Battery health state estimation model training and using method, device and equipment

Abstract

The invention discloses a model training method and device for battery health state estimation, equipment, a medium and a program product, and belongs to the field of battery management, and the method comprises the steps: obtaining a battery data set; taking the battery sensing signal under the first working condition as source domain input data, and extracting source domain features through an adaptive feature extraction network; taking the battery sensing signal under the second working condition as target domain input data, and extracting target domain features through an adaptive feature extraction network; calculating a distribution difference between the source domain feature and the target domain feature through a selective domain adaptive network; and training the adaptive feature extraction network and the selective domain adaptive network in order to enable the adaptive feature extraction network to extract domain invariant features in the source domain and the target domain by taking reduction of the distribution difference as a target. According to the method, on a server which can be a cloud server, a model is trained under multiple working conditions, domain invariant features are extracted, and battery health state estimation under the multiple working conditions is achieved.

Description

Technical Field

[0001] Embodiments of the present application relate to the field of battery management, and in particular to a method, apparatus, device, medium, and program product for training and using a model for estimating the health status of a battery. Background Art

[0002] Batteries, as a rechargeable power source, are an important component of many systems such as electric vehicles and smartphones. During the cycle of charge and discharge, the health of the battery tends to gradually deteriorate.

[0003] In the related art, a battery health state estimation model is established to estimate the battery health state, thereby determining whether the current battery health state is good.

[0004] However, the battery health status estimation model in the related art usually assumes that the battery operates under the same working conditions (ie, the same operating condition), which does not conform to the actual battery working conditions and may lead to errors in the judgment of the current health status of the battery. Summary of the invention

[0005] The present application provides a method, apparatus, device, medium and program product for model training and use of battery health state estimation, and the technical solution at least includes:

[0006] According to one aspect of an embodiment of the present application, a model training method for battery health state estimation is provided, the model includes an adaptive feature extraction network and a selective domain adaptation network, and the method includes:

[0007] Acquire a battery data set, the battery data set including battery sensor signals under at least two working conditions;

[0008] The battery sensor signal under the first working condition is used as the source domain input data, and the source domain features are extracted by the adaptive feature extraction network; the battery sensor signal under the second working condition is used as the target domain input data, and the target domain features are extracted by the adaptive feature extraction network;

[0009] The distribution difference between source domain features and target domain features is calculated through a selective domain adaptation network;

[0010] With the goal of reducing the distribution difference, the adaptive feature extraction network and the selective domain adaptation network are trained so that the adaptive feature extraction network extracts domain-invariant features in the source domain and the target domain;

[0011] Among them, domain-invariant features are used to represent the characteristics of the battery health status without being affected by working conditions.

[0012] According to another aspect of an embodiment of the present application, a method for using a model for estimating a battery health state is provided, wherein the model includes an adaptive feature extraction network and a battery health state estimation network, and the method includes:

[0013] Acquire a battery data set, the battery data set including a battery sensor signal under a second operating condition;

[0014] The battery sensor signal under the second working condition is used as the target domain input data, and the domain invariant features are extracted through an adaptive feature extraction network;

[0015] Predict the battery health state based on domain-invariant features through the battery health state estimation network;

[0016] Among them, domain-invariant features are used to represent the characteristics of the battery health status without being affected by working conditions.

[0017] According to another aspect of an embodiment of the present application, a model training device for battery health state estimation is provided, wherein the model includes an adaptive feature extraction network and a selective domain adaptation network, and the device includes:

[0018] An acquisition module, used to acquire a battery data set, where the battery data set includes battery sensor signals under at least two working conditions;

[0019] An extraction module, used to use the battery sensor signal under the first working condition as source domain input data, and extract the source domain features through an adaptive feature extraction network; use the battery sensor signal under the second working condition as target domain input data, and extract the target domain features through an adaptive feature extraction network;

[0020] A computation module for computing the distribution difference between source domain features and target domain features through a selective domain adaptation network;

[0021] A training module, for training the adaptive feature extraction network and the selective domain adaptation network with the goal of reducing the distribution difference, so that the adaptive feature extraction network extracts domain-invariant features in the source domain and the target domain;

[0022] Among them, domain-invariant features are used to represent the characteristics of the battery health status without being affected by working conditions.

[0023] According to another aspect of an embodiment of the present application, a device for using a model for estimating a battery health state is provided, wherein the model includes an adaptive feature extraction network and a battery health state estimation network, and the device includes:

[0024] An acquisition module, used for acquiring a battery data set, where the battery data set includes a battery sensor signal under a second working condition;

[0025] an extraction module, configured to use the battery sensor signal under the second working condition as target domain input data and extract domain invariant features through an adaptive feature extraction network;

[0026] A prediction module, used to predict the battery health state based on domain invariant features through a battery health state estimation network;

[0027] Among them, domain-invariant features are used to represent the characteristics of the battery health status without being affected by working conditions.

[0028] According to another aspect of an embodiment of the present application, a computer device is provided, which includes: a processor and a memory, wherein at least one program is stored in the memory; the processor is used to execute at least one program in the memory to implement the model training and use methods of battery health status estimation in the above-mentioned various aspects.

[0029] According to another aspect of an embodiment of the present application, a computer-readable storage medium is provided, in which at least one program is stored, and the at least one program is loaded and executed by a processor to implement model training and use methods for battery health status estimation as described in the above-mentioned aspects.

[0030] According to another aspect of an embodiment of the present application, a computer program product (or computer program) is provided, which includes computer instructions stored in a computer-readable storage medium, a processor obtains the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions to implement model training and use methods for battery health status estimation as described in the above-mentioned aspects.

[0031] The technical solution provided by the embodiments of the present application may have the following beneficial effects:

[0032] By training the adaptive feature extraction network and the selective domain adaptation network under multiple working conditions, the adaptive feature extraction network extracts domain-invariant features for estimating the battery health state, thereby realizing battery health state estimation under multiple working conditions and improving the versatility of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0034] Figure 1 A flowchart of a model training method for battery health state estimation provided by an exemplary embodiment of the present application is shown;

[0035] Figure 2 A flowchart of a model training method for battery health state estimation provided by an exemplary embodiment of the present application is shown;

[0036] Figure 3 A schematic diagram of a model training method for battery health state estimation provided by an exemplary embodiment of the present application is shown;

[0037] Figure 4 A schematic diagram of an adaptive feature extraction network provided by an exemplary embodiment of the present application is shown;

[0038] Figure 5 A schematic diagram of an adaptive feature transmission network provided by an exemplary embodiment of the present application is shown;

[0039] Figure 6 A schematic diagram of a selective domain adaptation network provided by an exemplary embodiment of the present application is shown;

[0040] Figure 7 A flowchart of a model training method for battery health state estimation provided by an exemplary embodiment of the present application is shown;

[0041] Figure 8 A flowchart of a model training method for battery health state estimation provided by an exemplary embodiment of the present application is shown;

[0042] Fig. 9 A schematic diagram of a model training method for battery health state estimation provided by an exemplary embodiment of the present application is shown;

[0043] Fig.10 A flowchart showing a method for using a model for estimating a battery health state provided by an exemplary embodiment of the present application is shown;

[0044] Fig.11 A schematic diagram showing a method for using a model for estimating a battery health state provided by an exemplary embodiment of the present application is shown;

[0045] Fig.12 A flowchart showing a method for using a model for estimating a battery health state provided by an exemplary embodiment of the present application is shown;

[0046] Fig.13 A block diagram of a model training device for battery health state estimation provided by an exemplary embodiment of the present application is shown;

[0047] Fig.14 A block diagram of a model using device for battery health state estimation provided by an exemplary embodiment of the present application is shown;

[0048] Fig.15 A schematic diagram of the structure of a computer device provided by an exemplary embodiment of the present application is shown. DETAILED DESCRIPTION

[0049] In order to make the objectives, technical solutions and advantages of the present application clearer, the implementation methods of the present application will be further described in detail below with reference to the accompanying drawings.

[0050] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The implementations described in the following exemplary embodiments do not represent all implementations consistent with the present application. Instead, they are merely examples of devices and methods consistent with some aspects of the present application as detailed in the appended claims.

[0051] The terms used in this application are for the purpose of describing specific embodiments only and are not intended to limit this application. The singular forms of "a", "said" and "the" used in this application and the appended claims are also intended to include plural forms unless the context clearly indicates other meanings. It should also be understood that the term "and / or" used herein refers to and includes any or all possible combinations of one or more associated listed items.

[0052] It should be noted that the object information (including but not limited to object device information, object personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the object or fully authorized by all parties, and the collection, use and processing of relevant data need to comply with relevant laws, regulations and standards of relevant countries and regions.

[0053] It should be understood that, although the terms first, second, etc. may be used in the present application to describe various information, these information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of the present application, the first parameter may also be referred to as the second parameter, and similarly, the second parameter may also be referred to as the first parameter. Depending on the context, the word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining".

[0054] In this application, a server can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms. The terminal can be a smart phone, tablet computer, laptop computer, desktop computer, smart speaker, smart watch, etc., but is not limited to this. The terminal and the server can be directly or indirectly connected via wired or wireless communication, which is not limited in this application.

[0055] Cloud computing is a computing model that distributes computing tasks on a resource pool composed of a large number of computers, allowing various application systems to obtain computing power, storage space and information services as needed. The network that provides resources is called a "cloud". From the user's perspective, the resources in the "cloud" are infinitely scalable and can be obtained at any time, used on demand, expanded at any time, and paid for by use.

[0056] As a cloud computing basic capability provider, a cloud computing resource pool platform will be established, referred to as a cloud platform, generally referred to as Infrastructure as a Service (IaaS), and various types of virtual resources will be deployed in the resource pool for external customers to choose to use. The cloud computing resource pool mainly includes: computing devices (virtualized machines, including operating systems), storage devices, and network devices.

[0057] According to the logical function division, the Platform as a Service (PaaS) layer can be deployed on the IaaS layer, and the Software as a Service (SaaS) layer can be deployed on the PaaS layer, or the SaaS layer can be directly deployed on the IaaS layer. The PaaS layer is a platform for software operation, such as databases, web containers, etc. The SaaS layer is a variety of business software, such as web portals, SMS mass senders, etc. Generally speaking, the SaaS layer and PaaS layer are upper layers relative to the IaaS layer.

[0058] As a rechargeable power source, batteries are an important component of many systems such as electric vehicles, smartphones, communication base stations, and data centers. During the cyclic charging and discharging process, the health status of the battery tends to gradually deteriorate. In related technologies, a battery health status estimation model is established to estimate the health status of the battery, thereby determining whether the current health status of the battery is good.

[0059] However, the battery health state estimation model in the related art usually assumes that the battery operates under the same working conditions (i.e., the same working condition), which does not conform to the actual battery working conditions. In reality, the battery usually operates under multiple working conditions. This battery health state estimation model applicable to the same working condition will lead to errors in the judgment of the current health state of the battery. Therefore, a method is needed to estimate the battery health state under multiple working conditions.

[0060] Figure 1 A flowchart of a training method for battery health state estimation provided by an exemplary embodiment of the present application is shown. The method uses a battery health state estimation model 100 (referred to as model 100) proposed in an embodiment of the present application. The model 100 includes: at least one of an unsupervised domain selection network 121, an adaptive feature extraction network 122, and a selective domain adaptive network 123. The method is executed by a computer device including a server, and the method includes:

[0061] Step 110: Preprocess the battery charge and discharge cycle data (multiple operating conditions).

[0062] In some embodiments, step 110 includes a multi-operating condition data sampling step and a data processing step.

[0063] In some embodiments, the battery sensing signal is collected by the sensor under multiple working conditions. The working condition is the working state of the device when it is running. For example, when the device is running, the multiple working conditions represent at least one of the working state at different ambient temperatures, the working state under different charging currents, and the working state under different discharge currents. The battery sensing signal is a signal about the battery state, including at least one of the battery voltage signal 111, the output current signal 112, and the battery temperature signal 113.

[0064] In each battery charge and discharge cycle, each type of battery sensor signal takes multiple sampling points to form sensor data. The number of sampling points is determined based on the accuracy of the sensor in the device, and the number of sampling points is proportional to the accuracy of the sensor. Exemplarily, in a battery charge and discharge cycle, such as one hour, 120 sampling points are taken for the battery voltage signal 111, the output current signal 112, and the battery temperature signal 113, respectively, to form sensor data with a length of 360 sampling points.

[0065] In some embodiments, the battery capacity data during the charge and discharge cycle is collected as the battery health status tag data of the above-mentioned battery sensor signal, and the battery health status tag data and the above-mentioned battery sensor signal are combined into a battery data set. The battery capacity data refers to the capacity of the battery and is used to indicate the battery health status. Each battery health status tag data and each group of battery sensor signals collected in a corresponding charge and discharge cycle constitute a group of data in the battery data set. Multiple groups of data corresponding to the same battery health status tag data constitute a set of data. A group of battery sensor signals as input data includes at least one of the battery voltage signal 111, the output current signal 112, and the battery temperature signal 113 at the same time.

[0066] Exemplarily, when the battery voltage signal 111 is 4 volts, the output current signal 112 is 1 ampere, and the battery temperature signal 113 is 30 degrees Celsius, the battery capacity data is 70%, and the corresponding battery health status label data is 0.7. The battery health status label data of 0.7, the battery voltage signal 111 of 4 volts, the output current signal 112 of 1 ampere, and the battery temperature signal 113 of 30 degrees Celsius form a set of data in the battery data set. Taking the number of sampling points 120 as an example, each battery health status label data corresponds to 120 sampling points until the battery health status label data indicates that the battery has failed. For example, the battery health status label data corresponding to the battery failure is 0.6, and the interval between each adjacent battery health status label data is 0.01, that is, 0.6, 0.61, 0.62...1, a total of 41 battery health status label data, and the battery data set has a total of 41*120=4920 groups of data.

[0067] In some embodiments, the above battery data set is divided into a training set 114 and a test set 115, wherein the training set 114 is used to train the model 100, and the test set 115 is used to test the effect after the model 100 is trained. The division ratio is set by a person skilled in the art, and can be a variety of different division ratios, which are not limited in this application. Exemplarily, the battery data set is divided into a training set 114 and a test set 115 at a ratio of 7:3, and the training set 114 and the test set 115 are standardized respectively to eliminate the dimensional differences between the data in the training set 114 and the test set 115.

[0068] Step 120: Extract domain-invariant features.

[0069] In the embodiment of the present application, a domain represents an operating condition, and one operating condition represents one domain. A domain with input data and battery health status label data is divided into a source domain; a domain with input data but no battery health status label data, or a domain with input data but only a small amount of input data corresponding to battery health status label data is divided into a target domain.

[0070] Domain-invariant features refer to features that are not affected by the domain. In the embodiments of the present application, they refer to features that are not affected by the operating conditions, that is, features that can characterize the health status of the battery under different operating conditions.

[0071] Since the change of working conditions will affect the performance of other battery health state estimation models (referred to as other models) similar to model 100, that is, other models trained with working condition 1 data only perform well in working condition 1, and their performance in other working conditions will decline. In order to solve this problem, the embodiment of the present application uses model 100 to extract domain-invariant features, which can effectively characterize the health state of the battery under different working conditions, that is, model 100 trained under working condition 1 still performs well in working condition 2. Therefore, even if the working conditions change, model 100 can still effectively realize the battery health estimation under the current working conditions.

[0072] In some embodiments, a convolutional autoencoder (CAE) is used to construct the unsupervised domain selection network 121.

[0073] CAE is a convolutional neural network-based autoencoder used to learn the compressed representation and reconstruction of input data. It encodes the input data through multiple convolutional layers and pooling layers, and then decodes it through deconvolution layers and upsampling layers to reconstruct the input data.

[0074] In some embodiments, for each battery data set under each source domain, a CAE model is trained using a training set, with the goal of minimizing the reconstruction error corresponding to each source domain, and the CAE model parameters are optimized to obtain the CAE model of each source domain.

[0075] The source domain refers to a data set of existing knowledge, which is used to train and adjust model parameters, such as a historical data set, a currently collected data set, and in the embodiment of the present application, the source domain refers to a data set under existing working conditions. The reconstruction error refers to the difference between the input data after encoding and decoding and the input data.

[0076] In some embodiments, for the target domain, a training set is selected and sequentially input into each source domain CAE model to obtain the reconstruction error corresponding to the target domain, and the source domain corresponding to the source domain CAE model with the smallest reconstruction error corresponding to the target domain is selected as the optimal source domain matching this target domain. The target domain is a data set that needs to be learned. In an embodiment of the present application, the target domain represents a data set under working conditions for battery health state estimation. The source domain and the target domain are independent of each other and there is no overlap. For example, in the case where the battery health state of five batteries in winter is needed to estimate the battery health state of four batteries in summer, the data set of five batteries in winter is the source domain, and the data set of four batteries in summer is the target domain.

[0077] In some embodiments, the battery dataset as input to the adaptive feature extraction network 122 includes a source domain and a target domain.

[0078] In some embodiments, the adaptive feature extraction network 122 includes a convolutional neural network (CNN) embedded with an adaptive feature transmission mechanism (AFTM) for extracting domain-invariant features between source domain and target domain features. AFTM performs pooling, screening, and other processing on source domain input data and target domain input data through network layers such as pooling layers, fully connected layers, and activation functions to extract source domain features and target domain features, respectively.

[0079] In some embodiments, the selective domain adaptation network 123 calculates the distribution difference between the source domain features and the target domain features. With the goal of reducing the distribution difference, the adaptive feature extraction network 122 and the selective domain adaptation network 123 are trained, for example, by using the maximum mean difference (MMD) for calculation, and by continuously reducing the distribution difference between the two, the adaptive feature extraction network 122 is helped to extract domain invariant features.

[0080] In some embodiments, based on the extracted domain-invariant features, the health status of the battery is estimated by a health index (HI) estimator. The HI estimator is a fully connected network followed by an activation function for mapping the HI value output by the HI estimator to 0-1. The domain-invariant features are input into the HI estimator, and the HI value is output. The error between the output HI value and the true HI value is calculated by the mean square error, and the HI estimator is optimized based on the error to achieve battery health status estimation.

[0081] Step 130: Construct a loss function for model training and parameter optimization.

[0082] In some embodiments, the training of the model 100 is divided into two stages: in stage 1, the loss function of the model 100 is the reconstruction error of the CAE model; in stage 2, the loss function of the model 100 consists of two parts: HI estimation error and MMD loss. The model is trained and parameter optimization is performed by reducing the reconstruction error, HI estimation error and MMD loss.

[0083] In some embodiments, the HI estimation error is calculated by at least one of mean square error (MSE), root mean square error (RMSE), and mean absolute error (MAE), which is not limited in the present application.

[0084] In summary, the method provided in this embodiment trains the adaptive feature extraction network and the selective domain adaptive network under multiple working conditions, so that the adaptive feature extraction network extracts domain-invariant features for estimating the battery health state, thereby realizing battery health state estimation under multiple working conditions and improving the versatility of the model.

[0085] The method provided in this embodiment also screens the input features through an adaptive feature extraction network, thereby optimizing the feature extraction function.

[0086] The method provided in this embodiment also selectively transfers source domain features to the target domain during transfer learning, thereby improving the transfer learning effect of the model.

[0087] Figure 2 A flowchart of a model training method for battery health state estimation provided by an exemplary embodiment of the present application is shown, and the method includes:

[0088] Step 210: Obtain a battery data set.

[0089] In some embodiments, the battery data set includes battery sensor signals under at least two operating conditions.

[0090] The working condition is the working state of the device during operation. For example, when the device is running, the multiple working conditions represent at least one of the working state at different ambient temperatures, the working state at different charging currents, and the working state at different discharging currents. The battery sensor signal is a signal about the battery state, including at least one of the battery voltage signal, the output current signal, and the battery temperature signal.

[0091] In some embodiments, the method of acquiring the battery data set includes: at least one of on-site collection, acquisition from historical records, and computer simulation calculation, which is not limited in the present application.

[0092] Step 222: Using the battery sensor signal under the first working condition as source domain input data, and extracting source domain features through an adaptive feature extraction network.

[0093] In an embodiment of the present application, the domain represents the working conditions, and the source domain represents a data set of existing knowledge, which is used to train and adjust model parameters, such as a historical data set and a currently collected data set. In an embodiment of the present application, the source domain represents a data set under existing working conditions.

[0094] For example, Figure 3 As shown, the battery data set 310 includes a battery sensor signal corresponding to an ambient temperature of 40 degrees Celsius, which is used as the source domain input data X s , the source domain features 1223 are extracted through the adaptive feature extraction network 122.

[0095] In some embodiments, the adaptive feature extraction network includes a CNN embedded with an AFTM, which is used to extract domain-invariant features between source domain and target domain features. The AFTM performs pooling, screening, and other processing on source domain input data through network layers such as pooling layers, fully connected layers, and activation functions to extract source domain features.

[0096] like Figure 3 As shown, the adaptive feature extraction network 122 is used for the source domain or the target domain. Figure 3 The two adaptive feature extraction networks 122 in the figure represent the case of being used in the source domain or the case of being used in the target domain, and do not represent the case of inputting the source domain or the target domain at the same time. Each source domain or target domain corresponds to an adaptive feature extraction network 122. Wherein, C corresponds to the number of types of the above-mentioned battery sensor signals, for example, C is 3, corresponding to the above-mentioned battery voltage signal, output current signal and battery temperature signal; t is the number of sampling points of each type of sensor signal, for example, t is 120; X s represents the source domain input data, X t Represents the target domain input data; Avg represents the average pooling layer, Max represents the maximum pooling layer, FC0 and FC1 represent two fully connected layers respectively; device 1221 represents the activation function.

[0097] like Figure 4 As shown, Figure 3 The adaptive feature extraction network 122 in the embodiment includes a CNN embedded with an AFTM 420. Figure 3 The previous layer of CNN (the previous convolution layer 410) and the next layer of CNN (the next convolution layer 450) are not shown in the figure. The output of the previous convolution layer 410 is used as input, passes through the intermediate convolution layer 430 and then the intermediate pooling layer 440, and is added to the data content processed by AFTM420 as the input of the next convolution layer 450. AFTM420 can be embedded in CNN only once, or it can be embedded in CNN only once. Figure 4The structure shown is used as a unit in the model 100. For example, when there are convolutional layer 1, convolutional layer 2, convolutional layer 3, and convolutional layer 4, a single AFTM can be embedded between any two layers of the above convolutional layers; when there are multiple AFTMs such as AFTM1 and AFTM2, a structure of convolutional layer 1, AFTM1, convolutional layer 2 and convolutional layer 3, AFTM2, and convolutional layer 4 can also be formed.

[0098] Among them, the structure of AFTM420 is as follows Figure 5 As shown in the figure: For the output features of each domain (the selected source domain and the target domain), that is, the input features F, the feature channel descriptor F is obtained by combining average pooling (Avg) and maximum pooling (Max) c , F c It is used to describe the overall feature information of the feature vectors of each domain. c After that, the channel weight w is obtained through the fully connected layer and activation function. For example, through two layers of fully connected layers ( Figure 3 FC0 and FC1 in the ) and Sigmoid nonlinear activation function to obtain the channel weight w. The two fully connected layers are used to convert F c Dimensionality increase and dimensionality reduction increase the nonlinear mapping capability and improve the effectiveness of the channel weight w. The activation function is used to map the channel weight w to between 0 and 1, so as to facilitate the subsequent multiplication operation between the channel weight w and the input feature F after the convolution layer 510. The process of obtaining the channel weight w can be expressed as follows:

[0099] w=σ(FC1(FC0(AVG_Pool(F)+MAX_Pool(F))))=σ(W1(W0(F c )));

[0100] Among them, σ is the Sigmoid nonlinear activation function, FC0 and FC1 are two-layer fully connected operations, AVG_Pool and MAX_Pool represent average pooling and maximum pooling operations respectively, F is the input feature F, and Represent the model parameters of the two fully connected layers, F c is the feature channel descriptor.

[0101] Since AFTM needs to be embedded in different network layers of CNN to realize feature transmission, the shape dimensions of its input features and output features are different. According to the input feature shape of the next layer of AFTM, a convolution layer with a convolution kernel size of 1×1 is used to change the shape of the AFTM input feature, that is, Figure 3 On this basis, the channel weight w obtained above is multiplied with the changed feature to obtain the output feature F' of AFTM, such as Figure 5As shown, the shape dimension of the input feature F is 4×6 (C×L), the shape dimension of the output feature F' is 5×4 (C'×L'), and feature cells with different numbers and blank feature cells represent different features.

[0102] In some embodiments, Figure 6 As shown, the adaptive feature extraction network 122 is assisted to extract domain invariant features by selecting the domain adaptation network 600. Wherein, c represents the number of channels of the feature, and l represents the length of the feature, for example, c is 4 and l is 4. Figure 6 The selective domain adaptation network 600 is shown with Figure 3 The selective domain adaptation network 123 shown represents the same network, the selective domain adaptation network 600 focuses on expressing the acquisition of channel weights and selective features, and the selective domain adaptation network 123 focuses on expressing the output of the adaptive feature extraction network 122 as input.

[0103] The source domain features 611 and the target domain features 612 obtained from the adaptive feature extraction network 122 are processed by a global pooling (average pooling or maximum pooling) operation. The source domain channel weights 631 and the target domain channel weights 632 are generated by a shared fully connected layer 620. Figure 5 The difference is that here the input feature length of the shared fully connected layer 620 is the same as the output feature length, and there is no change in the feature shape, for example Figure 5 The input feature length of the two fully connected layers is 4, and the output feature length is 5. Figure 6 The input feature length of the shared fully connected layer 620 is 4, and the output feature length is 4. The source domain channel weight 631 is multiplied by the source domain feature 611 to obtain the selective feature 641 of the source domain, and the target domain channel weight 632 is multiplied by the target domain feature 612 to obtain the selective feature 642 of the target domain.

[0104] Step 224: The battery sensor signal under the second working condition is used as the target domain input data, and the target domain features are extracted by an adaptive feature extraction network.

[0105] In some embodiments, the target domain is a data set that needs to be learned. In the embodiment of the present application, the target domain represents a data set under the working condition for battery health state estimation. The source domain and the target domain are independent of each other and do not overlap.

[0106] For example, Figure 3 As shown, the battery data set 310 includes the battery sensor signal corresponding to the ambient temperature of 50 degrees Celsius, which is used as the target domain input data X t , the target domain features 1224 are extracted through the adaptive feature extraction network 122.

[0107] In some embodiments, the adaptive feature extraction network includes a CNN embedded with an AFTM, which is used to extract domain-invariant features between source domain and target domain features. The AFTM performs pooling, screening, and other processing on the target domain input data through network layers such as pooling layers, fully connected layers, and activation functions to extract target domain features.

[0108] The principles of step 224 and step 223 are the same and will not be repeated here.

[0109] Step 230: Calculate the distribution difference between the source domain features and the target domain features through the selective domain adaptation network.

[0110] In some embodiments, the distribution difference indicates that the distribution conditions between the features are different, or that the distribution patterns between the features are inconsistent.

[0111] In some embodiments, the selective domain adaptation network performs pooling, screening, and other processing on source domain features and target domain features through network layers such as pooling layers, fully connected layers, and activation functions to obtain the selective features of the source domain and the selective features of the target domain, and calculate the distribution difference between the selective features of the source domain and the selective features of the target domain.

[0112] In some embodiments, the distribution difference between the selective features of the source domain and the selective features of the target domain is calculated, for example, by using MMD for calculation, and by continuously reducing the distribution difference between the two, the adaptive feature extraction network 122 is helped to extract domain-invariant features.

[0113] Step 240: With the goal of reducing the distribution difference, the adaptive feature extraction network and the selective domain adaptation network are trained so that the adaptive feature extraction network extracts domain-invariant features in the source domain and the target domain.

[0114] Among them, domain invariant features are used to represent the characteristics of the battery health status when it is not affected by the working conditions. Figure 1 The performance of other battery health state estimation models (referred to as other models) similar to the battery health state estimation model 100 (referred to as model 100) shown, that is, other models trained with working condition 1 data only perform well in working condition 1, and their performance in other working conditions will decrease. The embodiment of the present application uses model 100 to extract domain invariant features, and domain invariant features can effectively characterize the health state of the battery under different working conditions, that is, model 100 trained under working condition 1 still performs well in working condition 2, thereby improving the generalization of model 100.

[0115] In some embodiments, training is performed by setting a target threshold or setting the number of training times. In the embodiments of the present application, setting the number of training times is used as an example for illustration. For example, after training 3000 times, the distribution difference will be reduced to a smaller value to meet the training requirements.

[0116] In summary, the method provided in this embodiment trains the adaptive feature extraction network and the selective domain adaptive network under multiple working conditions, so that the adaptive feature extraction network extracts domain-invariant features for estimating the battery health state, thereby realizing battery health state estimation under multiple working conditions and improving the versatility of the model.

[0117] Figure 7 A flowchart of a model training method for battery health state estimation provided by an exemplary embodiment of the present application is shown, and the method includes:

[0118] Step 210: Obtain a battery data set.

[0119] Step 222: Using the battery sensor signal under the first working condition as source domain input data, and extracting source domain features through an adaptive feature extraction network.

[0120] Step 224: The battery sensor signal under the second working condition is used as the target domain input data, and the target domain features are extracted by an adaptive feature extraction network.

[0121] Step 231: By sharing the fully connected layer, the source domain channel weights corresponding to the source domain features and the target domain channel weights corresponding to the target domain features are calculated.

[0122] In some embodiments, the selective domain adaptation network further includes a source domain pooling layer and a target domain pooling layer.

[0123] The source domain features are pooled through the source domain pooling layer to obtain the pooled source domain features; the target domain features are pooled through the target domain pooling layer to obtain the pooled target domain features;

[0124] By sharing the fully connected layer, the source domain channel weights corresponding to the pooled source domain features and the target domain channel weights corresponding to the pooled target domain features are calculated.

[0125] In some embodiments, Figure 6 As shown, the pooling process includes at least one of maximum pooling and average pooling. Through the pooling operation, the number of parameters of the source domain features and the target domain features is reduced, and the computational complexity is reduced. The source domain channel weight 631 and the target domain channel weight 632 are generated by sharing the fully connected layer 620.

[0126] Step 232: Calculate source domain selective features based on source domain features and source domain channel weights; calculate target domain selective features based on target domain features and target domain channel weights.

[0127] In some embodiments, Figure 6 As shown, the source domain channel weight 631 is multiplied by the source domain feature 611 to obtain the selective feature 641 of the source domain, and the target domain channel weight 632 is multiplied by the target domain feature 612 to obtain the selective feature 642 of the target domain.

[0128] Step 233: Calculate the distribution difference between the source domain selective features and the target domain selective features.

[0129] In some embodiments, methods for calculating distribution differences include: MMD, Kullback-Leibler divergence, correlation alignment and other methods, which are not limited in the embodiments of the present application, and are described using MMD calculation as an example.

[0130] Step 240: With the goal of reducing the distribution difference, the adaptive feature extraction network and the selective domain adaptation network are trained so that the adaptive feature extraction network extracts domain-invariant features in the source domain and the target domain.

[0131] In summary, the method provided in this embodiment trains the adaptive feature extraction network and the selective domain adaptive network under multiple working conditions, so that the adaptive feature extraction network extracts domain-invariant features for estimating the battery health state, thereby realizing battery health state estimation under multiple working conditions and improving the versatility of the model.

[0132] The method provided in this embodiment also reduces the number of parameters of source domain features and target domain features through the pooling layer, reduces the computational complexity of the model, reduces and increases the dimension of the input features through the fully connected layer, improves the nonlinear mapping capability of the adaptive feature extraction network, and enhances the screening capability of the input features.

[0133] Figure 8 A flowchart of a model training method for battery health state estimation provided by an exemplary embodiment of the present application is shown, and the method includes:

[0134] Step 210: Obtain a battery data set.

[0135] Step 215: Based on the battery sensor signals corresponding to the at least two source domains and the battery sensor signals corresponding to the target domain, determine the optimal source domain that matches the target domain from among the at least two source domains.

[0136] In some embodiments, the battery data set includes: battery sensor signals corresponding to at least two source domains, and a battery sensor signal corresponding to the target domain.

[0137] (1) based on the battery sensor signal corresponding to each source domain in at least two source domains, training a source domain convolutional autoencoder model corresponding to each source domain;

[0138] In some embodiments, Fig. 9 As shown, CAE is used to construct an unsupervised domain selection network 121. CAE is an autoencoder based on a convolutional neural network, which is used to learn the compressed representation and reconstruction of input data. It encodes the input data through multiple convolutional layers and pooling layers, and then decodes it through deconvolution layers and upsampling layers to reconstruct the input data.

[0139] In some embodiments, for each battery data set under each source domain, a CAE model is trained using a training set, and the CAE model parameters are optimized with the goal of minimizing the reconstruction error corresponding to each source domain to obtain a CAE model of each source domain. The reconstruction error refers to the difference between the input data after encoding and decoding and the input data.

[0140] In some embodiments, Fig. 9 As shown, the adaptive feature extraction network 122 is used for the source domain or the target domain, and the target domain and the corresponding optimal source domain share one adaptive feature extraction network 122.

[0141] In some embodiments, based on the battery sensor signal corresponding to each source domain of at least two source domains, a source domain convolutional autoencoder model corresponding to each source domain is trained by a third loss function, and the third loss function is used to calculate the reconstruction error of the source domain convolutional autoencoder model corresponding to each source domain, and the formula is:

[0142]

[0143] in, represents the battery sensor signal input to the i-th CAE model, represents the battery sensor signal reconstructed by the i-th CAE model, n represents the total number of sampling points, represents the reconstruction error of the CAE model corresponding to the i-th source domain.

[0144] (2) Inputting the battery sensor signal corresponding to the target domain into the source domain convolutional autoencoder model corresponding to each source domain, and calculating the reconstruction error between the target domain and each source domain respectively;

[0145] (3) Determine the source domain with the smallest reconstruction error as the optimal source domain to match the target domain.

[0146] In some embodiments, for the target domain, a training set is selected and input into each source domain CAE model in turn to obtain the reconstruction error corresponding to the target domain, and the source domain corresponding to the source domain CAE model with the smallest reconstruction error corresponding to the target domain is selected as the optimal source domain matching this target domain.

[0147] exist Fig. 9 In the example, the source domains include DS1, DS2, DS3, ... DSm, a total of m source domains, and the target domains include DT1, DT2, DT3, ... DTn, a total of n target domains. The training set of DT1 is input into the m source domain CAE models corresponding to the m source domains, and the corresponding reconstruction errors are obtained. DS2 with the smallest reconstruction error is selected as the optimal source domain of DT1. Similarly, the source domains corresponding to each target domain are selected, DS1 corresponds to DT3, DS2 corresponds to DT1, DS3 corresponds to DTn, DSm corresponds to DT2, and m and n are natural numbers, which may be the same or different.

[0148] Step 2221: Use the battery sensor signal under the first working condition as source domain input data, and obtain a first source domain feature vector by convolving the source domain input data in the front convolution layer.

[0149] In some embodiments, the adaptive feature extraction network includes a cascaded front convolution layer, an adaptive feature transfer network, and a back convolution layer. The front convolution layer extracts a first source domain feature vector by convolving the battery sensor signal as the source domain input data as the input of the adaptive feature transfer network; the back convolution layer obtains the source domain feature by convolving the output of the adaptive feature transfer network.

[0150] In some embodiments, the adaptive feature transfer network (adaptive feature transfer mechanism AFTM) can be embedded in the CNN only once, or Figure 4 The structure shown is used as a unit in the model. For example, when there are convolutional layers 1, 2, 3, and 4, a single AFTM can be embedded between any two of the above convolutional layers; when there are multiple AFTMs such as AFTM1 and AFTM2, a structure of convolutional layer 1, AFTM1, convolutional layer 2 and convolutional layer 3, AFTM2, and convolutional layer 4 can also be formed.

[0151] Step 2222: Adaptively screen the first source domain feature vector through an adaptive feature transfer network to obtain a second source domain feature vector.

[0152] In some embodiments, the adaptive feature transfer network includes: an average pooling layer, a maximum pooling layer, at least two fully connected layers, and a first convolutional layer.

[0153] (1) performing pooling processing on the first source domain feature vector through an average pooling layer and a maximum pooling layer respectively and then adding the pooled features to obtain a first feature channel descriptor, where the first feature channel descriptor is used to describe the overall feature information of the first source domain feature vector;

[0154] In some embodiments, Fig. 9As shown, the first source domain feature vector is pooled by the average pooling layer (Avg) and the maximum pooling layer (Max), which reduces the number of parameters of the first source domain feature vector and reduces the computational complexity.

[0155] (2) processing the first feature channel descriptor through at least two fully connected layers to obtain a source domain channel weight, where the source domain channel weight is used to describe the weight of the source domain feature channel;

[0156] In some embodiments, source domain channel weights are obtained through a fully connected layer and an activation function.

[0157] like Fig. 9 As shown, the source domain channel weight is obtained through two layers of fully connected layers (FC0 and FC1) and Sigmoid nonlinear activation function. The two layers of fully connected layers are used to increase and reduce the dimension of the first feature channel descriptor, increase the nonlinear mapping capability, and improve the effectiveness of the source domain channel weight. The activation function is used to map the source domain channel weight to between 0 and 1, so as to facilitate the subsequent multiplication operation between the source domain channel weight and the input feature after the convolution layer 1222.

[0158] (3) convolving the first source domain feature vector through the first convolutional layer to obtain a third source domain feature vector;

[0159] In some embodiments, Fig. 9 As shown, the first convolution layer, namely the convolution layer 1222, convolves the first source domain feature vector through the convolution layer 1222 to obtain a third source domain feature vector.

[0160] (4) Adaptively filter the third source domain feature vector based on the source domain channel weight to obtain the second source domain feature vector.

[0161] In some embodiments, the source domain channel weight is multiplied by the third source domain feature vector to obtain a second source domain feature vector.

[0162] Step 2223: Obtain source domain features by convolving the second source domain feature vector in the post-convolutional layer.

[0163] In some embodiments, between the front convolution layer and the rear convolution layer, there are also: an intermediate convolution layer and an intermediate pooling layer.

[0164] Convolving the first source domain feature vector through the intermediate convolution layer to obtain a first intermediate convolution feature vector;

[0165] The first intermediate convolutional feature vector is pooled by the intermediate pooling layer to obtain a first intermediate pooled feature vector.

[0166] like Figure 4As shown, between the front convolution layer 410 and the rear convolution layer 450, there are also an intermediate convolution layer 430 and an intermediate pooling layer 440. The first source domain feature vector is convolved by the intermediate convolution layer 430 to obtain a first intermediate convolution feature vector; the first intermediate convolution feature vector is pooled by the intermediate pooling layer to obtain a first intermediate pooling feature vector, thereby reducing the number of parameters of the first intermediate convolution feature vector and reducing the computational complexity.

[0167] In some embodiments, step 2223 includes: convolving the sum of the second source domain feature vector and the first intermediate pooling feature vector to obtain the source domain feature.

[0168] Step 2241: Use the battery sensor signal under the second working condition as the target domain input data, and obtain the first target domain feature vector by convolving the target domain input data in the front convolution layer.

[0169] The principle of step 2241 is the same as that of step 2221 and will not be repeated here.

[0170] Step 2242: Adaptively screen the first target domain feature vector through an adaptive feature transfer network to obtain a second target domain feature vector.

[0171] In some embodiments, the adaptive feature transfer network includes: an average pooling layer, a maximum pooling layer, at least two fully connected layers, and a first convolutional layer;

[0172] Adaptively screening the first target domain feature vector through an adaptive feature transfer network to obtain a second target domain feature vector, including:

[0173] The first target domain feature vector is pooled by an average pooling layer and a maximum pooling layer respectively, and then added to obtain a second feature channel descriptor, where the second feature channel descriptor is used to describe the overall feature information of the first target domain feature vector;

[0174] Processing the second feature channel descriptor through at least two fully connected layers to obtain a target domain channel weight, where the target domain channel weight is used to describe the weight of the target domain feature channel;

[0175] Convolving the first target domain feature vector through the first convolution layer to obtain a third target domain feature vector;

[0176] Adaptive feature screening is performed on the third target domain feature vector based on the target domain channel weight to obtain a second target domain feature vector.

[0177] The principle of step 2242 is the same as that of step 2222 and will not be repeated here.

[0178] Step 2243: Obtain target domain features by convolving the second target domain feature vector in the post-convolutional layer.

[0179] In some embodiments, between the front convolution layer and the rear convolution layer, there are also: an intermediate convolution layer and an intermediate pooling layer.

[0180] Convolving the first target domain feature vector through the intermediate convolution layer to obtain a second intermediate convolution feature vector;

[0181] Performing pooling processing on the second intermediate convolutional feature vector through the intermediate pooling layer to obtain a second intermediate pooling feature vector;

[0182] In some embodiments, step 2243 includes: convolving the second target domain feature vector and the sum of the second intermediate pooled feature vector to obtain the target domain feature.

[0183] The principle of step 2243 is the same as that of step 2223 and will not be repeated here.

[0184] Step 230: Calculate the distribution difference between the source domain features and the target domain features through the selective domain adaptation network.

[0185] Step 242: Input the features of the battery sensor signal extracted by the adaptive feature extraction network into the battery health state estimation network for prediction to obtain the predicted battery health state.

[0186] In some embodiments, the battery data set includes actual battery health states corresponding to the battery sensor signals, and the model further includes a battery health state estimation network.

[0187] In some embodiments, Fig. 9 As shown, the battery health state estimation network includes a health indicator estimator (HI estimator) 124. Based on the extracted domain invariant features, the battery health state is estimated by the HI estimator 124. The HI estimator 124 is a fully connected network followed by an activation function for mapping the HI output by the HI estimator 124 to 0-1. The domain invariant features are input into the HI estimator 124, and the HI is output. The error between the output HI and the true HI is calculated by the mean square error, and the HI estimator 124 is optimized based on the error to achieve health state estimation.

[0188] Step 244: Based on the sum of the first loss function and the second loss function, the adaptive feature extraction network, the selective domain adaptation network, and the battery health state estimation network are trained so that the adaptive feature extraction network extracts domain invariant features in the source domain and the target domain.

[0189] Among them, the first loss function is used to calculate the error between the predicted battery health state and the actual battery health state, and the second loss function is used to calculate the distribution difference.

[0190] In some embodiments, the training of the model is divided into two stages. In stage 1, the loss function of the model is the reconstruction error of the CAE model; in stage 2, the loss function of the model consists of two parts: HI estimation error and MMD loss.

[0191] In some embodiments, the HI estimation error is calculated by at least one of MSE, RMSE, and MAE. The embodiment of the present application is not limited to this. MSE is used as an example for illustration, and its formula is:

[0192]

[0193] Among them, y j Indicates the true HI value, represents the HI value estimated by the model, n represents the total number of sampling points, L HI represents the HI estimation error.

[0194] The distribution difference problem between the selective features of the source domain and the selective features of the target domain is solved by minimizing the MMD loss. The MMD loss formula is:

[0195] L MMD =MMD(W(O s ),W(I t ));

[0196] Where W represents Figure 6 The feature channel weighting process in the embodiment, L MMD represents MMD loss, O s represents the source domain features, O t represents the target domain feature. MMD(a,b) represents the distribution difference between feature a and feature b, and the formula is:

[0197]

[0198] Among them, n s Represents the number of features in the source domain, n t represents the number of features in the target domain, φ is the Gaussian kernel function, ||·|| H represents a Hilbert space.

[0199] By adding the HI estimation error and the MMD loss, the loss function of the model is obtained, and its formula is:

[0200] L p2 =L HI +βL MMD ;

[0201] Among them, β is the loss term hyperparameter, which is used to balance the HI estimation error L HI and MMD loss L MMD , since the magnitude of HI estimation error and MMD loss is usually different, it is necessary to add a loss term hyperparameter, L p2 Represents the loss function of the model.

[0202] In summary, the method provided in this embodiment trains the adaptive feature extraction network and the selective domain adaptive network under multiple working conditions, so that the adaptive feature extraction network extracts domain-invariant features for estimating the battery health state, thereby realizing battery health state estimation under multiple working conditions and improving the versatility of the model.

[0203] The method provided in this embodiment also reduces the number of parameters of the input features through the pooling layer, reduces the computational complexity of the model, reduces and increases the dimension of the input features through the fully connected layer, improves the nonlinear mapping capability of the adaptive feature extraction network, and enhances the screening capability of the input features.

[0204] The method provided in this embodiment also determines the optimal source domain that matches the target domain through an unsupervised domain selection network, thereby improving the battery health status estimation accuracy of the model.

[0205] Fig.10 A flowchart of a method for using a model to estimate a battery health state provided by an exemplary embodiment of the present application is shown. The method includes:

[0206] Step 1010: Obtain a battery data set.

[0207] In some embodiments, the battery data set includes a battery sensor signal under a second operating condition.

[0208] Step 1010 and Figure 2 The principle of step 210 in the embodiment is the same and will not be repeated here.

[0209] Step 1020: Using the battery sensor signal under the second working condition as target domain input data, and extracting domain invariant features through an adaptive feature extraction network.

[0210] After the adaptive feature extraction network training is completed, the output domain remains unchanged. Other principles are the same as Figure 2 The principle of step 224 of the embodiment is the same and will not be repeated here.

[0211] Step 1030: Predict the battery health state based on the domain invariant features through the battery health state estimation network.

[0212] Among them, domain-invariant features are used to represent the characteristics of the battery health status without being affected by working conditions.

[0213] In some embodiments, Fig.11 As shown, the battery health state estimation network includes a health index estimator (HI estimator) 124. The HI estimator is a fully connected network followed by an activation function for mapping the HI value output by the HI estimator to 0-1. The domain invariant features are input into the HI estimator, and the HI value is output. The HI value is used to predict the battery health state.

[0214] In summary, the method provided in this embodiment predicts the battery health state based on domain-invariant features through a battery health state estimation network, thereby realizing battery health state estimation under multiple working conditions.

[0215] Fig.12 A flowchart of a method for using a model to estimate a battery health state provided by an exemplary embodiment of the present application is shown. The method includes:

[0216] Step 1010: Obtain a battery data set.

[0217] Step 1021: Use the battery sensor signal under the second working condition as the target domain input data, and obtain a fourth target domain feature vector by convolving the target domain input data in the front convolution layer.

[0218] In some embodiments, the adaptive feature extraction network includes a cascaded preceding convolutional layer, an adaptive feature transfer network, and a succeeding convolutional layer.

[0219] Step 1021 and Figure 8 The principle of step 2241 of the embodiment is the same and will not be repeated here.

[0220] Step 1022: Adaptively screen the fourth target domain feature vector through an adaptive feature transfer network to obtain a fifth target domain feature vector.

[0221] In some embodiments, the adaptive feature transfer network includes: an average pooling layer, a maximum pooling layer, at least two fully connected layers, and a first convolutional layer.

[0222] In some embodiments, the fourth target domain feature vector is pooled by an average pooling layer and a maximum pooling layer respectively and then added to obtain a third feature channel descriptor, and the third feature channel descriptor is used to describe the overall feature information of the fourth target domain feature vector;

[0223] The third feature channel descriptor is processed by at least two fully connected layers to obtain a target domain channel weight, where the target domain channel weight is used to describe the weight of the target domain feature channel;

[0224] Convolving the fourth target domain feature vector through the first convolution layer to obtain a sixth target domain feature vector;

[0225] Based on the target domain channel weights and the sixth target domain feature vector, a fifth target domain feature vector is obtained.

[0226] Step 1022 and Figure 8 The principle of step 2242 of the embodiment is the same and will not be repeated here.

[0227] Step 1023: Obtain domain invariant features by convolving the fifth target domain feature vector in a post-convolutional layer.

[0228] Step 1023 and Figure 8 The principle of step 2243 of the embodiment is the same and will not be repeated here.

[0229] Step 1030: Predict the battery health state based on the domain invariant features through the battery health state estimation network.

[0230] In summary, the method provided in this embodiment predicts the battery health state based on domain-invariant features through a battery health state estimation network, thereby realizing battery health state estimation under multiple working conditions.

[0231] The method provided in this embodiment also reduces the number of parameters of the input features through the pooling layer, reduces the computational complexity of the model, reduces and increases the dimension of the input features through the fully connected layer, improves the nonlinear mapping capability of the adaptive feature extraction network, and enhances the screening capability of the input features.

[0232] In the above embodiment, steps with the same sequence number can be considered as the same step. Figure 2 The corresponding embodiments, Figure 7 The corresponding embodiments, Figure 8 The corresponding embodiments, Fig.10 The corresponding embodiments and Fig.12 The corresponding embodiments may be implemented individually or in combination, and this application does not limit this.

[0233] In some embodiments, the effectiveness of the battery health status estimation model (referred to as the model for short) is verified through a test set. Exemplarily, the three indicators of RMSE, MAE and scoring function (SF) are used to evaluate the battery health estimation effect. The embodiments of the present application do not limit the indicators used for evaluation, and the above three indicators are used as examples for illustration.

[0234] For example, under operating condition 1 (discharge current 2 amperes, charging current 1.5 amperes, ambient temperature 24 degrees Celsius), operating condition 2 (discharge current 4 amperes, charging current 1.5 amperes, ambient temperature 43 degrees Celsius), and operating condition 3 (discharge current 1 ampere, charging current 1.5 amperes, ambient temperature 4 degrees Celsius), the above three operating conditions correspond to domain 1, domain 2, and domain 3, respectively.

[0235] In stage 1, for the specified target domain, the optimal source domain is selected through the unsupervised domain selection network. For example, if domain 1 is the target domain, domains 2 and 3 are the source domains. The source domain CAE models corresponding to domains 2 and 3 are trained respectively. The reconstruction errors of domains 2 and 3 are calculated using part of the data of domain 1. It can be seen from the experiment that the reconstruction error of domain 2 is smaller than that of domain 3, and the data reconstruction performance is better. Therefore, when domain 1 is the target domain, domain 2 is selected as the optimal source domain. Table 1 shows the selection results of the optimal source domain.

[0236] Table 1

[0237]

[0238] In order to prove the effectiveness of the unsupervised domain selection network, the present application embodiment conducted a migration task comparison experiment. The comparison results are shown in Table 2.

[0239] Table 2

[0240] Migration tasks RMSE MAE SF 2-1 0.1055 0.0804 0.9990 3-1 0.1226 0.0930 1.1454 1-2 0.0899 0.0700 0.2063 3-2 0.0593 0.0472 0.1417 1-3 0.3818 0.3012 1.6396 2-3 0.2441 0.1949 1.0490

[0241] When domain 2 is the target domain, the values ​​of the three indicators are smaller for the migration task "3-2", which means that the result of the migration task "3-2" is better than that of the migration task "1-2". The results are consistent with Table 1, which verifies the effectiveness of the unsupervised domain selection network in selecting the optimal source domain in the model.

[0242] In order to prove the effectiveness of the adaptive feature extraction network in the model, the embodiment of the present application was tested and compared with typical feature extraction networks: CNN, Recurrent Neural Network (RNN), and Residual Network (ResNet). The comparison results are shown in Table 3.

[0243] Table 3

[0244]

[0245]

[0246] In order to prove the effectiveness of the selective domain adaptation network in the model, the embodiment of the present application is tested and compared with typical transfer learning methods: domain adaptation neural network (Domain Adaptation Neural Network, DaNN), domain adversarial neural network (Domain Adversarial Neural Network, DANN), and correlation alignment (CORrelation ALignment, CORAL). The comparison results are shown in Table 4.

[0247] Table 4

[0248]

[0249] It can be seen from Tables 3 and 4 that the estimated loss (MAE, SF, RMSE) of the model in the embodiment of the present application is much smaller than that of other typical models in each migration task.

[0250] Fig.13 A block diagram of a model training device for battery health state estimation provided by an exemplary embodiment of the present application is shown, wherein the model includes an adaptive feature extraction network and a selective domain adaptive network, and the device includes: an acquisition module 1310, an extraction module 1320, a calculation module 1330, a training module 1340, a pooling module 1350, a convolution module 1360, a screening module 1370, a fully connected module 1380, a prediction module 1390 and a determination module 13100.

[0251] An acquisition module 1310 is used to acquire a battery data set, where the battery data set includes battery sensor signals under at least two working conditions;

[0252] The extraction module 1320 is used to use the battery sensor signal under the first working condition as source domain input data, and extract the source domain features through the adaptive feature extraction network; use the battery sensor signal under the second working condition as target domain input data, and extract the target domain features through the adaptive feature extraction network;

[0253] A calculation module 1330, configured to calculate a distribution difference between source domain features and target domain features through a selective domain adaptation network;

[0254] A training module 1340 is used to train the adaptive feature extraction network and the selective domain adaptation network with the goal of reducing the distribution difference, so that the adaptive feature extraction network extracts domain-invariant features in the source domain and the target domain;

[0255] Among them, domain-invariant features are used to represent the characteristics of the battery health status without being affected by working conditions.

[0256] In one possible design of this embodiment, the selective domain adaptation network includes a shared fully connected layer shared by the source domain and the target domain;

[0257] A calculation module 1330 is used to calculate the source domain channel weights corresponding to the source domain features and the target domain channel weights corresponding to the target domain features through a shared fully connected layer;

[0258] A calculation module 1330 is used to calculate a source domain selective feature based on a source domain feature and a source domain channel weight; and calculate a target domain selective feature based on a target domain feature and a target domain channel weight;

[0259] The calculation module 1330 is used to calculate the distribution difference between the source domain selective features and the target domain selective features.

[0260] In a possible design of this embodiment, the calculation module 1330 is used to calculate the distribution difference between the source domain selective feature and the target domain selective feature by using the maximum mean difference.

[0261] In a possible design of this embodiment, the selective domain adaptation network further includes a source domain pooling layer and a target domain pooling layer;

[0262] A pooling module 1350 is used to perform pooling processing on the source domain features through a source domain pooling layer to obtain pooled source domain features;

[0263] A pooling module 1350 is used to perform pooling processing on the target domain features through a target domain pooling layer to obtain pooled target domain features;

[0264] The calculation module 1330 is used to calculate the source domain channel weights corresponding to the pooled source domain features and the target domain channel weights corresponding to the pooled target domain features through a shared fully connected layer.

[0265] In a possible design of this embodiment, the adaptive feature extraction network includes a cascaded preceding convolutional layer, an adaptive feature transfer network, and a succeeding convolutional layer;

[0266] A convolution module 1360 is used to use the battery sensor signal under the first working condition as source domain input data, and obtain a first source domain feature vector by convolving the source domain input data in a front convolution layer;

[0267] A screening module 1370 is used to perform adaptive feature screening on the first source domain feature vector through an adaptive feature transmission network to obtain a second source domain feature vector;

[0268] The convolution module 1360 is used to obtain source domain features by convolving the second source domain feature vector in a post-convolution layer.

[0269] In a possible design of this embodiment, the adaptive feature transmission network includes: an average pooling layer, a maximum pooling layer, at least two fully connected layers and a first convolutional layer;

[0270] A pooling module 1350 is used to perform pooling processing on the first source domain feature vector through an average pooling layer and a maximum pooling layer respectively and then add the pooled data to obtain a first feature channel descriptor, where the first feature channel descriptor is used to describe the overall feature information of the first source domain feature vector;

[0271] A fully connected module 1380 is used to process the first feature channel descriptor through at least two fully connected layers to obtain a source domain channel weight, where the source domain channel weight is used to describe the weight of the source domain feature channel;

[0272] A convolution module 1360, configured to convolve the first source domain feature vector through a first convolution layer to obtain a third source domain feature vector;

[0273] The screening module 1370 is used to perform adaptive feature screening on the third source domain feature vector based on the source domain channel weight to obtain a second source domain feature vector.

[0274] In a possible design of this embodiment, the adaptive feature extraction network includes a cascaded preceding convolutional layer, an adaptive feature transfer network, and a succeeding convolutional layer;

[0275] A convolution module 1360 is used to use the battery sensor signal under the second working condition as the target domain input data, and obtain a first target domain feature vector by convolving the target domain input data in a front convolution layer;

[0276] A screening module 1370 is used to perform adaptive feature screening on the first target domain feature vector through an adaptive feature transmission network to obtain a second target domain feature vector;

[0277] The convolution module 1360 is used to obtain the target domain feature by convolving the second target domain feature vector in the post-convolution layer.

[0278] In a possible design of this embodiment, the adaptive feature transmission network includes: an average pooling layer, a maximum pooling layer, at least two fully connected layers and a first convolutional layer;

[0279] A pooling module 1350 is used to perform pooling processing on the first target domain feature vector through an average pooling layer and a maximum pooling layer respectively and then add them to obtain a second feature channel descriptor, where the second feature channel descriptor is used to describe the overall feature information of the first target domain feature vector;

[0280] A fully connected module 1380 is used to process the second feature channel descriptor through at least two fully connected layers to obtain a target domain channel weight, where the target domain channel weight is used to describe the weight of the target domain feature channel;

[0281] A convolution module 1360 is used to convolve the first target domain feature vector through a first convolution layer to obtain a third target domain feature vector;

[0282] The screening module 1370 is used to perform adaptive feature screening on the third target domain feature vector based on the target domain channel weight to obtain the second target domain feature vector.

[0283] In a possible design of this embodiment, between the front convolution layer and the rear convolution layer, there are also: an intermediate convolution layer and an intermediate pooling layer;

[0284] A convolution module 1360, configured to convolve the first source domain feature vector through an intermediate convolution layer to obtain a first intermediate convolution feature vector;

[0285] A pooling module 1350 is used to perform pooling processing on the first intermediate convolution feature vector through an intermediate pooling layer to obtain a first intermediate pooling feature vector;

[0286] The convolution module 1360 is used to convolve the sum of the second source domain feature vector and the first intermediate pooling feature vector to obtain the source domain feature.

[0287] In a possible design of this embodiment, between the front convolution layer and the rear convolution layer, there are also: an intermediate convolution layer and an intermediate pooling layer;

[0288] A convolution module 1360, configured to convolve the first target domain feature vector through an intermediate convolution layer to obtain a second intermediate convolution feature vector;

[0289] A pooling module 1350 is used to perform pooling processing on the second intermediate convolution feature vector through an intermediate pooling layer to obtain a second intermediate pooling feature vector;

[0290] The convolution module 1360 is used to convolve the second target domain feature vector and the sum of the second intermediate pooling feature vector to obtain the target domain feature.

[0291] In a possible design of this embodiment, the battery data set includes an actual battery health state corresponding to the battery sensor signal, and the model also includes a battery health state estimation network;

[0292] A prediction module 1390 is used to input the features of the battery sensor signal extracted by the adaptive feature extraction network into the battery health state estimation network for prediction to obtain a predicted battery health state;

[0293] A training module 1340 is used to train the adaptive feature extraction network, the selective domain adaptation network, and the battery health state estimation network based on the sum of the first loss function and the second loss function, so that the adaptive feature extraction network extracts domain invariant features in the source domain and the target domain;

[0294] Among them, the first loss function is used to calculate the error between the predicted battery health state and the actual battery health state, and the second loss function is used to calculate the distribution difference.

[0295] In a possible design of this embodiment, the battery data set includes: battery sensor signals corresponding to at least two source domains, and a battery sensor signal corresponding to the target domain;

[0296] The determination module 13100 is used to determine an optimal source domain matching the target domain from among the at least two source domains based on the battery sensor signals corresponding to the at least two source domains and the battery sensor signals corresponding to the target domain.

[0297] In a possible design of this embodiment, the training module 1340 is used to train a source domain convolutional autoencoder model corresponding to each source domain based on the battery sensor signal corresponding to each source domain in at least two source domains;

[0298] A calculation module 1330 is used to input the battery sensor signal corresponding to the target domain into the source domain convolutional autoencoder model corresponding to each source domain, and calculate the reconstruction error corresponding to each source domain between the target domain and each source domain;

[0299] The determination module 13100 is used to determine the source domain with the smallest reconstruction error as the optimal source domain matching the target domain.

[0300] In a possible design of this embodiment, the training module 1340 is used to train the source domain convolutional autoencoder model corresponding to each source domain in at least two source domains through a third loss function based on the battery sensor signal corresponding to each source domain, and the third loss function is used to calculate the reconstruction error of the source domain convolutional autoencoder model corresponding to each source domain.

[0301] Fig.14 A block diagram of a model using device for battery health state estimation provided by an exemplary embodiment of the present application is shown, the model includes an adaptive feature extraction network and a battery health state estimation network, and the device includes: an acquisition module 1410, an extraction module 1420, a prediction module 1430, a convolution module 1440, a screening module 1450, a pooling module 1460 and a fully connected module 1470.

[0302] An acquisition module 1410 is used to acquire a battery data set, where the battery data set includes a battery sensor signal under a second working condition;

[0303] An extraction module 1420, configured to use the battery sensor signal under the second operating condition as target domain input data and extract domain invariant features through an adaptive feature extraction network;

[0304] A prediction module 1430, configured to predict a battery health state based on domain invariant features through a battery health state estimation network;

[0305] Among them, domain-invariant features are used to represent the characteristics of the battery health status without being affected by working conditions.

[0306] In a possible design of this embodiment, the adaptive feature extraction network includes a cascaded preceding convolutional layer, an adaptive feature transfer network, and a succeeding convolutional layer;

[0307] A convolution module 1440 is used to use the battery sensor signal under the second working condition as target domain input data, and obtain a fourth target domain feature vector by convolving the target domain input data in a front convolution layer;

[0308] A screening module 1450 is used to perform adaptive feature screening on the fourth target domain feature vector through an adaptive feature transmission network to obtain a fifth target domain feature vector;

[0309] The convolution module 1440 is used to obtain domain invariant features by convolving the fifth target domain feature vector in a post-convolutional layer.

[0310] In a possible design of this embodiment, the adaptive feature transmission network includes: an average pooling layer, a maximum pooling layer, at least two fully connected layers and a first convolutional layer;

[0311] A pooling module 1460 is used to perform pooling processing on the fourth target domain feature vector through an average pooling layer and a maximum pooling layer respectively and then add them to obtain a third feature channel descriptor, where the third feature channel descriptor is used to describe the overall feature information of the fourth target domain feature vector;

[0312] A fully connected module 1470 is used to process the third feature channel descriptor through at least two fully connected layers to obtain a target domain channel weight, where the target domain channel weight is used to describe the weight of the target domain feature channel;

[0313] A convolution module 1440 is used to convolve the fourth target domain feature vector through the first convolution layer to obtain a sixth target domain feature vector;

[0314] The screening module 1450 is configured to obtain a fifth target domain feature vector based on the target domain channel weight and the sixth target domain feature vector.

[0315] An embodiment of the present application also provides a computer device, which includes: a processor and a memory, wherein at least one program is stored in the memory; the processor is used to execute at least one program in the memory to implement the model training and use methods for battery health status estimation provided by the above-mentioned method embodiments.

[0316] Fig.15 15 is a block diagram of a computer device 1500 provided by an exemplary embodiment of the present application. Generally, the computer device 1500 includes: a processor 1501 and a memory 1502 .

[0317] Processor 1501 may include one or more processing cores, such as a 4-core processor, an 8-core processor, etc. Processor 1501 may be implemented in at least one hardware form of digital signal processing (DSP), field programmable gate array (FPGA), and programmable logic array (PLA). Processor 1501 may also include a main processor and a coprocessor. The main processor is a processor for processing data in an awake state, also known as a central processing unit (CPU); the coprocessor is a low-power processor for processing data in a standby state. In some embodiments, processor 1501 may include a graphics processing unit (GPU), which is responsible for rendering and drawing the content to be displayed on the display screen. In some embodiments, processor 1501 may also include an artificial intelligence (AI) processor, which is used to process computing operations related to machine learning.

[0318] The memory 1502 may include one or more computer-readable storage media, which may be non-transitory. The memory 1502 may also include a high-speed random access memory, and a non-volatile memory, such as one or more disk storage devices, flash memory storage devices. In some embodiments, the non-transitory computer-readable storage medium in the memory 1502 is used to store at least one instruction, which is used to be executed by the processor 1501 to implement the model training and use method of battery health state estimation provided in the method embodiment of the present application.

[0319] In some embodiments, the server 1500 may also optionally include: an input interface 1503 and an output interface 1504. The processor 1501, the memory 1502, and the input interface 1503 and the output interface 1504 may be connected via a bus or a signal line. Each peripheral device may be connected to the input interface 1503 and the output interface 1504 via a bus, a signal line or a circuit board. The input interface 1503 and the output interface 1504 may be used to connect at least one peripheral device related to input / output to the processor 1501 and the memory 1502. In some embodiments, the processor 1501, the memory 1502, the input interface 1503, and the output interface 1504 are integrated on the same chip or circuit board; in some other embodiments, any one or two of the processor 1501, the memory 1502, the input interface 1503, and the output interface 1504 may be implemented on a separate chip or circuit board, which is not limited in the embodiments of the present application.

[0320] Those skilled in the art will appreciate that the structure shown above does not limit the computer device 1500 , and may include more or fewer components than shown in the figure, or combine certain components, or adopt a different component arrangement.

[0321] In an exemplary embodiment, a chip is also provided, which includes a programmable logic circuit and / or program instructions. When the chip runs on a computer device 1500, it is used to implement the model training and use method for battery health status estimation provided in an embodiment of the present application.

[0322] In an exemplary embodiment, a computer program product is also provided, the computer program product includes computer instructions, the computer instructions are stored in a computer-readable storage medium. The processor reads the computer instructions from the computer-readable storage medium, and the processor reads and executes the computer instructions from the computer-readable storage medium to implement the model training and use method for battery health state estimation provided in the embodiment of the present application.

[0323] In an exemplary embodiment, a computer-readable storage medium is also provided, in which at least one program is stored, and the at least one program is loaded and executed by a processor to implement the model training and use method for battery health status estimation provided in an embodiment of the present application.

[0324] A person skilled in the art will understand that all or part of the steps to implement the above embodiments may be accomplished by hardware or by instructing related hardware through a program, and the program may be stored in a computer-readable storage medium, and the above-mentioned storage medium may be a read-only memory, a disk or an optical disk, etc.

[0325] Those skilled in the art should be aware that in one or more of the above examples, the functions described in the embodiments of the present application can be implemented with hardware, software, firmware, or any combination thereof. When implemented using software, these functions can be stored in a computer-readable medium or transmitted as one or more instructions or codes on a computer-readable medium. Computer-readable media include computer storage media and communication media, wherein the communication media include any media that facilitates the transmission of a computer program from one place to another. The storage medium can be any available medium that a general or special-purpose computer can access.

[0326] The above are only optional embodiments of the present application and are not intended to limit the present application. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. A model training method for battery health state estimation, characterized in that: The model includes an adaptive feature extraction network and a selective domain adaptation network, and the method includes: Acquire a battery data set, wherein the battery data set includes battery sensor signals under at least two operating conditions; The battery sensor signal under the first working condition is used as source domain input data, and the source domain features are extracted by the adaptive feature extraction network; the battery sensor signal under the second working condition is used as target domain input data, and the target domain features are extracted by the adaptive feature extraction network; Calculating the distribution difference between the source domain features and the target domain features through the selective domain adaptation network; With the goal of reducing the distribution difference, the adaptive feature extraction network and the selective domain adaptation network are trained so that the adaptive feature extraction network extracts domain invariant features in the source domain and the target domain; The domain-invariant feature is used to represent the characteristics of the battery health status when it is not affected by the working conditions.

2. The method according to claim 1, characterized in that The selective domain adaptation network includes a shared fully connected layer shared by the source domain and the target domain; The calculating the distribution difference between the source domain feature and the target domain feature by the selective domain adaptation network includes: Calculating the source domain channel weight corresponding to the source domain feature and the target domain channel weight corresponding to the target domain feature through the shared fully connected layer; Based on the source domain features and the source domain channel weights, a source domain selective feature is calculated; based on the target domain features and the target domain channel weights, a target domain selective feature is calculated; The distribution difference between the source domain selective feature and the target domain selective feature is calculated.

3. The method according to claim 1 or 2, characterized in that: The adaptive feature extraction network includes a cascaded preceding convolutional layer, an adaptive feature transfer network and a succeeding convolutional layer; The method of using the battery sensor signal under the first working condition as source domain input data and extracting source domain features through the adaptive feature extraction network includes: Using the battery sensor signal under the first working condition as the source domain input data, and convolving the source domain input data through the previous convolution layer to obtain a first source domain feature vector; Performing adaptive feature screening on the first source domain feature vector through the adaptive feature transmission network to obtain a second source domain feature vector; The second source domain feature vector is convolved by the subsequent convolutional layer to obtain the source domain feature.

4. The method according to claim 3, characterized in that The adaptive feature transmission network includes: an average pooling layer, a maximum pooling layer, at least two fully connected layers and a first convolutional layer; The step of performing adaptive feature screening on the first source domain feature vector through the adaptive feature transmission network to obtain a second source domain feature vector includes: The first source domain feature vector is pooled by the average pooling layer and the maximum pooling layer respectively, and then added to obtain a first feature channel descriptor, where the first feature channel descriptor is used to describe the overall feature information of the first source domain feature vector; Processing the first feature channel descriptor through the at least two fully connected layers to obtain a source domain channel weight, where the source domain channel weight is used to describe the weight of the source domain feature channel; Convolving the first source domain feature vector through the first convolution layer to obtain a third source domain feature vector; Adaptively screening the third source domain feature vector based on the source domain channel weight to obtain the second source domain feature vector.

5. The method according to claim 1 or 2, characterized in that: The adaptive feature extraction network includes a cascaded preceding convolutional layer, an adaptive feature transfer network and a succeeding convolutional layer; The method of using the battery sensor signal under the second working condition as target domain input data and extracting target domain features through the adaptive feature extraction network includes: Using the battery sensor signal under the second working condition as the target domain input data, and convolving the target domain input data through the previous convolution layer to obtain a first target domain feature vector; Performing adaptive feature screening on the first target domain feature vector through the adaptive feature transmission network to obtain a second target domain feature vector; The second target domain feature vector is convolved through the subsequent convolutional layer to obtain the target domain feature.

6. The method according to claim 5, characterized in that The adaptive feature transmission network includes: an average pooling layer, a maximum pooling layer, at least two fully connected layers and a first convolutional layer; The step of performing adaptive feature screening on the first target domain feature vector through the adaptive feature transmission network to obtain a second target domain feature vector includes: The first target domain feature vector is pooled by the average pooling layer and the maximum pooling layer respectively, and then added to obtain a second feature channel descriptor, where the second feature channel descriptor is used to describe the overall feature information of the first target domain feature vector; Processing the second feature channel descriptor through the at least two fully connected layers to obtain a target domain channel weight, where the target domain channel weight is used to describe the weight of the target domain feature channel; Convolving the first target domain feature vector through the first convolution layer to obtain a third target domain feature vector; Adaptively screening the third target domain feature vector based on the target domain channel weight to obtain the second target domain feature vector.

7. The method according to claim 1 or 2, characterized in that: The battery data set includes an actual battery state of health corresponding to the battery sensor signal, and the model also includes a battery state of health estimation network; The method further comprises: Inputting the features of the battery sensor signal extracted by the adaptive feature extraction network into the battery health state estimation network for prediction to obtain a predicted battery health state; The step of training the adaptive feature extraction network and the selective domain adaptation network with the goal of reducing the distribution difference so that the adaptive feature extraction network extracts domain invariant features in the source domain and the target domain includes: Based on the sum of the first loss function and the second loss function, training the adaptive feature extraction network, the selective domain adaptation network, and the battery health state estimation network so that the adaptive feature extraction network extracts domain invariant features in the source domain and the target domain; The first loss function is used to calculate the error between the predicted battery health state and the actual battery health state, and the second loss function is used to calculate the distribution difference.

8. A method for using a model for battery health state estimation, characterized in that: The model includes an adaptive feature extraction network and a battery health state estimation network, and the method includes: Acquire a battery data set, wherein the battery data set includes a battery sensor signal under a second operating condition; Using the battery sensor signal under the second working condition as target domain input data, and extracting domain invariant features through the adaptive feature extraction network; Predicting the battery health state based on the domain invariant features by the battery health state estimation network; The domain-invariant feature is used to represent the characteristics of the battery health status when it is not affected by the working conditions.

9. The method according to claim 8, characterized in that The adaptive feature extraction network includes a cascaded preceding convolutional layer, an adaptive feature transfer network and a succeeding convolutional layer; The method of using the battery sensor signal under the second working condition as target domain input data and extracting domain invariant features through the adaptive feature extraction network includes: Using the battery sensor signal under the second working condition as the target domain input data, and convolving the target domain input data through the previous convolution layer to obtain a fourth target domain feature vector; Performing adaptive feature screening on the fourth target domain feature vector through the adaptive feature transmission network to obtain a fifth target domain feature vector; The fifth target domain feature vector is convolved by the subsequent convolutional layer to obtain the domain invariant feature.

10. The method according to claim 9, characterized in that The adaptive feature transmission network includes: an average pooling layer, a maximum pooling layer, at least two fully connected layers and a first convolutional layer; The step of performing adaptive feature screening on the fourth target domain feature vector through the adaptive feature transmission network to obtain a fifth target domain feature vector includes: The fourth target domain feature vector is pooled by the average pooling layer and the maximum pooling layer respectively, and then added to obtain a third feature channel descriptor, wherein the third feature channel descriptor is used to describe the overall feature information of the fourth target domain feature vector; Processing the third feature channel descriptor through the at least two fully connected layers to obtain a target domain channel weight, where the target domain channel weight is used to describe the weight of the target domain feature channel; Convolving the fourth target domain feature vector through the first convolution layer to obtain a sixth target domain feature vector; The fifth target domain feature vector is obtained based on the target domain channel weight and the sixth target domain feature vector.

11. A model training device for battery health state estimation, characterized in that: The model includes an adaptive feature extraction network and a selective domain adaptation network, and the device includes: An acquisition module, used to acquire a battery data set, wherein the battery data set includes battery sensor signals under at least two working conditions; An extraction module, used to use the battery sensor signal under the first working condition as source domain input data, and extract the source domain features through the adaptive feature extraction network; use the battery sensor signal under the second working condition as target domain input data, and extract the target domain features through the adaptive feature extraction network; A calculation module, configured to calculate a distribution difference between the source domain features and the target domain features through the selective domain adaptation network; A training module, configured to train the adaptive feature extraction network and the selective domain adaptation network with the goal of reducing the distribution difference, so that the adaptive feature extraction network extracts domain-invariant features in the source domain and the target domain; The domain-invariant feature is used to represent the characteristics of the battery health status when it is not affected by the working conditions.

12. A model using device for battery health state estimation, characterized in that: The model includes an adaptive feature extraction network and a battery health state estimation network, and the device includes: An acquisition module, configured to acquire a battery data set, wherein the battery data set includes a battery sensor signal under a second operating condition; an extraction module, configured to use the battery sensor signal under the second working condition as target domain input data, and extract domain invariant features through the adaptive feature extraction network; A prediction module, configured to predict the battery health state based on the domain invariant features through the battery health state estimation network; The domain-invariant feature is used to represent the characteristics of the battery health status when it is not affected by the working conditions.

13. A computer device, characterized in that: The computer device includes: a processor and a memory, wherein at least one program is stored in the memory; the processor is used to execute the at least one program in the memory to implement the model training method for battery health status estimation as described in any one of claims 1 to 7, and the model using method for battery health status estimation as described in any one of claims 8 to 10.

14. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores at least one program, and the at least one program is loaded and executed by a processor to implement the model training method for battery health state estimation as described in any one of claims 1 to 7, and the model using method for battery health state estimation as described in any one of claims 8 to 10.

15. A computer program product, characterized in that The computer program product includes computer instructions, which are stored in a computer-readable storage medium. The processor obtains the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions to implement the model training method for battery health state estimation as described in any one of claims 1 to 7, and the model using method for battery health state estimation as described in any one of claims 8 to 10.

Images