Battery system operation data generation method based on reconstruction type pre-training architecture

Abstract

The invention relates to the technical field of battery system data processing, in particular to a reconfiguration type pre-training architecture-based battery system operation data generation method, which comprises the following steps of: outputting virtual battery system operation data according to an encoder in a called pre-trained battery system data generation model; determining charging and discharging data in the operation data of the virtual battery system, and slicing the charging and discharging data according to a time dimension to form time sequence input data; setting a mask rate, and performing zero setting processing on any row and column data in the time sequence input data to form mask data; inputting the mask data into an encoder to obtain a representation vector; whether interpolation processing is carried out or not is judged according to the representation vector, if yes, interpolation processing is carried out on the representation vector, the representation vector obtained after interpolation processing serves as input data and is input into a reconstruction decoder, and corresponding reconstruction generation data are output, and if not, the representation vector is directly input into the reconstruction decoder, and the corresponding reconstruction generation data are output.

Description

Technical Field

[0001] The present invention relates to the technical field of battery system data processing, and in particular to a method for generating battery system operation data based on a reconstruction type pre-training architecture. Background Art

[0002] In the context of the booming development of the new energy industry, the safety performance of batteries as core components has become the focus of the industry. With the large-scale popularization of application scenarios such as electric vehicles and energy storage systems, the potential casualties, equipment damage and environmental risks caused by battery safety accidents (such as thermal runaway, internal short circuit, etc.) urgently require effective control measures. Achieving early warning before a fault through battery safety warning algorithms has become a key technical path to improve the safety of battery systems.

[0003] Key tasks such as battery safety warning and remaining life prediction highly rely on high-quality data support. However, in practical applications, there are often double bottlenecks of scarce labeled data and difficult acquisition of fault samples. The proportion of real fault label data corresponding to existing battery safety faults is low, resulting in low accuracy of subsequent battery safety warnings.

[0004] Based on this, there is an urgent need for a method for generating battery system operation data based on a reconstruction type pre-training architecture to solve the problem of low accuracy of battery safety warnings caused by insufficient battery operation data during the battery safety warning process in the prior art, and to achieve high data volume and high accuracy of battery operation data. Summary of the Invention

[0005] One of the purposes of the present invention is to provide a method for generating battery system operation data based on a reconstruction type pre-training architecture to solve the problem of low accuracy of battery safety warnings caused by insufficient battery operation data during the battery safety warning process in the prior art, and to achieve high data volume and high accuracy of battery operation data.

[0006] To achieve the above object, a method for generating battery system operation data based on a reconstruction type pre-training architecture is provided, including the following steps:

[0007] S1. Retrieve a pre-trained battery system data generation model from a database, and fix the network structure and weight parameters of the encoder in the battery system data generation model; the weight parameters include the weight parameters of the encoder and the reconstruction decoder;

[0008] S2. Output corresponding virtual battery system operation data according to the encoder in the retrieved pre-trained battery system data generation model;

[0009] S3. Determine the charge and discharge data in the virtual battery system operation data generated, and based on a preset data slicing strategy, slice the charge and discharge data according to the time dimension to form corresponding sequential input data i2;

[0010] S4. Arrange the formed sequential input data i2 in ascending order according to the time dimension of the segments, set a masking rate, and based on the masking rate, zero out any row and column data in the sequential input data i2 until the amount of zeroed data meets the masking rate. At this time, the sequential input data is the corresponding masked data m1;

[0011] S5. Input the masked data m1 into the encoder of the battery system data generation model to obtain 1 corresponding representation vector z2;

[0012] S6. Determine whether to perform interpolation processing based on the obtained representation vector z2. If so, based on a preset interpolation processing strategy, perform interpolation processing on the representation vector, and use the interpolated representation vector z2 ′ as input data, input it into the reconstruction decoder of the battery system data generation model, and output the corresponding reconstructed generated data r2 ′ , otherwise, directly input the representation vector into the reconstruction decoder and output the corresponding reconstructed generated data r2.

[0013] The technical principle and effect of this solution: In this solution, first, the battery system data generation model includes an encoder and a reconstruction decoder. The former maps the input data to a low-dimensional representation vector, and the latter realizes the reconstruction from the representation vector to the original data by learning the data distribution law. In step S1, the weight parameters of the encoder and decoder are fixed to avoid damaging the pre-trained model during subsequent training and ensure that the battery data feature distribution it has learned remains stable. Use the fixed encoder to directly generate virtual battery operation data (step S2). This process is based on the pre-trained model's ability to model the real data distribution and generates new samples that conform to the characteristics of the battery system through forward propagation.

[0014] Battery operation data has strong temporal correlation. By slicing in the time dimension (step S3), the charge and discharge data are divided into continuous temporal segments, retaining the time dependence of the data, which is convenient for subsequent sequence-based generation and enhancement operations.

[0015] In step S4, a masking rate is introduced to randomly zero out the rows and columns of the temporal data, simulating data loss or noise interference that may exist in the real scenario. This operation forces the model to infer the missing information through the features in the unmasked area, enhances the model's understanding of the data distribution, and provides training signals for subsequent interpolation.

[0016] After the encoder generates a representation vector for the mask data (step S5), the interpolation strategy is determined to be triggered (e.g., based on the distribution density of the representation vector or the diversity requirement of the generated data), and the representation vector is subjected to linear interpolation or Gaussian interpolation (step S6). The interpolation operation generates new sample points (e.g., z2 ′ ), expand the coverage of data distribution and further increase the diversity of generated data.

[0017] Both the original representation vector and the interpolated vector are mapped back to the data space through a reconstruction decoder with fixed parameters to generate reconstructed data containing the original features and the newly added features. This process uses the generation capability of the decoder to restore low-dimensional features into high-dimensional time series data that conforms to the laws of the battery system.

[0018] In this solution, the encoder of the pre-trained model directly generates virtual operating data that conforms to the characteristic distribution of the battery system, ensuring that the generated data inherits the physical rationality of the real data and avoiding the generation of "pseudo-data" (such as voltage jumps that violate the thermodynamic laws of the battery). The encoder directly generates virtual battery operating data, breaking through the time cost and sample limitations of real data collection. A single round can generate a sample size comparable to the original data size, set a mask rate for the time series data and perform interpolation processing. Each original data can derive multiple new data while improving the robustness of the generated data to noise, making it closer to the data quality in real industrial scenarios. Finally, the masked data is input into the decoder to output a new sample size, solving the problem of low accuracy of battery safety warnings caused by the small amount of battery operating data in the existing technology during battery safety warnings, and achieving high data volume and high accuracy of battery operating data.

[0019] By setting the mask rate and randomly setting the row and column data to zero, we can simulate noise scenarios such as sensor failure and missing data collection in real scenarios.

[0020] Furthermore, the preset data slicing strategy is:

[0021] According to the determined charge and discharge data, based on the preset data selection dimension, the charge and discharge data corresponding to the corresponding data selection dimension is selected from the charge and discharge data, and abnormal characters and invalid data are eliminated;

[0022] Set a slice length value, and according to the corresponding charge and discharge data and the slice length value, divide the charge and discharge data into charge and discharge data segments with the same slice length value. The charge and discharge data segments are the corresponding time series input data.

[0023] Beneficial effects: By eliminating abnormal characters (such as non-numerical noise and garbled codes generated by sensor communication errors) and invalid data (such as voltage / current values exceeding the physical limits of the battery), it is ensured that the charge and discharge data input into the model conforms to the actual physical laws. Based on the preset data selection dimensions (such as key physical quantities like voltage, current, temperature, SOC, etc.), data is screened to ensure that data from different batches and different acquisition cycles are consistent in terms of feature dimensions, solving the problem of dimension mismatch during multi-source data fusion.

[0024] The slicing operation requires that the lengths of the charge and discharge data segments be consistent, implicitly standardizing the sampling frequency of the original data. If there are problems with uneven sampling intervals in the original data (such as 1 second per time in some periods and 5 seconds per time in some periods), it is necessary to unify it to a fixed frequency through interpolation or resampling before slicing to avoid feature learning biases caused by time series misalignment.

[0025] Furthermore, the data selection dimensions include a time dimension, a single-cell voltage sequence dimension, a current dimension, an SOC dimension, and a temperature dimension.

[0026] Furthermore, the training strategy corresponding to the battery system data generation model pre-trained in S1 is as follows:

[0027] Step 1, retrieve historical real vehicle operation data from the cloud data platform of the vehicle, determine the SOC differences corresponding to each data in the historical real vehicle operation data, and select the historical charge and discharge data corresponding to the preset SOC threshold;

[0028] Step 2, based on the corresponding historical charge and discharge data and the corresponding data slicing strategy, perform slicing processing on the historical charge and discharge data, and output the historical input time series data corresponding to the historical charge and discharge data;

[0029] Step 3, construct a battery system data generation model, and the battery system data generation model includes an encoder and a reconstruction decoder;

[0030] Step 4, arrange the formed historical input time series data in ascending order according to the time dimension, randomly select the row indexes of a preset proportion, and replace all the column data corresponding to the corresponding rows with 0 to form the corresponding historical zeroed data;

[0031] Step 5, input the historical zeroed data into the encoder in the battery system data generation model to obtain 1 corresponding representation vector;

[0032] Step 6: Transfer the feature vector to the reconstruction decoder to output the corresponding reconstructed data. Calculate the reconstruction error between the reconstructed data and the feature vector, and determine whether the reconstruction error is less than the preset error threshold. If so, it is determined that the battery system data generation model has completed training. Otherwise, adjust the weight parameters of the encoder and the reconstruction decoder in the battery system data generation model according to the reconstruction error, and re-execute Step 4.

[0033] Beneficial effects: In this solution, by screening historical charge and discharge data that meet the preset SOC threshold, the typical working range during daily battery use is focused on, avoiding the interference of low-probability working conditions on model training, and enabling the model to preferentially learn the most valuable feature distributions. Screening data based on the SOC difference implicitly divides the charging and discharging process into stages, making the time-series data input into the model have clear physical meanings and facilitating the model to learn the parameter change laws in different stages.

[0034] Adopt an autoencoder (AE) architecture. Through the closed-loop training of "encoding - decoding - reconstruction error optimization", force the encoder to learn a low-dimensional feature vector that can uniquely represent the original data (for example, compress 1000-dimensional time-series data into a 50-dimensional feature vector), and the decoder learns the inverse mapping relationship from the feature vector to the original data.

[0035] This mechanism does not require manual labeling of labels and only uses the distribution law of the data itself to complete training, solving the core pain point of scarce labeled data in the battery field, especially suitable for scenarios where there is insufficient early data accumulation.

[0036] By calculating the error between the reconstructed data and the original input (such as mean squared error MSE, cosine similarity, etc.), backpropagate to update the weight parameters of the encoder and the decoder to ensure that the model gradually approaches the identity mapping of "input = output". After training is completed, fix the parameters of the encoder and the decoder (such as in Step S1), which can be directly used for tasks such as virtual data generation and masked reconstruction without retraining the model. That is, this training strategy systematically solves the training difficulties brought by "high-dimensionality, low-label, and strong time-series" of battery data through the pipeline design of "data screening - slice preprocessing - self-supervised reconstruction", not only endowing the model with powerful data generation and reconstruction capabilities, but also laying a foundation for subsequent data augmentation and model migration through lightweight feature representation, which is the premise and core driving force for the efficient operation of the entire technical solution.

[0037] Furthermore, the preset interpolation processing strategy is as follows:

[0038] According to the obtained feature vector, determine the vector dimension corresponding to the feature vector and judge the relationship between the vector dimension and the first preset threshold and the second preset threshold;

[0039] If the vector dimension is less than the first preset threshold, the corresponding interpolation processing rule is as follows:

[0040] z2 ′ = λ × z2 + (1 - λ) × z i

[0041] In the formula, z2 ′ is the representation vector after interpolation processing, z2 ′ ∈ C, z2 ∈ C, C is the category of the vector, λ is the interpolation factor, 0 ≤ λ ≤ 1; z i is the reference point vector corresponding to the reference point i randomly selected in the feature space of category C, z i ∈ C;

[0042] If the vector dimension is greater than or equal to the preset second threshold, the corresponding interpolation processing rule is as follows:

[0043]

[0044] In the formula, N is the number of reference points randomly selected in the feature space of category C, μ i is the weight value of the i-th reference point.

[0045] Beneficial effects: In this solution, by determining the vector dimension corresponding to the representation vector to select the corresponding interpolation processing rule, that is, strengthening the data coverage of the continuous feature evolution process in the low-dimensional space (such as the smooth transition of each charging stage), enabling the safety warning model to capture more refined state change trends (such as the tiny internal resistance fluctuation during the constant current to constant voltage transition), and actively generating complex failure scenarios with multi-factor coupling (such as the four-dimensional feature combination of "voltage sudden drop + temperature sudden rise + internal resistance sharp increase + gas production rate mutation" before thermal runaway), to solve the "data blind spot" problem caused by safety restrictions or high costs in real data. Brief Description of the Drawings

[0046] Figure 1 It is a logic block diagram of the method for generating battery system operation data based on the reconstruction pre-training architecture in Embodiment 1 of the present invention. Detailed Embodiments

[0047] The following is a further detailed description through specific embodiments:

[0048] Embodiment 1

[0049] A method for generating battery system operation data based on the reconstruction pre-training architecture is basically as Figure 1 shown, and includes the following steps:

[0050] S1. Retrieve the pre-trained battery system data generation model from the database, and fix the network structure and weight parameters of the encoder in the battery system data generation model; the weight parameters include the weight parameters of the encoder and the reconstruction decoder;

[0051] The training strategy corresponding to the pre-trained battery system data generation model in S1 is as follows:

[0052] Step 1. Retrieve the historical real vehicle operation data from the cloud data platform of the vehicle, determine the SOC difference corresponding to each data in the historical real vehicle operation data, and select the historical charge and discharge data corresponding to the preset SOC threshold; in this embodiment, the corresponding preset SOC threshold is the data segment where the SOC changes before and after by more than 60. For example, if the SOC of the first frame of data in the data is 10 and the SOC of the nth frame of data is 70, and their SOC change is greater than or equal to 60, then we will select such a data segment; at the same time, this difference can be modified according to the actual usage situation, is applicable to all types of batteries, and does not need to be adjusted according to different battery types. This is a manually set value.

[0053] Step 2. Based on the corresponding historical charge and discharge data and the corresponding data slicing strategy, slice the historical charge and discharge data to output the historical input time series data corresponding to the historical charge and discharge data;

[0054] Step 3. Construct a battery system data generation model, which includes an encoder and a reconstruction decoder;

[0055] Step 4. Arrange the formed historical input time series data in ascending order according to the time dimension, randomly select the row indexes of the preset proportion, and replace all the column data where the corresponding rows are located with 0 to form the corresponding historical zeroed data; in this embodiment, the preset proportion is 75%, and this proportion mainly comes from the best proportion obtained from multiple experiments, and the zeroing strategy is adopted for all the data used.

[0056] Step 5. Input the historical zeroed data into the encoder in the battery system data generation model to obtain 1 corresponding representation vector; in this embodiment, the encoder is usually a variational autoencoder (VAE) or a Transformer encoder. Among them, based on the Transformer architecture: 1 Patch partition module, which mainly divides the original input data into multiple patches for processing and supports custom dimensions. 1 1D structured position encoding module: mainly used to mark the positions of 1D data, such as time series and spectrogram data, and provide masked position information to accelerate the pre-training of the model.

[0057] A Transformer-structured encoder is used with a depth of 24 and 16 multi-head attention heads, and an Adapter layer is connected for efficient parameter fine-tuning and knowledge transfer.

[0058] The reconstructed decoder is an 8-layer Transformer. The block module structure of each layer is consistent with that of the encoder. RevIN (Reversible Instance Normalization) is used for normalization and denormalization operations in sequence modeling.

[0059] In step 6, the representation vector is passed to the reconstruction decoder, which outputs the corresponding reconstructed data. Based on the reconstructed data and the representation vector, the reconstruction error between the reconstructed data and the representation vector is calculated to determine whether the reconstruction error is less than a preset error threshold. If so, the battery system data generation model is considered to have completed training. Otherwise, the weight parameters of the encoder and reconstruction decoder in the battery system data generation model are adjusted based on the reconstruction error, and step 4 is re-executed. In this embodiment, more data is input in parallel for the large-scale pre-training in the above steps, thereby achieving model training.

[0060] S2. Generate an encoder in the model based on the retrieved pre-trained battery system data and output the corresponding virtual battery system operation data;

[0061] S3. Determine the charge and discharge data in the virtual battery system operation data based on the generated virtual battery system operation data, and slice the charge and discharge data according to the time dimension based on a preset data slicing strategy to form corresponding time series input data i2;

[0062] The preset data slicing strategy is:

[0063] According to the determined charge and discharge data, based on preset data selection dimensions, the charge and discharge data corresponding to the corresponding data selection dimensions are selected from the charge and discharge data, and abnormal characters and invalid data are eliminated; the data selection dimensions include time dimension, cell voltage sequence dimension, current dimension, SOC dimension, and temperature dimension.

[0064] Set the slice length value, and according to the corresponding charge and discharge data and the slice length value, slice the charge and discharge data according to the set slice length value into charge and discharge data segments with the same slice length value, and this charge and discharge data segment is the corresponding sequential input data. In this embodiment, the size of the segment length is mainly set according to the actual hardware situation, and does not need to be adjusted according to the battery operation cycle or data, mainly depending on the training hardware resources. However, in this embodiment, the data length is greater than 400. At this time, the data length mainly means that the number of charging data frames and the number of discharging data frames should be the same. For example, if there are 200 charging data frames, then there should also be 200 discharging data frames, and the total data length is 400.

[0065] S4. According to the formed sequential input data i2, sort it in ascending order according to the time dimension of the segment, and set the masking rate. Based on the masking rate, zero out any row and column data in the sequential input data i2 until the amount of zeroed data meets the masking rate. At this time, the sequential input data is the corresponding masked data m1; in this embodiment, the range of the masking rate is between 50% and 90%.

[0066] S5. Input the masked data m1 into the encoder of the battery system data generation model to obtain 1 corresponding representation vector z2;

[0067] S6. Determine whether to perform interpolation processing according to the obtained representation vector z2. If so, based on the preset interpolation processing strategy, perform interpolation processing on the representation vector, and use the interpolated representation vector z2 ′ as the input data and input it into the reconstruction decoder of the battery system data generation model to output the corresponding reconstructed generated data r2 ′ , otherwise, directly input the representation vector into the reconstruction decoder to output the corresponding reconstructed generated data r2.

[0068] The preset interpolation processing strategy is:

[0069] According to the obtained representation vector, determine the vector dimension corresponding to the representation vector and judge the relationship between the vector dimension and the first preset threshold and the second preset threshold;

[0070] If the vector dimension is less than the first preset threshold, the corresponding interpolation processing rule is:

[0071] z2 ′ =λ×z2+(1 - λ)×z i

[0072] In the formula, z2 ′ is the interpolated representation vector, z2 ′ ∈C, z2∈C, C is the category of the vector, λ is the interpolation factor, 0≤λ≤1; z iThe reference point vector corresponding to the reference point i randomly selected in the feature space of category C, z i ∈ C;

[0073] If the vector dimension is greater than or equal to a preset second threshold, the corresponding interpolation processing rule is:

[0074]

[0075] In the formula, N is the number of reference points randomly selected in the feature space of category C, and μ i is the weight value of the i-th reference point.

[0076] The above are only embodiments of the present invention. Common knowledge such as specific structures and characteristics in the solution is described in too much detail here. Those of ordinary skill in the art know all the common general technical knowledge in the technical field to which the invention belongs before the application date or the priority date, can know all the prior arts in this field, and have the ability to apply the conventional experimental means before this date. Those of ordinary skill in the art can, under the inspiration given in this application, complete and implement this solution in combination with their own abilities. Some typical well-known structures or well-known methods should not become obstacles for those of ordinary skill in the art to implement this application. It should be noted that for those skilled in the art, without departing from the structure of the present invention, several deformations and improvements can still be made, and these should also be regarded as the protection scope of the present invention, and these will not affect the implementation effect of the present invention and the practicality of the patent. The protection scope required by this application should be based on the content of its claims, and the specific implementation manners and the like recorded in the specification can be used to explain the content of the claims.

Claims

1. A method for generating battery system operation data based on a reconstructed pre-training architecture, characterized in that: The following steps are involved: S1. Retrieving a pre-trained battery system data generation model from a database, and fixing the network structure and weight parameters of the encoder in the battery system data generation model; the weight parameters include weight parameters of the encoder and reconstruction decoder; S2. Generate an encoder in the model based on the retrieved pre-trained battery system data and output the corresponding virtual battery system operation data; S3. Determine the charge and discharge data in the virtual battery system operation data based on the generated virtual battery system operation data, and slice the charge and discharge data according to the time dimension based on a preset data slicing strategy to form corresponding time series input data i2; S4. Arrange the formed time series input data i2 in ascending order according to the time dimension of the segments, set a masking rate, and set any row and column data in the time series input data i2 to zero based on the masking rate until the amount of zeroed data satisfies the masking rate. The time series input data at this point becomes the corresponding masking data m1. S5. Input the mask data m1 into the encoder of the battery system data generation model to obtain a corresponding representation vector z2; S6. Determine whether to perform interpolation processing based on the obtained representation vector z2. If so, perform interpolation processing on the representation vector based on a preset interpolation processing strategy, and use the interpolated representation vector z2 ′ as input data and input it into the reconstruction decoder of the battery system data generation model to output the corresponding reconstructed generated data r2 ′ , otherwise, directly input the representation vector into the reconstruction decoder to output the corresponding reconstructed generated data r2.

2. A method for generating battery system operation data based on a reconstructed pre-training architecture according to claim 1, characterized in that: The preset data slicing strategy is: According to the determined charge and discharge data, based on the preset data selection dimension, the charge and discharge data corresponding to the corresponding data selection dimension is selected from the charge and discharge data, and abnormal characters and invalid data are eliminated; Set a slice length value, and according to the corresponding charge and discharge data and the slice length value, divide the charge and discharge data into charge and discharge data segments with the same slice length value. The charge and discharge data segments are the corresponding time series input data.

3. A method for generating battery system operation data based on a reconstructed pre-training architecture according to claim 2, characterized in that: The data selection dimensions include a time dimension, a cell voltage sequence dimension, a current dimension, a SOC dimension, and a temperature dimension.

4. A method for generating battery system operation data based on a reconstructed pre-training architecture according to claim 3, characterized in that: The training strategy corresponding to the pre-trained battery system data generation model in S1 is: Step 1: retrieve historical real vehicle operation data from the vehicle's cloud data platform, determine the SOC difference corresponding to each data in the historical real vehicle operation data, and select the historical charge and discharge data corresponding to the preset SOC threshold; Step 2: Based on the corresponding historical charge and discharge data and the corresponding data slicing strategy, the historical charge and discharge data is sliced and processed, and the historical input time series data corresponding to the historical charge and discharge data is output; Step 3: construct a battery system data generation model, wherein the battery system data generation model includes an encoder and a reconstruction decoder; Step 4: Arrange the generated historical input time series data in ascending order according to the time dimension, randomly select a preset proportion of row indexes, and replace the column data of the corresponding rows with 0 to form the corresponding historical zero-set data; Step 5: Input the historical zero-set data into the encoder in the battery system data generation model to obtain a corresponding representation vector; Step 6: Transfer the feature vector to the reconstruction decoder to output the corresponding reconstructed data. Then, calculate the reconstruction error between the reconstructed data and the feature vector based on the reconstructed data and the feature vector. Determine whether the reconstruction error is less than the preset error threshold. If so, it is determined that the battery system data generation model is completed. Otherwise, adjust the weight parameters of the encoder and the reconstruction decoder in the battery system data generation model according to the reconstruction error, and re-execute Step 4.

5. A method for generating battery system operation data based on a reconstructed pre-training architecture according to claim 4, characterized in that: The preset interpolation processing strategy is as follows: Based on the obtained feature vector, determine the vector dimension corresponding to the feature vector and judge the relationship between the vector dimension and the first preset threshold and the second preset threshold; If the vector dimension is less than the first preset threshold, the corresponding interpolation processing rule is: z2 ′ = λ × z2 + (1 - λ) × z i where z2 ′ is the representation vector after interpolation processing, z2 ′ ∈C, z2 ∈ C, C is the category of the vector, λ is the interpolation factor, 0 ≤ λ ≤ 1; z i is the reference point vector corresponding to the reference point i randomly selected in the feature space of category C, z i ∈C; If the vector dimension is greater than or equal to the preset second threshold, the corresponding interpolation processing rule is: where N is the number of reference points randomly selected in the feature space of class C, and μ i is the weight value of the i-th reference point.

Images