Fault early warning method and system for battery pack based on data analysis

Abstract

The invention discloses a fault early warning method and system for a battery pack based on data analysis, and the method comprises the steps: obtaining asynchronous sampling data of voltage, current and temperature parameters from a plurality of data sources, mapping the asynchronous sampling data to a preset time grid through employing a linear interpolation algorithm, and generating time series data; inputting the time sequence data into a long short-term memory network, and calculating through forward propagation to obtain a hidden state sequence containing time dependence characteristics; calculating a feature weight of the hidden state sequence by adopting an attention mechanism, and if the feature weight is greater than a preset threshold value, reserving corresponding features to form a simplified feature set; calculating posterior probability distribution by adopting a Bayesian reasoning algorithm based on the simplified feature set, and determining a fault occurrence probability; and according to the fault occurrence probability, a dynamic threshold adjustment algorithm is adopted to update an early warning threshold, and if the fault probability exceeds the current early warning threshold, a graded early warning signal is triggered. According to the invention, a powerful guarantee is provided for safe and stable operation of power equipment.

Description

Technical Field

[0001] The present invention relates to the field of data processing, and in particular discloses a fault warning method and system for a battery pack based on data analysis. Background Art

[0002] As a core component in fields such as energy storage systems and electric vehicles, the operating stability and safety of a battery pack are crucial for system performance. Battery pack failures may lead to performance degradation, shortened lifespan, and even safety accidents. Therefore, developing an efficient fault warning method is of great significance. However, existing solutions have obvious limitations in processing complex operating data. Traditional methods mostly rely on a single data source or static thresholds, making it difficult to adapt to the dynamic characteristics of multi-source heterogeneous data during the operation of the battery pack, and they have insufficient processing capabilities for asynchronous sampling data and outliers, easily resulting in false alarms or missed alarms. In addition, existing models perform poorly in capturing long-term dependency relationships and quantifying prediction uncertainties, restricting the accuracy and reliability of the warning.

[0003] In the field of battery pack fault warning, the core challenges focus on the efficient processing of multi-source data, the accurate alignment of time series, and the robustness of dynamic warning. First, multi-source data includes various parameters such as voltage, current, and temperature. The data has a high dimension and there are asynchronous sampling problems, making it difficult to directly use for fault identification. Second, modeling long-term dependency relationships in time series data requires powerful algorithms, and traditional methods are difficult to balance real-time performance and prediction accuracy. Finally, the setting of dynamic thresholds and uncertainty quantification are the keys to achieving hierarchical warning, and existing methods lack the adaptive ability to complex working conditions. These technical factors not being resolved result in the limited applicability and accuracy of the fault warning system in complex scenarios.

[0004] Therefore, how to design a battery pack fault warning method based on multi-source data analysis, through efficient data processing, accurate time series modeling, and dynamic threshold management, to achieve highly robust and accurate fault identification and warning, has become the key issue of this research. Summary of the Invention

[0005] The present invention provides a fault warning method and system for a battery pack based on data analysis, aiming to solve at least one of the above-mentioned defects existing in the existing battery pack fault warning methods.

[0006] One aspect of the present invention relates to a fault warning method for a battery pack based on data analysis, including the following steps: Obtain asynchronous sampling data of voltage, current, and temperature parameters from multiple data sources, and use a linear interpolation algorithm to map the asynchronous sampling data to a preset time grid to generate time series data; Input the time series data into a long short-term memory network, and calculate the hidden state sequence containing time-dependent features through forward propagation; The attention mechanism is used to calculate the feature weights for the hidden state sequence. If the feature weight is greater than the preset threshold, the corresponding features are retained to form a refined feature set; Based on the refined feature set, the Bayesian inference algorithm is used to calculate the posterior probability distribution to determine the probability of a fault occurring; According to the probability of a fault occurring, the dynamic threshold adjustment algorithm is used to update the warning threshold. If the fault probability exceeds the current warning threshold, a hierarchical warning signal is triggered.

[0007] Furthermore, the steps of obtaining asynchronous sampling data of voltage, current, and temperature parameters from multiple data sources and using the linear interpolation algorithm to map the asynchronous sampling data to a preset time grid to generate time series data include: Obtain asynchronous sampling data of voltage parameters, current parameters, and temperature parameters from multiple data sources, and perform data acquisition at a preset sampling frequency to obtain the original asynchronous sampling data set; If the timestamps of the original asynchronous sampling data set are not aligned, the timestamp alignment algorithm is used. By calculating the differences between the timestamps of each data source and the preset time grid, the nearest grid points are determined to obtain the sampling data set with aligned timestamps; For the sampling data set with aligned timestamps, the linear interpolation algorithm is used to calculate the interpolations of the voltage parameters, current parameters, and temperature parameters at each sampling point on the preset time grid to obtain the interpolated time series data set; According to the interpolated time series data set, judge whether the interpolation accuracy meets the preset threshold. If not, adjust the linear interpolation step size and recalculate the interpolation to obtain the final time series data.

[0008] Furthermore, the steps of inputting the time series data into the long short-term memory network and calculating the hidden state sequence containing time-dependent features through forward propagation include: Obtain the original time series data from the time series database, perform normalization processing on the original time series data through a preprocessing tool, and use the mean normalization method to calculate the normalized value of each data point to obtain the normalized time series data; According to the normalized time series data, input it into the pre-established long short-term memory network, and calculate the cell state and hidden state at each time step through the forward propagation algorithm to obtain the hidden state sequence containing time-dependent features; If the dimension of the hidden state sequence exceeds the preset threshold, the hidden state sequence is dimensionally reduced through a principal component analysis tool to obtain the dimensionally reduced feature sequence; For the dimensionally reduced feature sequence, a sequence generation tool is used to reorganize the dimensionally reduced feature sequence, and the dependence relationship between features is calculated through the time window sliding method to obtain the final time-dependent feature sequence.

[0009] Further, the steps of calculating the feature weights for the hidden state sequence using the attention mechanism and retaining the corresponding features to form a refined feature set if the feature weight is greater than a preset threshold include: Obtain the original feature data from the hidden state sequence, calculate the attention score for each feature using the attention mechanism, and obtain the weight value of each feature through normalizing the attention score; Compare the weight value with the preset threshold. If the weight value is greater than the preset threshold, retain the corresponding feature to generate a preliminary refined feature set; For the preliminary refined feature set, perform dimensionality reduction processing using the principal component analysis algorithm, and obtain a refined feature set with a lower dimension by retaining the main variance direction; Through sequence analysis of the refined feature set, obtain the equidistant arrangement result, generate an optimized data representation, and obtain the final refined feature set.

[0010] Further, the steps of calculating the posterior probability distribution based on the refined feature set using the Bayesian inference algorithm and determining the probability of a fault occurrence include: Obtain the feature set from the training data, calculate the conditional probability of each feature using a statistical analysis tool, and obtain the conditional probability distribution corresponding to the refined feature set; According to the conditional probability distribution, calculate the posterior probability using the Bayesian inference algorithm in combination with the pre-established model parameters to obtain the posterior probability distribution.

[0011] If the posterior probability distribution is known, for the fault type in the real-time data, use a logical judgment tool to analyze the posterior probability distribution, judge the probability of a fault occurrence, and determine the fault probability value; Through a prediction result generation tool, associate the fault probability value with the fault type to generate a prediction result and obtain a fault prediction report.

[0012] Further, according to the probability of a fault occurrence, use a dynamic threshold adjustment algorithm to update the warning threshold. If the fault probability exceeds the current warning threshold, the steps of triggering a hierarchical warning signal include: Obtain the fault probability value from the real-time data stream, and perform standardization processing on the fault probability value through a data processing tool to obtain a standardized fault probability value; According to the standardized fault probability value, use the dynamic threshold adjustment algorithm in combination with the preset threshold update rule to calculate a new warning threshold and determine the updated warning threshold; If the standardized fault probability value exceeds the updated warning threshold, use a logical judgment tool to generate a corresponding hierarchical warning level in combination with the fault type and the preset hierarchical rule to obtain the hierarchical warning level; According to the graded warning level, a signal generation tool is used to associate the graded warning level with a preset signal template to generate a warning signal and obtain a triggered warning signal.

[0013] Another aspect of the present invention relates to a fault warning system for a battery pack based on data analysis, which is used to implement the above-mentioned fault warning method for a battery pack based on data analysis. The fault warning system for a battery pack based on data analysis includes: A generation module, configured to obtain asynchronous sampling data of voltage, current, and temperature parameters from multiple data sources, and map the asynchronous sampling data to a preset time grid by using a linear interpolation algorithm to generate time series data; An acquisition module, configured to input the time series data into a long short-term memory network and calculate a hidden state sequence containing time-dependent features through forward propagation; A processing module, configured to calculate feature weights for the hidden state sequence by using an attention mechanism. If the feature weight is greater than a preset threshold, the corresponding feature is retained to form a refined feature set; A determination module, configured to calculate a posterior probability distribution based on the refined feature set by using a Bayesian inference algorithm and determine the probability of a fault occurring; A triggering module, configured to update the warning threshold by using a dynamic threshold adjustment algorithm according to the probability of a fault occurring. If the fault probability exceeds the current warning threshold, a graded warning signal is triggered.

[0014] Furthermore, the generation module includes: A first acquisition unit, configured to obtain asynchronous sampling data of voltage parameters, current parameters, and temperature parameters from multiple data sources, and perform data acquisition at a preset sampling frequency to obtain an original asynchronous sampling data set; A second acquisition unit, configured to, if the timestamps of the original asynchronous sampling data set are not aligned, use a timestamp alignment algorithm to determine the nearest grid point by calculating the difference between the timestamps of each data source and the preset time grid, and obtain a sampled data set with aligned timestamps; A third acquisition unit, configured to, for the sampled data set with aligned timestamps, use a linear interpolation algorithm to calculate the interpolation of voltage parameters, current parameters, and temperature parameters at each sampling point on the preset time grid to obtain an interpolated time series data set; A fourth acquisition unit, configured to determine whether the interpolation accuracy meets a preset threshold according to the interpolated time series data set. If not, adjust the linear interpolation step size and recalculate the interpolation to obtain the final time series data.

[0015] Furthermore, the acquisition module includes: A fifth acquisition unit, configured to acquire original time series data from a time series database, perform normalization processing on the original time series data through a preprocessing tool, calculate the normalized value of each data point by using a mean normalization method, and obtain normalized time series data; A sixth acquisition unit, configured to input the normalized time series data into a pre-established long short-term memory network, and calculate the cell state and hidden state of each time step through a forward propagation algorithm, so as to obtain a hidden state sequence including time-dependent features; A seventh acquisition unit, configured to, if the dimension of the hidden state sequence exceeds a preset threshold, perform dimensionality reduction processing on the hidden state sequence through a principal component analysis tool, and obtain a dimensionality-reduced feature sequence; An eighth acquisition unit, configured to, for the dimensionality-reduced feature sequence, recombine the dimensionality-reduced feature sequence by using a sequence generation tool, and calculate the dependency relationship between features through a time window sliding method, so as to obtain a final time-dependent feature sequence.

[0016] Further, the processing module includes: A ninth acquisition unit, configured to acquire original feature data from the hidden state sequence, calculate the attention score of each feature by using an attention mechanism, and obtain the weight value of each feature through normalization processing of the attention score; A generation unit, configured to compare the weight value with a preset threshold, and if the weight value is greater than the preset threshold, retain the corresponding feature and generate a preliminary reduced feature set; A tenth acquisition unit, configured to, for the preliminary reduced feature set, perform dimensionality reduction processing by using a principal component analysis algorithm, and obtain a reduced feature set with a lower dimension by retaining the main variance direction; An eleventh acquisition unit, configured to perform sequence analysis on the reduced feature set, obtain an equally spaced arrangement result, generate an optimized data expression, and obtain a final reduced feature set.

[0017] The beneficial effects achieved by the present invention are as follows: The present invention provides a method and system for fault warning of a battery pack based on data analysis. Through linear interpolation and time series generation of multi-source asynchronous sampling data, combined with a long short-term memory network to extract time-dependent features, and using an attention mechanism to screen key features; on this basis, Bayesian inference is used to calculate the fault probability, and the warning threshold is dynamically adjusted to achieve hierarchical warning. The present invention effectively integrates multi-dimensional parameter information such as voltage, current, and temperature, fully excavates the data time series features, and improves the accuracy and timeliness of fault prediction; through the construction of a reduced feature set and dynamic threshold adjustment, the present invention can adapt to the complex and changeable operating environment of the power system, reduce the false alarm rate while ensuring the warning reliability, and provides a strong guarantee for the safe and stable operation of power equipment. Description of the Drawings

[0018] Figure 1 This is a flowchart showing an embodiment of a fault warning method for a battery pack based on data analysis according to the present invention. Detailed implementation manners

[0019] To better understand the above technical solution, the above technical solution will be described in detail below in conjunction with the accompanying drawings of the specification and specific implementation manners.

[0020] As Figure 1 shown, a first embodiment of the present invention proposes a fault warning method for a battery pack based on data analysis, including the following steps: Step S100: Obtain asynchronous sampling data of voltage, current, and temperature parameters from multiple data sources, and use a linear interpolation algorithm to map the asynchronous sampling data to a preset time grid to generate time series data.

[0021] Asynchronous sampling data refers to that during the data acquisition process, the sampling period or sampling frequency always remains fixed and does not change with the actual operating frequency of the system.

[0022] The linear interpolation algorithm is an interpolation method that constructs a linear relationship based on known data points and is used to estimate the value of any unknown point between two adjacent known points. The core idea of the linear interpolation algorithm is: assuming that the change trend between adjacent data points is linear, the interpolation calculation is completed by constructing a straight-line equation.

[0023] The preset time grid refers to a pre-set sequence of equally spaced time points, which is used to systematically plan the execution times of data acquisition, signal processing, or calculation tasks.

[0024] Time series data refers to a series of observations arranged in chronological order. Each data point is associated with a specific timestamp, and the time dimension is its core index. This type of data reflects the dynamic law of change over time by recording the states of variables at different time points.

[0025] Step S200: Input the time series data into a long short-term memory network, and obtain a hidden state sequence containing time-dependent features through forward propagation calculation.

[0026] The long short-term memory network (LSTM) is a special type of recurrent neural network (RNN) designed to solve the long-term dependence problem (i.e., gradient vanishing or explosion) of traditional RNNs. Its core idea is to achieve long-term memory and selective forgetting of information through a gating mechanism and a cell state.

[0027] Forward Propagation is the computational process in a neural network where data is passed layer by layer from the input layer to the output layer. The core objective of forward propagation is to generate a predicted output based on the current network parameters (weights, biases) and the input data.

[0028] Hidden State Sequence is a dynamic vector sequence composed of the hidden states at each time step in a recurrent neural network (RNN) and a long short-term memory network (LSTM) in chronological order. The core function of the hidden state sequence is to capture and transmit short-term dependencies in sequence data.

[0029] Step S300: Calculate the feature weights for the hidden state sequence using the attention mechanism. If the feature weight is greater than the preset threshold, retain the corresponding feature to form a reduced feature set.

[0030] Attention Mechanism is a strategy in machine learning that simulates the allocation of human cognitive resources. By dynamically assigning different weights to input information, the model can focus on the key parts when processing complex data. The core idea of the attention mechanism is to automatically filter and weight relevant features according to the current task requirements, suppressing the interference of irrelevant information.

[0031] Reduced Feature Set refers to a subset of features that are key and low in redundancy for the target task selected from the original data through dynamic weight assignment or dimensionality reduction. Step S400: Based on the reduced feature set, use the Bayesian inference algorithm to calculate the posterior probability distribution and determine the probability of a fault occurring.

[0032] Bayesian Inference Algorithm is a probability calculation method based on Bayes' theorem. It realizes statistical inference and model optimization by dynamically updating the hypothesis probability. The core of the Bayesian inference algorithm is to gradually correct the estimation of unknown parameters through the integration of prior knowledge and observed data, and finally output the posterior probability distribution as the decision-making basis.

[0033] Posterior Probability Distribution is a core concept in Bayesian statistics, referring to the result of updating the probability distribution of unknown parameters (or hypotheses) after observing the data. The essence of the posterior probability distribution is to quantify the uncertainty of parameters supported by data by integrating prior knowledge and observed data through Bayes' theorem.

[0034] Step S500: According to the probability of failure, a dynamic threshold adjustment algorithm is used to update the warning threshold. If the probability of failure exceeds the current warning threshold, a graded warning signal is triggered.

[0035] The dynamic threshold adjustment algorithm is a calculation method that automatically optimizes the decision boundary (threshold) based on the real-time data distribution or context environment. It aims to improve the robustness of the system in dynamic scenarios through feedback mechanisms or adaptive rules. The core of the dynamic threshold adjustment algorithm is to dynamically correct the threshold parameters by continuously monitoring key indicators (such as noise level, category imbalance, signal strength, etc.) to balance detection sensitivity and false positive rate. The ‌Graded Warning Signal‌ is a warning system that divides risks into multiple levels (usually four levels) based on the ‌harm ​​level‌, ‌urgency level‌ and ‌development trend‌ of meteorological disasters. It aims to guide the public to take differentiated defense measures through color identification (blue, yellow, orange, red) and quantitative standards. The core of the ‌Graded Warning Signal‌ is to achieve accurate matching of risk prediction and emergency response through clear level division.

[0036] Furthermore, the battery pack fault warning method based on data analysis proposed in this embodiment, step S100 includes: Step S110 , obtaining asynchronous sampling data of voltage parameters, current parameters, and temperature parameters from multiple data sources, and performing data acquisition using a preset sampling frequency to obtain an original asynchronous sampling data set.

[0037] Asynchronous sampling data acquisition is the core technology in multi-source sensor systems, which aims to obtain high-precision time series data from parameters such as voltage, current, temperature, etc. For example, in a battery pack monitoring scenario, the voltage sensor, current sensor, and temperature sensor collect data at different sampling frequencies, such as 100 times per second for the voltage sensor, 80 times per second for the current sensor, and 50 times per second for the temperature sensor.

[0038] Step S120: If the timestamps of the original asynchronous sampling data sets are not aligned, a timestamp alignment algorithm is used to calculate the difference between the timestamps of each data source and a preset time grid, determine the nearest grid point, and obtain a time-aligned sampling data set.

[0039] Due to differences in device hardware, the timestamps of the collected data are usually not aligned, making subsequent analysis difficult. This embodiment uses a preset sampling frequency, such as a uniform grid of 100 times per second, to collect the original asynchronously sampled data set. For example, voltage data may be recorded at timestamps of 1.01 seconds and 1.03 seconds, current data at 1.02 seconds and 1.04 seconds, and temperature data at 1.00 seconds and 1.05 seconds. Such timestamp deviations will result in reduced accuracy when data is fused.

[0040] In a possible implementation, the timestamp alignment algorithm maps data points to the nearest grid points by calculating the differences between the timestamps of each data source and a preset time grid.

[0041] Specifically, for the voltage data with a timestamp of 1.03 seconds, if the preset grid is 1.00 seconds, 1.01 seconds, 1.02 seconds, etc., then 1.03 seconds is mapped to the 1.03 - second grid point. For the temperature data with a timestamp of 1.05 seconds, it is mapped to the 1.05 - second grid point.

[0042] Step S130: For the sampled data set after time alignment, use the linear interpolation algorithm to calculate the interpolations of the voltage parameter, current parameter, and temperature parameter of each sampling point on the preset time grid, and obtain the interpolated time - series data set.

[0043] For the sampled data set after time alignment, the linear interpolation algorithm is used to calculate the values of each parameter on the preset time grid. For example, at the 1.02 - second grid point, the voltage data may be missing, but the voltage values at 1.01 seconds and 1.03 seconds are 220V and 221V respectively. Through linear interpolation, the voltage at 1.02 seconds can be estimated to be 220.5V. Similarly, the current and temperature parameters are interpolated in a similar way.

[0044] It should be noted that linear interpolation assumes that the parameter change is linear and is applicable to scenarios with smooth changes, such as the voltage fluctuation during the operation of a battery pack. The interpolated time - series data set contains complete voltage, current, and temperature values at grid points such as 1.00 seconds, 1.01 seconds, etc. This method significantly improves the integrity and consistency of the data, facilitating subsequent analysis.

[0045] Preferably, if a data point deviates too much from the grid point, such as exceeding 0.005 seconds, it can be marked as an abnormal point and excluded. This alignment method ensures the consistency of the data set in the time dimension and provides a reliable basis for subsequent interpolation. The beneficial effect is that the data set after time alignment can be directly used for multi - parameter fusion analysis, avoiding incorrect conclusions caused by time deviation.

[0046] Step S140: According to the interpolated time - series data set, determine whether the interpolation accuracy meets the preset threshold. If not, adjust the linear interpolation step size and recalculate the interpolation to obtain the final time - series data.

[0047] In one embodiment, the judgment of the interpolation accuracy is carried out by comparing the deviation between the interpolated value and the actual sampled value. For example, in a certain test, the interpolated voltage value is 220.5V at 1.02 seconds, while the actual sampled value is 220.4V, and the deviation is 0.1V, which is less than the preset threshold of 0.2V, so the accuracy meets the requirements. If the deviation exceeds the threshold, such as 0.3V, the linear interpolation step size is adjusted. For example, the interpolation step size is shortened from 0.01 second to 0.005 seconds, and the interpolation is recalculated to generate more refined time series data.

[0048] It can be understood that the step size adjustment increases the computational amount, but significantly improves the interpolation accuracy, especially in the scenario where parameters change rapidly.

[0049] The final time series data can be used for the state monitoring of the battery pack, such as detecting anomalies through the synchronous analysis of voltage and current, or predicting the overheating risk through temperature data. For example, the adjusted time series data can improve the accuracy of anomaly detection in the battery pack fault diagnosis. Compared with the unaligned original data, the interpolated data set is more consistent in the time dimension and the relationship between parameters is clearer. This method is not only applicable to the battery pack monitoring, but also can be extended to other industrial sensor systems, with universality.

[0050] The beneficial effect is that through time alignment and interpolation, the system can capture the parameter change trend more accurately and improve the fault warning ability.

[0051] Furthermore, the fault warning method for the battery pack based on data analysis proposed in this embodiment, step S200 includes: Step S210, obtain the original time series data from the time series database, perform standardization processing on the original time series data through a preprocessing tool, and calculate the normalized value of each data point by using the mean normalization method to obtain the standardized time series data.

[0052] Exemplarily, in the process of obtaining the original time series data from the time series database, the data acquisition of physical quantities such as voltage, current, and temperature is usually involved. Assume that in a certain battery pack monitoring system, the database stores the voltage and temperature data collected per second, but the data may have noise or missing due to different response times of the device sensors. The original data within a specific time period can be extracted through the query interface of the time series database. For example, the voltage data from April 1st to April 2nd, 2025, with one data point per second, a total of 86,400 points. Such data extraction provides the basis for subsequent processing.

[0053] In a possible implementation manner, when performing standardization processing on the original time series data, mean normalization is a commonly used method.

[0054] Assume that the voltage data ranges from 100 to 200 volts, with a mean of 150 volts and a standard deviation of 20 volts. Through mean normalization, each data point is subtracted by the mean of 150 and then divided by the standard deviation of 20, resulting in a normalized value range between -2.5 and 2.5. Such processing can eliminate the dimension difference and facilitate subsequent model training.

[0055] Step S220: According to the standardized time series data, input it into the pre-established long short-term memory network, and calculate the cell state and hidden state at each time step through the forward propagation algorithm to obtain a hidden state sequence containing time-dependent features.

[0056] For example, the mean of the temperature data may be 25 degrees Celsius and the standard deviation is 5 degrees Celsius. The normalized data distribution is more suitable for input into the neural network. Specifically, when the standardized time series data is input into the long short-term memory network, the network can capture the time-dependent relationship. Assume that the network contains 128 hidden units, and the input is the voltage and temperature data per second, with a time step of 10 seconds.

[0057] The long short-term memory network calculates the cell state at each time step through the forget gate, input gate, and output gate. For example, if the voltage suddenly changes to 160 volts at a certain time step, the network will judge whether this is a normal fluctuation or an abnormality based on the data trend in the previous 9 seconds. Therefore, the hidden state sequence contains time-dependent features and reflects the dynamic changes of the data.

[0058] Step S230: If the dimension of the hidden state sequence exceeds the preset threshold, then perform dimensionality reduction processing on the hidden state sequence through the principal component analysis tool to obtain the dimensionality-reduced feature sequence.

[0059] It should be noted that if the dimension of the hidden state sequence is too high, such as 128 dimensions, it may increase the computational complexity.

[0060] Through the dimensionality reduction processing by the principal component analysis tool, the main features can be retained. For example, it is found that the first 10 principal components explain 95% of the variance, so the sequence is reduced to 10 dimensions. The dimensionality-reduced feature sequence retains the key time-dependent information and at the same time reduces the computational burden of subsequent processing. For example, the feature changes during the voltage mutation can still be accurately reflected after dimensionality reduction.

[0061] In one embodiment, when using the sequence generation tool to reorganize the dimensionality-reduced feature sequence, the time window sliding method can further explore the dependency relationship between features.

[0062] Assume that the time window length is 5 seconds and the step size is 1 second. The sliding window will analyze the feature sequence second by second. For example, within a certain 5-second window, the voltage feature shows a gradually rising trend while the temperature feature remains stable, indicating that the device may be in a high-load state.

[0063] Step S240: For the dimension-reduced feature sequence, use a sequence generation tool to reorganize the dimension-reduced feature sequence, and calculate the dependency relationship between features through the time window sliding method to obtain the final time-dependent feature sequence.

[0064] By calculating the correlation of features within the window, a time-dependent feature sequence reflecting the device operating state can be generated. Such a sequence can provide clearer dynamic information for subsequent analysis.

[0065] Preferably, the above processing method significantly improves the usability of time series data through standardization, dimension reduction, and feature reorganization. For example, in a battery pack monitoring system, the generated feature sequence can be directly used for fault prediction or performance optimization. The implementation of each step is carried out around the characteristics of time series data, with strict logic and mutual support, ensuring the high efficiency and reliability of the entire process from data collection to feature generation.

[0066] Furthermore, the fault warning method for the battery pack based on data analysis proposed in this embodiment, step S300 includes: Step S310: Obtain the original feature data from the hidden state sequence, calculate the attention score of each feature using the attention mechanism, and obtain the weight value of each feature through normalizing the attention score.

[0067] In a possible implementation manner, when obtaining the original feature data from the hidden state sequence, it usually involves extracting features related to the time series from the output of the long short-term memory network. For example, assuming that the time series data in the battery pack data is being processed, the hidden state sequence may include features such as low voltage and high temperature changes.

[0068] The original feature data can be the hidden state vector at each time step, and the dimension may be relatively high, such as 100 dimensions. When extracting these features, it is necessary to ensure data integrity and avoid losing time-dependent information.

[0069] Specifically, when calculating the attention score of each feature using the attention mechanism, an importance weight will be assigned to each feature. For example, in the battery pack low voltage prediction scenario, the attention mechanism may focus on the features of recent low voltage and assign lower weights to earlier fluctuations. When implementing, a fully connected layer can be used to map the hidden state to the attention score, and then normalized through the softmax function to obtain the weight value between 0 and 1. Assuming that the attention score of a certain feature is 0.8, it indicates that it has a greater impact on the prediction result.

[0070] It should be noted that the normalization process ensures that the sum of the weights is 1, which is convenient for subsequent comparison.

[0071] Step S320: Compare the weight value with a preset threshold. If the weight value is greater than the preset threshold, retain the corresponding feature and generate a preliminary reduced feature set.

[0072] In one embodiment, by comparing the weight value with a preset threshold, features with weight values greater than the threshold are retained to generate a preliminary reduced feature set. Preferably, the threshold is set to 0.5, and features with weight values greater than 0.5 are retained. For example, among 100-dimensional features, there may be 60 features with weight values exceeding 0.5, and these features constitute the preliminary reduced feature set. This process reduces the interference of irrelevant features and improves the efficiency of data processing.

[0073] Step S330: For the preliminary reduced feature set, perform dimensionality reduction using the principal component analysis algorithm. By retaining the main variance direction, a reduced feature set with a lower dimension is obtained.

[0074] It can be understood that when performing dimensionality reduction on the preliminary reduced feature set using the principal component analysis algorithm, the goal is to retain the main variance direction. For example, a 60-dimensional preliminary reduced feature set may contain redundant information. Through principal component analysis, the dimension can be reduced to 20 dimensions, retaining 90% of the variance. In implementation, calculate the covariance matrix of the features, extract the first few principal components, and ensure that the core information is retained. This dimensionality reduction method can effectively reduce the computational complexity in the time series analysis of battery packs. For example, when performing sequence analysis on the reduced feature set to obtain equally spaced arrangement results, the temporal relationship between features can be analyzed through the sliding window method.

[0075] Assume the window size is 5 time steps, analyze the correlation of features within the window, and generate an equally spaced feature sequence. This method is suitable for capturing short-term trends of battery pack failures. The final reduced feature set can be used as the input for subsequent prediction models.

[0076] Step S340: Through sequence analysis of the reduced feature set, obtain equally spaced arrangement results, generate an optimized data representation, and obtain the final reduced feature set.

[0077] In one possible implementation, the optimized data representation is generated through the above steps and is applicable to the battery pack time series analysis scenario. For example, the final reduced feature set may contain 20-dimensional features, covering key information such as temperature trends and volatility. This feature set can be directly input into the prediction model to improve the model's ability to capture future fault predictions.

[0078] Furthermore, for the battery pack fault warning method based on data analysis proposed in this embodiment, step S400 includes: Step S410: Obtain a feature set from the training data, and use a statistical analysis tool to calculate the conditional probability of each feature to obtain the conditional probability distribution corresponding to the reduced feature set.

[0079] Exemplarily, in the field of time series analysis of battery pack fault warnings, when obtaining a feature set from training data, it usually involves extracting features related to the behavior of the battery pack from historical battery pack fault warning data. For example, when processing battery pack data, the feature set may include voltage, current, temperature parameters, etc. When obtaining these features, it is necessary to ensure that the data covers a sufficient time span, such as the battery pack data for the past year, to capture the fault patterns of the battery pack.

[0080] Specifically, through a data cleaning tool, the daily voltage, current, and temperature parameters can be extracted from the original battery pack data records to form an initial feature set containing hundreds of features.

[0081] In a possible implementation, when using a statistical analysis tool to calculate the conditional probability of each feature, a probability density estimation method can be used. For example, for the features of the daily voltage, current, and temperature parameters of the battery pack, analyze their distribution under specific conditions, such as under low voltage conditions.

[0082] Assume that the conditional probability distribution of the battery pack under low voltage conditions shows a normal distribution. When calculating, the historical data can be fitted through statistical software to obtain the conditional probability distribution of each feature. This method ensures that the feature importance is quantified and provides a basis for subsequent analysis.

[0083] Step S420: According to the conditional probability distribution, combine the pre-established model parameters through the Bayesian inference algorithm to calculate the posterior probability and obtain the posterior probability distribution.

[0084] It should be noted that when calculating the posterior probability through the Bayesian inference algorithm, it is particularly crucial to combine the pre-established model parameters. For example, the pre-established model may be a classifier based on historical data, including parameters under low voltage conditions. Assume that the model parameters indicate that the low voltage probability is 0.6. Combining the conditional probability of daily average low voltage, the posterior probability distribution can be derived through the Bayesian formula. For example, the posterior probability of a certain battery pack under the current low voltage shows that the low voltage probability is 0.75. This posterior probability distribution provides a reliable basis for fault prediction.

[0085] Step S430: If the posterior probability distribution is known, then for the fault types in the real-time data, use a logical judgment tool to analyze the posterior probability distribution, judge the probability of fault occurrence, and determine the fault probability value.

[0086] Specifically, for the fault types in real-time data, such as low battery pack voltage, when using a logical judgment tool to analyze the posterior probability distribution, a probability threshold can be set. For example, if the posterior probability is greater than 0.7, it is considered that there is a risk of abnormal fluctuation. Suppose the posterior probability of a low voltage of a certain battery pack in real-time data is 0.8, and the logical judgment tool will mark it as a high risk. When implementing, the probability value can be compared through a rule engine to quickly output the fault probability value.

[0087] Step S440: Through a prediction result generation tool, associate the fault probability value with the fault type to generate a prediction result and obtain a fault prediction report.

[0088] In one embodiment, when associating the fault probability value with the fault type through a prediction result generation tool, a structured prediction report can be generated. For example, if the fault type is "low battery pack voltage" and the fault probability value is 0.8, the prediction result generation tool will associate the two to generate a report including a timestamp, a risk category, and a risk level.

[0089] Preferably, the report is presented in a visual form, such as a battery pack risk heat map, for easy understanding by users.

[0090] It can be understood that this method forms a complete battery pack data analysis process from feature extraction to probability analysis and then to result generation. Each step supports each other, ensuring the accuracy and practicality of the prediction. For example, the conditional probability distribution provides a data basis for Bayesian inference, and the posterior probability distribution provides a basis for logical judgment. The finally generated prediction report provides an intuitive risk assessment for users. This strict logical relationship improves the analysis efficiency and reliability.

[0091] Furthermore, for the battery pack fault warning method based on data analysis proposed in this embodiment, step S500 includes: Step S510: Obtain the fault probability value from the real-time data stream, and perform standardization processing on the fault probability value through a data processing tool to obtain a standardized fault probability value.

[0092] In a possible implementation manner, the core of obtaining the fault probability value from the real-time data stream lies in ensuring the real-time and accurate data collection. For example, in a battery pack monitoring scenario, sensors collect battery pack voltage, current, and temperature parameters in real time, and a data processing tool extracts the fault probability value.

[0093] Suppose the voltage of a battery pack of a certain device is abnormal at a specific time point, and an initial fault probability value of 0.85 is generated in the data stream. The standardization processing maps this value to a unified interval for subsequent analysis.

[0094] Preferably, linear normalization can be used for standardization to convert probability values into the range of 0 to 1, eliminating the influence of dimension. For example, 0.85 may become 0.75 after standardization, ensuring data comparability with other devices. This method can effectively improve data consistency and provide reliable input for subsequent threshold adjustment.

[0095] Step S520: According to the standardized fault probability value, use a dynamic threshold adjustment algorithm in combination with a preset threshold update rule to calculate a new warning threshold and determine the updated warning threshold.

[0096] Specifically, the dynamic threshold adjustment algorithm combines with the preset threshold update rule to calculate a new warning threshold based on the standardized fault probability value. It can be understood that the dynamic threshold takes into account the fluctuations in the operating state of the battery pack. For example, when the battery pack is operating at a low voltage, the tolerance of the battery pack life may decrease, and the algorithm will dynamically adjust the threshold from 0.7 to 0.8 according to historical data and the current standardized probability value of 0.75.

[0097] Exemplarily, the adjustment rule may refer to the fault probability trend in the past 24 hours. If the trend is rising, the threshold is tightened. This way can flexibly adapt to changes in the device state and improve the pertinence of early warning.

[0098] Step S530: If the standardized fault probability value exceeds the updated warning threshold, use a logical judgment tool to generate a corresponding graded warning level in combination with the fault type and a preset grading rule, and obtain the graded warning level.

[0099] In one embodiment, if the standardized fault probability value exceeds the updated warning threshold, the logical judgment tool combines the fault type and the grading rule to generate a graded warning level. For example, the standardized probability value of 0.75 exceeds the threshold of 0.7. The logical judgment tool analyzes that the fault type is "low battery pack voltage" and classifies it as "medium" in the three-level warning according to the preset rule.

[0100] It should be noted that the grading rule may be based on the severity and influence range of the fault. For example, low battery pack voltage may affect the device life but does not cause immediate shutdown, so it is classified as medium. This grading method is convenient for maintenance personnel to quickly judge the fault priority.

[0101] Step S540: According to the graded warning level, use a signal generation tool to associate the graded warning level with a preset signal template to generate a warning signal and obtain a triggered warning signal.

[0102] For example, a signal generation tool associates graded warning levels with preset signal templates to generate warning trigger signals. In one possible implementation, the medium warning level corresponds to a yellow flashing signal. The signal generation tool maps "medium" to "yellow flashing + beeping sound" according to the template and displays it through the monitoring interface. Exemplarily, the signal may also include the fault type and suggested measures. For example, charging suggestion: when "low battery pack voltage" is detected, prompt the user to charge in time to avoid over-discharge; stop charge and discharge suggestion: when a fault trigger of level two or above is detected, forcefully stop charge and discharge to protect the battery (such as cutting off the high-voltage power supply). This method intuitively conveys warning information and facilitates quick response.

[0103] Preferably, the signal template supports customization. For example, the color or beeping sound intensity can be adjusted according to the requirements of the battery pack. This flexibility enhances the applicability of the system.

[0104] It can be understood that each of the above links is interlocked, forming a complete link from data acquisition to signal generation. For example, standardization processing ensures data comparability, dynamic thresholds adapt to the operating state, graded warnings clarify priorities, and signal generation intuitively conveys information. This logical progression ensures the accuracy and practicality of fault warnings. Especially in industrial scenarios, it can effectively improve the efficiency and response speed of battery pack management.

[0105] The present invention relates to a fault warning system for a battery pack based on data analysis, which is used to implement the above-mentioned fault warning method for a battery pack based on data analysis. The fault warning system for a battery pack based on data analysis includes a generation module, an acquisition module, a processing module, a determination module, and a trigger module. Among them, the generation module is used to obtain asynchronous sampling data of voltage, current, and temperature parameters from multiple data sources, and map the asynchronous sampling data to a preset time grid using a linear interpolation algorithm to generate time series data; the acquisition module is used to input the time series data into a long short-term memory network and obtain a hidden state sequence containing time-dependent features through forward propagation calculation; the processing module is used to calculate feature weights for the hidden state sequence using an attention mechanism. If the feature weight is greater than a preset threshold, the corresponding feature is retained to form a refined feature set; the determination module is used to calculate the posterior probability distribution based on the refined feature set using a Bayesian inference algorithm to determine the probability of a fault occurring; the trigger module is used to update the warning threshold using a dynamic threshold adjustment algorithm according to the probability of a fault occurring. If the fault probability exceeds the current warning threshold, a graded warning signal is triggered.

[0106] Furthermore, for the fault warning system of the battery pack based on data analysis provided in this embodiment, the generation module includes a first acquisition unit, a second acquisition unit, a third acquisition unit, and a fourth acquisition unit. Among them, the first acquisition unit is used to acquire asynchronous sampling data of voltage parameters, current parameters, and temperature parameters from multiple data sources, perform data acquisition at a preset sampling frequency, and obtain an original asynchronous sampling data set; the second acquisition unit is used to, if the timestamps of the original asynchronous sampling data set are not aligned, use a timestamp alignment algorithm to determine the nearest grid point by calculating the differences between the timestamps of each data source and the preset time grid, and obtain a sampling data set with aligned timestamps; the third acquisition unit is used to, for the sampling data set with aligned timestamps, use a linear interpolation algorithm to calculate the interpolations of the voltage parameters, current parameters, and temperature parameters at each sampling point on the preset time grid, and obtain an interpolated time series data set; the fourth acquisition unit is used to, according to the interpolated time series data set, determine whether the interpolation accuracy meets a preset threshold. If not, adjust the linear interpolation step size and recalculate the interpolation to obtain the final time series data.

[0107] Preferably, for the fault warning system of the battery pack based on data analysis provided in this embodiment, the acquisition module includes a fifth acquisition unit, a sixth acquisition unit, a seventh acquisition unit, and an eighth acquisition unit. Among them, the fifth acquisition unit is used to acquire original time series data from a time series database, perform standardization processing on the original time series data through a preprocessing tool, and calculate the normalized value of each data point using the mean normalization method to obtain a standardized time series data; the sixth acquisition unit is used to, according to the standardized time series data, input it into a pre-established long short-term memory network, and calculate the cell state and hidden state at each time step through the forward propagation algorithm to obtain a hidden state sequence containing time-dependent features; the seventh acquisition unit is used to, if the dimension of the hidden state sequence exceeds a preset threshold, perform dimensionality reduction processing on the hidden state sequence through a principal component analysis tool to obtain a dimensionality-reduced feature sequence; the eighth acquisition unit is used to, for the dimensionality-reduced feature sequence, use a sequence generation tool to reorganize the dimensionality-reduced feature sequence, and calculate the dependence relationship between features through the time window sliding method to obtain the final time-dependent feature sequence.

[0108] Furthermore, in the fault warning system of the battery pack based on data analysis provided in this embodiment, the processing module includes a ninth acquisition unit, a generation unit, a tenth acquisition unit, and an eleventh acquisition unit. Among them, the ninth acquisition unit is used to obtain the original feature data from the hidden state sequence, calculate the attention score of each feature using the attention mechanism, and obtain the weight value of each feature through normalizing the attention score; the generation unit is used to compare the weight value with a preset threshold. If the weight value is greater than the preset threshold, the corresponding feature is retained to generate a preliminary reduced feature set; the tenth acquisition unit is used to perform dimensionality reduction processing on the preliminary reduced feature set using the principal component analysis algorithm, and obtain a reduced feature set with a lower dimension by retaining the main variance direction; the eleventh acquisition unit is used to obtain an equidistant arrangement result through sequence analysis of the reduced feature set, generate an optimized data representation, and obtain a final reduced feature set.

[0109] Compared with the prior art, the fault warning method and system of the battery pack based on data analysis provided in this embodiment generate linear interpolation and time series through multi-source asynchronous sampling data, extract time-dependent features by combining long short-term memory networks, and use the attention mechanism to screen key features; on this basis, Bayesian inference is used to calculate the fault probability, and the warning threshold is dynamically adjusted to achieve hierarchical warning. This embodiment effectively integrates multi-dimensional parameter information such as voltage, current, and temperature, fully excavates the data time series features, and improves the accuracy and timeliness of fault prediction; through the construction of a reduced feature set and dynamic threshold adjustment, this embodiment can adapt to the complex and changeable operation environment of the power system, reduce the false alarm rate while ensuring the warning reliability, and provide a strong guarantee for the safe and stable operation of power equipment.

[0110] Although the preferred embodiments of the present invention have been described, those skilled in the art can make additional changes and modifications once they know the basic creative concept. Therefore, the appended claims are intended to be construed as including the preferred embodiments as well as all changes and modifications falling within the scope of the present invention. Obviously, those skilled in the art can make various changes and variations to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalent technologies, the present invention is also intended to include these modifications and variations.

Claims

1. A battery pack fault warning method based on data analysis, characterized in that: The following steps are involved: Acquire asynchronous sampling data of voltage, current and temperature parameters from multiple data sources, and map the asynchronous sampling data to a preset time grid using a linear interpolation algorithm to generate time series data; Inputting the time series data into a long short-term memory network, and obtaining a hidden state sequence containing time-dependent features through forward propagation calculation; The attention mechanism is used to calculate the feature weight of the hidden state sequence, and if the feature weight is greater than a preset threshold, the corresponding feature is retained to form a simplified feature set; Based on the simplified feature set, a Bayesian inference algorithm is used to calculate the posterior probability distribution to determine the probability of fault occurrence; According to the failure probability, a dynamic threshold adjustment algorithm is used to update the warning threshold. If the failure probability exceeds the current warning threshold, a graded warning signal is triggered.

2. The battery pack fault warning method based on data analysis as claimed in claim 1, characterized in that: The step of acquiring asynchronous sampling data of voltage, current and temperature parameters from multiple data sources, mapping the asynchronous sampling data to a preset time grid using a linear interpolation algorithm, and generating time series data comprises: Acquire asynchronous sampling data of voltage parameters, current parameters, and temperature parameters from multiple data sources, use a preset sampling frequency to collect data, and obtain an original asynchronous sampling data set; If the timestamps of the original asynchronous sampled data sets are not aligned, a timestamp alignment algorithm is used to calculate the difference between the timestamps of each data source and a preset time grid, determine the nearest grid point, and obtain a time-aligned sampled data set; For the time-aligned sampling data set, a linear interpolation algorithm is used to calculate the interpolation values ​​of the voltage parameters, current parameters, and temperature parameters of each sampling point on the preset time grid to obtain the interpolated time series data set; According to the interpolated time series data set, it is judged whether the interpolation accuracy meets the preset threshold. If not, the linear interpolation step size is adjusted and the interpolation is recalculated to obtain the final time series data.

3. The battery pack fault warning method based on data analysis as claimed in claim 1, characterized in that: The steps of inputting the time series data into a long short-term memory network and obtaining a hidden state sequence containing time-dependent features through forward propagation calculation include: Obtaining original time series data from a time series database, performing standardization processing on the original time series data through a preprocessing tool, and calculating a normalized value of each data point using a mean normalization method to obtain standardized time series data; According to the standardized time series data, the pre-established long short-term memory network is input, and the unit state and hidden state of each time step are calculated by the forward propagation algorithm to obtain the hidden state sequence containing time-dependent features; If the dimension of the hidden state sequence exceeds a preset threshold, the hidden state sequence is subjected to dimensionality reduction processing by a principal component analysis tool to obtain a feature sequence after dimensionality reduction; For the feature sequence after dimensionality reduction, a sequence generation tool is used to reorganize the feature sequence after dimensionality reduction, and the dependency relationship between features is calculated through a time window sliding method to obtain a final time-dependent feature sequence.

4. The battery pack fault warning method based on data analysis as claimed in claim 1, characterized in that: The step of using an attention mechanism to calculate feature weights for the hidden state sequence, and if the feature weights are greater than a preset threshold, retaining corresponding features to form a simplified feature set includes: Acquire original feature data from the hidden state sequence, calculate the attention score of each feature using an attention mechanism, and obtain a weight value of each feature by normalizing the attention score; Compare the weight value with a preset threshold value, and if the weight value is greater than the preset threshold value, retain the corresponding feature to generate a preliminary simplified feature set; A principal component analysis algorithm is used to perform dimensionality reduction processing on the preliminary simplified feature set, and a simplified feature set with a lower dimension is obtained by retaining the main variance direction; By performing sequence analysis on the simplified feature set, an equidistant arrangement result is obtained, an optimized data expression is generated, and a final simplified feature set is obtained.

5. The battery pack fault warning method based on data analysis as claimed in claim 1, characterized in that: Based on the simplified feature set, the steps of calculating the posterior probability distribution using the Bayesian inference algorithm and determining the probability of a fault occurrence include: Acquire a feature set from the training data, use a statistical analysis tool to calculate the conditional probability of each feature, and obtain a conditional probability distribution corresponding to the reduced feature set; According to the conditional probability distribution, the posterior probability is calculated by combining the pre-established model parameters through the Bayesian inference algorithm to obtain the posterior probability distribution; If the posterior probability distribution is known, a logical judgment tool is used to analyze the posterior probability distribution for the fault type in the real-time data, judge the probability of the fault, and determine the fault probability value; The fault probability value is associated with the fault type through a prediction result generating tool, a prediction result is generated, and a fault prediction report is obtained.

6. The battery pack fault warning method based on data analysis as claimed in claim 1, characterized in that: According to the probability of the failure, a dynamic threshold adjustment algorithm is used to update the warning threshold. If the probability of the failure exceeds the current warning threshold, the step of triggering a graded warning signal includes: Obtaining the fault probability value from the real-time data stream, and standardizing the fault probability value through a data processing tool to obtain a standardized fault probability value; According to the standardized fault probability value, a dynamic threshold adjustment algorithm is used in combination with a preset threshold update rule to calculate a new warning threshold and determine an updated warning threshold; If the standardized fault probability value exceeds the updated warning threshold, a corresponding graded warning level is generated by combining the fault type with the preset classification rules through a logical judgment tool to obtain a graded warning level; According to the graded warning level, a signal generation tool is used to associate the graded warning level with a preset signal template, generate a warning signal, and obtain a trigger warning signal.

7. A battery pack fault warning system based on data analysis, used to implement the battery pack fault warning method based on data analysis as claimed in any one of claims 1 to 6, characterized in that: The battery pack fault warning system based on data analysis includes: A generation module, used to obtain asynchronous sampling data of voltage, current and temperature parameters from multiple data sources, and map the asynchronous sampling data to a preset time grid using a linear interpolation algorithm to generate time series data; An acquisition module is used to input the time series data into a long short-term memory network, and obtain a hidden state sequence containing time-dependent features through forward propagation calculation; A processing module, configured to calculate feature weights of the hidden state sequence using an attention mechanism, and if the feature weight is greater than a preset threshold, retain the corresponding feature to form a simplified feature set; A determination module, configured to calculate a posterior probability distribution based on the simplified feature set using a Bayesian inference algorithm to determine a probability of a fault occurrence; The trigger module is used to update the warning threshold value according to the probability of the fault occurrence by adopting a dynamic threshold adjustment algorithm, and trigger a graded warning signal if the probability of the fault exceeds the current warning threshold value.

8. The battery pack fault warning system based on data analysis as claimed in claim 7, characterized in that: The generation module comprises: A first acquisition unit is used to acquire asynchronous sampling data of voltage parameters, current parameters, and temperature parameters from multiple data sources, and to acquire data using a preset sampling frequency to obtain an original asynchronous sampling data set; A second acquisition unit is configured to, if the timestamps of the original asynchronous sampling data set are not aligned, use a timestamp alignment algorithm to determine the nearest grid point by calculating the difference between the timestamps of each data source and a preset time grid, and obtain a sampling data set after time alignment; A third acquisition unit is used to calculate the interpolation values ​​of the voltage parameter, the current parameter, and the temperature parameter of each sampling point on the preset time grid by using a linear interpolation algorithm for the time-aligned sampling data set to obtain an interpolated time series data set; The fourth acquisition unit is used to determine whether the interpolation accuracy meets a preset threshold according to the interpolated time series data set. If not, the linear interpolation step size is adjusted and the interpolation is recalculated to obtain the final time series data.

9. The battery pack fault warning system based on data analysis as claimed in claim 7, characterized in that: The acquisition module comprises: A fifth acquisition unit is used to acquire original time series data from a time series database, perform standardization on the original time series data through a preprocessing tool, calculate a normalized value of each data point using a mean normalization method, and obtain standardized time series data; A sixth acquisition unit is used to input the pre-established long short-term memory network according to the standardized time series data, calculate the unit state and hidden state of each time step through the forward propagation algorithm, and obtain a hidden state sequence containing time-dependent features; a seventh acquisition unit, configured to, if the dimension of the hidden state sequence exceeds a preset threshold, perform dimensionality reduction processing on the hidden state sequence by a principal component analysis tool to obtain a feature sequence after dimensionality reduction; The eighth acquisition unit is used to reorganize the feature sequence after dimensionality reduction by using a sequence generation tool, calculate the dependency relationship between features by a time window sliding method, and obtain a final time-dependent feature sequence.

10. The battery pack fault warning system based on data analysis as claimed in claim 7, characterized in that: The processing module comprises: a ninth acquisition unit, configured to acquire original feature data from the hidden state sequence, calculate an attention score of each feature by using an attention mechanism, and obtain a weight value of each feature by normalizing the attention score; A generating unit, configured to compare the weight value with a preset threshold value, and if the weight value is greater than the preset threshold value, retain the corresponding feature to generate a preliminary reduced feature set; a tenth acquisition unit, configured to perform dimensionality reduction processing on the preliminary simplified feature set by using a principal component analysis algorithm, and obtain a simplified feature set with a lower dimension by retaining the main variance direction; The eleventh acquisition unit is used to perform sequence analysis on the simplified feature set to obtain an equidistant arrangement result, generate an optimized data expression, and obtain a final simplified feature set.

Images