Charging station charging pile anomaly detection method based on deep learning

Abstract

The invention discloses a charging station charging pile anomaly detection method based on deep learning. The method comprises the following steps: carrying out data preprocessing on current and voltage of a charging pile and data collected by a sensor; based on the preprocessed data, carrying out unsupervised learning on the time sequence lightweight adaptive network TSLANet, and learning the representation of the time sequence; on the basis of the pre-training model, binary classification model training is carried out; in the online prediction process, if abnormity exists, performing system alarm on the equipment, secondarily judging whether the equipment is abnormal or not by an operator, marking abnormal time sequence data, and storing a sample and a label in a database; the anomaly detection model is updated in a timed training mode; manual auditing is carried out on the basis that abnormity is detected by the model, the accuracy of abnormal sample capturing can be improved, the manual marking range is reduced, and then the labor cost of marking is reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of non-electrical variable detection of charging piles, and relates to a method for detecting abnormalities of charging piles in a charging station based on deep learning. Background Art

[0002] By detecting abnormal current and voltage of charging piles, safety accidents such as electrical fires and electric shocks can be effectively reduced. By detecting abnormal algorithms, potential safety hazards can be discovered in time, and alarms can be sent to charging station operation managers, so that measures can be taken to prevent accidents and ensure the safety of users and the public. The efficient and stable operation of charging piles directly affects the user experience. Abnormal detection and timely maintenance can ensure the good performance of charging piles and improve user satisfaction and loyalty to charging services. In addition, by performing regular abnormal detection and maintenance on charging piles, their service life can be effectively extended. This not only reduces the cost of frequent equipment replacement, but also helps to keep the equipment operating efficiently.

[0003] The current and voltage data of charging piles are mainly collected by industrial sensors. The abnormal detection of charging piles is mainly carried out by the operation and maintenance personnel of the charging station through regular or real-time monitoring and on-duty to troubleshoot problems. The operation and maintenance personnel rely on experience to judge and handle abnormal conditions by observing the operating conditions and related operating data. When external conditions and operating conditions change, it is difficult to predict, judge and handle abnormal conditions in a timely and accurate manner. Such anomaly detection methods are not only time-consuming and labor-intensive, but also difficult to effectively utilize a large amount of charging pile data, and the monitoring and prediction methods are not scalable and universal.

[0004] Traditional anomaly detection modeling methods mainly optimize control by establishing a mechanism model of charging pile equipment. However, the actual controlled process contains multiple variables, strong nonlinearity, and uncertainty, which makes it difficult to perform digital mechanism modeling, or the modeling accuracy is low, and the method is not universal. The anomaly detection method based on the mechanism model lacks a systematic review and investigation of the historical operation process of the charging pile equipment, and cannot detect equipment anomalies that may occur during the operation of the charging pile in real time.

[0005] Data-driven anomaly detection models require a large number of data samples and sample labels. Although anomaly detection models can dynamically and in real time detect anomalies in charging pile equipment, manual sample labeling is time-consuming, labor-intensive, and costly. On the other hand, data-driven machine learning models usually require a large number of abnormal sample point labels, which are expensive and difficult to obtain. Even if there are a large number of labeled data samples, feature engineering of data samples is time-consuming and labor-intensive, and greatly affects the effectiveness of the model.

[0006] 1. Reliance on manual experience and real-time monitoring: Traditional anomaly detection methods rely on operation and maintenance personnel to judge anomalies through observation and experience. This method is time-consuming and labor-intensive, and it is difficult to respond promptly and accurately when external conditions and operating conditions change. This method that relies on manual labor is inefficient and prone to misjudgment.

[0007] 2. Limitations of mechanism models: Modeling methods based on mechanism models have difficulties in dealing with multiple variables, strong nonlinearity and uncertainty, which makes it difficult to accurately model or the modeling accuracy is low. Methods based on mechanism models usually lack systematic analysis and review of the historical operation data of charging pile equipment, which limits the depth and breadth of fault prediction and diagnosis. In addition, mechanism models are not universal and are difficult to adapt to the needs of different types or new types of charging piles.

[0008] 3. Challenges of data-driven models: Although data-driven models can provide dynamic and real-time anomaly detection, this method requires a large number of data samples and manual labeling of abnormal samples, which is not only time-consuming and labor-intensive, but also costly. The difficulty and cost of data labeling limit the widespread application of this method.

[0009] 4. Data sample quality and feature engineering: Even if enough data samples are obtained, feature engineering of data is a complex and time-consuming process, which directly affects the performance and effect of the model. The quality of feature extraction and selection is a key factor in determining the success of the model. Summary of the invention

[0010] The technical problem to be solved by the present invention is to provide a charging pile anomaly detection method for a charging station based on deep learning, which makes full use of online abnormal samples and reduces the cost of personnel on duty.

[0011] In order to solve the above technical problems, the technical solution adopted by the present invention is: a method for detecting abnormalities of charging piles in a charging station based on deep learning, comprising the following steps: S1, data preprocessing: preprocessing the data collected by the current, voltage and sensor of the charging pile; S2, unsupervised learning; based on the preprocessed data, the time series lightweight adaptive network TSLANet is unsupervisedly learned to learn the representation of the time series; the adaptive network TSLANet first converts the time series data into PatchEmbedding, and converts the one-dimensional time series data into a two-dimensional structure; specifically: according to a window length and step size, it moves on the original time series, extracts multiple subsequences, organizes multiple subsequences into a patch, and then inputs the patch time series data into the Embedding layer for Embedding; in order to retain the time order information in the time series, positional encoding Positional Embedding is added to each patch; positional encoding is a learned parameter, or a fixed pattern based on sine and cosine functions; after the input time series Embedding extracts the corresponding features through N blocks, it passes through the linear layer to solve different types of problems, including regression or classification; S3, classification model training: Based on the pre-trained model, a binary classification model is trained. The classification model marks abnormal time series data as positive and normal time series data as negative. S4, prediction: In the online prediction process, the current and voltage data of the charging pile are first preprocessed in real time; S5, judgment: sending the processed data to the classification model to determine whether there is an abnormality; S6, system alarm; if there is an abnormality, the system alarm will be issued, and the operator will make a second judgment on whether it is abnormal, mark the abnormal time series data, and store the samples and labels in the database; S7, abnormal sample processing; the anomaly detection model is updated through scheduled training; the unsupervised model and binary classification model are retrained on a weekly or monthly basis based on the accumulated abnormal samples.

[0012] In S2, Block includes two structures, one is Adaptive SpectralBlock, namely Adaptive SpectralBlock ASB, and the other is Interactive Convolution Block, namely Interactive Convolution Block ICB.

[0013] In S2, the adaptive spectrum block ASB first performs fast Fourier transformations on the input time series features, and its formula is:

[0014] in, is a one-dimensional FFT operation, is the input multi-channel time series data; output Represents the characteristics of the original time series in the frequency domain.

[0015] In S2, the adaptive spectrum block ASB removes high-frequency noise; first, calculate The power spectrum of the data is used to identify the dominant frequency components. A learning parameter theta is then trained as the threshold. The threshold is adjusted according to the spectral characteristics of the data. Frequencies with power spectra higher than theta are retained, while other frequencies are filtered. The threshold theta selects the frequency threshold adaptively, and the adaptive spectrum block ASB customizes its filtering threshold according to each time series data set to improve the overall effectiveness of the model in processing various data scenarios. The formula is as follows:

[0016]

[0017] Among them, P is the frequency domain feature The power spectrum of is the frequency threshold.

[0018] In S2, two sets of learnable filters are used, one global filter Learn from the original frequency domain features, another local filter From the frequency domain features after adaptive filtering Then the filtered features are integrated; ;

[0019]

[0020] in are respectively learnable filter parameters; Finally, through inverse fast Fourier transform, From the frequency domain to the time domain, the output is .

[0021] In S2, the interactive convolution block ICB contains multiple parallel convolution blocks with different convolution kernel sizes to capture layout features and long-range dependencies; specifically, the first convolution layer is designed to capture fine-grained, local patterns in the data with a small kernel; the second layer is designed to identify broader and longer-range dependencies with a larger kernel; its formula is as follows:

[0022] ; ; in, They represent convolutional layers with different convolution kernel sizes. is the output of the interactive convolutional block ICB.

[0023] In S7, the anomaly detection step includes unsupervised training, binary classification model training, and online prediction.

[0024] In S7, unsupervised training involves pre-training on time series data using TSLANet’s self-supervised learning mechanism. In this phase, the model learns the intrinsic structure of the time series by masking a portion of the time series and attempting to reconstruct the masked portions. This part of the training uses the reconstruction error of the patch, i.e., the mean squared error (MSE), as the loss function and stochastic gradient descent as the derivation method. The mean squared error is the average of the squares of the differences between the predicted values ​​and the true values. It measures the average square of the difference between the model’s predicted values ​​and the true values. ; in, , Respectively refer to i The true value of the sample and the model predicted value.

[0025] In S7, the binary classification model training includes taking abnormal time series data as positive samples and non-abnormal time series data as negative samples. Due to the small number of positive samples and the imbalance between positive and negative samples, undersampling is used to randomly reduce the number of negative samples so that the ratio of positive and negative samples reaches 1:1. TSLANet is used for binary classification model training, and the loss function is Binary Cross-Entropy (BCE), and the stochastic gradient descent method is used as the derivation method. ; in, , Respectively refer to i The true value of the sample and the model predicted value.

[0026] In S7, online prediction includes first preprocessing the current and voltage data of the charging pile in real time, and then sending the processed data to the binary classification model to determine its prediction probability. Is it greater than 0.5? If it is greater than 0.5, there may be an abnormality and the system will alarm. The operator will then make a second judgment on whether it is abnormal, mark the abnormal time series data, store the samples and labels in the database, and accumulate the abnormal samples.

[0027] The main beneficial effects of the present invention are: It can make full use of the valuable data accumulated by the enterprise, with low data processing cost and low computing overhead. It is suitable for various industrial anomaly detection scenarios and has good scalability and versatility.

[0028] Unsupervised learning can effectively learn the characteristics of time series data itself and form a good feature extraction.

[0029] The anomaly detection model is learned based on the pre-trained model, which reduces the computational overhead of training and only requires fine-tuning based on the abnormal sample data.

[0030] The anomaly detection model is updated through regular training, which can make full use of online abnormal samples and reduce the cost of personnel on duty.

[0031] Manual review based on the anomalies detected by the model can improve the accuracy of capturing abnormal samples, reduce the scope of manual labeling, and thus reduce the labor cost of labeling. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0033] Figure 1 It is a flow chart of the present invention.

[0034] Figure 2 It is a structural diagram of the adaptive network TSLANet of the present invention. DETAILED DESCRIPTION

[0035] like Figure 1~Figure 2 A method for detecting abnormalities of charging piles in a charging station based on deep learning includes the following steps: S1, data preprocessing: preprocessing the data collected by the current, voltage and sensor of the charging pile; S2, unsupervised learning; based on the preprocessed data, the time series lightweight adaptive network TSLANet is unsupervisedly learned to learn the representation of the time series; the adaptive network TSLANet first converts the time series data into PatchEmbedding, and converts the one-dimensional time series data into a two-dimensional structure; specifically: according to a window length and step size, it moves on the original time series, extracts multiple subsequences, organizes multiple subsequences into a patch, and then inputs the patch time series data into the Embedding layer for Embedding; in order to retain the time order information in the time series, positional encoding Positional Embedding is added to each patch; positional encoding is a learned parameter, or a fixed pattern based on sine and cosine functions; after the input time series Embedding extracts the corresponding features through N blocks, it passes through the linear layer to solve different types of problems, including regression or classification; S3, classification model training: Based on the pre-trained model, a binary classification model is trained. The classification model marks abnormal time series data as positive and normal time series data as negative. S4, prediction: In the online prediction process, the current and voltage data of the charging pile are first preprocessed in real time; S5, judgment: sending the processed data to the classification model to determine whether there is an abnormality; S6, system alarm; if there is an abnormality, the system alarm will be issued, and the operator will make a second judgment on whether it is abnormal, mark the abnormal time series data, and store the samples and labels in the database; S7, abnormal sample processing; the anomaly detection model is updated through scheduled training; the unsupervised model and binary classification model are retrained on a weekly or monthly basis based on the accumulated abnormal samples.

[0036] In a preferred solution, in S2, Block includes two structures, one is an adaptive spectrum block Adaptive Spectral Block, namely, adaptive spectrum block ASB, and the other is an interactive convolution block Interactive Convolution Block, namely, interactive convolution block ICB.

[0037] In the preferred solution, in S2, the adaptive spectrum block ASB first performs Fast Fourier Transformations on the input time series features, and the formula is:

[0038] in, is a one-dimensional FFT operation, is the input multi-channel time series data; output Represents the characteristics of the original time series in the frequency domain.

[0039] In the preferred solution, in S2, the adaptive spectrum block ASB removes high-frequency noise; first, calculate The power spectrum of the data is used to identify the dominant frequency components. A learning parameter theta is then trained as the threshold. The threshold is adjusted according to the spectral characteristics of the data. Frequencies with power spectra higher than theta are retained, while other frequencies are filtered. The threshold theta selects the frequency threshold adaptively, and the adaptive spectrum block ASB customizes its filtering threshold according to each time series data set to improve the overall effectiveness of the model in processing various data scenarios. The formula is as follows:

[0040]

[0041] Among them, P is the frequency domain feature The power spectrum of is the frequency threshold.

[0042] In the preferred solution, in S2, two sets of learnable filters are used, one global filter Learn from the original frequency domain features, another local filter From the frequency domain features after adaptive filtering Then the filtered features are integrated; ;

[0043]

[0044] in are respectively learnable filter parameters; Finally, through inverse fast Fourier transform, From the frequency domain to the time domain, the output is .

[0045] In the preferred solution, in S2, the interactive convolution block ICB contains multiple parallel convolution blocks with different convolution kernel sizes to capture layout features and long-range dependencies; specifically: the first convolution layer is designed to capture fine-grained, local patterns in the data with a small kernel; the second layer is designed to identify broader, longer-range dependencies with a larger kernel; the formula is as follows:

[0046] ; ; in, They represent convolutional layers with different convolution kernel sizes. is the output of the interactive convolutional block ICB.

[0047] Preferably, the above method can make full use of the valuable data accumulated by the enterprise, has low data processing cost and low computing overhead, is suitable for various industrial anomaly detection scenarios, and has good scalability and versatility.

[0048] Unsupervised learning can effectively learn the characteristics of time series data itself and form a good feature extraction.

[0049] In a preferred solution, in S7, the anomaly detection step includes unsupervised training, binary classification model training and online prediction.

[0050] Preferably, in the above method, the anomaly detection model is learned on the basis of a pre-trained model, the computational overhead of training is reduced, and only fine-tuning according to the abnormal sample data is required.

[0051] The anomaly detection model is updated through regular training, which can make full use of online abnormal samples and reduce the cost of personnel on duty.

[0052] In the preferred scheme, in S7, the unsupervised training includes pre-training the time series data using the self-supervised learning mechanism of TSLANet; in this stage, the model learns the intrinsic structure of the time series by masking a part of the time series and trying to reconstruct these masked parts; this part of the training uses the reconstruction error of the patch, namely the mean square error MeanSquared Error, MSE as the loss function, and the stochastic gradient descent method as the derivation method; the mean square error is the average of the squares of the differences between the predicted values ​​and the true values; it measures the squares of the average differences between the model's predicted values ​​and the true values; ; in, , Respectively refer to i The true value of the sample and the model predicted value.

[0053] In the preferred solution, in S7, the binary classification model training includes taking abnormal time series data as positive samples and taking non-abnormal time series data as negative samples; due to the small number of positive samples and the imbalance of positive and negative samples, undersampling is used to randomly reduce the number of negative samples so that the ratio of positive and negative samples reaches 1:1; TSLANet is used for binary classification model training, the loss function is Binary Cross-Entropy loss function, BCE, and the stochastic gradient descent method is used as the derivation method; ; in, , Respectively refer to i The true value of the sample and the model predicted value.

[0054] In the preferred solution, in S7, the online prediction includes first pre-processing the current and voltage data of the charging pile in real time, and then sending the processed data to the binary classification model to determine its prediction probability. Is it greater than 0.5? If it is greater than 0.5, there may be an abnormality and the system will alarm. The operator will then make a second judgment on whether it is abnormal, mark the abnormal time series data, store the samples and labels in the database, and accumulate the abnormal samples.

[0055] In the above method, manual review is performed based on the anomalies detected by the model, which can improve the accuracy of capturing abnormal samples, reduce the scope of manual labeling, and thus reduce the labor cost of labeling.

[0056] The above-mentioned embodiments are only preferred technical solutions of the present invention and should not be regarded as limitations of the present invention. The embodiments and features in the embodiments of the present application can be arbitrarily combined with each other without conflict. The protection scope of the present invention shall be the technical solutions recorded in the claims, including the equivalent replacement solutions of the technical features in the technical solutions recorded in the claims. That is, equivalent replacement improvements within this scope are also within the protection scope of the present invention.

Claims

1. A method for detecting abnormalities of charging piles in charging stations based on deep learning, characterized in that: The steps include: S1, data preprocessing: preprocessing the data collected by the current, voltage and sensor of the charging pile; S2, unsupervised learning: Based on the preprocessed data, the time series lightweight adaptive network TSLANet is unsupervisedly learned to learn the representation of the time series; The adaptive network TSLANet first converts the time series data into PatchEmbedding, and transforms the one-dimensional time series data into a two-dimensional structure. Specifically, it moves on the original time series according to a window length and step size, extracts multiple subsequences, organizes multiple subsequences into a patch, and then inputs the patch time series data into the Embedding layer for Embedding. In order to retain the time order information in the time series, positional encoding Positional Embedding is added to each patch. Positional encoding is a learned parameter or a fixed pattern based on sine and cosine functions. After the input time series Embedding extracts the corresponding features through N blocks, it passes through the linear layer to solve different types of problems, including regression or classification. S3, classification model training: Based on the pre-trained model, a binary classification model is trained. The classification model marks abnormal time series data as positive and normal time series data as negative. S4, prediction: In the online prediction process, the current and voltage data of the charging pile are first preprocessed in real time; S5, judgment: sending the processed data to the classification model to determine whether there is an abnormality; S6, system alarm; if there is an abnormality, the system alarm will be issued, and the operator will make a second judgment on whether it is abnormal, mark the abnormal time series data, and store the samples and labels in the database; S7, abnormal sample processing; the anomaly detection model is updated through scheduled training; the unsupervised model and binary classification model are retrained on a weekly or monthly basis based on the accumulated abnormal samples.

2. The method for detecting abnormalities of charging piles in charging stations based on deep learning according to claim 1 is characterized in that: In S2, Block includes two structures, one is Adaptive Spectral Block, namely Adaptive Spectral Block ASB, and the other is Interactive Convolution Block, namely Interactive Convolution Block ICB.

3. The method for detecting abnormalities of charging piles in charging stations based on deep learning according to claim 2 is characterized in that: In S2, the adaptive spectrum block ASB first performs fast Fourier transformation on the input time series features, and its formula is: in, is a one-dimensional FFT operation, is the input multi-channel time series data; output Represents the characteristics of the original time series in the frequency domain.

4. The method for detecting abnormalities of charging piles in charging stations based on deep learning according to claim 2 is characterized in that: In S2, the adaptive spectrum block ASB removes high-frequency noise; first, calculate The power spectrum of the data is used to identify the dominant frequency components. A learning parameter theta is then trained as the threshold. The threshold is adjusted according to the spectral characteristics of the data. Frequencies with power spectra higher than theta are retained, while other frequencies are filtered; Threshold theta selects the frequency threshold adaptively, and the adaptive spectrum block ASB customizes its filtering threshold according to each time series data set, improving the overall effectiveness of the model in processing various data scenarios; The formula is as follows: Among them, P is the frequency domain feature The power spectrum of is the frequency threshold.

5. The method for detecting abnormalities of charging piles in charging stations based on deep learning according to claim 2 is characterized in that: In S2, two sets of learnable filters are used, one global filter Learn from the original frequency domain features, another local filter From the frequency domain features after adaptive filtering Then the filtered features are integrated; ； in are the learnable filter parameters respectively; Finally, through inverse fast Fourier transform, From the frequency domain to the time domain, the output is .

6. The method for detecting abnormalities of charging piles in charging stations based on deep learning according to claim 2 is characterized in that: In S2, the interactive convolution block ICB contains multiple parallel convolution blocks with different convolution kernel sizes to capture layout features and long-range dependencies; specifically, the first convolution layer is designed to capture fine-grained, local patterns in the data with a small kernel; the second layer is designed to identify broader and longer-range dependencies with a larger kernel; its formula is as follows: ； ； in, They represent convolutional layers with different convolution kernel sizes. is the output of the interactive convolutional block ICB.

7. The method for detecting abnormalities of charging piles in charging stations based on deep learning according to claim 1 is characterized in that: In S7, the anomaly detection step includes unsupervised training, binary classification model training, and online prediction.

8. The method for detecting abnormalities of charging piles in charging stations based on deep learning according to claim 7 is characterized in that: In S7, unsupervised training includes pre-training on time series data using TSLANet’s self-supervised learning mechanism. In this stage, the model learns the intrinsic structure of the time series by masking a portion of the time series and trying to reconstruct these masked portions. This part of the training uses the reconstruction error of the patch, i.e., Mean Squared Error (MSE), as the loss function and stochastic gradient descent as the derivation method. The mean squared error is the average of the squares of the differences between the predicted values ​​and the true values. It measures the average square of the difference between the model’s predicted values ​​and the true values. ； in, , Respectively refer to i The true value of the sample and the model predicted value.

9. The method for detecting abnormalities of charging piles in charging stations based on deep learning according to claim 7 is characterized in that: In S7, the binary classification model training includes taking abnormal time series data as positive samples and non-abnormal time series data as negative samples. Due to the small number of positive samples and the imbalance between positive and negative samples, undersampling is used to randomly reduce the number of negative samples so that the ratio of positive and negative samples reaches 1:

1. TSLANet is used for binary classification model training, and the loss function is the binary cross entropy loss function, BCE, and the stochastic gradient descent method is used as the derivation method. ； in, , Respectively refer to i The true value of the sample and the model predicted value.

10. The method for detecting abnormalities of charging piles in charging stations based on deep learning according to claim 7 is characterized in that: In S7, online prediction includes first preprocessing the current and voltage data of the charging pile in real time, and then sending the processed data to the binary classification model to determine its prediction probability. Is it greater than 0.5? If it is greater than 0.5, there may be an abnormality and the system will alarm. The operator will then make a second judgment on whether it is abnormal, mark the abnormal time series data, store the samples and labels in the database, and accumulate the abnormal samples.

Images