An artificial intelligence-based charging pile operation state internet of things management and control monitoring method

Abstract

The application relates to the technical field of artificial intelligence, in particular to a charging pile operation state Internet of Things management and control monitoring method based on artificial intelligence. The method comprises the following steps: collecting charging pile current data, labeling the operation state label of the charging pile current data, pre-processing the labeled charging pile current data, separating and fusing the steady-state characteristics and transient characteristics in the charging pile current data by adopting a double-channel adaptive filtering method, performing dynamic dimension reduction and feature optimization on the fused feature vectors by adopting a dynamic principal component analysis method based on time attenuation, training a deep neural network to obtain a trained deep neural network, collecting new charging pile current data, inputting the new charging pile current data into the trained deep neural network after the foregoing operation is performed, and obtaining an operation state label. The existing monitoring method has the problem of low recognition accuracy, and the charging pile operation state Internet of Things management and control monitoring method based on artificial intelligence provided by the application has high recognition accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence, and particularly relates to a charging pile operation state Internet of Things management and control monitoring method based on artificial intelligence. BACKGROUND

[0002] The charging pile, also known as an electric vehicle charging station or an electric vehicle power supply device, is a device for providing electric energy to an electric vehicle. The basic components include: a power module responsible for converting commercial power into a stable power source suitable for charging electric vehicles, a charging module for realizing electric energy storage and transmission, a user interface for providing user operation, a communication module responsible for data interaction with electric vehicles and a background monitoring system, etc.

[0003] In order to effectively prevent and reduce the occurrence of safety accidents and improve the user experience of charging piles, the operation state of all charging piles is monitored. The monitoring method mainly includes the following steps: real-time acquisition of key parameter data such as voltage, current, temperature, etc. of the charging pile through various types of sensors to determine the operation state of the charging pile, for example, using a current sensor to monitor the current size, using a temperature sensor to detect the temperature changes of the environment and the charging pile, and using a voltage sensor to measure the voltage level; through the key state parameter data of the charging pile, the real-time data is then transmitted to the remote monitoring center by various types of sensors through wireless networks, etc. The background monitoring system analyzes and identifies the data to determine whether the operation state of the charging pile is normal, and alarms the user or the operation and maintenance personnel through sound, light or network, etc. when the operation state of the charging pile is abnormal.

[0004] The existing monitoring method mainly has the following problems:

[0005] (1) The existing monitoring method uses sliding average filtering, wavelet filtering, fixed threshold denoising, etc. for signal processing, but the charging pile current data has both strong instantaneous pulse interference and periodic load fluctuation characteristics. Therefore, using the sliding average filtering method will blur the transient characteristics in the current signal, resulting in the loss of key fault pulse information; although the wavelet filtering method can partially preserve the transient characteristics, it will still introduce waveform distortion; the fixed threshold denoising method is also difficult to adapt to the load differences between different charging piles, and cannot effectively process such complex characteristics; resulting in incomplete fault feature extraction, leading to missed or misdiagnosed intelligent diagnosis models, seriously affecting the accuracy and reliability of the monitoring;

[0006] (2) The existing monitoring method uses principal component analysis method to reduce the dimension of data, and the principal component analysis method uses a fixed projection matrix for dimension reduction, but due to the non-stationary covariance structure of current data in the day-night load cycle, the characteristic projection based on the principal component analysis method cannot adapt to the time-varying characteristics of the load change; the data after dimension reduction cannot effectively capture the dynamic characteristics of the load change, resulting in a significant decrease in the recognition ability of the subsequent model to the time-varying abnormal state, affecting the real-time and accuracy of the monitoring;

[0007] (3) The existing monitoring method trains a deep neural network with data and uses the trained deep neural network to identify data. During training, the deep neural network is mainly initialized by a random initialization method, but it often takes a long time and is easy to fall into local optimization, resulting in poor classification performance of the deep neural network and affecting the reliability of the monitoring. During training, the attention mechanism is also used to enable the deep neural network to focus on key features or regions, but the existing attention mechanism fails to fully utilize the spectral correlation between load patterns, resulting in insufficient feature fusion effect and inability to accurately distinguish different charging pile operating states, which significantly reduces the recognition accuracy of the deep neural network to the abnormal state of the charging pile.

[0008] In summary, the existing monitoring method has the problem of low recognition accuracy. SUMMARY

[0009] The technical problem to be solved by the present application is to overcome the shortcomings of the prior art and provide a charging pile operating state Internet of Things management and control monitoring method based on artificial intelligence with high recognition accuracy.

[0010] To solve the above technical problems, the present application provides a charging pile operating state Internet of Things management and control monitoring method based on artificial intelligence, which comprises:

[0011] S1. Using a distributed charging pile data acquisition system constructed by Internet of Things technology to acquire charging pile current data;

[0012] S2. Labeling the operating state label of the charging pile current data to obtain labeled charging pile current data;

[0013] S3. Preprocessing the labeled charging pile current data;

[0014] The preprocessing includes data cleaning and standardization of the labeled charging pile current data;

[0015] S4. Adopting a dual-channel adaptive filtering method to separate and fuse the steady-state features and transient features in the charging pile current data;

[0016] S5. Adopting a dynamic principal component analysis method based on time decay to perform dynamic dimension reduction and feature optimization on the fused feature vectors;

[0017] S6. training the deep neural network to obtain a trained deep neural network;

[0018] S601. defining a structure of the deep neural network;

[0019] S602. initializing weights of the deep neural network based on a spectral clustering guided deep neural network weight initialization method;

[0020] S603. constructing a double-gated feature extraction unit;

[0021] S604. applying second-order differential regularization to hidden states of the deep neural network;

[0022] S605. performing graph attention enhancement at a spectral attention output layer;

[0023] S606. determining a classification result of the current iteration;

[0024] S607. calculating a loss function;

[0025] S608. performing error back propagation based on gradient descent and iterative optimization of model parameters;

[0026] S609. when a preset stopping iteration condition is reached, stopping iteration S603-S608;

[0027] S7. collecting new charging pile current data, performing operations of S3, S4 and S5 on the new charging pile current data, inputting processed new charging pile current data obtained to the trained deep neural network, and outputting, by the trained deep neural network, an operation state label corresponding to the new charging pile current data.

[0028] As a further improvement of the present application: S2 labels the operation state label of the charging pile current data to obtain the labeled charging pile current data, including: labeling the operation state label of the charging pile current data in each time window by a person skilled in the art in combination with operation and maintenance records, knowledge in the field, current waveform features and charging pile operation logs.

[0029] As a further improvement of the present application: S4 separates and fuses the steady-state features and transient features in the charging pile current data by using a double-channel adaptive filtering method, including:

[0030] S401. the main channel uses moving percentile filtering to retain steady-state features, and a median is calculated based on current values in a window to obtain a steady-state current component;

[0031] S402. The auxiliary channel extracts transient features through a gated pulse separator, based on the difference between the original current data and the steady-state current component, combined with a dynamic threshold and processed through a ReLU activation function and a sign function to obtain the transient current component;

[0032] S403. Fuse the steady-state current component and the transient current component to form a fused feature vector at the current moment.

[0033] As a further improvement of the present invention: S5 uses a dynamic principal component analysis method based on time decay to dynamically reduce the dimension and optimize the features of the fused feature vector, including:

[0034] S501. The incremental covariance matrix of the time decay factor is combined, and the covariance matrix of the current moment is updated based on the fused eigenvector and its transposed matrix and the covariance matrix of the previous moment and the time decay factor;

[0035] S502. Perform eigendecomposition on the covariance matrix through the eigenvector function to extract the principal components. Select the fused eigenvectors of the previous several moments according to the eigenvalue size to construct a projection matrix. Use the projection matrix to project the fused eigenvectors into the new feature space to obtain the projection data.

[0036] As a further improvement of the present invention: S602 initializing the deep neural network weights based on the deep neural network weight initialization method guided by spectral clustering includes:

[0037] S6021.Yes Sample set of projection data at time Perform spectral clustering and construct a similarity matrix based on the similarity between samples, where the similarity is calculated by the L2 norm between samples and the width parameter of the similarity;

[0038] S6022. Take the first K feature vectors to form the cluster label C, where K is the number of clusters; C is the cluster label set;

[0039] S6023. An initial weight matrix is ​​calculated based on the cluster centers, regularization parameters, and the identity matrix, and the initial weight matrix is ​​initialized using the cluster centers.

[0040] As a further improvement of the present invention: S603 constructing a dual-gated feature extraction unit includes:

[0041] S6031. Split the update gate into a time gate and a feature gate;

[0042] S6032. The candidate hidden state is calculated by combining the feature gate and the time gate through the hyperbolic tangent function and the weight matrix of the candidate hidden state, and the hidden state at the previous moment and the candidate hidden state are weighted and fused according to the value of the time gate to obtain the hidden state at the current moment.

[0043] As a further improvement of the application: S604, the second-order differential regularization is applied to the hidden state of the deep neural network, including:

[0044] The second-order smoothness penalty is used to force the hidden state to satisfy the second-order derivative approximation to be zero; based on the hidden state at the adjacent moment, the second-order difference of the hidden state is calculated and summed through the Frobenius norm to obtain the smooth regularization term.

[0045] Preferably, the calculation formula of the smooth regularization term is as follows:

[0046] ,

[0047] In the formula, T is the total length of the time sequence; is the Frobenius norm; is the hidden state at the moment t; is the hidden state at the moment t; is the hidden state at the moment t. is the hidden state at the moment t.

[0048] As a further improvement of the application: S605, the graph attention enhancement is performed in the spectral attention output layer, including:

[0049] S6051. The similarity graph is constructed using the cluster centers in the initialization stage, and the elements of the similarity graph are calculated based on the sum of squares of Euclidean norms between the cluster centers and the width parameter of the similarity graph;

[0050] S6052. The feature is enhanced by graph convolution, and the feature representation at the current moment is processed based on the inverse square root of the degree matrix and the similarity matrix to obtain the feature representation after graph convolution;

[0051] S6053. The attention score at each moment is calculated based on the transpose of the trainable attention vector, the attention weight matrix, the feature representation after graph convolution, and the trainable query vector, and the attention weight is obtained by the softmax operation;

[0052] S6054. The final output feature vector for classification is calculated based on the attention weight and the feature representation after graph convolution.

[0053] As a further improvement of the application: the calculation formula of the loss function in S607 is as follows: ​

[0054] ,

[0055] Where c is a positive integer; is the number of categories; is the true label of the c-th category; is the time weight; is the predicted class probability of the cth category; is the fault focusing parameter; Focus parameters; is the smoothing regularization term.

[0056] The beneficial effects of the present invention are as follows: The present invention provides an artificial intelligence-based Internet of Things management and monitoring method for charging pile operation status with high recognition accuracy.

[0057] This monitoring method uses a moving percentile filter combined with a main channel and a gated pulse separator for the auxiliary channel to successfully separate the steady-state and transient features in the current signal, which can better preserve the key transient pulse features in the current and effectively suppress noise. This method also dynamically updates the covariance matrix of the charging pile current data through a combination of an incremental covariance matrix and a time decay factor to adapt to changes in the charging pile load cycle, and can better capture the time-varying characteristics brought about by load changes. At the same time, when training a deep neural network, this monitoring method reduces the convergence time during training and improves the robustness of the network by using the load pattern clustering center to set the initial weights of the deep neural network; the graph attention mechanism is used in the spectral attention output layer, which takes into account the similarity between clustering centers, enhances the representation ability of features, and helps the network better understand the relationship between different load states. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 It is a principle block diagram of the present invention;

[0059] Figure 2 Comparison chart of feature retention effects of different filtering methods;

[0060] Figure 3 This is a comparison chart of dual-channel adaptive filtering performance;

[0061] Figure 4 Comparison chart of transient feature retention capabilities;

[0062] Figure 5 This is a comparison chart of noise suppression capabilities;

[0063] Figure 6 This is a spectrum analysis-comparison chart of the filtering effects of this technology;

[0064] Figure 7 This is an analysis diagram of the adaptability of the time-varying feature projection to the diurnal load pattern;

[0065] Figure 8 For projection effect contrast chart;

[0066] Figure 9 For clustering effect quantitative evaluation;

[0067] Figure 10 For different method charging pile state identification performance comparison chart. DETAILED DESCRIPTION

[0068] The specific embodiments of the application will be further described in detail below with reference to the accompanying drawings.

[0069] (1) MQTT protocol: Message Queuing Telemetry Transport, which means message queuing telemetry transport protocol, is a lightweight communication protocol designed for Internet of Things.

[0070] As shown in Figure 1 The application provides a kind of based on artificial intelligence's charging pile operating state Internet of Things management and control monitoring method, which includes:

[0071] S1. using the distributed charging pile data acquisition system constructed by Internet of Things technology to collect charging pile current data; Ensure that the charging pile current data under fast charging mode, slow charging mode, day and night load cycle, peak period, trough period is collected.

[0072] The distributed charging pile data acquisition system is constructed by Internet of Things technology, and the distributed charging pile data acquisition system includes embedded controller installed in the charging pile, and the embedded controller integrates high-precision current sensor and industrial communication module;

[0073] High-precision current sensor is a collection device for collecting charging pile current data under real-time working state of charging pile, and sampling rate is greater than or equal to 1kHz; Charging pile current data includes three-phase alternating current or direct current original signal;

[0074] Industrial communication module includes 4G, 5G or NB-IoT communication module, and the collection device transmits charging pile current data to cloud platform data center through industrial communication module; In order to ensure low delay and high reliability, data transmission adopts MQTT protocol;

[0075] S2. artificial mark charging pile current data running state label, get labeled charging pile current data;

[0076] Each time window of charging pile current data running state label is marked by combining operation and maintenance record, knowledge in the art, current waveform characteristics and charging pile operation log by those skilled in the art;

[0077] The operation and maintenance record includes a fault alarm record and a maintenance report.

[0078] The running state includes normal charging, overload protection, short circuit failure and standby state.

[0079] The running state label includes "0", "1", "2", "3", "0" corresponds to normal charging, "1" corresponds to overload protection, "2" corresponds to short circuit failure, and "3" corresponds to standby state.

[0080] S3. Preprocess the labeled charging pile current data;

[0081] The preprocessing includes data cleaning and standardization of the labeled charging pile current data.

[0082] The data cleaning includes: eliminating invalid data segments caused by sensor failure or communication interruption and repairing short-term missing points by linear interpolation; invalid data segments are all zero values or excessive range noise, etc.

[0083] The standardization includes: dimension unification processing, scaling the current amplitude to the interval [-1, 1], eliminating the range difference of different charging pile models; for the periodic load fluctuation characteristics of the current signal, a sliding window segmentation strategy is adopted to generate time sequence sample units, the window length can be set to 10 seconds, the step length can be set to 1 second, each sample contains 10000 sampling points, corresponding to a 10kHz sampling rate, and finally a sample-label mapping dataset is constructed.

[0084] S4. Adopt a dual-channel adaptive filtering method to separate and fuse the steady-state characteristics and transient characteristics in the charging pile current data.

[0085] S401. Adopt a dual-channel adaptive filtering method, the main channel adopts a moving percentile filter to retain the steady-state characteristics, and the median is calculated based on the current value in the window to obtain the steady-state current component.

[0086] The calculation formula of the steady-state current component at time t is:

[0087]

[0088] In the formula, represents the median filtering of the current value in the window, which is used to retain the steady-state current characteristics; is the original current data at time t; is the time index, and ; T is the total length of the time series; k is a positive integer; W is the window size. The calculation formula of the window size W is:

[0089] ​​​

[0090] ,

[0091] In the formula, fs is a sampling rate;

[0092] S402. The auxiliary channel extracts the transient feature through the gating pulse separator, and the transient current component is obtained based on the difference between the original current data and the steady-state current component, combined with the dynamic threshold and processed through the ReLU activation function and the sign function. The transient current component represents the burst fluctuation in the current signal;

[0093] The calculation formula of the transient current component at time t is:

[0094] ,

[0095] In the formula, f is the ReLU activation function; I(t) is the original current data at time t; is the dynamic threshold, which represents the threshold calculated based on the current difference at the current time; is the sign function, which outputs 1 when the input is greater than 0, -1 when the input is less than 0, and 0 when the input is equal to 0;

[0096] The calculation formula of the dynamic threshold is:

[0097] ,

[0098] In the formula, is the mean of the difference within the window W; is the standard deviation of the difference within the window W;

[0099] S403. The steady-state current component and the transient current component are fused to form a fused feature vector at the current time. The feature vector combines the steady-state and transient current features. The fusion of steady-state and transient features can comprehensively reflect the operating state of the charging pile, including periodic and burst load change characteristics;

[0100] The calculation formula of the fused feature vector at time t is:

[0101] ,

[0102] In the formula, is the transpose operation of the matrix;

[0103] ​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​The main channel uses the moving percentile filter to reserve the steady-state current component, can better capture the stable part of the current signal, and avoid the problem that the traditional sliding average filter blurs the transient characteristics. The auxiliary channel extracts the transient current component through the gated pulse separator. The dual-channel adaptive filtering method can clearly distinguish the transient fluctuation and the steady-state characteristics in the current signal, effectively separates the steady-state and transient characteristics in the charging pile current data, and can maintain the details of the signal and avoid information loss when facing complex signals with strong transient pulses and periodic load fluctuations.

[0104] Using the normalized current data, the processing effects of the sliding average filter, the wavelet filter and the dual-channel filter of the present technology on the same current signal are compared to verify the effect of the dual-channel filter of the present technology in feature reservation, and the performance differences of each method in retaining transient pulse characteristics and suppressing noise are analyzed. The experimental results are shown in Figure 2 The original signal contains a steady-state waveform and two key transient pulses, i.e., transient pulse 1 and transient pulse 2 in the figure. The conventional filtering method has obvious defects. The sliding average filter can smooth the noise but completely eliminates the transient pulse characteristics. The wavelet filter retains part of the transient characteristics but introduces waveform distortion. The dual-channel filter of the present technology cooperates through dual-channel processing, which not only completely retains the amplitude and shape characteristics of the transient pulse, but also effectively suppresses random noise while maintaining the smooth transition of the steady-state waveform, indicating that the moving percentile main channel accurately extracts the steady-state component, and the gated pulse separation auxiliary channel captures the transient characteristics, and the combination of the two realizes accurate analysis of the complex characteristics of the charging pile current.

[0105] Using another unnormalized current data, the performance advantage of the dual-channel filter of the present technology (upshift 300A) under complex working conditions is verified. The performances of the sliding average filter (upshift 100A), the fixed threshold denoising (upshift 200A) and the dual-channel filter of the present technology (upshift 300A) in retaining transient characteristics and suppressing noise are compared. Based on the real charging pile current signal, it contains steady-state load fluctuations, transient fault pulses and composite noise. The experimental results are shown in Figure 3 The time-domain waveform comparison shows that the present invention perfectly retains the transient pulse characteristics of the short-circuit fault, while the traditional method is either over-smoothed or residual noise.

[0106] From Figure 4 It can be seen that the dual-channel filter of the present technology has higher capture ability for key fault characteristics than other methods. From Figure 5 It can be seen that the signal-to-noise ratio improvement curve confirms that the noise suppression effect of the dual-channel filter of the present technology is the best. From Figure 6 It can be seen that the frequency spectrum analysis diagram proves from the frequency domain that the dual-channel filter of the present technology can not only reserve the fault characteristic frequency band (50-500Hz), but also effectively suppress high-frequency noise.

[0107] In summary, the application solves the contradiction between feature reservation and noise suppression in the traditional method through a dual-channel cooperative mechanism.

[0108] S5. A dynamic principal component analysis method based on time decay is used to perform dynamic dimension reduction and feature optimization on the fused feature vector.

[0109] S501. An incremental covariance matrix based on a time decay factor is combined with the fused feature vector at the current time and its transpose matrix, and the covariance matrix at the previous time and the time decay factor are combined to update the covariance matrix at the current time. The covariance matrix can describe the correlation between features.

[0110] The calculation formula of the covariance matrix at the current time is:

[0111] ,

[0112] In the formula, is a time decay factor that determines the influence weight of past information on the current update; is the covariance matrix at the current time; is the transpose matrix of

[0113] The calculation formula of the time decay factor

[0114] ,

[0115] S502. The covariance matrix is decomposed by a feature vector function to extract principal components, and the first several fused feature vectors at the current time are selected according to the size of the eigenvalue to construct a projection matrix. The fused feature vector is projected into a new feature space using the constructed projection matrix to obtain projection data.

[0116] The feature matrix at the current time is obtained by eigen decomposition of the covariance matrix. , The calculation formula of the feature matrix at the current time

[0117] ,

[0118] In the formula, is a feature vector function used to extract principal components from the covariance matrix, ​​​​​Eigenvalue and eigenvector are obtained by eigen-decomposition of the covariance matrix, and the first several eigenvectors of the current time are selected according to the size of the eigenvalue to construct a projection matrix, for example, the first five eigenvectors are selected;

[0119] For dimension reduction and key feature retention Projection data of the current time The calculation formula is:

[0120] ,

[0121] In the formula, represents The transpose of the sub-matrix composed of the first d columns in the feature matrix of the current time; d is the number of principal components retained, used to control the feature dimension after projection;

[0122] To verify the adaptability of the technology (time-varying projection) to the dynamic change of the charging pile day-night load mode, the distribution characteristics of the conventional method (static PCA) and the technology (time-varying projection) in the feature space are compared through a three-dimensional surface graph, and the PCA is the principal component analysis, and the Figure 7 , Figure 7 The principal component 1 and the principal component 2 form the horizontal coordinate axis, and the feature density is the vertical axis. The natural fluctuation of the charging pile load in a 24-hour period is simulated, and the results show that the feature distribution (purple surface) of the conventional method (static PCA) presents a diffuse state, indicating that the projection reference cannot adapt to the load change, resulting in feature drift, while the feature distribution (green surface) of the technology (time-varying projection) forms a steep and concentrated density peak in the feature space, proving that the technology (time-varying projection) updates the covariance matrix and the eigenvector in real time, so that the projection reference continuously tracks the load change and maintains the consistency of the feature representation, indicating that the technology (time-varying projection) can effectively solve the feature mismatch problem caused by the fixed projection matrix of the traditional method.

[0123] To verify the improvement of the state separation ability of the technology time-varying projection, the separability of the original feature, the conventional PCA and the technology time-varying projection in the feature space is compared, and the projection distribution of the three types of operating states (normal charging, overload protection, short-circuit fault) is visualized, as shown in Figure 8 The original feature has serious state aliasing, the conventional PCA has improved but the boundary is still fuzzy, and the technology time-varying projection makes the three types of states form a significantly separated clustering cluster in the feature space.

[0124] As shown in Figure 9As shown, the cluster evaluation index column chart further proves that the time-varying projection of the present technology leads in three key indicators of profile coefficient (representing the degree of intra-class closeness and inter-class separation), intra-class distance (reflecting the aggregation degree of samples of the same class), and inter-class distance (reflecting the separation degree of different classes), indicating that the incremental covariance matrix design has adaptive ability to the time-varying characteristics of the load.

[0125] S6. Training a deep neural network to obtain a trained deep neural network;

[0126] S601. Defining the structure of the deep neural network;

[0127] The deep neural network for charging pile operating state recognition adopts an end-to-end time sequence feature learning framework, the input layer receives preprocessed current window samples, the core feature extraction module is composed of three cascaded double-gated feature extraction units, the feature extraction layer is connected with a spectral attention output layer, and the global load mode features are fused through a clustering center guided graph attention mechanism;

[0128] The final output layer is a full connection layer, the number of neurons is equal to the number of state categories, and a Softmax activation function is used to generate a category probability distribution.

[0129] All hidden layers use a ReLU activation function, and residual connections are added between double-gated units to speed up gradient propagation.

[0130] S602. Initializing the weights of the deep neural network based on a spectral clustering guided deep neural network weight initialization method;

[0131] The initial weights are generated using load mode clustering centers, including:

[0132] S6021. Performing spectral clustering on the sample set of projection data at time t , constructing a similarity matrix based on the similarity between samples, wherein the similarity is calculated by the L2 norm between samples and a width parameter of the similarity;

[0133] The similarity between the i-th sample in the sample set of projection data at time t and the j-th sample in the sample set of projection data at time t is calculated as follows:

[0134] ,

[0135] wherein is a width parameter of the similarity, and the value can be set to =0.3.​​​​​ is the L2 norm; is the i-th sample in the sample set of the projection data at the time instant is the j-th sample in the sample set of the projection data at the time instant

[0136] S6022. Take the top K eigenvectors to form a clustering label C, K is the number of clusters; C is a set of clustering labels, representing the cluster category to which each sample belongs.

[0137] S6023. Calculate an initial weight matrix based on the clustering centers, a regularization parameter and an identity matrix, the initial weight matrix being initialized using the clustering centers;

[0138] The calculation formula of the initial weight matrix is:

[0139] ,

[0140] In the formula, k is a positive integer; is the k-th clustering center; is the transpose of is a regularization parameter, used to prevent overfitting and singular matrices, and the value can be set to = 0.01; is an identity matrix;

[0141] The calculation formula of the k-th clustering center is:

[0142] ,

[0143] In the formula, is a set of sample indexes in the k-th cluster; is the number of elements in the set of sample indexes in the k-th cluster; i is a positive integer; is the i-th projection data;

[0144] S603. Construct a double-gated feature extraction unit;

[0145] S6031. Split the update gate into a time gate and a feature gate;

[0146] The feature gate is based on the hidden state at the previous time instant and the projection data at the current time instant, and is processed by a Sigmoid activation function and a weight matrix and a bias vector of the feature gate to obtain a value of the feature gate, which is used to control the degree of feature update and focus on transient features;

[0147] ​​​The time gate is based on the differential features and the hidden state of the previous moment. It is processed through the Sigmoid activation function and the weight matrix and bias vector of the time gate to obtain the value of the time gate, which is used to control the degree of time information update and track slow-changing trends.

[0148] S6032. Combine the feature gate and the time gate, calculate the candidate hidden state through the hyperbolic tangent function and the weight matrix of the candidate hidden state, and perform a weighted fusion of the hidden state at the previous moment and the candidate hidden state according to the value of the time gate to obtain the hidden state at the current moment;

[0149] Feature gate is used to control the degree of feature update. The calculation formula is:

[0150] ,

[0151] Where, is the Sigmoid activation function; is the weight matrix of the feature gate; for The hidden state of the moment; for Projection data at the moment; is the bias vector of the feature gate;

[0152] The time gate is used to control the degree of time information update. The calculation formula is:

[0153] ,

[0154] Where, is the weight matrix of the time gate; is the differential feature; is the bias vector of the time gate;

[0155] Differential Features The calculation formula is:

[0156] ,

[0157] Where, for Projection data at the moment;

[0158] Candidate hidden states The calculation formula is:

[0159] ,

[0160] Where, is the hyperbolic tangent function; is the weight matrix of the candidate hidden state; is the Hadamard product;

[0161] Hidden state at the moment The calculation formula is:

[0162] ,

[0163] S604. Applying second-order differential regularization to the hidden state of a deep neural network;

[0164] A second-order smoothness penalty is used to force the hidden state to satisfy the second-order derivative to be approximately zero, which conforms to the smooth transition characteristics of normal current. Based on the hidden state at adjacent moments, the second-order difference of the hidden state is calculated and summed up through the Frobenius norm to obtain a smooth regularization term, which is used to constrain the hidden state.

[0165] The smoothing regularization term is used to control the smoothness of the hidden state. The calculation formula is:

[0166] ,

[0167] Where T is the total length of the time series; is the Frobenius norm; for The hidden state of the moment; for The hidden state of the moment; for The hidden state of the moment;

[0168] The smoothing regularization term calculates the second-order difference of the hidden state and uses the second-order smoothness penalty to constrain the change of the hidden state, so that the normal current data of the charging pile can transition smoothly, forcing the state change of the network output to meet the smooth transition characteristics of the current.

[0169] S605. Perform graph attention enhancement on the spectral attention output layer, including:

[0170] S6051. Construct a similarity graph using the cluster centers in the initialization phase, and calculate the elements of the similarity graph based on the Euclidean norm square between the cluster centers and the width parameter of the similarity graph;

[0171] Elements of similar graphs The calculation formula is:

[0172] ,

[0173] Where, is the square of the Euclidean norm; is the i-th cluster center; is the jth cluster center; is the width parameter of the similarity graph;

[0174] S6052. Enhance the features through graph convolution, process the feature representation at the current moment based on the inverse square root of the degree matrix and the similarity matrix, and obtain the feature representation after graph convolution;

[0175] Feature representation after graph convolution The calculation formula is:

[0176] ,

[0177] Where, is the inverse square root of the degree matrix; A is the similarity matrix, which represents the similarity between the cluster centers, and its elements are the elements of the similarity graph .

[0178] By combining graph convolution with spectral attention, the network is able to process the spectral correlation between charging pile load patterns while considering the local and global similarities of features, improving classification accuracy and helping the model better understand the relationship between different load states.

[0179] S6053. Calculate the attention score at each moment based on the transpose of the trainable attention vector, the attention weight matrix, the feature representation after graph convolution, and the trainable query vector, and obtain the attention weight through a softmax operation;

[0180] No. Moment attention score The calculation formula is:

[0181] ,

[0182] Where, is the transpose of the trainable attention vector; is the attention weight matrix; is the attention weight matrix; s is the trainable query vector;

[0183] Attention weight The calculation formula is:

[0184] ,

[0185] Where, is the attention score at the tth moment; is the attention score at the jth moment, The sum of the index representing the attention scores at all moments;

[0186] S6054. Calculate the final output feature vector for classification based on the attention weight and the feature representation after graph convolution;

[0187] The final output feature vector The calculation formula is:

[0188] ,

[0189] S606. Determine the classification result of this iteration;

[0190] Map the final feature vector generated by the spectral attention output layer to the classification category. First, map the dimension of to a 4-dimensional space through a fully connected layer, corresponding to 4 states, and then normalize the prediction probability distribution of each category through the Softmax function. Take the category index corresponding to the maximum probability as the predicted running state label of the current sample.

[0191] If the output vector is [0.02, 0.85, 0.08, 0.05], it is determined to be an overload protection state, corresponding to the label "1";

[0192] S607. Calculate the loss function;

[0193] Based on the number of categories, the real label and the predicted class probability of the category, combined with the time weight and the fault focus parameter, the focal parameter, the cross-entropy loss is weighted and adjusted, wherein the time weight is calculated according to the category average duration and the total time, so as to handle the data imbalance problem;

[0194] The calculation formula of the loss function is:

[0195] ,

[0196] In the formula, c is a positive integer; is the number of categories; is the real label of the cth category; is the time weight; is the predicted class probability of the cth category; is the fault focus parameter, which can be set to =2; is the focal parameter, which can be set to =2; is the logarithmic function, which is by default with base 10; is the smooth regularization term;

[0197] The calculation formula of the time weight is:

[0198] ,

[0199] wherein, is the average duration of the c-th class, representing the average event duration of the class, used in the time-weighted loss to handle data imbalance;

[0200] S608. Gradient descent-based error backpropagation and iterative optimization of model parameters are performed;

[0201] The loss function is calculated The gradients of the weights of each layer of the network are backpropagated layer by layer from the output layer to the input layer through the chain rule, and the loss function incorporates the gradient contribution of the second-order differential regularization term applied to the hidden state of the deep neural network;

[0202] The Adam optimizer is used to adaptively adjust the learning rate, and the weight matrix and bias vector are updated according to the gradient direction;

[0203] The update process includes gradient clipping to prevent gradient explosion, and the gradient clipping threshold is ±1.0;

[0204] S609. When the preset stopping iteration condition is reached, the iteration S603~S608 is stopped, and the model parameters with the optimal performance on the validation set are saved after training is terminated;

[0205] The training is stopped when any of the following conditions is met:

[0206] 1) The accuracy of the validation set has not improved for 10 consecutive iteration steps;

[0207] 2) The total number of training epochs reaches the preset upper limit, which is 10000 by default;

[0208] 3) The loss function value decreases at a rate lower than 0.001, and the window size is 5 iteration steps.

[0209] The comprehensive performance of different methods in the charging pile operating state recognition task is evaluated, and the accuracy and F1 score of SVM (Support Vector Machine), decision tree, LSTM (Long Short-Term Memory Network), GRU (Gated Recurrent Unit), CNN (Convolutional Neural Network method) and the present technology are compared intuitively through bar charts, as shown in Figure 10 The present technology leads other comparison methods in two core indicators, which reflects that the double-gated feature extraction unit effectively separates the steady-state drift and transient features, the weight initialization guided by spectral clustering accelerates model convergence, the second-order differential regularization applied to the hidden state of the deep neural network enhances the sensitivity to abnormal mutations, and the spectral attention mechanism optimizes the fusion of load pattern features, solving the problems of difficulty in capturing transient events in current data and scarcity of fault samples.

[0210] S7. Collect new charging pile current data, and input the processed new charging pile current data obtained after performing the operations of S3, S4, and S5 on the new charging pile current data into the trained deep neural network. The trained deep neural network outputs the running state label corresponding to the new charging pile current data. Different running state labels trigger different responses. "Overload protection" and "short circuit fault" trigger corresponding level alarm signals.

Claims

1. An artificial intelligence-based IoT management and monitoring method for charging pile operation status, characterized in that: include: S1. Collect charging pile current data using a distributed charging pile data acquisition system built using IoT technology; S2. Mark the operating status tag of the charging pile current data to obtain the marked charging pile current data; S3. Preprocess the labeled charging pile current data; Preprocessing includes data cleaning and standardization of the labeled charging pile current data; S4. Use a dual-channel adaptive filtering method to separate and fuse the steady-state and transient characteristics of the charging pile current data; S5. Use the dynamic principal component analysis method based on time decay to dynamically reduce the dimension and optimize the features of the fused feature vector; S6. training the deep neural network to obtain a trained deep neural network; S601. Define the structure of a deep neural network; S602. Initialize the deep neural network weights based on the deep neural network weight initialization method guided by spectral clustering; S603. Construct a dual-gated feature extraction unit; S604. Applying second-order differential regularization to the hidden state of a deep neural network; S605. Perform graph attention enhancement in the spectral attention output layer; S606. Determine the classification result of this iteration; S607. Calculate the loss function; S608. Perform error back propagation based on gradient descent and iterative optimization of model parameters; S609. When the preset stop iteration condition is reached, stop iteration S603~S608; S7. Collect new charging pile current data, perform S3, S4, and S5 on the new charging pile current data, and input the processed new charging pile current data into the trained deep neural network. The trained deep neural network outputs the operating status label corresponding to the new charging pile current data.

2. The method for controlling and monitoring the operation status of a charging pile based on artificial intelligence according to claim 1, characterized in that: The S2 labels the operating status label of the charging pile current data, and the labeled charging pile current data includes: the operating status label of the charging pile current data of each time window is labeled by a technician in this field based on operation and maintenance records, knowledge in this field, current waveform characteristics, and charging pile operation logs.

3. The method for controlling and monitoring the operation status of a charging pile based on artificial intelligence according to claim 1, characterized in that: The S4 uses a dual-channel adaptive filtering method to separate and fuse the steady-state characteristics and transient characteristics in the charging pile current data, including: S401. The main channel uses a moving percentile filter to retain steady-state characteristics and calculates the median based on the current value in the window to obtain the steady-state current component; S402. The auxiliary channel extracts transient features through a gated pulse separator, based on the difference between the original current data and the steady-state current component, combined with a dynamic threshold and processed through a ReLU activation function and a sign function to obtain the transient current component; S403. Fuse the steady-state current component and the transient current component to form a fused feature vector at the current moment.

4. The method for controlling and monitoring the operation status of a charging pile based on artificial intelligence according to claim 1, characterized in that: The S5 uses a dynamic principal component analysis method based on time decay to dynamically reduce the dimension and optimize the features of the fused feature vectors, including: S501. The incremental covariance matrix of the time decay factor is combined, and the covariance matrix of the current moment is updated based on the fused eigenvector and its transposed matrix and the covariance matrix of the previous moment and the time decay factor; S502. Perform eigendecomposition on the covariance matrix through the eigenvector function to extract the principal components. Select the fused eigenvectors of the previous several moments according to the eigenvalue size to construct a projection matrix. Use the projection matrix to project the fused eigenvectors into the new feature space to obtain the projection data.

5. The method for controlling and monitoring the operation status of a charging pile based on artificial intelligence according to claim 1, characterized in that: The S602 method of initializing the deep neural network weights based on the spectral clustering guided deep neural network weight initialization includes: S6021.Yes Sample set of projection data at time Perform spectral clustering and construct a similarity matrix based on the similarity between samples, where the similarity is calculated by the L2 norm between samples and the width parameter of the similarity; S6022. Take the first K feature vectors to form the cluster label C, where K is the number of clusters; C is the cluster label set; S6023. An initial weight matrix is ​​calculated based on the cluster centers, regularization parameters, and the identity matrix, and the initial weight matrix is ​​initialized using the cluster centers.

6. The method for controlling and monitoring the operation status of a charging pile based on artificial intelligence according to claim 1, characterized in that: The step S603 of constructing a dual-gated feature extraction unit includes: S6031. Split the update gate into a time gate and a feature gate; S6032. Combine the feature gate and the time gate, and calculate the candidate hidden state through the hyperbolic tangent function and the weight matrix of the candidate hidden state. According to the value of the time gate, the hidden state at the previous moment and the candidate hidden state are weighted and fused to obtain the hidden state at the current moment.

7. The method for controlling and monitoring the operation status of a charging pile based on artificial intelligence according to claim 1, characterized in that: The step S604 of applying second-order differential regularization to the hidden state of the deep neural network includes: A second-order smoothness penalty is used to force the hidden state to satisfy the second-order derivative to be approximately zero; based on the hidden state at adjacent moments, the second-order difference of the hidden state is calculated and summed up through the Frobenius norm to obtain the smooth regularization term.

8. The method for controlling and monitoring the operation status of a charging pile based on artificial intelligence according to claim 7, characterized in that: The smoothing regularization term The calculation formula is: ， Where T is the total length of the time series; is the Frobenius norm; for The hidden state of the moment; is the hidden state at time t; for The hidden state of the moment.

9. The method for controlling and monitoring the operation status of a charging pile based on artificial intelligence according to claim 1, characterized in that: The step S605 of performing graph attention enhancement in the spectral attention output layer includes: S6051. Construct a similarity graph using the cluster centers in the initialization phase, and calculate the elements of the similarity graph based on the Euclidean norm square between the cluster centers and the width parameter of the similarity graph; S6052. Enhance the features through graph convolution, process the feature representation at the current moment based on the inverse square root of the degree matrix and the similarity matrix, and obtain the feature representation after graph convolution; S6053. Calculate the attention score at each moment based on the transpose of the trainable attention vector, the attention weight matrix, the feature representation after graph convolution, and the trainable query vector, and obtain the attention weight through a softmax operation; S6054. The final output feature vector for classification is calculated based on the attention weights and the feature representation after graph convolution.

10. The method for controlling and monitoring the operation status of a charging pile based on artificial intelligence according to claim 1, characterized in that: The loss function in S607 The calculation formula is: ， Where c is a positive integer; is the number of categories; is the true label of the c-th category; is the time weight; is the predicted class probability of the cth category; is the fault focusing parameter; Focus parameters; is the smoothing regularization term.

Images