Lithium ion battery K value prediction method and system based on time-frequency domain feature fusion
By building a dual-channel LSTM twin network and an optimized LightGBM model, combining dynamic feature selection and mixed loss function, the problem of inaccurate prediction of lithium batteries in traditional methods is solved, and high-precision and efficient K-value detection is achieved.
Patent Information
- Application Number
- CN202510378937.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-28
- Publication Date
- 2025-07-22
AI Technical Summary
The prior art is difficult to accurately predict the K value of lithium batteries. Traditional methods rely on macro indicators and cannot capture microscopic changes. Data-driven methods are sensitive to noise. Traditional LSTM models have gradient vanishing problems in long-sequence prediction, resulting in inaccurate prediction and inefficient prediction.
A dual-channel LSTM twin network is built to extract the time domain and frequency domain signals. Combined with the optimized LightGBM model, dynamic feature selection and mixed loss functions are used to perform time-frequency domain feature fusion to achieve high-precision prediction.
The accuracy and efficiency of lithium battery K value prediction have been significantly improved, the RMSE has been reduced to 0.17%, the false detection rate has been reduced to 2.1%, and the detection efficiency has been improved by 3 times, which can more accurately reflect the changes in battery performance.
Smart Images
Figure CN120354056A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of lithium battery detection, and particularly relates to a method for predicting the K value of a lithium-ion battery based on the fusion of time-frequency domain features. Background Art
[0002] Among the numerous performance indicators of lithium batteries, the K value (voltage decay rate), as a key parameter reflecting the performance degradation of the battery, has extremely high research value and practical significance. The K value quantifies the voltage decay rate of the battery during charge and discharge, and can accurately reveal the internal performance changes of the battery from a microscopic level.
[0003] Although lithium battery technology has made great progress in the past few decades, there are mainly the following defects in K value prediction and related performance optimization:
[0004] (1) Limitations of traditional prediction methods: Traditional lithium battery performance prediction methods mainly rely on macroscopic indicators such as SOH (State of Health) and RUL (Remaining Useful Life). Although these methods can reflect the overall state of the battery to a certain extent, it is difficult to capture the subtle changes in the microscopic performance of the battery. The chemical reactions and physical processes inside the battery are extremely complex, and the microscopic changes are often the root causes of battery performance degradation. Relying solely on macroscopic indicators cannot deeply insight into these internal changes, and thus cannot accurately predict in advance the possible performance problems of the battery. In addition, although the model based on physical mechanism can theoretically explain the working principle and performance changes of the battery, this type of model requires calibration of a large number of complex parameters. In actual industrial scenarios, due to differences in battery production processes, diversity of usage environments, and different degrees of battery aging, it is extremely difficult to accurately obtain these parameters, resulting in the model being difficult to adapt to the actual situation, and the accuracy of the prediction results being greatly reduced. While the data-driven method has certain advantages in processing large-scale data, this method is very sensitive to noise. There are inevitably various noise interferences in the industrial production environment, and these noises will seriously affect the quality of the data, thereby interfering with the training and prediction of the model, resulting in the model lacking interpretability, and it is difficult for users to understand the decision-making process and basis of the model.
[0005] (2) Pain points of K value detection: According to the current GB / T 31486-2015 standard, K value detection requires a 28-day static test. This detection method has many disadvantages. The first is the long detection cycle. For production companies, this means that the efficiency of product shipments is greatly reduced, and a large number of products are piled up in warehouses, occupying a large amount of funds and storage space, affecting the company's capital turnover and production plans. Secondly, the testing cost is high. The testing fee for each batch is as high as 50,000 to 80,000 yuan. This is a considerable expense for large-scale production companies, increasing the company's production costs and reducing the market competitiveness of the products. More importantly, this detection method cannot monitor the status of the battery cells on the production line in real time. During the production process, the battery cells may have performance abnormalities due to various factors, and the problem can only be discovered after the 28-day static test. At this time, a large number of unqualified products may have been produced, causing serious waste of resources and economic losses.
[0006] (3) Technical bottleneck: The voltage decay process of lithium batteries presents nonlinear and time-varying characteristics, which makes it extremely difficult to accurately predict voltage decay. During the battery charging and discharging process, its internal chemical reaction rate, ion migration speed, etc. will change with time and usage conditions, resulting in complex and changeable laws of voltage decay, which is difficult to accurately describe and predict using traditional linear models. At the same time, industrial data has problems such as time series discontinuity and environmental noise interference. The data collected on the production line may have data missing and abnormal values due to equipment failure, signal transmission problems, etc., and electromagnetic interference, temperature changes and other factors in the environment will also cause noise interference to the data. These have brought great challenges to data analysis and model training, and seriously affected the accuracy and reliability of the prediction model. In addition, the traditional LSTM model is prone to the gradient vanishing problem when processing long sequence predictions. As the sequence length increases, the gradient of the model will gradually decrease during the back propagation process, resulting in the model being unable to effectively learn the long-term dependencies in the long sequence. For application scenarios such as lithium batteries that require long-term monitoring and prediction of performance, this defect of the traditional LSTM model limits its application effect in actual production. Summary of the invention
[0007] In view of the shortcomings of the prior art, the present invention proposes a lithium-ion battery K value prediction method based on time-frequency domain feature fusion. By constructing a dual-path LSTM twin network, it can extract features of time domain signals and frequency domain signals respectively, fully mine the useful information in the data, and combine with the optimized LightGBM model to further analyze and process the extracted time-frequency domain features to achieve high-precision prediction of the K value of the lithium battery.
[0008] To implement the above technical solution, the present invention provides a lithium-ion battery K value prediction method based on time-frequency domain feature fusion, which specifically includes the following steps:
[0009] S1. Collect the time-domain signals of the battery charging and discharging processes through multiple sensors to obtain the voltage, temperature, and current parameters of the battery during charging and discharging;
[0010] S2. Perform a fast Fourier transform on the time-domain signals to obtain frequency-domain features, specifically calculated through the formula to convert the time-domain signals into frequency-domain signals;
[0011] S3. Construct a hybrid prediction model containing LSTM and LightGBM. LSTM is used to extract temporal features, and LightGBM is used to further analyze and predict the features;
[0012] S4. Adopt a time-frequency domain feature fusion strategy for K-value prediction to fuse the time-domain features and frequency-domain features.
[0013] Preferably, in step S2, after converting the time-domain signals into frequency-domain signals, calculate the energy distribution of the signals at different frequencies through the formula ; then extract the amplitudes of the first N main frequency components, and determine the value of N according to actual requirements and signal characteristics to extract the amplitudes of the main frequency components in the signals.
[0014] Preferably, a dynamic feature selection mechanism is introduced in the process of converting the time-domain signals into frequency-domain signals. By calculating the importance index of each feature, and then according to the formula assign a weight w to each feature i , where a is a control parameter used to adjust the degree of change of the weight.
[0015] Preferably, a hybrid loss function is introduced in the process of converting the time-domain signals into frequency-domain signals. By setting different weights λ1, λ2, λ3, linearly combine the three loss functions of MSE, MAE, and Huber to obtain the hybrid loss function
[0016] Preferably, in step S3, the LSTM network includes a forget gate, which controls the retention and forgetting of information through the formula f t =σ(W f ·[h t-1, x t +b f ), where σ is the Sigmoid activation function, W f is the weight matrix, h t-1 is the hidden state at the previous moment, x t is the input at the current moment, and b f is the bias vector.
[0017] Preferably, in step S3, LightGBM adopts a leaf-wise growth strategy and selects the leaf node with the largest split gain for splitting each time.
[0018] Preferably, in step S3, the time domain branch consists of 3 layers of LSTM, with each layer containing 128 units, which is used to capture the long-term dependence and dynamic change characteristics in the time domain signal; the frequency domain branch consists of 2 layers of BiLSTM, with 64 units in each layer, and extracts the context information in the frequency domain signal by considering the forward and backward information of the sequence simultaneously.
[0019] Preferably, in step S1, the multi-sensor includes a voltage sensor, a temperature sensor, and a current sensor, and the sampling strategy adopts dynamic adaptive sampling, which automatically adjusts the sampling frequency according to the change of the battery voltage.
[0020] The present invention also provides a lithium-ion battery K value prediction system based on time-frequency domain feature fusion, including:
[0021] A data acquisition module, using a DAQ-2210 data acquisition card, which is used to acquire various parameters during the charging and discharging process of the battery, including voltage, temperature, and current;
[0022] A signal processing module, using a NIPXIe-5172 signal processing device, which performs denoising and filtering processing on the acquired signals;
[0023] A model inference module, deploying the above-mentioned lithium-ion battery K value prediction method based on time-frequency domain feature fusion, predicting the K value of the lithium battery according to the processed signals and the trained model, and outputting the prediction result.
[0024] The beneficial effects of the lithium-ion battery K value prediction method and system based on time-frequency domain feature fusion provided by the present invention are as follows:
[0025] (1) By constructing a dual-channel LSTM Siamese network, the present invention can extract features from time domain signals and frequency domain signals respectively, and fully mine the useful information in the data. The time domain branch consists of 3 layers of LSTM, with each layer containing 128 units, which can effectively capture the long-term dependence and dynamic change characteristics in the time domain signal; the frequency domain branch consists of 2 layers of BiLSTM, with 64 units in each layer, and can better extract the context information in the frequency domain signal by considering the forward and backward information of the sequence simultaneously. Combining with the optimized LightGBM model, further analyze and process the extracted time-frequency domain features to achieve high-precision prediction of the K value of the lithium battery. The LightGBM model adopts strategies such as feature importance weighting and histogram acceleration algorithm, which improves the prediction efficiency and accuracy of the model.
[0026] (2) The present invention innovatively introduces a dynamic feature selection mechanism and a hybrid loss function. The dynamic feature selection mechanism dynamically adjusts the feature weights according to the importance index of the features, enabling the model to more flexibly adapt to different battery states and data characteristics, and improving the prediction accuracy; the hybrid loss function consists of the mean square error (MSE), the mean absolute error (MAE), and the Huber loss function, comprehensively considering different types of errors, having stronger robustness to outliers, and improving the stability and prediction accuracy of the model.
[0027] (3) In the actual production line measurement, the present invention has achieved remarkable results. The RMSE is reduced to 0.17%, and compared with the traditional method, the prediction error is greatly reduced, and it can more accurately reflect the K value of the lithium battery; the detection efficiency is increased by 3 times, greatly improving the detection speed of the K value of the lithium battery on the production line and improving the production efficiency. Description of the Drawings
[0028] Figure 1 This is the flowchart of the present invention. Detailed Embodiments
[0029] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.
[0030] Embodiment 1: A method for predicting the K value of a lithium-ion battery based on time-frequency domain feature fusion.
[0031] Refer to Figure 1 As shown, a method for predicting the K value of a lithium-ion battery based on time-frequency domain feature fusion specifically includes the following steps:
[0032] Step 1: Collect the time-domain signals of the battery charging and discharging processes through multiple sensors to obtain the voltage, temperature, and current parameters of the battery during the charging and discharging processes.
[0033] Multi-source data acquisition is the basis for predicting the K value of the entire lithium battery, responsible for collecting key data during the charging and discharging process of the lithium battery. The multi-source data acquisition consists of multiple high-precision sensors and a data acquisition card. The sensors are responsible for sensing various physical parameters of the battery, and the data acquisition card converts the analog signals collected by the sensors into digital signals and transmits them to the subsequent data processing module. Its acquisition parameters are extremely rich, covering 12 key parameters such as voltage, temperature, current, internal resistance, and capacity. Taking the voltage parameter as an example, a high-precision voltage sensor can accurately measure the voltage change of the battery during the charging and discharging process, with an accuracy of up to ±1mV, which is crucial for capturing subtle electrochemical reaction changes inside the battery; the temperature sensor can monitor the temperature on the surface and inside of the battery in real time, with the accuracy controlled within ±0.5°C, because the change in battery temperature is closely related to the performance degradation of the battery, and too high a temperature may accelerate the aging of the battery. These parameters reflect the working state of the battery from multiple dimensions and provide comprehensive data support for subsequent analysis and prediction.
[0034] In terms of data acquisition, the present invention has clear and detailed acquisition dimensions, covering multiple key parameters during the charging and discharging process of the lithium battery. In addition to the voltage, temperature, current, internal resistance, and capacity mentioned above, it also includes parameters such as SOC (state of charge) and SOH (state of health). These parameters reflect the performance and state of the battery from different angles. The sampling strategy adopts dynamic adaptive sampling, automatically adjusting the sampling frequency according to the change of the battery voltage. When the change amount of the battery voltage ΔV is greater than 50mV, in order to be able to capture the rapid change of the voltage more precisely, the sampling frequency is set to 10Hz; in other working conditions, the sampling frequency is adjusted to 1Hz, which can not only ensure obtaining sufficient data details during critical periods but also reduce the data volume under normal conditions and improve the data processing efficiency. In terms of noise processing, the wavelet threshold denoising method is adopted. Based on the multi-resolution analysis characteristics of wavelet transform, the signal is decomposed into different frequency sub-bands, and then according to the characteristic differences between noise and signal in different sub-bands, by setting appropriate thresholds, the sub-bands where the noise is located are removed or the noise is suppressed, thereby effectively reducing the interference of noise on the data and improving the data quality.
[0035] Step 2: Perform a fast Fourier transform on the time-domain signal to obtain frequency-domain features, specifically calculated through the formula for calculation, converting the time-domain signal into a frequency-domain signal, and then calculating the energy distribution of the signal at different frequencies through the formula Finally, extract the amplitudes of the first N main frequency components, determine the value of N according to actual requirements and signal characteristics, and extract the amplitudes of the main frequency components in the signal.
[0036] Among them, in terms of time-domain feature extraction, in addition to common statistical features such as mean, variance, skewness, and kurtosis, the change trend of the signal is reflected through the difference sequence, and the data is smoothed through moving average to remove short-term fluctuations and highlight the long-term trend. For example, calculating the difference sequence of voltage can clearly show the change amount of voltage at each time step, so as to judge whether the charge and discharge rate of the battery is stable; moving average can process the current data to make the current curve smoother and facilitate the analysis of its long-term change law.
[0037] Frequency-domain feature extraction mainly relies on FFT transformation and power spectral density analysis. Through FFT transformation, the time-domain signal is converted into a frequency-domain signal to obtain the frequency components of the signal. For example, after performing FFT transformation on the voltage signal of the battery, the voltage amplitudes at different frequencies can be obtained, and these amplitudes reflect the electrochemical reaction processes at different frequencies inside the battery. Power spectral density analysis further calculates the energy distribution of the signal at different frequencies. By analyzing the power spectral density, the main frequency components in the signal and the change of these frequency components over time can be determined, providing more information for in-depth understanding of the battery performance.
[0038] Time-frequency fusion adopts wavelet packet decomposition technology, which can decompose the signal more finely and comprehensively analyze the signal in the time domain and frequency domain. Wavelet packet decomposition obtains the time-domain signals of different frequency sub-bands by decomposing the signal multiple times, so that the time-domain features and frequency-domain features of the signal can be extracted simultaneously at different frequencies. For example, for the temperature signal of the battery, through wavelet packet decomposition, the change trend, fluctuation situation, etc. of the temperature can be analyzed on different frequency sub-bands, and then these features are fused to obtain a more comprehensive and accurate time-frequency domain feature representation.
[0039] In the present invention, a dynamic feature selection mechanism is introduced in the process of converting the time-domain signal into the frequency-domain signal. The principle of the dynamic feature selection mechanism is to dynamically adjust the weights of features according to the feature importance index (FI). In lithium battery data, different features have different influence degrees on the K value, and this influence degree may change with factors such as the usage state of the battery and environmental conditions. By calculating the importance index of each feature, and then according to the formula assign a weight w to each feature i . Among them, a is a control parameter used to adjust the change degree of the weight. By dynamically adjusting the feature weights, the model can more flexibly adapt to different battery states and data characteristics, and improve the prediction accuracy.
[0040] The hybrid loss function consists of the mean squared error (MSE), mean absolute error (MAE), and Huber loss function. MSE can measure the average squared error between the predicted value and the true value, imposing a greater penalty on larger errors; MAE measures the average absolute error between the predicted value and the true value, treating all errors equally; the Huber loss function combines the advantages of MSE and MAE, using MSE when the error is small and MAE when the error is large, thus being more robust to outliers. In the present invention, by setting different weights λ1, λ2, λ3, these three loss functions are linearly combined to obtain the hybrid loss function This hybrid loss function can comprehensively consider different types of errors, improving the stability and prediction accuracy of the model.
[0041] The present invention innovatively introduces a dynamic feature selection mechanism and a hybrid loss function. The dynamic feature selection mechanism dynamically adjusts the feature weights according to the importance index of the features, enabling the model to more flexibly adapt to different battery states and data characteristics, and improving the prediction accuracy; the hybrid loss function consists of the mean squared error (MSE), mean absolute error (MAE), and Huber loss function, comprehensively considering different types of errors, being more robust to outliers, and improving the stability and prediction accuracy of the model.
[0042] Step 3: Construct a hybrid prediction model including LSTM and LightGBM. LSTM is used to extract time series features, and LightGBM is used to further analyze and predict the features. Among them, the LSTM network includes a forget gate, which controls the retention and forgetting of information through the formula f t = σ(W f ·[h t-1, x t +b f ), where σ is the Sigmoid activation function, W f is the weight matrix, h t-1 is the hidden state at the previous moment, x t is the input at the current moment, and b f is the bias vector. LightGBM adopts a leaf-wise growth strategy, and each time it selects the leaf node with the largest split gain for splitting.
[0043] The hybrid prediction model architecture consists of a front end and a back end. Among them, the front end is a dual-channel LSTM Siamese network. In the time-domain branch of the dual-channel LSTM Siamese network, the 3-layer LSTM structure can extract the features of the time-domain signal layer by layer, from the basic features at the bottom layer to the abstract features at the high layer. Each layer's 128 units can fully learn the complex patterns and long-term dependencies in the signal. The 2-layer BiLSTM structure in the frequency-domain branch can better capture the context information in the frequency-domain signal and enhance the ability to extract frequency-domain features by processing forward and backward information simultaneously. For example, when processing the charge and discharge current signal of a battery, the LSTM in the time-domain branch can learn the variation law of the current over time and the dependencies between different time steps; the BiLSTM in the frequency-domain branch can analyze the frequency components and phase relationships of the current signal from the frequency domain perspective, so as to more comprehensively understand the characteristics of the current signal.
[0044] The back end is an optimized LightGBM model. The optimized LightGBM model adopts a feature importance weighting strategy. According to the importance of features for K-value prediction, different weights are assigned to each feature, making the model pay more attention to the features that have a greater impact on the prediction result. At the same time, the histogram acceleration algorithm is adopted to discretize the continuous feature values into histogram form, greatly reducing the computational amount and improving the training and prediction speed of the model. For example, when processing a large amount of battery data, the histogram acceleration algorithm can quickly perform statistics and analysis on the features, thus accelerating the training process of the model, while the feature importance weighting can enable the model to more accurately capture the key features related to the K value and improve the prediction accuracy.
[0045] The hybrid prediction model architecture combines the advantages of deep learning and traditional machine learning, aiming to achieve high-precision prediction of the K value of lithium batteries. The front end of this architecture adopts a dual-channel LSTM Siamese network. Among them, the time-domain branch consists of 3 layers of LSTM, each layer containing 128 units, which can effectively capture the long-term dependencies and dynamic change features in the time-domain signal; the frequency-domain branch consists of 2 layers of BiLSTM, each layer having 64 units, and can better extract the context information in the frequency-domain signal by considering the forward and backward information of the sequence simultaneously. The back end adopts an optimized LightGBM model, which further analyzes and processes the time-frequency domain features extracted by the front end through strategies such as feature importance weighting and histogram acceleration algorithm, so as to achieve accurate prediction of the K value of lithium batteries. This hybrid model architecture gives full play to the advantages of LSTM in processing sequence data and the strengths of LightGBM in processing large-scale data and making fast predictions, and can improve the prediction efficiency while ensuring the prediction accuracy.
[0046] Step 4: Adopt a time-frequency domain feature fusion strategy for K-value prediction, and fuse the time-domain features and frequency-domain features.
[0047] The time-frequency domain feature fusion engine is one of the cores of the present invention. Its main function is to deeply analyze and extract features from the collected data, and organically fuse the features in the time domain and the frequency domain. During the charging and discharging process of lithium batteries, the laws of signals such as voltage and current changing with time contain rich information. Through time domain analysis, statistical features such as mean, variance, skewness, and kurtosis can be extracted, as well as dynamic features such as difference sequences and moving averages. Frequency domain analysis, on the other hand, converts time domain signals into frequency domain signals through methods such as the fast Fourier transform (FFT), so as to extract frequency domain features such as power spectral density and frequency spectrum peaks. The time-frequency domain feature fusion engine uses technologies such as wavelet packet decomposition to fuse time domain and frequency domain features, fully mining the hidden information in the data, and providing more representative and discriminative feature vectors for the subsequent prediction model.
[0048] The present invention shows significant advantages in the prediction of the K value of lithium batteries. These advantages are fully reflected through the comparison with traditional LSTM models in terms of indicators such as RMS and false detection rate. In terms of the RMSE indicator, the RMSE value of the traditional LSTM model is 0.38, while the RMSE value of the present invention is reduced to 0.17. This means that the error between the predicted value and the true value of the present invention is smaller, and the prediction result is more accurate. Taking a set of actual lithium battery K value prediction data as an example, when the traditional LSTM model predicts the K value of a certain batch of lithium batteries, the average error reaches 0.38, resulting in deviations in the performance evaluation of some batteries; while the average prediction error of the present invention is only 0.17, which can more accurately reflect the actual K value of the battery, providing a more reliable basis for the quality inspection and screening of the battery.
[0049] In terms of the false detection rate, the false detection rate of the traditional LSTM model is 8.2%, while the present invention reduces the false detection rate to 2.1%, effectively reducing false judgments. In the quality inspection of lithium batteries, false detection may lead to misjudging qualified battery cells as unqualified, or misjudging unqualified battery cells as qualified, which will cause economic losses to enterprises. The lower false detection rate of the present invention can ensure the accuracy of the battery cell quality inspection, improve product quality, and enhance the market competitiveness of enterprises.
[0050] Embodiment 2: A lithium-ion battery K value prediction system based on time-frequency domain feature fusion
[0051] The present invention also provides a lithium-ion battery K value prediction system based on time-frequency domain feature fusion, including:
[0052] A data acquisition module, using a DAQ-2210 data acquisition card, for acquiring various parameters during the charging and discharging process of the battery, including voltage, temperature, and current;
[0053] The signal processing module uses the NIPXIe-5172 signal processing device to denoise and filter the collected signals;
[0054] The model inference module deploys the lithium-ion battery K value prediction method based on time-frequency domain feature fusion described in the above-mentioned Embodiment 1. According to the processed signals and the trained model, it predicts the K value of the lithium battery and outputs the prediction result.
[0055] In the in-line actual measurement experiment of this system, significant results have been achieved. The RMSE has been reduced to 0.17%. Compared with the traditional system, the prediction error has been greatly reduced, and it can more accurately reflect the K value of the lithium battery; the detection efficiency has been increased by 3 times, greatly improving the detection speed of the K value of the lithium battery on the production line and improving the production efficiency.
[0056] The above are the preferred embodiments of the present invention, but the present invention should not be limited to the content disclosed in this embodiment and the drawings. Therefore, all equivalent or modified implementations completed without departing from the spirit disclosed by the present invention fall within the protection scope of the present invention.
Claims
1. A method for predicting the K value of a lithium-ion battery based on the fusion of time-frequency domain features, characterized in that Specifically, it includes the following steps: S1. Collect the time-domain signals of the battery charging and discharging processes through multiple sensors to obtain the voltage, temperature, and current parameters of the battery during charging and discharging; S2. Perform a fast Fourier transform on the time-domain signal to obtain frequency-domain features, specifically calculated through the formula to convert the time-domain signal into a frequency-domain signal; S3. Construct a hybrid prediction model containing LSTM and LightGBM. LSTM is used to extract time-series features, and LightGBM is used to further analyze and predict the features; S4. Adopt a time-frequency domain feature fusion strategy for K-value prediction to fuse time-domain features and frequency-domain features.
2. The method for predicting the K value of a lithium-ion battery based on time-frequency domain feature fusion according to claim 1, wherein: In the step S2, after converting the time-domain signal into a frequency-domain signal, the energy distribution of the signal at different frequencies is calculated through the formula ; then the amplitudes of the first N main frequency components are extracted, and the value of N is determined according to actual requirements and signal characteristics, and the amplitudes of the main frequency components in the signal are extracted.
3. The method for predicting the K value of a lithium-ion battery based on time-frequency domain feature fusion according to claim 1, wherein: Introduce a dynamic feature selection mechanism during the process of converting a time-domain signal into a frequency-domain signal. By calculating the importance index of each feature, and then according to the formula assign a weight w to each feature i , where a is a control parameter used to adjust the degree of change of the weight.
4. The method for predicting the K value of a lithium-ion battery based on time-frequency domain feature fusion according to claim 1, wherein: Introduce a hybrid loss function during the process of converting the time-domain signal into a frequency-domain signal. By setting different weights λ1, λ2, and λ3, linearly combine the three loss functions of MSE, MAE, and Huber to obtain the hybrid loss function 5. The method for predicting the K value of a lithium-ion battery based on time-frequency domain feature fusion according to claim 1, wherein: In the said step S3, the LSTM network includes a forget gate, and the retention and forgetting of information are controlled through the formula f t =σ(W f ·[h t-1, x t +b f ), where σ is the Sigmoid activation function, W f is the weight matrix, h t-1 is the hidden state at the previous moment, x t is the input at the current moment, and b f is the bias vector.
6. The method for predicting the K value of a lithium-ion battery based on time-frequency domain feature fusion according to claim 1, wherein: In step S3, LightGBM adopts a leaf-wise growth strategy and selects the leaf node with the largest split gain for splitting each time.
7. The method for predicting the K value of a lithium-ion battery based on time-frequency domain feature fusion according to claim 1, wherein: In step S3, the time-domain branch consists of 3 layers of LSTM, with 128 units in each layer, which is used to capture the long-term dependence and dynamic change features in the time-domain signal; The frequency-domain branch consists of 2 layers of BiLSTM, with 64 units in each layer. By considering the forward and backward information of the sequence simultaneously, it extracts the context information in the frequency-domain signal.
8. The method for predicting the K value of a lithium-ion battery based on time-frequency domain feature fusion according to claim 1, wherein: In step S1, the multiple sensors include a voltage sensor, a temperature sensor, and a current sensor. The sampling strategy adopts dynamic adaptive sampling, and automatically adjusts the sampling frequency according to the change of the battery voltage.
9. A lithium-ion battery K value prediction system based on time-frequency domain feature fusion, characterized in that It includes: A data acquisition module, using a DAQ-2210 data acquisition card, which is used to collect various parameters during the battery charging and discharging processes, including voltage, temperature, and current; A signal processing module, using a NIPXIe-5172 signal processing device, which performs denoising and filtering processing on the collected signals; A model inference module, deploying the method of any one of claims 1-8, predicting the K value of the lithium battery according to the processed signals and the trained model, and outputting the prediction result.