Health status assessment model training method

Abstract

A health state evaluation model training method includes the following steps: obtaining degradation data of a plurality of batteries, processing and dividing the degradation data into source client data sets and target client data sets; using data units in the source client data sets to train a local model to obtain a plurality of source client health evaluation models, modeling first feature information of the data units in the source client data sets by using a Gaussian mixture model to obtain a first Gaussian mixture model; extracting second feature information of the target client data by using the source client health evaluation models, and modeling the extracted second feature information by using the Gaussian mixture model to obtain a second Gaussian mixture model. The health state evaluation model training method protects client privacy and ensures the accuracy of model prediction.

Description

Technical Field

[0001] This invention relates to the field of model training technology, and in particular to a method for training a health status assessment model. Background Technology

[0002] Lithium-ion batteries, due to their high energy density, long lifespan, and low self-discharge rate, have become an essential component of portable electrical devices and electric vehicle energy storage systems. However, the performance of lithium-ion batteries inevitably deteriorates over time, leading to problems such as capacity and power reduction. In fact, as batteries age, their likelihood of failure increases, and they may even cause serious accidents, such as fires. Therefore, accurately monitoring and assessing the health status of lithium-ion batteries is of great significance for battery health management.

[0003] Existing methods for assessing the health status of lithium-ion batteries typically rely on centralized training models and large amounts of high-quality data. However, in real-world industrial applications, individual users possess very limited battery degradation data, making it difficult to independently develop high-performance health status assessment models. Furthermore, to protect their privacy and interests, users with similar batteries are often unwilling to share their private data, leading to data silos. Therefore, considering data privacy issues, adopting a centralized training model to obtain an ideal health status assessment model is impractical. Moreover, the operating conditions of different batteries often vary significantly, such as temperature and constant current rate during discharge, resulting in significant distributional differences in data from different users. This negatively impacts the accuracy of health status assessment. Summary of the Invention

[0004] In view of this, the purpose of this invention is to provide a method for training a health status assessment model that protects customer privacy and ensures the accuracy of model predictions.

[0005] This invention provides a method for training a health status assessment model, comprising the following steps:

[0006] Acquire degradation data from several batteries, process the degradation data, and divide it into source client dataset and target client dataset;

[0007] The local model is trained using data units in the source client dataset to obtain several source client health assessment models. The first feature information of the data units in the source client dataset is modeled using a Gaussian mixture model to obtain the first Gaussian mixture model.

[0008] The second feature information of the target client data is extracted using the source client health assessment model, and the extracted second feature information is modeled using a Gaussian mixture model to obtain a second Gaussian mixture model.

[0009] Unit test the corresponding source client health assessment model using data from the target client dataset to obtain the test error of the source client health assessment model;

[0010] The Bartlett distance between the source client data and the target client data is calculated using the parameters of the first and second Gaussian mixture models.

[0011] The contribution value of each source client health assessment model is calculated using Bach distance and test error. The aggregation weight corresponding to each source client health assessment model is obtained by calculating the contribution value. Based on the aggregation weight, several health assessment models are aggregated into a dynamic weighted model.

[0012] In one embodiment, the step of acquiring degradation data of several batteries, processing the degradation data, and dividing it into a source client dataset and a target client dataset further includes:

[0013] Acquire degradation data of several batteries, wherein the degradation data includes voltage, current and internal resistance during the degradation process;

[0014] The degraded data is then normalized.

[0015] The normalized degraded data is divided into source client datasets and target client datasets according to the operating conditions.

[0016] In one embodiment, the step of training its own local model using data units from the source client dataset to obtain several source client health assessment models, and then modeling using a Gaussian mixture model and the first feature information of the data units from the source client dataset to obtain a first Gaussian mixture model, further includes:

[0017] Retrieve the local model from the central server;

[0018] The data units in the source client dataset are input into the local model for training to obtain several source client health assessment models.

[0019] The first feature information of data units in the source client dataset is extracted using the source client health assessment model;

[0020] A first Gaussian mixture model is obtained by modeling using a Gaussian mixture model and utilizing the first feature information.

[0021] In one embodiment, the first Gaussian mixture model is represented as:

[0022]

[0023] Where j represents the number of sub-components in the first Gaussian mixture model, and They represent the kth one. The mean and variance of the j-th sub-component. It is the weight of the j-th sub-component.

[0024] In one embodiment, the first Gaussian mixture model is represented as:

[0025]

[0026] Where j represents the number of sub-components in the second Gaussian mixture model, This represents the second Gaussian mixture model fitted using the target client features extracted from the k-th source client health assessment model. and They represent the kth one. The mean and variance of the j-th sub-component. It is the weight of the j-th sub-component.

[0027] In one embodiment, the formula for calculating the test error is:

[0028]

[0029] in, Let y represent the test error of the k-th source client model, N represent the number of samples in the target client dataset, and y represent the test error of the k-th source client model. i express.

[0030] In one embodiment, the formula for calculating the Bartholomew's distance is:

[0031]

[0032] Among them, L k This represents the Bach distance between the k-th source client and the target client. yes The parameters, yes The parameters.

[0033] In one embodiment, the formula for calculating the contribution value of the source client health assessment model is as follows:

[0034]

[0035] Among them, Con k Let α be the contribution of the health assessment model of the k-th source client to the dynamic weighted model, and α be the value used to balance the weighted model. and L k A non-negative parameter.

[0036] In one embodiment, the formula for calculating the aggregate weight corresponding to each source client health assessment model is as follows:

[0037]

[0038] In one embodiment, the dynamic weighting model can be expressed as:

[0039]

[0040] in, This is a source client health assessment model.

[0041] The health status assessment model training method provided by this invention trains the model using its own private data, uses a Gaussian mixture model for feature extraction and modeling, and achieves feature sharing without touching the data. This can effectively solve the data privacy problem that is prone to occur in battery health status assessment tasks. Moreover, the aggregation method can overcome the impact of the distribution difference between the source client and the target client on the health status assessment model, thereby improving the accuracy of model prediction. Attached Figure Description

[0042] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1 This is a flowchart illustrating the training method for the health status assessment model provided by the present invention. Detailed Implementation

[0044] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some, not all, of the embodiments of the present invention. Based on the description of the present invention, all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of the present invention.

[0045] In the description of this invention, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms according to the specific circumstances.

[0046] The terms “upper,” “lower,” “left,” “right,” “front,” “back,” “top,” “bottom,” “inner,” and “outer,” etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product of the invention is in use. They are only for the convenience of description and simplification, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the present invention.

[0047] The terms “first,” “second,” “third,” etc., are used merely to distinguish elements with similar properties, not to indicate or imply relative importance or a specific order.

[0048] The terms “include,” “comprising,” or any other variation thereof are intended to cover non-exclusive inclusion, which includes not only the elements listed but also other elements not expressly listed.

[0049] Please see Figure 1 The health status assessment model training method provided by this invention includes the following steps:

[0050] Acquire degradation data from several batteries, process the degradation data, and divide it into source client dataset and target client dataset.

[0051] It is understandable that each battery has a corresponding client to collect degradation data, and the degradation data for each battery can be under the same operating condition, but the operating conditions for each battery are different. Different operating conditions can include constant current rates of 0.75C, 1C and 2C. The degradation data includes voltage, current and internal resistance during the degradation process.

[0052] Degraded data is normalized to simplify calculations;

[0053] The normalized degraded data is divided into source client dataset and target client dataset according to the operating conditions. The amount of data in the source client dataset and the target client dataset can be basically the same. Each source client dataset includes several data units, and each data unit corresponds to a single battery cell. The units in each data unit can be divided into training set and test set in a ratio of 8:2.

[0054] The local model is trained using data units from the source client dataset to obtain several source client health assessment models. A first Gaussian mixture model is obtained by modeling using the first feature information of the data units from the source client dataset.

[0055] To protect data privacy, a Gaussian mixture model is used to model the features extracted by the feature extractor.

[0056] It is understood that the above steps may further include:

[0057] Retrieve the local model from the central server;

[0058] Data units from the source client dataset are input into the local model for training to obtain several source client health assessment models.

[0059] It is known that the source client health assessment model can include a feature extractor-deep bidirectional gated recurrent unit and a health status evaluator-two fully connected layers.

[0060] The first feature information of data units in the source client dataset is extracted using the source client health assessment model.

[0061] Understandably, the first feature information can be obtained by extracting data units from the source client dataset using the feature extractor of the source client health assessment model.

[0062] A first Gaussian mixture model is obtained by modeling using a Gaussian mixture model and utilizing the first feature information. The first Gaussian mixture model can be represented as:

[0063]

[0064] Where j represents the number of sub-components in the first Gaussian mixture model, and They represent the kth one. The mean and variance of the j-th sub-component. It is the weight of the j-th sub-component.

[0065] Understandably, the parameters of the first Gaussian mixture model are used for subsequent feature information sharing, avoiding access to the source client data. The parameters of the source client health assessment model and the first Gaussian mixture model are sent to the central server for subsequent acquisition and processing.

[0066] The second feature information of the target client data is extracted using the source client health assessment model. The Gaussian mixture model is then used to model the data based on the extracted second feature information to obtain the second Gaussian mixture model.

[0067] It is known that the extraction method for the second feature information is the same as that for the extraction of the first feature information.

[0068] The source client health assessment model is unit tested using data from the target client dataset to obtain the test error of the source client health assessment model.

[0069] Understandably, the target client downloads the aforementioned source client health assessment models from the central server, and uses all data units in the target client dataset to test each source client health assessment model to obtain the test error of each source client health assessment model. The formula for the test error is:

[0070]

[0071] in, Let y represent the test error of the k-th source client model, N represent the number of samples in the target client dataset, and y represent the test error of the k-th source client model. i This represents the true label value of the i-th sample. This represents the evaluation value of the health assessment model for the k-th source client.

[0072] The Bartholomew's distance between the source client data and the target client data is calculated using the parameters of the first and second Gaussian mixture models.

[0073] Understandably, the formula for calculating Bach's distance is as follows:

[0074]

[0075] Among them, L k This represents the Bach distance between the k-th source client and the target client. yes The parameters, yes The parameters, Parameters and The meaning of the parameters is as described above.

[0076] The contribution value of each source client health assessment model is calculated using Bach distance and test error. The aggregation weight corresponding to each source client health assessment model is obtained by calculating the contribution value. Based on the aggregation weight, several health assessment models are aggregated into a dynamic weighted model.

[0077] Understandably, the formula for calculating contribution value can be expressed as:

[0078]

[0079] Among them, Con k Let α be the contribution of the health assessment model of the k-th source client to the dynamic weighted model, and α be the value used to balance the weighted model. and L k A non-negative parameter.

[0080] The formula for calculating the aggregate weight is:

[0081]

[0082] The parameters in the above formula can be found in the above description.

[0083] The dynamic weighted model can be expressed by the following formula:

[0084]

[0085] in, Let k be the source client health assessment model, and k be the number of source client health assessment models.

[0086] After training is complete, each client can download the dynamically weighted model from the central server. ， Furthermore, by inputting the battery degradation data that needs to be evaluated locally in the client into the dynamic weighted model, the battery health status assessment results can be obtained.

[0087] A test system was constructed, consisting of a computer (serving as a central server), a battery test bench, and 18650 lithium-ion batteries. The computer was used to collect battery degradation data. All lithium-ion batteries underwent the same charging mode: charging at a constant current rate of 0.5C under standard constant current / constant voltage mode until the voltage reached 4.2V, and then maintaining that voltage until the charging current dropped to 0.05A. The discharge voltage was set to 2.75V. Constant current rates of 0.75C, 1C, and 2C were considered under different operating conditions. Therefore, battery degradation data under three operating conditions were obtained.

[0088] Using data from two of the three operating conditions as local data for the source client and data from the third operating condition as local data for the target client, the task is to train a dynamically weighted model using the data from the two source clients and then use this dynamically weighted model to evaluate the health status of the target client's battery. The evaluation results of the proposed model and the comparative model (non-dynamically weighted model) are shown in the table below. RMSE and MAE represent the root mean square error and mean absolute error, respectively; the smaller the values ​​of these two indicators, the higher the evaluation accuracy of the model. Using data from two of the three operating conditions as source client data and data from the third operating condition as target client data, three different tasks can be formed, namely Task 1, Task 2, and Task 3. The table shows that the average RMSE (root mean square error) and MAE (mean absolute error) values ​​of the proposed model are 0.0141 and 0.0112, respectively, while the average RMSE and MAE values ​​of the comparative model are 0.0716 and 0.0566, respectively. The comparison shows that the evaluation accuracy of the proposed model is significantly higher than that of the non-dynamically weighted model. Therefore, the proposed model can accurately evaluate the battery's health status while protecting data privacy.

[0089]

[0090]

[0091] As described above, the health status assessment model training method provided by this invention trains the model using its own private data, uses a Gaussian mixture model for feature extraction and modeling, and achieves feature sharing without touching the data. This can effectively solve the data privacy problem that is prone to occur in battery health status assessment tasks. Furthermore, the aggregation method can overcome the impact of the distribution differences between the source client and the target client on the health status assessment model, thereby improving the accuracy of model prediction.

[0092] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. A method for training a health status assessment model, characterized in that, Includes the following steps: Acquire degradation data from several batteries, process the degradation data, and divide it into source client dataset and target client dataset; The local model is trained using data units from the source client dataset to obtain several source client health assessment models. A Gaussian mixture model is then used to model the first feature information of the data units in the source client dataset to obtain a first Gaussian mixture model, which is expressed as: in, This represents the number of sub-components in the first Gaussian mixture model. and They represent the kth one. The The mean and variance of each sub-component It is the first The weight of each sub-component; The second feature information of the target client data is extracted using the source client health assessment model, and a Gaussian mixture model is used to model the extracted second feature information to obtain a second Gaussian mixture model, which is expressed as: in, This represents the number of sub-components in the second Gaussian mixture model. This represents the second Gaussian mixture model fitted using the target client features extracted from the k-th source client health assessment model. and They represent the kth one. The The mean and variance of each sub-component It is the first The weight of each sub-component; The source client health assessment model is unit tested using data from the target client dataset to obtain the test error of the source client health assessment model. The formula for calculating the test error is as follows: in, This represents the test error of the k-th source client model. This indicates the number of samples in the target client's dataset. This represents the true label value of the i-th sample. This represents the evaluation value of the health assessment model for the k-th source client; The Barcol distance between the source client data and the target client data is calculated using the parameters of the first Gaussian mixture model and the second Gaussian mixture model. The formula for calculating the Barcol distance is as follows: in, This represents the Bach distance between the k-th source client and the target client. yes The parameters, yes Parameters; The contribution value of each source client health assessment model is calculated using Bach distance and test error. The aggregation weight corresponding to each source client health assessment model is obtained by calculating the contribution value. Based on the aggregation weight, several health assessment models are aggregated into a dynamic weighted model.

2. The health status assessment model training method as described in claim 1, characterized in that, The step of acquiring degradation data from several batteries, processing the degradation data, and dividing it into a source client dataset and a target client dataset further includes: Acquire degradation data of several batteries, wherein the degradation data includes voltage, current and internal resistance during the degradation process; The degraded data is then normalized. The normalized degraded data is divided into source client datasets and target client datasets according to the operating conditions.

3. The health status assessment model training method as described in claim 2, characterized in that, The step of training a local model using data units from the source client dataset to obtain several source client health assessment models, and modeling the first feature information of the data units in the source client dataset using a Gaussian mixture model to obtain a first Gaussian mixture model, further includes: Retrieve the local model from the central server; The data units in the source client dataset are input into the local model for training to obtain several source client health assessment models. The first feature information of data units in the source client dataset is extracted using the source client health assessment model; A first Gaussian mixture model is obtained by modeling using a Gaussian mixture model and utilizing the first feature information.

4. The health status assessment model training method as described in claim 3, characterized in that, The formula for calculating the contribution value of the source client health assessment model is as follows: in, Let the health assessment model of the k-th source client contribute to the dynamic weighted model. It is used for balance and A non-negative parameter.

5. The health status assessment model training method as described in claim 4, characterized in that, The formula for calculating the aggregate weight corresponding to each source client health assessment model is as follows: 。 6. The health status assessment model training method as described in claim 5, characterized in that, The dynamic weighted model can be expressed as: ， in, This is a source client health assessment model.

Images