Lithium battery health state prediction method based on BP neural network

Abstract

The invention provides a lithium battery health state prediction method based on a BP neural network, and the method comprises the steps: collecting an open-circuit voltage and an internal resistance signal of a lithium battery in real time, and inputting the open-circuit voltage and the internal resistance signal into a BP neural network input layer in combination with a pre-stored average life parameter and a real-time charge state parameter; the hidden layer adopts a hyperbolic tangent transfer function to execute nonlinear transformation, and the output layer generates residual capacity and health life prediction values through an S-type transfer function. And the system calculates the deviation between the predicted value and the measured value, if the deviation exceeds a preset threshold value, an error back propagation mechanism is started to iteratively update the weight and the bias item, the prediction and updating processes are repeated until the deviation is lower than the threshold value, and finally a health state prediction result is output. According to the method, three technical bottlenecks of battery degradation mutation response lag, edge deployment resource limitation and insufficient multi-source feature fusion are solved, and high-robustness prediction support is provided.

Description

Technical Field

[0001] This invention relates to the field of battery technology, and in particular to a method for predicting the health status of lithium batteries based on a BP neural network. Background Technology

[0002] In the field of lithium battery health state prediction, traditional methods generally rely on single electrochemical models or static machine learning algorithms, which are insufficient to cope with nonlinear abrupt changes and complex operating conditions during battery degradation. Existing technologies typically employ fixed threshold error correction mechanisms, which cannot dynamically respond to sudden changes in battery aging patterns. Furthermore, feature extraction is often limited to basic parameters such as voltage and current, neglecting the synergistic effects of ambient temperature and multi-scale discharge behavior, leading to predictions that deviate from the actual degradation trajectory. In addition, conventional neural network models face computational resource bottlenecks when deployed on edge devices; the high power consumption of floating-point operations clashes sharply with real-time requirements, limiting the engineering applicability of prediction systems. Although some studies have attempted to introduce signal denoising or feature fusion modules, they lack quantitative analysis of the dynamic correlation of temporal covariance, and the rigid model weight adjustment mechanisms make it difficult to accurately capture degradation inflection points. Summary of the Invention

[0003] This invention proposes a method for predicting the health status of lithium batteries based on a backpropagation neural network, comprising: S1. Real-time acquisition of open-circuit voltage and internal resistance signals of lithium batteries, and input of the acquired open-circuit voltage and internal resistance signals, as well as the pre-stored average life parameters and real-time state of charge parameters of lithium batteries, into the input layer of the BP neural network model. S2. The input layer data is processed by a nonlinear transformation through a hidden layer. The hidden layer uses a hyperbolic tangent transfer function. The output layer is processed by a linear weighting operation and activation operation on the output of the hidden layer. The output layer uses a sigmoid transfer function to generate the predicted values ​​of the remaining capacity and healthy life of the lithium battery. S3. Calculate the deviation between the predicted value and the measured value. When the deviation exceeds the preset threshold, start the error backpropagation mechanism to iteratively update the weight coefficients and bias terms of the neural network, and repeat the process of generating the predicted value and updating the parameters until the deviation is lower than the preset threshold, and output the final prediction result of the lithium battery health status.

[0004] In S1, the open-circuit voltage signal and internal resistance signal are preprocessed in real time by a dynamic noise reduction and feature enhancement processing module. The dynamic noise reduction and feature enhancement processing module adopts a cascaded architecture of adaptive wavelet transform and Kalman filter, and the processed signal is input to the input layer of the BP neural network model.

[0005] The output of the dynamic noise reduction and feature enhancement processing module is connected to the covariance analysis module. The covariance analysis module receives ambient temperature parameters and multi-scale discharge rate parameters, and dynamically calculates their correlation weights with open-circuit voltage signals and internal resistance signals, generating a fused feature vector that is input to the input layer of the BP neural network model.

[0006] The covariance analysis module extracts the time-series statistical features of ambient temperature parameters and multi-scale discharge rate parameters through a sliding time window, generates a dynamic covariance matrix that is strongly correlated with battery aging, and then weights and superimposes this matrix onto the original input feature vector.

[0007] In S2, the hidden layer is equipped with a real-time incremental weight adjustment unit based on reinforcement learning. When the lithium battery degradation mode changes abruptly, the weight adjustment unit dynamically fine-tunes the weight coefficients of the hidden layer to the output layer according to the historical prediction error.

[0008] The operation of the real-time incremental weight adjustment unit includes: monitoring the continuous deviation of the predicted value of the remaining capacity of the lithium battery online; when the rate of change of the deviation exceeds the preset mutation threshold, triggering the reinforcement learning network to generate a local weight correction amount and updating the key connection weight coefficients of the hidden layer.

[0009] In S3, the parameter update process of the error backpropagation mechanism is implemented through an embedded hardware acceleration interface. The embedded hardware acceleration interface integrates a low-precision fixed-point quantization algorithm to convert the floating-point operations of the weight coefficients and bias terms into fixed-point operations to adapt to the edge computing platform.

[0010] The execution flow of the low-precision fixed-point quantization algorithm in the embedded hardware acceleration interface is as follows: convert the floating-point values ​​of the weight coefficients and bias terms into integer format according to the dynamic scaling factor, maintain the consistency of the calculation precision during the backpropagation iteration process, and convert the integer format back to floating-point values ​​after the update is completed.

[0011] In S2, the output layer adopts a dual-channel independent activation structure. The first channel outputs the predicted value of the remaining capacity of the lithium battery, and the second channel outputs the predicted value of the healthy life. The weight coefficients and bias terms of each channel are optimized through a backpropagation mechanism.

[0012] In S3, the preset threshold is a dynamically adjusted threshold. The initial threshold is set based on the average lifespan parameter of the lithium battery, and it shrinks round by round according to the root mean square value of the historical deviation during the iterative update process until the lower limit of convergence accuracy is reached.

[0013] This invention effectively improves the signal-to-noise ratio of input data through a dynamic noise reduction and feature enhancement processing module, enhances feature representation capabilities by integrating multi-source environmental parameters through a covariance analysis module, adaptively corrects model parameters in response to degradation mutations through a real-time incremental weight adjustment unit, optimizes multi-task prediction through a dual-channel independent structure in the output layer, and adapts to edge computing platforms through an embedded hardware acceleration interface. This significantly improves the accuracy, real-time performance, and robustness of lithium battery health status prediction, enabling efficient battery monitoring and extending battery life. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating a lithium battery health status prediction method based on a BP neural network proposed in this invention. Detailed Implementation

[0015] To address the limitations of existing technologies in accurately predicting abrupt changes in battery degradation patterns in real time and adapting to edge computing environments, reference is made to... Figure 1 This invention provides a method for predicting the health status of lithium batteries based on a BP neural network, specifically including: S1. Real-time acquisition of open-circuit voltage and internal resistance signals of lithium batteries, and input of the acquired open-circuit voltage and internal resistance signals, as well as pre-stored average life parameters and real-time state of charge parameters of lithium batteries, into the input layer of the BP neural network model.

[0016] The open-circuit voltage signal and internal resistance signal are preprocessed in real time through a dynamic noise reduction and feature enhancement processing module. The dynamic noise reduction and feature enhancement processing module adopts a cascaded architecture of adaptive wavelet transform and Kalman filter, and the processed signal is input to the input layer of the BP neural network model.

[0017] Signal acquisition relies on a sensor array to synchronously capture battery operation data. An adaptive mechanism is used to dynamically adjust the denoising intensity to adapt to changes in environmental interference. The sensor output signal is first decomposed into multi-band sub-signals through an adaptive wavelet transform layer to remove high-frequency noise components. Subsequently, a Kalman filter layer estimates the signal state in real time. The signal trajectory is optimized through iterative prediction and correction steps, and finally a smooth feature vector is output. By addressing the error problem introduced by sensor noise, the signal-to-noise ratio of the input data is improved, thereby enhancing the convergence efficiency of the neural network.

[0018] In electric vehicle battery systems, this module can be integrated into the on-board controller, using an AD converter and FPGA chip to perform signal decomposition and filtering. The adaptive wavelet transform specifically employs the Daubechies wavelet basis to adaptively adjust the number of decomposition layers, optimizing denoising performance for different discharge conditions. During battery charge-discharge cycles, this filtering algorithm iteratively updates the mean and variance of the open-circuit voltage signal using state and observation equations.

[0019] Furthermore, the output of the dynamic noise reduction and feature enhancement processing module is connected to a covariance analysis module. The covariance analysis module receives ambient temperature parameters and multi-scale discharge rate parameters, and dynamically calculates their correlation weights with the open-circuit voltage signal and internal resistance signal, generating a fused feature vector which is then input to the input layer of the BP neural network model. Those skilled in the art will understand that the dynamic noise reduction and feature enhancement processing module, based on the principle of statistical correlation, fuses multi-source parameters to capture battery aging characteristics.

[0020] Ambient temperature parameters and multi-scale discharge rate parameters are acquired through the data acquisition unit. The covariance analysis module uses a sliding window to calculate the covariance matrix. The correlation weight calculation uses the least squares estimation method to determine the linear dependence between parameters. The fusion process weights and superimposes the original feature vectors with the covariance matrix to generate an enhanced input vector, which can solve the problem of single signal prediction bias, improve feature representation ability, and promote the accuracy of aging pattern recognition.

[0021] In industrial energy storage systems, this module can combine data from temperature and current sensors to calculate the covariance weight between temperature and voltage. Specifically, the ambient temperature parameter includes the battery surface temperature and the ambient temperature. The real-time temperature value of the battery's operating environment can be acquired through thermocouples or infrared sensors. The multi-scale discharge rate parameter is used to describe the battery's discharge behavior and can extract the discharge current and capacity change rate at different time granularities.

[0022] The covariance analysis module extracts the time-series statistical features of ambient temperature parameters and multi-scale discharge rate parameters through a sliding time window, generating a dynamic covariance matrix strongly correlated with battery aging. This matrix is ​​then weighted and superimposed onto the original input feature vector. The sliding time window principle is based on time series analysis, capturing short-term degradation trends through a fixed window size. Furthermore, the window size is preset to X-XX sampling periods. The module's internal algorithm calculates the mean, variance, and covariance coefficients of the parameters within the window. After the dynamic covariance matrix is ​​generated, the weighted superposition uses linear interpolation to adjust the weight factors, ensuring the feature vector is compatible with different operating conditions. This addresses the issue of time-series data fluctuations, enhancing the stability of the input layer data and improving the robustness of neural network inference.

[0023] S2. The input layer data is processed by a nonlinear transformation through a hidden layer using a hyperbolic tangent transfer function. The output layer is processed by a linear weighting operation and activation operation on the output of the hidden layer using a sigmoid transfer function to generate predicted values ​​for the remaining capacity and healthy lifespan of the lithium battery.

[0024] The principle behind the hidden layer performing nonlinear transformations on the input layer data lies in the hyperbolic tangent function smoothly mapping the input to a high-dimensional space, simulating the nonlinear relationship of battery degradation.

[0025] The input feature vector is first weighted and summed through a fully connected layer. The hyperbolic tangent activation function is used to calculate the output value of each neuron. The output layer uses linear weighting, and the weight coefficients are optimized through backpropagation. After activation by the sigmoid function, the output range is compressed to the interval [X, Y]. During the training phase, the input data is processed in batches, and the weights are iteratively fine-tuned to minimize the loss function, thereby solving the problem of insufficient fitting of the linear model. This is used to improve the prediction generalization ability and promote the accurate capture of complex aging patterns.

[0026] The hidden layer is equipped with a real-time incremental weight adjustment unit based on reinforcement learning. When the lithium battery degradation mode changes abruptly, the weight adjustment unit dynamically fine-tunes the weight coefficients from the hidden layer to the output layer based on historical prediction errors. Specifically, the real-time incremental weight adjustment unit based on reinforcement learning is an adaptive optimization mechanism. For example, in a drone battery system, the weight adjustment unit adjusts the weights online through the Q-learning algorithm to respond to sudden load changes. The weight adjustment unit can be based on a reinforcement learning framework and use reward signals to drive weight updates.

[0027] The weight adjustment unit continuously monitors the prediction error sequence, while historical error data is stored in a circular buffer. When a degradation mutation is detected, the unit triggers the policy network to generate an incremental update, which is directly applied to key weight connections. The fine-tuning process does not affect the overall network structure, but only locally corrects the weight values, thereby solving the model failure problem caused by sudden degradation events, so as to optimize the real-time performance of predictions and extend the battery life.

[0028] The operation of the real-time incremental weight adjustment unit specifically includes: online monitoring of the continuous deviation of the predicted value of the remaining capacity of the lithium battery; when the deviation change rate exceeds the preset mutation threshold, triggering the reinforcement learning network to generate a local weight correction amount and updating the key connection weight coefficients of the hidden layer. The reinforcement learning network specifically includes a decision model and uses a deep deterministic strategy gradient network to generate the weight correction amount in response to the deviation rate mutation event. The operation of the reinforcement learning network depends on the deviation change rate as a trigger signal, and the reinforcement learning network generates the optimal action through experience replay.

[0029] The continuous bias calculation uses the moving average method, and the mutation threshold is initialized to a multiple of the historical error standard deviation. After the policy network generates local corrections, the weight updates are only applied to the highly sensitive connections from the hidden layer to the output layer to avoid global retraining. The corrections of the reinforcement learning network can be achieved through gradient scaling to ensure computational efficiency, thereby solving the problem of weight rigidity, realizing online model adaptation, and improving the prediction response speed.

[0030] The output layer employs a dual-channel independent activation structure. The first channel outputs the predicted remaining capacity of the lithium battery, and the second channel outputs the predicted healthy lifespan. The weight coefficients and bias terms of each channel are optimized separately through a backpropagation mechanism. Specifically, this dual-channel independent activation structure is a parallel output architecture. The first channel uses a sigmoid function to output the percentage of remaining capacity, and the second channel uses linear activation to output the remaining number of iterations, thereby reducing task interference. Each channel has an independent fully connected layer. After initializing the weight coefficients and bias terms, the backpropagation mechanism calculates the gradient loss of each channel. The optimization process uses the Adam optimizer to independently update parameters, ensuring that the remaining capacity and healthy lifespan predictions are decoupled, resolving multi-task conflicts, improving prediction accuracy, and facilitating battery health management decisions.

[0031] S3. Calculate the deviation between the predicted value and the measured value. When the deviation exceeds the preset threshold, start the error backpropagation mechanism to iteratively update the weight coefficients and bias terms of the neural network, and repeat the process of generating the predicted value and updating the parameters until the deviation is lower than the preset threshold, and output the final prediction result of the lithium battery health status.

[0032] The principle of calculating the deviation between predicted and measured values ​​is based on the loss function (such as mean squared error) to quantify model error. The error backpropagation mechanism updates parameters through gradient descent. The measured values ​​are obtained through a battery tester. The deviation calculation adopts an online comparison module. When the deviation exceeds the standard, the backpropagation iteratively executes forward propagation to generate predicted values. The gradient calculation layer derives the partial derivatives of all weight coefficients and bias terms. The weight update step size is adjusted according to the learning rate. The loop is repeated until convergence, and the final prediction result is output. This solves the overfitting risk problem, improves the model's generalization ability, and promotes the reliability of long-term prediction.

[0033] The parameter update process of the error backpropagation mechanism is implemented through an embedded hardware acceleration interface. This interface integrates a low-precision fixed-point quantization algorithm, converting the floating-point operations of weight coefficients and bias terms into fixed-point operations to adapt to edge computing platforms. Specifically, the embedded hardware acceleration interface is a hardware optimization unit. In IoT gateway devices, this interface integrates an ARM Cortex-M series processor to accelerate backpropagation calculations. This interface can utilize fixed-point operations to reduce computational overhead, thereby solving the resource constraints of edge devices and reducing computational latency for real-time battery monitoring.

[0034] The low-precision fixed-point quantization algorithm specifically includes a numerical conversion method. This algorithm maps 32-bit floating-point weights to 8-bit integer format. The dynamic scaling factor can be dynamically adjusted based on the maximum absolute value. The execution flow of the low-precision fixed-point quantization algorithm in the embedded hardware acceleration interface is as follows: the floating-point values ​​of the weight coefficients and bias terms are converted to integer format according to the dynamic scaling factor. During the backpropagation iteration process, the consistency of the operation precision is maintained. After the update is completed, the integer format is converted back to the floating-point value.

[0035] Specifically, the dynamic scaling factor is a precision control parameter. It is dynamically calculated based on the weight distribution histogram to minimize conversion loss and optimize hardware efficiency through integer operations. During the conversion phase, a quantizer divides the floating-point number by the dynamic scaling factor for integer rounding. Backpropagation iterations use fixed-point multiply-accumulators to perform gradient calculations. Precision consistency is achieved through dynamic scaling factor compensation. In the rollback phase, the integer is multiplied by the dynamic scaling factor to restore the floating-point value, thus resolving the floating-point operation bottleneck problem, adapting to low-power platforms, and improving system deployment flexibility.

[0036] The preset threshold is a dynamically adjusted threshold. The initial threshold is set based on the average lifespan parameter of the lithium battery, and it shrinks round by round during the iterative update process according to the root mean square value of the historical deviation until the lower limit of convergence accuracy is reached. The dynamically adjusted threshold can be based on an adaptive learning strategy.

[0037] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting the health status of lithium batteries based on a BP neural network, characterized in that, include: S1. Real-time acquisition of open-circuit voltage and internal resistance signals of lithium batteries, and input of the acquired open-circuit voltage and internal resistance signals, as well as the pre-stored average life parameters and real-time state of charge parameters of lithium batteries, into the input layer of the BP neural network model. S2. The input layer data is processed by a nonlinear transformation through a hidden layer. The hidden layer uses a hyperbolic tangent transfer function. The output layer is processed by a linear weighting operation and activation operation on the output of the hidden layer. The output layer uses a sigmoid transfer function to generate the predicted values ​​of the remaining capacity and healthy life of the lithium battery. S3. Calculate the deviation between the predicted value and the measured value. When the deviation exceeds the preset threshold, start the error backpropagation mechanism to iteratively update the weight coefficients and bias terms of the neural network, and repeat the process of generating the predicted value and updating the parameters until the deviation is lower than the preset threshold, and output the final prediction result of the lithium battery health status.

2. The lithium battery health status prediction method as described in claim 1, characterized in that, In S1, the open-circuit voltage signal and internal resistance signal are preprocessed in real time by a dynamic noise reduction and feature enhancement processing module. The dynamic noise reduction and feature enhancement processing module adopts a cascaded architecture of adaptive wavelet transform and Kalman filter, and the processed signal is input to the input layer of the BP neural network model.

3. The lithium battery health status prediction method as described in claim 2, characterized in that, The output of the dynamic noise reduction and feature enhancement processing module is connected to the covariance analysis module. The covariance analysis module receives ambient temperature parameters and multi-scale discharge rate parameters, and dynamically calculates their correlation weights with open-circuit voltage signals and internal resistance signals, generating a fused feature vector that is input to the input layer of the BP neural network model.

4. The lithium battery health status prediction method as described in claim 3, characterized in that, The covariance analysis module extracts the time-series statistical features of ambient temperature parameters and multi-scale discharge rate parameters through a sliding time window, generates a dynamic covariance matrix that is strongly correlated with battery aging, and then weights and superimposes this matrix onto the original input feature vector.

5. The lithium battery health status prediction method as described in claim 1, characterized in that, In S2, the hidden layer is equipped with a real-time incremental weight adjustment unit based on reinforcement learning. When the lithium battery degradation mode changes abruptly, the weight adjustment unit dynamically fine-tunes the weight coefficients of the hidden layer to the output layer according to the historical prediction error.

6. The lithium battery health status prediction method as described in claim 5, characterized in that, The operation of the real-time incremental weight adjustment unit includes: monitoring the continuous deviation of the predicted value of the remaining capacity of the lithium battery online; when the rate of change of deviation exceeds the preset mutation threshold, triggering the reinforcement learning network to generate local weight correction amount and updating the key connection weight coefficients of the hidden layer.

7. The lithium battery health status prediction method as described in claim 1, characterized in that, In S3, the parameter update process of the error backpropagation mechanism is implemented through an embedded hardware acceleration interface. The embedded hardware acceleration interface integrates a low-precision fixed-point quantization algorithm to convert the floating-point operations of the weight coefficients and bias terms into fixed-point operations to adapt to the edge computing platform.

8. The lithium battery health status prediction method as described in claim 7, characterized in that, The execution flow of the low-precision fixed-point quantization algorithm in the embedded hardware acceleration interface is as follows: convert the floating-point values ​​of the weight coefficients and bias terms into integer format according to the dynamic scaling factor, maintain the consistency of the calculation precision during the backpropagation iteration process, and convert the integer format back to floating-point values ​​after the update is completed.

9. The lithium battery health status prediction method as described in claim 1, characterized in that, In S2, the output layer adopts a dual-channel independent activation structure. The first channel outputs the predicted value of the remaining capacity of the lithium battery, and the second channel outputs the predicted value of the healthy life. The weight coefficients and bias terms of each channel are optimized separately through the backpropagation mechanism.

10. The lithium battery health status prediction method as described in claim 1, characterized in that, In S3, the preset threshold is a dynamically adjusted threshold. The initial threshold is set based on the average lifespan parameter of the lithium battery, and it shrinks round by round according to the root mean square value of the historical deviation during the iterative update process until the lower limit of convergence accuracy is reached.

Images