Battery health state prediction method and device, storage medium and electronic device

Abstract

The application provides a battery health state prediction method and device, a storage medium and an electronic equipment, wherein the feature extraction network in the battery health state prediction model is used to extract associated feature data from battery characterization information, the associated feature data is feature data for deeply mining the association relationship between various parameters of the battery, and then the associated feature data is used as input data of a recurrent neural network; the recurrent neural network outputs battery health state prediction data after processing the associated feature data. In the process of predicting the battery health state, the feature extraction network is used to extract associated feature data with deep association, and then the associated feature data is used as input data of a time recurrent network, so that the accuracy of predicting the battery health state is improved.

Description

Technical Field

[0001] This invention relates to the field of battery technology, and in particular to a method and apparatus for predicting battery health status, a storage medium, and an electronic device. Background Technology

[0002] With the promotion of new energy sources, batteries are being used as energy storage systems in automobiles, small and medium-sized unmanned aerial vehicles, and various small and medium-sized machines. As battery usage time increases, performance degradation is inevitable. To better assess the lifespan of energy storage systems, it is necessary to predict the battery's State of Health (SOH).

[0003] Currently, the method for predicting battery SOH usually involves extracting physical discrete features from the battery's charge and discharge curves, such as voltage difference, temperature change rate, and voltage inflection point kurtosis. Then, machine learning methods are used to predict SOH based on these physical discrete features. However, the traditional battery SOH prediction process cannot correlate the intrinsic relationships between various battery variables, resulting in low accuracy in the predicted battery SOH. Summary of the Invention

[0004] In view of this, embodiments of the present invention provide a method and apparatus for predicting battery health status, a storage medium and an electronic device. The present invention obtains associated feature data by using a feature extraction network for deep mining, and uses this data as input data for a recurrent neural network, so that the recurrent neural network can predict the battery health status based on the feature data, effectively improving the accuracy of the prediction.

[0005] To achieve the above objectives, the embodiments of the present invention provide the following technical solutions:

[0006] A method for predicting battery health status, comprising:

[0007] Obtain battery characterization information within a preset time range, including information on voltage, current, temperature, and battery capacity.

[0008] The feature extraction network in the pre-trained battery health status prediction model is used to extract associated feature data from the battery characterization information;

[0009] The recurrent neural network in the battery health status prediction model is used to process the associated feature data to obtain the battery health status prediction data.

[0010] Optionally, in the above method, the step of extracting associated feature data from the battery representation information using the feature extraction network in the pre-trained battery health state prediction model includes:

[0011] The feature extraction network is invoked to process the battery characterization information to obtain initial feature data;

[0012] The initial feature data is input into the fully connected layer of the battery health state prediction model, so that the fully connected layer performs dimensionality reduction on the initial feature data to obtain associated feature data.

[0013] Optionally, the recurrent neural network described above can be one of a long short-term memory recurrent neural network, a gated recurrent unit, or a temporal convolutional neural network.

[0014] Optionally, in the above method, processing the associated feature data using the recurrent neural network in the battery health status prediction model to obtain the battery health status prediction data includes:

[0015] The recurrent neural network is triggered to process the associated feature data based on preset network weights, and output the battery health status prediction data.

[0016] Optionally, the training process for the battery health status prediction model, as described above, includes:

[0017] Obtain battery experimental data;

[0018] The battery experimental data is input into the state prediction model, and the state prediction model is trained until the state prediction model meets the preset convergence condition to obtain the pre-trained model.

[0019] Obtain battery operating condition data;

[0020] By applying the battery operating condition data, the network weights of the pre-trained model are adjusted to obtain a battery health status prediction model.

[0021] The above methods may also include:

[0022] The battery health state prediction model is pruned to obtain a pruned battery health state prediction model.

[0023] Optionally, the above method includes inputting the battery experimental data into the state prediction model and training the state prediction model until the state prediction model meets a preset convergence condition to obtain a pre-trained model, comprising:

[0024] Select the target data from the battery experiment data;

[0025] The target data is input into the state prediction model, which processes the target data and outputs prediction information corresponding to the target data. The prediction information includes health prediction data and health change rate.

[0026] Based on the predicted information and the preset supervision signal corresponding to the target data, determine whether the state prediction model satisfies the convergence condition;

[0027] When it is determined that the state prediction model does not meet the convergence condition, the network weights of the state prediction model are adjusted based on the prediction information and the supervision signal, and then the process returns to the step of selecting target data in the battery experimental data.

[0028] When it is determined that the state prediction model satisfies the convergence condition, the state prediction model is determined as a pre-trained model.

[0029] Optionally, in the above method, the step of processing the target data by the state prediction model and outputting prediction information corresponding to the target data includes:

[0030] The feature extraction network in the state prediction model is invoked to process the target data and output the extracted data.

[0031] The extracted data is dimensionality-reduced using the fully connected layer in the state prediction model, and the dimensionality-reduced data is output.

[0032] Based on the dimensionality reduction data, the rate of change in health status is determined, and the recurrent neural network in the state prediction model is used to process the dimensionality reduction data to output health prediction data.

[0033] Both the rate of change in health status and the health prediction data are determined as prediction information.

[0034] Optionally, in the above method, adjusting the network weights of the state prediction model based on the prediction information and the supervision signal includes:

[0035] Based on the rate of change of health status in the predicted information and the preset rate of change of health status in the supervision signal, the network weights of the feature extraction network in the state prediction model are adjusted.

[0036] Based on the health prediction data in the prediction information and the preset health data in the supervision signal, the network weights of the recurrent neural network and the network weights of the feature extraction network in the state prediction model are adjusted.

[0037] A battery health state prediction device, comprising:

[0038] The first acquisition unit is used to acquire battery characterization information of the battery within a preset time range, the battery characterization information including information on voltage, current, temperature and battery capacity;

[0039] The extraction unit is used to extract associated feature data from the battery characterization information using the feature extraction network in the pre-trained battery health state prediction model;

[0040] The processing unit is used to process the associated feature data using the recurrent neural network in the battery health status prediction model to obtain the battery health status prediction data.

[0041] Optionally, in the aforementioned apparatus, the extraction unit includes:

[0042] The calling subunit is used to call the feature extraction network to process the battery characterization information and obtain initial feature data;

[0043] The dimensionality reduction subunit is used to input the initial feature data into the fully connected layer of the battery health state prediction model, so that the fully connected layer performs dimensionality reduction processing on the initial feature data to obtain associated feature data.

[0044] Optionally, the recurrent neural network in the aforementioned device is one of a long short-term memory recurrent neural network, a gated recurrent unit, or a temporal convolutional neural network.

[0045] Optionally, the processing unit in the aforementioned apparatus includes:

[0046] The trigger subunit is used to trigger the recurrent neural network to process the associated feature data based on preset network weights and output the battery health status prediction data.

[0047] The aforementioned apparatus may optionally further include:

[0048] The second acquisition unit is used to acquire battery experimental data;

[0049] The training unit is used to input the battery experimental data into the state prediction model and train the state prediction model until the state prediction model meets the preset convergence condition, thus obtaining a pre-trained model.

[0050] The third acquisition unit is used to acquire battery operating condition data;

[0051] The adjustment unit is used to adjust the network weights of the pre-trained model using the battery operating condition data to obtain a battery health status prediction model.

[0052] The aforementioned apparatus may optionally further include:

[0053] The pruning unit is used to prune the battery health state prediction model to obtain the pruned battery health state prediction model.

[0054] Optionally, the training unit in the aforementioned apparatus includes:

[0055] Select sub-unit, used to select target data in the battery experimental data;

[0056] The processing subunit is used to input the target data into the state prediction model, so that the state prediction model processes the target data and outputs prediction information corresponding to the target data, the prediction information including health prediction data and health change rate;

[0057] The judgment subunit is used to determine whether the state prediction model satisfies the convergence condition based on the prediction information and the preset supervision signal corresponding to the target data.

[0058] The adjustment subunit is used to adjust the network weights of the state prediction model based on the prediction information and the supervision signal when it is determined that the state prediction model does not meet the convergence condition, and then return to the step of selecting target data in the battery experimental data.

[0059] A sub-unit is defined to determine the state prediction model as a pre-trained model when it is determined that the state prediction model satisfies the convergence condition.

[0060] Optionally, the processing subunit in the aforementioned apparatus includes:

[0061] The calling module includes calling the feature extraction network in the state prediction model to process the target data and output the extracted data;

[0062] The dimensionality reduction module is used to perform dimensionality reduction processing on the extracted data using the fully connected layer in the state prediction model, and output dimensionality-reduced data.

[0063] The determination module is used to determine the rate of change of health status based on the dimensionality reduction data, and to process the dimensionality reduction data using the recurrent neural network in the state prediction model to output health prediction data.

[0064] Both the rate of change in health status and the health prediction data are determined as prediction information.

[0065] Optionally, the adjustment subunit in the aforementioned apparatus includes:

[0066] The first adjustment module is used to adjust the network weights of the feature extraction network in the state prediction model based on the health change rate in the predicted information and the preset health change rate in the supervision signal.

[0067] The second adjustment module is used to adjust the network weights of the recurrent neural network and the feature extraction network in the state prediction model based on the health prediction data in the prediction information and the preset health data in the supervision signal.

[0068] A storage medium comprising stored instructions, wherein, when the instructions are executed, the device in which the storage medium resides executes the battery health state prediction method as described above.

[0069] An electronic device includes a memory and one or more instructions, wherein one or more instructions are stored in the memory and configured to be executed by one or more processors as described above for predicting battery health status.

[0070] Compared with the prior art, the present invention has the following advantages:

[0071] This invention provides a method, apparatus, storage medium, and electronic device for predicting battery health status. In this method, a feature extraction network within a battery health status prediction model extracts correlated feature data from battery characterization information. This correlated feature data is feature data that deeply mines the correlations between various battery parameters. This correlated feature data is then used as input data for a recurrent neural network (RNN). The RNN processes the correlated feature data and outputs predicted battery health status data. In the process of predicting battery health status, the feature extraction network extracts deeply correlated feature data, which is then used as input data for a recurrent neural network, thereby improving the accuracy of battery health status prediction. Attached Figure Description

[0072] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0073] Figure 1 This is a flowchart of a method for predicting the state of harm (SOH) of a battery using existing technologies.

[0074] Figure 2 A flowchart illustrating another method for predicting battery SOH using existing technologies;

[0075] Figure 3 A flowchart illustrating a method for predicting battery health status according to an embodiment of the present invention;

[0076] Figure 4A flowchart illustrating the training process of the battery health status prediction model provided in this embodiment of the invention;

[0077] Figure 5 This is a flowchart of a method for training a state prediction model according to an embodiment of the present invention;

[0078] Figure 6 This is a flowchart illustrating the process of training a battery health state prediction model in the method provided in this embodiment of the invention;

[0079] Figure 7 This is a structural example diagram of the battery health state prediction model provided in an embodiment of the present invention;

[0080] Figure 8 Another structural example diagram of the battery health state prediction model provided in the embodiments of the present invention;

[0081] Figure 9 A schematic diagram of the structure of a battery health status prediction device provided in an embodiment of the present invention;

[0082] Figure 10 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0083] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0084] In this application, the terms "comprising," "including," or any other variations thereof are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0085] Reference Figure 1 This is a flowchart illustrating existing methods for predicting battery state of harmonics (SOH). The flowchart utilizes characteristic changes in voltage, current, and temperature from the battery's charge / discharge curves, such as voltage difference, temperature change rate, and voltage kurtosis, and employs machine learning methods to predict the battery's SOH. (Refer to...) Figure 2This is a flowchart illustrating another method for predicting battery state-of-the-art (SOH) using existing technologies. The flowchart uses a historical SOH sequence of the battery for prediction, and a time-series prediction model is used to process the historical SOH sequence to predict the SOH of the battery in the next cycle. Current prediction methods cannot perform deeper feature mining on variables such as battery voltage, current, and temperature, nor can they uncover the inherent logical relationships between these variables. This results in limited scenario adaptability and generalization performance of existing prediction methods. Furthermore, the process of using time-series prediction algorithms to predict historical SOH information cannot guarantee the accuracy of historical information and lacks feature input based on the battery's intrinsic information, thus exhibiting certain limitations. Therefore, the results of predicting battery SOH using traditional methods are inaccurate.

[0086] To improve the accuracy of predicting battery state of health (SOH), this invention provides a method, apparatus, storage medium, and electronic device for predicting battery SOH. In the process of predicting battery SOH, a feature extraction network is used to extract associated feature data, which is then used as input data to a recurrent neural network to obtain predicted data. Using a feature extraction network allows for in-depth analysis of the relationships between various variables in the battery, thereby obtaining associated feature data. The extracted associated features are then used as input data to the recurrent neural network to predict battery SOH, thus improving the accuracy of the prediction results.

[0087] This invention can be used in a wide range of general-purpose or special-purpose computing environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor devices, distributed computing environments including any of the above devices, etc. The method provided by this invention can be applied in a battery management system, where the processor of the battery management system can be the executing entity.

[0088] Reference Figure 3 The following is a flowchart of a method for predicting battery health status according to an embodiment of the present invention, which is described in detail below:

[0089] S101. Obtain battery characterization information within a preset time range.

[0090] It should be noted that the battery characterization information obtained may be one or multiple.

[0091] Preferably, battery characterization information of the battery over multiple charge-discharge cycles within a preset time range is obtained. The preset time range can be set according to actual needs. The preset time range is a historical time period, such as battery characterization information within 24 hours. Here, 24 hours is the time interval between the current time and the historical time, and the duration between the historical time and the current time is 24 hours.

[0092] Battery characterization information includes information on the voltage, current, temperature, and battery capacity of the cell at various moments during the current charge-discharge cycle. Preferably, battery characterization information includes information on the changes in various variables such as battery voltage, current, temperature, and battery capacity.

[0093] Battery characterization information consists of cyclic data of cell charging and discharging. This data can be viewed as a matrix without time-series information, taking into account the spatial correlation and coupling between cell observation attribute data.

[0094] S102. Use the feature extraction network in the pre-trained battery health status prediction model to extract associated feature data from the battery characterization information.

[0095] Preferably, the feature extraction network can be a residual convolutional neural network (ResNet), a feature extraction network with a stacked one-dimensional convolutional neural network structure, etc. Furthermore, the feature extraction network can also be called a deep feature extraction network. The feature extraction network can perform in-depth analysis of battery characterization information, thereby extracting correlation feature data containing deep logical relationships between various variables such as battery voltage, current, temperature, and battery capacity.

[0096] Furthermore, the process of extracting associated feature data from battery characterization information is as follows:

[0097] The feature extraction network is invoked to process the battery characterization information to obtain initial feature data;

[0098] The initial feature data is input into the fully connected layer in the battery health state prediction model, which then performs dimensionality reduction on the initial feature data to obtain associated feature data.

[0099] It should be noted that when there are multiple layers in the feature extraction network, the weights in the feature extraction network are shared; furthermore, the fully connected layer can be understood as a fully connected network (FC), which can reduce the dimensionality of the initial feature data to a one-dimensional short sequence feature (F, m×1). Thus, associated feature data can be obtained. At this time, the associated feature data is a high-dimensional feature obtained after deep mining.

[0100] S103. The recurrent neural network in the battery health status prediction model is used to process the associated feature data to obtain the battery health status prediction data.

[0101] Preferably, the recurrent neural network of the present invention is one of Long Short Term Memory Network (LSTM), Gated Recurrent Unit (GRU), or Temporal Convolutional Network (TCN). Furthermore, the recurrent neural network of the present invention includes, but is not limited to, the above three networks, and other neural networks that can be used to process time series are also applicable to the present invention.

[0102] Preferably, the recurrent neural network in the battery health status prediction model processes the associated feature data based on preset network weights and outputs battery health status prediction data. Preferably, the preset network weights of the recurrent neural network are the network weights of the recurrent neural network after the battery health status prediction model has been trained.

[0103] Furthermore, the obtained battery health status prediction data can be the predicted battery SOH value, which can reflect the battery's health status.

[0104] In the method provided by this invention, a feature extraction network in a battery health status prediction model is used to extract associated feature data from battery characterization information. This associated feature data is feature data that deeply mines the correlation between various parameters of the battery. Then, this associated feature data is used as input data for a recurrent neural network. The recurrent neural network processes the associated feature data and outputs predicted battery health status data. In the process of predicting battery health status, the feature extraction network extracts deeply correlated feature data, and then uses this correlated feature data as input data for a recurrent neural network, thereby improving the accuracy of battery health status prediction.

[0105] Reference Figure 4 The flowchart below shows the training process of the battery health state prediction model provided in this embodiment of the invention.

[0106] S201. Obtain battery experimental data.

[0107] The battery experimental data of the present invention are experimental data of cell charging and discharging under multiple working conditions. Furthermore, the battery experimental data includes experimental battery characterization information of multiple charge and discharge cycles. The experimental battery characterization information can be a matrix without time sequence information. The experimental battery characterization information includes, but is not limited to, parameter changes of voltage, current, temperature and battery capacity of the cell during the charge and discharge cycle. Furthermore, each experimental battery characterization information also includes, but is not limited to, information such as battery SOC value, preset health change rate and preset SOH value.

[0108] The experimental battery characterization information includes information such as the battery's voltage, current, temperature, and battery capacity at multiple times.

[0109] For example, battery experimental data can be represented as X = [x1, x2, ..., x n ], where x n This represents the experimental battery characterization information for the nth charge-discharge cycle.

[0110] Furthermore, the changes in voltage, current, temperature, and battery capacity in the experimental battery characterization information can be represented in matrix form. Specifically, for example, the experimental battery characterization information x for the nth charge-discharge cycle. n For example:

[0111]

[0112] in, This represents the voltage at time T in the nth cycle. This represents the current at time T in the nth cycle. This represents the temperature at time T in the nth cycle. This represents the battery capacity at time T of the nth cycle.

[0113] S202. Input the battery experimental data into the state prediction model and train the state prediction model until the state prediction model meets the preset convergence condition to obtain the pre-trained model.

[0114] It should be noted that the state prediction model provided by this invention includes a feature extraction network, a fully connected layer, and a recurrent neural network. For a description of the feature extraction network, the fully connected layer, and the recurrent neural network, please refer to the description above, and it will not be repeated here.

[0115] Reference Figure 5 The following is a flowchart of a method for training a state prediction model according to an embodiment of the present invention, which is described in detail below:

[0116] S301. Select the target data from the battery experiment data.

[0117] Preferably, the battery experimental data includes multiple experimental battery characterization information. Target data is selected from each experimental battery characterization information. Preferably, the target data can be selected according to the order in which the experimental battery characterization information is sorted. For example, the experimental battery characterization information can be sorted according to the order of charge-discharge cycles. Specifically, the experimental battery characterization information belonging to the first charge-discharge cycle is placed first, followed by the experimental battery characterization information belonging to the second charge-discharge cycle, and then the experimental battery characterization information belonging to the third charge-discharge cycle, and so on. This invention will not elaborate further.

[0118] S302. Input the target data into the state prediction model, so that the state prediction model processes the target data and outputs prediction information corresponding to the target data.

[0119] It should be noted that the prediction information includes health prediction data and the rate of change in health status. Preferably, the health prediction data can be the predicted SOH value.

[0120] The process by which the state prediction model processes the target data and outputs prediction information corresponding to the target data is as follows:

[0121] S401. Call the feature extraction network in the state prediction model to process the target data and output the extracted data.

[0122] S402. Use the fully connected layer in the state prediction model to perform dimensionality reduction on the extracted data and output the dimensionality-reduced data.

[0123] S403. Based on the dimensionality-reduced data, determine the rate of change in health status, and use the recurrent neural network in the state prediction model to process the dimensionality-reduced data and output health prediction data.

[0124] Health prediction data refers to the SOH value predicted by the state prediction model based on the input data.

[0125] S404. Both the rate of change in health status and the health prediction data are determined as prediction information.

[0126] It should be noted that the extracted data is feature data that has not yet been dimensionality reduced.

[0127] The process for determining the health change rate is explained as follows: Preferably, the health change rate y = ΔSOH * DOD. When the target data is the data from the first charge-discharge cycle of the cell, the health change rate of the target data is determined to be 0. When the target data is not the data from the first charge-discharge cycle of the cell, it is necessary to determine the SOH difference ΔSOH of the target data. Further, the SOH difference is based on the difference between the preset SOH value of the current target data and the preset SOH value of the previous target data, specifically: ΔSOH = current SOH - previous SOH. Further, the current SOH is the preset SOH value of the current target data, and the previous SOH is the preset SOH value of the previous target data. After determining the ΔSOH of the current target data, it is necessary to determine the depth of discharge (DOD) of the current target data. Further, the DOD of the current target data is the difference between the maximum SOC and the minimum SOC value in the target data.

[0128] Furthermore, dimensionality-reduced data is data extracted from the target data. Therefore, dimensionality-reduced data contains data for each parameter that determines the rate of change in health status. The corresponding data can be directly applied from the dimensionality-reduced data to determine the rate of change in health status of the target data.

[0129] S303. Based on the prediction information and the preset supervision signal corresponding to the target data, determine whether the state prediction model meets the convergence condition; if it is determined that the state prediction model does not meet the convergence condition, execute S304; if it is determined that the state prediction model meets the convergence condition, execute S305.

[0130] Each target data point has a monitoring signal, which includes a preset health change rate and preset health data. Preferably, both the preset health change rate and preset health data are preset based on the actual situation. Furthermore, the preset health data can be represented as a preset SOH.

[0131] Based on the health change rate and health prediction data in the prediction information, as well as the preset health change rate and preset health data in the preset supervision signal corresponding to the target data, it can be determined whether the loss function of the state prediction model has converged. When it is determined that the loss function has converged, it is determined that the state prediction model meets the convergence condition. When it is determined that the loss function has not converged, it is determined that the state prediction model has not met the convergence condition.

[0132] For example, the process of determining whether the loss function of the state prediction model has converged can be as follows: apply a preset first function to determine the difference between the rate of change of health in the prediction information and the preset rate of change of health in the supervision signal corresponding to the target data, and obtain the first function value; apply a preset second function to determine the difference between the health prediction data in the prediction information and the preset health data in the supervision signal corresponding to the target data, and obtain the second function value; when it is determined that the first function value is stably less than or equal to the first preset convergence value, and the second function value is stably less than or equal to the second preset convergence value, it is determined that the loss function of the state prediction model has converged; otherwise, it is determined that the loss function of the state prediction model has not converged.

[0133] Preferably, there are multiple ways to determine whether the first function value is stably less than or equal to the first preset convergence value. For example, when the number of times the first function value is less than or equal to the first preset convergence value reaches a preset number, it can be determined that the first function value is stably less than or equal to the first preset convergence value. Other methods will not be described in detail here. The method for determining whether the second function is stably less than or equal to the second convergence value can refer to the process of determining whether the first function value is stably less than or equal to the first preset convergence value, which will not be described in detail here.

[0134] The first predicted convergence value and the second preset convergence value can be the same.

[0135] Furthermore, there are many other ways to determine whether a state prediction model meets the convergence condition, which will not be illustrated in this invention. There are also multiple ways to determine whether the loss function of the state prediction model converges, which will not be illustrated in this invention.

[0136] S304. Adjust the weights of the state prediction model based on the prediction information and supervision signals, and then return to execute S301.

[0137] Preferably, the network weights of the feature extraction network in the state prediction model are adjusted based on the health change rate in the predicted information and the preset health change rate in the supervision signal; for example, the difference between the health change rate and the preset health change rate is determined as the first difference, and then the network weights of the feature extraction network are adjusted based on the first difference.

[0138] Based on the health prediction data in the prediction information and the preset health data in the supervision signal, the network weights of the recurrent neural network and the feature extraction network in the state prediction model are adjusted; for example, the difference between the health prediction data and the preset health data is determined as the second difference, and then the network weights of the feature extraction network and the recurrent neural network are adjusted based on the second difference.

[0139] This invention uses two types of information to adjust the network weights of the recurrent neural network and the feature extraction network, so that the trained feature extraction network and recurrent neural network can output data that is more suitable for prediction, resulting in higher prediction accuracy.

[0140] Preferably, after adjusting the weights of the state prediction model, it is necessary to return to execute S301 until the state prediction model is trained to meet the convergence condition.

[0141] S305. The state prediction model is determined as a pre-trained model.

[0142] This invention trains a state prediction model using battery experimental data to obtain a relatively satisfactory prediction model. However, battery cells experience various operating conditions during actual use, and experimental data cannot fully cover all operating conditions of battery cells during use. Therefore, in order to improve the prediction accuracy of the model and enable the model to predict batteries under various operating conditions, it is necessary to adjust the obtained model to achieve higher prediction accuracy.

[0143] S203, Obtain battery operating condition data.

[0144] Battery operating condition data includes data generated by the cells during actual application, including actual battery characterization information after multiple charge-discharge cycles.

[0145] Actual battery characterization information includes changes in battery voltage, current, temperature, and capacity. It should be noted that the actual battery characterization information can be found in the previous section on experimental battery characterization information, and will not be repeated here.

[0146] S204. Using battery operating condition data, adjust the network weights of the pre-trained model to obtain a battery health status prediction model.

[0147] Battery operating condition data is input into a pre-trained model to train the model, thereby adjusting the network weights of the pre-trained model and obtaining a battery health status prediction model.

[0148] It should be noted that the process of training a pre-trained model can be referred to the process of training a state prediction model, and will not be repeated here.

[0149] It should be noted that the data used in the two model training processes are different. The battery experimental data was used to train the state prediction model. Using battery experimental data can quickly train the model and shorten the training time. On the other hand, battery operating condition data is used to improve the prediction accuracy of the model, so that the final model can be applied to various actual operating conditions.

[0150] The pre-trained model is trained using battery operating condition data, which is data generated during the actual application of batteries. By using the actual usage data of batteries, the network weights of the pre-trained model are adjusted to make the pre-trained model more suitable for various scenarios of battery application in real time, thereby improving the prediction accuracy of the model under various operating conditions.

[0151] Furthermore, the battery health state prediction model can be pruned to obtain a pruned model. To reduce the computational burden on the platform running the model, pruning can be applied to platforms with lower computing power. Preferably, the feature extraction network and recurrent neural network in the battery health state prediction model are pruned, removing some weights that have less impact on the model results, thus reducing the computational burden on the platform.

[0152] Furthermore, when the platform has sufficient computing power, it can directly use battery health status prediction models.

[0153] The supervision signal used in the training process of this invention includes two data: preset health change rate and preset health data. The preset health change rate reflects the health loss at the current discharge depth due to the data changes in the current charge-discharge cycle, and removes the influence of the accumulation effect of time series. Together with the original time series supervision branch, the two-way supervision enables the high-dimensional features to have both spatial and temporal information enhancement, thereby making the prediction accuracy of the trained model higher.

[0154] The method provided by this invention trains a state prediction model using battery experimental data and actual battery operating condition data, ultimately obtaining a battery health state prediction model. The experimental data serves as standard data, allowing for rapid model training. After initial training, the weights in the trained model are adjusted using actual battery operating condition data. This ensures that the final model can cover various complex operating conditions in real-world battery applications, guaranteeing its practical usability. Therefore, this invention uses battery experimental data and actual battery operating condition data to train the model, effectively shortening the training time while ensuring the final model is applicable to predicting the state of batteries in actual use. Furthermore, by pruning the model, the computing power required for the platform supporting the model can be reduced, allowing the model to be used on platforms with lower computing power, effectively expanding the model's application scope.

[0155] Reference Figure 6 The flowchart for training a battery health state prediction model in the method provided in this embodiment of the invention is described in detail below:

[0156] The battery health state prediction model can be trained on a cloud platform. Furthermore, the battery health state prediction model includes a residual convolutional neural network, a fully connected layer, and an LSTM. The charging and discharging data under actual cell operating conditions and the multi-condition charging and discharging experimental data of the cell are both uploaded to the cloud platform as training data. Then, the residual convolutional neural network and LSTM in the battery health state prediction model are trained using the training data so as to finally obtain a battery health state prediction model that meets the conditions.

[0157] Preferably, the model training process can be divided into four parts, as detailed below:

[0158] 1) Use experimental data for pre-training.

[0159] Experimental data from multi-condition charge-discharge of battery cells can accurately estimate the state of harm (SOH) for each cycle, improving the model's accuracy in predicting SOH. Cloud platforms can provide powerful computing capabilities for training deep learning models, enhancing their accuracy and generalization ability. First, a training model is built on the cloud platform, inputting battery characterization information X = [x1, x2, ..., xn] for n charge-discharge cycles. n Battery characterization information includes changes in voltage, current, temperature, and battery capacity during charge-discharge cycles. i For example:

[0160]

[0161] Among them, V i T This represents the voltage at time T in the i-th cycle. Let T represent the current at time T during the i-th cycle. i T This represents the temperature at time T of the i-th cycle. This represents the battery capacity at time T of the i-th cycle.

[0162] Preferably, the input n charge-discharge cycle data are obtained by sliding window of battery cycle data throughout the entire life cycle. For example, if there are N full life cycle battery cycle data, the sliding window length is n, and the sliding step size is s, then (Nn) / s input data can be obtained, and the dimension of each input data is n*T*4.

[0163] Reference Figure 7 This is a structural example diagram of the battery health state prediction model provided in an embodiment of the present invention, combined with... Figure 6 and Figure 7 The training process of the battery health status prediction model is explained.

[0164] Reference Figure 7The input matrix in the figure represents battery characterization information. The time series model is constructed using LSTM. Each battery characterization information X is input into a feature extraction network, which can be a residual convolutional neural network. Different layers of the residual convolutional neural network share network weights. The residual convolutional neural network reduces the output features to one-dimensional short sequence features (F, m×1) through a fully connected network, thus obtaining high-dimensional semantic features (i.e., the associated feature data mentioned above). Secondly, the fully connected layer inputs the obtained high-dimensional semantic features into the Long Short-Term Memory (LSTM) neural network, enabling the LSTM to perform time series prediction based on the high-dimensional semantic features and output the predicted SOH. Furthermore, the output SOH can be represented by Y. On the other hand, the fully connected layer outputs the health change rate based on the high-dimensional semantic features. The health change rate is the product of the SOH change rate and the cell discharge depth, i.e., y = ΔSOH * DOD, where y is the health change rate, ΔSOH is the SOH change rate, and DOD is the current cell discharge depth. Furthermore, ΔSOH is the difference between the current SOH and the previous SOH. Furthermore, the current SOH is the preset SOH corresponding to the data currently input into the feature extraction network, and the previous SOH is the preset SOH corresponding to the data previously input into the feature extraction network.

[0165] As shown in the figure, during the training process, each time battery representation information is input into the model, the model processes the battery representation information and outputs a supervision signal. This supervision signal can be used as the supervision branch of the model. When the model has not yet met the convergence condition, the output SOH and health change rate are used as supervision signals to adjust the weights of the feature extraction network and LSTM. Thus, a certain loss function can be applied to train the model based on this supervision signal, so that the feature extraction network can extract high-dimensional features to enhance semantic information.

[0166] 2) Fine-tune using actual operating data.

[0167] The charging and discharging modes of battery cells under actual operating conditions are more complex, and the decay law of battery cell life will change under complex operating conditions. We use some accurate prediction data of actual operating conditions under SOH to fine-tune the pre-trained model.

[0168] Preferably, actual working condition data can be input into the model after it has been trained using experimental data to further train the model and adjust the network weights in the model.

[0169] For example, the model weights trained using experimental data can be saved. These model weights include the network weights of the residual convolutional neural network, fully connected layers, and LSTM network. The saved model weights can then be updated in another model on the cloud platform. The pre-trained weights can then be input into this model, and the actual working data can be input into the model to train it. This further trains the model's weights, ultimately resulting in a fully trained model.

[0170] 3) Model pruning or lightweighting.

[0171] The model trained using real-world working data will be pruned or lightweighted. This mainly involves pruning the residual convolutional neural network and LSTM network in the model, removing some weights that have a small impact on the model results, and reducing the computational pressure on the platform for a large number of weighted models.

[0172] 4) Embedded platform applications.

[0173] Preferably, after pruning or lightweighting the model, the model can be put into use. Specifically, the pruned model can be imported into an embedded platform (such as a battery management system) so that the platform can use the imported model to perform SOH prediction. Preferably, the platform should have a certain AI algorithm computing capability.

[0174] Reference Figure 8 This is another structural example diagram of the battery health state prediction model provided in an embodiment of the present invention. This structure is relative to... Figure 7 The structure shown adds an LSTM hidden layer h(t), and outputs Y = SOH through a fully connected layer. For details on the implementation, please refer to [link / reference needed]. Figure 7 The explanation will not be repeated here.

[0175] This invention treats charge-discharge cycle data as a matrix without temporal information. Considering the spatial correlation and coupling between cell observation attribute data, it uses a residual convolutional neural network (ResNet) for deep mining of statistical data features to extract high-dimensional feature information. Thus, memoryless extraction of representational information during battery charge-discharge processes can be achieved using deep networks, exponentially reducing the configuration requirements for large data storage space and lowering costs. When using deep networks to extract high-dimensional features, this invention adds a health rate of change layer as a supervisory signal to update the feature weights during the training of the deep network. This health rate of change reflects the health loss of the data changes in the current cycle at the current discharge depth and removes the influence of temporal accumulation effects. Combined with the predicted SOH from the original time-series output as a supervisory signal, these two supervisory signals enhance the high-dimensional features with both spatial and temporal information. This invention also combines the ability of time series network models to update and predict historical data, enhancing the influence of SOH on the cumulative effect of battery loss. The time series network model of this invention differs from traditional time series prediction methods that use the current time series as input and the time-shifted time series as output. This invention uses extracted high-dimensional features (i.e., the associated feature data mentioned above) as input and the current cycle point SOH as output. By using the extracted high-dimensional features as input to the time series network model, the accuracy of prediction is improved.

[0176] and Figure 3 Correspondingly, the present invention also provides a battery health status prediction device, which is used to support Figure 3 Specifically, the device can be configured within a battery management system.

[0177] Reference Figure 9 The above is a schematic diagram of a battery health status prediction device provided in an embodiment of the present invention, and is described in detail below:

[0178] The first acquisition unit 501 is used to acquire battery characterization information of the battery within a preset time range, the battery characterization information including information on voltage, current, temperature and battery capacity.

[0179] Extraction unit 502 is used to extract associated feature data from the battery characterization information using the feature extraction network in the pre-trained battery health state prediction model;

[0180] The processing unit 503 is used to process the associated feature data using the recurrent neural network in the battery health status prediction model to obtain the battery health status prediction data.

[0181] In the apparatus provided by this invention, a feature extraction network in a battery health status prediction model extracts associated feature data from battery characterization information. This associated feature data is feature data that deeply mines the correlation between various parameters of the battery. Then, this associated feature data is used as input data for a recurrent neural network. The recurrent neural network processes the associated feature data and outputs predicted battery health status data. In the process of predicting battery health status, the feature extraction network extracts deeply correlated feature data, and then uses this correlated feature data as input data for a recurrent neural network, thereby improving the accuracy of battery health status prediction.

[0182] In another embodiment provided by the present invention, the extraction unit of the device can be configured as follows:

[0183] The calling subunit is used to call the feature extraction network to process the battery characterization information and obtain initial feature data;

[0184] The dimensionality reduction subunit is used to input the initial feature data into the fully connected layer of the battery health state prediction model, so that the fully connected layer performs dimensionality reduction processing on the initial feature data to obtain associated feature data.

[0185] In another embodiment of the present invention, the recurrent neural network of the device is one of a long short-term memory recurrent neural network, a gated recurrent unit, or a temporal convolutional neural network.

[0186] In another embodiment provided by the present invention, the processing unit 503 of the device includes:

[0187] The trigger subunit is used to trigger the recurrent neural network to process the associated feature data based on preset network weights and output the battery health status prediction data.

[0188] In another embodiment provided by the present invention, the device may also be configured as follows:

[0189] The second acquisition unit is used to acquire battery experimental data;

[0190] The training unit is used to input the battery experimental data into the state prediction model and train the state prediction model until the state prediction model meets the preset convergence condition, thus obtaining a pre-trained model.

[0191] The third acquisition unit is used to acquire battery operating condition data;

[0192] The adjustment unit is used to adjust the network weights of the pre-trained model using the battery operating condition data to obtain a battery health status prediction model.

[0193] In another embodiment provided by the present invention, the device may also be configured as follows:

[0194] The pruning unit is used to prune the battery health state prediction model to obtain the pruned battery health state prediction model.

[0195] In another embodiment of the present invention, the training unit of the device includes:

[0196] Select sub-unit, used to select target data in the battery experimental data;

[0197] The processing subunit is used to input the target data into the state prediction model, so that the state prediction model processes the target data and outputs prediction information corresponding to the target data, the prediction information including health prediction data and health change rate;

[0198] The judgment subunit is used to determine whether the state prediction model satisfies the convergence condition based on the prediction information and the preset supervision signal corresponding to the target data.

[0199] The adjustment subunit is used to adjust the network weights of the state prediction model based on the prediction information and the supervision signal when it is determined that the state prediction model does not meet the convergence condition, and then return to the step of selecting target data in the battery experimental data.

[0200] A sub-unit is defined to determine the state prediction model as a pre-trained model when it is determined that the state prediction model satisfies the convergence condition.

[0201] In another embodiment provided by the present invention, the processing subunit of the device includes:

[0202] The calling module includes calling the feature extraction network in the state prediction model to process the target data and output the extracted data;

[0203] The dimensionality reduction module is used to perform dimensionality reduction processing on the extracted data using the fully connected layer in the state prediction model, and output dimensionality-reduced data.

[0204] The determination module is used to determine the rate of change of health status based on the dimensionality reduction data, and to process the dimensionality reduction data using the recurrent neural network in the state prediction model to output health prediction data.

[0205] Both the rate of change in health status and the health prediction data are determined as prediction information.

[0206] In another embodiment of the present invention, the adjustment subunit of the device includes:

[0207] The first adjustment module is used to adjust the network weights of the feature extraction network in the state prediction model based on the health change rate in the predicted information and the preset health change rate in the supervision signal.

[0208] The second adjustment module is used to adjust the network weights of the recurrent neural network and the feature extraction network in the state prediction model based on the health prediction data in the prediction information and the preset health data in the supervision signal.

[0209] The specific working process of each unit, subunit, and module in the battery health state prediction device disclosed in the above embodiments can be found in the corresponding content of the battery health state prediction device method disclosed in the above embodiments of the present invention, and will not be repeated here.

[0210] This invention also provides a storage medium that includes stored instructions, wherein when the instructions are executed, the device containing the storage medium is controlled to perform the aforementioned battery health state prediction method.

[0211] This invention also provides an electronic device, the structural schematic of which is shown below. Figure 10 As shown, it specifically includes a memory 601 and one or more instructions 602, wherein one or more instructions 602 are stored in the memory 601 and configured to be executed by one or more processors 603 to perform the above-mentioned battery health state prediction method.

[0212] The specific implementation processes and derivative methods of the above embodiments are all within the protection scope of this invention.

[0213] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for system or system embodiments, since they are basically similar to method embodiments, the description is relatively simple, and relevant parts can be referred to the descriptions in the method embodiments. The systems and system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0214] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0215] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for predicting battery health status, characterized in that, include: Obtain battery characterization information within a preset time range, including information on voltage, current, temperature, and battery capacity. The feature extraction network in the pre-trained battery health status prediction model is used to extract associated feature data from the battery characterization information; The recurrent neural network in the battery health status prediction model is used to process the associated feature data to obtain the battery health status prediction data. The training process of the battery health status prediction model includes: Obtain battery experimental data; The battery experimental data is input into the state prediction model, and the state prediction model is trained until the state prediction model meets the preset convergence condition to obtain the pre-trained model. Obtain battery operating condition data; By applying the battery operating condition data, the network weights of the pre-trained model are adjusted to obtain a battery health status prediction model.

2. The method according to claim 1, characterized in that, The feature extraction network in the pre-trained battery health status prediction model extracts associated feature data from the battery characterization information, including: The feature extraction network is invoked to process the battery characterization information to obtain initial feature data; The initial feature data is input into the fully connected layer of the battery health state prediction model, so that the fully connected layer performs dimensionality reduction on the initial feature data to obtain associated feature data.

3. The method according to claim 1, characterized in that, The recurrent neural network is one of a long short-term memory recurrent neural network, a gated recurrent unit, or a temporal convolutional neural network.

4. The method according to claim 1, characterized in that, The process of using a recurrent neural network in the battery health status prediction model to process the associated feature data to obtain the battery health status prediction data includes: The recurrent neural network is triggered to process the associated feature data based on preset network weights, and output the battery health status prediction data.

5. The method according to claim 1, characterized in that, Also includes: The battery health state prediction model is pruned to obtain the pruned battery health state prediction model.

6. The method according to claim 1, characterized in that, The step of inputting the battery experimental data into the state prediction model and training the state prediction model until the state prediction model meets the preset convergence condition to obtain a pre-trained model includes: Select the target data from the battery experiment data; The target data is input into the state prediction model, which processes the target data and outputs prediction information corresponding to the target data. The prediction information includes health prediction data and health change rate. Based on the predicted information and the preset supervision signal corresponding to the target data, determine whether the state prediction model satisfies the convergence condition; When it is determined that the state prediction model does not meet the convergence condition, the network weights of the state prediction model are adjusted based on the prediction information and the supervision signal, and then the process returns to the step of selecting target data in the battery experimental data. When it is determined that the state prediction model satisfies the convergence condition, the state prediction model is determined as a pre-trained model.

7. The method according to claim 6, characterized in that, The step of enabling the state prediction model to process the target data and output prediction information corresponding to the target data includes: The feature extraction network in the state prediction model is invoked to process the target data and output the extracted data. The extracted data is dimensionality-reduced using the fully connected layer in the state prediction model, and the dimensionality-reduced data is output. Based on the dimensionality reduction data, the rate of change in health status is determined, and the recurrent neural network in the state prediction model is used to process the dimensionality reduction data to output health prediction data. Both the rate of change in health status and the health prediction data are determined as prediction information.

8. The method according to claim 6, characterized in that, Adjusting the network weights of the state prediction model based on the prediction information and the supervision signal includes: Based on the rate of change of health in the predicted information and the preset rate of change of health in the supervision signal, the network weights of the feature extraction network in the state prediction model are adjusted. Based on the health prediction data in the prediction information and the preset health data in the supervision signal, the network weights of the recurrent neural network and the network weights of the feature extraction network in the state prediction model are adjusted.

9. A device for predicting battery health status, characterized in that, include: The first acquisition unit is used to acquire battery characterization information of the battery within a preset time range, the battery characterization information including information on voltage, current, temperature and battery capacity; The extraction unit is used to extract associated feature data from the battery characterization information using the feature extraction network in the pre-trained battery health state prediction model; The processing unit is used to process the associated feature data using the recurrent neural network in the battery health status prediction model to obtain the battery health status prediction data. It also includes: a second acquisition unit, used to acquire battery experimental data; The training unit is used to input the battery experimental data into the state prediction model and train the state prediction model until the state prediction model meets the preset convergence condition, thus obtaining a pre-trained model. The third acquisition unit is used to acquire battery operating condition data; The adjustment unit is used to adjust the network weights of the pre-trained model using the battery operating condition data to obtain a battery health status prediction model.

10. A storage medium, characterized in that, The storage medium includes stored instructions, wherein, when the instructions are executed, the device in which the storage medium resides executes the battery health state prediction method as described in any one of claims 1-8.

11. An electronic device, characterized in that, It includes a memory, and one or more instructions, wherein one or more instructions are stored in the memory and configured to be executed by one or more processors as described in any one of claims 1-8.

Images