Power battery health state monitoring method and system based on multi-source feature fusion and dynamic weight

Abstract

The application relates to the field of battery capacity monitoring and discloses a power battery health state monitoring method and system based on multi-source feature fusion and dynamic weight. t , a current sequence I t and a temperature sequence T t , performing sliding window filtering and outlier rejection on the original data to obtain a pretreatment data set, extracting a time domain feature vector and a frequency domain feature vector from the pretreatment data set to constitute a multi-source feature matrix, constructing a dynamic weight distribution model, calculating a feature weight vector according to a current SOC value of the battery, performing Hadamard product calculation on the multi-source feature matrix and the feature weight vector to generate a weighted feature matrix, inputting the weighted feature matrix into a pre-trained hybrid neural network and outputting a health state parameter; the technical advantage of the application is that the multi-physical field coupling signs before capacity sudden drop and the dynamic weight model can be captured, the battery capacity sudden drop process and the health state thereof can be accurately predicted.

Description

Technical Field

[0001] This invention relates to the field of battery capacity monitoring, and in particular to a method and system for monitoring the health status of power batteries based on multi-source feature fusion and dynamic weighting. Background Technology

[0002] In electric new energy vehicles, accurate State of Health (SOH) monitoring helps battery management systems optimize charging and discharging strategies, effectively extending battery life and reducing user costs. Simultaneously, real-time monitoring of battery health status can provide early warnings of potential risks, enhancing vehicle safety. For energy storage power stations, reliable health status monitoring facilitates refined management of battery packs. By understanding the health status of each individual battery cell, battery operating conditions can be rationally allocated, maximizing the overall efficiency of the battery pack, improving the stability and economy of the energy storage system, and continuously improving the accuracy and efficiency of SOH monitoring. This can promote the large-scale application of power batteries in areas such as secondary use, fostering the sustainable development of the new energy industry chain, and contributing to the achievement of "dual carbon" goals (carbon reduction and emission reduction).

[0003] Current methods for monitoring the health status of power batteries primarily target lithium-ion batteries, relying on features such as voltage plateau slope and capacity increment analysis. However, these methods cannot monitor the health of liquid metal batteries, which exhibit significantly different aging mechanisms. Liquid metal batteries experience sudden, abrupt capacity drops, unlike the linear, gradual decreases seen in lithium-ion batteries. These drops originate from the instantaneous formation of irreversible solid deposits at the liquid electrode interface, leading to irreversible loss of active material. Traditional slope-based methods also fail in monitoring the health status of liquid metal batteries. The core reason is insufficient feature sensitivity, resulting in minimal changes in voltage / current slopes before and after the capacity drop, failing to provide effective early warnings. Furthermore, the battery can still operate after the drop, exhibiting nonlinear decoupling and multi-physics coupling between the internal resistance growth rate and capacity decay rate. Anomalies in features such as temperature gradients and harmonic components are already apparent before the drop, but the reliability of a single feature is low. Therefore, current technologies lack a monitoring scheme for dynamic weight allocation and fusion of multi-source features in liquid metal batteries, making it impossible to predict sudden drop events in advance. Summary of the Invention

[0004] This device provides a method and system for monitoring the health status of power batteries based on multi-source feature fusion and dynamic weighting, solving the technical problem of the lack of accurate and real-time monitoring of the health status of liquid metal batteries in existing technologies. The specific implementation method is as follows:

[0005] A method for monitoring the health status of power batteries based on multi-source feature fusion and dynamic weights includes the following steps:

[0006] S100: Real-time acquisition of charge and discharge cycle data of the power battery, including voltage sequence. Current sequence Based on the temperature sequence Tt, the original data is subjected to sliding window filtering and outlier removal to obtain a preprocessed dataset. ;

[0007] S200, from the preprocessed dataset Extracting temporal feature vectors and frequency domain eigenvectors To form a multi-source feature matrix ;

[0008] S300. Construct a dynamic weight allocation model and calculate the feature weight vector based on the current SOC value of the battery. Multi-source feature matrix With feature weight vector Perform Hadamard product calculation to generate a weighted feature matrix. ;

[0009] S400, Weighted feature matrix Input a pre-trained LSTM-Transformer hybrid neural network and output health state parameters. ,in Indicates the capacity decay rate. This indicates the growth rate of internal resistance.

[0010] Furthermore, in step S200, the extraction step of the time-domain feature vector includes:

[0011] S211. Calculate the voltage range:

[0012] ;

[0013] Where max(V) t ) represents the maximum value in the voltage sequence, min(V) t () represents the minimum value in the voltage sequence;

[0014] S212. Calculate the average rate of change of current:

[0015] ;

[0016] Where I t+1 Let n be the current sequence for the next time step, and n be the total sampling time.

[0017] S213. Calculate the temperature gradient using the Sobel operator:

[0018] .

[0019] Furthermore, in step S200, the extraction of the frequency domain feature vector includes:

[0020] S221. Perform a Fast Fourier Transform on the voltage sequence to extract the amplitude of the third harmonic:

[0021] ;

[0022] in For Fast Fourier Transform, Ω is the fundamental frequency of charging and discharging;

[0023] S222. Calculate the power spectral density of the current sequence and extract the peak values:

[0024] ;

[0025] Where N is the total number of sampling points, f is the frequency variable, and fs is the sampling frequency.

[0026] Furthermore, the dynamic weight allocation model described in step S300 is as follows:

[0027] ;

[0028] Where i and j represent the indexes of the feature dimensions, i specifically refers to the particular feature dimension for which the weights are currently being calculated, j represents the traversal variable in the summation operation, and k i The sensitivity coefficient of the i-th dimension feature is determined through calibration experiments; SOC is the current state of charge of the battery, calculated in real time using the ampere-hour integration method.

[0029] The process of determining the sensitivity coefficient ki through calibration experiments is as follows:

[0030] S311. Measure the partial derivative of characteristic Fi with respect to SOH at multiple SOC operating points:

[0031] ;

[0032] Where m = 1, 2, 3…M represents the index of the SOC working point; SOC m This represents the state of charge value at the m-th operating point;

[0033] S312. Solve for the sensitivity coefficient using linear regression:

[0034] .

[0035] Furthermore, the structure of the LSTM-Transformer hybrid neural network in step S400 includes: an input layer, an LSTM layer, a Transformer encoding layer, and a fully connected layer; the pre-training in step S400 employs a transfer learning strategy, specifically including:

[0036] S411. Pre-train an LSTM-Transformer hybrid neural network on a standard battery aging dataset;

[0037] S412. The LSTM-Transformer hybrid neural network is transferred to the target battery model through the Domain Adaptive Module (DA-MDD).

[0038] A power battery health status monitoring system based on multi-source feature fusion and dynamic weighting includes a processor and a storage medium; the processor loads and executes instructions and data in the storage medium to implement a power battery health status monitoring method based on multi-source feature fusion and dynamic weighting.

[0039] In summary, the beneficial effects provided by this invention are:

[0040] 1. This invention achieves high-precision early warning of sudden capacity drops by fusing time-domain (voltage difference, temperature gradient) and frequency-domain (harmonic amplitude, PSD peak) features to capture multi-physics coupling signs before a sudden capacity drop.

[0041] 2. The dynamic weights of this invention improve robustness. The dynamic weight allocation model adjusts the feature contribution in real time based on SOC, which solves the feature sensitivity drift problem under different working conditions. Compared with the fixed weight model, the prediction error is lower.

[0042] 3. This invention enables rapid deployment across different models. By using an LSTM-Transformer hybrid neural network that combines the capabilities of temporal modeling (LSTM) and global dependency capture (Transformer), the transfer learning strategy DA-MDD reduces the training data requirements of the target domain by 80%, and can adapt to different liquid metal battery variants. Attached Figure Description

[0043] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0044] Figure 2 This is a schematic diagram of the hardware device operation of the present invention.

[0045] Explanation of reference numerals in the attached figures:

[0046] 401. Monitoring equipment; 402. Processor; 403. Storage medium. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0048] Before formally describing the present invention, a general description of the invention's solution is provided for ease of understanding. The following description, in conjunction with the appendix, is further detailed. Figure 1-2This application will be described in further detail.

[0049] This application discloses a method and system for monitoring the health status of power batteries based on multi-source feature fusion and dynamic weighting.

[0050] Example 1

[0051] refer to Figure 1 This embodiment discloses a method for monitoring the health status of power batteries based on multi-source feature fusion and dynamic weights, including the following steps:

[0052] S100: Real-time acquisition of charge and discharge cycle data of the power battery, including voltage sequence V. t Current sequence I t and temperature sequence T t The original data is subjected to sliding window filtering and outlier removal to obtain a preprocessed dataset. .

[0053] It should be noted that the preprocessing in step S100 includes:

[0054] S111. A median filter with a window width of N = 10 sampling periods is used to process the voltage sequence V. t ;

[0055] S112, Current sequence I based on Z-score normalization method t Normalize:

[0056] ;

[0057] Where mean represents the average value and std represents the standard deviation;

[0058] S113. Eliminating temperature series T using the 3σ criterion t Outliers in the data.

[0059] S200, from the preprocessed dataset Extracting temporal feature vectors and frequency domain eigenvectors To form a multi-source feature matrix .

[0060] Firstly, it should be noted that the extraction of the time-domain feature vector in step S200 includes:

[0061] S211. Calculate the voltage range:

[0062] ;

[0063] Where max(Vt) is the maximum value in the voltage sequence, and min(Vt) is the minimum value in the voltage sequence;

[0064] S212. Calculate the average rate of change of current:

[0065] ;

[0066] It+1 represents the current sequence at the next moment, and n represents the total sampling time.

[0067] S213. Calculate the temperature gradient using the Sobel operator:

[0068] .

[0069] Secondly, it should be noted that the extraction of frequency domain feature vectors in step S200 includes:

[0070] S221. Perform a Fast Fourier Transform on the voltage sequence to extract the amplitude of the third harmonic:

[0071] ;

[0072] in For Fast Fourier Transform, Ω is the fundamental frequency of charging and discharging, through... Calculation, where Tcycle represents the charge / discharge cycle;

[0073] S222. Calculate the power spectral density of the current sequence and extract the peak values:

[0074] ;

[0075] Where N is the total number of sampling points, f is the frequency variable, and fs is the sampling frequency.

[0076] Therefore, the multi-source feature matrix is:

[0077] =[F1,F2,F3,F4,F5]=[ , ].

[0078] S300. Construct a dynamic weight allocation model and calculate the feature weight vector based on the current SOC value of the battery. Multi-source feature matrix With feature weight vector Perform Hadamard product calculation to generate a weighted feature matrix. ;

[0079] It should be noted that the dynamic weight allocation model in step S3 is as follows:

[0080] ;

[0081] Where i and j both represent the index of the feature dimension, but i specifically refers to the specific feature dimension for which the weight is currently being calculated, and j represents the traversal variable in the summation operation; ki is the sensitivity coefficient of the i-th feature, which is determined through calibration experiments; SOC is the current state of charge of the battery, which is calculated in real time using the ampere-hour integration method.

[0082] Specifically, the process of determining the sensitivity coefficient ki through calibration experiments is as follows:

[0083] S311. Measure the partial derivative of characteristic Fi with respect to SOH at multiple SOC operating points:

[0084] ;

[0085] Where m = 1, 2, 3…M represents the index of the SOC operating point; SOCm is the state of charge value of the m-th operating point;

[0086] S312. Solve for the sensitivity coefficient using linear regression:

[0087] .

[0088] S400, Weighted feature matrix Input a pre-trained LSTM-Transformer hybrid neural network, and output health state parameters:

[0089] ;

[0090] in Indicates the capacity decay rate. This indicates the growth rate of internal resistance.

[0091] First, it should be noted that the structure of the LSTM-Transformer hybrid neural network includes: an input layer, an LSTM layer, a Transformer encoding layer, and a fully connected layer.

[0092] Specifically, the input layer receives a 128-dimensional feature vector; the LSTM layer has 64 hidden units, and the activation function is:

[0093] ;

[0094] Transformer encoding layer: 4 attention heads, positional encoding uses:

[0095] ;

[0096] The fully connected layer is used to output 2D SOH parameters.

[0097] Secondly, it should be noted that the pre-training in step S400 employs a transfer learning strategy:

[0098] S411. Pre-train an LSTM-Transformer hybrid neural network on a standard battery aging dataset;

[0099] S412. The LSTM-Transformer hybrid neural network is transferred to the target battery model through the Domain Adaptive Module (DA-MDD).

[0100] After step S00, when Greater than 20% and When the battery level exceeds 25%, a battery balancing command is triggered, and a health status visualization report is generated, including the SOH degradation curve and remaining lifespan prediction.

[0101] Example 2

[0102] refer to Figure 1 This embodiment discloses the application of the power battery health status monitoring method to the monitoring of a liquid metal battery system with a rated capacity of 100Ah, including the following steps:

[0103] Step S1000, Data Acquisition and Preprocessing:

[0104] The voltage sequence (sampling frequency 10Hz), current sequence, and temperature sequence were collected for one charge-discharge cycle.

[0105] Sliding window filtering: The voltage sequence is filtered using median filtering with a window width of N=10 to eliminate impulse noise;

[0106] Current normalization: Z-score normalization, where the parameters are set to μ=0, σ=1;

[0107] Temperature outlier removal: The 3σ criterion is used to remove data points that exceed ±3 standard deviations.

[0108] Step S2000, Multi-source feature extraction:

[0109] The specific temporal features collected are as follows:

[0110] Voltage range of data acquisition: = 0.32V;

[0111] Average rate of change of current collected: = 0.15A / s;

[0112] Temperature gradient acquisition: = 0.8°C / mm;

[0113] The specific frequency domain features collected are as follows:

[0114] Acquisition voltage: FFT third harmonic amplitude =0.05V, Tcycle=3600s;

[0115] Peak value of current power spectral density collected: =0.12W / Hz.

[0116] Step S3000, Dynamic Weight Allocation:

[0117] With the current SOC at 60%, the sensitivity coefficient k obtained through calibration experiments is [0.8, 1.2, 0.5, 1.0, 1.5].

[0118] Taking weight vector w1 as an example: w1 = exp(0.8×60) / Σexp(k j ×SOC) ≈ 0.18.

[0119] Step S4000, SOH prediction:

[0120] Weighted feature input pre-trained LSTM-Transformer network (structure: 128-dimensional input layer → 64-unit LSTM layer → 4-head attention Transformer encoding layer → 2-output fully connected layer);

[0121] Output: Capacity decay rate =25.3%, internal resistance growth rate =28.7%;

[0122] Warning triggered: >20% and When the SOH level exceeds 25%, the balance control is activated, and a health status visualization report is generated, which includes the SOH degradation curve and remaining service life prediction.

[0123] Example 3

[0124] refer to Figure 2 This embodiment discloses a power battery health status monitoring system based on multi-source feature fusion and dynamic weighting. The hardware configuration of the monitoring system includes a monitoring device 401, a processor 402, and a storage medium 403. In this structure, the processor 402 and the storage medium 403 are mounted on the monitoring device 401, and the monitoring device 401 is implemented based on the power battery health status monitoring method based on multi-source feature fusion and dynamic weighting.

[0125] The processor 402 loads and executes the instructions and data in the storage medium 403, and the storage medium 403 stores the instructions and data. In this structure, the implementation of the processor 402 and the storage medium 403 is based on the power battery health status monitoring method of multi-source feature fusion and dynamic weight in Embodiment 1.

[0126] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for monitoring the health status of power batteries based on multi-source feature fusion and dynamic weighting, characterized in that, Includes the following steps: S100: Real-time acquisition of charge and discharge cycle data of the power battery, including voltage sequence V. t Current sequence I t and temperature sequence T t The original data is subjected to sliding window filtering and outlier removal to obtain a preprocessed dataset. ; S200, from the preprocessed dataset Extracting temporal feature vectors and frequency domain eigenvectors To form a multi-source feature matrix ; For voltage range, The average rate of change of current. For temperature gradient, The amplitude of the third harmonic. Calculate the peak power spectral density for the current sequence; The voltage range, temperature gradient, voltage harmonic amplitude and current peak characteristics of the power battery charge and discharge cycle were collected to capture the multi-physics coupling signs before the capacity drop; S300. Construct a dynamic weight allocation model and calculate the feature weight vector based on the current SOC value of the power battery. Multi-source feature matrix With feature weight vector Perform Hadamard product calculation to generate a weighted feature matrix. The dynamic weight allocation model adjusts the feature contribution in real time based on SOC; S400, Weighted feature matrix Input a pre-trained LSTM-Transformer hybrid neural network, and output health state parameters: ； in Indicates the capacity decay rate. Indicates the internal resistance growth rate; Once the S500, capacity decay rate, and internal resistance growth rate meet the standards, they are used to trigger the equalization command of the power battery, and at the same time generate a health status visualization report containing the SOH degradation curve and the remaining service life prediction. The dynamic weight allocation model described in step S300 is as follows: ； Where i specifically refers to the index of a particular feature dimension in the current weight calculation, j represents the traversal variable in the summation operation, and k i The sensitivity coefficient of the i-th feature is determined through calibration experiments; SOC is the current state of charge of the battery, calculated in real time using the ampere-hour integration method. The process of determining the sensitivity coefficient ki through calibration experiments is as follows: S311. Measure the partial derivative of characteristic Fi with respect to SOH at multiple SOC operating points, where SOH represents the battery's state of health. ； Where m = 1, 2, 3…M represents the index of the SOC working point; SOC m This represents the state of charge value at the m-th operating point; S312. The sensitivity coefficient is calculated using the following formula: 。 2. The method for monitoring the health status of a power battery based on multi-source feature fusion and dynamic weights according to claim 1, characterized in that, In step S200, the extraction step of the time-domain feature vector includes: S211. Calculate the voltage range: ； Where max(V) t ) represents the maximum value in the voltage sequence, min(V) t () represents the minimum value in the voltage sequence; S212. Calculate the average rate of change of current: ； Where I t+1 Let be the current sequence at time t+1, and n be the total sampling time; S213. Calculate the temperature gradient using the Sobel operator: 。 3. The method for monitoring the health status of a power battery based on multi-source feature fusion and dynamic weights according to claim 1, characterized in that: In step S200, the extraction of the frequency domain feature vector includes: S221. Perform a Fast Fourier Transform on the voltage sequence to extract the amplitude of the third harmonic: ； in For Fast Fourier Transform, Ω is the fundamental frequency of charging and discharging; S222. Calculate the power spectral density of the current sequence and extract the peak values: ； Where N is the total number of sampling points, f is the frequency variable, and fs is the sampling frequency.

4. The method for monitoring the health status of a power battery based on multi-source feature fusion and dynamic weights according to claim 1, characterized in that, The structure of the LSTM-Transformer hybrid neural network described in step S400 includes: an input layer, an LSTM layer, a Transformer encoding layer, and a fully connected layer.

5. The method for monitoring the health status of a power battery based on multi-source feature fusion and dynamic weights as described in claim 4, characterized in that, The pre-training described in step S400 employs a transfer learning strategy, specifically including: S411. Pre-train an LSTM-Transformer hybrid neural network on a standard battery aging dataset; S412. The LSTM-Transformer hybrid neural network is transferred to the target battery model through the Domain Adaptive Module (DA-MDD).

6. A power battery health status monitoring system based on multi-source feature fusion and dynamic weighting, characterized in that, The monitoring system employs the power battery health status monitoring method based on multi-source feature fusion and dynamic weighting as described in any one of claims 1 to 5.

Images