New energy automobile battery missing signal interpolation method and device

Abstract

The invention provides a new energy automobile battery missing signal interpolation method and equipment. The method comprises the steps of data acquisition and preprocessing, extraction of time features from data collected by a battery, standardization processing of numerical value features, one-hot coding of classification features, and determination of an interpolation data range by using a statistical test method. Model construction: carrying out feature extraction and coding by adopting a large language model, and constructing forward and reverse data groups; model training: training the data by using a deep learning model, evaluating the performance of the model by using a mean square error and a cross entropy loss function, and performing parameter optimization; interpolation operation: performing interpolation on the missing time period by using the forward and reverse models, and fusing results to obtain a final interpolation result; and result verification: verifying the validity and reliability of the interpolation data through statistical indexes and actual application. The scheme has good accuracy and data adaptability and high data processing efficiency.

Description

Technical Field

[0001] This invention relates to the field of new energy vehicle battery data processing technology, specifically to a method and device for interpolating missing signals in new energy vehicle batteries. Background Technology

[0002] With the rapid development of new energy vehicles, the performance and safety of power batteries, as core components, are receiving increasing attention. Batteries are highly sensitive to voltage and temperature: excessively high or low voltage can lead to overcharging or over-discharging, damaging battery performance and lifespan; abnormal temperatures can trigger safety accidents such as thermal runaway. Therefore, battery safety monitoring systems have emerged to effectively manage and control battery packs by monitoring the voltage and temperature of individual cells in real time. By balancing battery state and optimizing charging and discharging strategies, not only can thermal runaway and overcharging / over-discharging be prevented, ensuring safe vehicle operation, but battery life can also be extended. Furthermore, accurately estimating battery state and maintaining optimal operating temperature can improve battery performance, optimize vehicle energy management, increase driving range, and reduce energy consumption.

[0003] However, due to the long transmission link of battery signal data from the vehicle to the battery safety monitoring system, and its susceptibility to communication signal interference, battery signal uploads are often lost, resulting in data frame drops and quality degradation. In voltage monitoring, missing time-series data disrupts data continuity, affecting the accurate assessment of voltage change trends (such as voltage fluctuations during charging or driving). Reduced sample size introduces errors in calculating voltage statistics, potentially leading to incorrect assessments of battery health. In temperature monitoring, missing data results in incomplete temperature records, making it difficult to monitor temperature trends in real time and potentially delaying thermal runaway warnings. This also hinders the evaluation of heat transfer and dissipation within the battery pack, reducing temperature management efficiency. For the entire battery safety monitoring system, missing data not only increases processing complexity but also requires additional algorithms to handle missing values. This can introduce new errors, reducing the reliability of battery status assessments, causing false alarms or missed alarms, ultimately impacting vehicle safety and battery lifespan.

[0004] For the reasons mentioned above, it is necessary to process missing values ​​in the time-series data generated by the battery signal in order to improve the monitoring quality. Summary of the Invention

[0005] The purpose of this invention is to provide a missing signal interpolation scheme based on new energy vehicle batteries. This scheme utilizes a self-attention mechanism to simultaneously focus on different positions in the input sequence, capturing long-term dependencies and global features in the time series. This scheme can better capture the global features of the time series and is particularly suitable for processing time series data with complex patterns and long-term trends, thereby solving the problems mentioned in the background art.

[0006] Specifically, the present invention provides the following technical solutions: Method and steps: Phase 1: Data Acquisition and Preprocessing S1. Data Acquisition: Extract time information, such as hours and days, from the "cumulative driving time" as time features. The time step corresponding to each data point can be calculated in hours. Acquire time-series data with missing values ​​collected from the new energy vehicle battery; the time-series data with missing values ​​is data obtained at equal time intervals with a certain sampling frequency, and the data contains missing time periods.

[0007] S2: Data Standardization. The continuous numerical characteristics (average vehicle speed, battery cell temperature, battery cell voltage) are standardized using the following formula:

[0008] Where x is the original data, μ is the mean of the feature data, and σ is the standard deviation of the feature data. The standardized data will have zero mean and unit variance, which helps to improve the efficiency and performance of model training.

[0009] The classification feature of "battery type" is encoded using one-hot encoding. Given n battery types, each battery type will be represented as a binary vector of length n.

[0010] S3: Determine the range of data to be imputed. Using statistical tests, such as the Mann-Kendall test, determine the data to be imputed based on the missing time period. This test can identify trend inflection points in the data, thus allowing for the reasonable selection of valid data segments before and after the missing values ​​as the basis for imputation.

[0011] Phase Two: Model Building S4: Data Feature Extraction and Encoding. Feature selection is performed, with "cumulative driving time," "average vehicle speed," "battery cell temperature," "battery cell voltage," and "battery charge / discharge state" used as the main feature inputs to the time series model. "Battery charge / discharge state" can be considered a binary feature (0 indicates charging, 1 indicates discharging). "Battery type" is used as a static classification feature input to the model, and "battery manufacturing date" is used as an auxiliary time feature input (e.g., it can be converted to battery lifespan).

[0012] The data interpolation model comprises an encoder and a decoder. The encoder receives features and temporal information from past time steps, while the decoder predicts feature values ​​for future time steps. The data to be interpolated is then processed using a Transformer model for feature extraction and encoding. The encoder portion of the Transformer model extracts features from the data through a self-attention mechanism, calculated using the following formula:

[0013] Where Q is the query matrix, K is the key matrix, and V is the value matrix. It is the key dimension. This mechanism enables the model to focus on the correlations between different time steps in the data and capture long-term dependencies.

[0014] During feature extraction, more types of features are integrated, including static features of the battery: battery type and manufacturing date; and dynamic features of the battery's charge and discharge states.

[0015] Preferably, the data interpolation model structure is as follows: the interpolated data first enters the feature embedding layer, and then is connected to the self-attention layer; the output data of the self-attention layer is sent to the first residual connection and layer normalization layer, while the input data of the self-attention layer is directly sent to the first residual connection and layer normalization layer; the output data of the residual connection and layer normalization layer is sent to the feedforward neural network, and the output data of the feedforward neural network is again connected to the second residual connection and layer normalization layer, while the input data of the feedforward neural network is sent to the second residual connection and layer normalization layer; the output data of the second residual connection and layer normalization layer serves as the feature sequence.

[0016] Preferably, the imputation data is formed as follows: static categorical features are mapped to fixed-dimensional vectors through a feature embedding layer; temporal features are mapped to continuous vectors through a feature embedding layer; and sinusoidal position encoding is used to add time step order to the output data of the feature embedding layer to obtain the feature embedding vector. Dynamic numerical features are not input into the feature embedding layer; they are directly concatenated with the feature embedding vector and then input into the self-attention layer.

[0017] S5: Data Set Construction. Based on the preset time step, the obtained feature sequences are divided into forward and reverse data sets in the forward and reverse time directions, respectively. This step aims to capture data patterns from different time perspectives, providing a comprehensive data perspective for model training.

[0018] Phase 3: Model Training S6: Model Training. Using deep learning models, such as Long Short-Term Memory (LSTM) networks or Transformer variants, the forward and backward data sets for each component are divided into training and validation sets according to a predefined ratio. The forward and backward models are then trained according to the set training parameters. During training, the Mean Squared Error (MSE) is typically used as the loss function to evaluate model performance. The calculation formula is as follows:

[0019] in, It is the actual value. It is a predicted value. This refers to the number of samples. By optimizing this loss function, the model can learn patterns and regularities in the data to minimize the difference between the predicted and actual values.

[0020] For classification targets (such as battery charge / discharge state), the cross-entropy (CE) loss function can be used. For the predicted probability p and the true label y, the cross-entropy loss is...

[0021] In the above formula, Indicates the time point.

[0022] Next, we will continue to optimize the model's parameters. In this embodiment, the preferred parameter update method is as follows: First-order moment estimation:

[0023] Second-order moment estimation:

[0024] First moment of bias correction:

[0025] The second moment of bias correction:

[0026] Parameter update:

[0027] in, and These are parameters First-order moment estimates and second-order moment estimates and It is the attenuation factor. It's the learning rate. It is a constant used for numerical stability.

[0028] Fourth stage: Interpolation operation S7: Data Imputation. The missing time intervals are interpolated one by one using both forward and backward models. The interpolated data is added to the time vector for the next interpolation step, until the entire missing time interval is imputed. During the interpolation process, a value for each time step is generated and added to the time vector for use in the next time step, ensuring the continuity and consistency of the interpolation.

[0029] Forward model: predictive interpolation is performed from the past to the future time direction; Inverse model: predictive interpolation is performed from the future to the past.

[0030] S8: Result Fusion. The results from the forward model and the backward model are fused to obtain the final interpolation result. The preferred fusion method is:

[0031] in, This is the final interpolation result. It is the interpolation result of the forward model. This is the interpolation result of the inverse model, where α is the fusion weight, ranging from [0,1]. By adjusting the value of α, the contribution ratio of the forward and inverse model results in the final interpolation can be controlled, thus obtaining a more accurate interpolation result.

[0032] Phase 5: Result Verification S9. Verify the interpolated data results, use the model to predict the feature values ​​of future time steps, and output the prediction results.

[0033] Statistical evaluation: The prediction results are evaluated by calculating the MASE and sMAPE indices to assess the model's predictive interpolation performance.

[0034] Practical application verification: The interpolated data was applied to the battery management system of new energy vehicles to verify its effectiveness and reliability in performance evaluation and fault prediction tasks.

[0035] The present invention also provides a new energy vehicle battery missing signal interpolation device, which includes a processor and a memory. The processor calls computer instructions in the memory to execute the new energy vehicle battery missing signal interpolation method as described above.

[0036] Compared with existing technologies, this solution has at least the following advantages: High efficiency: By setting up a large language model, this solution enhances parallel computing capabilities, enabling rapid processing of long sequence data and significantly improving interpolation efficiency.

[0037] Accuracy: By capturing long-term dependencies and feature correlations in time series data, this model can generate more accurate imputation values ​​and better reflect the true changing patterns of the data.

[0038] Adaptability: This solution is suitable for various types of time series data, and performs particularly well when dealing with high-dimensional, long-series data, demonstrating strong adaptability and versatility.

[0039] Robustness: Combining the interpolation results of forward and backward models can effectively reduce the bias that may be caused by a single model, and improve the stability and robustness of the interpolation results. Attached Figure Description

[0040] 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 some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 This is a flowchart of the interpolation method according to an embodiment of the present invention.

[0042] Figure 2 This is a flowchart of the model processing in an embodiment of the present invention.

[0043] Figure 3 The diagram shows the structure of the data interpolation model and illustrates the processing steps in step S4. Detailed Implementation

[0044] The method of the present invention will be described in detail below through specific embodiments: Implementation scenario: In this embodiment, there is a set of battery signal time series data of new energy vehicles. This dataset contains the following characteristics: Cumulative driving time (in hours); average vehicle speed (in kilometers per hour); battery cell temperature (in degrees Celsius); battery cell voltage (in volts); battery type; battery manufacturing date; battery charge / discharge status.

[0045] The data sampling frequency is set to, for example, once per minute, but due to sensor malfunction, some data is missing for a certain period of time. The goal of this embodiment is to use the method of the present invention to interpolate this missing data.

[0046] Combination Figure 1 , Figure 2 As shown, the specific steps of this embodiment can be implemented in the following manner: Phase 1: Data Acquisition and Preprocessing S1. Collect voltage, current, and temperature data of multiple battery packs in new energy vehicles during actual operation, with a sampling frequency of once every 10 seconds; clean the collected data to remove outliers and missing values; divide the processed data into training, validation, and test sets according to a ratio of 70%, 15%, and 15%, respectively. The time series data collected above is time series data with missing values, obtained at equal time intervals with a certain sampling frequency, and there are missing time periods in the data.

[0047] S2. First, clean the data to remove obvious outliers and noise.

[0048] Secondly, the data is standardized by calculating the mean and standard deviation of each feature, and converting the data to a form with zero mean and unit variance. The standardization formula is:

[0049] in, x It is the raw data. μ It is the mean of the data. σ This is the standard deviation of the data. Standardized data will have zero mean and unit variance, which helps improve the efficiency and performance of model training.

[0050] In this stage, the dataset needs to be processed in batches, and the process is as follows: A) Preprocess and integrate the training data. These datasets include data upload time, vehicle speed, battery cell temperature, and vehicle status, etc. Then, integrate this data with different characteristics into a unified dataset, ensuring that each point in time in the dataset contains values ​​of all the aforementioned features. This provides a complete data foundation for subsequent model training. B) Constructing Input Features for the Training Sample Dataset. The processed data is used to construct the input features for the model. The input at each time point contains multiple features, covering state duration, average vehicle speed, battery cell temperature, battery voltage, etc. These features comprehensively reflect the vehicle's operating status and battery-related indicators at various time points, providing sufficient input information for subsequent model prediction and other tasks.

[0051] S3. Perform the Mann-Kendall test (MK test) on the data before and after the missing time period to identify the trend abrupt change point, so as to reasonably select the effective data segment before and after the missing value as the basis for imputation.

[0052] Select the effective data segment based on the trend inflection point, and then select the preceding and following data segments to participate in the interpolation based on the effective data segment.

[0053] In this embodiment, the preferred method for detecting trend abrupt change points is as follows: (1) Perform MK test on the data windows of 200 time points before and after the missing time period; (2) If the Z value of the MK test statistic changes significantly at a certain time point, then the point can be regarded as a trend change point. In this embodiment, |Z|>1.96 is taken, which means that it is significant at the 95% confidence level. Points that meet this condition are selected as change points.

[0054] In a more preferred embodiment, the selection method for the preceding and following data segments involved in interpolation is as follows: After identifying the trend abrupt change point, the preceding and following data segments for interpolation are automatically selected according to the following rules: (1) Data segment length setting: Set a minimum data length threshold. (For example, 200 time points) to ensure the model has sufficient contextual information for imputation; set a maximum data length threshold. To avoid introducing too much irrelevant information.

[0055] (2) Data segment selection method: The preceding data is searched backward from the starting time point of the missing segment to find the first trend inflection point. If the time length between this inflection point and the starting time point of the missing segment is within [ , If the data falls within the specified range, that segment is selected as the preceding data segment. If no mutation point is detected, the segment preceding the missing segment is selected by default. The first time point is used as the first data segment; the second data segment is selected by searching backwards from the end time point of the missing segment to find the first trend inflection point. If the time length between this inflection point and the end time point of the missing segment is within […]. , If the data falls within the specified range, that segment will be selected as the subsequent data segment. If no mutation point is detected, the segment following the missing segment will be selected by default. Each time point is used as subsequent data.

[0056] Phase Two: Model Building S4. In this step, data feature extraction and encoding are performed, and the preprocessed data is input into the data interpolation model. Feature selection is conducted, with "cumulative driving time," "average vehicle speed," "battery cell temperature," "battery cell voltage," and "battery charge / discharge state" as the main time series features input into the model. Among them, "battery charge / discharge state" can be regarded as a binary feature (0 indicates charging, 1 indicates discharging); "battery type" is input into the model as a static classification feature, and "battery manufacturing date" is used as an auxiliary time feature input (for example, it can be converted into battery life).

[0057] The model comprises an encoder and a decoder. The encoder receives features and temporal information from past time steps, while the decoder predicts feature values ​​for future time steps. The encoder portion of the model extracts features from the data using a self-attention mechanism, generating a high-dimensional feature representation. The self-attention mechanism is calculated using the following formula:

[0058] in, Q It is a query matrix. K It is a key matrix. V It is a value matrix. It is the dimension of the key.

[0059] The preprocessed data is encoded using time series coding techniques to extract features such as trends and periodicity from the time series. Training samples are generated using a sliding window with a window size of 100, meaning each sample contains 100 historical data points and the corresponding target value (the battery signal value at the 101st time point).

[0060] Feature extraction is primarily based on the model's self-attention mechanism: ① Map the numerical features in the time series to a high-dimensional space through linear transformation; ② Add position codes to indicate the order of time steps; ③ Embed time features into a high-dimensional space; ④ Calculate the query (Q), key (K), and value (V), and extract features using the self-attention formula; then perform a nonlinear transformation on the output of the self-attention through a feedforward neural network, and apply residual connections and layer normalization after each sub-layer to stabilize the training process.

[0061] The encoder processing procedure in this part is as follows: Figure 3 As shown, it is used to extract time series features, capture the dependencies between signals such as voltage and temperature, and provide a high-dimensional representation for subsequent interpolation.

[0062] In this embodiment, the model structure is as follows: The input features first enter the feature embedding layer, which performs steps ①-④ above to embed various types of feature data into a high-dimensional space and embed temporal codes; then it is connected to the self-attention layer; the output data of the self-attention layer is sent to the first residual connection and layer normalization layer, and at the same time, the input data of the self-attention layer is directly sent to the first residual connection and layer normalization layer; the output data of the residual connection and layer normalization layer is sent to the feedforward neural network, and the output data of the feedforward neural network is again connected to the second residual connection and layer normalization layer, and at the same time, the input data of the feedforward neural network is sent to the second residual connection and layer normalization layer; the output data of the second residual connection and layer normalization layer is used as the output features.

[0063] In the model of this embodiment, feature embedding is implemented by an encoder: static category features, such as battery type, are mapped to fixed-dimensional vectors through feature embedding; time features, such as month and hour, are mapped to continuous vectors through a time embedding layer or sinusoidal position encoding; and the time step position is represented by sinusoidal position encoding to indicate the order of time steps.

[0064] In this embodiment, dynamic numerical features such as voltage and temperature are continuous variables and are directly input into the model without feature embedding processing. All feature embedding vectors are concatenated with the dynamic numerical features and used as input to the Transformer encoder (i.e., the self-attention layer).

[0065] In this embodiment, more types of features can be integrated during the feature extraction process, including static features of the battery such as battery type and manufacturing date, and dynamic numerical features such as battery charge and discharge status.

[0066] S5. Based on the preset time step, the obtained feature sequence is divided into forward and reverse data sets in the forward and reverse time directions, respectively. This step aims to capture data patterns from different time perspectives, providing a comprehensive data perspective for model training.

[0067] In this embodiment, the time step is set to 12 (i.e., 12 minutes), and the feature sequence is processed into forward data groups and reverse data groups respectively. The forward data groups are arranged in chronological order, while the reverse data groups are arranged in reverse chronological order.

[0068] Input data format: ① Time series data: The input data for the inverse model is a reverse arrangement of time series data. The forward series is [ , , , ..., The reverse sequence is [] , , , ..., The data format is [sequence length, number of features].

[0069] ② Time features: The time features are arranged in reverse order to match the reversed time series data. The data format is [sequence length, number of time features].

[0070] ③ Static features: Static features are the same as those in the forward model, and the data format is [number of static features].

[0071] ④ Dynamic Features: Dynamic features are arranged in reverse order, and the data format is [sequence length, number of dynamic features]. Phase 3: Model Training S6. Divide the forward data set and the reverse data set into a training set (80%) and a validation set (20%).

[0072] The deep learning model was trained using the Adam optimization algorithm, with a learning rate of 0.001, a batch size of 32, and 100 training epochs. During model training, the mean squared error (MSE) was used as the loss function, calculated using the following formula:

[0073] in, It is the actual value. It is a predicted value. It refers to the number of samples.

[0074] For classification targets (such as battery charge / discharge state), the cross-entropy (CE) loss function can be used. For the predicted probability p and the true label y, the cross-entropy loss is...

[0075] In the above formula, This indicates the number of data points.

[0076] Next, we will optimize the model parameters. The preferred parameter update method is as follows: First-order moment estimation:

[0077] Second-order moment estimation:

[0078] First moment of bias correction:

[0079] The second moment of bias correction:

[0080] Parameter update:

[0081] in, and These are parameters First and second moment estimates, and It is the attenuation factor. It's the learning rate. It is a constant used for numerical stability.

[0082] During training, the loss function of the validation set is monitored to prevent overfitting.

[0083] Fourth stage: Interpolation operation S7. Using both forward and backward models, interpolation is performed on the missing time periods one by one. The interpolated data is added to the time vector at the next interpolation time, until the entire missing time period is interpolated. During the interpolation process, a value for each time step is generated and added to the time vector for interpolation at the next time step, ensuring the continuity and consistency of the interpolation.

[0084] The missing time period is interpolated using the trained forward and backward models. During the interpolation process, the data obtained from each interpolation is added to the time vector of the next interpolation moment to ensure the continuity and consistency of the interpolation, until the entire missing time period is completely interpolated.

[0085] S8. Interpolation results obtained from the forward model Interpolation results obtained from the inverse model Perform weighted average fusion. The final interpolation result is represented by the formula:

[0086] Here, α is the fusion weight, with a value range of [0, 1]. The final result is the complete time-series data interpolation, which is then inversely normalized back to the original data range.

[0087] Phase 5: Result Verification S9. Verify the interpolated data results, use the model to predict the feature values ​​of future time steps, and output the prediction results.

[0088] Statistical evaluation: The prediction results are evaluated by calculating the MASE and sMAPE indices to assess the model's predictive interpolation performance.

[0089] Practical application verification: The interpolated data was applied to the battery management system of new energy vehicles to verify its effectiveness and reliability in performance evaluation and fault prediction tasks.

[0090] The solution provided by this invention, by combining the powerful feature extraction capabilities of a model with the efficient fitting capabilities of a deep learning model, achieves accurate and efficient imputation of missing values ​​in high-dimensional, long-sequence time series data. Compared with traditional methods, this solution has significant advantages in processing complex time series data, providing reliable data support for battery management, performance evaluation, and fault prediction of new energy vehicles. Experimental results show that this method performs excellently on multiple evaluation metrics, demonstrating broad application prospects and practical value.

[0091] In another embodiment, the present invention also provides a new energy vehicle battery missing signal interpolation device, the device including a processor and a memory, wherein the processor calls computer instructions in the memory to execute the new energy vehicle battery missing signal interpolation method provided in the above embodiment.

[0092] This specification can be described in the general context of computer-executable instructions that are executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform a specific task or implement a specific abstract data type. This specification can also be practiced in distributed computing environments, where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside on local and remote computer storage media, including storage devices.

[0093] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0094] The above description is merely an embodiment of this specification and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of this application should be included within the scope of the claims of this application.

Claims

1. A new energy vehicle battery absence signal interpolation method, characterized in that, The method comprises: S1, acquiring time series data with missing values collected by a new energy vehicle battery; S2, preprocessing the time series data to obtain processed time series data; S3, detecting trend mutation points in the processed time series data, selecting effective data segments based on the trend mutation points, and determining interpolation data of missing time period data according to the effective data segments; S4, extracting and encoding features of the interpolation data through a data interpolation model to form a feature sequence; S5, based on a preset time step, obtaining a forward data group and a reverse data group in a forward time direction and a reverse time direction respectively according to the feature sequence; S6, training a forward model and a reverse model respectively by using the forward data group and the reverse data group; S7, interpolating the missing time period data by using the forward model and the reverse model respectively to obtain interpolation data, wherein the interpolation data obtained this time is added to the processed time series data as a data source for calculating interpolation data next time until interpolation of all missing time period data is completed; S8, fusing the interpolation data obtained by the forward model and the reverse model to obtain final interpolation data.

2. The method of claim 1, wherein, The preprocessing in S2 comprises: data cleaning to remove outliers and noise points; The cleaned data is standardized.

3. The method of claim 1, wherein, The S3 further comprises: verifying the data before and after the missing time period in the processed time series data to determine the trend mutation points; After determining the trend mutation points, the front segment data and the rear segment data participating in interpolation are determined within a predetermined range of each trend mutation point, thereby forming the interpolation data.

4. The method of claim 1, wherein, In S6, the forward model and the reverse model are constructed based on the same deep learning model; in the training process, the mean square error is used as the loss function: wherein, is the true value, is the predicted value, is the number of samples.

5. The method of claim 1, wherein, In S8, the specific fusion method is: Wherein, α is the fusion weight, and the value range is [0, 1]; the complete time series data interpolation result is finally obtained, and it is converted back to the original data range through inverse standardization.

6. The method of claim 1, wherein, After S8, it further comprises: S9, verifying the model interpolation result, and using the root mean square error, the mean absolute error and the determination coefficient between the interpolation data and the true data to quantitatively evaluate the interpolation effect.

7. The method of claim 3, wherein, The determination method of the front segment data and the rear segment data participating in interpolation is: A maximum data length threshold of a preset data segment and a minimum data length threshold ; The data length is in units of time points; searching forward from the start time point of the missing time period, a first trend mutation point is found, if the data length between the trend mutation point and the start time point of the missing time period is within the range of [ , ], the segment of data is selected as the front segment of data; if no mutation point is detected, the data of the start time point of the missing time period and the previous time points are taken as the front segment of data;​ searching backward from the end time point of the missing time period, a first trend mutation point is found, if the data length between the trend mutation point and the end time point of the missing time period is within the range of [ , ], the segment of data is selected as the later segment of data; if no mutation point is detected, the data of the time point behind the end time point of the missing time period by is taken as the later segment of data.​​ 8. The method of claim 1, wherein, The data interpolation model structure is as follows: The interpolation data first enters the feature embedding layer, and then connects the self-attention layer; the output data of the self-attention layer is sent to the first residual connection and the layer normalization layer, and at the same time, the input data of the self-attention layer is directly sent to the first residual connection and the layer normalization layer; the output data of the residual connection and the layer normalization layer is sent to the feedforward neural network, and the output data of the feedforward neural network is connected to the second residual connection and the layer normalization layer again, and at the same time, the input data of the feedforward neural network is sent to the second residual connection and the layer normalization layer; the output data of the second residual connection and the layer normalization layer is the feature sequence.

9. The method of claim 8, wherein, The feature embedding layer is realized by an encoder; The participation in the interpolation data formation manner is that: the static category features are mapped into fixed dimension vectors through a feature embedding layer; the time features are mapped into continuous vectors through a feature embedding layer; and the sine position coding manner is that the time step sequence is added to the output data of the feature embedding layer to obtain the feature embedding vector. The dynamic numerical features are not input into the feature embedding layer, but are directly spliced with the feature embedding vector and then input into the self-attention layer.

10. A new energy vehicle battery absence signal interpolation device, characterized in that, The device comprises a processor and a memory, and the processor invokes computer instructions in the memory to execute the new energy vehicle battery missing signal interpolation method in any one of claims 1-9.

Images