New energy charging pile fault prediction method and device and medium

Abstract

The invention relates to the technical field of new energy equipment fault prediction. The new energy charging pile fault prediction method comprises the steps that historical charging behavior information of a target charging pile is obtained, a charging behavior portrait model is constructed based on the historical charging behavior information, and current charging behavior information is collected in real time in the operation process of the target charging pile; current charging behavior information is input into the charging behavior portrait model to obtain behavior deviation characteristics, electrical operation information of a target charging pile is obtained, normalization processing and time sequence analysis are carried out on the operation information to obtain electrical characteristics, and the behavior deviation characteristics and the electrical characteristics are input into a trained multi-mode fault prediction model. And obtaining a fault risk level, if the fault risk level exceeds a set early warning threshold, generating corresponding early warning information, and sending the early warning information to a charging pile management platform. The method has the effect of realizing high-precision and low-delay fault early warning.

Description

Technical Field

[0001] This invention relates to the technical field of fault prediction for new energy equipment, and in particular to a method, device and medium for fault prediction of new energy charging piles. Background Technology

[0002] For predicting faults in new energy charging stations, regular manual inspections are typically used. Staff conduct on-site inspections of the charging stations at set intervals, checking for external damage, loose components, etc. Additionally, simple threshold alarm mechanisms are set up to issue an alarm when certain electrical parameters, such as voltage or current, exceed preset normal ranges. Some fault diagnosis methods are based on fixed rules, using pre-defined logical rules to determine whether a charging station has malfunctioned.

[0003] However, the aforementioned existing technologies have significant drawbacks. Regular manual inspections cannot achieve real-time monitoring, making it difficult to detect potential faults promptly, and are also costly and inefficient. Simple threshold alarm mechanisms can only detect situations significantly exceeding normal ranges, failing to accurately warn of some gradual faults or situations within the normal range but exhibiting abnormal trends. Fault diagnosis based on fixed rules lacks flexibility, struggles to adapt to the complex and ever-changing operating conditions of charging piles, and cannot effectively predict faults in new energy charging piles. Summary of the Invention

[0004] To achieve high-precision, low-latency fault early warning, this application provides a method, device, and medium for predicting faults in new energy charging piles.

[0005] The above-mentioned objective of this application is achieved through the following technical solution: A method for predicting faults in new energy charging piles, the method comprising: Obtain historical charging behavior information of the target charging pile, and construct a charging behavior profile model based on the historical charging behavior information; During the operation of the target charging pile, the current charging behavior information is collected in real time and input into the charging behavior profile model to obtain behavior deviation features; The electrical operation information of the target charging pile is obtained, and the operation information is normalized and analyzed by time series to obtain electrical characteristics; The behavioral deviation features and electrical features are input into the trained multimodal fault prediction model to obtain the fault risk level. If the fault risk level exceeds the set warning threshold, a corresponding warning message is generated and sent to the charging pile management platform.

[0006] By adopting the above technical solution, a charging behavior profile model is constructed based on the historical charging behavior information of the target charging pile. During operation, current charging behavior information is collected in real time. By comparing the difference between the current behavior and the historical typical behavior, behavioral deviation features are extracted. At the same time, the electrical operation information is normalized and time series analyzed to extract stable electrical feature information. The two types of features are fused and input into the trained multimodal fault prediction model for inference and judgment. This can effectively improve the sensitivity and accuracy of fault risk identification of new energy charging piles. In the model training stage, the labeled historical multimodal samples are used to construct training subsets and validation subsets. Deep learning optimization is carried out using the cross-entropy loss function and backpropagation mechanism, so that the model can fully learn the fault mode features in charging behavior and electrical signals. This improves the ability to identify abnormal trends under complex working conditions such as gradual and nonlinearity, thereby making up for the shortcomings of manual inspection, simple thresholds or fixed rule strategies, realizing early fault identification and intelligent early warning, and enhancing the stable operation guarantee capability of new energy charging piles in changing environments.

[0007] In a preferred embodiment, this application can be further configured as follows: obtaining historical charging behavior information of the target charging pile and constructing a charging behavior profile model based on the historical charging behavior information includes: The historical charging behavior information is cleaned and time-aligned to obtain behavioral feature parameters. Based on the aforementioned behavioral feature parameters, a charging behavior time series sample is constructed. An unsupervised clustering algorithm is used to perform similarity clustering on the time series sample to identify typical patterns of charging behavior. The typical patterns are feature-encoded to construct a charging behavior profile model that characterizes the typical charging habits of the target charging pile.

[0008] By adopting the above technical solution, historical charging behavior information is first cleaned and time-aligned to eliminate missing, duplicate, or abnormal data, and the recording format of different time granularities is unified to ensure the integrity and consistency of the behavior data. Then, time series samples are constructed based on the processed behavior feature parameters to make charging behaviors comparable and logically sequential. Next, unsupervised clustering algorithms are used to perform cluster analysis on the samples, and representative typical behavior patterns are identified based on features such as charging duration, start time, and charging power. By vectorizing and encoding each typical pattern, it is transformed into a fixed-length feature representation, thereby constructing a behavior profile model that represents the charging habits of the target charging pile. This profile model can reflect historical behavior patterns and stable characteristics, providing a reliable foundation for subsequent behavior deviation identification and risk assessment, and effectively overcoming the problem of limited predictive ability caused by the lack of user behavior modeling in traditional methods.

[0009] In a preferred embodiment, this application can be further configured as follows: inputting the current charging behavior information into the charging behavior profile model to obtain behavior deviation features includes: The current charging behavior information is processed by feature extraction to construct a current behavior feature vector; The Euclidean distance between the current behavior feature vector and the feature vector of the typical charging behavior pattern represented in the charging behavior profile model is calculated to obtain the behavior deviation metric. The behavior deviation metric is compared with a set deviation judgment threshold. If the behavior deviation metric exceeds the set deviation judgment threshold, the current charging behavior is determined to have an abnormal deviation, and the behavior deviation metric is used as the behavior deviation feature.

[0010] By adopting the above technical solution, feature extraction processing is performed on the current charging behavior information to construct a numerical vector representing the current behavior characteristics. Then, Euclidean distance is calculated between the vector and the feature vector of the pre-encoded typical charging behavior pattern in the charging behavior profile model to quantify the similarity difference between the current behavior and the historical typical behavior. Subsequently, the behavior deviation metric is compared with the set deviation judgment threshold. This allows for timely identification of potential abnormal trends when there are minor abnormal deviations in the charging behavior. The behavior deviation metric is used as a behavior deviation feature to provide a highly sensitive behavior state indicator for the subsequent multimodal fusion and fault prediction inference stages, thereby enhancing the accuracy and response efficiency of charging behavior anomaly identification.

[0011] In a preferred embodiment, this application can be further configured as follows: The step of calculating the Euclidean distance between the current behavior feature vector and the feature vector of the typical charging behavior pattern represented in the charging behavior profile model to obtain a behavior deviation metric includes: Based on the clustering training of multiple typical behavioral feature vectors obtained in the charging behavior profile model, a set of reference behavior vectors is constructed. Perform Euclidean distance calculations between the current behavior feature vector and each typical behavior feature vector in the reference behavior vector set, resulting in multiple distance results; According to the preset deviation evaluation rules, the behavior deviation metric value used to characterize the degree of difference between the current behavior and the historical pattern is determined from the multiple distance results.

[0012] By adopting the above technical solution, and leveraging multiple typical behavioral feature vectors formed through cluster training in the charging behavior profiling model, a set of reference behavioral vectors is constructed. This allows the current behavioral feature vector to be systematically compared with historical charging patterns. During the similarity measurement process, the Euclidean distance between the current behavioral feature vector and each reference behavioral vector is calculated one by one to comprehensively quantify the degree of difference between various behaviors. Based on multiple distance results, the most representative behavioral deviation metric value is extracted under the influence of deviation evaluation rules. This achieves an accurate characterization of the degree of difference between the current charging behavior and existing patterns, establishing a stable and reliable quantitative foundation for subsequent fault trend identification and risk assessment, and improving the sensitivity and accuracy of fault prediction.

[0013] In a preferred example, this application can be further configured such that the trained multimodal fault prediction model previously includes: Obtain a historical multimodal training sample set with labeled fault states, and divide the historical multimodal training sample set into a training subset and a verification subset according to a preset ratio; According to the preset training configuration, the training subset is input into the multilayer neural network model, and the error between the output of the multilayer neural network model and the training subset is calculated using the cross-entropy loss function. Based on the error, the parameters of the multilayer neural network model are adjusted using the backpropagation algorithm, and the parameters of the multilayer neural network model are iteratively optimized using the stochastic gradient descent algorithm to obtain the multilayer neural network model under training. Determine whether the performance index change trend of the multi-layer neural network model in training on the validation subset reaches the preset convergence condition. If the convergence condition is reached, stop training and obtain the trained multi-modal fault prediction model.

[0014] By adopting the above technical solution, supervised learning training is carried out using a historical multimodal training sample set with labeled fault states. The training data is composed of multi-source information such as behavioral features and electrical features. During the training phase, the data is divided into training subsets and validation subsets according to a preset ratio, which helps to ensure the independence of model training and validation. The difference between the model prediction and the true label is measured by the cross-entropy loss function, and the gradient information of each layer of the network is accurately calculated by the backpropagation algorithm. Then, the model weights and bias parameters are gradually optimized by combining the stochastic gradient descent algorithm. A validation subset is introduced during the model training process for performance monitoring. The convergence of the model is judged by the loss curve or accuracy curve on the validation set, thereby avoiding overfitting or underfitting. Finally, a well-trained multimodal fault prediction model with high generalization ability and fault discrimination ability under multimodal input conditions is obtained, which effectively improves the fault identification efficiency and prediction accuracy of new energy charging piles under complex working conditions.

[0015] In a preferred embodiment, this application can be further configured as follows: inputting the behavioral deviation features and the electrical features into a trained multimodal fault prediction model to obtain a fault risk level includes: The behavioral deviation features and the electrical features are embedded and mapped to construct a unified multi-dimensional fusion feature vector; The fused feature vector is input into the trained multimodal fault prediction model, which performs fault risk inference on the multidimensional fused feature vector and outputs the fault risk level corresponding to the current charging pile status.

[0016] By adopting the above technical solution, feature unification of data modalities from different sources is achieved through embedding mapping. After aligning behavioral deviation features and electrical features in the semantic space, they are combined into a fusion feature vector. This effectively integrates multi-dimensional information during the model input stage, enabling the model to simultaneously perceive feature information at both the user behavior level and the equipment operation level. In the trained multi-modal fault prediction model, deep neural networks are used to jointly model and discriminate the fusion features. This allows the output of the fault risk level to not only reflect electrical anomalies but also capture potential fault risks brought about by behavioral changes. This improves the ability to identify early complex fault symptoms, enhances the prediction accuracy of non-obvious faults, and overcomes the problem that existing methods can only make judgments based on a single anomaly threshold or static rules.

[0017] In a preferred embodiment, this application can be further configured as follows: the fused feature vector is input into the trained multimodal fault prediction model, wherein the trained multimodal fault prediction model performs fault risk inference on the multidimensional fused feature vector and outputs the fault risk level corresponding to the current charging pile state, including: The fused feature vector is input into the input layer of the trained multimodal fault prediction model, and then the feature is mapped and deep feature is extracted sequentially through the fully connected hidden layer of the trained multimodal fault prediction model to obtain high-dimensional semantic features. Based on the high-dimensional semantic features, the risk level output score is calculated using the risk level determination structure in the output layer of the trained multimodal fault prediction model, and the fault risk level is determined based on the risk level output score.

[0018] By adopting the above technical solution, the fused feature vector is input into the input layer of the trained multimodal fault prediction model. It then passes through multiple fully connected hidden layers for nonlinear feature mapping and deep feature extraction, resulting in high-dimensional semantic features. Subsequently, in the output layer, a pre-trained risk level determination structure is used to perform multi-class probability estimation on the high-dimensional semantic features, calculate the output score corresponding to each level, and combine the relative magnitudes of the scores of each level to determine the most likely fault risk level under the current state. This enables end-to-end risk level inference of the fused features, improving the intelligence level of fault prediction and the detail of level determination.

[0019] The second objective of this invention is achieved through the following technical solution: A fault prediction device for new energy charging piles, the device comprising: The behavior profile modeling module is used to obtain historical charging behavior information of the target charging pile and to build a charging behavior profile model based on the historical charging behavior information. The behavior deviation analysis module is used to collect current charging behavior information in real time during the operation of the target charging pile, and input the current charging behavior information into the charging behavior profile model to obtain behavior deviation features; An electrical feature extraction module is used to acquire the electrical operation information of the target charging pile, perform normalization processing and time series analysis on the operation information, and obtain electrical features. The early warning information sending module is used to input the behavioral deviation features and the electrical features into a trained multimodal fault prediction model to obtain the fault risk level. If the fault risk level exceeds the set early warning threshold, the module generates corresponding early warning information and sends the early warning information to the charging pile management platform.

[0020] By adopting the above technical solution, a charging behavior profile model is constructed based on the historical charging behavior information of the target charging pile. During operation, current charging behavior information is collected in real time. By comparing the difference between the current behavior and the historical typical behavior, behavioral deviation features are extracted. At the same time, the electrical operation information is normalized and time series analyzed to extract stable electrical feature information. The two types of features are fused and input into the trained multimodal fault prediction model for inference and judgment. This can effectively improve the sensitivity and accuracy of fault risk identification of new energy charging piles. In the model training stage, the labeled historical multimodal samples are used to construct training subsets and validation subsets. Deep learning optimization is carried out using the cross-entropy loss function and backpropagation mechanism, so that the model can fully learn the fault mode features in charging behavior and electrical signals. This improves the ability to identify abnormal trends under complex working conditions such as gradual and nonlinearity, thereby making up for the shortcomings of manual inspection, simple thresholds or fixed rule strategies, realizing early fault identification and intelligent early warning, and enhancing the stable operation guarantee capability of new energy charging piles in changing environments.

[0021] The above-mentioned objective three of this application is achieved through the following technical solution: A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the above-described method for predicting faults in new energy charging piles.

[0022] The fourth objective of this application is achieved through the following technical solution: A computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described method for predicting faults in new energy charging piles.

[0023] In summary, this application includes at least one of the following beneficial technical effects: 1. A charging behavior profile model is constructed based on the historical charging behavior information of the target charging pile. During operation, the current charging behavior information is collected in real time. By comparing the difference between the current behavior and the historical typical behavior, the behavior deviation features are extracted. At the same time, the electrical operation information is normalized and time series analysis is performed to extract stable electrical feature information. The two types of features are fused and input into the trained multimodal fault prediction model for inference and judgment. This can effectively improve the sensitivity and accuracy of fault risk identification of new energy charging piles. In the model training stage, the labeled historical multimodal samples are used to construct training subsets and validation subsets. The cross-entropy loss function and backpropagation mechanism are used for deep learning optimization, so that the model can fully learn the fault mode features in charging behavior and electrical signals. This improves the ability to identify abnormal trends under complex working conditions such as gradual and nonlinearity, thereby making up for the shortcomings of manual inspection, simple threshold or fixed rule strategy, realizing early fault identification and intelligent early warning, and enhancing the stable operation guarantee capability of new energy charging piles in changing environments. 2. By using an embedding mapping method to unify features of data modalities from different sources, behavioral deviation features and electrical features are aligned in the semantic space and combined into a fusion feature vector. This effectively integrates multi-dimensional information during the model input stage, enabling the model to simultaneously perceive feature information at both the user behavior level and the equipment operation level. In the trained multi-modal fault prediction model, a deep neural network is used to jointly model and discriminate the fusion features, so that the output of the fault risk level can not only reflect electrical anomalies but also capture potential fault risks brought about by behavioral changes. This improves the ability to identify early complex fault symptoms, enhances the prediction accuracy of non-obvious faults, and overcomes the problem that existing methods can only make judgments based on a single anomaly threshold or static rules. 3. The fused feature vector is input into the input layer of the trained multimodal fault prediction model. It then passes through multiple fully connected hidden layers for nonlinear feature mapping and deep feature extraction. The extraction results form high-dimensional semantic features. Subsequently, in the output layer, the trained risk level determination structure is used to perform multi-class probability estimation on the high-dimensional semantic features, calculate the output score corresponding to each level, and combine the relative magnitude of the scores of each level to determine the most likely fault risk level in the current state. This enables end-to-end risk level inference of the fused features, improving the intelligence level of fault prediction and the detail of level determination. Attached Figure Description

[0024] Figure 1 This is a flowchart of a fault prediction method for new energy charging piles in one embodiment of this application; Figure 2 This is a flowchart illustrating the implementation of step S10 in a fault prediction method for new energy charging piles according to an embodiment of this application. Figure 3This is a flowchart illustrating the implementation of step S20 in a fault prediction method for new energy charging piles according to an embodiment of this application. Figure 4 This is a flowchart illustrating the implementation of step S202 in a fault prediction method for new energy charging piles according to an embodiment of this application. Figure 5 This is a flowchart illustrating the implementation of step S40 in a fault prediction method for new energy charging piles according to an embodiment of this application. Figure 6 This is a flowchart illustrating the implementation of step S40 in a fault prediction method for new energy charging piles according to an embodiment of this application. Figure 7 This is a flowchart illustrating the implementation of step S406 in a fault prediction method for new energy charging piles according to an embodiment of this application. Figure 8 This is a schematic diagram of a fault prediction device for a new energy charging pile according to one embodiment of this application; Figure 9 This is a schematic diagram of a device according to one embodiment of this application. Detailed Implementation

[0025] The present application will be further described in detail below with reference to the accompanying drawings.

[0026] In one embodiment, such as Figure 1 As shown, this application discloses a method for predicting faults in new energy charging piles, which specifically includes the following steps: S10: Obtain historical charging behavior information of the target charging pile and build a charging behavior profile model based on the historical charging behavior information.

[0027] In this embodiment, historical charging behavior information refers to the charging process data recorded by the target charging pile during its past operating cycles. Specifically, this includes parameters such as charging start time, charging end time, charging duration, charging capacity, charging current, voltage change curve, charging interruption status, user identification tag, charging frequency, and interval. The charging behavior profiling model is a model structure built based on feature parameters extracted from historical charging behavior information. This model uses an unsupervised clustering algorithm to classify different charging behavior samples into multiple typical charging pattern categories based on feature similarity. Each category corresponds to a cluster center represented by a feature vector. All cluster centers form a feature vector set, used to describe the charging behavior patterns and preferences exhibited by the target charging pile in different time periods. This model can be used for similarity comparison and anomaly deviation measurement of current behavior.

[0028] Specifically, by reading the historical charging behavior information recorded by the target charging pile within a preset operating cycle, the start time, end time, charging duration, voltage value, current value, and charging amount of each charging process are extracted sequentially. The extracted data is arranged in chronological order, and records with time misalignment, missing fields, or abnormal values ​​are cleaned and corrected. After data cleaning, multiple sets of behavioral parameters are constructed based on the various charging behavior information to describe the characteristics of charging actions. While ensuring that the behavioral parameters have a uniform time granularity and data format, the multiple sets of behavioral parameters are aggregated and combined according to the behavioral patterns presented by each set of parameters in different charging cycles to construct a charging behavior profile model that characterizes the charging behavior characteristics of the target charging pile in different historical cycles.

[0029] S20: During the operation of the target charging pile, collect the current charging behavior information in real time, and input the current charging behavior information into the charging behavior profile model to obtain the behavior deviation characteristics.

[0030] Specifically, after the target charging pile starts charging, the start time, end time, charging duration, voltage change, current change, and charging capacity are continuously recorded. The collected current charging behavior information is organized in the same format as the historical behavior parameters. Duplicates generated during the recording process are removed, and the time field is corrected in order to maintain consistent time accuracy. The organized current charging behavior information is constructed into a current behavior parameter group based on the data values ​​in the dimensions of time, voltage, current, and capacity. Then, the current behavior parameter group is mapped into the input according to the behavior parameter dimensions used when constructing the charging behavior profile model. Finally, the charging behavior profile model outputs behavior deviation features to characterize the degree of difference between the current behavior and the historical behavior pattern.

[0031] S30: Obtain the electrical operation information of the target charging pile, normalize the operation information and perform time series analysis to obtain electrical characteristics.

[0032] In this embodiment, electrical operation information refers to various real-time data related to electrical status collected by the target charging pile during the charging process. Specifically, it includes information such as AC input voltage, input current, DC output voltage, output current, power factor, rectifier unit temperature, charging module conversion efficiency, cable contact resistance, changes in power metering values, and grid voltage fluctuations.

[0033] Specifically, during the continuous operation of the target charging pile, electrical operation information is recorded at each fixed time interval. The recorded electrical operation information is arranged in chronological order to form an original operation sequence. Time misalignment or data abrupt changes are corrected to ensure that all electrical parameters correspond to consistent sampling points under the same time reference. Subsequently, normalization is performed according to the value range of different electrical parameters to make parameter values ​​in different dimensions comparable. After processing, a time series structure is constructed for the normalized operation data according to time sequence. By observing the changes of different electrical parameters in the time dimension, the correlation patterns between electrical parameters are determined. Combined with the operation continuity in different time periods, the numerical variation trend, periodic change characteristics, and fluctuation relationship between voltage and current are extracted to form electrical features that can be used for subsequent analysis.

[0034] S40: Input the behavioral deviation features and electrical features into the trained multimodal fault prediction model to obtain the fault risk level. If the fault risk level exceeds the set warning threshold, generate the corresponding warning information and send the warning information to the charging pile management platform.

[0035] In this embodiment, the trained multimodal fault prediction model refers to a deep neural network model constructed based on a historical multimodal training sample set with labeled fault states. This model is trained by jointly modeling behavioral deviation features and electrical features. It possesses the ability to perform feature inference and risk level determination for the current operating status of charging piles. Its internal structure typically includes an input layer, multiple fully connected hidden layers, and an output layer. Training is performed using a loss function such as cross-entropy, and parameter optimization is achieved through backpropagation and stochastic gradient descent algorithms. During training, a validation subset is used to monitor performance and ensure model convergence. Ultimately, the model is used in actual operation to receive fused feature vectors and output the corresponding fault risk level. The charging pile management platform is a backend information processing platform used to centrally receive, display, and manage the operating status data and early warning information of each charging pile. It is typically deployed on a local server or in a cloud environment. It establishes data connections with multiple charging piles through communication interfaces and has the ability to receive behavioral deviation information, electrical feature information, fault risk levels, and early warning signals uploaded by the charging piles. It also supports functions such as visual display of charging pile status, recording and querying of early warning logs, and dispatching and tracking management of fault events, providing real-time status monitoring and decision support for operation and maintenance personnel.

[0036] Specifically, after acquiring behavioral deviation features and electrical features, the two types of features are organized into feature vectors with the same dimensional structure according to a pre-set input format. These vectors are then concatenated and combined to form joint feature data for fault prediction analysis. This joint feature data is input into a pre-trained multimodal fault prediction model. The model's internal feature processing structure performs layer-by-layer transformation and parameter parsing on the input features, outputting a fault risk level value corresponding to the current operating status of the charging pile. This value is then compared with a preset risk judgment threshold. If the threshold is exceeded, a warning information data packet containing the charging pile identifier, time information, and risk level information is constructed based on the judgment result. This warning information is then sent to the charging pile management platform through the communication interface between the charging pile and the management platform.

[0037] In one embodiment, such as Figure 2 As shown, in step S10, the historical charging behavior information of the target charging pile is obtained, and a charging behavior profile model is constructed based on the historical charging behavior information, including: S101: Perform data cleaning and time alignment on historical charging behavior information to obtain behavior feature parameters.

[0038] Specifically, the historical charging behavior information recorded by the target charging pile within a set time period is obtained, and data items such as start time, end time, charging amount, voltage, and current of each charging are extracted. Records with null values, duplicate values, or abnormal values ​​are eliminated or corrected according to preset conditions. By comparing the recorded time with the system timestamp, time misalignment in the charging records is identified and corrected. All charging behavior records are organized using a uniform time granularity to ensure that all data have consistent time accuracy. The cleaned charging records are arranged according to the charging sequence, and behavioral characteristic values ​​such as charging duration, average voltage, average current, and unit time energy are calculated for each record. Finally, the values ​​extracted from each record are combined into multiple behavioral characteristic parameters.

[0039] S102: Construct charging behavior time series samples based on behavioral feature parameters, and use an unsupervised clustering algorithm to perform similarity clustering on the time series samples to identify typical patterns of charging behavior.

[0040] In this embodiment, unsupervised clustering algorithm refers to an algorithm that does not rely on pre-labeled data and automatically divides charging behavior time series samples into several classes by analyzing the similarity between data samples, so that samples in the same class have high similarity in the feature space and large differences between different classes. In this embodiment, K-means clustering algorithm can be preferred to train the behavior feature vector sequence to extract several typical patterns of representative charging behavior.

[0041] Specifically, the behavioral feature parameters obtained after data cleaning and time alignment are arranged in the order of charging occurrence to construct a time series sample containing multiple continuous charging behaviors. Each time series sample consists of behavioral features corresponding to multiple adjacent time points. After construction, the similarity between samples is calculated based on the numerical distribution and change trend of behavioral features in each time series sample using a distance metric. An unsupervised clustering method is selected to classify and group all time series samples. During the clustering process, no label information is preset; different charging behavior categories are dynamically divided only based on the feature similarity between samples. After clustering, representative charging behavior change patterns in historical behavior are identified according to the grouping results, and these patterns are used as typical charging behavior patterns of the target charging pile.

[0042] S103: Encode the typical patterns to construct a charging behavior profile model that represents the typical charging habits of the target charging pile.

[0043] Specifically, the typical charging behavior patterns identified in the clustering results are used as input data. Representative behavioral feature parameter combinations are extracted from each typical pattern, including but not limited to key dimension parameters such as charging duration range, voltage and current stability interval, and charging efficiency per unit time. These parameters are converted into feature vectors with fixed structures according to a unified encoding rule. Each dimension in the vector is assigned a numerical weight corresponding to its physical meaning to maintain consistency in expression between different dimensions. The multiple feature vectors after encoding are used as the basic elements for model construction. A mapping set is established by combining the feature vectors corresponding to various typical patterns. This mapping set is used to represent the charging behavior characteristics that repeatedly occur in different charging cycles of the target charging pile. Furthermore, this mapping set is summarized into a complete charging behavior profile model.

[0044] In one embodiment, such as Figure 3 As shown, in step S20, the current charging behavior information is input into the charging behavior profile model to obtain behavior deviation features, including: S201: Perform feature extraction processing on the current charging behavior information to construct the current behavior feature vector.

[0045] Specifically, charging data items consistent with the historical behavior modeling process are extracted from the current charging process record, including charging start time, charging end time, total charging duration, average voltage, average current, and charging amount per unit time. By preprocessing these raw charging data, the corresponding statistical values ​​and variation ranges are calculated to give different parameters a unified scale and dimensional structure. Then, according to the field order used when constructing the historical behavior feature parameters, the above-processed data are arranged in sequence to form a feature data group. Each dimension in the feature data group is normalized to maintain a consistent data format and numerical range with the typical feature vectors in the historical behavior profile model. Finally, the current charging behavior state is represented by a fixed dimensional structure to construct the current behavior feature vector.

[0046] S202: Calculate the Euclidean distance between the current behavior feature vector and the feature vector of the typical charging behavior pattern represented in the charging behavior profile model to obtain the behavior deviation metric.

[0047] Specifically, the distance between the constructed current behavior feature vector and the feature vectors of multiple typical charging behavior patterns encoded in the charging behavior profile model is calculated one by one. The Euclidean distance formula is used to square the numerical differences between the current behavior feature vector and each typical pattern feature vector in each dimension and sum them. Then, the square root of the sum is calculated to obtain the distance value between the two. The above steps are performed on all typical pattern vectors to obtain multiple Euclidean distance results. Each distance result is used to represent the degree of difference between the current behavior and the corresponding typical pattern. Finally, the minimum value is selected from all Euclidean distance results as a quantitative indicator of the degree of deviation of the current behavior, that is, the behavior deviation metric.

[0048] S203: Compare the behavior deviation metric with the set deviation judgment threshold. If the behavior deviation metric exceeds the set deviation judgment threshold, it is determined that the current charging behavior has an abnormal deviation, and the behavior deviation metric is used as the behavior deviation feature.

[0049] Specifically, the behavioral deviation metric calculated using Euclidean distance is numerically compared with a preset deviation judgment threshold. The deviation judgment threshold is determined by statistical analysis of a large amount of historical charging behavior deviation data and represents the maximum acceptable range of difference between the current behavior and the typical pattern. During the comparison, if the behavioral deviation metric is not greater than the deviation judgment threshold, the current charging behavior is considered normal and no behavioral deviation modeling is performed. If the behavioral deviation metric is greater than the deviation judgment threshold, the current charging behavior is considered to have deviated significantly from the historical typical pattern, and the behavioral deviation metric is directly used as the behavioral deviation feature. In one embodiment, such as Figure 4As shown, in step S202, the Euclidean distance is calculated between the current behavior feature vector and the feature vector of the typical charging behavior pattern represented in the charging behavior profile model to obtain the behavior deviation metric, including: S2021: Construct a reference behavior vector set based on multiple typical behavior feature vectors obtained from cluster training in the charging behavior profile model.

[0050] Specifically, several typical charging behavior patterns identified through unsupervised clustering algorithms during the construction of the charging behavior profile model are selected. Feature parameters are extracted from core charging record samples within each typical pattern. Extracted fields include parameters closely related to behavioral stability, such as charging duration, voltage, current, and unit charge. Following a unified encoding rule, these parameters are converted into feature vectors with the same dimensional structure. Each feature vector corresponds to a specific typical charging behavior pattern. All obtained typical behavior feature vectors are identified and combined according to cluster categories to form a complete set of reference behavior vectors. S2022: Calculate the Euclidean distance between the current behavior feature vector and each typical behavior feature vector in the reference behavior vector set, resulting in multiple distance results.

[0051] Specifically, for each feature value in the current behavior feature vector, the difference is calculated between it and the corresponding dimension value of each typical behavior feature vector in the reference behavior vector set. Then, the difference in all dimensions is squared, and the squared values ​​are accumulated. Finally, the square root operation is performed on the accumulated result to obtain the Euclidean distance between the current behavior feature vector and the typical behavior feature vector. The above steps are repeated to calculate the Euclidean distance between the current behavior feature vector and all typical behavior feature vectors in the reference behavior vector set one by one, resulting in multiple distance results representing the degree of difference between the current behavior and various historical typical patterns.

[0052] S2023: Based on the preset deviation evaluation rules, determine the behavioral deviation metric value from multiple distance results to characterize the degree of difference between the current behavior and the historical pattern.

[0053] In this embodiment, the preset deviation evaluation rule refers to the use of the minimum distance criterion in the process of constructing the behavior deviation metric, that is, selecting the distance with the smallest value among all Euclidean distance calculation results as the behavior deviation metric value between the current behavior and the historical typical behavior pattern.

[0054] Specifically, firstly, the Euclidean distance is calculated between the current behavior feature vector and each typical behavior feature vector in the reference behavior vector set, resulting in a distance result set containing multiple Euclidean distance values. Each distance value represents the degree of difference between the current behavior and a certain typical behavior pattern in the feature space. Then, each Euclidean distance value in the set is read sequentially, and the currently read distance value is compared with the recorded minimum distance value. If the current distance value is less than the recorded minimum distance value, the current distance value is used to replace the original minimum distance value. The same comparison operation is continued for the next distance value until all Euclidean distance values ​​are traversed. The minimum distance value retained is the behavior deviation metric.

[0055] In one embodiment, such as Figure 5 As shown, before step S40, i.e., the trained multimodal fault prediction model, the following steps are included: S401: Obtain the historical multimodal training sample set with labeled fault states, and divide the historical multimodal training sample set into a training subset and a validation subset according to a preset ratio.

[0056] In this embodiment, the historical multimodal training sample set refers to the data set collected and labeled during the long-term operation of the charging pile. Each training sample in this set contains multiple types of data covering the same time window, specifically including two types of information: charging behavior data and electrical operation data. The charging behavior data includes charging start time, duration, charging power changes, user operation mode, etc., while the electrical operation data includes raw electrical parameters such as voltage, current, power factor, and temperature. In addition, each sample also corresponds to a fault status label generated by the inspection of maintenance personnel or the automatic diagnosis results of the system.

[0057] Specifically, firstly, historical multimodal data covering normal and abnormal states is extracted from the charging pile backend monitoring system. The multimodal data includes charging behavior data and electrical operation data collected within a unified time window. The charging behavior data includes the start time, duration, and charging current change sequence of each charging task. The electrical operation data includes raw electrical parameters such as voltage, current, temperature, and power factor within the sampling period. At the same time, combined with fault record logs, corresponding fault labels marked by manual inspection or system diagnostic tools are attached to each sample to form labeled multimodal training samples. Then, the sample set is divided into a preset ratio of 80% and 20%. 80% of the samples form the training subset for parameter training of the neural network model, and 20% of the samples form the validation subset for performance evaluation during the training process.

[0058] S402: Based on the preset training configuration, input the training subset into the multilayer neural network model, and use the cross-entropy loss function to calculate the error between the output of the multilayer neural network model and the training subset.

[0059] Specifically, the model structure is first initialized according to the neural network framework. This multi-layer neural network model consists of an input layer, several hidden layers, and an output layer. The number of nodes in the input layer is consistent with the feature dimension of each sample in the training subset, and the number of nodes in the output layer corresponds to the number of fault level categories. Each hidden layer is configured with a preset number of neurons and uses ReLU as the activation function. The forward propagation path between the layers is constructed through a fully connected method. Before training begins, training configuration parameters are loaded, including batch size, learning rate, number of training rounds, optimization algorithm type, and regularization coefficient. Then, the training subset is divided into multiple small batches of samples according to the set batches. In each round of training, the small batches of samples are input into the model one by one. The model output results are obtained through the forward propagation process. The difference between the output results and the corresponding labels in the training subset is then input into the cross-entropy loss function to calculate the average loss value of each batch of samples. This loss value is used as a measure of the error between the current model prediction result and the true label.

[0060] S403: Based on the error, adjust the parameters of the multilayer neural network model using the backpropagation algorithm, and iteratively optimize the parameters of the multilayer neural network model using the stochastic gradient descent algorithm to obtain the multilayer neural network model under training.

[0061] Specifically, after completing the forward propagation of each batch of training subset samples and calculating the cross-entropy loss value, the partial derivatives of the loss value with respect to the parameters of each layer of the model are backpropagated using the chain rule. The error signal is passed from the output layer to the input layer in sequence, thereby obtaining the gradient information corresponding to each weight parameter. After obtaining the gradients of all parameters, the stochastic gradient descent algorithm is used to adjust the weights and bias parameters of each layer in the neural network. In each iteration, the current parameter value is subtracted from the product of the corresponding gradient and the learning rate according to the learning rate to realize the descent update of the loss function value. After all mini-batch samples have completed one traversal, the model parameter state and loss change of that round are recorded. The above forward and backpropagation iterative process is continuously executed until the set number of training rounds or loss convergence condition is reached. At this time, the model enters the training completion stage and forms a multi-layer neural network model in training.

[0062] S404: Determine whether the performance index change trend of the multi-layer neural network model under training on the validation subset has reached the preset convergence condition. If the convergence condition is reached, stop training and obtain the trained multi-modal fault prediction model.

[0063] Specifically, after each training round, forward inference is performed on the sample data in the validation subset using the updated model parameters of the current round. The cross-entropy loss value and fault classification accuracy on the validation subset of that round are recorded and calculated. A performance change curve is constructed based on the decrease in loss value and the increase in accuracy in consecutive training rounds. If the decrease in loss value is lower than the set minimum change threshold in several consecutive rounds, or if the accuracy does not improve in several consecutive rounds, the model is considered to have stabilized, and the preset convergence condition is met. If the convergence condition is met, the training process is terminated, and the model parameters of the current training round are saved as the final training result to form a trained multimodal fault prediction model. If the above convergence condition is not met, the next round of training is continued until the convergence criterion is met or the maximum number of training rounds is reached.

[0064] In one embodiment, such as Figure 6 As shown, in step S40, the behavioral deviation features and electrical features are input into the trained multimodal fault prediction model to obtain the fault risk level, including: S405: Embedding and mapping behavioral deviation features and electrical features to construct a unified multidimensional fusion feature vector.

[0065] Specifically, a first set of embedding mapping parameters is set according to the numerical dimension of the behavioral deviation features. A weight matrix is ​​used to perform linear projection on each dimension of the behavioral deviation features. The original behavioral deviation feature matrix and the embedding weight matrix are calculated by matrix multiplication to generate a first mapping result. Then, an activation function is used to process the first mapping result to introduce nonlinear expressive power, such as the ReLU function. A second set of embedding mapping parameters is set according to the numerical dimension of the electrical features. The matrix multiplication and activation function processing are performed in the same way as the behavioral deviation features to generate a second mapping result. Subsequently, the first mapping result and the second mapping result are concatenated along the feature dimension direction. The two embedded feature vectors are concatenated in sequence into a unified feature vector to construct a multidimensional fused feature vector.

[0066] Furthermore, the ReLU function is an activation function used to perform nonlinear mapping processing. By definition, the ReLU function performs a "maximum value" operation on the input data, that is, it compares the input data with zero element by element and takes the larger value. If an input is positive, it outputs the positive value; if an input is zero or negative, it outputs zero. Using this function can keep the gradient constant when the input is positive and return negative values ​​to zero, thereby achieving nonlinear expressive power and accelerating the convergence of neural networks.

[0067] S406: Input the fused feature vector into the trained multimodal fault prediction model. The trained multimodal fault prediction model performs fault risk inference on the multidimensional fused feature vector and outputs the fault risk level corresponding to the current charging pile status.

[0068] Specifically, the fused feature vector is input into the input layer of the trained multimodal fault prediction model according to a preset data format. The input layer completes the initial reception and normalization of the fused features, which are then passed to multiple fully connected hidden layers. In each hidden layer, linear weighting and nonlinear activation operations are performed. Linear weighting transforms the input features through a weight matrix and a bias term, while nonlinear activation is performed using the ReLU function. Higher-order feature information is extracted layer by layer, and dimensionality increase and semantic feature enhancement are completed. Finally, the output of the last hidden layer is sent to the output layer. The output layer calculates the prediction score corresponding to each fault risk level according to a preset classification structure, and uses the Softmax function to normalize multiple prediction scores into probability values. The risk level with the highest probability value is selected as the fault risk level of the current charging pile and output.

[0069] In one embodiment, such as Figure 7 As shown, in step S406, the fused feature vector is input into the trained multimodal fault prediction model. The trained multimodal fault prediction model performs fault risk inference on the multidimensional fused feature vector and outputs the fault risk level corresponding to the current charging pile status, including: S4061: Input the fused feature vector into the input layer of the trained multimodal fault prediction model, and then perform feature mapping and deep feature extraction through the fully connected hidden layer of the trained multimodal fault prediction model to obtain high-dimensional semantic features.

[0070] Specifically, the fused feature vector is input into the input layer of the trained multimodal fault prediction model according to a set format. Normalization is performed on the input data in the input layer, including linear scaling of the feature values ​​in each dimension to ensure they fall within a predefined numerical range, facilitating stable training and inference calculations in subsequent network layers. The normalized fused feature vector is then sequentially passed to multiple fully connected hidden layers. In each hidden layer, the input vector is first multiplied by the corresponding weight matrix using matrix multiplication, with a bias term added. The resulting linear transformation is then fed into an activation function for nonlinear transformation. The ReLU function is used, specifically performing a max(0,x) operation on each element to suppress the propagation of negative signals and retain positive values. After continuous mapping and nonlinear processing through multiple hidden layers, a more discriminative deep feature representation is extracted, ultimately yielding high-dimensional semantic features with high-order semantic feature expression capabilities. S4062: Based on the high-dimensional semantic features, the risk level determination structure in the output layer of the trained multimodal fault prediction model is used to calculate the risk level output score, and the fault risk level is determined based on the risk level output score.

[0071] Specifically, the high-dimensional semantic feature vector is input into the output layer of the trained multimodal fault prediction model. The output layer contains a classification structure for risk level determination. The classification structure includes a fully connected output node set and a Softmax normalization function. The fully connected output node set receives the high-dimensional semantic feature vector and multiplies it with a preset output weight matrix by matrix multiplication, then adds a bias vector to generate an initial score set corresponding to each fault risk level. The Softmax normalization function performs exponential mapping and normalization on the initial score set to calculate the probability value of each fault risk level. Specifically, it performs an exponential operation on each initial score and divides it by the sum of the exponents of all score values ​​to obtain the normalized output score. Finally, the risk level with the highest output score is taken as the fault risk level corresponding to the current charging pile status.

[0072] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0073] In one embodiment, a fault prediction device for new energy charging piles is provided, which corresponds one-to-one with the fault prediction method for new energy charging piles described in the above embodiments. For example... Figure 8 As shown, this new energy charging pile fault prediction device includes a behavior profiling modeling module, a behavior deviation analysis module, an electrical feature extraction module, and an early warning information transmission module. Detailed descriptions of each functional module are as follows: The behavior profiling modeling module is used to obtain historical charging behavior information of the target charging pile and build a charging behavior profiling model based on the historical charging behavior information. The behavior deviation analysis module is used to collect current charging behavior information in real time during the operation of the target charging pile, and input the current charging behavior information into the charging behavior profile model to obtain behavior deviation characteristics; The electrical feature extraction module is used to acquire the electrical operation information of the target charging pile, perform normalization processing and time series analysis on the operation information, and obtain electrical features. The early warning information sending module is used to input behavioral deviation features and electrical features into the trained multimodal fault prediction model to obtain the fault risk level. If the fault risk level exceeds the set early warning threshold, the corresponding early warning information is generated and sent to the charging pile management platform.

[0074] Optional, the behavioral profiling modeling module includes: The behavior data preprocessing submodule is used to clean and time-align historical charging behavior information to obtain behavior feature parameters. The typical pattern recognition submodule is used to construct charging behavior time series samples based on behavioral feature parameters, and to perform similarity clustering on the time series samples using an unsupervised clustering algorithm to identify typical patterns of charging behavior. The profile model construction submodule is used to encode the features of typical patterns and construct a charging behavior profile model to represent the typical charging habits of the target charging pile.

[0075] Optional, the behavioral deviation analysis module includes: The current behavior feature extraction submodule is used to extract features from the current charging behavior information and construct the current behavior feature vector. The behavior deviation calculation submodule is used to calculate the Euclidean distance between the current behavior feature vector and the feature vector of the typical charging behavior pattern represented in the charging behavior profile model to obtain the behavior deviation metric. The behavior anomaly determination submodule is used to compare the behavior deviation metric with the set deviation determination threshold. If the behavior deviation metric exceeds the set deviation determination threshold, the current charging behavior is determined to have an abnormal deviation, and the behavior deviation metric is used as the behavior deviation feature.

[0076] Optional, the behavior deviation calculation submodule includes: The reference behavior vector construction unit is used to construct a set of reference behavior vectors based on multiple typical behavior feature vectors obtained by cluster training in the charging behavior profile model. The Euclidean distance calculation unit is used to calculate the Euclidean distance between the current behavior feature vector and each typical behavior feature vector in the reference behavior vector set, resulting in multiple distance results. The behavior deviation assessment unit is used to determine the behavior deviation metric value, which characterizes the degree of difference between the current behavior and the historical pattern, from multiple distance results according to the preset deviation evaluation rules.

[0077] Optionally, the warning information sending module includes the following: The sample set preparation module is used to obtain a historical multimodal training sample set with labeled fault states, and divide the historical multimodal training sample set into a training subset and a validation subset according to a preset ratio; The model training configuration module is used to input the training subset into the multilayer neural network model according to the preset training configuration, and to calculate the error between the output of the multilayer neural network model and the training subset using the cross-entropy loss function. The parameter optimization module is used to adjust the parameters of the multilayer neural network model based on the error using the backpropagation algorithm, and to iteratively optimize the parameters of the multilayer neural network model using the stochastic gradient descent algorithm to obtain the multilayer neural network model under training. The training convergence judgment module is used to determine whether the performance index change trend of the multi-layer neural network model in training on the validation subset has reached the preset convergence condition. If the convergence condition is reached, training is stopped and a trained multimodal fault prediction model is obtained.

[0078] Optionally, the early warning information sending module includes: The feature fusion construction submodule is used to embed and map behavioral deviation features and electrical features to construct a unified multi-dimensional fusion feature vector; The fault risk inference submodule is used to input the fused feature vector into the trained multimodal fault prediction model. The trained multimodal fault prediction model performs fault risk inference on the multidimensional fused feature vector and outputs the fault risk level corresponding to the current charging pile status.

[0079] Optionally, the fault risk reasoning submodule includes: The deep feature extraction submodule is used to input the fused feature vector into the input layer of the trained multimodal fault prediction model, and then perform feature mapping and deep feature extraction through the fully connected hidden layer of the trained multimodal fault prediction model to obtain high-dimensional semantic features. The risk level determination submodule is used to calculate the risk level output score based on the risk level determination structure in the output layer of the trained multimodal fault prediction model, according to the high-dimensional semantic features, and to determine the fault risk level based on the risk level output score.

[0080] For specific limitations regarding the fault prediction device for new energy charging piles, please refer to the limitations of the fault prediction method for new energy charging piles mentioned above, which will not be repeated here. Each module in the aforementioned fault prediction device for new energy charging piles can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0081] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 9As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a method for predicting faults in new energy charging piles.

[0082] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps: Obtain historical charging behavior information of the target charging pile, and build a charging behavior profile model based on the historical charging behavior information; During the operation of the target charging pile, the current charging behavior information is collected in real time and input into the charging behavior profile model to obtain the behavior deviation characteristics; Obtain the electrical operation information of the target charging pile, normalize the operation information and perform time series analysis to obtain electrical characteristics; The behavioral deviation features and electrical features are input into the trained multimodal fault prediction model to obtain the fault risk level. If the fault risk level exceeds the set warning threshold, the corresponding warning information is generated and sent to the charging pile management platform.

[0083] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor: Obtain historical charging behavior information of the target charging pile, and build a charging behavior profile model based on the historical charging behavior information; During the operation of the target charging pile, the current charging behavior information is collected in real time and input into the charging behavior profile model to obtain the behavior deviation characteristics; Obtain the electrical operation information of the target charging pile, normalize the operation information and perform time series analysis to obtain electrical characteristics; The behavioral deviation features and electrical features are input into the trained multimodal fault prediction model to obtain the fault risk level. If the fault risk level exceeds the set warning threshold, the corresponding warning information is generated and sent to the charging pile management platform.

[0084] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0085] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0086] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A new energy charging pile fault prediction method, characterized in that, The new energy charging pile fault prediction method comprises: obtaining historical charging behavior information of a target charging pile, and constructing a charging behavior portrait model based on the historical charging behavior information; during operation of the target charging pile, real-time collection of current charging behavior information, and input of the current charging behavior information into the charging behavior portrait model to obtain behavior deviation characteristics; obtaining electrical operation information of the target charging pile, normalizing and time series analyzing the operation information to obtain electrical characteristics; inputting the behavior deviation characteristics and the electrical characteristics into a trained multi-modal fault prediction model to obtain a fault risk level, and if the fault risk level exceeds a set warning threshold, generating corresponding warning information and sending the warning information to a charging pile management platform.

2. The new energy charging pile fault prediction method according to claim 1, characterized in that, The method comprises: data cleaning and time alignment processing of the historical charging behavior information to obtain behavior characteristic parameters; constructing charging behavior time series samples based on the behavior characteristic parameters, using an unsupervised clustering algorithm to perform similarity clustering on the time series samples, and identifying typical modes of charging behavior; feature encoding of the typical modes to construct the charging behavior portrait model for representing typical charging habits of the target charging pile.

3. The new energy charging pile fault prediction method according to claim 1, characterized in that, The method comprises: feature extraction processing of the current charging behavior information to construct a current behavior feature vector; Euclidean distance calculation of the current behavior feature vector and the feature vector of the typical charging behavior mode represented in the charging behavior portrait model to obtain a behavior deviation measurement value; comparison of the behavior deviation measurement value with a set deviation judgment threshold, and if the behavior deviation measurement value exceeds the set deviation judgment threshold, it is determined that the current charging behavior deviates abnormally, and the behavior deviation measurement value is taken as the behavior deviation characteristic.

4. The new energy charging pile fault prediction method according to claim 3, characterized in that, The method comprises: constructing a reference behavior vector set based on a plurality of typical behavior feature vectors obtained by clustering training in the charging behavior portrait model; Euclidean distance calculation of the current behavior feature vector and each typical behavior feature vector in the reference behavior vector set to obtain a plurality of distance results; determination of the behavior deviation measurement value representing the difference degree between the current behavior and the historical mode from the plurality of distance results according to a preset deviation evaluation rule.

5. The new energy charging pile fault prediction method according to claim 1, characterized in that, The trained multi-modal fault prediction model previously comprises: obtaining a historical multi-modal training sample set with labeled fault states, and dividing the historical multi-modal training sample set into a training subset and a validation subset according to a preset proportion; According to a preset training configuration, the training subset is input into a multi-layer neural network model, and a cross-entropy loss function is used to calculate the error between the output of the multi-layer neural network model and the training subset; According to the error, the parameters of the multi-layer neural network model are adjusted by a back propagation algorithm, and the parameters of the multi-layer neural network model are iteratively optimized by a stochastic gradient descent algorithm to obtain a trained multi-layer neural network model; It is judged whether the performance index change trend of the trained multi-layer neural network model on the verification subset reaches a preset convergence condition, and if the convergence condition is reached, the training is stopped to obtain the trained multi-modal fault prediction model.

6. The new energy charging pile fault prediction method according to claim 1, characterized in that, The behavior deviation feature and the electrical feature are input into the trained multi-modal fault prediction model to obtain a fault risk level, which comprises: The behavior deviation feature and the electrical feature are embedded and mapped to construct a unified multi-dimensional fusion feature vector; The fusion feature vector is input into the trained multi-modal fault prediction model, and the trained multi-modal fault prediction model performs fault risk reasoning on the multi-dimensional fusion feature vector to output the fault risk level corresponding to the current charging pile state.

7. The new energy charging pile fault prediction method according to claim 6, characterized in that, The fusion feature vector is input into the trained multi-modal fault prediction model, and the trained multi-modal fault prediction model performs fault risk reasoning on the multi-dimensional fusion feature vector to output the fault risk level corresponding to the current charging pile state, which comprises: The fusion feature vector is input into the input layer of the trained multi-modal fault prediction model, and is sequentially mapped and deep feature extracted by the fully connected hidden layer of the trained multi-modal fault prediction model to obtain high-dimensional semantic features; According to the high-dimensional semantic features, the risk level determination structure in the output layer of the trained multi-modal fault prediction model is used to calculate the risk level output score, and based on the risk level output score, the fault risk level is determined.

8. A new energy charging pile fault prediction device, characterized in that, The new energy charging pile fault prediction device comprises: A behavior portrait modeling module is configured to acquire historical charging behavior information of a target charging pile and construct a charging behavior portrait model based on the historical charging behavior information; A behavior deviation analysis module is configured to acquire current charging behavior information in real time during operation of the target charging pile and input the current charging behavior information into the charging behavior portrait model to obtain behavior deviation features; An electrical feature extraction module is configured to acquire electrical operation information of the target charging pile, perform normalization processing and time series analysis on the operation information, and obtain electrical features; An early warning information sending module is configured to input the behavior deviation features and the electrical features into a trained multi-modal fault prediction model to obtain a fault risk level, and if the fault risk level exceeds a set early warning threshold, corresponding early warning information is generated and sent to a charging pile management platform.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, The processor executes the computer program to implement the steps of the new energy charging pile fault prediction method according to any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 9. The computer program is executed by the processor to implement the steps of the new energy charging pile fault prediction method according to any one of claims 1 to 7.

Images