A battery state of charge estimation method and device fusing ultrasonic sensing and transfer learning and a medium

Abstract

This invention discloses a method, device, and medium for estimating the state of charge (SOC) of a battery by integrating ultrasonic sensing and transfer learning. Belonging to the field of battery technology, the estimation method includes: acquiring ultrasonic signals, voltage, current, and temperature data of the battery during charging and discharging; extracting time-domain and frequency-domain features from the ultrasonic signals; constructing a deep learning network model including a feature extractor and a state estimator and training it using a source domain dataset; designing a transfer learning strategy based on physical feature mapping to transfer the weights of the battery source domain model to the target domain model, and introducing a physical feature distribution alignment mechanism between different battery domains; testing the target battery domain dataset based on the target domain model, and outputting the battery's SOC estimation result. The method disclosed in this invention solves the problem that existing ultrasonic sensing-based battery SOC estimation techniques do not meet the accuracy and robustness requirements for battery estimation in multiple domains.

Description

Technical Field

[0001] This invention belongs to the field of battery technology, and more specifically, relates to a method, device and medium for estimating the state of charge of a battery by integrating ultrasonic sensing and transfer learning. Background Technology

[0002] Lithium-ion batteries, as the core energy carrier in electric vehicles, energy storage systems, and portable electronic devices, directly impact the overall system performance through their safety, reliability, and lifecycle management. Among these, the State of Charge (SoC), a key parameter reflecting the remaining battery capacity, is crucial for precise energy management and extending battery lifespan. Currently, mainstream SoC estimation methods rely on acquiring external signals such as voltage, current, and temperature, and indirectly extrapolating them by constructing equivalent circuit or electrochemical models. However, these methods only characterize the battery state from a macroscopic electrical perspective, failing to deeply reflect the microscopic physical and chemical changes within the battery. This results in limited accuracy and insufficient robustness under complex operating conditions.

[0003] In recent years, ultrasonic sensing technology, with its non-invasive and highly sensitive characteristics, has provided a new technical approach for sensing the internal state of batteries. The propagation characteristics of ultrasonic signals inside batteries—including time-of-flight, amplitude attenuation, and spectral shift—can effectively characterize the evolution of electrode material interface properties, demonstrating great potential as a novel signal source for battery state monitoring. However, existing ultrasonic detection methods mostly rely on manually designed features for signal analysis, making it difficult to fully extract the high-dimensional time-frequency domain information contained within. Furthermore, due to the significant differences in the ultrasonic response distribution of batteries at different aging stages, with different chemical compositions, and under different ambient temperatures, using a single-domain model for SoC estimation easily leads to a decrease in model generalization ability and difficulty in guaranteeing estimation accuracy. Therefore, how to overcome the limitations of traditional ultrasonic signal processing methods and achieve high-precision and robust estimation of battery state of charge across operating conditions and life stages has become an important research topic in current intelligent battery management technology. Summary of the Invention

[0004] To address the shortcomings and improvement needs of existing technologies, this invention provides a method and apparatus for estimating the state of charge (SoC) of a battery that integrates ultrasonic sensing and transfer learning. The purpose is to solve the problem that existing ultrasonic sensing-based battery SoC estimation technologies do not meet the accuracy and robustness requirements for battery estimation in multiple domains.

[0005] To achieve the above objectives, according to one aspect of the present invention, a method for estimating the state of charge of a battery integrating ultrasonic sensing and transfer learning is provided, comprising: S1: Collect ultrasonic signals, voltage, current and temperature data of the battery during charging and discharging, and extract time-domain and frequency-domain features from the ultrasonic signals; S2: Construct a deep learning network model that includes a feature extractor and a state estimator and train it using the source domain dataset; S3: Design a transfer learning strategy based on physical feature mapping, transfer the weights of the battery source domain model in S2 to the target domain model, and introduce a physical feature distribution alignment mechanism between different battery domains; S4: Based on the target domain model obtained in S3, input the target battery domain dataset for testing, and output the battery state of charge estimation results.

[0006] In summary, the above-described technical solutions conceived in this invention can achieve the following beneficial effects: First, in terms of multi-source information fusion and deep feature extraction, this invention breaks through the limitations of traditional SoC estimation methods that rely solely on macroscopic electrical parameters such as voltage, current, and temperature, and innovatively introduces ultrasonic sensing information into the estimation model. By constructing a deep feature extractor composed of a convolutional neural network (CNN) and a bidirectional long short-term memory network (Bi-LSTM), this invention can simultaneously perform local pattern mining and temporal coupling feature modeling on multi-source input signals, fully extracting the internal and external state information of the battery under different operating conditions, and providing a richer and more reliable data foundation for subsequent high-precision SoC estimation.

[0007] Secondly, regarding cross-domain transfer estimation and model generalization capabilities, this invention addresses the issue of decreased estimation accuracy of traditional models under different environmental temperatures, chemical compositions, and aging conditions by introducing a transfer learning strategy and designing an efficient parameter fine-tuning mechanism. During the transfer process, by freezing the CNN layers and fully connected layers in the source domain model that have learned general features, only the Bi-LSTM layer responsible for temporal modeling is updated, enabling the model to quickly adapt to new conditions driven by a small number of labeled samples in the target domain. This mechanism not only effectively solves the problem of distribution differences in cross-domain scenarios and significantly improves the accuracy and robustness of SoC estimation, but also avoids the risk of overfitting the model in new tasks.

[0008] Third, regarding small-sample adaptation and engineering practicality, this invention can achieve efficient model adaptation even when there are only a very small number of labeled samples in the target domain. This significantly reduces the reliance on data collection and manual annotation under new conditions, saving considerable time and economic costs. Furthermore, the fine-tuning transfer strategy algorithm is simple and fast, meeting the real-time and reliability requirements of estimation results in practical applications, and possesses good deployability and broad engineering application prospects.

[0009] In summary, this invention integrates ultrasonic sensing and transfer learning to construct a high-precision, robust, and low-data-dependency battery SoC estimation method, providing effective technical support for improving the performance of intelligent battery management systems. Attached Figure Description

[0010] Figure 1 A flowchart of a battery state-of-charge estimation method integrating ultrasonic sensing and transfer learning provided in an embodiment of the present invention; Figure 2 A schematic diagram of the time domain and frequency domain of the ultrasonic signal after Fast Fourier Transform provided in an embodiment of the present invention; Figure 3 A schematic diagram of a deep learning network model provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the model transfer learning process provided in an embodiment of the present invention; Figure 5 This is a comparison of SoC estimation results curves on the target domain dataset using different migration schemes when migrating a lithium battery from 25 degrees Celsius to 40 degrees Celsius under UDDS conditions, as provided in the embodiments of the present invention. Detailed Implementation

[0011] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0012] In this invention, the terms "first," "second," etc. (if present) in the invention and the accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0013] This invention discloses a battery state-of-charge estimation method that integrates ultrasonic sensing and transfer learning, such as... Figure 1 As shown, it includes: S1: Collect ultrasonic signals, voltage, current, and temperature data of the battery during charging and discharging, and extract time-domain and frequency-domain features from the ultrasonic signals. It should be noted that the ultrasonic signals here are signals generated by the battery during charging and discharging through ultrasonic acquisition. The ultrasonic signals directly reflect the evolution of key physical characteristics inside the battery, which is reflected in the changes in parameters such as ultrasonic signal flight time, amplitude / energy attenuation, and spectral shift.

[0014] Furthermore, time-domain and frequency-domain feature extraction is performed on the ultrasonic signal, specifically including: denoising and normalizing the ultrasonic signal s(t) to obtain time-domain and frequency-domain signal features. The time-domain features include the time of flight, representing the propagation time of the ultrasonic signal from transmission to reception. The time of flight is defined as: ,in, For launch time, The arrival time of the main peak of the received signal; the peak amplitude of the ultrasonic signal is defined as: Ultrasonic signal energy, used to measure the overall intensity of a time-domain signal, is defined as: ,in Indicate the next deadline. Indicates the upper cutoff time; the frequency domain characteristics of ultrasonic signals include: for ultrasonic signals s(t) Performing a Fast Fourier Transform yields the following spectrum function: The amplitude of the main frequency is defined as: Spectral energy is defined as: , f 1 indicates the lower cutoff frequency. f 2 indicates the upper cutoff frequency.

[0015] S2: Construct a deep learning network model that includes a feature extractor and a state estimator and train it using the source domain dataset; Specifically, such as Figure 3 As shown, the deep learning network model described above includes a feature extractor and a state estimator. The feature extractor uses a one-dimensional convolutional neural network and a bidirectional long short-term memory network to learn the deep coupling features of ultrasonic signals, voltage, current, and temperature data; the state estimator uses a fully connected network to capture the time-series dependence of battery state and regress the output battery state of charge estimate.

[0016] Specifically, a one-dimensional convolutional neural network consists of convolutional layers and pooling layers, and its calculation formula is as follows:

[0017] in, : No. l The first layer k Each feature map in time t The output value at that location, Indicates the first Layer in channel On time t The input signal at the location, : No. l The first layer k Each feature map in time t The output intermediate value at that point, Indicates the first Layer in channel On time t’ The final output after pooling, For the first Layer The weights of each convolutional kernel, For bias terms, p For pooling step size, Related to the padding method, BatchNorm represents normalization operation, and ReLU represents non-linear activation operation. That is, the output of the convolutional layer is processed by batch normalization operation and non-linear activation function operation before entering the pooling layer.

[0018] The bidirectional Long Short-Term Memory (LSTM) network consists of a forward LSTM and a backward LSTM, used to further extract forward and backward sequence dependency features in the time dimension, realizing dynamic modeling of the signal on a global time scale. Its calculation formula is as follows:

[0019] in, For the Sigmoid function, This represents the Hadamard product. 、 、 These are the forget gate, input gate, and output gate, respectively. For the state of the memory unit, For output of the hidden layer, W and b These are the weight matrix and bias terms that the model needs to learn.

[0020] The formula for calculating a fully connected network is: .in, Represents the input vector. and These are the weight matrices and bias terms for each layer. This is the SoC estimate of the battery's state of charge output by the network.

[0021] During training, input samples from the source domain dataset and their corresponding real SoC labels are used for supervised training based on the mean squared error loss function, which is defined as follows:

[0022] in, For the sample size, These are the model's predicted values. This represents the actual SoC value. The network parameters are updated using backpropagation and optimization algorithms until the loss function converges.

[0023] Those skilled in the art should understand that the convolutional neural networks and bidirectional long short-term memory networks listed above are only one implementation method and are not limitations of the present invention. Other neural networks that can achieve feature learning and bidirectional long short-term memory networks are also applicable.

[0024] S3: Design a transfer learning strategy based on physical feature mapping, transfer the weights of the battery source domain model in step S2 to the target domain model, and introduce a physical feature distribution alignment mechanism between different battery domains; Specific transfer learning strategies based on physical feature mapping include introducing ultrasonic signals as the core feature mapping method. When ultrasound propagates inside a power battery, its acoustic parameters (such as sound velocity and amplitude) can highly sensitively reflect the evolution of Young's modulus of the graphite anode under different temperatures, aging states, and chemical compositions. This strong physical correlation makes ultrasonic features highly transferable between battery systems of different types. By capturing Young's modulus, a common physical indicator, a mapping relationship between the source and target domains can be effectively established, thereby significantly reducing the feature search space of the target domain model during the transfer process and limiting model complexity.

[0025] To better understand the process of the transfer learning strategy, the process will be described in detail, such as... Figure 4 As shown, the weights of the source domain model trained in step S2 are first fully transferred to the target domain model to retain the underlying general feature representations learned from the large-scale data in the source domain (such as the fundamental frequency features of ultrasonic waves, signal envelopes, etc.). To reduce the need for labeled samples in the target domain while ensuring the stability of the underlying physical features, a layered freezing strategy is adopted. Specifically, the parameters of the convolutional neural network layer and the fully connected layer remain unchanged, locking the captured fundamental acoustic spatial features, and only setting the bidirectional long short-term memory network layer to a trainable state. Using a small number of labeled samples in the target domain, the bidirectional long short-term memory network layer is fine-tuned to capture the temporal evolution of the Young's modulus of the graphite anode with charge-discharge cycles. By adjusting the parameters of this layer, feature distribution alignment under specific electrochemical systems in the target domain is achieved, ensuring that the model still possesses high-precision state estimation capabilities under small sample conditions.

[0026] Through the aforementioned transfer learning strategy, cross-system battery state estimation with "small sample size and high accuracy" is achieved. Compared with the comparative model, the method of this invention significantly reduces cross-domain estimation error, verifying the effectiveness of introducing ultrasonic features and transfer learning mechanisms in improving model robustness and generalization ability.

[0027] S4: Based on the target domain model obtained in step S3, input the target battery domain dataset for testing, and output the battery state of charge estimation results.

[0028] Specifically, the real-time ultrasonic signal, voltage, current and temperature data of the target battery under test are preprocessed and feature extracted according to the method in step S1; the extracted features are input into the target domain model after transfer learning in step S3; the target domain model performs forward propagation calculation through its feature extractor and state estimator, and outputs the state of charge estimation result of the target battery under test in real time.

[0029] It should be noted that the source domain dataset and the target domain dataset are based on data collected from batteries under standard operating conditions (25°C, 100% SOH) through dynamic operating conditions (including Dynamic Stress Test Cycle (DST), Federal City Driving Cycle (FUDS), Urban Driving Cycle (UDDS), Federal Test Procedure (FTP), Highway Fuel Economy Test Cycle (HWFET), New York City Cycle (NYC), Los Angeles City Cycle (LA92), and US High-Speed ​​Acceleration Cycle (US06), etc.). The target domain dataset is based on data collected from batteries under different ambient temperatures (e.g., 10°C or 40°C), different aging levels (95% SOH or 90% SOH), and different chemical compositions (e.g., LFP or NMC) through the same dynamic operating conditions and three constant current discharge (0.5CC, 1CC, 2CC) operating conditions. The number of labeled samples in the target domain dataset is much smaller than that in the source domain dataset.

[0030] In summary, this invention provides a battery state-of-charge (SoC) estimation method that integrates ultrasonic sensing and transfer learning, achieving high-precision and highly generalizable SoC estimation under varying environmental temperatures, chemical compositions, and aging levels. By introducing ultrasonic sensing technology, acoustic response information of internal battery state changes is obtained, making the model more sensitive to electrochemical and structural changes at the feature level. Furthermore, by combining deep learning and transfer learning strategies, adaptive alignment of feature distributions between the source and target domains is achieved, significantly improving the accuracy and stability of the model in cross-domain prediction scenarios.

[0031] To further illustrate the battery state-of-charge estimation method integrating ultrasonic sensing and transfer learning provided by this invention, a specific embodiment is described in detail below: This embodiment uses a lithium iron phosphate battery as the research object, with an actual capacity of 1.3Ah, operating temperatures of 25°C and 40°C, and charge / discharge cutoff voltages of 3.65V and 2.5V, respectively. During the test, voltage, current, temperature, and ultrasonic signal data were simultaneously collected using a battery testing system. The ultrasonic transducer used had a center frequency of 2–3MHz and a sampling frequency of 10MHz. The source domain dataset consisted of data from eight typical dynamic operating conditions collected at 25°C, including DST, FUDS, UDDS, FTP, HWFET, NYC, LA92, and US06 conditions. The target domain dataset consisted of the same eight dynamic operating conditions and three constant current operating conditions (0.5CC, 1CC, and 2CC) collected at 40°C. The labeled dataset for the target domain included the three constant current operating conditions, used for model fine-tuning during the transfer learning phase. The eight dynamic operating conditions in the target domain served as an unlabeled test set to verify the model's cross-domain generalization performance.

[0032] like Figure 2 As shown, Figure 2 Figure (a) shows the time-domain waveforms of the ultrasonic signal of the battery at 0% SoC and 100% SoC. Figure 2 (b) in the diagram represents the corresponding frequency domain signal. The main frequencies where the frequency domain peaks are selected based on the frequency domain diagram are 2.281MHz and 2.424MHz. Figure 2 As shown, with the change of SoC, the time domain peak and main frequency peak of the ultrasonic signal both change significantly, indicating that the internal acoustic characteristics of the battery change significantly with the state of charge.

[0033] The deep learning model structure in this invention is as follows: Figure 3 As shown, it consists of two parts: a feature extractor and a state estimator. The input data is a normalized multimodal feature vector, which includes the time-domain features of the ultrasound signal (e.g., time-to-flight (TTOF), peak amplitude Amax, signal energy Et), frequency-domain features (e.g., amplitude of dominant frequency 1 Af1, amplitude of dominant frequency 2 Af2, spectral energy Ef), voltage, current, and temperature signals.

[0034] The feature extractor consists of a one-dimensional convolutional neural network and a bidirectional long short-term memory network, used to learn deeply coupled features from multi-source inputs. The one-dimensional convolutional neural network is responsible for extracting local temporal structure features and pattern distributions, gradually capturing the higher-order representations of features through convolution and pooling operations to obtain common features in different domains; the bidirectional long short-term memory network further explores the time-series dependencies, realizing dynamic modeling of the battery's state of charge (SOC) change trend. The state estimator adopts a multi-layer fully connected network structure, integrating the temporal features output by the feature extractor and regressing the SOC at the corresponding time moment.

[0035] During the model training phase, the model was first pre-trained using the source domain dataset. 80% of the samples in the source domain dataset were used for training, and 20% were used for validation. The mean squared error (MSE) loss function was used, and the Adam optimization algorithm was employed. The search space for hyperparameters was defined in Table 1, and tuning was performed using the OPTUNA automatic hyperparameter tuning platform. The optimal values ​​were determined based on the performance on the validation set. Finally, the initial learning rate of the source domain model was set to 5 × 10⁻⁶. - The batch size is 512, and the number of training epochs is 300. An early stopping strategy is employed during training: training automatically stops when the validation set loss no longer decreases for 20 consecutive epochs to prevent overfitting. The convolutional neural network consists of one convolutional layer with 32 kernels and a kernel size of 3; a one-layer bidirectional long short-term memory network with 64 hidden units; and a two-layer fully connected network with 32 and 1 nodes respectively, using ReLU activation and linear activation for the output layer. After model training, the optimal weight parameters are saved for use in the transfer learning phase.

[0036] Table 1 Hyperparameter Search Space

[0037] In the transfer learning phase, such as Figure 4 As shown, the model parameters trained in the source domain are transferred to the target domain model. Specifically, the parameters of the convolutional and fully connected layers in the feature extractor are retained and frozen, and only the weights of the bidirectional long short-term memory network layers are fine-tuned to adapt to the differences in the distribution characteristics of the target domain. During transfer training, mean squared error is still used as the loss function to achieve feature space distribution alignment, thereby improving the model's cross-domain generalization performance. Three labeled constant-current conditions in the target domain dataset are used for feature distribution alignment and model fine-tuning, with 80% of the samples used for training and 20% for validation. The initial learning rate of the target domain model is set to 2×10⁻⁴, the batch size is 512, and the number of training epochs is 100.

[0038] To verify the effectiveness of the method of this invention, two comparative models were designed under the same test conditions. The first model is one without transfer learning, where the model trained in the source domain is directly used for state of charge estimation of the target domain data without fine-tuning using labeled small sample data. The second model uses only traditional features, i.e., no ultrasonic signal is input, only voltage, current, and temperature signals are used as model input, and the transfer learning process is consistent with the method of this invention. The hyperparameter settings of the three models are the same, the only difference being whether transfer learning is performed and whether ultrasonic features are introduced.

[0039] This invention embodiment selects UDDS working conditions from the target domain dynamic working condition dataset for demonstration, and the test results are as follows: Figure 5As shown, the results indicate that the mean absolute error (MAE) of the method described in this invention is 1.071%, and the root mean square error (RMSE) is 1.398%, while the MAE of the model without transfer learning is 8.959% and the RMSE is 11.765%, and the MAE of the traditional feature model is 4.454% and the RMSE is 5.829%. Compared with the comparative model, the method described in this invention improves the estimation accuracy by approximately 88% and 76%, respectively, and significantly reduces the cross-domain estimation error, verifying the effectiveness of introducing ultrasonic features and transfer learning mechanisms in improving the robustness and generalization ability of the model.

[0040] Example 2: This invention discloses an apparatus for implementing the battery state-of-charge estimation method integrating ultrasonic sensing and transfer learning in Embodiment 1, comprising: Acquisition module: Used to acquire ultrasonic signals, voltage, current and temperature data of the battery during charging and discharging, and to extract time-domain and frequency-domain features of the ultrasonic signals; Training module: Used to build deep learning network models including feature extractors and state estimators and train them using source domain datasets; Transfer module: Used to design transfer learning strategies based on physical feature mapping, transfer the weights of the battery source domain model to the target domain model, and introduce a physical feature distribution alignment mechanism between different battery domains; Output module: Based on the target domain model, input the target battery domain dataset for testing, and output the battery's state of charge estimation results.

[0041] The device in this embodiment is used to implement the battery state of charge estimation method that integrates ultrasonic sensing and transfer learning in Embodiment 1. Therefore, the specific implementation of this device can be found in the embodiment section of the battery state of charge estimation method that integrates ultrasonic sensing and transfer learning mentioned above. The specific implementation can be referred to the description of the corresponding embodiments, and will not be elaborated here.

[0042] Example 3: The present invention also provides a computer storage medium, wherein the storage medium may be a magnetic disk, an optical disk, or a read-only memory (ROM). ROM (ROM) or RAM (Random Access Memory) are all types of memory.

[0043] The computer storage medium stores a program for a battery state of charge estimation method that integrates ultrasonic sensing and transfer learning. When the program is executed by a processor, it implements the flow steps of the above-described embodiment of the battery state of charge estimation method program that integrates ultrasonic sensing and transfer learning.

[0044] Those skilled in the art will clearly understand that the techniques in the embodiments of the present invention can be implemented using software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solutions in the embodiments of the present invention, or the parts that contribute to the prior art, can be embodied in the form of a software product, which is stored in a storage medium such as a USB flash drive, external hard drive, or read-only memory (ROM). The method includes various media capable of storing program code, such as only memory, random access memory (RAM), magnetic disks or optical disks, and several instructions to cause a computer terminal (which may be a personal computer, server, or a second terminal, network terminal, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0045] In summary, this invention discloses a battery state-of-charge (SoC) estimation method, device, and medium that integrates ultrasonic sensing and transfer learning. It overcomes the limitations of traditional SoC estimation methods that rely solely on macroscopic electrical parameters such as voltage, current, and temperature, innovatively introducing ultrasonic sensing information into the estimation model. By constructing a deep feature extractor composed of a convolutional neural network (CNN) and a bidirectional long short-term memory network (Bi-LSTM), this invention can simultaneously perform local pattern mining and temporal coupling feature modeling on multi-source input signals, fully extracting the internal and external state information of the battery under different operating conditions, providing a richer and more reliable data foundation for subsequent high-precision SoC estimation.

[0046] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for estimating the state of charge of a battery by integrating ultrasonic sensing and transfer learning, characterized in that, include: S1: Collect ultrasonic signals, voltage, current and temperature data of the battery during charging and discharging, and extract time-domain and frequency-domain features from the ultrasonic signals; S2: Construct a deep learning network model that includes a feature extractor and a state estimator and train it using the source domain dataset; S3: Design a transfer learning strategy based on physical feature mapping, transfer the weights of the battery source domain model in S2 to the target domain model, and introduce a physical feature distribution alignment mechanism between different battery domains; S4: Based on the target domain model obtained in S3, input the target battery domain dataset for testing, and output the battery state of charge estimation results.

2. The battery state-of-charge estimation method integrating ultrasonic sensing and transfer learning according to claim 1, characterized in that, The evolution of the ultrasonic signal in S1 is related to the key physical properties inside the battery, which is reflected in the changes in the ultrasonic signal flight time, amplitude / energy attenuation and spectral shift parameters.

3. The battery state-of-charge estimation method integrating ultrasonic sensing and transfer learning according to claim 1, characterized in that, Time-domain function of ultrasonic signal in S1 s(t) Performing a Fast Fourier Transform, the spectral function is obtained as follows: The amplitude of the main frequency is: The spectral energy is: , f 1 indicates the lower cutoff frequency. f 2 indicates the upper cutoff frequency.

4. The battery state-of-charge estimation method integrating ultrasonic sensing and transfer learning according to claim 1, characterized in that, In S2, the feature extractor uses a one-dimensional convolutional neural network and a bidirectional long short-term memory network to learn the deep coupling features of ultrasonic signals, voltage, current and temperature data; the state estimator uses a fully connected network to capture the time-series dependence of battery state and regress the output battery state of charge estimate.

5. The battery state-of-charge estimation method integrating ultrasonic sensing and transfer learning according to claim 4, characterized in that, A one-dimensional convolutional neural network consists of convolutional layers and pooling layers, and its calculation formula is as follows: in, : No. l The first layer k Each feature map in time t The output value at that location, Indicates the first Layer in channel On time t The input signal at the location, : No. l The first layer k Each feature map in time t The output intermediate value at that point, Indicates the first Layer in channel On time t’ The final output after pooling, For the first Layer The weights of each convolutional kernel, For bias terms, p For pooling step size, Depending on the filling method, BatchNorm represents normalization operation, and ReLU represents non-linear activation operation.

6. The battery state-of-charge estimation method integrating ultrasonic sensing and transfer learning according to claim 4, wherein the bidirectional long short-term memory network consists of a forward LSTM and a backward LSTM, used to further extract forward and backward sequence dependency features in the time dimension, realizing dynamic modeling of the signal on a global time scale, and its calculation formula is as follows: in, For the Sigmoid function, This represents the Hadamard product. 、 、 These are the forget gate, input gate, and output gate, respectively. For the state of the memory unit, For output of the hidden layer, W and b These are the weight matrix and bias terms that the model needs to learn.

7. The battery state-of-charge estimation method integrating ultrasonic sensing and transfer learning according to claim 1, characterized in that, The transfer learning strategy based on physical feature mapping includes: introducing ultrasound signals as the core feature mapping method, and establishing a mapping relationship between the source domain and the target domain by capturing the physical index of Young's modulus.

8. An apparatus for estimating the state of charge of a battery by integrating ultrasonic sensing and transfer learning, characterized in that, include: Acquisition module: Used to acquire ultrasonic signals, voltage, current and temperature data of the battery during charging and discharging, and to extract time-domain and frequency-domain features of the ultrasonic signals; Training module: Used to build deep learning network models including feature extractors and state estimators and train them using source domain datasets; Transfer module: Used to design transfer learning strategies based on physical feature mapping, transfer the weights of the battery source domain model to the target domain model, and introduce a physical feature distribution alignment mechanism between different battery domains; Output module: Based on the target domain model, input the target battery domain dataset for testing, and output the battery's state of charge estimation results.

9. A computer-readable storage medium, characterized in that, A readable storage medium stores a battery state of charge estimation method program that integrates ultrasonic sensing and transfer learning. When executed by a processor, the battery state of charge estimation method program that integrates ultrasonic sensing and transfer learning implements the steps of the battery state of charge estimation method that integrates ultrasonic sensing and transfer learning as claimed in any one of claims 1 to 7.

Images