New energy automobile performance index evaluation method

Abstract

The invention relates to the technical field of new energy automobiles, and particularly provides a new energy automobile performance index evaluation method which comprises the following steps: calculating a probability density function of each parameter through a data distribution extraction technology according to a multi-dimensional alignment data set to obtain a parameter distribution feature set; according to the dimension reduction feature vector set, potential abnormal distribution features are generated by using a variational auto-encoder through a hidden variable extraction technology, and an abnormal distribution feature set is obtained; according to the abnormal distribution feature set, an abnormal sensitivity technology is adopted to detect distribution offset, and an abnormal offset score is obtained; according to an abnormal time period positioning result, calculating high-temperature duration to obtain defect duration data; querying associated thermal diffusion coefficients and historical data features in the knowledge graph according to defect types in the early warning information to obtain a supplementary acquisition parameter suggestion list; and according to the refined defect identification result, analyzing the thermal state of the battery through a thermal load distribution technology, and judging whether a final early warning report containing thermal characteristics is generated or not.

Description

Technical Field

[0001] The present invention relates to the technical field of new energy vehicles, and in particular to a method for evaluating performance indicators of new energy vehicles. Background Art

[0002] As a key research area, new energy vehicle performance evaluation is directly related to product quality, safety, and market competitiveness, and its importance is self-evident. With the rapid development of the new energy vehicle industry, performance evaluation is not only essential for verifying technical reliability but also a core support for promoting industrial upgrades and ensuring user safety. However, this field still faces numerous challenges in meeting increasingly complex testing requirements, and breakthroughs are urgently needed.

[0003] Currently, traditional evaluation methods rely primarily on manual inspection and simple rule-based analysis, which are inadequate for processing massive amounts of real-time data. Manual inspection struggles to quickly capture complex correlations between parameters and can easily overlook hidden abnormal patterns, leading to delayed defect discovery and missed opportunities for optimal action. This inefficiency and limitations no longer meet the high standards required for modern new energy vehicle evaluation.

[0004] Against this backdrop, the core challenges faced during the evaluation process have become increasingly prominent. The most prominent technical factors include modeling complex correlations between parameters, accurately identifying anomaly patterns, and systematically expressing accumulated knowledge. Because parameters such as battery temperature, voltage, and current dynamically interact with each other, traditional methods struggle to accurately capture their inherent patterns. Furthermore, the diversity and hidden nature of anomaly patterns further complicate identification. The lack of effective knowledge graph support makes it difficult to translate experience into reusable analytical insights. These unresolved technical issues directly contribute to inefficient defect identification and insufficient early warning capabilities.

[0005] Therefore, building parameter correlation models within massive amounts of real-time evaluation data to reveal key influencing relationships, mine diverse and hidden anomaly patterns, and intelligently identify and warn of defect characteristics through knowledge graphs has become a key challenge in improving the efficiency and accuracy of new energy vehicle performance evaluation. Solving this problem requires not only algorithmic innovation but also higher requirements for collaborative design of system architectures. Summary of the Invention

[0006] The present invention provides a new energy vehicle performance index evaluation method, which mainly includes:

[0007] The battery temperature, voltage, and current data streams are collected in real time by sensors, and the parameters are subjected to sliding window filtering using time series smoothing technology to obtain a smoothed time series data set.

[0008] Based on the smoothed time series data set, the temperature, voltage, and current timestamps are synchronized and aligned using data consistency technology to obtain a multi-dimensional aligned data set.

[0009] According to the multi-dimensional aligned data set, the probability density function of each parameter is calculated by data distribution extraction technology to obtain the parameter distribution feature set;

[0010] According to the parameter distribution feature set, the principal component analysis method is used to perform linear dimension reduction on temperature, voltage and current to obtain a dimension reduction feature vector set reflecting the temperature gradient and heat load distribution.

[0011] According to the reduced dimension feature vector set, the latent variable extraction technology is used to generate potential abnormal distribution features using the variational autoencoder to obtain the abnormal distribution feature set;

[0012] Based on the abnormal distribution feature set, the abnormal sensitivity technology is used to detect the distribution shift and obtain the abnormal shift score;

[0013] If the anomaly offset score exceeds the preset thermal anomaly threshold, the slope before and after the anomaly timestamp is calculated using the temperature mutation rate technology to obtain the anomaly time period positioning result;

[0014] Based on the abnormal time period positioning results, the high temperature duration is calculated using the heat accumulation time technology to obtain the defect duration data;

[0015] Based on the defect duration data, the thermal correlation coefficient is matched with the known defect patterns in combination with the pre-established knowledge graph to obtain the defect classification label and severity score.

[0016] If the severity score exceeds the safety threshold determined based on the uniformity of heat distribution, a warning message containing the defect type and time period is generated through the rule engine;

[0017] Based on the defect type in the warning information, the associated thermal diffusion coefficient and historical data features in the knowledge graph are queried to obtain a list of recommended supplementary acquisition parameters;

[0018] Adjust the sensor acquisition range according to the recommended list of supplementary acquisition parameters, obtain new real-time data streams, and update the reduced-dimensional feature vector set using feature compression ratio technology to obtain refined defect recognition results;

[0019] Based on the refined defect identification results, the battery thermal status is analyzed using thermal load distribution technology to determine whether a final warning report containing thermal characteristics should be generated.

[0020] Optionally, the real-time data streams of battery temperature, voltage, and current are collected by sensors, and sliding window filtering is performed on each parameter using a time series smoothing technique to obtain a smoothed time series data set, including:

[0021] The battery temperature, voltage, and current data are collected in real time through sensors to generate a data stream containing a time series and obtain the original data set;

[0022] The sliding window filtering technology is used to smooth the time series in the original data set to obtain a smoothed time series data set;

[0023] For the smoothed time series data set, the changing trend of each parameter in the sliding window is calculated to obtain the parameter change sequence;

[0024] If the battery temperature in the parameter change sequence exceeds the preset threshold, the abnormal state is determined by the anomaly detection algorithm to obtain an abnormal flag sequence;

[0025] According to the abnormal label sequence, the clustering algorithm is used to group the smooth time series data set to obtain the grouped data set;

[0026] Obtain the abnormal groups in the grouped data set, calculate the statistical characteristics of the parameters in each group, and obtain the characteristic sequence;

[0027] By comparing the characteristic sequence with the preset threshold, the battery status is judged and the status assessment result is obtained.

[0028] Optionally, the step of synchronously aligning the temperature, voltage, and current timestamps using a data consistency technology based on the smoothed time series data set to obtain a multi-dimensional aligned data set includes:

[0029] Through time series analysis, the original timestamp sequence is obtained from the temperature data, voltage data and current data to obtain the initial multidimensional data set;

[0030] Smoothing technology is used to smooth the time series in the initial multidimensional data set to obtain smoothed time series data;

[0031] For smoothed time series data, the differences between timestamps are used to determine the synchronization alignment requirements and obtain the timestamp deviation set;

[0032] Through data consistency technology, the timestamp deviation set is adjusted, and the temperature data, voltage data, and current data are synchronously aligned to obtain a preliminary aligned data set;

[0033] Obtain multidimensional data features from the preliminary aligned dataset, perform cluster analysis using the K-means algorithm, determine data synchronization accuracy, and obtain an optimized aligned dataset;

[0034] Based on the optimized aligned dataset, outliers in the time series are detected. If the outliers exceed the preset threshold, they are repaired through interpolation to obtain the final aligned dataset.

[0035] Extract the synchronized timestamp sequence from the final aligned dataset to generate a multidimensional aligned dataset.

[0036] Optionally, the method of calculating the probability density function of each parameter by using a data distribution extraction technique based on the multi-dimensional aligned data set to obtain a parameter distribution feature set includes:

[0037] By aligning the data set in multiple dimensions, we can obtain the data distribution, calculate the parameter density using extraction techniques, and obtain the probability function.

[0038] For the probability function, calculate the distribution characteristics of each parameter and determine the feature set;

[0039] If the feature set is complete, the multi-dimensional alignment is verified through the dataset source to determine the consistency of the alignment dimensions and obtain the alignment result;

[0040] Based on the alignment results, a computational process is used to analyze the parameter density changes and obtain distribution feature updates;

[0041] By updating the distribution characteristics, the stability of the probability function is judged and the parameter calculation optimization is obtained;

[0042] Optimize parameter calculations, obtain technical application adjustments, and determine the final feature set;

[0043] Through the final feature set, extraction technology is used to analyze data distribution trends and obtain business optimization solutions.

[0044] Optionally, the linear dimension reduction of temperature, voltage, and current is performed using principal component analysis based on the parameter distribution feature set to obtain a dimension reduction feature vector set reflecting the temperature gradient and heat load distribution, including:

[0045] Obtain temperature data, voltage data, and current data through parameter distribution to form an initial feature set;

[0046] The principal component analysis method is used to perform linear dimensionality reduction on the initial feature set to obtain a reduced dimensionality feature vector set;

[0047] Extract the characteristic components corresponding to the temperature gradient from the reduced-dimensional feature vector set to determine the temperature change trend;

[0048] Determine the distribution characteristics of the heat load based on the temperature change trend and obtain the heat load distribution vector;

[0049] If the component of the heat load distribution vector exceeds the preset threshold, the distribution characteristics are adjusted through the vector set to obtain an optimized characteristic set;

[0050] For the optimized feature set, cluster analysis method is used to divide the relevant areas of temperature gradient and heat load to obtain regional distribution results;

[0051] The mapping relationship of dimensionality reduction features is generated through the regional distribution results to determine the final vector set.

[0052] Optionally, the step of generating potential abnormal distribution features using a variational autoencoder through latent variable extraction technology based on the reduced dimension feature vector set to obtain an abnormal distribution feature set includes:

[0053] Generate feature vectors by reducing dimensionality features, and use variational autoencoders to extract latent variable sets from feature vectors to obtain potential distribution data;

[0054] Analyze the abnormal distribution characteristics from the underlying distribution data to obtain a preliminary representation of the abnormal distribution;

[0055] If the initial representation of the abnormal distribution exceeds the preset threshold, the latent distribution is adjusted through the autoencoder to obtain the optimized distribution characteristics;

[0056] Based on the optimized distribution characteristics, vector processing technology is used to generate a feature set to determine the final abnormal distribution feature set;

[0057] Verify the feature set through extraction techniques to determine whether it conforms to the expected characteristics of the underlying distribution;

[0058] Obtain the verified feature set, use the clustering algorithm to group the abnormal distribution features, and obtain the classified feature subset;

[0059] For the classified feature subsets, the distribution patterns of each group of features are analyzed through statistical tools to obtain a detailed description of the abnormal distribution.

[0060] Optionally, the detecting distribution shift using an abnormal sensitivity technique based on the abnormal distribution feature set to obtain an abnormal shift score includes:

[0061] Obtain data input through abnormal distribution feature set, and obtain feature set using preset extraction method;

[0062] Analyze abnormal distribution based on feature set and use statistical methods to obtain distribution characteristics;

[0063] Detect abnormal sensitivity through distribution characteristics and obtain sensitivity analysis results using sensitivity technology;

[0064] The distribution shift is determined based on the sensitivity analysis results. If the shift exceeds a preset threshold, the shift status is determined.

[0065] The distribution shift is analyzed by shift status to obtain the shift score;

[0066] The degree of abnormality is determined based on the deviation score, and the classification algorithm is used to obtain the output result;

[0067] The detection technology is adjusted based on the output results to obtain the optimized offset score.

[0068] Optionally, if the anomaly offset score exceeds a preset thermal anomaly threshold, the slope before and after the anomaly timestamp is calculated using a temperature mutation rate technique to obtain an abnormal time period positioning result, including:

[0069] If the offset score exceeds the preset threshold, the slope before and after the timestamp is calculated using the temperature mutation rate technology to obtain the abnormal time period positioning result;

[0070] Through mutation rate analysis, temperature mutation characteristics are extracted from timestamps to obtain mutation trend data;

[0071] According to the mutation trend data, the linear regression algorithm is used to calculate the slope values ​​before and after time to obtain the slope change sequence;

[0072] If there is a mutation point in the slope change sequence, the abnormal time is located by the timestamp to obtain the abnormal time range;

[0073] Obtain the offset score data within the abnormal time range, use mean calculation to determine the boundary points of the abnormal time period, and obtain the boundary time series;

[0074] For the boundary time series, the abnormal time period is divided by clustering algorithm to obtain the final positioning result;

[0075] Anomaly detection indicators are extracted from the final positioning results, the integrity of the anomaly time period is determined, and verification data is obtained.

[0076] Optionally, the method of calculating the duration of high temperature based on the abnormal time period positioning result by using a heat accumulation time technique to obtain defect duration data includes:

[0077] Obtain abnormal time and time period data through abnormal time period positioning results, use preset thresholds to determine the abnormal time range, and obtain positioning results;

[0078] Extract time period features from the positioning results, use heat accumulation time technology to calculate the high temperature duration, and obtain heat accumulation data;

[0079] Analyze the duration of high temperature based on heat accumulation data and determine the duration value;

[0080] If the duration exceeds the preset threshold, the defect duration data is obtained by matching the defect data with the duration;

[0081] Use time technology to iteratively calculate the defect duration data to obtain the calculation process results;

[0082] By combining the calculation process results with technical methods, the accuracy of the result analysis is judged and the final defect duration data is obtained;

[0083] The time trend of defect duration is generated based on the result analysis, and the support vector machine algorithm is used to classify the trend to obtain the classification result.

[0084] Optionally, the defect duration data is combined with a pre-established knowledge graph to match thermal correlation coefficients with known defect patterns to obtain defect classification labels and severity scores, including:

[0085] The original input sequence is obtained through the defect duration and duration data, and the duration data is standardized using preprocessing tools to obtain a standardized sequence;

[0086] Extract feature vectors from the canonical sequence, match the thermal correlation coefficients in the knowledge graph, and determine the relevance weight of the feature vectors;

[0087] Compare the defect pattern with the known defect based on the correlation weight. If the correlation weight exceeds the preset threshold, it is judged as a matching pattern and a preliminary classification result is obtained;

[0088] Based on the preliminary classification results, combined with the severity of the defect mode analysis, a linear regression algorithm is used to generate a score value to obtain a score sequence;

[0089] The final classification label is determined through the scoring sequence and classification results, and the labels are grouped using a clustering algorithm to obtain a group label set;

[0090] Obtain a group label set, analyze the pattern distribution characteristics of each group in the group label set, and obtain a distribution feature description;

[0091] According to the distribution characteristics, the defect mode stability of each group is judged. If the stability is lower than the preset standard, the group label set is optimized through iterative adjustment to obtain the final label set.

[0092] Optionally, if the severity score exceeds a safety threshold determined based on heat distribution uniformity, a rule engine generates warning information including defect type and time period, including:

[0093] Collect heat distribution data through sensors, calculate uniformity indicators, and obtain preliminary results;

[0094] If the uniformity index exceeds the preset threshold, a judgment condition is generated through comparison to determine the abnormal state;

[0095] Based on the abnormal state, the rule engine is used to analyze the thermal distribution data and obtain the defect type;

[0096] Extract time period information from defect type and obtain time range through timestamp mark;

[0097] Integrate defect types and time periods through the rule engine to generate warning information content;

[0098] Use message queues to push warning information, determine the receiving status, and obtain confirmation results;

[0099] According to the confirmation results, the log files are recorded and stored in the database to obtain the processing records.

[0100] Optionally, according to the defect type in the warning information, the associated thermal diffusion coefficient and historical data features are queried in the knowledge graph to obtain a list of supplementary acquisition parameter suggestions, including:

[0101] Extract defect types through early warning information and determine query targets;

[0102] Obtain the thermal diffusion coefficient associated with the defect type from the knowledge graph to obtain preliminary feature data;

[0103] Match historical data according to thermal diffusion coefficient to obtain corresponding data features;

[0104] Use data feature analysis to supplement parameters and determine whether the acquisition conditions are met. If so, generate parameter recommendations;

[0105] Generate acquisition parameters through parameter suggestions and determine the optimized acquisition plan;

[0106] Build a suggestion list from the collected parameters and output the final result;

[0107] Based on the relationship between the suggestion list and the knowledge graph, update the query results and improve the knowledge graph content.

[0108] Optionally, adjusting the sensor acquisition range according to the supplementary acquisition parameter suggestion list, acquiring a newly added real-time data stream, and updating the reduced-dimensionality feature vector set through feature compression ratio technology to obtain a refined defect recognition result include:

[0109] Adjust the sensor acquisition range by supplementing the collection parameter suggestion list to obtain real-time data streams;

[0110] Update the data stream range according to the new content of the real-time data stream to obtain an extended data set;

[0111] Use feature compression ratio technology to process the expanded data set and determine the initial dimension reduction feature set;

[0112] If the initial dimensionality reduction feature set matches the preset threshold, the feature vector is updated through principal component analysis to obtain a refined feature vector set;

[0113] Perform classification processing on the refined feature vector set to determine whether there are defect features;

[0114] Obtain defect characteristics and compare them with historical data, and determine defect identification results using a logistic regression algorithm;

[0115] Refinement processing is performed based on the defect identification results to obtain the final output data.

[0116] Optionally, analyzing the battery thermal state by using a thermal load distribution technology based on the refined defect identification result to determine whether to generate a final warning report containing thermal characteristics includes:

[0117] Obtain refined results through defect identification, process data using a preset algorithm, and obtain the initial heat load distribution;

[0118] Extract load distribution features from the initial thermal load distribution, perform cluster analysis on the feature values, and determine the battery thermal state category;

[0119] If the battery thermal state category exceeds the preset threshold, the spatial variation trend of the heat load is calculated using distribution technology to determine the degree of thermal abnormality;

[0120] According to the degree of thermal anomaly, thermal characteristic data related to heat load is obtained, and a support vector machine algorithm is used to generate a feature vector;

[0121] Analyze the correlation between thermal characteristics and battery thermal state through feature vectors to determine whether to trigger warning conditions;

[0122] If the warning condition is triggered, a final data set containing thermal characteristics is generated based on the thermal analysis results;

[0123] Using preset templates, key indicators are extracted from the final data set and warning-related information is output.

[0124] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0125] The present invention discloses a new energy vehicle performance index evaluation method for intelligent early warning of battery thermal anomalies. The method collects battery temperature, voltage, and current data in real time, uses time series smoothing and data consistency technology for preprocessing, and then uses principal component analysis and variational autoencoder to extract abnormal distribution characteristics. Based on abnormal sensitivity technology, distribution offset is detected, and abnormal time periods are located in combination with temperature mutation rate and heat accumulation time. Known defect patterns are matched through knowledge graphs, the severity is evaluated, and early warning information is generated. Furthermore, the sensor acquisition range is dynamically adjusted according to the defect type to achieve accurate defect identification and thermal status analysis. The present invention effectively solves the problems of real-time monitoring, precise positioning, and intelligent early warning of battery thermal anomalies, improves battery safety and reliability, and is of great significance to the optimization of battery management systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0126] Figure 1 The present invention is a flowchart of a new energy vehicle performance index evaluation method.

[0127] Figure 2 Schematic diagram of a new energy vehicle performance index evaluation method of the present invention.

[0128] Figure 3 This is another schematic diagram of a new energy vehicle performance index evaluation method of the present invention. DETAILED DESCRIPTION

[0129] The technical solutions of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0130] like Figure 1-3 In this embodiment, a new energy vehicle performance index evaluation method may specifically include:

[0131] S101. Collect battery temperature, voltage, and current data streams in real time through sensors, and use time series smoothing technology to perform sliding window filtering on each parameter to obtain a smoothed time series data set.

[0132] Sensors collect battery temperature, voltage, and current data in real time, generating a data stream containing a time series to produce the original dataset. Sliding window filtering is used to smooth the time series in the original dataset to produce a smoothed time series dataset. For the smoothed time series dataset, the changing trends of each parameter within the sliding window are calculated to produce a parameter change sequence. If the battery temperature in the parameter change sequence exceeds a preset threshold, an anomaly detection algorithm is used to determine the abnormal state, producing an anomaly marker sequence. Based on the anomaly marker sequence, a clustering algorithm is used to group the smoothed time series dataset to produce a grouped dataset. Abnormal groups within the grouped dataset are obtained, and the statistical characteristics of the parameters within each group are calculated to produce a feature sequence. The feature sequence is compared with the preset threshold to determine the battery status and obtain a status assessment result.

[0133] For example, the temperature, voltage, and current data streams of the battery are collected in real time by sensors. Assuming the sampling frequency is once per second, data is collected continuously for 10 seconds to obtain an initial data set. The temperature data is [23, 25, 27, 26, 28, 29, 20, 21, 22, 23], the voltage data is [70, 71, 72, 73, 74, 75, 76, 77, 78, 79], and the current data is [50, 51, 52, 53, 54, 55, 56, 57, 58, 59]. To eliminate noise and smooth the data, a sliding window average filter with a window size of 3 was used to smooth the temperature data. First, the average of the first three data points was calculated: (23 + 25 + 27) / 3 = 25, resulting in the first smoothed value of 25. The window was then slid back one position to calculate (25 + 27 + 26) / 3 = 26, resulting in the second smoothed value of 26. This process continued, ultimately resulting in the smoothed temperature data of [25, 26, 27, 28, 29, 20, 21, 22]. Similarly, a sliding window average filter with a window size of 3 was used to smooth the voltage and current data, resulting in the smoothed voltage data of [71, 72, 73, 74, 75, 76, 77, 78] and the current data of [51, 52, 53, 54, 55, 56, 57, 58]. This sliding window average filter effectively reduced random noise in the data, resulting in a smoother time series dataset, providing a more reliable data foundation for subsequent battery state analysis and prediction.

[0134] S102 : Based on the smoothed time series data set, the temperature, voltage, and current timestamps are synchronously aligned using data consistency technology to obtain a multi-dimensional aligned data set.

[0135] Through time series analysis, the original timestamp sequence is obtained from the temperature data, voltage data, and current data to obtain an initial multidimensional dataset. Smoothing processing technology is used to smooth the time series in the initial multidimensional dataset to obtain smoothed time series data. For the smoothed time series data, the differences between timestamps are used to determine the synchronization alignment requirements and obtain a timestamp deviation set. Using data consistency technology, the timestamp deviation set is adjusted to synchronize the temperature data, voltage data, and current data to obtain a preliminary aligned dataset. The multidimensional data features in the preliminary aligned dataset are obtained, and cluster analysis using the K-means algorithm is used to determine the data synchronization accuracy to obtain an optimized aligned dataset. Based on the optimized aligned dataset, outliers in the time series are detected. If the outliers exceed the preset threshold, they are repaired through interpolation to obtain the final aligned dataset. The synchronized timestamp sequence is extracted from the final aligned dataset to generate a multidimensional aligned dataset.

[0136] For example, in the process of smoothing the time series data set, it is first necessary to preprocess the data in the three dimensions of temperature, voltage, and current to ensure that the data acquisition frequency is consistent. Assuming that the sampling frequency of temperature in the original data is 1 time per second, voltage is 2 times per second, and current is 3 times per second, the temperature data is interpolated using a linear interpolation algorithm to increase its sampling frequency to 3 times per second to ensure synchronization with the current data. The specific operation is that for temperature data points T1 and T2, where T1 = 23°C and T2 = 25°C, the time interval is 1 second, and the new data point after interpolation is T5 = 24°C. Next, the voltage data is downsampled, and the sliding average algorithm is used to reduce the sampling frequency of the voltage data from 2 times per second to 1 time per second. For example, the original voltage data V1 = 221V, V2 = 223V, and the downsampled voltage data V = 222V. Finally, a timestamp alignment algorithm is used to align the timestamps of temperature, voltage, and current. For example, if the current timestamp is 2023-10-01 12:00:00, the temperature data is 24°C, the voltage data is 222V, and the current data is 1A, these are combined into a single multidimensional aligned data record. This process ensures that the resulting time series dataset is completely consistent in the temporal dimension, providing a reliable foundation for subsequent multidimensional data analysis.

[0137] S103 . Calculate the probability density function of each parameter using data distribution extraction technology based on the multi-dimensional aligned data set to obtain a parameter distribution feature set.

[0138] By aligning the dataset multidimensionally, we obtain the data distribution. Extraction techniques are used to calculate parameter density and generate a probability function. Based on the probability function, the distribution characteristics of each parameter are calculated to determine the feature set. If the feature set is complete, the multidimensional alignment is verified using the dataset source to determine the consistency of the alignment dimensions and obtain the alignment result. Based on the alignment result, a calculation process is used to analyze changes in parameter density and obtain updated distribution characteristics. Based on the updated distribution characteristics, the stability of the probability function is determined, resulting in parameter calculation optimization. Based on the parameter calculation optimization, technical application adjustments are made to determine the final feature set. Using the final feature set, extraction techniques are used to analyze data distribution trends and determine a business optimization solution.

[0139] Exemplarily, in a multi-dimensional aligned dataset, the probability density function of each parameter is first calculated using a data distribution extraction technique.

[0140] For example, for a data set containing temperature and humidity, the kernel density estimation (KDE) algorithm can be used to estimate the probability density function of each parameter. Assuming that the temperature data range is 10°C to 30°C and the humidity data range is 40% to 80%, using the Gaussian kernel function and a bandwidth of 5, the probability density functions of temperature and humidity can be calculated through the KDE algorithm. Then, these probability density functions are used to extract parameter distribution feature sets, such as calculating the mean temperature value to be 20°C, the standard deviation to be 5°C, and the mean humidity value to be 60%, and the standard deviation to be 10%. These feature sets can be further used for model training or data classification.

[0141] For example, in weather forecasting models, these feature sets can help identify different weather patterns, such as high temperature and high humidity or low temperature and low humidity. In this way, the parameter distribution feature sets of multi-dimensional aligned datasets can provide strong support for subsequent data analysis and decision-making.

[0142] S104. Based on the parameter distribution feature set, a principal component analysis method is used to perform linear dimensionality reduction on the temperature, voltage, and current to obtain a dimensionality reduction feature vector set reflecting the temperature gradient and heat load distribution.

[0143] Temperature data, voltage data, and current data are obtained through parameter distribution to form an initial feature set. Principal component analysis is used to perform linear dimensionality reduction on the initial feature set to obtain a reduced-dimensionality feature vector set. Feature components corresponding to the temperature gradient are extracted from the reduced-dimensionality feature vector set to determine the temperature change trend. The distribution characteristics of the heat load are determined based on the temperature change trend to obtain the heat load distribution vector. If the components of the heat load distribution vector exceed the preset threshold, the distribution characteristics are adjusted through the vector set to obtain an optimized feature set. For the optimized feature set, cluster analysis is used to divide the relevant areas of temperature gradient and heat load to obtain the regional distribution results. The mapping relationship of the reduced-dimensionality features is generated through the regional distribution results to determine the final vector set.

[0144] For example, in the parameter distribution feature set of temperature, voltage, and current, the original data is first standardized so that each parameter has the same dimension.

[0145] For example, consider temperature data ranging from 20°C to 100°C, voltage from 200V to 400V, and current from 1A to 5A. Using a normalization formula, the data is transformed into a distribution with a mean of 0 and a variance of 1. Principal component analysis (PCA) is then used to reduce the dimensionality of the normalized data. This involves calculating the covariance matrix, assuming it is [[0, 8, 6], [8, 0, 7], [6, 7, 0]], and then solving for its eigenvalues ​​and eigenvectors. Assuming the eigenvalues ​​are [5, 3, 2], the corresponding eigenvectors are [[6, 7, 4], [-7, 1, 7], [3, -7, 6]]. The first two principal components, whose cumulative contribution exceeds 90%, are selected. The original data is projected onto these two principal components to obtain the reduced eigenvector set.

[0146] For example, the original data point [30°C, 250V, 2A] becomes [8, -2] after projection. This vector reflects the key characteristics of the temperature gradient and heat load distribution. This method achieves linear dimensionality reduction of multidimensional parameters, providing a concise and effective feature representation for subsequent thermal management analysis.

[0147] S105. Based on the dimensionality reduction feature vector set, a variational autoencoder is used to generate potential abnormal distribution features through latent variable extraction technology to obtain an abnormal distribution feature set.

[0148] A feature vector is generated by reducing the dimensionality of the features. A variational autoencoder is used to extract the latent variable set from the feature vector to obtain the underlying distribution data. The abnormal distribution characteristics are analyzed from the underlying distribution data to obtain a preliminary representation of the abnormal distribution. If the preliminary representation of the abnormal distribution exceeds a preset threshold, the autoencoder is used to adjust the underlying distribution to obtain an optimized distribution feature. Based on the optimized distribution feature, vector processing techniques are used to generate a feature set to determine the final abnormal distribution feature set. The feature set is verified using extraction techniques to determine whether it meets the expected characteristics of the underlying distribution. After obtaining the verified feature set, a clustering algorithm is used to group the abnormal distribution features to obtain a classified feature subset. For the classified feature subset, statistical tools are used to analyze the distribution patterns of each feature group to obtain a detailed description of the abnormal distribution.

[0149] For example, the input high-dimensional feature vector set is first reduced in dimension using principal component analysis (PCA). For example, a 100-dimensional feature vector is reduced to 20 dimensions, retaining 95% of the variance information. This yields the reduced feature vector set. Next, a variational autoencoder (VAE) is used to extract latent variables. Both the encoder and decoder are configured as three-layer fully connected neural networks, with 128, 64, and 32 neurons per layer, respectively. Reluctant linear unit (ReLU) activation functions are used, and the latent variable dimension is set to 10. The VAE model is trained to learn the latent representation of the data and generate a latent variable distribution. During training, the Adam optimizer is used with a learning rate of 0.01, a batch size of 64, and 100 epochs. The loss function is a weighted sum of reconstruction loss and KL divergence, with weights of 1 and 1, respectively. After training, the latent variable distribution is extracted, and its mean and variance are calculated. For example, the mean is [5, -3, 2, ...] and the variance is [1, 2, 15, ...]. Then, by calculating the difference between the latent variable and the normal sample distribution, the Mahalanobis distance is used as the anomaly score. For example, if a sample has an anomaly score of 5, which exceeds the threshold of 8, it is judged to be anomaly. Finally, the latent variable features of all abnormal samples are clustered and analyzed using the K-means algorithm with K=3. This yields an abnormal distribution feature set. For example, the cluster centers are [[6,-4,3], [7,-5,4], [8,-6,5]], with the number of samples in each cluster being 50, 30, and 20, respectively. This completes the generation and analysis of abnormal distribution features.

[0150] S106. Based on the abnormal distribution feature set, an abnormal sensitivity technology is used to detect distribution shift and obtain an abnormal shift score.

[0151] Data input is obtained using an anomaly distribution feature set, and a feature set is obtained using a preset extraction method. The anomaly distribution is analyzed based on the feature set, and distribution characteristics are derived using statistical methods. The distribution characteristics are used to detect anomaly sensitivity, and sensitivity analysis results are obtained using sensitivity techniques. The sensitivity analysis results determine distribution shift. If the shift exceeds a preset threshold, the shift status is determined. The shift status is used to analyze the distribution shift and obtain a shift score. The shift score determines the degree of anomaly, and a classification algorithm is used to obtain the output result. The output results are used to adjust the detection technology and obtain an optimized shift score.

[0152] For example, in the process of constructing the abnormal distribution feature set, the data is first clustered using the Gaussian mixture model (GMM). Assuming that the data consists of three Gaussian distributions with means of 5, 0, and 5, and variances of 5, the expectation maximization (EM) algorithm is used to iteratively optimize the parameters, ultimately obtaining the probability of belonging to each data point. Next, the Mahalanobis distance of each data point is calculated based on these probabilities. Assuming that the eigenvector of a data point is [0, 5], its corresponding Mahalanobis distance is 8. If the distance is greater than a preset threshold of 0, it is marked as an outlier. In the abnormal sensitivity analysis, the Kullback-Leib l er (KL) divergence is used to measure distribution deviation. Assuming that the KL divergence between the current distribution and the baseline distribution is 3, if it is greater than a threshold of 2, it is considered to have a significant deviation. Finally, the abnormal deviation score is calculated based on the degree of deviation. Assuming the degree of deviation is 5, the score is 5 multiplied by the weight coefficient 8, resulting in a final score of 2, which is used to evaluate the abnormal state of the system.

[0153] S107: If the anomaly offset score exceeds the preset thermal anomaly threshold, the slope before and after the anomaly timestamp is calculated using the temperature mutation rate technology to obtain the anomaly time period positioning result.

[0154] If the offset score exceeds the preset threshold, the slope before and after the timestamp is calculated using temperature mutation rate technology to obtain the abnormal time period location result. Through mutation rate analysis, the temperature mutation characteristics are extracted from the timestamp to obtain mutation trend data. Based on the mutation trend data, a linear regression algorithm is used to calculate the slope values ​​before and after the time to obtain a slope change sequence. If there is a mutation point in the slope change sequence, the abnormal time is located by the timestamp to obtain the abnormal time range. The offset score data within the abnormal time range is obtained, and the boundary points of the abnormal time period are determined using mean calculation to obtain the boundary time series. Based on the boundary time series, the abnormal time period is divided into the abnormal time period using a clustering algorithm to obtain the final location result. Anomaly detection indicators are extracted from the final location result to determine the integrity of the abnormal time period and obtain verification data.

[0155] For example, during thermal anomaly detection, the thermal anomaly threshold is set at 85 points. When the anomaly excursion score reaches 90 points, the system determines that the threshold has been exceeded. The temperature mutation rate technique is then used for calculation. The specific algorithm is as follows: temperature data is selected from the 5 minutes before and 5 minutes after the anomaly timestamp, and their average slope is calculated.

[0156] For example, the temperature data for the first five minutes is [20, 25, 20, 25, 20], and the temperature data for the last five minutes is [25, 20, 25, 20, 25]. Linear regression analysis shows that the slope for the first five minutes is 5°C / minute, and the slope for the last five minutes is 4°C / minute. Based on the significant change in the slope, the system locates the abnormal time period as 5 minutes before and after the abnormal timestamp, for a total of 10 minutes. This process, implemented through an automated algorithm, ensures the precise location of the abnormal time period and provides reliable data support for further anomaly analysis.

[0157] S108. Based on the abnormal time period positioning result, the high temperature duration is calculated using the heat accumulation time technology to obtain the defect duration data.

[0158] The abnormal time and time period data are obtained through the abnormal time period positioning results. The abnormal time range is judged using a preset threshold to obtain the positioning results. The time period characteristics are extracted from the positioning results. The high temperature duration is calculated using the thermal accumulation time technology to obtain the thermal accumulation data. The duration of the high temperature is analyzed based on the thermal accumulation data to determine the duration value. If the duration exceeds the preset threshold, the defect data is matched with the duration to obtain the defect duration data. The defect duration data is iteratively calculated using time technology to obtain the calculation process results. By combining the calculation process results with technical methods, the accuracy of the result analysis is judged to obtain the final defect duration data. Based on the result analysis, a time trend of the defect duration is generated. The support vector machine algorithm is used to classify the trend to obtain the classification result.

[0159] For example, in the abnormal time period location results, suppose sensors detected an abnormally high temperature on a device between 10:00 AM and 2:00 PM on October 1, 2023, reaching a peak of 120°C. Using the thermal accumulation time technique, the device's normal operating temperature threshold is set at 80°C. The thermal accumulation time is calculated when the temperature exceeds this threshold. According to the temperature curve, the device first exceeded the threshold at 10:30 AM and continued to fall below it until 1:45 PM, resulting in a high temperature duration of 3 hours and 15 minutes. Using the thermal accumulation time algorithm, the difference between the duration of the high temperature and the time the temperature exceeded the threshold is integrated to calculate the value: Thermal accumulation time = Σ(Δt × (T(t) - T_threshold)), where Δt is the time interval, T(t) is the temperature at time t, and T_threshold is the temperature threshold. Assuming temperature data is recorded every 5 minutes, the calculated thermal accumulation time is 450 minutes. Further analysis of the defect duration reveals the empirical formula: Defect duration = heat accumulation time / k, where k is the thermal conductivity of the equipment material. Assuming k is 10°C / minute, the defect duration is 45 minutes. This calculation and analysis accurately assesses the duration of high temperatures during abnormal periods and the extent of the defect, providing data support for subsequent maintenance and troubleshooting.

[0160] S109. Based on the defect duration data, the thermal correlation coefficient is matched with the known defect pattern in combination with the pre-established knowledge graph to obtain the defect classification label and severity score.

[0161] The original input sequence is obtained through the defect duration and duration data, and the duration data is standardized using a preprocessing tool to obtain a standard sequence. Feature vectors are extracted from the standard sequence, matched with the thermal correlation coefficient in the knowledge graph, and the correlation weight of the feature vector is determined. The defect pattern of the known defect is compared based on the correlation weight. If the correlation weight exceeds the preset threshold, it is judged as a matching pattern to obtain a preliminary classification result. Based on the preliminary classification results, the severity of the defect pattern is analyzed and a linear regression algorithm is used to generate a scoring value to obtain a scoring sequence. The final classification label is determined based on the scoring sequence and classification results, and the labels are grouped using a clustering algorithm to obtain a group label set. The group label set is obtained, and for each group in the group label set, its pattern distribution characteristics are analyzed to obtain a distribution feature description. Based on the distribution feature description, the stability of the defect pattern of each group is judged. If the stability is lower than the preset standard, the group label set is optimized through iterative adjustment to obtain the final label set.

[0162] For example, based on the defect duration data, the defects are first associated with known patterns by matching thermal correlation coefficients through the knowledge graph.

[0163] For example, if a system defect lasts 48 hours, the system will match it with the "memory leak" pattern in the knowledge graph. The matching thermal correlation coefficient is 85, indicating that the defect is highly correlated with the memory leak pattern. Based on the matching results, the system then uses a weighted scoring algorithm to classify the defect and assess its severity.

[0164] Specifically, the system will assign different weights based on dimensions such as the duration of the defect, the scope of impact, and the historical repair time. Taking this defect as an example, the weight of its duration is 40%, the weight of its scope of impact is 30%, and the weight of its historical repair time is 30%. Assuming that the scope of impact is the "core module" and the historical repair time is 24 hours, the severity score obtained through weighted calculation is: 48*4+9 (corresponding score of the core module)*3+24*3=33 points. According to the preset scoring threshold, 33 points is classified as a "high severity" defect. Finally, the system automatically generates a defect classification label "memory leak" and marks the severity as "high" to facilitate subsequent priority processing and resource allocation. The entire process is automated through information technology without the need for human intervention, ensuring the efficiency and accuracy of defect management.

[0165] S1010. If the severity score exceeds a safety threshold determined based on heat distribution uniformity, a warning message including defect type and time period is generated through a rule engine.

[0166] Sensors collect thermal distribution data and calculate uniformity indicators to obtain preliminary results. If the uniformity indicator exceeds the preset threshold, a comparison generates a judgment condition to determine an abnormal state. Based on the abnormal state, a rules engine analyzes the thermal distribution data to determine the defect type. Time period information is extracted from the defect type, and the time range is obtained through timestamp marking. The rules engine integrates the defect type and time period to generate warning information. Warning information is pushed through a message queue, and the reception status is determined to obtain a confirmation result. The confirmation result is recorded in a log file and stored in a database to obtain a processing record.

[0167] For example, in thermal uniformity analysis, infrared thermal imaging technology is used to collect temperature data from the device surface to generate a thermal distribution map. This image is then smoothed using a Gaussian filter to eliminate noise. Next, a pixel-based temperature difference calculation method is used to calculate the temperature difference between each pixel and its surrounding pixels. If the average temperature difference in a region exceeds a set safety threshold (e.g., 5°C), the region is deemed to have uneven thermal distribution.

[0168]

[0169] Where T_ij represents the temperature value of the current pixel, i and j represent the coordinate positions of the pixel, the summation term represents the temperature values ​​of the eight surrounding pixels, the coefficient 1 / 8 represents the average value, and ΔT_ij represents the average temperature difference between the current pixel and the surrounding pixels.

[0170]

[0171] Where M and N represent the number of rows and columns of the image, respectively. ΔT_ij represents the temperature difference of each pixel. The sum of the values ​​is divided by the total number of pixels MN to obtain the average temperature difference of the region.

[0172]

[0173] Here, γ represents the judgment result, and T_th represents the set temperature threshold. When the regional average temperature difference exceeds the threshold, γ equals 1, indicating uneven heat distribution; otherwise, γ equals 0, indicating uniform temperature distribution. A weighted algorithm is used to calculate the severity score, using the temperature difference, regional area, and duration as input parameters with weights of 4, 3, and 3, respectively, to calculate a comprehensive score. If the score exceeds the preset threshold (e.g., 80 points), the rule engine is triggered to generate an early warning message. The rule engine matches the scoring result with the defect signature library to determine the defect type (e.g., local overheating, poor heat dissipation, etc.) and, combined with the timestamp, generates an early warning message containing the defect type and time period, such as "local overheating, October 15, 2023, 14:30 to 15:00." The early warning message is sent to the monitoring system via a message queue, triggering an automatic alarm and recording a log for subsequent analysis and processing.

[0174] S1011. Based on the defect type in the warning information, query the associated thermal diffusion coefficient and historical data features in the knowledge graph to obtain a list of recommended supplementary acquisition parameters.

[0175] Defect types are extracted from early warning information to determine the query target. The thermal diffusion coefficient associated with the defect type is obtained from the knowledge graph to generate preliminary feature data. Historical data is matched against the thermal diffusion coefficient to obtain corresponding data features. Data feature analysis is used to supplement parameters to determine whether acquisition conditions are met. If so, parameter recommendations are generated. Acquisition parameters are generated based on the parameter recommendations, and an optimized acquisition plan is determined. A list of recommendations is constructed from the acquisition parameters, and the final results are output. Based on the relationship between the recommendation list and the knowledge graph, the query results are updated to improve the knowledge graph content.

[0176] For example, according to the defect type in the warning information, the associated thermal diffusion coefficient is first queried through the knowledge graph. For example, the thermal diffusion coefficient of a certain material is 8mm 2 / s, combined with historical data characteristics, such as the thermal diffusivity of the material in a high temperature environment varies from 7 to 9 mm2 / s, and further analyze its thermal diffusion behavior at different temperatures. Through algorithm calculation, such as using finite element analysis method to simulate the thermal diffusion process of the material under a specific temperature gradient, it is concluded that when the temperature gradient is 50℃ / mm, the thermal diffusion coefficient is 85mm 2 / s. Based on these data, a list of supplementary acquisition parameter recommendations is generated, including a recommended acquisition temperature range of 200°C to 300°C and an acquisition frequency of 10 times per second to ensure data accuracy and completeness. At the same time, combined with abnormal points in historical data, such as the abnormal increase in thermal diffusion coefficient to 95mm at 250°C in a certain acquisition, 2 / s, it is recommended to increase the acquisition density around 250°C to capture possible abnormal phenomena. In this way, the acquisition parameters can fully reflect the thermal diffusion characteristics of the material and provide reliable data support for subsequent analysis.

[0177] S1012. Adjust the sensor acquisition range according to the recommended list of supplementary acquisition parameters, obtain the newly added real-time data stream, and update the reduced-dimensionality feature vector set through feature compression ratio technology to obtain a refined defect recognition result.

[0178] Adjust the sensor acquisition range by supplementing the recommended acquisition parameter list to acquire a real-time data stream. Update the data stream range based on the new content in the real-time data stream to obtain an expanded data set. Use feature compression ratio technology to process the expanded data set and determine a preliminary dimensionality reduction feature set. If the preliminary dimensionality reduction feature set matches the preset threshold, update the feature vectors using principal component analysis to obtain a refined feature vector set. Classify the refined feature vector set to determine whether defect features exist. Compare the defect features with historical data and determine the defect identification results using a logistic regression algorithm. Refine the defect identification results to obtain the final output data.

[0179] For example, during the sensor acquisition range adjustment process, the sensor's acquisition range is first adjusted from the original 0-100 mm to 0-150 mm based on the supplementary acquisition parameter recommendation list to ensure coverage of a wider detection area. This adjustment enables the sensor to capture additional real-time data streams, such as defect signals detected at 120 mm. Next, feature compression technology is used to process the new data. Specifically, the principal component analysis (PCA) algorithm is used to compress the original data from 100 dimensions to 20 dimensions, retaining 95% of the variance information. This process is achieved by calculating the covariance matrix and extracting the first 20 principal components, significantly reducing data storage and computational complexity. Subsequently, the updated set of reduced-dimensionality feature vectors is combined with a support vector machine (SVM) classifier for defect identification. The kernel function is the radial basis function (RBF), the penalty parameter C is set to 0, and the kernel parameter γ is set to 0. Through this series of operations, the system can more accurately identify defect types, for example, reducing the original false positive rate from 5% to 2%, significantly improving the accuracy and efficiency of defect identification.

[0180] S1013. Based on the refined defect identification results, analyze the battery thermal state using thermal load distribution technology to determine whether to generate a final warning report containing thermal characteristics.

[0181] Obtain refined results through defect identification, use a preset algorithm to process the data, and obtain the initial thermal load distribution. Extract load distribution characteristics from the initial thermal load distribution, perform cluster analysis on the eigenvalues, and determine the battery thermal state category. If the battery thermal state category exceeds the preset threshold, calculate the spatial variation trend of the thermal load through distribution technology to determine the degree of thermal abnormality. According to the degree of thermal abnormality, obtain thermal feature data related to the thermal load, and use the support vector machine algorithm to generate feature vectors. Analyze the correlation between thermal features and battery thermal state through feature vectors to determine whether the warning conditions are triggered. If the warning conditions are triggered, generate a final data set containing thermal features based on the thermal state analysis results. Use preset templates to extract key indicators from the final data set and output warning-related information.

[0182] For example, in the process of refined defect identification, the voltage, current and temperature data of the battery are first collected by high-precision sensors, and the internal defects of the battery are identified by combining deep learning algorithms such as convolutional neural networks (CNN), with an identification accuracy of more than 95%. Subsequently, the thermal load distribution technology is used to simulate the thermal field distribution of the battery under different working conditions through finite element analysis (FEA), and the temperature gradient of each part of the battery is calculated. For example, when the discharge rate is 2C, the surface temperature of the battery can reach up to 45°C, while the internal hot spot temperature may reach 60°C. Based on these data, the heat conduction equation and heat diffusion model are used to further analyze the thermal state of the battery to determine whether there is a risk of thermal runaway.

[0183] For example, if the internal battery temperature exceeds 65°C and the temperature gradient exceeds 10°C / cm, the system automatically generates a final warning report containing thermal characteristics. The report details the temperature anomaly area, heat load distribution, and potential risk level, providing a scientific basis for subsequent battery management and maintenance. The entire process uses automated data processing and algorithmic analysis to ensure the timeliness and accuracy of warnings.

[0184] Although the present invention has been described in detail above using general descriptions and specific embodiments, it will be apparent to those skilled in the art that modifications and improvements may be made thereto. Therefore, such modifications and improvements, without departing from the spirit of the present invention, are intended to be within the scope of protection claimed herein.

Claims

1. A new energy vehicle performance index evaluation method, characterized in that: The method comprises: The battery temperature, voltage, and current data streams are collected in real time by sensors, and the parameters are subjected to sliding window filtering using time series smoothing technology to obtain a smoothed time series data set. Based on the smoothed time series data set, the temperature, voltage, and current timestamps are synchronized and aligned using data consistency technology to obtain a multi-dimensional aligned data set. According to the multi-dimensional aligned data set, the probability density function of each parameter is calculated by data distribution extraction technology to obtain the parameter distribution feature set; According to the parameter distribution feature set, the principal component analysis method is used to perform linear dimension reduction on temperature, voltage and current to obtain a dimension reduction feature vector set reflecting the temperature gradient and heat load distribution. According to the reduced dimension feature vector set, the latent variable extraction technology is used to generate potential abnormal distribution features using the variational autoencoder to obtain the abnormal distribution feature set; Based on the abnormal distribution feature set, the abnormal sensitivity technology is used to detect the distribution shift and obtain the abnormal shift score; If the anomaly offset score exceeds the preset thermal anomaly threshold, the slope before and after the anomaly timestamp is calculated using the temperature mutation rate technology to obtain the anomaly time period positioning result; Based on the abnormal time period positioning results, the high temperature duration is calculated using the heat accumulation time technology to obtain the defect duration data; Based on the defect duration data, the thermal correlation coefficient is matched with the known defect patterns in combination with the pre-established knowledge graph to obtain the defect classification label and severity score. If the severity score exceeds the safety threshold determined based on the uniformity of heat distribution, a warning message containing the defect type and time period is generated through the rule engine; Based on the defect type in the warning information, the associated thermal diffusion coefficient and historical data features in the knowledge graph are queried to obtain a list of recommended supplementary acquisition parameters; Adjust the sensor acquisition range according to the recommended list of supplementary acquisition parameters, obtain new real-time data streams, and update the reduced-dimensional feature vector set using feature compression ratio technology to obtain refined defect recognition results; Based on the refined defect identification results, the battery thermal status is analyzed using thermal load distribution technology to determine whether a final warning report containing thermal characteristics should be generated.

2. The new energy vehicle performance index evaluation method according to claim 1, characterized in that: The battery temperature, voltage, and current data streams are collected in real time by sensors, and each parameter is subjected to sliding window filtering using time series smoothing technology to obtain a smoothed time series data set, including: The battery temperature, voltage, and current data are collected in real time through sensors to generate a data stream containing a time series and obtain the original data set; The sliding window filtering technology is used to smooth the time series in the original data set to obtain a smoothed time series data set; For the smoothed time series data set, the changing trend of each parameter in the sliding window is calculated to obtain the parameter change sequence; If the battery temperature in the parameter change sequence exceeds the preset threshold, the abnormal state is determined by the anomaly detection algorithm to obtain an abnormal flag sequence; According to the abnormal label sequence, the clustering algorithm is used to group the smooth time series data set to obtain the grouped data set; Obtain the abnormal groups in the grouped data set, calculate the statistical characteristics of the parameters in each group, and obtain the characteristic sequence; By comparing the characteristic sequence with the preset threshold, the battery status is judged and the status assessment result is obtained.

3. The new energy vehicle performance index evaluation method according to claim 1, characterized in that: The method uses data consistency technology to synchronize and align the temperature, voltage, and current timestamps based on the smoothed time series data set to obtain a multi-dimensional aligned data set, including: Through time series analysis, the original timestamp sequence is obtained from the temperature data, voltage data and current data to obtain the initial multidimensional data set; Smoothing technology is used to smooth the time series in the initial multidimensional data set to obtain smoothed time series data; For smoothed time series data, the differences between timestamps are used to determine the synchronization alignment requirements and obtain the timestamp deviation set; Through data consistency technology, the timestamp deviation set is adjusted, and the temperature data, voltage data, and current data are synchronously aligned to obtain a preliminary aligned data set; Obtain multidimensional data features from the preliminary aligned dataset, perform cluster analysis using the K-means algorithm, determine data synchronization accuracy, and obtain an optimized aligned dataset; Based on the optimized aligned dataset, outliers in the time series are detected. If the outliers exceed the preset threshold, they are repaired through interpolation to obtain the final aligned dataset. Extract the synchronized timestamp sequence from the final aligned dataset to generate a multidimensional aligned dataset.

4. The new energy vehicle performance index evaluation method according to claim 1, characterized in that: The method calculates the probability density function of each parameter using data distribution extraction technology based on the multi-dimensional aligned data set to obtain a parameter distribution feature set, including: By aligning the data set in multiple dimensions, we can obtain the data distribution, calculate the parameter density using extraction techniques, and obtain the probability function. For the probability function, calculate the distribution characteristics of each parameter and determine the feature set; If the feature set is complete, the multi-dimensional alignment is verified through the dataset source to determine the consistency of the alignment dimensions and obtain the alignment result; Based on the alignment results, a computational process is used to analyze the parameter density changes and obtain distribution feature updates; By updating the distribution characteristics, the stability of the probability function is judged and the parameter calculation optimization is obtained; Optimize parameter calculations, obtain technical application adjustments, and determine the final feature set; Through the final feature set, extraction technology is used to analyze data distribution trends and obtain business optimization solutions.

5. The new energy vehicle performance index evaluation method according to claim 1, characterized in that: According to the parameter distribution feature set, the principal component analysis method is used to perform linear dimensionality reduction on the temperature, voltage, and current to obtain a dimensionality reduction feature vector set reflecting the temperature gradient and heat load distribution, including: Obtain temperature data, voltage data, and current data through parameter distribution to form an initial feature set; The principal component analysis method is used to perform linear dimensionality reduction on the initial feature set to obtain a reduced dimensionality feature vector set; Extract the characteristic components corresponding to the temperature gradient from the reduced-dimensional feature vector set to determine the temperature change trend; Determine the distribution characteristics of the heat load based on the temperature change trend and obtain the heat load distribution vector; If the component of the heat load distribution vector exceeds the preset threshold, the distribution characteristics are adjusted through the vector set to obtain an optimized characteristic set; For the optimized feature set, cluster analysis method is used to divide the relevant areas of temperature gradient and heat load to obtain regional distribution results; The mapping relationship of dimensionality reduction features is generated through the regional distribution results to determine the final vector set.

6. The new energy vehicle performance index evaluation method according to claim 1, characterized in that: According to the reduced dimension feature vector set, the potential abnormal distribution features are generated by using the variational autoencoder through the latent variable extraction technology to obtain the abnormal distribution feature set, including: Generate feature vectors by reducing dimensionality features, and use variational autoencoders to extract latent variable sets from feature vectors to obtain potential distribution data; Analyze the abnormal distribution characteristics from the underlying distribution data to obtain a preliminary representation of the abnormal distribution; If the initial representation of the abnormal distribution exceeds the preset threshold, the latent distribution is adjusted through the autoencoder to obtain the optimized distribution characteristics; Based on the optimized distribution characteristics, vector processing technology is used to generate a feature set to determine the final abnormal distribution feature set; Verify the feature set through extraction techniques to determine whether it conforms to the expected characteristics of the underlying distribution; Obtain the verified feature set, use the clustering algorithm to group the abnormal distribution features, and obtain the classified feature subset; For the classified feature subsets, the distribution patterns of each group of features are analyzed through statistical tools to obtain a detailed description of the abnormal distribution.

7. The new energy vehicle performance index evaluation method according to claim 1, characterized in that: The abnormal distribution feature set is used to detect the distribution shift using an abnormal sensitivity technique to obtain an abnormal shift score, including: Obtain data input through abnormal distribution feature set, and obtain feature set using preset extraction method; Analyze abnormal distribution based on feature set and use statistical methods to obtain distribution characteristics; Detect abnormal sensitivity through distribution characteristics and obtain sensitivity analysis results using sensitivity technology; The distribution shift is determined based on the sensitivity analysis results. If the shift exceeds a preset threshold, the shift status is determined. The distribution shift is analyzed by shift status to obtain the shift score; The degree of abnormality is determined based on the deviation score, and the classification algorithm is used to obtain the output result; The detection technology is adjusted based on the output results to obtain the optimized offset score.

8. The new energy vehicle performance index evaluation method according to claim 1, characterized in that: If the anomaly offset score exceeds the preset thermal anomaly threshold, the slope before and after the anomaly timestamp is calculated using the temperature mutation rate technology to obtain the anomaly time period positioning result, including: If the offset score exceeds the preset threshold, the slope before and after the timestamp is calculated using the temperature mutation rate technology to obtain the abnormal time period positioning result; Through mutation rate analysis, temperature mutation characteristics are extracted from timestamps to obtain mutation trend data; According to the mutation trend data, the linear regression algorithm is used to calculate the slope values ​​before and after time to obtain the slope change sequence; If there is a mutation point in the slope change sequence, the abnormal time is located by the timestamp to obtain the abnormal time range; Obtain the offset score data within the abnormal time range, use mean calculation to determine the boundary points of the abnormal time period, and obtain the boundary time series; For the boundary time series, the abnormal time period is divided by clustering algorithm to obtain the final positioning result; Anomaly detection indicators are extracted from the final positioning results, the integrity of the anomaly time period is determined, and verification data is obtained.

9. The new energy vehicle performance index evaluation method according to claim 1, characterized in that: The above method calculates the duration of high temperature based on the abnormal time period positioning result through the heat accumulation time technology to obtain the defect duration data, including: Obtain abnormal time and time period data through abnormal time period positioning results, use preset thresholds to determine the abnormal time range, and obtain positioning results; Extract time period features from the positioning results, use heat accumulation time technology to calculate the high temperature duration, and obtain heat accumulation data; Analyze the duration of high temperature based on heat accumulation data and determine the duration value; If the duration exceeds the preset threshold, the defect duration data is obtained by matching the defect data with the duration; Use time technology to iteratively calculate the defect duration data to obtain the calculation process results; By combining the calculation process results with technical methods, the accuracy of the result analysis is judged and the final defect duration data is obtained; The time trend of defect duration is generated based on the result analysis, and the support vector machine algorithm is used to classify the trend to obtain the classification result.

10. The new energy vehicle performance index evaluation method according to claim 1, characterized in that: The defect duration data is combined with the pre-established knowledge graph to match the thermal correlation coefficient with the known defect pattern to obtain the defect classification label and severity score, including: The original input sequence is obtained through the defect duration and duration data, and the duration data is standardized using preprocessing tools to obtain a standardized sequence; Extract feature vectors from the canonical sequence, match the thermal correlation coefficients in the knowledge graph, and determine the relevance weight of the feature vectors; Compare the defect pattern with the known defect based on the correlation weight. If the correlation weight exceeds the preset threshold, it is judged as a matching pattern and a preliminary classification result is obtained; Based on the preliminary classification results, combined with the severity of the defect mode analysis, a linear regression algorithm is used to generate a score value to obtain a score sequence; The final classification label is determined through the scoring sequence and classification results, and the labels are grouped using a clustering algorithm to obtain a group label set; Obtain a group label set, analyze the pattern distribution characteristics of each group in the group label set, and obtain a distribution feature description; According to the distribution characteristics, the defect mode stability of each group is judged. If the stability is lower than the preset standard, the group label set is optimized through iterative adjustment to obtain the final label set.

Images