Battery health state estimation system based on multi-scale feature fusion extraction

Abstract

The invention relates to the technical field of battery health management, and provides a battery health state estimation system based on multi-scale feature fusion extraction, which comprises a data acquisition module, a feature processing module, a deep learning model module and a result display module, the data acquisition module is used for acquiring battery data in real time; the feature processing module is used for extracting multi-scale features of the battery data; the deep learning module is used for constructing and training a deep learning model and outputting an actual SOH value; and the result display module is used for displaying the actual SOH value. The application is passed. The method has the advantages of high precision and strong robustness, and is suitable for monitoring and estimating the battery health state of an electric vehicle, an energy storage system and other scenes.

Description

Technical Field

[0001] This invention belongs to the field of battery health management technology, and in particular relates to a battery health status estimation system based on multi-scale feature fusion extraction. Background Technology

[0002] Existing real-time battery health monitoring (SOH) technology is mainly used in battery management systems (BMS) for electric vehicles and energy storage power stations, supporting key tasks such as battery safety early warning, energy optimization scheduling, and remaining life prediction.

[0003] In electric vehicles, accurate State of Harmony (SOH) estimation can avoid safety hazards caused by overcharging and over-discharging, and improve the accuracy of range estimation. In energy storage power station scenarios, it directly affects the battery pack's balanced management efficiency and system-level life cycle cost control.

[0004] However, existing SOH systems generally suffer from the following defects, including insufficient decoupling of multi-scale features, only processing single-cycle capacity decay or single-cycle voltage fluctuations, ignoring the synergistic effect of micro (single-cycle dynamics) and macro (cross-cycle trend) features; weak key feature recognition capability, and the inability of a single CNN / LSTM model to simultaneously capture spatial local features and temporal dependencies, resulting in large errors. Summary of the Invention

[0005] In view of the above-mentioned deficiencies of the prior art, this invention proposes a battery health state estimation system based on multi-scale feature fusion extraction to solve the problems of multi-scale feature decoupling and key feature identification. The technical solution designed in this invention includes: The module includes a data acquisition module, a feature processing module, a deep learning model module, and a results display module. The data acquisition module is used to collect battery data in real time; The feature processing module is used to extract multi-scale features from battery data; The deep learning module is used to build and train a deep learning model and output the actual SOH value; The results display module is used to display the actual SOH value.

[0006] Preferably, the data acquisition module further includes: The battery data is then cleaned and normalized.

[0007] Preferably, the multi-scale features extracted from the battery data include: Microscopic and macroscopic battery degradation characteristics were extracted from the battery data.

[0008] Preferably, the microcellular degradation characteristics include: The data source is the original sequence of voltage and / or current at the second level within a single charge-discharge cycle in battery data. The original sequence is decomposed into high-resolution segments by a sliding window to capture short-time dynamic features within the cycle as microscopic battery degradation features.

[0009] Preferably, the macroscopic battery degradation characteristics include: The data source is full life cycle data from battery data, and the overall aging degree is quantified by voltage difference as a macroscopic characteristic of battery degradation.

[0010] Preferably, constructing the deep learning model includes: The deep learning model architecture consists of convolutional neural networks, long short-term memory networks, and attention mechanisms.

[0011] Preferably, the deep learning training module includes: The multi-scale features of the input battery data are fed into the convolutional neural network to output a feature vector. The feature vector is then fed into a long short-term memory network to output a SOH prediction value. The SOH prediction value is then denormalized to convert it into the actual SOH value.

[0012] Beneficial effects: This application proposes a battery health state estimation system based on multi-scale feature fusion extraction. First, data is collected and preprocessed, then the collected data is used to train the network, and the health state of the lithium battery is tested using the trained model. Multi-timescale feature extraction can extract multi-dimensional aging features. Through refined sequence processing and convolutional attention mechanism, the model's ability to analyze complex aging patterns is improved. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of a preferred embodiment of the present invention; Figure 2 This is a schematic diagram of the operation flow of a preferred embodiment of the present invention; Figure 3 This is a schematic diagram of the battery charge-discharge cycle voltage curve of a preferred embodiment of the present invention; Figure 4 This is a schematic diagram of multi-timescale feature extraction according to a preferred embodiment of the present invention; Figure 5 This is a schematic diagram comparing the estimation results of various algorithms for 100%-80% SOH of a blue No. 2 battery according to a preferred embodiment of the present invention; Figure 6 This is a schematic diagram showing the error comparison of multiple models according to a preferred embodiment of the present invention. Detailed Implementation

[0014] The embodiments of the present invention will be described in detail below. The embodiments described below are implemented based on the technical solution of the present invention, and detailed implementation methods and specific operation processes are given. However, the protection scope of the present invention is not limited to the embodiments described below.

[0015] This invention designs a battery health state estimation system based on multi-scale feature fusion extraction, such as... Figure 1-2 As shown, the technical solution specifically includes: The module includes a data acquisition module, a feature processing module, a deep learning model module, and a results display module. The data acquisition module is used to collect battery data in real time; The feature processing module is used to extract multi-scale features from battery data; The deep learning module is used to build and train deep learning models and output actual SOH values; The results display module is used to show the actual SOH value.

[0016] Specifically, the experimental subjects were six square lithium iron phosphate batteries with a nominal capacity of 100Ah, including three from BYD and three from EVE Energy. The accelerated aging conditions were an ambient temperature of 50℃, a 1C charge / discharge current of 100A, a charging cutoff voltage of 3.65V, and a discharging cutoff voltage of 2.6V. After each cycle, the batteries were left to stand for 30 minutes, and the voltage, current, number of cycles, and periodically measured capacity data were recorded as real-time battery data.

[0017] In addition, the experimental equipment and materials include: Test platform: Programmable battery tester (supports 1C constant current charge and discharge). Environmental chamber: Temperature controlled at 50℃±2℃ to simulate high-temperature accelerated aging conditions.

[0018] The loop steps include: Discharge phase: Discharge at a constant current of 100A until the voltage reaches 2.6V, and record the discharge time and voltage curve; Settling phase: After the discharge is complete, let it stand for 30 minutes until the voltage stabilizes; Charging phase: Charge at a constant current of 100A until the voltage reaches 3.65V, and record the charging time and voltage curve; Data recording: Voltage and current data are collected every 1 second, and the actual capacity is measured and SOH is calculated every 50 cycles.

[0019] like Figure 3 The image shows the charge-discharge cycle voltage curve of the blue No. 1 battery. Different colored curves represent different numbers of cycles (blue for the initial cycle and red for later cycles), showing that as the number of cycles increases, the time it takes for the battery to reach the cutoff voltage shortens, visually reflecting the decreasing trend of SOH (State of Harm).

[0020] Preferably, the data acquisition module further includes: Perform data cleaning and normalization on the battery data.

[0021] Specifically, data cleaning includes removing abnormal data points with voltages outside the normal range (2.6V-3.8V); filling in a small number of missing data points using linear interpolation; and normalization processing includes standardizing voltage and current data to the [0,1] interval to accelerate model convergence.

[0022] Preferably, the multi-scale features of the battery data are extracted, including: Microscopic and macroscopic battery degradation characteristics were extracted from the battery data.

[0023] Preferably, the micro-cell degradation characteristics include: The data source is the original sequence of voltage and / or current at the second level within a single charge-discharge cycle in battery data. The original sequence is decomposed into high-resolution segments by a sliding window to capture short-time dynamic features within the cycle as microscopic battery degradation features.

[0024] Preferably, the macroscopic battery degradation characteristics include: The data source is full life cycle data from battery data, and the overall aging degree is quantified by voltage difference as a macroscopic characteristic of battery degradation.

[0025] Specifically, battery degradation is a multi-factor coupled process, influenced by both short-term operating conditions (such as charging rate and temperature) and long-term aging trends (such as capacity decay and increased internal resistance). Analysis at a single timescale often fails to fully capture these complex dynamics. For example, focusing only on short-term characteristics within a single cycle may overlook long-term capacity decay trends, while analyzing only long-term characteristics across cycles may miss the short-term impact of operating conditions on battery performance. Therefore, it is necessary to combine micro and macro scales to improve prediction accuracy. Figure 4 As shown, battery degradation characteristics are divided into two time scales: micro and macro. A hierarchical feature system is constructed. The data source for the micro scale (single cycle dynamics) is a 1-second sampling frequency, which captures short-term dynamic features such as polarization voltage fluctuations and lithium-ion diffusion kinetics within the cycle. The data source for the macro scale (inter-cycle dynamics) is hundreds to thousands of cycles, and the voltage difference is the voltage difference between the current cycle and the first cycle.

[0026] Equation 1 shows the matrix for extracting the cycle characteristic parameters of a single battery. The elements within the matrix can be changed to adapt to different operating conditions. This represents the i-th voltage value measured within that cycle.

[0027] (1) This represents the voltage difference between the current cycle and the first cycle, as shown in Formula 2: (2) This represents the nth current value measured within the first cycle. Similar to the voltage value, it reflects the magnitude of the current in the circuit within that cycle and can be used to analyze the circuit's operating state in different cycles.

[0028] This represents the difference in current between the current in this cycle and the first cycle, which helps to understand the trend of current variation across different cycles. As shown in Formula 3: (3) This indicates the temperature corresponding to that cycle itself.

[0029] This represents the difference between the temperature of this cycle and the temperature of the first cycle. As shown in Formula 4: (4) This represents the estimated battery health status value obtained within that cycle.

[0030] In addition, for microscopic battery degradation characteristics, single-cycle voltage and current sequences are recorded to generate short-time sequences; for macroscopic battery degradation characteristics, multi-cycle feature vector sequences are used (each feature vector contains voltage difference, current difference, and sequence length N=10), and finally, medium- and long-term sequences are generated.

[0031] Preferably, constructing a deep learning model includes: The deep learning model architecture consists of convolutional neural networks, long short-term memory networks, and attention mechanisms.

[0032] Preferably, the deep learning module includes: The multi-scale features of the input battery data are fed into the convolutional neural network to output a feature vector. The feature vector is then fed into a long short-term memory network to output a SOH prediction value. The SOH prediction value is then denormalized to convert it into the actual SOH value.

[0033] Specifically, Convolutional Neural Networks (CNN), Long Short-Term Memory Networks (LSTM), and attention mechanisms are the three core modules of deep learning, including: Convolutional Neural Network (CNN): Input is high-resolution charge / discharge curve data (voltage, current, cycle count). Local degradation features are extracted through two convolutional layers (kernel size 3, channel numbers 16 and 32 respectively) and pooling layers (max pooling, window size 2, stride 2). Convolutional operations can identify local patterns in the charging curve (such as the smoothness of voltage plateaus), while pooling operations reduce data dimensionality and suppress noise. The output is a low-dimensional feature vector representing short-term dynamic degradation patterns within a single cycle.

[0034] Long Short-Term Memory (LSTM) network: Input features extracted by CNN and historical cycle data. LSTM (64 hidden units) models long-term dependencies across multiple cycles. LSTM's forget gate, input gate, and output gate mechanisms effectively filter noise and retain key information related to degradation, making it particularly suitable for capturing the slow decline trend of battery capacity. Output is a feature vector encoding the long-term degradation trend.

[0035] The forget gate determines how much information from the previous time step's cell state is retained in the current time step. It calculates the forgetting factor using the sigmoid activation function. (5) in This is the hidden state from the previous moment. This is the current input.

[0036] The forget gate prevents irrelevant information from interfering with the current prediction by "forgetting" it. In SOH estimation, the forget gate can filter out short-term noise unrelated to battery aging.

[0037] The input gate controls how much of the current input information is updated in the cell state. Its calculation method is as follows: (6) And combined with candidate status (7) The input gate selectively introduces new information to ensure that features related to SOH (such as capacity decay trends) are stored.

[0038] The output gate determines how much information from the current cell state is output as a hidden state. Its calculation method is as follows: (8) Hidden state (9) The output gate converts the current memory into network output for use by subsequent layers.

[0039] LSTM cell states The calculation method is as follows: Through joint updates using the forget gate and the input gate, the computation is... (10) This mechanism enables LSTM to retain important information over a long period of time while dynamically adjusting for the effects of short-term inputs.

[0040] Attention Mechanism: Input the output features of CNN and LSTM. A self-attention mechanism dynamically adjusts the weights of different time steps or features, highlighting the most important degradation points for SOH estimation (such as the rapid capacity decay stage). The attention mechanism uses query (Q), key (K), and value (V) vectors, calculates an attention score using a softmax function, and generates a context vector. Core formula of the attention mechanism: (11) Output: The weighted feature vector, used for final SOH prediction.

[0041] Fully connected layer: Maps the output of the attention mechanism to the SOH value (scalar), and ensures the stability of the nonlinear mapping through an activation function (such as ReLU).

[0042] In addition, the model training configuration includes: Loss function: Root Mean Square Error (RMSE), the formula is: (12) Optimizer: Adam optimizer, initial learning rate 0.001, decays by 0.1 every 20 epochs (decay factor 0.1). Hyperparameter tuning: Determining the optimal parameter combination through grid search. Kernel size: 3 (balancing feature extraction capability and computational complexity); LSTM unit count: 64 (balancing long-term dependency capture and overfitting risk). Batch size: 32 (mini-batch gradient descent accelerates convergence).

[0043] Training termination strategy: Early stopping method. If the validation set loss does not decrease for 10 consecutive epochs, training is stopped to prevent overfitting.

[0044] In addition, the data input to the deep learning module is divided into the following categories: Training set: 70% recurring data; Validation set: 30% circular data.

[0045] In addition, this application also provides performance verification and comparison, including: Evaluation indicators: Root Mean Square Error (RMSE): Measures the square root of the mean square of the deviations between the predicted and actual values. The formula is:

[0046] Mean Absolute Error (MAE): Reflects the average absolute deviation between the predicted and actual values. The formula is: .

[0047] like Figure 5 As shown, the hybrid model can predict and track the true SOH value very well. Figure 6 As shown, the hybrid model significantly outperforms CNN and LSTM in both RMSE and MAE across all data points, especially in MAE (e.g., MAE of BLUE_3 is only 0.0040, and MAE of BLACK_2 is 0.0048), demonstrating the best performance. The hybrid model performs optimally in the prediction task, significantly outperforming CNN and LSTM. It has an advantage in reducing prediction error, making it particularly suitable for the data scenario in this experiment. CNN and LSTM exhibit higher errors on some data points, requiring further optimization or model tuning for specific data.

[0048] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A battery health state estimation system based on multi-scale feature fusion extraction, characterized in that, include: The module includes a data acquisition module, a feature processing module, a deep learning model module, and a results display module. The data acquisition module is used to collect battery data in real time; The feature processing module is used to extract multi-scale features from battery data; The deep learning module is used to build and train a deep learning model and output the actual SOH value; The results display module is used to display the actual SOH value.

2. The battery health state estimation system based on multi-scale feature fusion extraction according to claim 1, characterized in that, The data acquisition module also includes: The battery data is then cleaned and normalized.

3. The battery health state estimation system based on multi-scale feature fusion extraction according to claim 1, characterized in that, The multi-scale features extracted from the battery data include: Microscopic and macroscopic battery degradation characteristics were extracted from the battery data.

4. The battery health state estimation system based on multi-scale feature fusion extraction according to claim 3, characterized in that, The microscopic battery degradation characteristics include: The data source is the original sequence of voltage and / or current at the second level within a single charge-discharge cycle in battery data. The original sequence is decomposed into high-resolution segments by a sliding window to capture short-time dynamic features within the cycle as microscopic battery degradation features.

5. The battery health state estimation system based on multi-scale feature fusion extraction according to claim 3, characterized in that, The macroscopic battery degradation characteristics include: The data source is full life cycle data from battery data, and the overall aging degree is quantified by voltage difference as a macroscopic characteristic of battery degradation.

6. The battery health state estimation system based on multi-scale feature fusion extraction according to claim 1, characterized in that, The construction of the deep learning model includes: The deep learning model architecture consists of convolutional neural networks, long short-term memory networks, and attention mechanisms.

7. The battery health state estimation system based on multi-scale feature fusion extraction according to claim 1, characterized in that, The deep learning training module includes: The multi-scale features of the input battery data are fed into the convolutional neural network to output a feature vector. The feature vector is then fed into a long short-term memory network to output a SOH prediction value. The SOH prediction value is then denormalized to convert it into the actual SOH value.

Images