Residual life prediction model training method and device, computer equipment and storage medium

Abstract

The invention discloses a residual life prediction model training method and device, computer equipment and a storage medium, relates to the technical field of power battery management, and is used for solving the problem of low prediction accuracy caused by the fact that source domain selection mainly depends on expert experience and lacks interpretable judgment on migration effectiveness and possibly causes negative migration risks in the related technologies. The method comprises the following steps: acquiring multi-dimensional data characteristic indexes corresponding to databases of various types of power batteries; according to the multi-dimensional data characteristic index, determining a source domain and an auxiliary domain from a database of each type of power batteries; and performing cross-domain model training based on the source domain and the auxiliary domain to obtain a residual life prediction model.

Description

Technical Field

[0001] This invention relates to the field of power battery management technology, and in particular to a method, apparatus, computer equipment, and storage medium for training a remaining life prediction model. Background Technology

[0002] With the rapid development of electric vehicles and energy storage systems, power battery life prediction has become a key technology for ensuring the safe and reliable operation of power battery systems. Accurate life prediction can not only prevent safety accidents caused by power battery failure, but also optimize power battery maintenance strategies and significantly improve system economy. However, due to the short R&D cycle and high testing costs of new power batteries, it is often difficult to obtain sufficient aging data, which leads to a serious small sample dilemma for traditional data-driven models.

[0003] To overcome data bottlenecks, deep learning-based transfer learning methods have attracted widespread attention. This method assists in predicting the lifespan of novel power batteries (target domain) by transferring aging knowledge from existing power batteries (source domain). The accuracy of its predictions is highly dependent on the quality of the source domain selection. In related technologies, source domain selection mainly relies on expert experience, lacking interpretable judgment of the transfer's effectiveness, potentially leading to negative transfer risks and low prediction accuracy. Summary of the Invention

[0004] Based on this, a training method, apparatus, computer equipment, and storage medium for a remaining lifetime prediction model are provided to solve the problem that source domain selection in related technologies mainly relies on expert experience, lacks interpretable judgment on the effectiveness of transfer, may have negative transfer risk, and leads to low prediction accuracy.

[0005] In a first aspect, the present invention provides a method for training a remaining lifespan prediction model, the method comprising: Obtain multi-dimensional data characteristic indices corresponding to the databases of various types of power batteries, wherein the multi-dimensional data characteristic indices include at least the data abundance index, the label coverage index, and the data quality index; Based on the multi-dimensional data characteristic index, the source domain and auxiliary domain are determined from the database of each type of power battery; A cross-domain model is trained based on the source domain and the auxiliary domain to obtain a remaining lifetime prediction model.

[0006] In one embodiment, before obtaining the multi-dimensional data characteristic index corresponding to the database of each type of power battery, the method further includes: Collect historical battery data for different types of power batteries during each charging process, as well as the historical remaining life corresponding to the historical battery data; The historical battery data is used as sample data, and the historical remaining lifespan corresponding to the historical battery data is used as the remaining lifespan label value of the sample data. Based on the sample data and the remaining life tag values ​​of the sample data, a database of each type of power battery is formed.

[0007] In one embodiment, obtaining the multi-dimensional data characteristic index corresponding to the database of various types of power batteries includes: Obtain the total number of samples, time span, and number of operating condition types in the database for each type of power battery; Based on the total number of samples, the time span, and the number of operating condition types, the data abundance index corresponding to the database of each type of power battery is calculated.

[0008] In one embodiment, obtaining the multi-dimensional data characteristic index corresponding to the database of various types of power batteries further includes: The rule compliance rate of each type of power battery is calculated based on preset rules, wherein the preset rules include at least a preset numerical range; Obtain the time series corresponding to all sample data in the database of each type of power battery, calculate the actual interval between adjacent sample data based on the time series, and calculate the time continuity score corresponding to the database of each type of power battery based on the actual interval. Based on the rule compliance rate and the time continuity score, the data quality index corresponding to the database of each type of power battery is calculated.

[0009] In one embodiment, the step of training a cross-domain model based on the source domain and the auxiliary domain to obtain a remaining lifetime prediction model includes: Common features are extracted from the source domain and the auxiliary domain respectively to obtain the source domain feature matrix and the auxiliary domain feature matrix; Temporal modeling is performed on the source domain feature matrix and the auxiliary domain feature matrix to obtain the source domain temporal feature matrix and the auxiliary domain temporal feature matrix, wherein the source domain temporal feature matrix and the auxiliary domain temporal feature matrix are distributed and aligned. Regression prediction is performed based on the source domain time series feature matrix and the auxiliary domain time series feature matrix to obtain the remaining lifetime prediction value; The total loss value is determined based on the predicted remaining lifetime value and the remaining lifetime label value in the source domain, and the model is trained based on the total loss value to obtain the remaining lifetime prediction model.

[0010] In one embodiment, the step of extracting common features from the source domain and the auxiliary domain respectively to obtain a source domain feature matrix and an auxiliary domain feature matrix includes: Local temporal features are extracted from the source domain and the auxiliary domain respectively to obtain shallow features of the source domain and shallow features of the auxiliary domain; Cross-sensor joint feature extraction is performed on the shallow features of the source domain and the shallow features of the auxiliary domain respectively to obtain intermediate features of the source domain and intermediate features of the auxiliary domain. Cross-period degradation trend feature extraction is then performed on the intermediate features of the source domain and the intermediate features of the auxiliary domain respectively to obtain mid-layer features of the source domain and mid-layer features of the auxiliary domain. Lifetime decay features are extracted from the mid-layer features of the source domain and the mid-layer features of the auxiliary domain, respectively, to obtain deep features of the source domain and deep features of the auxiliary domain; Global temporal features are extracted from the deep features of the source domain and the deep features of the auxiliary domain respectively to obtain global temporal features of the source domain and global temporal features of the auxiliary domain. The source domain global temporal features and the auxiliary domain global temporal features are mapped to a common feature space to obtain the source domain feature matrix and the auxiliary domain feature matrix.

[0011] In one embodiment, determining the total loss value based on the predicted remaining lifetime value and the remaining lifetime label value in the source domain includes: Obtain the auxiliary domain labels corresponding to the auxiliary domain temporal feature matrix, and perform domain discrimination on the auxiliary domain temporal feature matrix to obtain the domain label prediction results; Calculate the deviation between the predicted domain label and the auxiliary domain label to obtain a first loss value; The deviation between the predicted remaining lifetime value and the labeled remaining lifetime value is calculated to obtain the second loss value; The total loss value is calculated based on the first loss value and the second loss value.

[0012] In a second aspect, the present invention provides a training apparatus for a remaining lifetime prediction model, the apparatus comprising: The acquisition module is used to acquire multi-dimensional data characteristic indices corresponding to the database of various types of power batteries, wherein the multi-dimensional data characteristic indices include at least the data abundance index, the label coverage index, and the data quality index. The determination module is used to determine the source domain and auxiliary domain from the database of each type of power battery based on the multi-dimensional data characteristic index. The model training module is used to perform cross-domain model training based on the source domain and the auxiliary domain to obtain the remaining lifetime prediction model.

[0013] Thirdly, the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the remaining lifetime prediction model training method of the first aspect described above.

[0014] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the remaining lifetime prediction model training method of the first aspect described above.

[0015] The aforementioned training method, apparatus, computer equipment, and storage medium for the remaining life prediction model, by quantitatively evaluating the multi-dimensional data characteristics corresponding to the databases of various types of power batteries, intelligently selects the source domain and auxiliary domain, thereby avoiding the negative transfer risk that may occur when relying on expert experience, and performs cross-domain model training based on the selected source domain and auxiliary domain to construct a remaining life prediction model with high generalization ability, which can significantly improve the accuracy of remaining life prediction for power batteries. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating a method for training a remaining lifetime prediction model in one embodiment. Figure 2 This is a structural block diagram of a training device for a remaining lifetime prediction model in one embodiment; Figure 3 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. The specific operational methods in the method embodiments can also be applied to the device embodiments or system embodiments. It should be noted that in the description of this invention, "multiple" is understood as "at least two". "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing together, or B existing alone. A connected to B can represent: A and B directly connected, or A and B connected through C. Furthermore, in the description of this invention, terms such as "first" and "second" are used only for distinguishing purposes and should not be construed as indicating or implying relative importance or order.

[0018] In this invention, the acquisition, transmission, storage, and use of data all comply with the requirements of relevant national laws and regulations.

[0019] Before introducing the training method for the remaining lifetime prediction model provided by this invention, the technical background of this invention will be described in detail below for ease of understanding.

[0020] Deep learning-based transfer learning methods have attracted widespread attention. This approach assists in predicting the lifespan of novel power batteries (target domain) by transferring aging knowledge from existing power batteries (source domain). The accuracy of the prediction is highly dependent on the quality of the source domain selection. In related technologies, source domain selection mainly relies on expert experience, lacking interpretable judgment of the transfer's effectiveness, potentially leading to negative transfer risks and low prediction accuracy.

[0021] Therefore, in one embodiment, such as Figure 1 The diagram shows a flowchart of a method for training a remaining lifespan prediction model, which includes the following steps: S101, obtain the multi-dimensional data characteristic index corresponding to the database of various types of power batteries; S102, Based on the multi-dimensional data characteristic index, the source domain and auxiliary domain are determined from the database of various types of power batteries; S103, based on the source domain and auxiliary domain, performs cross-domain model training to obtain the remaining lifetime prediction model.

[0022] The multi-dimensional data characteristic index includes at least the data abundance index, label coverage index, and data quality index. The label coverage index represents the proportion of first samples with remaining life label values ​​in the database of each type of power battery. The auxiliary domain refers to the subset of data used for domain adaptation training, which is partitioned from the database of each type of power battery based on the multi-dimensional data characteristic index to simulate the distribution difference between the target domain and the source domain in real migration scenarios.

[0023] For example, multi-dimensional data characteristic indices corresponding to databases of various types of power batteries are obtained. Based on these indices, source and auxiliary domains are determined from the databases of each type of power battery. For instance, target scores are calculated for each type of power battery database based on the multi-dimensional data characteristic indices, and the database with the highest target score is determined as the source domain, while the other databases are determined as auxiliary domains. Cross-domain model training is then performed based on the source and auxiliary domains to obtain a remaining life prediction model.

[0024] By using the above methods, the multi-dimensional data characteristics of databases corresponding to various types of power batteries are quantitatively evaluated, and source and auxiliary domains are intelligently selected to avoid the negative transfer risk that may occur when relying on expert experience. Based on the selected source and auxiliary domains, cross-domain model training is performed to build a high-generalization remaining life prediction model, which can significantly improve the accuracy of remaining life prediction for power batteries.

[0025] In one embodiment, exemplarily described, before obtaining the multi-dimensional data characteristic indices corresponding to the databases of various types of power batteries in S101, the method further includes: Collect historical battery data for different types of power batteries during each charging process, as well as the historical remaining lifespan corresponding to the historical battery data.

[0026] For example, first, determine the nominal parameters of the power battery, such as its rated capacity, charging cut-off voltage, and discharging cut-off voltage. Then, use a standard constant current / constant voltage charging protocol to perform charge-discharge cycles on the power battery. The charging process can be as follows: constant current charging is performed at a preset charging rate, and the power battery voltage slowly increases with charging time. When the power battery voltage reaches the charging cut-off voltage, the power battery enters the constant voltage charging stage. In this stage, the voltage remains constant at the charging cut-off voltage, and the charging current gradually decreases over time until it falls below the lower limit of the constant voltage current, at which point charging stops. The discharging process can be as follows: the power battery discharges with a constant current, and the voltage gradually decreases. When the voltage drops to the discharging cut-off voltage, discharging stops. This is considered one charge-discharge cycle of the power battery.

[0027] As the number of charge-discharge cycles increases, the capacity of the power battery gradually decreases. When the State of Health (SOH) of the power battery drops to a preset threshold (e.g., 0.7, the specific value of which depends on the situation and is not limited here) in a charge-discharge cycle, the power battery is considered to have failed, that is, the power battery has reached the end of its life (EOL). Here, SOH is the ratio of the actual usable capacity (the product of discharge current and discharge time) of the power battery to its rated capacity in a charge-discharge cycle. Therefore, the remaining useful life (RUL) of the power battery can be defined as the number of charge-discharge cycles that the power battery can still perform when its SOH degrades to EOL. Thus, primary characteristic parameters such as current, voltage, and temperature (ambient temperature or the battery's own temperature during operation) are recorded during each charging process as historical battery data, and the historical RUL corresponding to the historical battery capacity data is calculated based on the historical battery capacity data during each charging process.

[0028] Then, historical battery data is used as sample data, and the historical RUL corresponding to the historical battery data is used as the RUL tag value of the sample data. Thus, a database of various types of power batteries is formed based on the sample data and the RUL tag value of the sample data.

[0029] Using the methods described above, we systematically collected charge and discharge data for the entire life cycle of various power batteries and their corresponding RUL tags, and constructed a standardized database for multiple types of power batteries, providing a high-quality sample foundation for training data-driven RUL prediction models.

[0030] In one embodiment, it is exemplarily illustrated that the multi-dimensional data characteristic indices corresponding to the databases of various types of power batteries obtained in S101 include, but are not limited to: Obtain the total number of samples, time span, and number of operating condition types in the database for each type of power battery. The time span refers to the time range covered by the sample data in the database for each type of power battery. The number of operating condition types refers to the various combinations of working conditions encountered by the power battery during use. For example, the database needs to cover cross combinations of different charging rates and temperatures (such as fast charging at high temperatures, slow charging at low temperatures, etc., where 0.2C is slow charging, 1C is standard, 2C is fast charging, -20~0℃ is low temperature, 25℃ is normal temperature, and 40~60℃ is high temperature). The specific operating conditions depend on the situation and are not limited here.

[0031] Then, based on the total number of samples, time span, and number of operating conditions, the data abundance index corresponding to the database of each type of power battery is calculated.

[0032] For example, determine the first preset weight corresponding to the total number of samples, the second preset weight corresponding to the time span, and the third preset weight corresponding to the number of operating condition types. Simultaneously, normalize the total number of samples, the time span, and the number of operating condition types to obtain the normalized total number of samples, the normalized time span, and the normalized number of operating condition types. The normalization method can be deviation standardization. The following example illustrates the normalization of the sample number: Determine the minimum value N of the total number of samples N in the database for each type of power battery. min and maximum value N max Based on N min and N max Normalize N in the database of various types of power batteries to obtain the normalized total number of samples Norm(N) = (NN) min ) / (N max -N min ).

[0033] Therefore, based on the first preset weight, the second preset weight, and the third preset weight, a weighted summation operation is performed on Norm(N), the normalized time span, and the normalized number of operating condition types to obtain the data abundance index corresponding to the database of each type of power battery.

[0034] Using the above methods, the total number of samples, time span, and number of operating conditions in the databases of various types of power batteries are quantitatively statistically analyzed. Based on these multidimensional indicators, a standardized data abundance index is calculated, thereby providing a quantifiable decision-making basis for source domain selection in transfer learning.

[0035] In one embodiment, exemplarily illustrating that obtaining the multi-dimensional data characteristic indices corresponding to the databases of various types of power batteries in S101 further includes: First, the rule compliance rate for each type of power battery in the database is calculated based on preset rules. These preset rules include at least a preset numerical range (e.g., voltage range of 2.5-4.2V, charging current less than 3C, temperature range of -20-60℃; specific rules vary depending on the situation and are not limited here). For example, the number of second samples in the database for each type of power battery that fall within the preset numerical range is counted, and this second sample number is divided by the total number of samples to obtain the rule compliance rate for each type of power battery in the database.

[0036] Simultaneously, the time series corresponding to all sample data in the database of each type of power battery is obtained, and the actual interval between adjacent sample data is calculated based on the time series, and the time continuity score corresponding to the database of each type of power battery is calculated based on the actual interval.

[0037] For example, obtain the theoretical intervals (calculated based on the theoretical sampling frequency) corresponding to all sample data in the database of each type of power battery, and then count the percentage of intervals where the actual interval is equal to the theoretical interval, and determine this percentage as the time continuity score.

[0038] Finally, based on the rule compliance rate and time continuity score, the data quality index corresponding to the database of each type of power battery is calculated. For example, the fourth preset weight corresponding to the rule compliance rate and the fifth preset weight corresponding to the time continuity score are obtained, and the rule compliance rate and time continuity score are weighted and summed based on the fourth preset weight and the fifth preset weight to obtain the data quality index corresponding to the database of each type of power battery.

[0039] Using the above methods, the rule compliance rate of databases for each type of power battery is calculated, and the time continuity score is quantified by combining time series analysis. Finally, the data quality index corresponding to the databases of each type of power battery is generated, realizing the automated and standardized evaluation of power battery data quality, thereby providing a quantifiable decision basis for source domain selection in transfer learning.

[0040] In one embodiment, it is exemplarily illustrated that in S102, the source domain and auxiliary domain are determined from the database of various types of power batteries based on multi-dimensional data characteristic indices, including but not limited to: The first objective weight corresponding to the data abundance index, the second objective weight corresponding to the label coverage index, and the third objective weight corresponding to the data quality index are determined. Based on the first objective weight, the second objective weight, and the third objective weight, the data abundance index, the label coverage index, and the data quality index are weighted and summed to obtain the target scores corresponding to the databases of each type of power battery. The power battery database with the highest score among the target scores is determined as the source domain, and the other power battery databases are determined as auxiliary domains.

[0041] Using the above method, the multi-dimensional characteristic indices of the power battery database are weighted and scored to obtain the target score. The source domain is then automatically selected based on the target score, realizing data-driven objective domain selection. This effectively overcomes the subjectivity and one-sidedness of the traditional expert experience method and significantly improves the representativeness of source domain selection in transfer learning.

[0042] In one embodiment, it is exemplarily illustrated that in S103, cross-domain model training based on the source domain and auxiliary domain is performed to obtain a remaining life prediction model, including but not limited to: Construct an initial prediction model, which includes a common feature extractor E. c Domain Feature Extractor E d The source domain and auxiliary domain are used to train the initial prediction model, thereby obtaining the remaining life prediction model.

[0043] For example, the source domain is represented as Where s represents the source domain index, This represents the i-th sample data in the source domain. Represents the RUL label value corresponding to the i-th sample data in the source domain, n s The sample size in the source domain is represented by ρ. The marginal probability distribution of sample matching in the source domain can be expressed as ρ. s For p auxiliary fields, it can be represented as Where t represents the auxiliary field index, This represents the j-th sample data in the m-th auxiliary field. This represents the RUL label value corresponding to the j-th sample data in the m-th auxiliary domain, where n t，m The marginal probability distribution of the m-th auxiliary domain can be represented by ρ. t，m This indicates that p is an integer greater than 0.

[0044] First, regarding D s The time series X in s and D t，m The time series X in t，m Through a sliding window T of the same length w The data is divided to obtain source domain time-series blocks. and auxiliary domain timing blocks Simultaneously, a one-dimensional 16-layer deep convolutional neural network model (One Dimensional Visual Geometry Group 16, 1D-VGG16) was constructed as an E c For the input and Common features are extracted to obtain the source domain feature matrix. and auxiliary domain feature matrix , of which E c Other convolutional neural network models (such as residual networks, lightweight deep learning models, etc.) can also be used, depending on the situation, and no specific model is specified here.

[0045] Then, and The source-auxiliary domain characteristic matrix pair - Enter into E d Temporal modeling is performed to obtain the source domain temporal feature matrix. and auxiliary domain temporal feature matrix Among them, E d It can be composed of multiple bidirectional long short-term memory (BiLSTM) networks or multiple LSTM networks; the specific network structure depends on the situation and is not limited here. Thus, through E... d Learning within multiple scales - The time characteristics, and and Mapping to different feature subspaces yields domain-invariant features. and .

[0046] Then and Input the corresponding r to perform regression prediction and obtain the RUL predicted value. Each - The corresponding r are not shared between each other. r can be a fully connected regression predictor. The specific structure of r depends on the situation and is not limited here.

[0047] Finally, based on the RUL predicted value and D s The total loss value L is determined by the RUL label value, and the parameters of the initial prediction model are optimized based on minimizing L to obtain the RUL prediction model.

[0048] Using the above methods, common features are extracted and time-series modeling is performed on the source and auxiliary domains to achieve distribution alignment between the source and auxiliary domains. By combining regression prediction and loss optimization, an adaptive cross-domain migration RUL prediction model is constructed, which significantly improves the RUL prediction accuracy under different power battery types or operating conditions.

[0049] In one embodiment, it is exemplarily illustrated that common features are extracted from the source domain and the auxiliary domain to obtain a source domain feature matrix and an auxiliary domain feature matrix, including but not limited to: The source and auxiliary domains are windowed separately to obtain source domain time series blocks and auxiliary domain time series blocks. These blocks are then input into 1D-VGG16, which outputs source domain feature matrices and auxiliary domain feature matrices. 1D-VGG16 contains 13 convolutional layers and 3 fully connected layers, all of which are one-dimensional convolutional layers. The feature extraction process can be as follows: Two convolutional layers with K=3 kernels and 64 channels are used to extract local temporal features (such as voltage jitter and current peaks within a single charge / discharge cycle) from the source and auxiliary domain temporal blocks, respectively, to obtain shallow features in the source and auxiliary domains. Then, a first pooling layer is used to pool these shallow features, resulting in pooled shallow features in the source and auxiliary domains, thereby reducing the dimensionality of the shallow features and enhancing their local translation invariance.

[0050] Two convolutional layers with K=3 and 128 channels are used to perform cross-sensor joint feature extraction (such as voltage-temperature dynamic correlation, current-voltage-temperature multi-physical quantity coupling, etc.) on the shallow features of the source domain and the shallow features of the auxiliary domain after pooling, respectively, to obtain intermediate features of the source domain and the auxiliary domain. Then, a second pooling layer is used to pool the intermediate features of the source domain and the auxiliary domain to obtain pooled intermediate features of the source domain and the auxiliary domain, thereby reducing the dimensionality of the intermediate features and enhancing the translation invariance.

[0051] Three convolutional layers with K=3 and 256 channels are used to extract cross-cycle degradation trend features (such as the capacity decay slope of a power battery) from the pooled intermediate features of the source domain and the pooled intermediate features of the auxiliary domain, respectively, to obtain the intermediate features of the source domain and the auxiliary domain. A third pooling layer is then used to further pool these intermediate features, resulting in pooled intermediate features of the source domain and the auxiliary domain, thereby reducing the dimensionality of the intermediate features and enhancing their translation invariance.

[0052] By using three convolutional layers with K=3 and 512 channels, lifetime decay features are extracted from the pooled mid-level features of the source domain and the pooled mid-level features of the auxiliary domain, respectively, to obtain deep features of the source domain and the auxiliary domain. Then, a fourth pooling layer is used to further pool these deep features, resulting in pooled deep features of the source domain and the auxiliary domain, thereby reducing the dimensionality of the deep features and enhancing their translation invariance.

[0053] Three convolutional layers with K=3 and 512 channels are used to extract global temporal features from the pooled deep features of the source domain and the pooled deep features of the auxiliary domain, respectively. This yields global temporal features of the source domain and the auxiliary domain. Global temporal features refer to comprehensive high-order features that can fully characterize the lifespan degradation of the power battery throughout its entire lifecycle. A fifth pooling layer is then used to pool these global temporal features, resulting in pooled global temporal features of the source domain and the auxiliary domain, thereby reducing the dimensionality of the global temporal features and enhancing their translation invariance.

[0054] By using three fully connected layers FC 1024, the pooled global temporal features of the source domain and the pooled global temporal features of the auxiliary domain are mapped to a common feature space to obtain the source domain feature matrix and the auxiliary domain feature matrix.

[0055] The above method abstracts the degradation information of the power battery from local to global levels based on the hierarchical feature extraction strategy, thereby decoupling the life decay features across operating conditions and batteries. Furthermore, the feature distribution of the two domains is forcibly aligned by mapping the common feature space, reducing the deviation caused by sensor differences, environmental noise, or different operating conditions, and significantly improving the generalization ability of the model.

[0056] In one embodiment, it is illustrated by way of example that the total loss value is determined based on the RUL predicted value and the RUL label value in the source domain, including but not limited to: Obtain the auxiliary domain label corresponding to the auxiliary domain time series feature matrix. For example, use the auxiliary domain index t as the auxiliary domain label, or map t to the value 0 and use 0 as the auxiliary domain label. The specific method of obtaining the label depends on the situation and is not limited here.

[0057] Then, a domain discriminator is used to perform domain discrimination on the auxiliary domain temporal feature matrix to obtain the domain label prediction result. The domain discriminator can be a multi-layer gated recurrent unit (GRU). The multi-layer GRU performs domain discrimination in a progressive manner: first, the bottom-layer GRU receives the auxiliary domain temporal feature matrix and learns the basic time series pattern; then, its output is passed to the upper-layer GRU for higher-level abstraction and discrimination. Each layer of GRU captures domain-specific information at different time scales through a gating mechanism. Finally, the output of the top-layer GRU, after passing through a classification layer, yields the domain label prediction result. The entire process aims to distinguish whether the input comes from the source domain or the auxiliary domain, and to correlate with E. d Adversarial training is employed to facilitate the learning of domain-invariant features.

[0058] Finally, the deviation between the domain label prediction result and the auxiliary domain label is calculated to obtain the first loss value L1, and the deviation between the RUL prediction value and the RUL label value is calculated to obtain the second loss value L2. Thus, based on L1 and L2, L = L2 - L1 is calculated.

[0059] Using the above method, the domain discrimination loss of the auxiliary domain temporal feature matrix is ​​calculated, and the total loss value is obtained by combining it with the RUL prediction loss. This achieves collaborative training of domain adaptation and RUL prediction, effectively reducing the distribution difference between the source domain and the auxiliary domain, while improving the prediction accuracy of the RUL prediction model in the auxiliary domain.

[0060] It should be understood that, although Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order in which these steps are executed, and they can be performed in other orders. Furthermore, Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0061] In one embodiment, such as Figure 2 As shown, a remaining lifespan prediction model training device is provided, including: an acquisition module 201, a determination module 202, and a model training module 203, wherein: The acquisition module 201 is used to acquire the multi-dimensional data characteristic index corresponding to the database of various types of power batteries. The multi-dimensional data characteristic index includes at least the data abundance index, the label coverage index, and the data quality index. The determination module 202 is used to determine the source domain and auxiliary domain from the database of various types of power batteries based on the multi-dimensional data characteristic index. The model training module 203 is used to perform cross-domain model training based on the source domain and the auxiliary domain to obtain the remaining life prediction model.

[0062] In one embodiment, the acquisition module 201 is further configured to: Collect historical battery data for different types of power batteries during each charging process, as well as the historical remaining lifespan corresponding to the historical battery data; Historical battery data is used as sample data, and the historical remaining lifespan corresponding to the historical battery data is used as the remaining lifespan label value of the sample data. A database of various types of power batteries is formed based on sample data and the remaining life tag values ​​of the sample data.

[0063] In one embodiment, the acquisition module 201 is further configured to: Obtain the total number of samples, time span, and number of operating condition types in the database for each type of power battery; Based on the total number of samples, time span, and number of operating conditions, the data abundance index corresponding to the database of each type of power battery is calculated.

[0064] In one embodiment, the acquisition module 201 is further configured to: The rule compliance rate of each type of power battery database is calculated based on preset rules, wherein the preset rules include at least a preset numerical range. Obtain the time series corresponding to all sample data in the database of each type of power battery, calculate the actual interval between adjacent sample data based on the time series, and calculate the time continuity score corresponding to the database of each type of power battery based on the actual interval. Based on rule compliance rate and time continuity score, calculate the data quality index corresponding to the database of each type of power battery.

[0065] In one embodiment, the model training module 203 is used for: Common features are extracted from the source domain and the auxiliary domain respectively to obtain the source domain feature matrix and the auxiliary domain feature matrix; Temporal modeling is performed on the source domain feature matrix and the auxiliary domain feature matrix to obtain the source domain temporal feature matrix and the auxiliary domain temporal feature matrix, wherein the source domain temporal feature matrix and the auxiliary domain temporal feature matrix are aligned in distribution; Regression prediction is performed based on the source domain time series feature matrix and the auxiliary domain time series feature matrix to obtain the remaining lifetime prediction value; The total loss value is determined based on the remaining lifetime prediction value and the remaining lifetime label value in the source domain, and the model is trained based on the total loss value to obtain the remaining lifetime prediction model.

[0066] In one embodiment, the model training module 203 is further configured to: Local temporal features are extracted from the source domain and the auxiliary domain respectively to obtain shallow features of the source domain and shallow features of the auxiliary domain; Cross-sensor joint feature extraction is performed on the shallow features of the source domain and the shallow features of the auxiliary domain to obtain the intermediate features of the source domain and the intermediate features of the auxiliary domain. Cross-period degradation trend feature extraction is then performed on the intermediate features of the source domain and the intermediate features of the auxiliary domain to obtain the middle features of the source domain and the middle features of the auxiliary domain. Lifetime decay features are extracted from the mid-layer features of the source domain and the mid-layer features of the auxiliary domain, respectively, to obtain deep features of the source domain and deep features of the auxiliary domain. Global temporal features are extracted from the deep features of the source domain and the deep features of the auxiliary domain respectively to obtain the global temporal features of the source domain and the global temporal features of the auxiliary domain. By performing common feature space mapping on the global temporal features of the source domain and the global temporal features of the auxiliary domain respectively, the feature matrices of the source domain and the auxiliary domain are obtained.

[0067] In one embodiment, the model training module 203 is further configured to: Obtain the auxiliary domain labels corresponding to the auxiliary domain temporal feature matrix, and perform domain discrimination on the auxiliary domain temporal feature matrix to obtain the domain label prediction results; The deviation between the predicted domain label and the auxiliary domain label is calculated to obtain the first loss value; The deviation between the predicted remaining lifetime and the labeled remaining lifetime is calculated to obtain the second loss value; Calculate the total loss value based on the first loss value and the second loss value.

[0068] Specific limitations regarding the training device for the remaining lifetime prediction model can be found in the limitations on the training method for the remaining lifetime prediction model mentioned above, and will not be repeated here. Each module in the aforementioned training device for the remaining lifetime prediction model can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0069] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 3 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The database stores training data for a remaining lifetime prediction model. The network interface communicates with external terminals via a network connection. When the computer program is executed by the processor, it implements a remaining lifetime prediction model training method. The display screen can be an LCD screen or an e-ink display screen. The input device can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device casing, or an external keyboard, touchpad, or mouse.

[0070] Those skilled in the art will understand that Figure 3The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the computer device to which the present invention is applied. A specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0071] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps: Obtain multi-dimensional data characteristic indices corresponding to the databases of various types of power batteries. Among them, the multi-dimensional data characteristic indices include at least the data abundance index, the label coverage index, and the data quality index. Based on multi-dimensional data characteristic indices, the source domain and auxiliary domain are determined from the database of various types of power batteries; A remaining lifetime prediction model is obtained by training a cross-domain model based on the source domain and the auxiliary domain.

[0072] In one embodiment, the processor, when executing a computer program, also performs the following steps: Collect historical battery data for different types of power batteries during each charging process, as well as the historical remaining lifespan corresponding to the historical battery data; Historical battery data is used as sample data, and the historical remaining lifespan corresponding to the historical battery data is used as the remaining lifespan label value of the sample data. A database of various types of power batteries is formed based on sample data and the remaining life tag values ​​of the sample data.

[0073] In one embodiment, the processor, when executing a computer program, also performs the following steps: Obtain the total number of samples, time span, and number of operating condition types in the database for each type of power battery; Based on the total number of samples, time span, and number of operating conditions, the data abundance index corresponding to the database of each type of power battery is calculated.

[0074] In one embodiment, the processor, when executing a computer program, also performs the following steps: The rule compliance rate of each type of power battery database is calculated based on preset rules, wherein the preset rules include at least a preset numerical range. Obtain the time series corresponding to all sample data in the database of each type of power battery, calculate the actual interval between adjacent sample data based on the time series, and calculate the time continuity score corresponding to the database of each type of power battery based on the actual interval. Based on rule compliance rate and time continuity score, calculate the data quality index corresponding to the database of each type of power battery.

[0075] In one embodiment, the processor, when executing a computer program, also performs the following steps: Common features are extracted from the source domain and the auxiliary domain respectively to obtain the source domain feature matrix and the auxiliary domain feature matrix; Temporal modeling is performed on the source domain feature matrix and the auxiliary domain feature matrix to obtain the source domain temporal feature matrix and the auxiliary domain temporal feature matrix, wherein the source domain temporal feature matrix and the auxiliary domain temporal feature matrix are aligned in distribution; Regression prediction is performed based on the source domain time series feature matrix and the auxiliary domain time series feature matrix to obtain the remaining lifetime prediction value; The total loss value is determined based on the remaining lifetime prediction value and the remaining lifetime label value in the source domain, and the model is trained based on the total loss value to obtain the remaining lifetime prediction model.

[0076] In one embodiment, the processor, when executing a computer program, also performs the following steps: Local temporal features are extracted from the source domain and the auxiliary domain respectively to obtain shallow features of the source domain and shallow features of the auxiliary domain; Cross-sensor joint feature extraction is performed on the shallow features of the source domain and the shallow features of the auxiliary domain to obtain the intermediate features of the source domain and the intermediate features of the auxiliary domain. Cross-period degradation trend feature extraction is then performed on the intermediate features of the source domain and the intermediate features of the auxiliary domain to obtain the middle features of the source domain and the middle features of the auxiliary domain. Lifetime decay features are extracted from the mid-layer features of the source domain and the mid-layer features of the auxiliary domain, respectively, to obtain deep features of the source domain and deep features of the auxiliary domain. Global temporal features are extracted from the deep features of the source domain and the deep features of the auxiliary domain respectively to obtain the global temporal features of the source domain and the global temporal features of the auxiliary domain. By performing common feature space mapping on the global temporal features of the source domain and the global temporal features of the auxiliary domain respectively, the feature matrices of the source domain and the auxiliary domain are obtained.

[0077] In one embodiment, the processor, when executing a computer program, also performs the following steps: Obtain the auxiliary domain labels corresponding to the auxiliary domain temporal feature matrix, and perform domain discrimination on the auxiliary domain temporal feature matrix to obtain the domain label prediction results; The deviation between the predicted domain label and the auxiliary domain label is calculated to obtain the first loss value; The deviation between the predicted remaining lifetime and the labeled remaining lifetime is calculated to obtain the second loss value; Calculate the total loss value based on the first loss value and the second loss value.

[0078] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor: Obtain multi-dimensional data characteristic indices corresponding to the databases of various types of power batteries. Among them, the multi-dimensional data characteristic indices include at least the data abundance index, the label coverage index, and the data quality index. Based on multi-dimensional data characteristic indices, the source domain and auxiliary domain are determined from the database of various types of power batteries; A remaining lifetime prediction model is obtained by training a cross-domain model based on the source domain and the auxiliary domain.

[0079] In one embodiment, when the computer program is executed by a processor, it also performs the following steps: Collect historical battery data for different types of power batteries during each charging process, as well as the historical remaining lifespan corresponding to the historical battery data; Historical battery data is used as sample data, and the historical remaining lifespan corresponding to the historical battery data is used as the remaining lifespan label value of the sample data. A database of various types of power batteries is formed based on sample data and the remaining life tag values ​​of the sample data.

[0080] In one embodiment, when the computer program is executed by a processor, it also performs the following steps: Obtain the total number of samples, time span, and number of operating condition types in the database for each type of power battery; Based on the total number of samples, time span, and number of operating conditions, the data abundance index corresponding to the database of each type of power battery is calculated.

[0081] In one embodiment, when the computer program is executed by a processor, it also performs the following steps: The rule compliance rate of each type of power battery database is calculated based on preset rules, wherein the preset rules include at least a preset numerical range. Obtain the time series corresponding to all sample data in the database of each type of power battery, calculate the actual interval between adjacent sample data based on the time series, and calculate the time continuity score corresponding to the database of each type of power battery based on the actual interval. Based on rule compliance rate and time continuity score, calculate the data quality index corresponding to the database of each type of power battery.

[0082] In one embodiment, when the computer program is executed by a processor, it also performs the following steps: Common features are extracted from the source domain and the auxiliary domain respectively to obtain the source domain feature matrix and the auxiliary domain feature matrix; Temporal modeling is performed on the source domain feature matrix and the auxiliary domain feature matrix to obtain the source domain temporal feature matrix and the auxiliary domain temporal feature matrix, wherein the source domain temporal feature matrix and the auxiliary domain temporal feature matrix are aligned in distribution; Regression prediction is performed based on the source domain time series feature matrix and the auxiliary domain time series feature matrix to obtain the remaining lifetime prediction value; The total loss value is determined based on the remaining lifetime prediction value and the remaining lifetime label value in the source domain, and the model is trained based on the total loss value to obtain the remaining lifetime prediction model.

[0083] In one embodiment, when the computer program is executed by a processor, it also performs the following steps: Local temporal features are extracted from the source domain and the auxiliary domain respectively to obtain shallow features of the source domain and shallow features of the auxiliary domain; Cross-sensor joint feature extraction is performed on the shallow features of the source domain and the shallow features of the auxiliary domain to obtain the intermediate features of the source domain and the intermediate features of the auxiliary domain. Cross-period degradation trend feature extraction is then performed on the intermediate features of the source domain and the intermediate features of the auxiliary domain to obtain the middle features of the source domain and the middle features of the auxiliary domain. Lifetime decay features are extracted from the mid-layer features of the source domain and the mid-layer features of the auxiliary domain, respectively, to obtain deep features of the source domain and deep features of the auxiliary domain. Global temporal features are extracted from the deep features of the source domain and the deep features of the auxiliary domain respectively to obtain the global temporal features of the source domain and the global temporal features of the auxiliary domain. By performing common feature space mapping on the global temporal features of the source domain and the global temporal features of the auxiliary domain respectively, the feature matrices of the source domain and the auxiliary domain are obtained.

[0084] In one embodiment, when the computer program is executed by a processor, it also performs the following steps: Obtain the auxiliary domain labels corresponding to the auxiliary domain temporal feature matrix, and perform domain discrimination on the auxiliary domain temporal feature matrix to obtain the domain label prediction results; The deviation between the predicted domain label and the auxiliary domain label is calculated to obtain the first loss value; The deviation between the predicted remaining lifetime and the labeled remaining lifetime is calculated to obtain the second loss value; Calculate the total loss value based on the first loss value and the second loss value.

[0085] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0086] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0087] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A method for training a remaining lifespan prediction model, characterized in that, The method includes: Obtain multi-dimensional data characteristic indices corresponding to the databases of various types of power batteries, wherein the multi-dimensional data characteristic indices include at least the data abundance index, the label coverage index, and the data quality index; Based on the multi-dimensional data characteristic index, the source domain and auxiliary domain are determined from the database of each type of power battery; A cross-domain model is trained based on the source domain and the auxiliary domain to obtain a remaining lifetime prediction model.

2. The method according to claim 1, characterized in that, Before obtaining the multi-dimensional data characteristic indices corresponding to the databases of various types of power batteries, the process also includes: Collect historical battery data for different types of power batteries during each charging process, as well as the historical remaining life corresponding to the historical battery data; The historical battery data is used as sample data, and the historical remaining lifespan corresponding to the historical battery data is used as the remaining lifespan label value of the sample data. Based on the sample data and the remaining life tag values ​​of the sample data, a database of each type of power battery is formed.

3. The method according to claim 1, characterized in that, The process of obtaining multi-dimensional data characteristic indices corresponding to the databases of various types of power batteries includes: Obtain the total number of samples, time span, and number of operating condition types in the database for each type of power battery; Based on the total number of samples, the time span, and the number of operating condition types, the data abundance index corresponding to the database of each type of power battery is calculated.

4. The method according to claim 1, characterized in that, The method of obtaining multi-dimensional data characteristic indices corresponding to the databases of various types of power batteries also includes: The rule compliance rate of each type of power battery is calculated based on preset rules, wherein the preset rules include at least a preset numerical range; Obtain the time series corresponding to all sample data in the database of each type of power battery, calculate the actual interval between adjacent sample data based on the time series, and calculate the time continuity score corresponding to the database of each type of power battery based on the actual interval. Based on the rule compliance rate and the time continuity score, the data quality index corresponding to the database of each type of power battery is calculated.

5. The method according to claim 1, characterized in that, The method of training a cross-domain model based on the source domain and the auxiliary domain to obtain a remaining lifetime prediction model includes: Common features are extracted from the source domain and the auxiliary domain respectively to obtain the source domain feature matrix and the auxiliary domain feature matrix; Temporal modeling is performed on the source domain feature matrix and the auxiliary domain feature matrix to obtain the source domain temporal feature matrix and the auxiliary domain temporal feature matrix, wherein the source domain temporal feature matrix and the auxiliary domain temporal feature matrix are distributed and aligned. Regression prediction is performed based on the source domain time series feature matrix and the auxiliary domain time series feature matrix to obtain the remaining lifetime prediction value; The total loss value is determined based on the predicted remaining lifetime value and the remaining lifetime label value in the source domain, and the model is trained based on the total loss value to obtain the remaining lifetime prediction model.

6. The method according to claim 5, characterized in that, The step of extracting common features from the source domain and the auxiliary domain respectively to obtain the source domain feature matrix and the auxiliary domain feature matrix includes: Local temporal features are extracted from the source domain and the auxiliary domain respectively to obtain shallow features of the source domain and shallow features of the auxiliary domain; Cross-sensor joint feature extraction is performed on the shallow features of the source domain and the shallow features of the auxiliary domain respectively to obtain intermediate features of the source domain and intermediate features of the auxiliary domain. Cross-period degradation trend feature extraction is then performed on the intermediate features of the source domain and the intermediate features of the auxiliary domain respectively to obtain mid-layer features of the source domain and mid-layer features of the auxiliary domain. Lifetime decay features are extracted from the mid-layer features of the source domain and the mid-layer features of the auxiliary domain, respectively, to obtain deep features of the source domain and deep features of the auxiliary domain; Global temporal features are extracted from the deep features of the source domain and the deep features of the auxiliary domain respectively to obtain global temporal features of the source domain and global temporal features of the auxiliary domain. The source domain global temporal features and the auxiliary domain global temporal features are mapped to a common feature space to obtain the source domain feature matrix and the auxiliary domain feature matrix.

7. The method according to claim 5, characterized in that, The step of determining the total loss value based on the predicted remaining lifetime value and the remaining lifetime label value in the source domain includes: Obtain the auxiliary domain labels corresponding to the auxiliary domain temporal feature matrix, and perform domain discrimination on the auxiliary domain temporal feature matrix to obtain the domain label prediction results; Calculate the deviation between the predicted domain label and the auxiliary domain label to obtain a first loss value; The deviation between the predicted remaining lifetime value and the labeled remaining lifetime value is calculated to obtain the second loss value; The total loss value is calculated based on the first loss value and the second loss value.

8. A training device for a remaining lifespan prediction model, characterized in that, The device includes: The acquisition module is used to acquire multi-dimensional data characteristic indices corresponding to the database of various types of power batteries, wherein the multi-dimensional data characteristic indices include at least the data abundance index, the label coverage index, and the data quality index. The determination module is used to determine the source domain and auxiliary domain from the database of each type of power battery based on the multi-dimensional data characteristic index. The model training module is used to perform cross-domain model training based on the source domain and the auxiliary domain to obtain the remaining lifetime prediction model.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

Images