Battery online remote status monitoring method and system based on autonomous learning

Abstract

The present invention provides a method and system for online remote status monitoring of batteries based on autonomous learning, which relates to the field of battery monitoring technology. The method and system include: obtaining historical monitoring data of source and target domains corresponding to a target battery, constructing a time series data set, constructing a first prediction model and a second prediction model, determining a loss value of the prediction model and optimizing parameters of the prediction model through a pre-selected optimizer to obtain an initial prediction model; training the initial prediction model on source domain data to obtain pre-trained parameters, constructing a parameter migration graph network with the pre-trained parameters as nodes and updating the nodes, obtaining a migration strategy function by determining the dependency relationship between parameters, migrating the pre-trained parameters to the initial prediction model to obtain a decay prediction model; calculating the maximum mean difference, constructing a second loss function, updating the hyperparameters of the decay prediction model and adapting the decay prediction model to the target domain to obtain a capacity decay prediction result.

Description

Technical Field

[0001] The present invention relates to the technical field of battery monitoring, and in particular to a method and system for online remote status monitoring of a battery based on autonomous learning. Background Art

[0002] With the development of battery technology and the continuous expansion of its application fields, batteries have been widely used in various devices and systems. In order to ensure the reliability and safety of batteries, effective status monitoring is crucial. Traditional battery monitoring methods mainly rely on manual regular inspection and maintenance, which has problems such as low monitoring efficiency and poor real-time performance.

[0003] In recent years, with the development of technologies such as the Internet of Things, big data, and artificial intelligence, remote online monitoring has become a growing trend in battery condition monitoring. Several existing solutions for remote online battery monitoring exist. These solutions typically use sensors to collect battery parameters such as voltage, current, and temperature, and transmit the data via wireless communication to a remote server for analysis and processing.

[0004] However, existing remote online battery monitoring technologies still have some limitations. For example, most solutions rely on pre-set thresholds and rules to determine status, lacking adaptability and flexibility. For complex battery systems, existing solutions struggle to achieve comprehensive and accurate status estimation and fault diagnosis. Existing solutions often require extensive manual intervention, such as parameter setting and model maintenance, resulting in insufficient autonomy and intelligence.

[0005] Therefore, there is an urgent need for a solution that can adaptively learn the state characteristics of batteries, achieve accurate state estimation and fault diagnosis, reduce manual participation, and improve the autonomy and intelligence level of the system. Summary of the Invention

[0006] The embodiments of the present invention provide a method and system for online remote status monitoring of a battery based on autonomous learning, which can at least solve some of the problems existing in the prior art.

[0007] A first aspect of an embodiment of the present invention provides a method for online remote status monitoring of a battery based on autonomous learning, comprising:

[0008] Acquire historical monitoring data of the source domain and the target domain corresponding to the target battery, construct a time series data set, construct a first prediction model corresponding to the source domain and a second prediction model corresponding to the target domain based on the time series data set, determine the loss value of the prediction model by combining the first loss function, and optimize the parameters of the prediction model using a pre-selected optimizer to obtain an initial prediction model;

[0009] The initial prediction model is trained on source domain data to obtain pre-trained parameters, a parameter migration graph network is constructed using the pre-trained parameters as nodes, and nodes are updated, a migration strategy function is obtained by determining dependencies between parameters based on the updated nodes, and the pre-trained parameters are migrated to the initial prediction model based on the migration strategy function to obtain a decay prediction model;

[0010] For the attenuation prediction model, the maximum mean difference is obtained by calculating the difference in the battery data feature distribution in the source domain and the target domain. A second loss function corresponding to the target domain model is constructed based on the maximum mean difference. Based on the second loss function, the hyperparameters of the attenuation prediction model are updated through backpropagation and the attenuation prediction model is adapted to the target domain to obtain the capacity attenuation prediction result.

[0011] In an optional embodiment,

[0012] Acquire historical monitoring data of the source domain and the target domain corresponding to the target battery, construct a time series data set, construct a first prediction model corresponding to the source domain and a second prediction model corresponding to the target domain based on the time series data set, determine the loss value of the prediction model in combination with the first loss function, and optimize the parameters of the prediction model using a pre-selected optimizer to obtain an initial prediction model including:

[0013] Acquire source domain and target domain historical monitoring data corresponding to the target battery, and construct a source domain time series dataset and a target domain time series dataset, wherein the historical monitoring data includes charge and discharge current, voltage, and temperature;

[0014] The first prediction model is constructed based on the source domain time series dataset, and the second prediction model is constructed based on the target domain time series dataset, wherein each prediction model includes an input layer, a prediction layer, a fully connected layer, and an output layer, wherein the input layer receives the source domain time series dataset and extracts long-term and short-term dependencies in the time series data through the prediction layer, and performs feature mapping on the long-term and short-term dependencies through the fully connected layer;

[0015] Define a first loss function corresponding to the first prediction model and the second prediction model, calculate the difference between the model prediction value and the true value based on the first loss function, obtain the loss value corresponding to the first prediction model and the second prediction model, iteratively optimize the prediction model parameters in combination with the loss value through an optimization algorithm, and obtain the initial prediction model on the source domain and the target domain.

[0016] In an optional embodiment,

[0017] The first loss function is shown in the following formula:

[0018]

[0019] Among them, L() represents the loss function value, N represents the number of samples, and y i The i-th element represents the true value, m i represents the normalization constant, Represents the i-th element of the predicted value.

[0020] In an optional embodiment,

[0021] The initial prediction model is trained on source domain data to obtain pre-trained parameters, a parameter migration graph network is constructed with the pre-trained parameters as nodes, and nodes are updated, a migration strategy function is obtained by determining dependencies between parameters based on the updated nodes, and the pre-trained parameters are migrated to the initial prediction model based on the migration strategy function to obtain the attenuation prediction model, which includes:

[0022] Based on a pre-trained initial prediction model, source domain data is obtained and the initial prediction model is trained on the source domain data to obtain pre-trained parameters. A parameter transition graph network is constructed using the pre-trained parameters as nodes. Each node in the parameter transition graph network is updated using a preset graph convolutional neural network, aggregating neighbor information of each node and modeling dependencies between nodes to obtain an updated node representation.

[0023] Based on the updated node representation, the dependency weights between nodes are calculated and the pre-trained parameters are mapped to the weighted sum of the dependency nodes corresponding to each node to obtain a migration strategy function;

[0024] Based on the migration strategy function, the pre-trained parameters are migrated to the initial prediction model through a layer-by-layer attenuation migration method to obtain an initial attenuation prediction model. Based on the target domain data, the hyperparameters in the initial attenuation prediction model are adjusted by minimizing the empirical risk to obtain the attenuation prediction model.

[0025] In an optional embodiment,

[0026] The migration strategy function is shown in the following formula:

[0027]

[0028] in, represents the parameters of the original task j, φ() represents the parameters after migration, m represents the number of target tasks, is the feature representation of the original task j, is the feature representation of the target task k, is the parameter of the target task k, and l represents the index of each task in the target task set.

[0029] In an optional embodiment,

[0030] For the attenuation prediction model, a maximum mean difference is obtained by calculating the difference in battery data feature distribution in the source domain and the target domain, a second loss function corresponding to the target domain model is constructed based on the maximum mean difference, and based on the second loss function, hyperparameters of the attenuation prediction model are updated through back propagation and the attenuation prediction model is adapted to the target domain, to obtain a capacity attenuation prediction result including:

[0031] Adding source domain data and target domain data to the attenuation prediction model, extracting the hidden state of the current time step as the feature representation of the current time step at each time step, aggregating the feature representation of each time step to obtain a sequence feature vector, and generating a source domain feature set and a target domain feature set based on the sequence feature vector;

[0032] For the attenuation prediction model, mapping the source domain feature set and the target domain feature set into a reproducing kernel Hilbert space by a mapping function and calculating empirical means of the source domain and the target domain in the reproducing kernel Hilbert space, and calculating the maximum mean difference between the battery data features in the source domain data and the target domain data based on the empirical means;

[0033] constructing a second loss function based on the maximum mean difference, calculating the prediction loss and the maximum mean difference loss in each training session based on the second loss function and determining a model loss value, updating hyperparameters in the attenuation prediction model using a gradient descent algorithm in combination with the model loss value to minimize the model loss value, and repeating the update until the model converges;

[0034] The converged model is used to predict the battery capacity attenuation and obtain the capacity attenuation prediction result.

[0035] In an optional embodiment,

[0036] The maximum mean difference between the battery data features in the source domain data and the target domain data is calculated as follows:

[0037]

[0038] Among them, MMD() represents the maximum mean difference, D s represents the source domain feature set, D t represents the target domain feature set, n s represents the number of samples in the source domain, represents the source domain sample weight, f() represents the mapping function, represents the rth sample in the source domain, n t represents the number of samples in the target domain, represents the target domain sample weight, represents the rth sample in the source domain, ‖‖ H represents the norm in Hilbert space.

[0039] A second aspect of an embodiment of the present invention provides an online remote state monitoring system for batteries based on autonomous learning, comprising:

[0040] The first unit is configured to obtain historical monitoring data of a source domain and a target domain corresponding to a target battery, construct a time series data set, respectively construct a first prediction model corresponding to the source domain and a second prediction model corresponding to the target domain based on the time series data set, determine a loss value of the prediction model in combination with a first loss function, and optimize the parameters of the prediction model using a pre-selected optimizer to obtain an initial prediction model;

[0041] The second unit is configured to train the initial prediction model on source domain data to obtain pre-trained parameters, construct a parameter migration graph network with the pre-trained parameters as nodes and update the nodes, obtain a migration strategy function based on the updated nodes by determining dependencies between parameters, and migrate the pre-trained parameters to the initial prediction model based on the migration strategy function to obtain a decay prediction model;

[0042] The third unit is used to obtain the maximum mean difference of the attenuation prediction model by calculating the difference in battery data feature distribution in the source domain and the target domain, construct a second loss function corresponding to the target domain model based on the maximum mean difference, and update the hyperparameters of the attenuation prediction model through back propagation based on the second loss function, so as to adapt the attenuation prediction model to the target domain to obtain a capacity attenuation prediction result.

[0043] According to a third aspect of the embodiments of the present invention,

[0044] An electronic device is provided, comprising:

[0045] processor;

[0046] a memory for storing processor-executable instructions;

[0047] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0048] According to a fourth aspect of the embodiments of the present invention,

[0049] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0050] In the present invention, by acquiring the historical monitoring data of the source domain and the target domain corresponding to the target battery and constructing a time series data set, it is helpful to make full use of the historical operation information of the battery, capture the dynamic changes and evolution laws of capacity attenuation, and provide rich data support for subsequent model training and optimization. By constructing the first prediction model and the second prediction model on the source domain and the target domain data respectively, and optimizing the parameters in combination with the first loss function, an initial prediction model is obtained, which is helpful to learn the general laws and specific patterns of battery capacity attenuation on the source domain and the target domain data respectively, laying the foundation for subsequent knowledge transfer and domain adaptation. By transferring the pre-trained Parameters are migrated to the initial prediction model to obtain an attenuation prediction model, which helps to migrate the general rules and feature representations learned in the source domain data to the target domain model, improve the model's adaptability and prediction performance on the target domain data, and realize cross-domain knowledge migration and generalization. By calculating the distribution difference of battery data features in the source domain and the target domain, the maximum mean difference is obtained, and a second loss function is constructed based on the maximum mean difference, which helps to measure the difference between the source domain and target domain data distributions, quantify the degree of distribution offset between domains, and provide guidance and constraints for domain adaptation. In summary, the present invention provides an efficient and reliable technical means for battery health status monitoring and life prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 Schematic diagram of the flow of a method for online remote status monitoring of a battery based on autonomous learning according to an embodiment of the present invention;

[0052] Figure 2 The figure is a structural diagram of a battery online remote status monitoring system based on autonomous learning according to an embodiment of the present invention. DETAILED DESCRIPTION

[0053] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0054] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0055] Figure 1 FIG. 1 is a flow chart of a method for online remote status monitoring of a battery based on autonomous learning according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0056] S1. Obtain historical monitoring data from the source and target domains corresponding to the target battery, construct a time series dataset, and construct a first prediction model corresponding to the source domain and a second prediction model corresponding to the target domain based on the time series dataset. Determine the loss value of the prediction model using a first loss function and optimize the prediction model parameters using a pre-selected optimizer to obtain an initial prediction model.

[0057] The time series dataset refers to a series of data points arranged in chronological order, usually including a timestamp and an associated observation value, which can be continuous or discrete. The source domain and the target domain are usually nouns used in domain adaptation. The source domain is a dataset that already has labels, and the target domain is a new dataset to which the model is to be generalized. The optimizer is an algorithm for training a neural network model, which is responsible for adjusting the model parameters according to the loss function to minimize the loss.

[0058] In an optional embodiment,

[0059] Acquire historical monitoring data of the source domain and the target domain corresponding to the target battery, construct a time series data set, construct a first prediction model corresponding to the source domain and a second prediction model corresponding to the target domain based on the time series data set, determine the loss value of the prediction model in combination with the first loss function, and optimize the parameters of the prediction model using a pre-selected optimizer to obtain an initial prediction model including:

[0060] Acquire source domain and target domain historical monitoring data corresponding to the target battery, and construct a source domain time series dataset and a target domain time series dataset, wherein the historical monitoring data includes charge and discharge current, voltage, and temperature;

[0061] The first prediction model is constructed based on the source domain time series dataset, and the second prediction model is constructed based on the target domain time series dataset, wherein each prediction model includes an input layer, a prediction layer, a fully connected layer, and an output layer, wherein the input layer receives the source domain time series dataset and extracts long-term and short-term dependencies in the time series data through the prediction layer, and performs feature mapping on the long-term and short-term dependencies through the fully connected layer;

[0062] Define a first loss function corresponding to the first prediction model and the second prediction model, calculate the difference between the model prediction value and the true value based on the first loss function, obtain the loss value corresponding to the first prediction model and the second prediction model, iteratively optimize the prediction model parameters in combination with the loss value through an optimization algorithm, and obtain the initial prediction model on the source domain and the target domain.

[0063] The sensor collects monitoring data such as the charge and discharge current, voltage, and temperature of the target battery. The collected monitoring data is sorted by timestamp to construct the source domain time series dataset and the target domain time series dataset.

[0064] A first prediction model is constructed based on a source domain time series dataset, the source domain time series dataset is input into the input layer of the first prediction model, long-term and short-term dependencies in the time series data are extracted through the prediction layer of the first prediction model, the extracted long-term and short-term dependencies are subjected to feature mapping through a fully connected layer, the mapped features are output to the output layer, and the output result of the first prediction model is obtained; a second prediction model is constructed based on a target domain time series dataset, the target domain time series dataset is input into the input layer of the second prediction model, long-term and short-term dependencies in the time series data are extracted through the prediction layer of the second prediction model, the extracted long-term and short-term dependencies are subjected to feature mapping through a fully connected layer, the mapped features are output to the output layer, and the output result of the second prediction model is obtained;

[0065] Based on the output of the prediction model and the true value, a first loss function is constructed. The first loss function is used to measure the degree of difference between the predicted value and the true value. The output of the first prediction model and the true value of the source domain are substituted into the first loss function to calculate the loss value of the first prediction model. The output of the second prediction model and the true value of the target domain are substituted into the first loss function to calculate the loss value of the second prediction model.

[0066] Select a suitable optimization algorithm, such as a gradient descent algorithm, backpropagate the loss value of the first prediction model, and use the optimization algorithm to update the parameters of the first prediction model. Backpropagate the loss value of the second prediction model, and use the optimization algorithm to update the parameters of the second prediction model. Repeat the parameter update until the preset maximum number of iterations is reached or the model converges. After multiple iterative optimizations, the first prediction model achieves better performance in the source domain, and the initial prediction model in the source domain is obtained. The second prediction model achieves better performance in the target domain, and the initial prediction model in the target domain is obtained.

[0067] In this embodiment, by obtaining the historical monitoring data of the source domain and target domain corresponding to the target battery, constructing the source domain time series dataset and the target domain time series dataset helps to comprehensively characterize the state characteristics and working status of the battery. Constructing a prediction model helps to make full use of historical monitoring data and learn the inherent laws and characteristic representations of the battery state evolution. By constructing a first loss function, it helps to improve the accuracy and generalization ability of the prediction model, and obtain an initial prediction model with excellent performance in the source domain and the target domain. In summary, this embodiment can provide technical support for battery management and prediction, and help improve the use efficiency and life of the battery.

[0068] In an optional embodiment,

[0069] The first loss function is shown in the following formula:

[0070]

[0071] Among them, L() represents the loss function value, N represents the number of samples, and y i The i-th element represents the true value, m i represents the normalization constant, Represents the i-th element of the predicted value.

[0072] In this function, by introducing the logarithmic function, the ratio relationship between the predicted value and the true value is converted into a difference relationship, which helps to amplify the difference between the predicted value and the true value, so that the model pays more attention to the relative difference between the predicted value and the true value, and improves the sensitivity and discrimination ability of the model. By calculating the relative entropy of the predicted value and the true value respectively, and adding the two parts to form a symmetrical loss function form, it helps to balance the importance of the predicted value and the true value, avoid the model from being too biased towards one aspect, and improve the stability and robustness of the model. By introducing the normalization constant, it helps to eliminate the influence of different feature dimensions, making the optimization process of the model smoother and more efficient. In summary, this embodiment helps to improve the sensitivity, stability, training efficiency and generalization ability of the prediction model, and lays a good foundation for subsequent model optimization and status monitoring.

[0073] S2. Training the initial prediction model on source domain data to obtain pre-trained parameters, constructing a parameter migration graph network using the pre-trained parameters as nodes and updating the nodes, determining dependencies between parameters based on the updated nodes to obtain a migration strategy function, and migrating the pre-trained parameters to the initial prediction model based on the migration strategy function to obtain a decay prediction model;

[0074] The pre-trained parameters refer to the parameters of a model trained on a task. The parameter migration network is a neural network structure used to migrate parameters between a source domain and a target domain. The migration strategy function is a function used to calculate the weight or offset of each parameter during the parameter migration process. The migration method of the model parameters can be dynamically adjusted according to the similarity between the source domain and the target domain. The attenuation prediction model is a model used to predict attenuation behavior and is commonly used in the field of battery health management.

[0075] In an optional embodiment,

[0076] The initial prediction model is trained on source domain data to obtain pre-trained parameters, a parameter migration graph network is constructed with the pre-trained parameters as nodes, and nodes are updated, a migration strategy function is obtained by determining dependencies between parameters based on the updated nodes, and the pre-trained parameters are migrated to the initial prediction model based on the migration strategy function to obtain the attenuation prediction model, which includes:

[0077] Based on a pre-trained initial prediction model, source domain data is obtained and the initial prediction model is trained on the source domain data to obtain pre-trained parameters. A parameter transition graph network is constructed using the pre-trained parameters as nodes. Each node in the parameter transition graph network is updated using a preset graph convolutional neural network, aggregating neighbor information of each node and modeling dependencies between nodes to obtain an updated node representation.

[0078] Based on the updated node representation, the dependency weights between nodes are calculated and the pre-trained parameters are mapped to the weighted sum of the dependency nodes corresponding to each node to obtain a migration strategy function;

[0079] Based on the migration strategy function, the pre-trained parameters are migrated to the initial prediction model through a layer-by-layer attenuation migration method to obtain an initial attenuation prediction model. Based on the target domain data, the hyperparameters in the initial attenuation prediction model are adjusted by minimizing the empirical risk to obtain the attenuation prediction model.

[0080] Obtain a pre-trained initial prediction model, obtain source domain data, and train the initial prediction model on the source domain data to obtain pre-trained parameters. By training the initial prediction model on the source domain data, its parameters are adapted to the characteristic distribution of the source domain data. Using the pre-trained parameters as nodes, a parameter migration graph network is constructed. Each node represents a set of pre-trained parameters, and the connections between nodes represent the dependencies between the parameters. Each node in the parameter migration graph network is updated using a preset graph convolutional neural network. The graph convolutional neural network aggregates the neighbor information of each node and models the dependencies between nodes. The representation of each node is updated through a graph convolution operation to obtain an updated node representation.

[0081] Based on the updated node representation, the dependency weights between nodes are calculated, and the pre-trained parameters are mapped to the weighted sum of the dependent nodes corresponding to each node. Through the weighted sum operation, a migration strategy function is obtained, which maps the pre-trained parameters to the initialization parameters of the target domain model.

[0082] Based on the migration strategy function, the pre-trained parameters are migrated layer by layer to the initial prediction model. During the migration process, an attenuation factor is introduced. As the number of network layers increases, the attenuation factor gradually decreases. The initial attenuation prediction model is obtained through the layer-by-layer attenuation migration method. The parameters of this model are a weighted combination of the pre-trained parameters and the randomly initialized parameters of the target domain. The target domain data, i.e., the status monitoring data of the target battery, is obtained and input into the initial attenuation prediction model. By minimizing the empirical risk, the hyperparameters in the initial attenuation prediction model, such as the learning rate and regularization coefficient, are adjusted. Through hyperparameter tuning, a better performance attenuation prediction model is obtained.

[0083] In this embodiment, by training the initial prediction model on the source domain data, pre-training parameters are obtained, which helps to capture the general knowledge and feature representations in the source domain data and provide a good initialization foundation for subsequent knowledge transfer. By constructing a parameter migration graph network with pre-training parameters as nodes and using a graph convolutional neural network to update the nodes, aggregating neighbor information and modeling the dependencies between nodes, it helps to explore the potential associations and dependency structures between pre-training parameters, extract richer and more abstract feature representations, and enhance the adaptability and generalization ability of knowledge transfer. The pre-training parameters are migrated to the initial prediction model through a layer-by-layer attenuation migration method to obtain an initial attenuation prediction model, which helps to gradually adapt to the characteristics of the target domain data while retaining the source domain knowledge and realize smooth and progressive knowledge transfer. In summary, this embodiment improves the performance of the model in the target domain, enhances the adaptability and generalization ability of the model, and lays the foundation for efficient and accurate battery status monitoring.

[0084] In an optional embodiment,

[0085] The migration strategy function is shown in the following formula:

[0086]

[0087] in, represents the parameters of the original task j, φ() represents the parameters after migration, m represents the number of target tasks, is the feature representation of the original task j, is the feature representation of the target task k, is the parameter of the target task k, and l represents the index of each task in the target task set.

[0088] In this embodiment, the attention weight between the original task and the target task is obtained by calculating the inner product between the feature representation of the original task and the feature representation of the target task, and applying an exponential function for normalization, which helps to quantify the correlation and importance between different tasks and realize adaptive parameter migration. By introducing the exponential function in the inner product calculation of the feature representation, the similarity difference between tasks is amplified, which helps to highlight the distinction between tasks, so that the attention weight is more concentrated on the target task with a higher correlation with the original task, thereby improving the pertinence and effectiveness of knowledge transfer. By normalizing the exponential inner product of all target tasks in the denominator, the probability distribution characteristics of the attention weight are ensured, which helps to balance the contribution of different target tasks, avoid the parameter migration of individual tasks being too dominant, and improve the stability and robustness of knowledge migration. In summary, this embodiment improves the generalization ability and adaptability of the model while realizing efficient and accurate knowledge transfer.

[0089] S3. For the attenuation prediction model, the maximum mean difference is obtained by calculating the difference in the distribution of battery data features in the source domain and the target domain, and a second loss function corresponding to the target domain model is constructed based on the maximum mean difference. Based on the second loss function, the hyperparameters of the attenuation prediction model are updated through backpropagation and the attenuation prediction model is adapted to the target domain to obtain a capacity attenuation prediction result.

[0090] The data feature distribution difference refers to the different distribution of features between different data sets. The maximum mean difference is a method for measuring the distribution difference between two data sets. The back propagation is an optimization algorithm for training neural networks. It calculates the gradient of the loss function with respect to the model parameters and updates the parameters along the direction of the gradient to minimize the loss function.

[0091] In an optional embodiment,

[0092] For the attenuation prediction model, a maximum mean difference is obtained by calculating the difference in battery data feature distribution in the source domain and the target domain, a second loss function corresponding to the target domain model is constructed based on the maximum mean difference, and based on the second loss function, hyperparameters of the attenuation prediction model are updated through back propagation and the attenuation prediction model is adapted to the target domain, to obtain a capacity attenuation prediction result including:

[0093] Adding source domain data and target domain data to the attenuation prediction model, extracting the hidden state of the current time step as the feature representation of the current time step at each time step, aggregating the feature representation of each time step to obtain a sequence feature vector, and generating a source domain feature set and a target domain feature set based on the sequence feature vector;

[0094] For the attenuation prediction model, mapping the source domain feature set and the target domain feature set into a reproducing kernel Hilbert space by a mapping function and calculating empirical means of the source domain and the target domain in the reproducing kernel Hilbert space, and calculating the maximum mean difference between the battery data features in the source domain data and the target domain data based on the empirical means;

[0095] constructing a second loss function based on the maximum mean difference, calculating the prediction loss and the maximum mean difference loss in each training session based on the second loss function and determining a model loss value, updating hyperparameters in the attenuation prediction model using a gradient descent algorithm in combination with the model loss value to minimize the model loss value, and repeating the update until the model converges;

[0096] The converged model is used to predict the battery capacity attenuation and obtain the capacity attenuation prediction result.

[0097] The source domain data and the target domain data are organized and aligned according to time steps, and are input into the decay prediction model. The model begins to process the data and extract features. At each time step, the decay prediction model extracts the hidden state of the current time step as the feature representation of the current time step. The feature representation extracted at each time step is aggregated, for example, by averaging or maximum pooling, to obtain the feature vector of the entire sequence. The aggregated sequence feature vectors are stored as the source domain feature set and the target domain feature set respectively.

[0098] Define a mapping function, such as a Gaussian kernel function, to map the source domain feature set and the target domain feature set into the reproducing kernel Hilbert space. In the reproducing kernel Hilbert space, calculate the empirical mean of the source domain feature set and the target domain feature set to obtain the representation of the source domain and the target domain in the space. Calculate the difference between the empirical means of the source domain and the target domain in the reproducing kernel Hilbert space, use the maximum mean difference to measure the distribution difference between the source domain and the target domain in the feature space, and obtain the maximum mean difference between the battery data features in the source domain data and the target domain data.

[0099] A second loss function is constructed based on the maximum mean difference to measure the difference in data distribution between the source and target domains. In each training session, the prediction loss of the attenuation prediction model on the source and target domain data is calculated. At the same time, the maximum mean difference loss, i.e., the value of the second loss function, is calculated. The prediction loss and the maximum mean difference loss are added together to obtain the total loss value of the model, i.e., the model loss value.

[0100] The gradient descent algorithm is used to calculate the gradient according to the model loss value. The calculated gradient is used to update the hyperparameters in the attenuation prediction model, such as weights and biases. Through continuous iterative updates, the model loss value gradually decreases and the model performance continues to improve. The update process is repeated until the model converges, that is, the loss value stabilizes at a low level.

[0101] The converged attenuation prediction model is applied to the target domain data to predict the battery capacity attenuation. The model outputs the predicted capacity attenuation value based on the input battery data characteristics, and obtains the final capacity attenuation prediction result, which is used to evaluate the health status and remaining life of the battery.

[0102] In this embodiment, by adding source domain and target domain data to the attenuation prediction model, extracting feature representations at each time step, and aggregating to obtain sequence feature vectors, it helps to make full use of the temporal information of the source domain and target domain data, capture the dynamic changes and evolution laws of the data, and provide rich feature representations for subsequent domain adaptation and model optimization. By mapping the source domain feature set and the target domain feature set into the reproducing kernel Hilbert space through a mapping function, and calculating the empirical mean, it helps to transform the original feature space into a higher-dimensional and more expressive feature space, thereby enhancing the distinguishability and generalization ability of the features. By constructing a second loss function based on the maximum mean difference, and calculating the prediction loss and the maximum mean difference loss in each training to obtain the model loss value, it helps to incorporate the distribution difference of the source domain and target domain data into the model optimization target, prompting the model to minimize the distribution difference between domains while minimizing the prediction error, thereby improving the domain adaptability and generalization performance of the model. In summary, this embodiment effectively improves the prediction performance and domain adaptability of the model, provides an efficient and accurate technical means for battery capacity attenuation prediction, and has important practical application value.

[0103] In an optional embodiment,

[0104] The maximum mean difference between the battery data features in the source domain data and the target domain data is calculated as follows:

[0105]

[0106] Among them, MMD() represents the maximum mean difference, D s represents the source domain feature set, D t represents the target domain feature set, n s represents the number of samples in the source domain, represents the source domain sample weight, f() represents the mapping function, represents the rth sample in the source domain, n t represents the number of samples in the target domain, represents the target domain sample weight, represents the rth sample in the source domain, ‖‖ H represents the norm in Hilbert space.

[0107] In this embodiment, the source domain samples and the target domain samples are mapped into the reproducing kernel Hilbert space through a mapping function, which helps to convert the original data features into high-dimensional, nonlinear feature representations, enhance the expressiveness and distinguishability of the features, and provide a suitable feature space for subsequent calculations. By calculating the Hilbert space norm difference between the source domain empirical mean and the target domain empirical mean, the maximum mean difference value is obtained, which helps to measure the distribution difference between the source domain data and the target domain data in the feature space, quantify the degree of distribution offset between domains, and provide guidance and constraints for subsequent domain adaptation and model optimization. In summary, this embodiment encourages the model to learn domain-invariant feature representations and prediction rules, improves the generalization ability and prediction performance of the model, and realizes effective cross-domain knowledge transfer and capacity decay prediction.

[0108] Figure 2 FIG. 1 is a structural diagram of a battery online remote status monitoring system based on autonomous learning according to an embodiment of the present invention. Figure 2 As shown, the system includes:

[0109] The first unit is configured to obtain historical monitoring data of a source domain and a target domain corresponding to a target battery, construct a time series data set, respectively construct a first prediction model corresponding to the source domain and a second prediction model corresponding to the target domain based on the time series data set, determine a loss value of the prediction model in combination with a first loss function, and optimize the parameters of the prediction model using a pre-selected optimizer to obtain an initial prediction model;

[0110] The second unit is configured to train the initial prediction model on source domain data to obtain pre-trained parameters, construct a parameter migration graph network with the pre-trained parameters as nodes and update the nodes, obtain a migration strategy function based on the updated nodes by determining dependencies between parameters, and migrate the pre-trained parameters to the initial prediction model based on the migration strategy function to obtain a decay prediction model;

[0111] The third unit is used to obtain the maximum mean difference of the attenuation prediction model by calculating the difference in battery data feature distribution in the source domain and the target domain, construct a second loss function corresponding to the target domain model based on the maximum mean difference, and update the hyperparameters of the attenuation prediction model through back propagation based on the second loss function, so as to adapt the attenuation prediction model to the target domain to obtain a capacity attenuation prediction result.

[0112] According to a third aspect of the embodiments of the present invention,

[0113] An electronic device is provided, comprising:

[0114] processor;

[0115] a memory for storing processor-executable instructions;

[0116] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0117] According to a fourth aspect of the embodiments of the present invention,

[0118] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0119] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0120] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A battery online remote status monitoring method based on autonomous learning, characterized in that: include: Acquire historical monitoring data of the source domain and the target domain corresponding to the target battery, construct a time series dataset, construct a first prediction model corresponding to the source domain and a second prediction model corresponding to the target domain based on the time series dataset, determine the loss value of the prediction model in combination with a first loss function, and optimize the parameters of the prediction model using a pre-selected optimizer to obtain an initial prediction model in the source domain and an initial prediction model in the target domain; The initial prediction model in the source domain is trained on source domain data to obtain pre-trained parameters, a parameter migration graph network is constructed with the pre-trained parameters as nodes, and the nodes are updated. Based on the updated nodes, a migration strategy function is obtained by determining dependencies between parameters. The pre-trained parameters are migrated to the initial prediction model in the target domain based on the migration strategy function to obtain a decay prediction model. For the attenuation prediction model, the maximum mean difference is obtained by calculating the difference in the battery data feature distribution in the source domain and the target domain. A second loss function corresponding to the target domain model is constructed based on the maximum mean difference. Based on the second loss function, the hyperparameters of the attenuation prediction model are updated through backpropagation and the attenuation prediction model is adapted to the target domain to obtain the capacity attenuation prediction result.

2. The method according to claim 1, characterized in that Acquire historical monitoring data of the source domain and the target domain corresponding to the target battery, construct a time series data set, construct a first prediction model corresponding to the source domain and a second prediction model corresponding to the target domain based on the time series data set, determine the loss value of the prediction model in combination with a first loss function, and optimize the parameters of the prediction model using a pre-selected optimizer, respectively obtaining an initial prediction model on the source domain and an initial prediction model on the target domain, including: Acquire source domain and target domain historical monitoring data corresponding to the target battery, and construct a source domain time series dataset and a target domain time series dataset, wherein the historical monitoring data includes charge and discharge current, voltage, and temperature; The first prediction model is constructed based on the source domain time series dataset, and the second prediction model is constructed based on the target domain time series dataset, wherein each prediction model includes an input layer, a prediction layer, a fully connected layer, and an output layer, wherein the input layer receives the source domain time series dataset and extracts long-term and short-term dependencies in the time series data through the prediction layer, and performs feature mapping on the long-term and short-term dependencies through the fully connected layer; Define a first loss function corresponding to the first prediction model and the second prediction model, measure the difference between the model prediction distribution and the true distribution based on the first loss function, obtain the loss values ​​corresponding to the first prediction model and the second prediction model, iteratively optimize the prediction model parameters in combination with the loss values ​​through an optimization algorithm, and obtain the initial prediction model on the source domain and the target domain.

3. The method according to claim 1, characterized in that The initial prediction model in the source domain is trained on source domain data to obtain pre-trained parameters, a parameter migration graph network is constructed with the pre-trained parameters as nodes and nodes are updated, a migration strategy function is obtained by determining dependencies between parameters based on the updated nodes, and the pre-trained parameters are migrated to the initial prediction model in the target domain based on the migration strategy function to obtain the attenuation prediction model, which includes: Based on a pre-trained initial prediction model on the source domain, source domain data is obtained and the initial prediction model on the source domain is trained on the source domain data to obtain pre-trained parameters. A parameter transition graph network is constructed using the pre-trained parameters as nodes. Each node in the parameter transition graph network is updated using a preset graph convolutional neural network, neighbor information of each node is aggregated, and dependencies between nodes are modeled to obtain an updated node representation. Based on the updated node representation, the dependency weights between nodes are calculated and the pre-trained parameters are mapped to the weighted sum of the dependency nodes corresponding to each node to obtain a migration strategy function; Based on the migration strategy function, the pre-trained parameters are migrated to the initial prediction model on the target domain through a layer-by-layer attenuation migration method to obtain an initial attenuation prediction model. Based on the target domain data, the hyperparameters in the initial attenuation prediction model are adjusted by minimizing the empirical risk to obtain the attenuation prediction model.

4. The method according to claim 1, wherein The migration strategy function is shown in the following formula: ； in, Represents the original task j Parameters, φ() Represents the parameters after migration, m Indicates the number of target tasks, For the original task j The feature representation of For target tasks k The feature representation of For target tasks k Parameters, l Represents the index of each task in the target task collection.

5. The method according to claim 1, characterized in that For the attenuation prediction model, a maximum mean difference is obtained by calculating the difference in battery data feature distribution in the source domain and the target domain, a second loss function corresponding to the target domain model is constructed based on the maximum mean difference, and based on the second loss function, hyperparameters of the attenuation prediction model are updated through back propagation and the attenuation prediction model is adapted to the target domain, to obtain a capacity attenuation prediction result including: Adding source domain data and target domain data to the attenuation prediction model, extracting the hidden state of the current time step as the feature representation of the current time step at each time step, aggregating the feature representation of each time step to obtain a sequence feature vector, and generating a source domain feature set and a target domain feature set based on the sequence feature vector; For the attenuation prediction model, mapping the source domain feature set and the target domain feature set into a reproducing kernel Hilbert space by a mapping function and calculating empirical means of the source domain and the target domain in the reproducing kernel Hilbert space, and calculating the maximum mean difference between the battery data features in the source domain data and the target domain data based on the empirical means; constructing a second loss function based on the maximum mean difference, calculating the prediction loss and the maximum mean difference loss in each training session based on the second loss function and determining a model loss value, updating hyperparameters in the attenuation prediction model using a gradient descent algorithm in combination with the model loss value to minimize the model loss value, and repeating the update until the model converges; The converged model is used to predict the battery capacity attenuation and obtain the capacity attenuation prediction result.

6. The method according to claim 5, characterized in that The maximum mean difference between the battery data features in the source domain data and the target domain data is calculated as follows: ； in, MMD() represents the maximum mean difference, D s represents the source domain feature set, D t represents the target domain feature set, n s represents the number of samples in the source domain, represents the source domain sample weight, f() represents the mapping function, Indicates the first r samples, n t represents the number of samples in the target domain, represents the target domain sample weight, Indicates the first r samples, ‖‖ H represents the norm in Hilbert space.

7. A battery online remote status monitoring system based on autonomous learning, used to implement the method according to any one of claims 1 to 6, characterized in that: include: The first unit is configured to obtain historical monitoring data of a source domain and a target domain corresponding to a target battery, construct a time series data set, and respectively construct a first prediction model corresponding to the source domain and a second prediction model corresponding to the target domain based on the time series data set, determine a loss value of the prediction model in combination with a first loss function, and optimize the parameters of the prediction model using a pre-selected optimizer, thereby obtaining an initial prediction model in the source domain and an initial prediction model in the target domain. The second unit is configured to train the initial prediction model in the source domain on the source domain data to obtain pre-trained parameters, construct a parameter migration graph network with the pre-trained parameters as nodes and update the nodes, obtain a migration strategy function based on the updated nodes by determining the dependency relationship between the parameters, and migrate the pre-trained parameters to the initial prediction model in the target domain based on the migration strategy function to obtain a decay prediction model; The third unit is used to obtain the maximum mean difference of the attenuation prediction model by calculating the difference in battery data feature distribution in the source domain and the target domain, construct a second loss function corresponding to the target domain model based on the maximum mean difference, and update the hyperparameters of the attenuation prediction model through back propagation based on the second loss function, so as to adapt the attenuation prediction model to the target domain to obtain a capacity attenuation prediction result.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images