Multi-scene lithium ion battery fault diagnosis method and system

Abstract

The invention discloses a multi-scene lithium ion battery fault diagnosis method and system, and the method comprises the steps: data preparation: carrying out the basic performance test of a battery, obtaining the parameters of the battery, constructing a first-order RC equivalent circuit model, simulating a fault, judging the precision of the model, and simulating the batteries of different fault types; data processing: carrying out preprocessing and feature extraction on the battery data from different scenes, and carrying out fault mode identification by adopting multi-feature combination; and model construction and training: using a consistency regularization semi-supervised learning model, using labeled data to train a K-means model, testing and re-labeling unlabeled data, and realizing iterative training of the model. The first-order RC equivalent circuit model is constructed to generate data, so that a large amount of experiment cost can be saved, and the data can be obtained without depending on a large amount of actual fault experiments.

Description

Technical Field

[0001] The patent of this invention belongs to the field of lithium-ion battery fault diagnosis technology, and specifically relates to a multi-scenario lithium-ion battery fault diagnosis method and system. Background Art

[0002] Lithium-ion batteries, with their high energy density and long cycle life, are widely used in electric vehicles, energy storage systems, rail transit, and other fields. However, ensuring accurate diagnosis of internal short circuit (ISC) faults and consistency anomalies in lithium-ion battery systems is crucial to maintaining the high safety and long life of battery systems.

[0003] Currently, ISC faults are hidden and difficult to diagnose directly through voltage signals. Inconsistencies are ubiquitous across all battery cells, making them difficult to decouple from other faults. Furthermore, the characteristics of ISC and consistency anomalies are highly similar, complicating accurate diagnosis. Traditional fault diagnosis methods have limitations when dealing with such complex issues: 1. Relying on a large amount of labeled data for training, but in actual applications, it is costly and difficult to obtain a large amount of accurately labeled fault data.

[0004] 2. For different fault types with similar characteristics, the diagnostic accuracy is low and cannot meet the high requirements for lithium-ion battery fault diagnosis in practical applications. Summary of the Invention

[0005] In view of the shortcomings of the prior art, the present invention provides a multi-scenario lithium-ion battery fault diagnosis method, which includes: Data preparation: Perform basic performance tests on the battery to obtain battery parameters and build a first-order RC Equivalent circuit model, simulates faults and determines model accuracy, while simulating batteries with different fault types; Data processing: preprocessing and feature extraction of battery data from different scenarios, Use multiple features to identify fault patterns; Model construction and training: Use a consistency regularized semi-supervised learning model, train the K-means model with labeled data, test and relabel the unlabeled data, and implement iterative training of the model.

[0006] Furthermore, the data preparation specifically includes: providing a battery pack comprising a plurality of battery cells; Obtain battery parameters by performing basic performance tests on the battery, wherein the basic performance tests include capacity test, low current test, and internal resistance test; Build an equivalent circuit model to simulate the operating conditions of the series system, and define and simulate ISC, CA, and RA faults; The accuracy of the model is determined using measured voltage data. The internal resistance, capacity, and SOC of all batteries are modified to simulate inconsistencies. Gaussian distribution is used as the parameter distribution method to simulate batteries with different fault types.

[0007] Furthermore, the data processing specifically includes: applying the same data preprocessing method to battery data from different scenarios to extract corresponding characteristic operating conditions; Data preprocessing includes uniform resampling to a specific Hz and adding simulated noise to match the noise characteristics of actual working conditions; Standardize features, extract sliding windows, and process extreme values ​​to enhance feature discrimination; The voltage change is calculated, the curve is divided into voltage segments, and the start and end voltage points are specified. The fault mode is identified by combining four features: TSM, correlation coefficient, ranking score and voltage range.

[0008] Furthermore, the model construction and training specifically include: Use a consistency regularized semi-supervised learning model to train labeled data and test and relabel unlabeled data; Use labeled data to train the K-means model. Use labeled samples from simulation data to extract four features: TSM, voltage score, correlation coefficient, and voltage range, and train the K-means clustering model. Use PCA dimensionality reduction visualization to distinguish normal batteries from faulty batteries; Using consistency regularization, we add small perturbations to the unlabeled actual running data, input the trained K-means model, and compare the prediction results before and after the perturbation. If the results are consistent, the data is considered to have been successfully anti-perturbated and is added to the model iterative training as a new label sample.

[0009] Furthermore, the TSM calculation method is as follows: a sliding window method is used for the time series voltage data, local features are extracted according to a fixed window size and step size, and then the dimension effect is eliminated through normalization processing to calculate the cumulative voltage deviation of the battery in the charge and discharge stage, which is used to reflect the deviation trend of the battery voltage curve from the normal state.

[0010] Furthermore, the voltage score is calculated as follows: based on the performance of the battery voltage during the charge and discharge process, the voltage data is extracted using a sliding window, an exponential function is used to enhance the sensitivity of voltage fluctuations, and a quantitative score is performed based on the degree to which the voltage deviates from the normal range to highlight the fault characteristics.

[0011] Furthermore, the correlation coefficient is calculated by calculating the correlation coefficient between the single cell voltage curve and the module average voltage curve, extracting the correlation coefficients of different time periods by a sliding window method, and taking the extreme values ​​as features to reduce redundant information.

[0012] Furthermore, the voltage range is calculated by calculating the voltage extremes of the battery within a charge and discharge segment, taking the difference between the beginning and the end of the voltage data, and then normalizing it by the maximum and minimum values ​​to highlight the voltage fluctuation differences of different fault types.

[0013] On the other hand, the present invention also provides a multi-scenario lithium-ion battery fault diagnosis system, the system comprising: The data preparation module is used to perform data preparation steps, obtain battery parameters by performing basic performance tests on the battery, build a first-order RC equivalent circuit model, simulate faults and determine the model accuracy, and simulate batteries with different fault types; The data processing module is used to perform data processing steps, pre-process and extract features of battery data from different scenarios, and use a combination of multiple features to identify fault patterns; The model building and training module is used to perform the model building and training steps, using a consistency regularized semi-supervised learning model, training the K-means model with labeled data, testing and relabeling the unlabeled data, and realizing iterative training of the model.

[0014] Furthermore, the data preparation module, data processing module and model building and training module are implemented by a computer program, and the computer program is stored in a computer-readable storage medium.

[0015] The present invention has the following technical effects: 1. The present invention generates data by constructing a first-order RC equivalent circuit model, which can save a lot of experimental costs and does not need to rely on a large number of actual fault experiments to obtain data.

[0016] 2. The fault characteristics of the present invention are constructed based on the physical properties of the battery, such as voltage change, voltage extreme difference, etc., which are physically explainable and easy to understand the basis of fault diagnosis.

[0017] 3. The present invention uses a consistency regularized semi-supervised model for fault diagnosis training, which effectively solves the problem of unlabeled data in practical applications. By adding disturbances to the unlabeled data for relabeling, it fully utilizes the information of the unlabeled data, thereby improving the generalization ability of the model and the accuracy of fault diagnosis.

[0018] 4. The method of the present invention has been verified under various fault types. The recognition accuracy for internal short circuit and normal batteries reached 100%, the accuracy for small capacity and large internal resistance was 96.7%, and the model F1 value was 0.99; in the unlabeled data test, the average fault accuracy of the test set was 83.6%, the precision was 88.7%, the F1 value was 0.909, and the disturbance resistance of the semi-supervised model reached 90.85%. It can realize accurate diagnosis of typical faults such as internal short circuit, large internal resistance, and small capacity of lithium-ion batteries, and has high practical application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a fault diagnosis flow chart of the present invention; Figure 2 It is the equivalent circuit model and internal short circuit model of the present invention; Figure 3 It is a simulation curve diagram of an internal short-circuit battery and a simulation diagram of an inconsistent battery of the present invention; Figure 4 It is a characteristic curve diagram of the present invention; Figure 5 It is the clustering result of the present invention. DETAILED DESCRIPTION

[0020] The present invention will be described in detail below with reference to the accompanying drawings. Embodiment 1:

[0021] like Figures 1 to 5 As shown, a multi-scenario lithium-ion battery fault diagnosis method includes the following steps: S1, data preparation module, is used to perform data preparation steps, obtain battery parameters by performing basic performance tests on the battery, build a first-order RC equivalent circuit model, simulate faults and judge the model accuracy, and simulate batteries with different fault types.

[0022] S11 , providing a battery pack including a plurality of battery cells.

[0023] S12, obtaining battery parameters by performing basic performance tests on the battery, including but not limited to capacity test, low current test, and internal resistance test.

[0024] S13: Build equivalent circuit models, including but not limited to the Rint model and first-order RC model. Simulate the operating conditions of the series system and define and simulate ISC, CA (capacity abnormality), and RA (internal resistance abnormality) faults.

[0025] S14, using the measured voltage data to determine the accuracy of the model, wherein the internal resistance, capacity, and SOC (state of charge) of all batteries are modified to simulate inconsistency, and parameter distribution methods include but are not limited to Gaussian distribution to simulate batteries with different fault types.

[0026] S2, the data processing module, is used to execute data processing steps, preprocess and extract features of battery data from different scenarios, and use a combination of multiple features to perform fault mode identification.

[0027] S21, using the same data preprocessing method for battery data from different scenarios to extract corresponding characteristic operating conditions.

[0028] S22, data preprocessing includes uniform resampling to a specific Hz and adding simulated noise (such as Gaussian noise) to match the noise characteristics of actual working conditions.

[0029] S23, standardizes features (Z-score), performs sliding window extraction, and performs extreme value processing to enhance feature discrimination.

[0030] S24 calculates voltage changes, divides the curve into voltage segments, and specifies start and end voltage points. Fault pattern identification is performed by combining four features: TSM (Time Shift Measurement), correlation coefficient, ranking score, and voltage range.

[0031] S3, the model building and training module, is used to perform the model building and training steps, using a consistency regularized semi-supervised learning model, training the K-means model with labeled data, testing and relabeling the unlabeled data, and realizing iterative training of the model.

[0032] S31, uses a consistency regularized semi-supervised learning model to train labeled data and test and relabel unlabeled data.

[0033] S32, the specific process is: use labeled data to train the K-means model, use the labeled samples in the simulation data (internal short circuit, large internal resistance, small capacity, normal) to extract four features: TSM, voltage score, correlation coefficient, and voltage range, and train the K-means clustering model.

[0034] S33, through PCA dimensionality reduction visualization, normal batteries and faulty batteries can be clearly distinguished, and the clustering accuracy of various fault types reaches 96.7%-100%.

[0035] S34 uses consistency regularization to add small disturbances (such as Gaussian noise) to the unlabeled actual running data, inputs the trained K-means model, and compares the prediction results before and after the disturbance; if the results are consistent, the data is considered to be successfully anti-disturbance, and it is added as a new label sample to the model iterative training.

[0036] S35, through the consistency regularization strategy, forces the model to learn a robust representation of the data and improves its adaptability to noise and uncertainty in actual data.

[0037] Example 2: S1 Experimental Preparation S11 used the Arbin battery test system on a 1P15S 100Ah lithium iron phosphate battery pack system to simulate ISC, CA, and RA failures at 25°C. S12 builds a first-order RC equivalent circuit model (ECM) based on the MATLAB / Simulink platform, expands the fault range to a continuous value, simulates internal short circuits (connecting resistors of different resistance values ​​in parallel) through experiments, and generates voltage curves for batteries with large internal resistance, small capacity, and normal batteries through simulation. A data set is constructed that includes 30 internal short circuit modules, 30 modules with large internal resistance, 30 modules with small capacity, and 910 normal modules.

[0038] S13 obtains 2145 actual operating module voltage data from two energy storage power stations. This data has no fault tags.

[0039] S2 data preprocessing and working condition extraction S21 resamples the simulation data and actual data uniformly, adjusting the sampling frequency to 1 / 300Hz S22 adds simulated noise, such as Gaussian noise, to the data so that its noise characteristics match the actual operating conditions.

[0040] S23 standardizes data features using the Z-score standardization method; uses the sliding window method to extract data features and performs extreme value processing to enhance the distinguishability of data features.

[0041] S24 calculates the battery voltage change, divides the voltage curve into different voltage segments, and clearly specifies the start and end voltage points of each segment. S3 feature extraction S31TSM (trend shift) calculation: A sliding window method is used on time series voltage data to extract local features according to a fixed window size and step size, and then normalization is performed to eliminate dimensionality effects. The cumulative voltage deviation of the battery during the charge and discharge phase is calculated to reflect the deviation trend of the battery voltage curve from the normal state. S32 voltage scoring calculation: Based on the battery voltage performance during the charge and discharge process, the voltage data is extracted using a sliding window, and an exponential function is used to enhance the sensitivity to voltage fluctuations. A quantitative score is assigned based on the degree to which the voltage deviates from the normal range, highlighting fault characteristics. S33 Correlation coefficient calculation: Calculate the correlation coefficient between the single cell voltage curve and the module average voltage curve. Use the sliding window method to extract the correlation coefficients of different time periods, and take the extreme values ​​as features to reduce redundant information. S34 voltage range calculation: Calculates the voltage range of the battery within a charge and discharge segment, takes the difference between the beginning and end of the voltage data, and then normalizes it by the maximum and minimum values ​​to highlight the voltage fluctuation differences for different fault types. Establishment and classification of S4 semi-supervised learning model S41 uses labeled samples (internal short circuit, large internal resistance, small capacity, normal) in simulation data to extract four features: TSM, voltage score, correlation coefficient, and voltage range, and train the K-means clustering model. S42 uses PCA dimensionality reduction to visualize training results, verifying the differentiation between normal and faulty batteries and ensuring that the clustering accuracy of various fault types reaches 96.7%-100%. S43 uses consistency regularization to add small perturbations, such as Gaussian noise, to unlabeled real-world running data. S44 inputs the disturbed data into the trained K-means model and compares the prediction results before and after the disturbance. If the results are consistent, the data is judged to be successfully anti-disturbance, and it is added to the model as a new label sample for iterative training. Through the consistency regularization strategy, the model's adaptability to noise and uncertainty in actual data is improved.

[0042] S5 Experimental Results Verification The S51 was tested with labeled data, and the results showed that the accuracy rate for identifying internal short-circuit and normal batteries reached 100%, the accuracy rate for small capacity and large internal resistance was 96.7%, and the model F1 value was 0.99. S52 was tested on unlabeled data, and the average fault accuracy of the test set was 83.6%, the precision was 88.7%, the F1 value was 0.909, and the disturbance resistance of the semi-supervised model reached 90.85%. The verification model has strong robustness to actual data and can accurately diagnose various typical faults of lithium-ion batteries. The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

[0043] The information disclosed in the background technology section of this document is only intended to deepen the understanding of the overall background technology of the present invention, and should not be regarded as an admission or any form of suggestion that the information constitutes the prior art already known to those skilled in the art.

Claims

1. A multi-scenario lithium-ion battery fault diagnosis method, characterized in that: The method comprises: Data preparation: Perform basic performance tests on the battery to obtain battery parameters, construct a first-order RC equivalent circuit model, simulate faults and determine model accuracy, and simulate batteries with different fault types. Data processing: pre-processing and feature extraction of battery data from different scenarios, and fault pattern recognition using multi-feature combination; Model construction and training: Use a consistency regularized semi-supervised learning model, train the K-means model with labeled data, test and relabel the unlabeled data, and implement iterative training of the model.

2. The method according to claim 1, wherein The data preparation specifically includes: providing a battery pack comprising a plurality of battery cells; Obtain battery parameters by performing basic performance tests on the battery, wherein the basic performance tests include capacity test, low current test, and internal resistance test; Build an equivalent circuit model to simulate the operating conditions of the series system, and define and simulate ISC, CA, and RA faults; The accuracy of the model is determined using measured voltage data. The internal resistance, capacity, and SOC of all batteries are modified to simulate inconsistencies. Gaussian distribution is used as the parameter distribution method to simulate batteries with different fault types.

3. The method according to claim 1, wherein The data processing specifically includes: applying the same data preprocessing method to battery data from different scenarios to extract corresponding characteristic operating conditions; Data preprocessing includes uniform resampling to a specific Hz and adding simulated noise to match the noise characteristics of actual working conditions; Standardize features, extract sliding windows, and process extreme values ​​to enhance feature discrimination; The voltage change is calculated, the curve is divided into voltage segments, and the start and end voltage points are specified. The fault mode is identified by combining four features: TSM, correlation coefficient, ranking score and voltage range.

4. The method according to claim 1, wherein The model construction and training specifically include: Use a consistency regularized semi-supervised learning model to train labeled data and test and relabel unlabeled data; Use labeled data to train the K-means model. Use labeled samples from simulation data to extract four features: TSM, voltage score, correlation coefficient, and voltage range, and train the K-means clustering model. Use PCA dimensionality reduction visualization to distinguish normal batteries from faulty batteries; Using consistency regularization, we add small disturbances to the unlabeled actual running data, input the trained K-means model, and compare the prediction results before and after the disturbance. If the results are consistent, the data is considered to be successfully anti-disturbance and is added to the model iterative training as a new label sample.

5. The method according to claim 4, wherein The TSM calculation method is as follows: using a sliding window method on time series voltage data, extracting local features according to a fixed window size and step size, then eliminating dimensionality effects through normalization, and calculating the cumulative voltage deviation of the battery during the charge and discharge phase to reflect the deviation trend of the battery voltage curve from the normal state.

6. The method according to claim 4, wherein The voltage score is calculated by performing a sliding window extraction on the voltage data based on the battery voltage performance during the charge and discharge process, taking an exponential function to enhance the sensitivity of voltage fluctuations, and performing a quantitative score based on the degree to which the voltage deviates from the normal range to highlight the fault characteristics.

7. The method according to claim 4, wherein The correlation coefficient is calculated by calculating the correlation coefficient between the single cell voltage curve and the module average voltage curve, extracting the correlation coefficients of different time periods by a sliding window method, and taking the extreme value as a feature to reduce redundant information.

8. The method according to claim 4, wherein The voltage range is calculated by calculating the voltage extremes of the battery within a charge and discharge segment, taking the difference between the beginning and end of the voltage data, and then normalizing it by the maximum and minimum values ​​to highlight the voltage fluctuation differences of different fault types.

9. A multi-scenario lithium-ion battery fault diagnosis system, characterized in that: The system comprises: The data preparation module is used to perform data preparation steps, obtain battery parameters by performing basic performance tests on the battery, build a first-order RC equivalent circuit model, simulate faults and determine the model accuracy, and simulate batteries with different fault types; The data processing module is used to perform data processing steps, pre-process and extract features of battery data from different scenarios, and use a combination of multiple features to identify fault patterns; The model building and training module is used to perform the model building and training steps, using a consistency regularized semi-supervised learning model, training the K-means model with labeled data, testing and relabeling the unlabeled data, and realizing iterative training of the model.

10. The system according to claim 9, wherein The data preparation module, data processing module and model building and training module are implemented by computer programs, and the computer programs are stored in a computer-readable storage medium.

Images