Lithium battery capacity prediction method based on DRSN-Transform algorithm

Abstract

The invention discloses a lithium battery capacity prediction method based on a DRSN-Transformer algorithm, and the method comprises the following steps: S1, adding Gaussian white noise with different signal-to-noise ratios into an original signal, carrying out the time-frequency domain transformation of a capacity signal of the white noise, and constructing a time-frequency graph; s2, carrying out feature extraction and capacity prediction by using a DRSN-Transform model; s3, noise reduction is carried out on the DRSN-Transform model by using a soft threshold module in the difference depth residual shrinkage network, and important features on different subspaces in the DRSN-Transform model are concerned by using multiple heads of attention in the Transform; s4, analyzing a battery capacity estimation method of the lithium battery capacity prediction method based on the DRSN-Transform algorithm; s5, the academic property of the lithium battery capacity prediction method based on the DRSN-Transform algorithm is analyzed; s6, analyzing the uniqueness of the lithium battery capacity prediction method based on the DRSN-Transform algorithm; the safety and reliability of the battery system are improved, the service life of the battery is prolonged, the cost is reduced, preventive measures and key warning systems are triggered, and catastrophic results are avoided.

Description

Technical Field

[0001] The present invention belongs to the technical field of lithium battery capacity prediction, and in particular relates to a lithium battery capacity prediction method based on a DRSN-Transformer algorithm. Background Art

[0002] With the development of new energy technology, lithium batteries, as core energy storage devices, are of vital importance to the stability of the entire new energy system in terms of reliability and lifespan. In response to the problems faced by battery capacity prediction and analysis, such as online analysis requirements, sensitive characteristics of influencing factors, and insufficient detection accuracy, the research team proposed a battery capacity decomposition trajectory prediction algorithm based on time-frequency analysis and image feature extraction. The algorithm is based on digital signal processing and deep learning technology, using DWT, VMD, EMD and its derived digital signal processing technology to extract key representations of capacity decay from the time series data of lithium battery cycle charge and discharge, and then input the results into the improved DRSN model for feature extraction and capacity prediction, to build a PHM system such as Figure 1 .

[0003] Over the past decade, the global energy market has seen a dramatic increase in the adoption of lithium-ion batteries (LIBs) as a reliable fuel source for electric vehicles (EVs), electronic devices, and medical devices.

[0004] According to statistics reported by the Electric Vehicle Administration (EDTA), the sales of electric vehicles in the US market increased from 345 in 2010 to 601,600 in 2022, with a total of 1.8 million electric vehicles during the 12-year sales period. As a key component for powering electric vehicles and renewable energy systems (RES), the energy management of battery packs directly affects their performance under various operating conditions. Due to their high energy density, long service life, and no memory effect, lithium-ion (Li-ion) batteries have become the first choice for electric vehicles and renewable energy, and their application range has also expanded to various fields such as robots, automatic guided vehicles (AGVs), and consumer electronics.

[0005] As the demand for lightweight, sustainable, and longer lifecycle energy storage devices continues to increase, lithium-ion batteries have become a universal solution due to their lighter weight, higher energy density, relatively low self-discharge rate, and longer lifecycle. Despite their huge advantages over traditional fossil fuels, lithium-ion batteries still have reliability and safety issues. There have been reports of lithium-ion batteries in electric vehicles, mobile phones, and energy storage systems occasionally exploding due to their high flammability, so a safer method is needed to further accelerate the deployment of electric vehicles.

[0006] PHM systems can monitor the state of the battery through both prognostic and health monitoring frameworks. The task of health monitoring is to detect potential degradation and prevent potential failures, while prognostics are responsible for predicting how long the product will be on the road to failure. PHM technology usually consists of three main components: condition monitoring, data acquisition, and health diagnostics. Condition monitoring focuses on battery performance and derives key aspects of battery performance such as voltage, current, charge and discharge capacity. Data acquisition obtains the required performance indicators, and health diagnostics is responsible for estimating the health of the battery.

[0007] The project aims to use digital signal processing and deep learning technologies to focus on battery capacity degradation, thereby improving the safety and reliability of battery systems, extending battery life, reducing maintenance costs, and triggering preventive measures and key warning systems to intervene in time when battery capacity rapidly decays, and possibly avoid catastrophic consequences. Summary of the invention

[0008] The present invention provides a lithium battery capacity prediction method based on the DRSN-Transformer algorithm, which improves the safety and reliability of the battery system, prolongs the battery life, reduces maintenance costs and plays a role in triggering preventive measures and key warning systems, performs timely intervention when the battery capacity decays rapidly, and may avoid catastrophic consequences.

[0009] The technical solution of the present invention is as follows:

[0010] A lithium battery capacity prediction method based on DRSN-Transformer algorithm comprises the following steps:

[0011] Step S1: adding Gaussian white noise with different signal-to-noise ratios to the original signal, performing time-frequency domain transformation on the capacity signal of the white noise, and constructing a time-frequency diagram;

[0012] Step S2: Use the DRSN-Transformer model for feature extraction and capacity prediction;

[0013] Step S3: Use the soft threshold module in the difference deep residual shrinkage network to denoise the DRSN-Transformer model, and use the multi-head attention in the Transformer to focus on important features in different subspaces of the DRSN-Transformer model;

[0014] Step S4: analyzing the battery capacity estimation method of the lithium battery capacity prediction method based on the DRSN-Transformer algorithm;

[0015] Step S5: Analyze the academic nature of the lithium battery capacity prediction method based on the DRSN-Transformer algorithm;

[0016] Step S6: Analyze the uniqueness of the lithium battery capacity prediction method based on the DRSN-Transformer algorithm.

[0017] Preferably, step S3 is used to enhance the overall feature extraction capability of the model.

[0018] Preferably, step S4 includes: a direct measurement method, an analysis-based method, a SOC-based method, and a data-driven method;

[0019] Analysis-based methods include incremental curve analysis method, differential voltage curve analysis method, differential thermal analysis method, mechanical stress analysis method, and EIS analysis method.

[0020] Preferably, the data-driven method comprises the following steps:

[0021] Step S41: collecting data and preprocessing data;

[0022] Step S42: Performing offline training on the data according to the preprocessed data;

[0023] Step S43: Estimating the battery capacity online based on the offline training data.

[0024] Preferably, S5 includes the following sub-steps:

[0025] Sub-step S51: Analyze the innovation of cross-integration of multiple fields;

[0026] Sub-step S52: Analyze the experiment.

[0027] Preferably, step S52 includes the following sub-steps:

[0028] Sub-step S521: exploring the performance in analyzing capacity regeneration phenomenon based on three signal decomposition techniques DWT, EMD and VMD and other digital signal processing algorithms;

[0029] Sub-step S522: Implement the Matlab signal MultiresolutionAnalyzer toolbox based on python 3.7 according to the performance;

[0030] Sub-step S523: According to the Matlab signal Multiresolution Analyzer toolbox, a capacity decay model is constructed based on DRSN-Transformer, parameters are set, and DRSN is used to remove redundant noise and reduce noise interference.

[0031] Preferably, uniqueness includes:

[0032] The innovation of predicting battery capacity based on the DRSN-Transformer algorithm, the innovation of noise processing in time-frequency domain transformation, the innovation of combining DRSN and Transformer, and the innovation of comparison with existing literature.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] 1. The present invention aims to utilize digital signal processing and deep learning technology to focus on battery capacity degradation, thereby improving the safety and reliability of the battery system and extending the battery life.

[0035] 2. The present invention reduces maintenance costs and acts as a trigger for preventive measures and critical warning systems.

[0036] 3. The present invention provides timely intervention when the battery capacity rapidly decays and may avoid catastrophic consequences. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0038] Figure 1 This is a PHM system diagram of the present invention.

[0039] Figure 2 FIG. 4 is a diagram of the NASA data collection process of the present invention.

[0040] FIG3( a ) is a flow chart of the overall experiment of the present invention.

[0041] FIG3( b ) is a diagram showing the overall experimental details of the present invention.

[0042] FIG. 3( c ) is a diagram showing the overall experimental details of the present invention.

[0043] Figure 4 It is a schematic diagram of the application of the data-driven method of the present invention.

[0044] Figure 5 This is a flow chart of the data-driven capacity estimation method of the present invention.

[0045] FIG6( a ) is a DWT-based diagram of the extracted IMFs / RES of the present invention.

[0046] FIG6( b ) is an EMD-based diagram of the extracted IMFs / RES of the present invention.

[0047] FIG6( c ) is a VMD-based diagram of the extracted IMFs / RES of the present invention.

[0048] FIG. 7( a ) is a diagram showing the soft thresholding result of the present invention.

[0049] FIG. 7( b ) is a partial derivative diagram of the soft threshold of the present invention.

[0050] FIG8( a ) is a structural design diagram of the 2D-DRSN of the present invention.

[0051] FIG8( b ) is a diagram showing the architecture of the residual shrinkage unit of the present invention.

[0052] Fig. 9 It is a schematic diagram of the 2D-RSBU structure of the present invention. DETAILED DESCRIPTION

[0053] The present invention is described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those of ordinary skill in the art, several variations and improvements may be made without departing from the concept of the present invention. These all belong to the present invention.

[0054] A lithium battery capacity prediction method based on DRSN-Transformer algorithm comprises the following steps:

[0055] Step S1: adding Gaussian white noise with different signal-to-noise ratios to the original signal, performing time-frequency domain transformation on the capacity signal of the white noise, and constructing a time-frequency diagram;

[0056] Step S2: Use the DRSN-Transformer model for feature extraction and capacity prediction;

[0057] Step S3: Use the soft threshold module in the difference deep residual shrinkage network to denoise the DRSN-Transformer model, and use the multi-head attention in the Transformer to focus on important features in different subspaces of the DRSN-Transformer model;

[0058] Step S4: analyzing the battery capacity estimation method of the lithium battery capacity prediction method based on the DRSN-Transformer algorithm;

[0059] Step S5: Analyze the academic nature of the lithium battery capacity prediction method based on the DRSN-Transformer algorithm;

[0060] Step S6: Analyze the uniqueness of the lithium battery capacity prediction method based on the DRSN-Transformer algorithm.

[0061] Step S3 of this embodiment is used to enhance the overall feature extraction capability of the model.

[0062] Step S4 of this embodiment includes: a direct measurement method, an analysis-based method, an SOC-based method, and a data-driven method;

[0063] Analysis-based methods include incremental curve analysis method, differential voltage curve analysis method, differential thermal analysis method, mechanical stress analysis method, and EIS analysis method.

[0064] The data-driven method of this embodiment includes the following steps:

[0065] Step S41: collecting data and preprocessing data;

[0066] Step S42: Performing offline training on the data according to the preprocessed data;

[0067] Step S43: Estimating the battery capacity online based on the offline training data.

[0068] S5 of this embodiment includes the following sub-steps:

[0069] Sub-step S51: Analyze the innovation of cross-integration of multiple fields;

[0070] Sub-step S52: Analyze the experiment.

[0071] Step S52 of this embodiment includes the following sub-steps:

[0072] Sub-step S521: exploring the performance in analyzing capacity regeneration phenomenon based on three signal decomposition techniques DWT, EMD and VMD and other digital signal processing algorithms;

[0073] Sub-step S522: Implement the Matlab signal MultiresolutionAnalyzer toolbox based on python 3.7 according to the performance;

[0074] Sub-step S523: According to the Matlab signal Multiresolution Analyzer toolbox, a capacity decay model is constructed based on DRSN-Transformer, parameters are set, and DRSN is used to remove redundant noise and reduce noise interference.

[0075] Unique features of this implementation include:

[0076] The innovation of predicting battery capacity based on the DRSN-Transformer algorithm, the innovation of noise processing in time-frequency domain transformation, the innovation of combining DRSN and Transformer, and the innovation of comparison with existing literature.

[0077] When this implementation plan is implemented,

[0078] The present invention proposes a method based on the DRSN-Transformer algorithm, which can accurately predict the capacity decay curve of lithium batteries under high noise conditions. First, Gaussian white noise with different signal-to-noise ratios (SNR) is added to the original signal, the noisy capacity signal is transformed in the time-frequency domain, and the time-frequency diagram is constructed. Then, the DRSN-Transformer model is used for feature extraction and capacity prediction. The soft threshold module in the differential deep residual shrinkage network (DRSN) can achieve better noise reduction effect. At the same time, the multi-head attention in the Transformer can focus on important features in different subspaces, thereby enhancing the overall feature extraction capability of the model. The overall process can be simplified into three parts:

[0079] 1. Data visualization and analysis based on lithium battery data and statistical science

[0080] 2. Time-frequency domain analysis based on digital signal processing

[0081] 3. Feature extraction and capacity prediction based on DRSN.

[0082] To enhance the rigor and scientificity of the experimental results, the experiment will use a random cycle charge and discharge battery data set from NASA in the United States. The data collection process is as follows: Figure 2 shown.

[0083] The overall experimental process is shown in Figure 3(a);

[0084] Figure 3(b) and Figure 3(c) are detailed diagrams of the overall experimental process.

[0085] 1. Cutting-edge analysis

[0086] 1. Current developments and challenges in the field

[0087] Considering the complexity of battery degradation, it is still challenging for BMS to accurately predict the capacity of on-board batteries. Researchers have made great efforts to solve this problem. This section will briefly introduce the advantages and disadvantages of battery capacity estimation methods in various research literatures. Among them, the methods are mainly divided into direct measurement methods, analysis-based methods, SOC-based methods and data-driven methods. Their principles and current processes will be explained in detail in the following subsections.

[0088] 1.1 Direct measurement method

[0089] Direct measurement methods require that the battery be fully charged or discharged under specific conditions. Currently, various standards from the International Electrotechnical Commission (IEC), the International Organization for Standardization (ISO), and the Institute of Electrical and Electronics Engineers Standards Association (IEEE-SA) have been proposed for testing the capacity of lithium-ion batteries under standard conditions. However, this test procedure is quite strict compared to the working environment of the battery pack in actual applications. The battery needs to be soaked at a predetermined temperature for at least 12 hours to ensure thermal stability, which requires the battery temperature to change less than 1°C within a 1-hour time interval. The current and voltage measurement accuracy should be less than + / -1%, and the time measurement is less than + / -0.1%. Therefore, it is impractical to always meet such stringent requirements in battery applications, which limits these test methods to laboratory tests as a reference.

[0090] 1.2 Analysis-based methods

[0091] a. Incremental curve analysis method

[0092] The IC curve mainly extracts the change of voltage with the degradation of lithium-ion batteries, so it can be used for aging mechanism analysis. However, a very low current rate is required to obtain the IC curve for battery diagnosis, but using the IC curve to analyze battery degradation also requires professional knowledge about the electrochemical reactions inside the battery. If the reaction mechanism inside the battery is not understood, it may lead to misjudgment of the results. At the same time, because the IC curve is obtained by derivation of the voltage curve, it is very sensitive to noise. Any measurement noise will be amplified, which may cause the curve to have a false signal or fail to accurately reflect the actual state of the battery.

[0093] b. Differential voltage curve analysis method

[0094] This method calculates the change in battery capacity within a constant voltage interval to obtain a dQ / dV-V curve. Each peak of the curve represents an electrochemical reaction, the peak point represents the phase change point of the material, and the area enclosed by the curve and the horizontal axis represents the capacity charged or discharged during the phase change. Its advantage is that it uses direct measurement of BMS and has low implementation cost. A common problem with the differential voltage curve analysis method is that it cannot effectively handle measurement noise in reality.

[0095] c. Differential thermal analysis

[0096] Although compared with IC and DV curves, the DT curve can be applied to large current charging or discharging. However, since DTA measures the temperature difference between the sample and the reference material, this difference is usually very small, which makes the DT curve sensitive to measurement noise. Any electrical noise or environmental noise in the experimental system (such as temperature fluctuations, mechanical vibrations, etc.) may interfere with the signal and reduce the sensitivity and accuracy of the measurement. It can be used for aging mechanism analysis because it is crucial to the heat generation during the electrochemical reaction. In addition, the connection between the DT curve and battery aging still needs to be studied.

[0097] d. Mechanical stress analysis

[0098] Although mechanical stress is considered an effective method for battery testing, the main limitation is that it requires specially designed expansion measurement equipment, such as displacement sensors, pressure sensors, test fixtures, etc. As a result, the cost increases, and its implementation into the battery pack must be carefully considered so as not to destroy its original design. At the same time, for dynamic stress analysis (such as fatigue testing, etc.), these instabilities are particularly obvious and will affect the reliability of the data. The increase in noise will reduce the signal-to-noise ratio (SNR) of the signal, making it difficult to distinguish the real stress signal from the noise. This may make it difficult to accurately identify stress concentration points or stress change trends.

[0099] e.EIS analysis method

[0100] Battery EIS has a strong potential to reflect electrochemical reactions in the frequency domain and is expected to have great potential in on-board BMS applications. However, most EIS-related battery degradation analysis is based on commercial electrochemical workstations, which are accurate but costly. Given the size and weight of electrochemical workstations, they are difficult to use directly in an EV environment, and given the cost of current BMS, examples of hardware design for EIS measurements are still very limited. Another problem is that EIS measurements are susceptible to noise. In EIS measurements, the signal applied to the system is usually a small-amplitude AC signal to avoid excessive perturbations to the electrochemical system. Low-amplitude signals are inherently susceptible to noise. Noise reduces the signal-to-noise ratio (SNR) of the signal, causing the true impedance signal to be masked by noise, thereby affecting the accuracy of the impedance data.

[0101] 1.3 Based on data-driven approach

[0102] With the rapid development of the Internet of Things and artificial intelligence, daily operational measurements of battery systems are easily recorded into cloud platforms and can be further used for cloud-to-edge estimation calculations. Data-driven approaches are characterized by relying on large data sets to make decisions and do not require specific battery models. In data-driven approaches, as long as sufficiently representative samples are available, models can be used to map the data without predetermining a definitive model in advance. For batteries, the operational measurements that can be collected may contain aging information. The degree of aging can be reflected by certain characteristics during the charging and discharging process, and the data-driven approach builds an approximate model to match the true aging situation with this information. Figure 4 It is a schematic diagram of the application of data-driven approach.

[0103] It mainly includes three processes: data collection and preprocessing, offline training and online estimation. The main purpose of data collection is to measure the voltage, current and temperature of the battery pack during operation. Then, the data-driven model can be trained offline using a high-computing processor, and the trained model can be implemented in the BMS for online estimation, such as Figure 5 shown.

[0104] In this type of approach, the key lies in data processing, extraction of key features, and model training. Although it has high accuracy, data-driven models are rarely used in current BMS. The reason is that the performance of data-driven methods is closely related to the characteristics and quality of the measurement data set. There is a lack of open source data sets from real BESS applications to train data-driven models. In reality, it is also difficult to obtain labeled data sets, and it is costly to measure all required data sets from experimental test platforms. Another point is that data-driven methods lack interpretability, and the credibility of the estimated results may be questionable for practical applications. On this basis, noisy data may cause the model to learn wrong patterns or associations during training, thereby affecting the generalization ability and prediction accuracy of the model. Especially in supervised learning, noisy labels (i.e., wrong label data) can lead to model misleading and ultimately reduce the performance of the model. Noise may change the actual distribution of the data, causing overfitting of certain abnormal patterns during model training. In this way, the model may not show good adaptability when facing real noise-free data.

[0105] 2. Research innovation

[0106] In response to the problems of the above research methods, the team developed a hybrid onboard capacity estimation framework of data-driven methods and other technologies. This onboard capacity estimation framework shows significant advantages in dealing with noise by combining data-driven methods with time-frequency domain signal decomposition methods. The time-frequency domain signal decomposition method can analyze signals at multiple scales, extract features that are sensitive to the health status of the battery, and effectively filter out noise interference. It can retain key information even in a high-noise environment and ensure the reliability of the data-driven model. The fusion of multiple technologies not only enhances the noise suppression capability, but also effectively corrects anomalies in the data and reduces the negative impact of noise on the model. In addition, the application of a single data processing program simplifies noise management, allowing filtering and noise reduction processing to be performed more uniformly and efficiently, further improving data quality. The framework also has dynamic adjustment capabilities, which can monitor the battery operating status in real time and adjust parameter settings according to the noise level to ensure that the system can still accurately estimate the battery capacity in a complex environment. By automatically extracting health status features, researchers do not need to have an in-depth understanding of the complex battery degradation mechanism, thereby simplifying the research process and reducing the possibility of noise introduction. In addition, the method predicts capacity regeneration phenomena more accurately and maintains good prediction performance even in the presence of noise. Therefore, this framework can not only effectively deal with the noise challenges in lithium-ion battery capacity estimation, but also improve the prediction accuracy and robustness of the model, providing a more reliable solution for practical applications.

[0107] 2. Academic Analysis

[0108] 1. Innovation of cross-disciplinary integration

[0109] This project integrates cutting-edge technologies from multiple fields, involving disciplines such as signal processing, deep learning, and electrochemistry. Traditional lithium-ion battery health status monitoring usually relies on single signal feature extraction in the time domain or frequency domain, which makes it difficult to fully capture the complex dynamic characteristics of the battery. This project effectively solves the problem of decomposition and prediction of capacity regeneration phenomena by introducing time-frequency image analysis and combining nonlinear signal processing techniques such as discrete wavelet transform (DWT) and empirical mode decomposition (EMD). This method significantly improves the model's adaptability to non-stationary and nonlinear signals, and demonstrates superior performance in dealing with the complex dynamic behaviors of battery aging and capacity attenuation. The cross-domain approach effectively improves the accuracy and flexibility of lithium-ion battery capacity prediction. This technical integration not only provides richer information extraction methods in lithium battery health management, but also provides a new research path for state monitoring and prediction of complex systems.

[0110] 2. Experimental analysis

[0111] (a) Exploring the performance of three signal decomposition techniques, DWT, EMD and VMD, and other digital signal processing algorithms in analyzing capacity regeneration phenomena; as shown in Figures 6(a), 6(b) and 6(c).

[0112] (II) Implementation based on python3.7 and Matlab signal MultiresolutionAnalyzer toolbox

[0113] VMD and MED decomposition techniques are superior to DWT decomposition techniques due to their adaptability. These techniques do not require a mother-wavelet selection process. The performance of the DWT analysis method depends on the accurate selection of the mother wavelet function. After removing the noise from the regeneration capacity curve, the experiment will use the de-noised data to train the DRSN model to predict the future trajectory of the capacity, that is, add Gaussian noise of different SNRs to the original battery signal, and perform a time-frequency transform on the noisy signal to generate a time-frequency diagram as the input sample.

[0114] In order to study the influence of noise signals on the prediction effect of the model, the vibration signal is taken as an example. First, Gaussian white noise with different signal-to-noise ratios is added to the original vibration signal. Then, the vibration signal containing noise is transformed by continuous wavelet transform to generate a two-dimensional color wavelet time-frequency diagram. At the same time, in order to avoid affecting the classification results, the legend and coordinate system are set to not be displayed. Finally, the image is grid-normalized and compressed to generate a pixel format of 128×128×3.

[0115] (III) Build a capacity decay model based on DRSN-Transformer, set parameters, use DRSN to remove redundant noise, and reduce noise interference

[0116] Zhao Minghang et al. developed a new deep learning method, called Deep Residual Shrinkage Network (DRSN), to improve the ability of deep learning methods to learn features from strong noisy signals and achieve high accuracy in fault diagnosis and data analysis. In the past 20 years, soft thresholding has often been used as a key step in signal denoising algorithms. Typically, the signal is converted to a domain. In this domain, features close to zero are unimportant. Then, soft thresholding sets these features close to zero to zero. For example, wavelet thresholding is a classic signal denoising algorithm that usually includes three steps: wavelet decomposition, soft thresholding, and wavelet reconstruction. In order to ensure the effect of signal denoising, a key task of wavelet thresholding is to design a filter. This filter can convert useful information into relatively large features and convert noise-related information into features close to zero. However, designing such a filter requires a lot of expertise in signal processing and is often very difficult. Deep learning provides a new way to solve this problem. These filters can be automatically optimized by the back-propagation algorithm instead of being designed by experts. Therefore, the combination of soft thresholding and deep learning is an effective way to eliminate noise information and obtain strong discriminative features. The equation for soft thresholding is expressed as follows

[0117]

[0118] Where x is the input feature, y is the output feature, and τ is the threshold (i.e., a positive number). Soft thresholding sets features close to zero directly to zero, rather than setting negative features to zero like ReLU, so negative and useful features can be retained. The result of soft thresholding is shown in Figure 7(a). It can be seen that the derivative of the output of soft thresholding with respect to the input is either 1 or 0, so it is also effective in avoiding the problem of gradient vanishing and gradient exploding. Its partial derivative is shown in Figure 7(b).

[0119] The 2D-DRSN structure design is shown in Figure 8(a) and Figure 8(b);

[0120] This residual module builds a special module to estimate the threshold required for soft thresholding. In this special module, global mean pooling is applied to the absolute value of the feature map to obtain a one-dimensional vector. Then, this one-dimensional vector is input into a two-layer fully connected network to obtain a scaling parameter. The Sigmoid function normalizes this scaling parameter to between 0 and 1, expressed as:

[0121]

[0122] Among them, z is the output of the two fully connected layers in RSBU-CS, and α is the corresponding scaling parameter. Then, this scaling parameter, multiplied by the average of the absolute values ​​of the feature map, is used as the threshold. This arrangement is due to the fact that the threshold of soft thresholding must not only be positive but also not too large. If the threshold is greater than the feature with the largest absolute value in the feature map, then the output of soft thresholding will all be 0. In summary, the threshold in RSBU-CS is expressed as:

[0123] τ=α·a i,j,c |x i,j,c | Formula (3)

[0124] Among them, i, j, and c represent the width, length, and channel number of the feature map x, respectively. In this way, the threshold can be controlled within a suitable range and the output features will not be all zero.

[0125] The present invention converts the battery capacity loss signal into a two-dimensional time-frequency diagram as an input sample through time-frequency domain transformation. At the same time, RSBU is designed for a one-dimensional time series signal at the beginning. In order to adapt to the input sample, the one-dimensional convolution in RSBU is changed to a two-dimensional convolution and named 2D-RSBU. Fig. 9 shown.

[0126] The only difference is that RSBU replaces the normal residual building block. A certain number of RSBUs are stacked so that noise-related features are gradually reduced. Another advantage is that the threshold is automatically learned instead of being manually set by an expert, so no expertise in signal processing is required when implementing DRSN.

[0127] 3. Uniqueness Analysis

[0128] 1. Battery capacity prediction based on DRSN-Transformer algorithm

[0129] Innovation: A battery capacity prediction algorithm combining Transformer and Deep Residual Shrinkage Network (DRSN) is proposed. This method can accurately identify key features under high noise conditions, which is not possible with other traditional methods. The residual shrinkage feature of DRSN helps to reduce the impact of noise on feature extraction, while the attention mechanism of Transformer enhances the ability to focus on important features, thereby improving the accuracy of prediction.

[0130] 2. Noise processing in time-frequency domain transformation

[0131] Innovation: Noise with different signal-to-noise ratios is added to the original signal, and then time-frequency domain transformations such as continuous wavelet transform (CWT) are applied to construct the time-frequency diagram. This method enables the algorithm to process data under various noise levels and effectively capture the time-frequency characteristics in the signal, while facilitating the detection of noise filtering effects in subsequent studies. Compared with traditional time-domain analysis, this time-frequency domain analysis method can more comprehensively reflect the dynamic changes and noise characteristics in the signal, thereby improving the accuracy and robustness of battery capacity prediction.

[0132] 3. Combination of DRSN and Transformer

[0133] Innovation: Combining the soft threshold module (used to process noise) with the multi-head attention mechanism in the Transformer module enhances the model's attention to important features in different subspaces. This combination not only improves the ability to extract features, but also enhances the model's robustness to noise. In this way, the model can still extract key battery status features when processing noisy data, thereby improving prediction performance.

[0134] 4. Comparison with existing literature

[0135] Innovation: The study found that most existing studies focus on time domain analysis, but rarely involve comprehensive research in the environmental domain. In addition, many prediction algorithms are complex in design and require a lot of prior knowledge, which limits their practical application effects. The time-frequency domain analysis algorithm developed by the team not only fills this research gap, but also simplifies the complexity of the prediction model, making the model more practical in practical applications.

Claims

1. A lithium battery capacity prediction method based on DRSN-Transformer algorithm, characterized in that: The following steps are involved: Step S1: adding Gaussian white noise with different signal-to-noise ratios to the original signal, performing time-frequency domain transformation on the capacity signal of the white noise, and constructing a time-frequency diagram; Step S2: Use the DRSN-Transformer model for feature extraction and capacity prediction; Step S3: Use the soft threshold module in the difference deep residual shrinkage network to denoise the DRSN-Transformer model, and use the multi-head attention in the Transformer to focus on important features in different subspaces of the DRSN-Transformer model; Step S4: analyzing the battery capacity estimation method of the lithium battery capacity prediction method based on the DRSN-Transformer algorithm; Step S5: Analyze the academic nature of the lithium battery capacity prediction method based on the DRSN-Transformer algorithm; Step S6: Analyze the uniqueness of the lithium battery capacity prediction method based on the DRSN-Transformer algorithm.

2. The lithium battery capacity prediction method based on the DRSN-Transformer algorithm according to claim 1 is characterized in that: The step S3 is used to enhance the overall feature extraction capability of the model.

3. The lithium battery capacity prediction method based on the DRSN-Transformer algorithm according to claim 1 is characterized in that: The step S4 includes: a direct measurement method, an analysis-based method, a SOC-based method, and a data-driven method; The analysis-based methods include incremental curve analysis method, differential voltage curve analysis method, differential thermal analysis method, mechanical stress analysis method, and EIS analysis method.

4. The lithium battery capacity prediction method based on the DRSN-Transformer algorithm according to claim 3 is characterized in that: The data-driven method comprises the following steps: Step S41: collecting data and preprocessing data; Step S42: Performing offline training on the data according to the preprocessed data; Step S43: Estimating the battery capacity online based on the offline training data.

5. The lithium battery capacity prediction method based on the DRSN-Transformer algorithm according to claim 1, characterized in that: The S5 comprises the following sub-steps: Sub-step S51: Analyze the innovation of cross-integration of multiple fields; Sub-step S52: Analyze the experiment.

6. The lithium battery capacity prediction method based on the DRSN-Transformer algorithm according to claim 5, characterized in that: The step S52 includes the following sub-steps: Sub-step S521: exploring the performance in analyzing capacity regeneration phenomenon based on three signal decomposition techniques DWT, EMD and VMD and other digital signal processing algorithms; Sub-step S522: Implement the Matlab signal MultiresolutionAnalyzer toolbox based on python 3.7 according to the performance; Sub-step S523: According to the Matlab signal Multiresolution Analyzer toolbox, a capacity decay model is constructed based on DRSN-Transformer, parameters are set, and DRSN is used to remove redundant noise and reduce noise interference.

7. The lithium battery capacity prediction method based on the DRSN-Transformer algorithm according to claim 1, characterized in that: The uniqueness includes: The innovation of predicting battery capacity based on the DRSN-Transformer algorithm, the innovation of noise processing in time-frequency domain transformation, the innovation of combining DRSN and Transformer, and the innovation of comparison with existing literature.

Images