Charging pile fault prediction method and system based on multi-modal time series data analysis

Abstract

The invention discloses a charging pile fault prediction method based on multi-modal time series data analysis, and the method comprises the steps: obtaining the original operation data of a charging pile, and carrying out the preprocessing of the original operation data, and obtaining an initial data set; according to the initial data set, dynamic change indexes are extracted, and a time sequence feature set is obtained; extracting feature vectors according to the time sequence feature set to obtain an independent feature representation set; according to the independent feature representation set, performing weighted fusion to obtain a fusion feature result; in combination with a preset double-layer full-connection neural network, performing fault prediction to obtain a preliminary fault prediction result; according to the preliminary fault prediction result, performing global search and adjustment on a preset double-layer full-connection neural network to obtain an optimized parameter combination; according to the optimization parameter combination, configuration information is updated, training is executed, and prediction confidence is obtained; and if the prediction confidence is lower than a preset prediction threshold, returning to obtain an optimization parameter combination for iteration, and otherwise, outputting a final fault prediction result.

Description

Technical Field

[0001] This invention relates to the field of charging pile fault prediction technology, and in particular to a charging pile fault prediction method and system based on multimodal time series data analysis. Background Technology

[0002] Currently, with the rapid development of the new energy vehicle industry, charging piles, as core nodes in the electric vehicle energy replenishment network, directly affect the safety and availability of the entire energy system. Charging piles are located outdoors or in semi-enclosed environments for extended periods, facing challenges such as high-frequency use, fluctuating electrical loads, and complex environmental conditions, making them highly susceptible to failure. To achieve long-term reliable operation of the equipment, the concept of Predictive Fault and Health Management (PHM) has been gradually introduced internationally, aiming to achieve early identification, trend prediction, and proactive maintenance during equipment operation, thereby improving the overall stability and intelligent operation and maintenance level of the system.

[0003] In existing technologies, PHM applications for charging piles are still in the exploratory stage. Current mainstream research mostly adopts single-modal feature analysis or empirical judgment based on fixed rules, lacking the ability to deeply fuse and process multi-source heterogeneous data. During actual operation, charging piles generate multi-modal time-series data such as voltage, current, and temperature. These data exhibit strong nonlinearity, dynamic correlation, and time-series dependence, making it difficult to reflect the true health status of the equipment using only a single channel. Furthermore, in the fault prediction modeling process, traditional methods rely heavily on manual experience or static optimization strategies in key stages such as feature modeling, parameter configuration, and model training, easily getting trapped in local optima and failing to meet the high-precision prediction requirements of complex application scenarios. This problem is particularly prominent in the parameter adaptability and real-time response capability of PHM systems.

[0004] Therefore, existing technologies cannot effectively fuse multimodal time series data and adaptively optimize prediction models, making it difficult to meet the actual needs of charging pile PHM. Summary of the Invention

[0005] This invention provides a method and system for predicting charging pile faults based on multimodal time series data analysis, so as to achieve effective fusion of multimodal time series data and adaptive optimization of prediction models, and meet the needs of high-precision pre-diagnosis and health management in charging pile PHM applications.

[0006] Firstly, in order to solve the above-mentioned technical problems, the present invention provides a charging pile fault prediction method based on multimodal time series data analysis, comprising: Obtain the raw operating data of the charging piles and perform preprocessing on the raw operating data to obtain an initial data set with a unified format; Based on the initial data set, dynamic change indicators for each time period are extracted to obtain a time-series feature set; Based on the time-series feature set, feature vectors of each modality data are extracted to obtain an independent feature representation set; Based on the set of independent feature representations, a weighted fusion operation is performed to obtain a fusion feature result with multi-dimensional information correlation. Based on the fusion feature results, combined with a preset two-layer fully connected neural network, fault trend prediction inference is performed to obtain preliminary fault prediction results. Based on the preliminary fault prediction results, a global search and adjustment are performed on the preset two-layer fully connected neural network to obtain an optimized parameter combination. Based on the optimized parameter combination, the configuration information of the preset two-layer fully connected neural network is updated, and training is performed on the fused feature results to obtain the prediction confidence. If the prediction confidence is lower than the preset prediction threshold, the process of iterating through the step of performing a global search and adjustment on the preset two-layer fully connected neural network based on the fault prediction result to obtain an optimized parameter combination is returned; otherwise, the final fault prediction result is output.

[0007] Preferably, the step of acquiring the original operating data of the charging pile and performing preprocessing on the original operating data to obtain an initial data set with a unified format includes: Add timestamps to the original running data to obtain an original record set with a time-series structure; Based on the original record set, the data of different modalities are processed to unify the units and missing values ​​are filled in to obtain an intermediate dataset with a consistent format. The intermediate dataset is anomaly-marked, and out-of-range abnormal data points are removed or labeled to obtain a filtered dataset containing anomaly information; Based on the filtered dataset, the data is segmented and organized into time window segments to obtain an initial dataset with a unified format.

[0008] Preferably, the step of extracting dynamic change indicators for each time period based on the initial data set to obtain a time-series feature set includes: The initial data set is segmented by combining a preset time window interval to obtain a segmented time sequence set; Dynamic feature analysis is performed on the modal data within the time series fragment set to extract the rate of change, extreme value range, and fluctuation frequency, resulting in a dynamic feature dataset with changing trends. Based on the dynamic feature dataset, time-dependent segments that exceed a preset dynamic threshold are identified and anomaly marking is performed to obtain a feature segment set containing anomaly markers. The trend consistency adjustment is performed on the feature fragment set to obtain the optimized time series feature set.

[0009] Preferably, the step of extracting feature vectors from each modality of data based on the temporal feature set to obtain an independent feature representation set includes: Based on the time-series feature set, current, voltage, and temperature are extracted to obtain modal data; Based on the modal data, perform high-dimensional mapping to expand the modal feature dimensions and obtain a high-dimensional feature dataset; Based on the high-dimensional feature dataset, abnormal feature fragments exceeding a preset abnormal threshold are identified and labeled to obtain a labeled feature set; Based on the set of labeled features, the modal data is calibrated to obtain an independent feature representation set.

[0010] Preferably, the step of performing a weighted fusion operation based on the set of independent feature representations to obtain a fusion feature result with multi-dimensional information correlation includes: Based on the set of independent feature representations, the weight coefficients of each modality feature vector are calculated and then combined in a weighted manner to obtain a weighted feature matrix. Based on the weighted feature matrix, dimensional unification and time alignment processing are performed to calibrate feature deviations between different modalities, resulting in a fusion feature result with multi-dimensional information correlation.

[0011] Preferably, the step of performing fault trend prediction inference based on the fused feature results and combining them with a preset two-layer fully connected neural network to obtain preliminary fault prediction results includes: Based on the fusion feature results, extract the feature subsets related to time-series dependencies, classify and organize them to obtain the classified feature groups; Based on the feature grouping, the correlation strength between each feature is calculated, feature combinations whose dynamic changes exceed a preset abnormal threshold are identified, and abnormal marking is performed to obtain abnormal feature groups; Based on the abnormal features grouping, compare with known fault modes in preset historical data, extract key mode fragments, and determine the mode features associated with the fault. The pattern features are input into a preset two-layer fully connected neural network for training and inference to obtain preliminary fault prediction results.

[0012] Preferably, the step of performing a global search and adjustment on the preset two-layer fully connected neural network based on the preliminary fault prediction results to obtain an optimized parameter combination includes: Based on the preliminary fault prediction results, the uncertainty distribution of the prediction output is analyzed, and the output subset related to abnormal fluctuations is extracted to obtain the abnormal distribution range. Based on the range of the abnormal distribution, identify the key parameter combinations in the current model parameters that are related to the local optimum, and obtain the subset of parameters to be optimized; Based on the subset of parameters to be optimized, a global search is performed to adjust the parameter weights, optimize the correlation configuration between parameters, and obtain an optimized parameter combination.

[0013] Secondly, the present invention provides a charging pile fault prediction system based on multimodal time series data analysis, comprising: The data acquisition module is used to acquire the original operating data of the charging pile and perform preprocessing on the original operating data to obtain an initial data set with a unified format. The time-series feature module is used to extract dynamic change indicators within each time period based on the initial data set to obtain a time-series feature set; An independent feature module is used to extract feature vectors of each modality data based on the time-series feature set to obtain an independent feature representation set; The feature fusion module is used to perform a weighted fusion operation based on the set of independent feature representations to obtain a fusion feature result with multi-dimensional information correlation. The preliminary prediction module is used to perform fault trend prediction inference based on the fused feature results and combined with a preset two-layer fully connected neural network to obtain preliminary fault prediction results. The parameter optimization module is used to perform a global search and adjustment on the preset two-layer fully connected neural network based on the preliminary fault prediction results to obtain an optimized parameter combination. The configuration update module is used to update the configuration information of the preset two-layer fully connected neural network according to the optimized parameter combination, and to perform training on the fused feature results to obtain the prediction confidence. The prediction and judgment module is used to iterate the steps of performing a global search and adjustment on the preset two-layer fully connected neural network to obtain an optimized parameter combination based on the fault prediction result if the prediction confidence is lower than a preset prediction threshold; otherwise, it outputs the final fault prediction result.

[0014] Thirdly, the present invention also provides an electronic device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor executes the computer program to implement the charging pile fault prediction method based on multimodal time series data analysis as described above.

[0015] Fourthly, the present invention also provides a computer-readable storage medium comprising a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to execute the charging pile fault prediction method based on multimodal time-series data analysis described above.

[0016] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention constructs a unified processing flow for multimodal time-series data such as current, voltage, and temperature, sequentially performing time-series feature extraction, modal feature separation, and high-dimensional mapping, and weightedly fusing different modal features, effectively solving the problems of isolated processing of multi-source data and insufficient feature expression capabilities in existing technologies. Because anomaly identification and trend correction mechanisms are introduced during feature extraction, the model's response to sudden states and boundary conditions is enhanced, realizing dynamic modeling and deep behavioral characterization of equipment operating status. Therefore, it can improve the sensitivity to potential fault characteristics, providing more timely and accurate front-end data support for PHM (Predictive Health Management).

[0017] (2) In the fault prediction modeling stage, this invention adopts a well-defined two-layer fully connected prediction structure and introduces key pattern recognition and fault precursor extraction mechanisms, enabling the model to focus on feature combinations strongly correlated with historical patterns. By constructing a mapping relationship between feature groups and historical patterns, the model's ability to identify fault trends under nonlinear and variable operating conditions is further improved. Compared with traditional models based on fixed feature inputs, the prediction module of this invention can adapt to the state characteristics under different working stages, realizing the diverse expansion of fault pre-diagnosis logic, thereby improving the prediction breadth and coverage in the health management process.

[0018] (3) To address the issues of local optima risk and insufficient generalization ability in prediction models, this invention introduces an iterative feedback mechanism driven by confidence, combining parameter search and configuration update operations to form a closed-loop optimization process. This scheme not only supports global search of the parameter space but also automatically triggers a new round of training when the model confidence is insufficient, ensuring that the model always outputs prediction results in a high-reliability state. Experiments show that this mechanism can effectively reduce the influence of the model's sensitivity to initial parameters, improve the prediction stability and robustness in complex and variable environments, and effectively enhance the health status judgment accuracy and adaptive performance of the PHM system. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the charging pile fault prediction method based on multimodal time series data analysis provided in the first embodiment of the present invention; Figure 2 This is a schematic diagram of the charging pile fault prediction system based on multimodal time series data analysis provided in the second embodiment of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Reference Figure 1 The first embodiment of the present invention provides a charging pile fault prediction method based on multimodal time series data analysis, including the following steps: S11, Obtain the original operating data of the charging pile, and perform preprocessing on the original operating data to obtain an initial data set with a unified format; S12, Based on the initial data set, extract dynamic change indicators for each time period to obtain a time-series feature set; S13, Based on the time-series feature set, extract the feature vectors of each modality data to obtain an independent feature representation set; S14, Based on the set of independent feature representations, perform a weighted fusion operation to obtain a fusion feature result with multi-dimensional information correlation; S15, Based on the fusion feature results, combined with a preset two-layer fully connected neural network, fault trend prediction inference is performed to obtain preliminary fault prediction results; S16, Based on the preliminary fault prediction results, perform a global search and adjustment on the preset two-layer fully connected neural network to obtain an optimized parameter combination; S17, based on the optimized parameter combination, update the configuration information of the preset two-layer fully connected neural network, and perform training on the fused feature result to obtain the prediction confidence; S18, if the prediction confidence is lower than the preset prediction threshold, then return to the step of performing a global search and adjustment on the preset two-layer fully connected neural network based on the fault prediction result to obtain an optimized parameter combination, and otherwise output the final fault prediction result.

[0022] In step S11, it is necessary to obtain the original operating data of the charging pile and perform preprocessing on the original operating data to obtain an initial data set with a unified format, including: Add timestamps to the original running data to obtain an original record set with a time-series structure; Based on the original record set, the data of different modalities are processed to unify the units and missing values ​​are filled in to obtain an intermediate dataset with a consistent format. The intermediate dataset is anomaly-marked, and out-of-range abnormal data points are removed or labeled to obtain a filtered dataset containing anomaly information; Based on the filtered dataset, the data is segmented and organized into time window segments to obtain an initial dataset with a unified format.

[0023] In a specific embodiment, to effectively monitor the operating status of charging piles and subsequently model fault prediction, it is first necessary to acquire the raw operating data of the charging piles. This raw data includes, but is not limited to, various key operating parameters such as charging current, voltage, and equipment surface temperature. The acquisition frequency can be set to once per minute, with a continuous sampling period of no less than 24 hours. To ensure the data is usable for time-series analysis, timestamps need to be introduced during the data acquisition phase through a pre-defined sensor interface. Specifically, an IoT acquisition gateway can be used to uniformly encapsulate data collected from current sensors (such as Hall current transmitters), voltage acquisition modules, and temperature sensors (such as NTC thermistors), and add a standard UTC format time field to each data record, ultimately generating a raw record set with a continuous time index. For example, in a certain acquisition, the system records a current of 14.2A, a voltage of 225.6V, a temperature of 29.8℃, and a timestamp of 2025-06-30 14:00:00.

[0024] Specifically, to eliminate the differences in dimensions and inconsistencies in data units between different modalities in the original record set, the data for each modality needs to be standardized. The Z-score standardization method is used, subtracting the mean from each data category and dividing by the standard deviation to convert it into a dimensionless form. Taking current as an example, if the mean of current data in a certain sampling segment is 12.0A and the standard deviation is 1.5A, then the original value 14.2A can be standardized to (14.2-12.0) / 1.5 = 1.47. For missing values, linear interpolation is used to impute the missing values, ensuring that all modal fields are complete at each time stamp. For example, if temperature is missing in the data at the 35th minute, with the data before and after being 28.4℃ and 28.8℃ respectively, it can be imputed to 28.6℃. After the above processing, the resulting intermediate dataset maintains structural consistency and a unified data format, making it suitable for subsequent analysis.

[0025] For example, after obtaining the intermediate dataset, to further improve data quality, anomaly labeling is required to identify sudden changes or abnormal fluctuations in data. The system presets reasonable fluctuation ranges for each mode or uses statistical rules (such as the 3σ principle) for initial judgment. For example, for standardized voltage data, if the value at a certain moment exceeds the mean ± 3 times the standard deviation, it can be considered abnormal; if the standardized voltage at a certain point is 3.8 in actual acquisition, which exceeds the statistical range, it can be automatically marked as abnormal.

[0026] In one specific embodiment, to organize the above data into a time-series segment structure with analytical significance, the selected dataset needs to be segmented based on time windows. The time window can be set according to the actual charging cycle; common configurations are 10-minute or 30-minute sliding windows, with each window containing a fixed number of data points. Taking a 10-minute window as an example, the data within each window is independently numbered and packaged, while maintaining the original time index. The data in each window is repackaged according to modality dimensions to form a unified structure, with field formats including window start time, end time, mean, maximum, minimum, and fluctuation range for each modality. In this way, an initial dataset with a unified structure and traceable time series can be obtained, laying the foundation for subsequent feature extraction and predictive modeling.

[0027] In step S12, dynamic change indicators for each time period need to be extracted based on the initial data set to obtain a time-series feature set, including: The initial data set is segmented by combining a preset time window interval to obtain a segmented time sequence set; Dynamic feature analysis is performed on the modal data within the time series fragment set to extract the rate of change, extreme value range, and fluctuation frequency, resulting in a dynamic feature dataset with changing trends. Based on the dynamic feature dataset, time-dependent segments that exceed a preset dynamic threshold are identified and anomaly marking is performed to obtain a feature segment set containing anomaly markers. The trend consistency adjustment is performed on the feature fragment set to obtain the optimized time series feature set.

[0028] In a specific embodiment, to extract time-series features reflecting changes in the charging pile's operating status from the initial dataset, the initial dataset first needs to be segmented according to time windows. This time window can be flexibly set according to the system monitoring granularity; commonly used intervals include 5 minutes, 10 minutes, or 30 minutes. Taking 10 minutes as an example, a set of data is extracted from each 10-minute continuous operation record to form a time-series segment. If the system samples data once per minute, each segment contains 10 modal data points such as current, voltage, and temperature. By iterating through and sliding to extract all time windows segment by segment, a segmented dataset containing multiple time-series segments can be formed, each segment having an independent time index and modal data structure.

[0029] Specifically, for each segmented time period, dynamic feature extraction is performed to characterize the operational fluctuations of the charging pile within that time period. The dynamic indicators used by the system include, but are not limited to, rate of change (such as the derivative of current change), extreme value range (the difference between the maximum and minimum values), and fluctuation frequency (the number of changes per unit time). For example, if the current rises from 11A to 14A within a 10-minute window and then falls back to 12A, the rate of change is (14-11) / Δt, the extreme value range is 3A, and the fluctuation frequency can be defined as the number of changes exceeding ±0.5A. The system extracts these indicators for the three modal dimensions of current, voltage, and temperature, ultimately generating a dynamic feature dataset with trend descriptions for further analysis of state change behavior.

[0030] For example, to identify potential abnormal trends, after obtaining the dynamic feature dataset, thresholds need to be set for anomaly identification and labeling. The system presets reasonable ranges for each type of indicator; for example, the normal rate of change of current should be less than 0.3A / minute. If a certain window rate reaches 0.5A / minute, it can be judged as an abnormal segment with excessively fast dynamics. Simultaneously, a time-dependent constraint is introduced; if an abnormal segment appears continuously for more than three windows, it is considered a trend anomaly. Anomaly labeling can use Boolean bits or additional identifier fields, such as "flag=1" indicating that the current segment has abnormal characteristics. The final feature segment set is the set of suspicious time-series segments containing anomaly identifiers.

[0031] In one specific embodiment, to further improve feature stability and overall analysis accuracy, trend consistency adjustment is also required for the labeled feature segment set. The adjustment logic mainly targets critical intervals or misjudgment situations. By comparing the trend change direction and magnitude of adjacent time segments, it determines whether to retain the current anomaly label. For example, if a window is judged as an anomaly, but the windows before and after it are normal and the trend change is gradual, the system can determine that the current segment is an occasional disturbance and cancel the label; conversely, if three consecutive windows show the same trend, their label weight is increased and they are merged into one anomaly period. The final adjustment result constitutes the optimized temporal feature set, which not only retains key dynamic features but also improves the temporal consistency of anomaly identification, providing a more reliable input basis for subsequent modeling and prediction.

[0032] It should be noted that, to effectively identify abnormal dynamic behaviors in time-series segments, the system introduces preset dynamic thresholds as the basis for judgment. These dynamic thresholds are set separately for different modalities of data, specifically including current change rate thresholds, voltage fluctuation amplitude thresholds, and temperature rise rate thresholds, used to identify drastic changes in data within a short period. Taking current data as an example, under normal operating conditions of the charging pile, the current change rate does not exceed 0.2A / minute; therefore, the system sets the preset dynamic threshold for the current change rate to ±0.3A / minute. If the absolute value of the current derivative calculated within a certain time window exceeds this threshold, the segment is marked as abnormal. Similarly, for voltage fluctuations, the system can set the fluctuation amplitude threshold to ±10V; for temperature, the temperature rise rate threshold is set to 1.0℃ / 10 minutes. Each threshold can be initialized based on historical operating data statistics or can be adaptively updated according to specific usage scenarios. Through the configuration of the above dynamic thresholds, the system can quickly identify abnormal operating trends during the feature extraction stage, providing data basis for subsequent pre-diagnostic analysis.

[0033] In step S13, feature vectors for each modality data need to be extracted based on the time-series feature set to obtain an independent feature representation set, including: Based on the time-series feature set, current, voltage, and temperature are extracted to obtain modal data; Based on the modal data, perform high-dimensional mapping to expand the modal feature dimensions and obtain a high-dimensional feature dataset; Based on the high-dimensional feature dataset, abnormal feature fragments exceeding a preset abnormal threshold are identified and labeled to obtain a labeled feature set; Based on the set of labeled features, the modal data is calibrated to obtain an independent feature representation set.

[0034] In a specific embodiment, to further enhance the model's ability to represent the operating status of charging piles, it is necessary to extract modal data such as current, voltage, and temperature based on the aforementioned optimized temporal feature set, and use these as basic data sources for subsequent high-dimensional feature construction. Specifically, the system first extracts key statistical indicators for each modality at different time segments from the temporal feature set, such as the average value and maximum rate of change of current, the stability range of voltage, and the upward trend of temperature. The extracted modal data is organized in vector form, retaining the corresponding time index, forming a basic set of three types of modal features, laying the input foundation for subsequent high-dimensional representation.

[0035] Specifically, to enhance the model's ability to recognize complex dynamic structures within modes, high-dimensional mapping processing needs to be performed on the aforementioned modal data. The system employs a multi-dimensional mapping strategy, expanding the original features into a higher-dimensional feature space through methods such as combination functions, derivative calculations, and cross-term expansion. For example, for current data, the system can construct a three-dimensional representation from "average value," "variation amplitude," and "fluctuation frequency," and then introduce the ratio of the rate of change with voltage to form a five-dimensional feature vector. For example, if the average current is 13.2A, the variation range is 3.4A, and the fluctuation frequency is 2 times / 10 minutes, then the original features are 13.2, 3.4, 2.0. This can be further expanded into a five-dimensional vector: 13.2, 3.4, 2.0, (13.2 / 220.5), (3.4 × temperature rise rate). All modalities are constructed in this way, ultimately forming a high-dimensional feature dataset with a unified structure, used to enhance the model's input expressive power.

[0036] For example, to ensure the usability and stability of high-dimensional features, the system needs to identify anomalous feature segments. Specifically, anomaly threshold ranges are set for each dimension of the high-dimensional feature dataset, using historical data statistics to set upper limits such as 4.0A for current variation and 2.5℃ / 10 minutes for temperature rise rate. If a segment shows a current variation of 5.1A, it is considered anomalous, and the corresponding high-dimensional vector is marked as an anomalous feature. Marking can be achieved by adding a flag=1 field or by adding an anomalous index array to form a set of marked features, which will be the focus of subsequent calibration operations.

[0037] In one specific embodiment, to mitigate the interference of anomalous segments on the overall modeling accuracy, the system needs to perform modal calibration on the aforementioned marked feature set. This processing, based on temporal consistency and intermodal coupling, employs methods such as weighted smoothing, neighborhood equalization, or reconstruction substitution to adjust the values ​​of anomalous feature segments. Taking the rate of change of current as an example, if it abnormally rises to 6 A / min over a certain period, the system can use a weighted median method based on the values ​​taken before and after the current window to correct it to 4.2 A / min, which is closer to the contextual trend. Simultaneously, the system considers the temperature trend to determine if there are any linked anomalies, and performs joint modal correction if necessary. Finally, the calibrated feature vector set becomes an independent feature representation set, which can be used as the fusion input for the next stage, ensuring that the model maintains accuracy and stability under complex operating conditions.

[0038] In one specific embodiment, to identify abnormal change segments in high-dimensional feature data, the system sets preset abnormal thresholds based on historical operational data statistics to determine whether there are out-of-limit behaviors in the modal feature vector. For the current mode, the set threshold for change amplitude is 4.0 amperes, and the threshold for change rate is 0.35 amperes / minute; for the voltage mode, the set threshold for instantaneous drop amplitude is 12 volts, and the set threshold for average fluctuation rate is 5%; for the temperature mode, the set threshold for temperature rise rate is 1.8℃ / 10 minutes. These thresholds are derived from statistical analysis of 2000 sets of normal and abnormal charging pile operational data, covering various seasons, charging power levels, and usage scenarios from different equipment manufacturers. For example, if the current rapidly increases from 8.5A to 14.0A within 10 minutes over a certain time window, with a change amplitude of 5.5A, exceeding the preset threshold of 4.0A, this segment can be marked as an abnormal segment and included in subsequent calibration processing. This abnormal judgment logic is embedded in the feature extraction subsystem as the core basis for the screening and marking mechanism.

[0039] In step S14, a weighted fusion operation needs to be performed based on the set of independent feature representations to obtain a fusion feature result with multi-dimensional information correlation, including: Based on the set of independent feature representations, the weight coefficients of each modality feature vector are calculated and then combined in a weighted manner to obtain a weighted feature matrix. Based on the weighted feature matrix, dimensional unification and time alignment processing are performed to calibrate feature deviations between different modalities, resulting in a fusion feature result with multi-dimensional information correlation.

[0040] In a specific embodiment, to integrate modal feature information such as current, voltage, and temperature, the system performs weighted fusion processing based on a pre-constructed set of independent feature representations, aiming to obtain fused feature results that reflect multi-dimensional correlations. Specifically, the system first analyzes the importance of each type of modal feature to determine the relative contribution of each mode to the prediction task. This weight allocation process can be based on historical prediction accuracy performance and statistical results of the correlation between modal features and fault precursors. For example, in a charging pile application scenario, the detection of current changes has a stronger early indication role in fault risk; therefore, the system can set the fusion weight of the current mode to 0.5, the voltage mode to 0.3, and the temperature mode to 0.2. Subsequently, the system performs a weighted combination operation on the feature vectors of each mode according to the above weights to generate a weighted feature matrix. This matrix integrates the changing trends, extreme value ranges, and fluctuation information of each mode, forming a unified structured fused feature expression.

[0041] Specifically, to ensure consistency in numerical scale and temporal distribution across different modal data, the system further performs dimensionality unification and time alignment processing on the weighted feature matrix. Dimensionality unification addresses the issue of inconsistent dimensions between modalities; for example, current is measured in amperes while temperature is expressed in degrees Celsius. Direct fusion could lead to numerical bias. Therefore, the system uses normalization and scaling to unify all modal feature values ​​into a relatively comparable range. Simultaneously, considering the different sampling frequencies and misaligned acquisition times of multimodal data, the system employs a unified time axis alignment mechanism. Using the current mode as a benchmark, other modal data are time-matched using nearest-neighbor interpolation or sliding synchronization to ensure that each set of fused features corresponds to the device's operating status at the same point in time. Through these processes, the final fused feature results exhibit good numerical balance and temporal consistency, providing high-quality input for subsequent prediction models.

[0042] In step S15, based on the fused feature results, and combined with a preset two-layer fully connected neural network, fault trend prediction inference is performed to obtain preliminary fault prediction results, including: Based on the fusion feature results, extract the feature subsets related to time-series dependencies, classify and organize them to obtain the classified feature groups; Based on the feature grouping, the correlation strength between each feature is calculated, feature combinations whose dynamic changes exceed a preset abnormal threshold are identified, and abnormal marking is performed to obtain abnormal feature groups; Based on the abnormal features grouping, compare with known fault modes in preset historical data, extract key mode fragments, and determine the mode features associated with the fault. The pattern features are input into a preset two-layer fully connected neural network for training and inference to obtain preliminary fault prediction results.

[0043] In one specific embodiment, to predict the fault trends of charging pile equipment, the system performs predictive inference based on the constructed fusion feature results and a preset two-layer fully connected neural network structure, outputting preliminary fault judgment results. Specifically, the system first extracts the feature subset with the strongest time correlation from the fusion feature results. These features include voltage fluctuation period, current change slope, temperature gradient trend, etc., which have obvious temporal correlation with the equipment operating status. Subsequently, the system classifies and organizes the extracted subset according to the physical source and statistical characteristics of the data, dividing it into power feature group, temperature rise feature group, and fluctuation frequency feature group to enhance the network structure's ability to understand the input data.

[0044] In one specific embodiment, to effectively calculate the correlation strength between modal features, the system constructs a set of feature pairs within the same time window based on key feature vectors such as temperature, current, and voltage extracted from the fused feature results, using a time alignment method. Subsequently, the system compares the change trends of each pair of modal features within the same time slice to determine whether their fluctuation directions are consistent. For example, if the temperature value continuously rises within a certain time window, and the current value also shows an upward trend, the system considers these two features to have a positive correlation trend. The system further sets a sliding time window and statistically analyzes the change directions of each feature pair in multiple consecutive time segments. If, within the entire sliding window interval, the two features rise or fall synchronously for more than 70% of the time period, the pair of features is considered to have a strong correlation strength.

[0045] Furthermore, the system can introduce a variation amplitude comparison mechanism, which, based on the consistency analysis of feature directions, further calculates the similarity of their variation amplitudes. If the difference in the variation of two features within a unit of time is small (e.g., within ±10%), they are judged as a high-match combination. Through this dual-rule judgment mechanism of "directional consistency + similar variation amplitude," the system can achieve a textual and logical quantitative judgment of the strength of feature association without relying on traditional correlation coefficient formulas, providing a reliable basis for subsequent fault feature screening and model input.

[0046] Specifically, the system further analyzes the interrelationships within each feature group and assesses the strength of their correlation. This process is based on the similarity or degree of coordinated change in the trends of each feature over a recent period. For example, when the current rises rapidly and is accompanied by drastic temperature changes, the system can determine that these two types of features have a strong correlation before the fault occurs. The system marks these highly correlated combinations and determines whether they exceed preset abnormal thresholds (such as a temperature rise rate exceeding 10℃ / min or a current fluctuation exceeding 15%). If the conditions are met, they are grouped into abnormal feature groups.

[0047] Next, the system extracts the change trajectories from the abnormal feature groups and compares them with a pre-set historical sample library to identify segments that are highly consistent with typical fault modes. Taking a real-world operating scenario as an example, the system detects that the voltage continuously drops beyond a set threshold within 3 minutes, and the temperature rises sharply in a stepwise manner. After comparison, this highly matches the "poor output contact" type of fault mode. Based on this, the system extracts key features of this type of mode (such as multiple periodic fluctuations and temperature transition characteristics) as the mode feature input for the final inference.

[0048] Finally, the system inputs the aforementioned pattern features into a pre-configured two-layer fully connected neural network for forward propagation training and real-time inference. This neural network includes two hidden layers, each with 64 and 32 nodes respectively, employing the ReLU activation function to enhance nonlinear mapping capabilities. During inference, the system performs weighted synthesis of the input features and outputs a preliminary fault prediction result for the current device, such as a prediction of "high-risk abnormal trend" or "possible short-term voltage instability." This prediction result will serve as the basis for subsequent optimization training and feedback adjustments.

[0049] It should be noted that, to achieve fault trend prediction based on the multimodal characteristics of charging piles, the system pre-constructs a two-layer fully connected neural network model. This model consists of an input layer, two hidden layers, and an output layer, using a fixed configuration and pre-trained with historical datasets. The input layer receives multi-dimensional vector data that has undergone feature fusion, specifically including a 24-dimensional fused feature vector containing the mean current, standard deviation, maximum temperature, and voltage change rate per unit time, after time-series segmentation processing.

[0050] In terms of neural network structure, the first hidden layer has 64 neurons, and the second hidden layer has 32 neurons, both using the ReLU activation function to improve the ability to express nonlinear feature combinations. The output layer uses one Sigmoid activation node to normalize the probability of fault occurrence. The network uses binary cross-entropy as the loss function, and the optimization method is the Adam optimizer. The initial learning rate is set to 0.001, and the batch size is 64.

[0051] For example, the system constructs the training dataset as follows: Sample data containing complete label information is selected from historical charging pile operation logs. A fused feature vector is extracted within each 5-minute time window, and the window is labeled to indicate whether any fault phenomena such as abnormal high temperature, sudden voltage drop, or abnormal output interruption have occurred within that window. The fault labels in the training set are binary encoded (0 represents normal, 1 represents abnormal). Approximately 8000 sets of labeled fused feature samples were collected, of which 70% were used for training, 15% for validation, and 15% for testing.

[0052] During training, the model reads training samples in a predefined batch order, continuously performing forward and backward propagation to optimize network weight parameters. After each training epoch, the system records the current loss function value and the validation set prediction accuracy. If the validation set accuracy shows no improvement (fluctuation less than 0.5%) for 10 consecutive training epochs, or if the loss function has stabilized below 0.15, the system determines that the training process has converged and stops iterating early. If the above conditions are not met even after reaching the maximum number of training epochs (e.g., 100 epochs), the system will record the current state and output the best-performing model as the baseline model.

[0053] After training, the system freezes the network structure and uses it as a preliminary judgment model in the fault prediction process. It performs the inference process of fusing feature inputs and risk probability outputs, providing a basic reference for subsequent parameter optimization and model adjustment.

[0054] It should be noted that the preset historical data refers to the actual operation record data collected and stored during the operation of the charging pile. The data sources include the on-site deployed power acquisition module, environmental sensors, control unit log system, and upper computer data management platform. The collected historical data contains time-series information from multiple continuous operating cycles, covering various states such as daily operation, before and after a fault, and during equipment maintenance, and is representative and complete.

[0055] Specifically, the modal information included in the historical data includes, but is not limited to: charging output current, real-time voltage values, interface temperature, charging duration, load power factor, user start-up interval, and ambient temperature and humidity. For example, at an electric bus charging station in a certain city, operational data from April 2023 to March 2024 was selected, totaling over 12,000 hours of operational records. These logs were synchronized with the upper-level platform via the CAN interface of the charging pile control board. Data was sampled at 5-minute intervals, resulting in 12 sets of records per hour and 288 sets per day, totaling over 1 million data samples per station.

[0056] To generate the labeled samples needed for training, the system marks confirmed fault events in historical records (such as automatic power-off due to overheating, voltage drops, and connection anomalies) with time points. For example, between 14:35 and 14:50 on September 18, 2023, a device experienced a sudden drop in output voltage to 190V, which remained unstable for 5 minutes. Subsequently, the system triggered its protection mechanism to automatically shut down the power; this record is numbered #20230918_293. Such samples are labeled as fault windows and paired with multimodal information such as temperature, current, and load status during the same period to form complete samples. Time windows without anomalies are included in the training set as normal samples for balanced training of the model.

[0057] The pre-defined historical data constructed using the above methods possesses temporal consistency, modal integrity, and event traceability, providing a high-quality reference foundation for subsequent fault mode identification, trend training, and anomaly comparison.

[0058] In step S16, based on the preliminary fault prediction results, a global search and adjustment of the preset two-layer fully connected neural network is performed to obtain an optimized parameter combination, including: Based on the preliminary fault prediction results, the uncertainty distribution of the prediction output is analyzed, and the output subset related to abnormal fluctuations is extracted to obtain the abnormal distribution range. Based on the range of the abnormal distribution, identify the key parameter combinations in the current model parameters that are related to the local optimum, and obtain the subset of parameters to be optimized; Based on the subset of parameters to be optimized, a global search is performed to adjust the parameter weights, optimize the correlation configuration between parameters, and obtain an optimized parameter combination.

[0059] In a specific embodiment, to improve the judgment accuracy of the charging pile fault prediction model, step S16 requires a global search and adjustment of the parameters of the preset two-layer fully connected neural network. This aims to overcome the problems of insufficient accuracy or high uncertainty in the current preliminary prediction results. This processing flow takes the preliminary prediction results as input, systematically analyzes their uncertainty distribution, constructs an optimized logical path, and outputs a set of updated parameter configurations, thereby improving the model's ability to identify boundary states and weak feature anomalies.

[0060] Specifically, the system first analyzes the distribution of confidence scores in the preliminary prediction results, paying particular attention to the middle region where predicted values ​​are concentrated between 0.4 and 0.6. This region reflects strong uncertainty in the model under certain time periods or feature combinations. For example, in one inference process, approximately 18% of the predicted samples had confidence scores between 0.45 and 0.55. Further analysis, combined with their input characteristics, revealed that these samples often corresponded to abnormal situations such as low load at night, high temperatures, or sudden fluctuations. Based on this, the system extracts these samples with large fluctuations in prediction confidence scores as an "abnormal output subset" for subsequent targeted optimization.

[0061] Subsequently, based on the aforementioned subset of abnormal outputs, the system traced and analyzed the current model's weight configuration and hyperparameter settings, identifying the parameter regions that led to local optima. For example, it was found that the weight coefficient corresponding to the temperature feature in the first hidden layer was too low (less than 0.05), and since historical analysis showed that temperature had an impact on fault judgment under high load, this weight configuration was included in the subset of parameters to be optimized. Other key parameters identified included the learning rate (initially set to 0.01), the activation function type (originally set to ReLU), and the feature normalization method (originally maximum scaling).

[0062] Building upon this, the system employs a global search method based on an evolutionary strategy to combine and adjust the aforementioned key parameters. For example, the system initializes a candidate solution set containing 100 different weight combinations and performs a fast training-evaluation loop for each configuration. In each iteration, the system scores the model based on three metrics: prediction accuracy on the validation set, F1-score, and confidence stability, selecting the optimal parameter combination to update the model. For instance, after three iterations, an optimized parameter configuration is obtained: the temperature weight is increased to 0.12, the learning rate is adjusted to 0.005, and the activation function is changed to LeakyReLU. Ultimately, the proportion of predictions with a confidence greater than 0.8 in the test set increases from 71% to 85%.

[0063] Ultimately, through this global search and adjustment process, the system obtains a set of optimal network parameter configurations for the current scenario, effectively improving the model's generalization ability and the stability of fault prediction, laying the foundation for the next step of training updates and final result output. The entire process is adaptive, iterative, and data-driven, making it suitable for charging pile PHM applications under different site conditions and load conditions.

[0064] In step S17, the configuration information of the preset two-layer fully connected neural network needs to be updated according to the optimized parameter combination, and training is performed on the fused feature results to obtain the prediction confidence, including: In a specific embodiment, in step S17, the structure and weight configuration of the original two-layer fully connected neural network are updated based on the optimized parameter combination obtained in the previous step S16. The training process is then re-executed based on the extracted fusion feature results to obtain the current model's prediction confidence for various fault trends. This process aims to ensure that the model, under the optimized configuration, can fully learn the key correlation patterns in the fusion features, thereby improving the accuracy and reliability of its judgments on future states.

[0065] Specifically, the system first writes the optimized parameter combination into the neural network's configuration file, including but not limited to the updated weight initialization values, learning rate, activation function type, regularization strength, and batch size. For example, if the optimization result output in step S16 is: the first layer weight matrix is ​​updated to W1*, the activation function is switched to LeakyReLU, and the learning rate is set to 0.005, then the system replaces the data at the corresponding positions in the original network structure with these parameters and reconstructs and initializes the network.

[0066] Subsequently, the system uses the fused feature result as input to the training dataset to retrain the model. This fused feature result contains a high-dimensional feature vector generated by weighted fusion of modal information such as current, voltage, and temperature, exhibiting clear fault precursor trend characteristics. During training, the system employs supervised learning, using historical label data as the target value, performing forward and backward propagation operations, and gradually optimizing the loss function value through gradient descent. For example, the training process is set to a maximum of 200 epochs; training is stopped early if the validation set loss value no longer decreases within 10 consecutive epochs. The loss function used is the cross-entropy function, and the optimizer is the Adam algorithm.

[0067] After training, the system calculates the confidence score of the predicted output for each fused feature sample. This confidence score represents the probability that the current model classifies a specific input as a fault state. For example, after a certain round of training, the model output may contain different confidence scores such as 0.93, 0.85, and 0.47. The system uses these confidence scores to determine the reliability of the prediction, providing a basis for subsequent iterative adjustments.

[0068] Through this step, the model achieves retraining under optimized parameter configuration based on the fused feature representations, thereby improving the accuracy and reliability of the output prediction results and laying a crucial foundation for the stable output of the final fault trend. The entire process ensures the model's continuous learning and performance enhancement capabilities in a dynamic adaptive environment, meeting the reliability and real-time requirements of the charging pile PHM system.

[0069] In step S18, if the prediction confidence is lower than a preset prediction threshold, the step of performing a global search and adjustment on the preset two-layer fully connected neural network based on the fault prediction result to obtain an optimized parameter combination needs to be iterated; otherwise, the final fault prediction result is output, including: In a specific embodiment, in step S18, it is necessary to determine whether the prediction ability of the current neural network meets the preset performance requirements based on the prediction confidence value obtained in the previous step S17. If the confidence is lower than the prediction threshold set by the system, it indicates that the model's judgment of the current input sample lacks stability or accuracy. In this case, it is necessary to return to step S16 and re-perform a global search and adjustment of the model parameters to continuously optimize the network performance in an iterative manner. Otherwise, the current model output result is considered reliable, and the final fault prediction result is directly output for subsequent maintenance decisions or early warning mechanisms.

[0070] Specifically, the system first sets a set of prediction confidence thresholds as judgment criteria. For example, in the charging pile PHM application, the system presets the confidence threshold to 0.85, meaning that if the probability of the model predicting a certain input sample as "fault" or "non-fault" is less than 85%, the result is considered unstable and further optimization of the model structure and parameters is required. At this point, the system automatically jumps to step S16, re-analyzes the abnormal distribution in the current model output, extracts the corresponding key parameter subset, performs global parameter search again, and updates the network structure.

[0071] After returning to S16 to perform parameter adjustments, the system retrains the network according to the process in step S17 and obtains new prediction confidence again. This judgment-optimization-training process can be executed cyclically within a set maximum number of iterations. For example, it can be limited to a maximum of 5 rounds of optimization iterations to prevent excessive consumption of computing resources or getting trapped in local optima.

[0072] When the system detects that the prediction confidence level after the latest training is greater than or equal to 0.85, or the number of optimization attempts reaches the preset maximum limit, the system will output the current network output as the final prediction result. This prediction result may include: fault type prediction label (such as "module overheating"), occurrence probability (such as 0.91), associated modes (such as temperature + current), etc., to support the operation and maintenance system in making early maintenance scheduling or alarm response.

[0073] This step's judgment and iteration mechanism enables dynamic adaptive tuning of model performance, avoiding misjudgments caused by insufficient one-time training or suboptimal initial parameters. This mechanism enhances the model's stability and fault tolerance under complex multimodal data, improving the overall reliability and practical deployment value of the charging pile PHM system.

[0074] In summary, this invention provides a charging pile fault prediction method and system based on multimodal time series data analysis, so as to achieve effective fusion of multimodal time series data and adaptive optimization of prediction models, thereby meeting the needs of high-precision pre-diagnosis and health management in charging pile PHM applications.

[0075] Reference Figure 2 The second embodiment of the present invention provides a charging pile fault prediction system based on multimodal time series data analysis, comprising: The data acquisition module is used to acquire the original operating data of the charging pile and perform preprocessing on the original operating data to obtain an initial data set with a unified format. The time-series feature module is used to extract dynamic change indicators within each time period based on the initial data set to obtain a time-series feature set; An independent feature module is used to extract feature vectors of each modality data based on the time-series feature set to obtain an independent feature representation set; The feature fusion module is used to perform a weighted fusion operation based on the set of independent feature representations to obtain a fusion feature result with multi-dimensional information correlation. The preliminary prediction module is used to perform fault trend prediction inference based on the fused feature results and combined with a preset two-layer fully connected neural network to obtain preliminary fault prediction results. The parameter optimization module is used to perform a global search and adjustment on the preset two-layer fully connected neural network based on the preliminary fault prediction results to obtain an optimized parameter combination. The configuration update module is used to update the configuration information of the preset two-layer fully connected neural network according to the optimized parameter combination, and to perform training on the fused feature results to obtain the prediction confidence. The prediction and judgment module is used to iterate the steps of performing a global search and adjustment on the preset two-layer fully connected neural network to obtain an optimized parameter combination based on the fault prediction result if the prediction confidence is lower than a preset prediction threshold; otherwise, it outputs the final fault prediction result.

[0076] It should be noted that the charging pile fault prediction system based on multimodal time series data analysis provided in this embodiment of the invention is used to execute all the process steps of the charging pile fault prediction method based on multimodal time series data analysis in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0077] This invention also provides an electronic device. The electronic device includes a processor, a memory, and a computer program stored in the memory and executable on the processor, such as a preliminary prediction program. When the processor executes the computer program, it implements the steps described in the various embodiments of the charging pile fault prediction method based on multimodal time-series data analysis, for example... Figure 1 The step S11 shown. Alternatively, when the processor executes the computer program, it implements the functions of each module / unit in the above-described device embodiments, such as the parameter optimization module.

[0078] For example, the computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the electronic device.

[0079] The electronic device may be a desktop computer, laptop, handheld computer, or smart tablet, etc. The electronic device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the above components are merely examples of electronic devices and do not constitute a limitation on the electronic device. It may include more or fewer components than described above, or combine certain components, or different components. For example, the electronic device may also include input / output devices, network access devices, buses, etc.

[0080] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the electronic device, connecting all parts of the electronic device via various interfaces and lines.

[0081] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory and by calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0082] Wherein, if the modules / units integrated in the electronic device are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0083] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0084] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for predicting charging pile faults based on multimodal time series data analysis, characterized in that... include: Obtain the raw operating data of the charging piles and perform preprocessing on the raw operating data to obtain an initial data set with a unified format; Based on the initial data set, dynamic change indicators for each time period are extracted to obtain a time-series feature set; Based on the time-series feature set, feature vectors of each modality data are extracted to obtain an independent feature representation set; Based on the set of independent feature representations, a weighted fusion operation is performed to obtain a fusion feature result with multi-dimensional information correlation. Based on the fusion feature results, combined with a preset two-layer fully connected neural network, fault trend prediction inference is performed to obtain preliminary fault prediction results. Based on the preliminary fault prediction results, a global search and adjustment are performed on the preset two-layer fully connected neural network to obtain an optimized parameter combination. Based on the optimized parameter combination, the configuration information of the preset two-layer fully connected neural network is updated, and training is performed on the fused feature results to obtain the prediction confidence. If the prediction confidence is lower than the preset prediction threshold, the process of iterating through the step of performing a global search and adjustment on the preset two-layer fully connected neural network based on the fault prediction result to obtain an optimized parameter combination is returned; otherwise, the final fault prediction result is output.

2. The charging pile fault prediction method based on multimodal time series data analysis according to claim 1, characterized in that, The process involves acquiring the raw operating data of the charging piles and performing preprocessing on the raw operating data to obtain an initial data set with a unified format, including: Add timestamps to the original running data to obtain an original record set with a time-series structure; Based on the original record set, the data of different modalities are processed to unify the units and missing values ​​are filled in to obtain an intermediate dataset with a consistent format. The intermediate dataset is anomaly-marked, and out-of-range abnormal data points are removed or labeled to obtain a filtered dataset containing anomaly information; Based on the filtered dataset, the data is segmented and organized into time window segments to obtain an initial dataset with a unified format.

3. The charging pile fault prediction method based on multimodal time series data analysis according to claim 1, characterized in that, The step of extracting dynamic change indicators for each time period based on the initial data set to obtain a time-series feature set includes: The initial data set is segmented by combining a preset time window interval to obtain a segmented time sequence set; Dynamic feature analysis is performed on the modal data within the time series fragment set to extract the rate of change, extreme value range, and fluctuation frequency, resulting in a dynamic feature dataset with changing trends. Based on the dynamic feature dataset, time-dependent segments that exceed a preset dynamic threshold are identified and anomaly marking is performed to obtain a feature segment set containing anomaly markers. The trend consistency adjustment is performed on the feature fragment set to obtain the optimized time series feature set.

4. The charging pile fault prediction method based on multimodal time series data analysis according to claim 1, characterized in that, The step of extracting feature vectors from each modality data based on the temporal feature set to obtain an independent feature representation set includes: Based on the time-series feature set, current, voltage, and temperature are extracted to obtain modal data; Based on the modal data, perform high-dimensional mapping to expand the modal feature dimensions and obtain a high-dimensional feature dataset; Based on the high-dimensional feature dataset, abnormal feature fragments exceeding a preset abnormal threshold are identified and labeled to obtain a labeled feature set; Based on the set of labeled features, the modal data is calibrated to obtain an independent feature representation set.

5. The charging pile fault prediction method based on multimodal time series data analysis according to claim 1, characterized in that, The step of performing a weighted fusion operation based on the set of independent feature representations to obtain a fusion feature result with multi-dimensional information relationships includes: Based on the set of independent feature representations, the weight coefficients of each modality feature vector are calculated and weighted together to obtain a weighted feature matrix; Based on the weighted feature matrix, dimensional unification and time alignment processing are performed to calibrate feature deviations between different modalities, resulting in a fusion feature result with multi-dimensional information correlation.

6. The charging pile fault prediction method based on multimodal time series data analysis according to claim 1, characterized in that, Based on the fused feature results, and combined with a preset two-layer fully connected neural network, fault trend prediction inference is performed to obtain preliminary fault prediction results, including: Based on the fusion feature results, extract the feature subsets related to time-series dependencies, classify and organize them to obtain the classified feature groups; Based on the feature grouping, the correlation strength between each feature is calculated, feature combinations whose dynamic changes exceed a preset abnormal threshold are identified, and abnormal marking is performed to obtain abnormal feature groups; Based on the abnormal features grouping, compare with known fault modes in preset historical data, extract key mode fragments, and determine the mode features associated with the fault. The pattern features are input into a preset two-layer fully connected neural network for training and inference to obtain preliminary fault prediction results.

7. The charging pile fault prediction method based on multimodal time series data analysis according to claim 1, characterized in that, The step of performing a global search and adjustment on the preset two-layer fully connected neural network based on the preliminary fault prediction results to obtain an optimized parameter combination includes: Based on the preliminary fault prediction results, the uncertainty distribution of the prediction output is analyzed, and the output subset related to abnormal fluctuations is extracted to obtain the abnormal distribution range. Based on the range of the abnormal distribution, identify the key parameter combinations in the current model parameters that are related to the local optimum, and obtain the subset of parameters to be optimized; Based on the subset of parameters to be optimized, a global search is performed to adjust the parameter weights, optimize the correlation configuration between parameters, and obtain an optimized parameter combination.

8. A charging pile fault prediction system based on multimodal time series data analysis, characterized in that, include: The data acquisition module is used to acquire the original operating data of the charging pile and perform preprocessing on the original operating data to obtain an initial data set with a unified format. The time-series feature module is used to extract dynamic change indicators within each time period based on the initial data set to obtain a time-series feature set; An independent feature module is used to extract feature vectors of each modality data based on the time-series feature set to obtain an independent feature representation set; The feature fusion module is used to perform a weighted fusion operation based on the set of independent feature representations to obtain a fusion feature result with multi-dimensional information correlation. The preliminary prediction module is used to perform fault trend prediction inference based on the fused feature results and combined with a preset two-layer fully connected neural network to obtain preliminary fault prediction results. The parameter optimization module is used to perform a global search and adjustment on the preset two-layer fully connected neural network based on the preliminary fault prediction results to obtain an optimized parameter combination. The configuration update module is used to update the configuration information of the preset two-layer fully connected neural network according to the optimized parameter combination, and to perform training on the fused feature results to obtain the prediction confidence. The prediction and judgment module is used to iterate the steps of performing a global search and adjustment on the preset two-layer fully connected neural network to obtain an optimized parameter combination based on the fault prediction result if the prediction confidence is lower than a preset prediction threshold; otherwise, it outputs the final fault prediction result.

9. An electronic device, characterized in that, The system includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor, when executing the computer program, implements the charging pile fault prediction method based on multimodal time series data analysis as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to perform the charging pile fault prediction method based on multimodal time series data analysis as described in any one of claims 1 to 7.

Images