as follows:

[0056]

[0057] In the above formula, m is the total number of samples; This represents the true value of the battery health status corresponding to the i-th sample; Let be the predicted value of the battery health status corresponding to the i-th sample; This represents the true value of the remaining useful life corresponding to the i-th sample. Let be the predicted remaining useful life for the i-th sample.

[0058] Furthermore, in step S5, the state evaluation results obtained in step S4, including the SOC estimation results, SOH estimation results, and RUL prediction results and their anomaly indicators, are fused with the high-dimensional feature vector constructed in step S3 to form an enhanced feature vector, as shown in the following formula:

[0059] Let the high-dimensional feature vector constructed in step S3 be... , dimension And let the SOC value, RUL value, SOH value, and anomaly flag vector obtained in step S4 be respectively... , , and , It is A binary vector of dimensions (e.g., [SOC anomaly, SOH anomaly, RUL anomaly, voltage anomaly...]).

[0060] The enhanced feature vector E is formed through a concatenation operation:

[0061]

[0062] Where [* ; *] denotes vector concatenation, and the dimension of the enhanced feature vector E is . +3+ .

[0063] Further, step S6 is implemented as follows:

[0064] The enhanced feature vector E is shaped into a one-dimensional signal and input into the CNN model, and the following operations are performed:

[0065] (1) Convolutional layer operations

[0066] The output of the k-th feature map of the l-th convolutional layer is:

[0067]

[0068] Where * represents a one-dimensional convolution operation; It is the input of the m-th feature map of the (l-1)-th layer; These are the convolution kernel weights that connect the m-th feature map to the k-th feature map; It is the bias term of the k-th feature map; () is an activation function (such as ReLU, Leaky);

[0069] (2) Pooling layer operations:

[0070]

[0071] Pooling refers to either max pooling or average pooling, used for dimensionality reduction and enhancing feature invariance.

[0072] (3) Output layer calculation:

[0073] The Softmax output layer calculates a probability vector based on the output of the pooling layer operations:

[0074]

[0075] Where n represents the total number of fault categories; These represent the probabilities that the model considers the input data to belong to fault type 1, 2, ..., n, respectively; the sum of all probabilities is 1, i.e. + +...+ =1;

[0076] (4) Fault classification decision:

[0077] pass The function finds the category with the highest probability from the probability vector y and uses it as the final fault diagnosis result:

[0078]

[0079] The function returns the index number of the category corresponding to the highest probability value.

[0080] Furthermore, the method of the present invention is applicable to lead-acid batteries and lithium batteries, and can realize the estimation of battery state of charge (SOC), the estimation of battery state of health (SOH), the assessment of battery remaining service life (RUL), and the intelligent diagnosis of internal battery faults.

[0081] The present invention also provides an electronic device including a processor and a memory, wherein the memory stores a computer program that, when executed by the processor, causes the processor to perform the steps of the method described above.

[0082] The present invention also provides a computer-readable storage medium having stored thereon computer program instructions that can be executed by a processor, wherein when the processor executes the computer program instructions, it can implement the steps of the method described above.

[0083] Compared with the prior art, the present invention has the following beneficial effects: The present invention achieves accurate, real-time and comprehensive assessment of the health status of lead-acid batteries and lithium batteries and early fault diagnosis through an automated, multi-parameter fusion artificial intelligence method. Attached Figure Description

[0084] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0085] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings.

[0086] This invention provides a method for constructing a battery health status assessment and fault diagnosis model, comprising the following steps:

[0087] Step S1: Multi-source data acquisition and preprocessing: Acquire multi-dimensional data including voltage, current, impedance parameters, AC internal resistance, DC internal resistance, phase angle, temperature, humidity, and cycle number. At the same time, apply an AC sweep frequency signal from 0.1Hz to 10kHz to the battery to obtain electrochemical impedance spectroscopy data, and normalize the electrochemical impedance spectroscopy data.

[0088] Step S2, Multi-source feature correction and compensation: For the data collected in step S1, establish correction models based on environment or usage conditions to perform compensation and correction.

[0089] Step S3: Construct high-dimensional feature vectors: Based on the data corrected in step S2, construct high-dimensional feature vectors;

[0090] Step S4, Health Status Prediction Model Construction: A hierarchical fusion ensemble learning framework is adopted to construct a support vector machine (SVM) model for battery state of charge (SOC) estimation and a neural network (NN) model for battery health status (SOH) estimation and battery remaining life (RUL) prediction, respectively. The SVM model and the NN model share the high-dimensional feature vector extracted from step S2, and the SOC estimation results, SOH estimation results, and RUL prediction results are consistently integrated and interactively corrected through a result coordination layer.

[0091] Step S5: Correlation between state assessment results and fault diagnosis: The state assessment results obtained in step S4, including the SOC estimation results, SOH estimation results, and RUL prediction results and their anomaly indicators, are fused with the high-dimensional feature vector constructed in step S3 to form an enhanced feature vector;

[0092] Step S6, Fault diagnosis based on CNN model: The enhanced feature vector obtained in step S5 is used to classify internal battery faults through convolutional layer operations, pooling layer operations, output layer calculation, backpropagation and optimization of the CNN model.

[0093] Step S7, Model Training, Optimization and Embedded Deployment: Train and optimize the model using historical battery data, quantize the model and deploy it to the embedded MCU to achieve real-time diagnostic inference.

[0094] The following is a detailed implementation process of the present invention.

[0095] like Figure 1 As shown in the example, this method for constructing a battery health status assessment and fault diagnosis model includes the following steps:

[0096] Step S1: Multi-source data acquisition and preprocessing

[0097] Data was collected across multiple dimensions, including voltage, current, temperature, humidity, and historical cycle count. Simultaneously, electrochemical impedance spectroscopy (EIS) data was acquired by applying an AC sweep signal from 0.1 Hz to 10 kHz to the battery. The raw EIS data were then normalized using the following formula:

[0098]

[0099] in, and These are the mean and standard deviation of the impedance data for this batch, respectively, to eliminate the influence of dimensions.

[0100] Step S2: Multi-source feature correction and compensation

[0101] For the voltage, current, temperature, humidity, and electrochemical impedance spectroscopy data collected in step S1, correction models based on environmental or usage conditions are established to compensate for and correct the characteristic quantities. Specifically, this includes:

[0102] (1) Voltage and current correction model:

[0103] A dynamic compensation model based on the temperature-current-voltage coupling relationship is established to correct voltage drift and current fluctuations caused by temperature changes. The formula is as follows:

[0104]

[0105]

[0106] in, Here, T represents the reference temperature, H represents the ambient temperature, and H represents the ambient humidity. This is the temperature compensation coefficient. This is the temperature and humidity coupling coefficient; , V and I represent the voltage and current for ambient temperature detection, respectively; V and I represent the corrected voltage and current, respectively.

[0107] (2) Impedance parameter correction model:

[0108] A humidity compensation model is established for the impedance parameters, and the formula is as follows:

[0109]

[0110] in, This represents the real part of the impedance measured under ambient humidity. This is a temperature and humidity coupling compensation function. This is the real part of the corrected impedance;

[0111] (3) Temperature correction model for electrochemical impedance spectroscopy:

[0112] Using an Arrhenius-type temperature correction model, electrochemical impedance spectroscopy data measured at different temperatures are uniformly corrected to equivalent values ​​at a reference temperature (e.g., 25°C). The specific formula is as follows:

[0113]

[0114] The impedance value is the actual measured value at ambient temperature T (it can be the entire electrochemical impedance spectroscopy, i.e., the EIS spectrum, or the impedance at a certain frequency point). The reference temperature obtained for calibration Impedance value at (e.g., 25℃); The activation energy is an intrinsic parameter of the material, representing the sensitivity of the electrochemical process to temperature. This value needs to be determined through preliminary experiments. is the universal gas constant, a fixed value of 8.314 J·mol⁻¹·K⁻¹.

[0115] Step S3: Construct high-dimensional feature vectors

[0116] Time-domain features such as capacity decay rate, voltage drop slope, and internal resistance change rate are extracted from the discharge curve; Nyquist curve feature points such as the real part minimum, imaginary part maximum, phase angle at characteristic frequency points, and relaxation time are extracted from the EIS spectrum to construct a high-dimensional feature vector.

[0117]

[0118] The meanings of each parameter are shown in Table 1:

[0119] Table 1

[0120]

[0121] Step S4: Construction of Health Status Prediction Model

[0122] A hierarchical fusion ensemble learning framework is adopted to construct dedicated prediction models for different characteristics of battery health state (SOH), state of charge (SOC), and remaining service life (RUL).

[0123] (1) Support Vector Machine (SVM) model for SOC estimation

[0124] Model selection:

[0125] Given the strong nonlinear relationship between SOC and parameters such as battery voltage, current, and temperature, although these parameters change smoothly in the short term, a Support Vector Regression (SVR) model is adopted, incorporating... - Optimize using an insensitive loss function.

[0126] Optimize the objective function:

[0127]

[0128] The constraints are:

[0129]

[0130] Furthermore, a radial basis function (RBF) kernel is used, with the following expression:

[0131]

[0132] In the above formula, w is the weight vector; b is the bias term; and ||w||² is the margin maximization term. Here, is the error penalty term, representing the sum of the total costs for all samples violating the constraints; This is a slack variable, representing the degree of deviation when the deviation between the predicted and actual values ​​of the sample exceeds a preset insensitive interval; Let be the slack variable for the i-th sample; Let be the dual relaxation variable for the i-th sample; m be the total number of samples; and C be the penalty factor. This represents the feature vector of the i-th input sample (i.e., the i-th data point). For reference sample; Represents the i-th input sample The corresponding target value (label value); (·) is a high-dimensional space function; γ is the kernel function parameter; exp() is the natural exponential function, which converts the distance into a similarity value in the interval [0,1].

[0133] The meanings of each symbol are shown in Table 2:

[0134] Table 2

[0135]

[0136] (2) Neural Network (NN) Models for SOH Estimation and RUL Prediction

[0137] SOH and RUL are relevant long-term degradation metrics, so a multi-output neural network is designed for joint learning.

[0138] ① Network structure: It adopts a structure with a shared bottom feature extraction layer and two independent output layers in the upper layer.

[0139] Input layer: receives feature vectors:

[0140]

[0141] Shared hidden layers (2-3 layers): Use fully connected layers with the following activation function:

[0142]

[0143] Output layer 1 (SOH): Using the Sigmoid activation function:

[0144]

[0145] The output range is (0,1), corresponding to 0%~100% SOH.

[0146] Output layer 2 (RUL): using a linear activation function: Predict the remaining cycle period or number of days.

[0147] ② Loss function: A combined loss function is used to optimize the prediction accuracy of both SOH and RUL.

[0148]

[0149] Mean Squared Error (MSE):

[0150]

[0151] Mean Absolute Error (MAE):

[0152]

[0153] In the above formula, α and β are hyperparameters used to balance the weights of the two tasks. m is the total number of samples involved in the calculation; This represents the true value of the battery health status corresponding to the i-th sample; Let be the predicted value of the battery health status corresponding to the i-th sample; This represents the true value of the remaining useful life corresponding to the i-th sample. Let be the predicted remaining useful life for the i-th sample.

[0154] (3) Deep fusion mechanism of hierarchical fusion integrated learning framework

[0155] One of the core innovations of this invention is that it does not simply use SVM and NN models side by side, but designs a hierarchical fusion mechanism that enables them to work together deeply at the feature level and the result level, thereby overcoming the limitations of a single model and improving the accuracy and robustness of the overall evaluation system.

[0156] ① Overview of Integration Motivation and Architecture

[0157] Although a battery's State of Charge (SOC) and State of Health (SOH) differ on a time scale, they are physically closely related. A severely degraded battery (low SOH) will have a reduced full-charge capacity, and its accurate range of SOC will inevitably be limited. Traditional independent models cannot utilize this physical constraint. This framework introduces an "outcome coordination layer" to encode this physical common sense into the system, ensuring that the output results are logically consistent.

[0158] The entire fusion architecture is divided into three layers: feature preprocessing and allocation layer, model parallel execution layer, and result collaborative output layer.

[0159] ② Feature preprocessing and allocation layer

[0160] This layer is responsible for intelligently allocating the high-dimensional feature vector F generated in step S3. Based on the task characteristics of SOC and SOH / RUL, it divides it into two optimal feature subsets:

[0161] SOC feature subset It mainly includes voltage (V), current (I), temperature (T) and their rate of change (V / I). These parameters can quickly respond to the instantaneous state of the battery and are strongly correlated with the state of charge (SOC).

[0162] SOH / RUL Feature Subset In addition to including some basic parameters, it focuses more on reflecting the characteristics of long-term aging, such as the number of cycles ( AC / DC internal resistance ( ) and electrochemical impedance spectroscopy characteristics ( , These characteristics directly reflect the battery's capacity decay and internal resistance increase.

[0163] By assigning features, each model can focus on the most relevant information, thereby improving the prediction efficiency and accuracy in its respective field.

[0164] ③ Parallel execution and collaborative training of models

[0165] The SVM and NN models run in parallel during inference without blocking each other. However, during the training phase, this invention employs an innovative two-stage collaborative training strategy to strengthen the connections between the models:

[0166] Phase 1 - Independent Pre-training:

[0167] SVM model uses Train with real SOC labels to optimize its regression performance.

[0168] NN model use The system is trained using real SOH and RUL labels to optimize its combined loss function. .

[0169] Phase Two - Collaborative Fine-tuning:

[0170] Fix the parameters of the pre-trained SVM model. During the training of the NN model, use the SOC estimate of the SVM model for each training sample. As a new feature, compared with the original feature subset The features are concatenated to form a new input feature vector. =[ , ].

[0171] use The neural network (NN) model is fine-tuned. This stage enables the NN model to "sense" the current state of battery (SOC) when evaluating SOH and RUL. For example, the model can learn that certain aging characteristics of the battery may be more pronounced in the low SOC range. This complementary information greatly enhances the context-awareness of SOH and RUL predictions.

[0172] ④ Result Collaborative Output Layer

[0173] This is the final step in achieving model fusion and is crucial for ensuring the reasonableness of the results. After obtaining the SVM model output... and the output of the NN model Then, the system executes the following consistency verification logic:

[0174]

[0175] This logic is based on a solid physical principle: a battery's current state of charge (SOC) cannot exceed the capacity limit defined by its maximum state of health (SOH). For example, if a battery's SOH has decayed to 80% (i.e., its maximum capacity is only 80% of its rated capacity), then its SOC can only reach a maximum of 80%. This correction mechanism automatically corrects any physical paradoxes that the model may produce, significantly improving the reliability and engineering practicality of the system output.

[0176] Step S5: Correlation between condition assessment results and fault diagnosis

[0177] The core of this step is feature fusion, which combines different types of information into an enhanced feature vector.

[0178] set up:

[0179] The high-dimensional feature vector obtained from step S3 has a dimension of .

[0180] : SOC (Remaining Capacity) Predicted Value.

[0181] : SOH (Health Status) Predicted Value.

[0182] : RUL (Remaining Useful Life) Predicted Value.

[0183] An anomaly flag vector is a... A binary vector of dimensions (e.g., [SOC anomaly, SOH anomaly, RUL anomaly, voltage anomaly...]).

[0184] The enhanced feature vector E is formed through a concatenation operation:

[0185]

[0186] Where [* ; *] denotes vector concatenation (* represents a vector). Therefore, the dimension of the enhanced feature vector E is . +3+ .

[0187] This fusion process combines the original physical characteristics with the evaluation conclusions and diagnostic signals of the upper-level model, providing more comprehensive input for subsequent fault classification.

[0188] Step S6: Fault diagnosis based on CNN model

[0189] The core of this step is to use convolutional neural networks to extract and classify features from the enhanced feature vectors.

[0190] Suppose the enhanced feature vector E is reshaped into a one-dimensional "signal" and input into the CNN model, performing the following operation:

[0191] (1) Convolutional layer operations

[0192] The output of the k-th feature map of the l-th convolutional layer is:

[0193]

[0194] Where * represents a one-dimensional convolution operation; It is the input of the m-th feature map of the (l-1)-th layer; These are the convolution kernel weights that connect the m-th feature map to the k-th feature map; It is the bias term of the k-th feature map; () is the activation function;

[0195] (2) Pooling layer operations:

[0196]

[0197] Pooling refers to either max pooling or average pooling.

[0198] (3) Output layer calculation:

[0199] The Softmax output layer calculates a probability vector based on the output of the pooling layer operations:

[0200]

[0201] Where n represents the total number of fault categories; These represent the probabilities that the model considers the input data to belong to fault type 1, 2, ..., n, respectively; the sum of all probabilities is 1, i.e. + +...+ =1;

[0202] (4) Fault classification decision:

[0203] pass The function finds the category with the highest probability from the probability vector y and uses it as the final fault diagnosis result:

[0204]

[0205] The function returns the index number of the category corresponding to the highest probability value.

[0206] (5) Specific examples

[0207] Suppose that this invention defines four fault states and assigns them numbers:

[0208] Category 1 ( ):"normal"

[0209] Category 2 ( ): "Internal short circuit"

[0210] Category 3 ( ):: "Lithium plating"

[0211] Category 4 ( ): "Loose connection"

[0212] Scene:

[0213] The model calculates a probability vector from a set of input data.

[0214] y = [0.02, 0.85, 0.10, 0.03]

[0215] =0.02 → The probability of this being "normal" is 2%.

[0216] =0.85 → The probability of it being an "internal short circuit" is 85%.

[0217] =0.10 → The probability of it belonging to "lithium plating" is 10%.

[0218] =0.03 → The probability of "loose connection" is 3%.

[0219] Applying decision rules:

[0220] arg max([0.02, 0.85, 0.10, 0.03])=2

[0221] Because the second probability value (0.85) is the largest, its corresponding index is 2. Therefore, the final fault diagnosis result of the model is: Category 2, i.e., "internal short circuit".

[0222] Step 7: Model Training, Optimization, and Embedded Deployment

[0223] This step completes the entire training and optimization process of the model, and adapts it to embedded hardware through quantization and compression to achieve real-time on-site diagnostics. Specifically, it includes three parts: training optimization, quantization deployment, and performance verification.

[0224] (1) Model training and optimization

[0225] ① Training data preparation

[0226] Using over 1000 sets of battery data covering different aging levels (SOH 30%-100%) and different fault types (lithium dendrites, electrolyte drying, etc.), the data is divided into a training set (for model parameter learning), a validation set (for hyperparameter tuning), and a test set (for final performance evaluation) in a 7:2:1 ratio. The data includes multi-dimensional time-domain data and EIS frequency-domain data collected in step 1, as well as corresponding SOC / SOH / RUL labels and fault type labels.

[0227] ② Overfitting suppression strategy

[0228] Dropout regularization: Add dropout layers to the shared hidden layers of the NN model and the fully connected layers of the CNN model, with a dropout rate of 0.2, to randomly disable some neurons and prevent the model from over-relying on local features.

[0229] L2 regularization: Adds a weight decay term to the model optimization objective to suppress excessively large weight parameters and reduce model complexity.

[0230] Data augmentation: Slightly rotate (±5°) and add noise (Gaussian noise, standard deviation 0.01) to the EIS map samples to increase the diversity of training data and improve the model's generalization ability.

[0231] ③ Training process monitoring

[0232] The TensorBoard monitors the loss values ​​(combined loss of the NN model, cross-entropy loss of the CNN model) and evaluation metrics (SOC estimated MAE < 2%, SOH estimated MAE < 1.5%, fault classification accuracy > 92%) of the training and validation sets in real time. Training is stopped when the validation set loss does not decrease for 5 consecutive epochs to avoid overfitting.

[0233] (2) Model Quantization and Embedded Deployment

[0234] ① Model compression and quantization

[0235] Model pruning: The KRP (Kernel Row Pruning) pruning method is used to prune the convolutional kernels of the CNN model row by row, retaining the top 80% of the rows with the highest contribution (measured by the absolute value of the gradient), removing redundant parameters, and reducing the number of model parameters by more than 60%.

[0236] Precision quantization: Using the STM32Cube.AI toolchain, the trained floating-point (FP32) model is quantized into an 8-bit integer (INT8) model. The quantization process uses the min-max quantization algorithm to compress the model size to 1 / 4 of the original size (e.g., from 20MB to below 5MB) while ensuring that the precision loss is less than 3%.

[0237] ② Embedded hardware adaptation

[0238] Target hardware selection: Select STM32F401 series MCU (84MHz, 96KB RAM, 512KB Flash) to meet the requirements of model storage (INT8 model <5MB) and real-time computing (inference time <100ms);

[0239] Code generation and deployment: Install the X-CUBE-AI extension package in STM32CubeMX, import the quantized model file (.h5 format), and automatically generate C code for model inference; integrate the code with the driver code of the terminal data acquisition module (voltage / current sensor, EIS measurement module), and download it to the Flash of the STM32F401 MCU.

[0240] ③ Real-time inference process

[0241] Data acquisition: The MCU acquires voltage, current, and temperature data through the ADC interface (sampling rate 1kHz), and controls the EIS module through the SPI interface to acquire impedance data from 0.1Hz to 10kHz (acquiring once every 10s).

[0242] Data preprocessing: In the MCU, perform EIS normalization calculation in step 1 (based on the pre-stored batch mean μz and standard deviation σz) and feature extraction in step 2 (simplified time-domain / frequency-domain feature calculation, such as voltage drop slope and impedance real part mean).

[0243] Model inference: Call the generated inference function, input the preprocessed feature vector, and output the predicted values ​​of SOC / SOH / RUL and the probability of fault type;

[0244] Output results: The diagnostic results (such as "SOH: 85%, RUL: 500 cycles, fault: none") are uploaded to the host computer or display terminal via the UART interface to complete real-time diagnosis.

[0245] The present invention also provides an electronic device including a processor and a memory, wherein the memory stores a computer program that, when executed by the processor, causes the processor to perform the steps of the method described above.

[0246] The present invention also provides a computer-readable storage medium having stored thereon computer program instructions that can be executed by a processor, wherein when the processor executes the computer program instructions, it can implement the steps of the method described above.

[0247] The above are preferred embodiments of the present invention. Any changes made to the technical solution of the present invention that do not exceed the scope of the technical solution of the present invention shall fall within the protection scope of the present invention.

Claims

1. A method for constructing a battery health status assessment and fault diagnosis model, characterized in that, Includes the following steps: Step S1: Multi-source data acquisition and preprocessing: Acquire multi-dimensional data including voltage, current, impedance parameters, AC internal resistance, DC internal resistance, phase angle, temperature, humidity, and cycle number. At the same time, apply an AC sweep frequency signal from 0.1Hz to 10kHz to the battery to obtain electrochemical impedance spectroscopy data, and normalize the electrochemical impedance spectroscopy data. Step S2, Multi-source feature correction and compensation: For the data collected in step S1, establish correction models based on environment or usage conditions to perform compensation and correction. Step S3: Construct high-dimensional feature vectors: Based on the data corrected in step S2, construct high-dimensional feature vectors; Step S4, Construction of Health Status Prediction Model: A hierarchical fusion ensemble learning framework is adopted to construct a support vector machine (SVM) model for estimating the state of charge (SOC) and a neural network (NN) model for estimating the state of health (SOH) and predicting the remaining battery life (RUL). The SVM model and the NN model share the high-dimensional feature vector extracted from step S2, and the SOC estimation results, SOH estimation results, and RUL prediction results are consistently integrated and interactively corrected through a result coordination layer. Step S5: Correlation between state assessment results and fault diagnosis: The state assessment results obtained in step S4, including the SOC estimation results, SOH estimation results, and RUL prediction results and their anomaly indicators, are fused with the high-dimensional feature vector constructed in step S3 to form an enhanced feature vector; Step S6, Fault diagnosis based on CNN model: The enhanced feature vector obtained in step S5 is used to classify internal battery faults through convolutional layer operations, pooling layer operations, output layer calculation, backpropagation and optimization of the CNN model. Step S7, Model Training, Optimization and Embedded Deployment: Train and optimize the model using historical battery data, quantize the model and deploy it to the embedded MCU to achieve real-time diagnostic inference.

2. The method for constructing a battery health status assessment and fault diagnosis model according to claim 1, characterized in that, In step S1, the formula for normalizing the electrochemical impedance spectroscopy data is: in, and , , represent the mean and standard deviation of the corresponding batch of electrochemical impedance spectroscopy data, respectively, and Z represents the electrochemical impedance spectroscopy data before normalization. These are the normalized electrochemical impedance spectroscopy data.

3. The method for constructing a battery health status assessment and fault diagnosis model according to claim 1, characterized in that, In step S2, compensation and correction are performed on the data collected in step S1 by establishing correction models based on environmental or usage conditions, specifically including: (1) Voltage and current correction model: A dynamic compensation model based on the temperature-current-voltage coupling relationship is established to correct voltage drift and current fluctuations caused by temperature changes. The formula is as follows: in, Here, T represents the reference temperature, H represents the ambient temperature, and H represents the ambient humidity. This is the temperature compensation coefficient. This is the temperature and humidity coupling coefficient; , V and I represent the voltage and current for ambient temperature detection, respectively; V and I represent the corrected voltage and current, respectively. (2) Impedance parameter correction model: A humidity compensation model is established for the impedance parameters, and the formula is as follows: in, This represents the real part of the impedance measured under ambient humidity. This is a temperature and humidity coupling compensation function. This is the real part of the corrected impedance; (3) Temperature correction model for electrochemical impedance spectroscopy: An Arrhenius-type temperature correction model was used to uniformly correct the electrochemical impedance spectroscopy data measured at different temperatures to the equivalent values ​​at the reference temperature. The specific formula is as follows: The impedance value is the actual value measured at ambient temperature T. The reference temperature obtained for calibration The impedance value below; Activation energy is a characteristic parameter of the material itself. This is the universal gas constant.

4. The method for constructing a battery health status assessment and fault diagnosis model according to claim 1, characterized in that, In step S3, the constructed high-dimensional feature vector is represented as follows: Where V is voltage, I is current, T is temperature, and H is humidity. The number of loops. Due to internal communication barriers, DC internal resistance, The rate of change of voltage. The rate of change of current, (@f1) represents the real part of the impedance at the predetermined frequency f1. The imaginary part of the impedance at a predetermined frequency f1, The phase angle at the predetermined frequency f3.

5. The method for constructing a battery health status assessment and fault diagnosis model according to claim 1, characterized in that, In step S4, the hierarchical fusion ensemble learning framework includes a feature preprocessing layer, a model parallel layer, and a result collaboration layer; Feature preprocessing layer: Performs unified preprocessing on the high-dimensional feature vector F and generates a feature subset for SOC estimation. and feature subsets for SOH estimation / RUL prediction : Where V is voltage, I is current, T is temperature, and H is humidity. The number of loops. Due to internal communication barriers, DC internal resistance, The rate of change of voltage. The rate of change of current, (@f1) represents the real part of the impedance at the predetermined frequency f1. The imaginary part of the impedance at a predetermined frequency f1, The phase angle at a predetermined frequency f3; Parallel Model Layer: Support Vector Machine (SVM) Model Reception SOC estimation is performed by a neural network (NN) model. SOH estimation and RUL prediction were performed, with the two models running in parallel. Results Collaboration Layer: Performs a consistency check on the SOC value output by the Support Vector Machine (SVM) model and the SOH value output by the Neural Network (NN) model. If the SOC value exceeds the SOH value, the SOC value is limited to the SOH value to ensure that the evaluation results are reasonable.

6. The method for constructing a battery health status assessment and fault diagnosis model according to claim 1, characterized in that, In step S4, the neural network NN model introduces a two-stage collaborative training strategy during the training phase: Phase 1: Independent training, using individual methods and Train the Support Vector Machine (SVM) model and the Neural Network (NN) model until convergence; The second stage: Collaborative fine-tuning. The parameters of the Support Vector Machine (SVM) model are fixed, and the SOC estimate output by the SVM model is added as a new feature. After being spliced ​​together, they are used as input to the neural network (NN) model to fine-tune the NN model so that its SOH estimation and RUL prediction can sense and adapt to the dynamic changes of SOC.

7. The method for constructing a battery health status assessment and fault diagnosis model according to claim 1, characterized in that, In step S4, a Support Vector Machine (SVM) model and a Support Vector Regression (SVR) model are used to introduce... - Insensitive loss function optimization, the objective function is: The constraints are: Furthermore, a radial basis function (RBF) kernel is used, with the following expression: In the above formula, w is the weight vector; b is the bias term; and ||w||² is the margin maximization term. This is the error penalty term, representing the sum of the total costs for all samples violating the constraints; This is a slack variable, representing the degree of deviation when the deviation between the predicted and actual values ​​of the sample exceeds a preset insensitive interval; Let be the slack variable for the i-th sample; Let be the dual relaxation variable of the i-th sample; m be the total number of samples; and C be the penalty factor. This represents the i-th input sample; For reference sample; Represents the i-th input sample The corresponding target value; (·) is a high-dimensional space function; γ is the kernel function parameter; exp() is the natural exponential function, which converts the distance into a similarity value in the interval [0,1].

8. The method for constructing a battery health status assessment and fault diagnosis model according to claim 1, characterized in that, In step S4, the neural network (NN) model adopts a structure with a shared bottom-level feature extraction layer and two independent output layers in the upper layers; the input layer receives high-dimensional feature vectors; the shared hidden layer consists of 2-3 fully connected layers with the activation function being... Activation function; Output layer 1 is used for SOH estimation, using the Sigmoid activation function, with an output range of (0,1), corresponding to 0%~100% SOH; Output layer 2 is used for RUL prediction, using a linear activation function, outputting the remaining battery cycle life or number of days; The loss function used is a combined loss function: in Mean square error, The mean absolute error is given, and α and β are hyperparameters. It is expressed as follows: It is expressed as follows: In the above formula, m is the total number of samples; This represents the true value of the battery health status corresponding to the i-th sample; Let be the predicted value of the battery health status corresponding to the i-th sample; This represents the true value of the remaining useful life corresponding to the i-th sample. Let be the predicted remaining useful life for the i-th sample.

9. The method for constructing a battery health status assessment and fault diagnosis model according to claim 1, characterized in that, In step S5, the state assessment results obtained in step S4, including the SOC estimation results, SOH estimation results, and RUL prediction results and their anomaly indicators, are fused with the high-dimensional feature vector constructed in step S3 to form an enhanced feature vector, as shown in the following formula: Let the high-dimensional feature vector constructed in step S3 be... , dimension And let the SOC value, RUL value, SOH value, and anomaly flag vector obtained in step S4 be respectively... , , and , It is A binary vector of dimension 1; The enhanced feature vector E is formed through a concatenation operation: Where [* ; *] denotes vector concatenation, and the dimension of the enhanced feature vector E is . +3+ .

10. The method for constructing a battery health status assessment and fault diagnosis model according to claim 1, characterized in that, Step S6 is implemented as follows: The enhanced feature vector E is shaped into a one-dimensional signal and input into the CNN model, and the following operations are performed: (1) Convolutional layer operations The output of the k-th feature map of the l-th convolutional layer is: Where * represents a one-dimensional convolution operation; It is the input of the m-th feature map of the (l-1)-th layer; These are the convolution kernel weights that connect the m-th feature map to the k-th feature map; It is the bias term of the k-th feature map; () is the activation function; (2) Pooling layer operations: Pooling refers to either max pooling or average pooling. (3) Output layer calculation: The Softmax output layer calculates a probability vector based on the output of the pooling layer operations: Where n represents the total number of fault categories; These represent the probabilities that the model considers the input data to belong to fault type 1, 2, ..., n, respectively; the sum of all probabilities is 1, i.e. + +...+ =1; (4) Fault classification decision: pass The function finds the category with the highest probability from the probability vector y and uses it as the final fault diagnosis result: The function returns the index number of the category corresponding to the highest probability value.

Images