Battery state of health estimation method and system based on feature fusion convolutional neural network
By employing a convolutional neural network structure with feature fusion in the estimation of the health status of lithium-ion batteries, the problems of lack of intrinsic connections between features and insufficient data preprocessing are solved, and higher estimation accuracy is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SUN YAT SEN UNIV
- Filing Date
- 2023-06-27
- Publication Date
- 2026-08-04
AI Technical Summary
Existing methods for estimating the health status of lithium-ion batteries lack further design of the intrinsic relationships between features and suffer from insufficient data preprocessing, resulting in limited estimation accuracy.
A convolutional neural network structure based on feature fusion is adopted. By expanding the original voltage data with various features and designing a feature fusion module in the model, the learning ability of the intrinsic relationship between features is improved.
It improves the accuracy of lithium-ion battery health status estimation, with mean absolute error and mean absolute percentage error not exceeding 0.0028 and 0.32%, respectively.
Smart Images
Figure CN116774089B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence technology applications, and in particular to a method and system for estimating battery health status based on feature fusion using a convolutional neural network. Background Technology
[0002] Lithium-ion batteries possess advantages such as high energy density, high power density, long cycle life, and low self-discharge, leading to their rapid development and widespread application in numerous fields, including electric vehicles, mobile phones, and other mobile electronic devices. Lithium-ion batteries typically experience performance degradation during aging, primarily manifested as a reduction in maximum usable capacity and power decay. Generally, a battery is considered to have reached the end of its life when its current capacity drops below 80% of its initial capacity. In particular, the usable capacity decreases rapidly as the battery nears the end of its life.
[0003] To ensure the overall safety and reliability of the battery system, the battery management system (BMS) needs to accurately and reliably monitor the battery's health status. Health status is typically defined in a BMS as an indicator of the battery's health condition, calculated as the ratio of the battery's current maximum usable capacity to its initial rated capacity. Accurate battery health status estimation is crucial for monitoring the battery's current state, ensuring safe operation and timely replacement before the battery reaches the end of its lifespan.
[0004] A common approach to estimating battery health status is to collect voltage, current, and temperature data generated during battery operation to estimate the battery health status factor. This is because it is difficult to directly obtain the battery health status estimation factor in real-world environments; it requires recording data of the battery from a fully charged state to a fully discharged state at a fixed discharge rate and then calculating the battery's current maximum capacity. However, in practical engineering applications, the complex scenarios in which lithium-ion batteries are used, the uncertainties inherent in data acquisition, and the complex nonlinear aging characteristics of the batteries themselves all pose significant challenges to the accurate estimation of lithium-ion battery health status.
[0005] Current methods for estimating battery health status are mainly divided into model-based methods and data-driven methods. Model-based methods can be further divided into those based on equivalent circuit models, electrochemical models, and empirical degradation models. In equivalent circuit models, the model typically consists of voltage sources, resistors, capacitors, and other components to simulate the battery's state, such as RC equivalent circuit models, fractional-order equivalent circuit models, and impedance spectrum growth models, and is combined with filtering algorithms to estimate the battery's health status. Equivalent circuit models are generally simple in principle and computationally inexpensive, but often have limited estimation accuracy. Electrochemical model-based methods typically describe the internal physical and chemical mechanisms of battery degradation through a series of partial differential equations, such as pseudo-two-dimensional models (P2D) and their simplified versions; however, this method generally faces difficulties in parameter extraction and complex equation calculations. Empirical degradation models, on the other hand, establish empirical expressions for battery aging behavior by analyzing historical capacity decay data, which can be used to describe the overall capacity degradation trend of the battery. Although parameter identification is simple in empirical degradation models, they struggle to explain battery capacity regeneration phenomena, i.e., local fluctuations in battery capacity. In summary, because different chemical reactions occur inside a battery under varying operating conditions, model-based methods struggle to create accurate battery models, and simply estimating model parameters leads to limited accuracy in estimating battery health. (This entire section could be further optimized.)
[0006] In recent years, scholars both domestically and internationally have conducted a series of studies on the health status estimation of lithium-ion batteries based on data-driven methods. These methods are built upon a large amount of battery operating data and do not require prior knowledge of the battery's internal chemical mechanisms for health status estimation. These methods can be divided into two categories: traditional machine learning models and deep learning models. In traditional machine learning models, such as support vector machines, Gaussian process regression, and random forests, features related to battery health status typically need to be designed manually. For example, features related to battery health status can be designed based on voltage data, including terminal voltage values, fixed voltage interval charge / discharge times, and the slope / area / radius of curvature of the voltage curve. Traditional machine learning methods are usually simple and interpretable; however, in manually designed feature extraction, only information from a single data point is often utilized, losing feature information from continuous sequence data, thus affecting the effectiveness of battery health status estimation. In contrast, deep learning models typically use convolutional neural networks, long short-term memory networks, etc., to automatically extract features from massive amounts of data and complete model training in an end-to-end manner, thereby solving the problem of insufficient data utilization in manual feature extraction.
[0007] Convolutional neural networks (CNNs) possess characteristics such as sparse connections, weight sharing, and pooling operations, which help reduce the number of model parameters, model complexity, and alleviate problems such as overfitting, thus resulting in better model generalization ability. The basic structure of a CNN includes convolutional layers, pooling layers, and fully connected layers. CNNs with different topologies composed of different network layers exhibit varying performance.
[0008] In the field of lithium-ion battery health state estimation, the application of convolutional neural networks mainly faces the following problems:
[0009] (1) Most existing models lack the ability to learn the intrinsic relationships between features when multiple features are used as inputs. This manifests as follows: all features are directly combined into a matrix as input to the convolutional neural network structure, and then all features are processed using the same structure. While the weight-sharing characteristic of convolutional neural networks allows the model to learn the relationships between features, this approach is too implicit and still has room for further optimization.
[0010] (2) Most research studies often use neural networks that lack sufficient preprocessing of the input feature data, resulting in insufficient learning performance of the network structure on the original data input features. Although deep learning models often adopt an end-to-end model structure, they are limited by the learning capabilities of existing model structures. Improving the quality of the learning samples can effectively improve the model's learning performance on data features. Summary of the Invention
[0011] To address the problems existing in the prior art, this invention provides a method and system for estimating the state of health of lithium-ion batteries based on feature fusion convolutional neural networks. By expanding multiple voltage features of the original voltage data and designing a new convolutional neural network structure with a feature fusion layer, the accuracy of estimating the state of health of lithium-ion batteries is improved.
[0012] The technical solution adopted in the estimation method of this invention is: a battery health status estimation method based on feature fusion using a convolutional neural network, comprising the following steps:
[0013] Acquire sampling data of the battery whose health status needs to be estimated;
[0014] The sampled data is subjected to one-dimensional feature extraction processing to obtain multiple one-dimensional features after data preprocessing;
[0015] A health status estimation model is constructed based on a convolutional neural network. The health status estimation model includes a feature extraction module, a feature fusion module, and a regression estimation module. The feature fusion module performs feature fusion on multiple feature maps generated by the feature extraction module by adding them bit by bit. The regression estimation module is used to extract the information after feature fusion.
[0016] The one-dimensional feature is input into the health status estimation model to estimate the battery health status, and the health status estimation result is obtained.
[0017] Preferably, the sampled data undergoes one-dimensional feature extraction processing to obtain multiple preprocessed one-dimensional features, including:
[0018] Feature extraction processing is performed on the sampled data to obtain discharge capacity data and voltage data within multiple cycles;
[0019] The first function curve is constructed based on the discharge capacity data and voltage data;
[0020] Differentiate the voltage data based on the discharge capacity data to obtain the rate of change of the discharge capacity data, which serves as the second function curve.
[0021] The third function curve is obtained by comparing the discharge capacity data in each cycle with the discharge capacity data in the first cycle.
[0022] Preferably, the feature extraction module includes three parallel sub-model structures, which are used to extract features from the one-dimensional features of each input separately to obtain the corresponding feature maps; in the feature fusion module, the feature maps from the three sub-model structures are added bit by bit to achieve feature fusion between the inputs.
[0023] The estimation system of the present invention adopts the following technical solution: a convolutional neural network-based battery health status estimation system based on feature fusion, comprising the following modules:
[0024] The data sampling module is used to acquire sampling data of the battery whose health status needs to be estimated.
[0025] The preprocessing module is used to perform one-dimensional feature extraction on the sampled data to obtain multiple preprocessed one-dimensional features.
[0026] The model building module is used to build a health status estimation model based on a convolutional neural network. The health status estimation model includes a feature extraction module, a feature fusion module, and a regression estimation module. The feature fusion module performs feature fusion on multiple feature maps generated by the feature extraction module by adding them bit by bit. The regression estimation module is used to extract the information after feature fusion.
[0027] The state estimation module is used to input the one-dimensional features into the health state estimation model to estimate the battery health state and obtain the health state estimation result.
[0028] The present invention has the following advantages and effects compared with the prior art:
[0029] 1. First, based on the collected voltage data generated by the battery during operation, various voltage feature extensions are performed as data preprocessing to make the original voltage data more effectively utilized. Then, by incorporating a feature fusion module into the convolutional neural network model, the inherent relationships between various features can be better learned. Through multiple voltage feature extensions of the original voltage data and a newly designed convolutional neural network structure with a feature fusion layer, the accuracy of lithium-ion battery health state estimation is improved.
[0030] 2. On the one hand, placing the feature fusion module at the beginning of the model might result in insufficient extraction of features from each input during feature fusion; on the other hand, placing the feature fusion module at the end of the model, in deeper network structures, where the learned information is often more abstract semantic information, would pose a greater challenge to feature fusion. Therefore, this invention places the feature fusion module in a more central position within the entire model.
[0031] 3. By using multiple preprocessed one-dimensional voltage features, the information input to the convolutional neural network model is richer than that input using only the original voltage features, which helps the model learn more information. Attached Figure Description
[0032] Figure 1 This is a flowchart of the battery health state estimation method based on feature fusion provided in the embodiments of this application;
[0033] Figure 2 This is a network architecture diagram of the health status estimation model provided in the embodiments of this application;
[0034] Figure 3 These are the specific parameter settings for the network architecture diagram of the health status estimation model provided in the embodiments of this application;
[0035] Figure 4 This is a comparison chart of the model health status estimate and the actual value provided in the embodiments of this application. Detailed Implementation
[0036] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.
[0037] Example 1
[0038] Reference Figure 1 This embodiment provides a battery health status estimation method based on feature fusion using a convolutional neural network, including the following steps:
[0039] S101. Obtain sampling data of the battery whose health status needs to be estimated.
[0040] In this embodiment, this step involves data acquisition and processing during the discharge process of the battery to be evaluated for its health status, to obtain discharge data; the discharge data is then sampled for discharge capacity and voltage data to obtain sampled data.
[0041] In this embodiment, a deep learning model is built using observational data generated by the battery during operation to directly and quickly estimate the battery's health status. The data generated during battery operation may include discharge data such as current, voltage, capacity, internal resistance, and temperature detected by a battery capacity tester. Sampling data is then obtained by sampling discharge capacity and voltage data from the discharge data.
[0042] It should be noted that in actual operation, it is difficult to obtain all the information about the battery from a fully charged state to a fully discharged state. This is because, in real-world applications, batteries rarely reach a fully discharged state. In this case, a method of setting a fixed voltage range can be used to collect battery operation data uniformly.
[0043] S102. Perform one-dimensional feature extraction processing on the sampled data to obtain multiple preprocessed one-dimensional features. In this embodiment, the specific process includes the following:
[0044] Feature extraction processing is performed on the sampled data to obtain discharge capacity data and voltage data within multiple cycles;
[0045] The first function curve is constructed based on the discharge capacity data and voltage data;
[0046] Differentiate the voltage data based on the discharge capacity data to obtain the rate of change of the discharge capacity data, which serves as the second function curve.
[0047] The third function curve is obtained by comparing the discharge capacity data in each cycle with the discharge capacity data in the first cycle.
[0048] In this embodiment, the discharge capacity and voltage data of the battery within one cycle are extracted from the sampling data of the battery to be estimated for its health status, and are respectively represented by Q. dLet Q1(v) represent the discharge capacity-discharge voltage curve over one cycle of the battery in its state of health estimation. The discharge capacity is considered a function of the discharge voltage, with the discharge voltage as the horizontal axis. For ease of description, Q1(v) represents the time-series data of the discharge capacity-discharge voltage curve over one cycle, yielding the first function curve. Next, the first derivative of the discharge capacity with respect to voltage over one discharge cycle is calculated to obtain the discharge capacity Q. d The rate of change of v, i.e., the second function curve, is denoted as Q2(v). The formula for calculating the second function curve is shown below:
[0049]
[0050] In the formula, j represents the j-th sampling point, Q d (j) represents the discharge capacity data at the j-th sampling point, and v(j) represents the voltage data at the j-th sampling point.
[0051] Furthermore, considering the changes in discharge capacity and voltage data with the cycle period, the discharge capacity data is subtracted. The time-series discharge capacity data for each cycle within the predicted battery life is subtracted from the data for the first cycle to obtain the third function curve, denoted as Q3(v). The formula for calculating the third function curve is shown below:
[0052] Q3 = Q di -Q d1
[0053] Among them, Q di and Q d1 These represent the discharge capacity curve data for the i-th and 1-th cycles, respectively.
[0054] Since the amount of data collected from the battery in actual sampling may vary in different cycle periods, data generated during battery operation within a fixed voltage range will be uniformly used for subsequent model training. Simultaneously, 100 equidistant voltage sampling points are set between the upper and lower limits of the battery discharge voltage, and the corresponding discharge capacity data for each voltage sampling point is obtained through interpolation fitting. Ultimately, the three voltage function curves Q1(v), Q2(v), and Q3(v) are three 1*100 one-dimensional data sets.
[0055] S103. Construct a health status estimation model based on a convolutional neural network. The health status estimation model includes a feature extraction module, a feature fusion module, and a regression estimation module. The convolutional neural network adopts a self-built network architecture. This step includes the following processes:
[0056] A feature extraction module is constructed, which includes multiple convolutional layers and pooling layers;
[0057] A feature fusion module is constructed, which uses a point-by-point addition method to fuse multiple feature maps generated by the feature extraction module.
[0058] A regression estimation module is constructed, which includes multiple fully connected layers for extracting information after feature fusion;
[0059] A battery health status estimation model is constructed based on the feature extraction module, feature fusion module, and regression estimation module.
[0060] In this embodiment of the invention, the model is based on a convolutional neural network (CNN) structure for feature extraction. A CNN is a feedforward multi-layer network that exhibits excellent performance in extracting features from high-dimensional data. Its typical features include sparse connections, weight sharing, and pooling operations, which allow the CNN structure to significantly reduce the number of training parameters, model complexity, and the risk of overfitting, thereby achieving better generalization ability. The basic structure of a CNN includes convolutional layers, pooling layers, and fully connected layers.
[0061] In this embodiment, battery health status is estimated by independently constructing a battery health status estimation model based on a convolutional neural network. The model structure is referenced from [reference needed]. Figure 2 The corresponding network structure parameters are referenced. Figure 3 The model takes three 1*100 one-dimensional data points as input, which are then normalized using z-scores before being fed into the model. The final output is an estimate of the battery's health status.
[0062] This embodiment builds a battery health status estimation model based on a convolutional neural network. First, a feature extraction module is constructed, which comprises the first six layers of the model. Specific parameter settings are described in [reference needed]. Figure 3 The L1-L6 feature extraction module comprises three parallel sub-model structures, each extracting a separate feature map for each input's one-dimensional feature. This mitigates the problem of insufficient attention to certain features when all features are processed together. Each sub-model structure has six layers, including multiple convolutional layers and a max-pooling layer following each convolutional layer. This structure effectively extracts low-level features. Furthermore, each convolutional layer in the sub-model structure uses a 1x3 kernel with a stride of 1. This design of multiple small convolutional kernels is believed to increase network depth without increasing the number of parameters, resulting in better model performance. The max-pooling layer in the sub-model employs overlapping pooling to improve generalization ability; its window size is 1x3 with a stride of 2.
[0063] In the feature fusion module of this embodiment, the feature maps from the previous three sub-model structures will be... Figure 3The L7 layer is subjected to a point-by-point addition operation to achieve feature fusion between inputs. Point-by-point addition serves as a connection method, linking different layers as fused features. In the feature fusion module of this embodiment, feature maps from Q1(v), Q2(v), and Q3(v) are added, allowing information from multiple features to interact and thus improving model performance. Figure 3 In L6, Q1(v), Q2(v), and Q3(v) each generate their own corresponding feature map X. i Y i and Z i These feature maps are all a×b in size, which is the basis for their point-by-point addition; through L7 feature fusion, new feature maps R will be generated corresponding to the features of Q1(v), Q2(v), and Q3(v). i =X i +Y i +Z i As a form of information interaction and supplement, it enriches the information representation capability of the feature maps because the new feature maps are directly related to the feature maps extracted from Q1(v), Q2(v), and Q3(v). This direct and simple approach enables information interaction among multiple features, improving network performance. Without... Figure 3 The L7's bitwise addition operation only allows information exchange between layers when the convolutional kernels process information from all preceding feature maps. In the subsequent two convolutional layers, information is exchanged as described above. Figure 3 L8-L9 will be used to further extract information from the fusion of the three input features.
[0064] In this embodiment, a zero-padding strategy is applied to all convolutional and pooling layers to counteract the effect of size shrinkage during feature map computation. After each convolutional layer, a rectified linear unit (RELU) is introduced as an activation function, whose nonlinear characteristics can be used to saturate the output or limit the generated output.
[0065] This embodiment uses a convolutional neural network to build a battery health status estimation model. Specific parameter settings are detailed in [reference needed]. Figure 3First, a feature extraction module is constructed, consisting of three identical sub-models with a total of 6 layers. The first convolutional layer has 6 kernels (1x3) with a stride of 1, followed by a ReLU activation function, resulting in a feature map of size 1x100x6. Next, a second convolutional layer with 6 kernels (1x3) and a stride of 1, followed by another ReLU activation function, also yields a feature map of size 1x100x6. Then, a first max pooling layer with a 1x3 pooling window and a stride of 2 outputs a feature map of size 1x50x6. Finally, a third convolutional layer with 16 kernels (1x3) and a stride of 1, followed by another ReLU activation function, results in a feature map of size 1x50x16. Next, the fourth convolutional layer has 16 kernels (1x3 size) and a stride of 1, followed by a ReLU activation function, resulting in a feature map of size 1x50x16. Then, a second max-pooling layer with a pooling window size of 1x3 and a stride of 2 outputs a feature map of size 1x25x6. Next, a feature fusion module is built, which first adds the feature maps generated by the previous three sub-modules of the same size, one by one, at corresponding positions, resulting in a new feature map of size 1x25x6. Next, the fifth convolutional layer has 128 kernels (1x3 size) and a stride of 1, followed by a ReLU activation function, resulting in a feature map of size 1x25x128. Finally, a sixth convolutional layer has 128 kernels (1x3 size) and a stride of 1, followed by a ReLU activation function, resulting in a feature map of size 1x25x128. Finally, a regression estimation module is constructed, consisting of a first fully connected layer and a second fully connected layer. The feature map obtained after the sixth convolutional layer is first fed into the first fully connected layer, which contains 10 neurons. Next, it is fed into the second fully connected layer, which has one neuron. The regression estimation module extracts information after feature fusion. Through the fully connected layers, it transforms the two-dimensional feature map output by the convolution into a one-dimensional vector, which is then used to output the estimated health state value. Each node in the fully connected layer is connected to all nodes in the previous layer, effectively combining all the features from the previous layer and thus efficiently extracting the information after feature fusion.
[0066] S104. Input the one-dimensional feature into the health status estimation model to estimate the battery health status and obtain the health status estimation result.
[0067] Preferably, prior to the battery health state estimation in this step, the health state estimation method of this embodiment further includes pre-training a battery health state estimation model, specifically including:
[0068] The model parameters of the battery health state estimation model are tuned using an optimization algorithm to obtain a trained battery health state estimation model.
[0069] In this embodiment, before the battery health state estimation model estimates the battery's health state, it needs to be trained and tested. The Adam algorithm is chosen as the optimization algorithm for training the health state estimation model. Compared to stochastic gradient descent, the Adam algorithm adds first and second moments and sets specific adaptive learning rates for different parameters. The initial learning rate is set to 0.001, and after 100 training iterations, the learning rate is set to 0.00001, with a total of 200 training iterations. Experiments were then conducted on a battery degradation dataset to test the model. The experiments show that the battery health state estimation model has reliable estimation accuracy in battery health state estimation, as shown in the estimation results. Figure 4 As shown. After training and testing the battery health state estimation model as described above, the battery health state can be directly estimated using the data generated during the battery's current operation.
[0070] Combined with appendix Figure 1 The estimation process of this invention specifically includes: performing feature expansion on the sampled data of the battery whose health status needs to be estimated, so that the single feature of the original data becomes three features containing more information, which are then input into the battery health status estimation model built on a convolutional neural network to obtain a reliable battery health status estimation result.
[0071] In this embodiment of the invention, only less than 40% of the complete discharge data in the current battery cycle needs to be collected to obtain excellent battery health state estimation results. In experimental data, the average absolute error and average absolute percentage error do not exceed 0.0028 and 0.32%, respectively, improving the accuracy of the battery health state estimation results.
[0072] In summary, the embodiments of the present invention have the following advantages:
[0073] This invention applies the high-performance convolutional neural network (CNN) model, which excels in computer vision, natural language processing, and semantic segmentation, to battery health state estimation. To fully utilize the model's learning efficiency when multiple features are used as input, a feature fusion module is specifically designed to facilitate further communication between feature information. Simultaneously, to better utilize the collected raw discharge voltage data, it is expanded into three voltage features, enriching the model's input. The data preprocessing operations and methods for further learning the internal relationships among multiple input features emphasized in this invention further improve the accuracy of battery health state estimation. To evaluate its technical effectiveness, this invention conducts experiments on a battery degradation dataset using the proposed evaluation method. This method requires less than 40% of the complete discharge data collected during the current battery cycle, with the average absolute error not exceeding 0.0028 and the average absolute percentage error not exceeding 0.32%.
[0074] Example 2
[0075] Based on the same inventive concept as Embodiment 1, this embodiment provides a convolutional neural network-based battery health state estimation system based on feature fusion, comprising the following modules:
[0076] The data sampling module is used to acquire sampling data of the battery whose health status needs to be estimated.
[0077] The preprocessing module is used to perform one-dimensional feature extraction on the sampled data to obtain multiple preprocessed one-dimensional features.
[0078] The model building module is used to build a health status estimation model based on a convolutional neural network. The health status estimation model includes a feature extraction module, a feature fusion module, and a regression estimation module. The feature fusion module performs feature fusion on multiple feature maps generated by the feature extraction module by adding them bit by bit. The regression estimation module is used to extract the information after feature fusion.
[0079] The state estimation module is used to input the one-dimensional features into the health state estimation model to estimate the battery health state and obtain the health state estimation result.
[0080] The preprocessing process of the preprocessing module includes:
[0081] Feature extraction processing is performed on the sampled data to obtain discharge capacity data and voltage data within multiple cycles;
[0082] The first function curve is constructed based on the discharge capacity data and voltage data;
[0083] Differentiate the voltage data based on the discharge capacity data to obtain the rate of change of the discharge capacity data, which serves as the second function curve.
[0084] The third function curve is obtained by comparing the discharge capacity data in each cycle with the discharge capacity data in the first cycle.
[0085] In this embodiment, the feature extraction module includes three parallel sub-model structures, which are used to extract features from the one-dimensional features of each input separately to obtain the corresponding feature maps. In the feature fusion module, the feature maps from the three sub-model structures are added bit by bit to achieve feature fusion between the inputs.
[0086] Each module in this embodiment is used to implement the corresponding steps in embodiment 1, and the detailed implementation process can be found in embodiment 1.
[0087] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.
Claims
1. A method for estimating battery health status based on feature fusion using a convolutional neural network, characterized in that, Includes the following steps: Acquire sampling data of the battery whose health status needs to be estimated; The sampled data is subjected to one-dimensional feature extraction processing to obtain multiple one-dimensional features after data preprocessing; A health status estimation model is constructed based on a convolutional neural network. The health status estimation model includes a feature extraction module, a feature fusion module, and a regression estimation module. The feature fusion module fuses multiple feature maps generated by the feature extraction module by adding them point by point; the regression estimation module is used to extract the information after feature fusion. The one-dimensional feature is input into the health status estimation model to estimate the battery health status, and the health status estimation result is obtained. One-dimensional feature extraction is performed on the sampled data to obtain multiple preprocessed one-dimensional features, including: Feature extraction processing is performed on the sampled data to obtain discharge capacity data and voltage data within multiple cycles; The first function curve is constructed based on the discharge capacity data and voltage data; Differentiate the voltage data based on the discharge capacity data to obtain the rate of change of the discharge capacity data, which serves as the second function curve. The third function curve is obtained by comparing the discharge capacity data in each cycle with the discharge capacity data in the first cycle. The feature extraction module includes three parallel sub-model structures, which are used to extract features from the one-dimensional features of each input separately to obtain the corresponding feature maps. In the feature fusion module, the feature maps from the three sub-model structures are added bit by bit to achieve feature fusion between the inputs. Each sub-model structure includes multiple convolutional layers and a max-pooling layer set after the convolutional layers to effectively extract low-level features; each convolutional layer in the sub-model structure uses a design with multiple small convolutional kernels to increase the depth of the network without increasing the number of model parameters.
2. The battery health status estimation method according to claim 1, characterized in that, Before estimating the battery health status, the health status estimation method further optimizes the model parameters of the battery health status estimation model using an optimization algorithm to obtain a trained battery health status estimation model.
3. The battery health status estimation method according to claim 2, characterized in that, The Adam algorithm was selected as the optimization algorithm.
4. A battery health status estimation system based on feature fusion using a convolutional neural network, characterized in that, Includes the following modules: The data sampling module is used to acquire sampling data of the battery whose health status needs to be estimated. The preprocessing module is used to perform one-dimensional feature extraction on the sampled data to obtain multiple preprocessed one-dimensional features. The model building module is used to build a health status estimation model based on a convolutional neural network. The health status estimation model includes a feature extraction module, a feature fusion module, and a regression estimation module. The feature fusion module fuses multiple feature maps generated by the feature extraction module by adding them point by point; the regression estimation module is used to extract the information after feature fusion. The state estimation module is used to input the one-dimensional features into the health state estimation model to estimate the battery health state and obtain the health state estimation result. The preprocessing process in the preprocessing module includes: Feature extraction processing is performed on the sampled data to obtain discharge capacity data and voltage data within multiple cycles; The first function curve is constructed based on the discharge capacity data and voltage data; Differentiate the voltage data based on the discharge capacity data to obtain the rate of change of the discharge capacity data, which serves as the second function curve. The third function curve is obtained by comparing the discharge capacity data in each cycle with the discharge capacity data in the first cycle. The feature extraction module includes three parallel sub-model structures, which are used to extract features from the one-dimensional features of each input separately to obtain the corresponding feature maps. In the feature fusion module, the feature maps from the three sub-model structures are added bit by bit to achieve feature fusion between the inputs. Each sub-model structure includes multiple convolutional layers and a max-pooling layer set after the convolutional layers to effectively extract low-level features; each convolutional layer in the sub-model structure uses a design with multiple small convolutional kernels to increase the depth of the network without increasing the number of model parameters.