An intelligent detection method for tram charging equipment

Abstract

The present invention discloses an intelligent detection method for tram charging equipment, which relates to the technical field of power equipment fault detection. By obtaining the real-time operation data of the target charging equipment and intercepting the data sequence of a preset duration, a real-time operation signal is obtained; a variety of signals are filtered and aligned to obtain a first feature matrix; the first feature matrix is normalized to obtain a second feature matrix; the second feature matrix is used as the input of the fault detection model to obtain the fault detection result. By introducing a bidirectional gated recurrent unit convolutional neural network, the expression ability of the model for complex data patterns is enhanced, the dependence on manual feature engineering is reduced, and the detection effect is optimized. Moreover, the end-to-end training mode of the model and the comprehensive analysis of multiple signals improve the adaptability and discrimination ability for different fault modes, effectively solving the problem of insufficient accuracy of traditional methods.

Description

Technical Field

[0001] The present invention relates to the technical field of power equipment fault detection, and particularly relates to an intelligent detection method for tram charging equipment. Background Art

[0002] Electric vehicles have become the trend of future automotive industry development. The development of electric vehicles is inseparable from charging infrastructure. Currently, charging infrastructure mainly includes DC charging piles, AC charging piles, on-board chargers, charging stations, etc. Among them, DC charging piles are equipped with charging modules, and the charging power can reach dozens to hundreds of kilowatts. The charging rate is fast and can meet the requirements of high-power fast charging. However, the working voltage of the charging module is high, including a large number of power electronic devices and electrolytic capacitors, which are prone to aging failure and device faults under high voltage and repeated temperature cycles, becoming the component with the highest failure rate in DC charging piles. It is of great significance to monitor the possible faults of the charging module of the charging pile.

[0003] Currently, the fault detection of charging modules mainly classifies through traditional machine learning models (such as support vector machines, random forests, etc.). However, these methods need to rely on manual feature engineering, cannot fully mine the deep patterns of data, and have poor adaptability to data distribution changes, resulting in low accuracy of detection results. Summary of the Invention

[0004] The purpose of the present invention is to solve the problem of low accuracy mentioned in the above background art, and propose an intelligent detection method for tram charging equipment.

[0005] In the first aspect of the implementation of the present invention, an intelligent detection method for tram charging equipment is provided. The method includes:

[0006] Obtain the real-time operation data of the target charging equipment, and intercept the data sequence of the preset duration to obtain the real-time operation signal; the real-time operation signal includes a variety of voltage signals and current signals;

[0007] Filter and align a variety of signals to obtain the first feature matrix;

[0008] Perform normalization processing on the first feature matrix to obtain the second feature matrix;

[0009] Use the second feature matrix as the input of the fault detection model to obtain the fault detection result; the fault detection model is composed of a bidirectional gated recurrent unit and a convolutional neural network.

[0010] Optionally, the charging module of the target charging equipment includes a front-stage circuit and a rear-stage circuit; the front-stage circuit adopts a three-phase VIENNA rectifier structure; the rear-stage circuit adopts a full-bridge LLC resonant converter structure.

[0011] Optionally, the post-stage circuit includes a full-bridge inverter circuit, a resonant cavity circuit, a transformer module, and a rectifier and filter circuit;

[0012] The operating signals include the current signals of the three-phase input in the pre-stage circuit, the voltage signal of the full-bridge inverter circuit, the current signal of the resonant cavity circuit, and the voltage signal of the rectifier and filter circuit.

[0013] Optionally, the fault detection model includes a data fusion module, a feature extraction and enhancement module, a feature fusion module, and a classification and prediction module; where:

[0014] The data fusion module is used to perform feature fusion on the input data through a bidirectional gated recurrent unit to obtain a hidden state vector;

[0015] The feature extraction and enhancement module is used to perform multi-scale feature extraction on the hidden state vector through multiple convolutional branches and perform attention weighting to obtain multiple feature tensors;

[0016] The feature fusion module is used to perform global average pooling and concatenation on multiple feature tensors to obtain a target feature vector;

[0017] The classification and prediction module is used to classify the target feature vector through a fully connected layer to obtain a fault detection result.

[0018] Optionally, the feature extraction and enhancement module includes a multi-scale differentiation layer and multiple convolutional branches, namely the first convolutional branch, the second convolutional branch, and the third convolutional branch;

[0019] The multi-scale differentiation layer is used to perform max-pooling and average-pooling on the hidden state vector to obtain a first vector and a second vector; the first convolutional branch receives the first vector as input; the second convolutional branch receives the second vector as input; the third convolutional branch receives the hidden state vector as input;

[0020] The target convolutional branch is used to perform downsampling processing on the input of this branch through one-dimensional convolution and max-pooling to obtain a first feature tensor; and perform feature enhancement on the first feature tensor through an attention module to obtain a second feature tensor; the target convolutional branch is any one of the multiple convolutional branches.

[0021] Optionally, the downsampling structures of each convolutional branch are the same; the attention module calculates attention weights through two fully connected layers;

[0022] The calculation process of the target convolutional branch is: ; where F2 is the output of the target convolutional branch; X is the input of the target convolutional branch; F1 is the first feature tensor; f is an operation operator, with subscripts Conv and MP representing one-dimensional convolution and max pooling respectively, and the superscripts representing the convolutional kernel size or pooling window size, and the stride is 2 for both; W is the channel attention weight; GAP represents global average pooling; S1 and S2 are the weight parameters of two fully connected layers; ReLu and sigmoid are activation functions.

[0023] Optionally, the operation process of the feature fusion module includes:

[0024] Performing global average pooling on each channel of the target second feature tensor to obtain a temporary vector; the target second feature tensor is any one of multiple second feature tensors;

[0025] Concatenating multiple temporary vectors to obtain a target feature vector.

[0026] Optionally, the classification and prediction module uses a multi-layer perceptron and the Softmax function for classification and prediction.

[0027] Advantages of the present invention:

[0028] The present invention proposes an intelligent detection method for electric vehicle charging equipment, which includes: obtaining real-time operation data of a target charging equipment, and intercepting a data sequence of a preset duration to obtain a real-time operation signal; the real-time operation signal includes multiple voltage signals and current signals; filtering and aligning the multiple signals to obtain a first feature matrix; performing normalization processing on the first feature matrix to obtain a second feature matrix; using the second feature matrix as the input of a fault detection model to obtain a fault detection result; the fault detection model is composed of a bidirectional gated recurrent unit and a convolutional neural network.

[0029] By introducing a bidirectional gated recurrent unit and a convolutional neural network, the expression ability of the model for complex data patterns is enhanced, the dependence on manual feature engineering is reduced, and the detection effect is optimized. Moreover, the end-to-end training method of the model and the comprehensive analysis of multiple signals improve the adaptability and discrimination ability for different fault patterns, effectively solving the problem of insufficient accuracy of traditional methods. Description of the Drawings

[0030] The present invention will be further described below with reference to the accompanying drawings.

[0031] Figure 1 It is a flowchart of an intelligent detection method for electric vehicle charging equipment provided by an embodiment of the present invention;

[0032] Figure 2 It is a circuit topology diagram of a charging module provided by an embodiment of the present invention;

[0033] Figure 3 The embodiment of the present invention provides a structural schematic diagram of a fault detection model. Specific implementation manners

[0034] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described with reference to the accompanying drawings in the embodiments of the present invention. Apparently, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0035] The embodiment of the present invention provides an intelligent detection method for tram charging equipment. Refer to Figure 1 , Figure 1 which is a flowchart of an intelligent detection method for tram charging equipment provided by the embodiment of the present invention. The method includes the following steps:

[0036] S101, obtain the real-time operation data of the target charging equipment, and intercept the data sequence of a preset duration to obtain a real-time operation signal.

[0037] S102, filter and align multiple signals to obtain a first feature matrix.

[0038] S103, perform normalization processing on the first feature matrix to obtain a second feature matrix.

[0039] S104, use the second feature matrix as the input of the fault detection model to obtain a fault detection result.

[0040] Among them, the real-time operation signal includes multiple voltage signals and current signals; the fault detection model is composed of a bidirectional gated recurrent unit and a convolutional neural network; the fault detection result includes multiple open-circuit faults.

[0041] Based on the intelligent detection method for tram charging equipment provided by the embodiment of the present invention, by introducing a bidirectional gated recurrent unit and a convolutional neural network, the expression ability of the model for complex data patterns is enhanced, and at the same time, the dependence on manual feature engineering is reduced, and the detection effect is optimized. Moreover, the end-to-end training method of the model and the comprehensive analysis of multiple signals improve the adaptability and discrimination ability for different fault patterns, effectively solving the problem of insufficient accuracy of traditional methods.

[0042] In one implementation manner, by obtaining the real-time operation data of the tram charging equipment and processing it, potential fault problems can be discovered in time, maintenance can be carried out in advance, and safety hazards caused by faults can be avoided.

[0043] In one implementation, a low-pass filter can be used to denoise the running signal. Different signals are aligned according to timestamps through linear interpolation or spline interpolation, so that different signals have corresponding values at the same time point. By filtering and aligning various voltage signals and current signals, noise can be removed, data redundancy can be reduced, the quality of the signals can be ensured, and the input data of the fault detection model can be made more accurate and efficient.

[0044] In one implementation, the maximum and minimum values of the training data set of the fault detection model are used to normalize the running signal data.

[0045] In one implementation, this method can be applied to different types of tram charging devices and different usage scenarios, and has strong versatility and scalability. See Figure 2 , Figure 2 FIG. is a circuit topology diagram of a charging module provided by an embodiment of the present invention. The charging module of the target charging device includes a front-stage circuit and a rear-stage circuit; the front-stage circuit adopts a three-phase VIENNA rectifier structure; the rear-stage circuit adopts a full-bridge LLC resonant converter structure. The rear-stage circuit includes a full-bridge inverter circuit, a resonant cavity circuit, a transformer module, and a rectifier and filter circuit.

[0046] Among the faults of the charging module of the charging pile, the proportion of open-circuit faults and short-circuit faults of power devices is relatively high. The short-circuit fault time of power devices is short and the harm is great. Usually, it is detected and protected through the design of the hardware circuit of the drive module, or the short-circuit fault is converted into an open-circuit fault through a fast fuse. After the open-circuit fault of the power device occurs, the system can still run, but the voltage and current are distorted, accelerating the aging and failure of non-fault power devices and endangering the safe and reliable operation of the system. Therefore, the diagnosis and identification of the open-circuit fault of the power device of the charging module is of great significance for improving the operation reliability of the module. Multiple open-circuit faults can be used as classification labels, and normal operation can be used as a special fault label to label the training data set. The training process of the fault detection model is as follows:

[0047] Step 1, construct a simulation model of the charging module of the target charging device.

[0048] Step 2, according to the charging module simulation model, within a preset parameter range, perform multiple fault simulations, collect the running signals when different faults occur, and generate labeled data in combination with the injected fault types.

[0049] Step 3, normalize the feature values of the labeled data according to the maximum and minimum values to establish a data set.

[0050] Step 4, according to the data set, with the goal of maximizing the accuracy, train a preset neural network model to obtain a fault detection model.

[0051] In one embodiment, for a charging module as shown in Figure 2 , the operating signals include the current signals (Ia, Ib, and Ic) of the three-phase input in the front-stage circuit, the voltage signal (voltage U1 between endpoints p1 and p2) of the full-bridge inverter circuit, the current signal (Ir) of the resonant cavity circuit, and the voltage signal (voltage U2 between endpoints p3 and p4) of the rectifier filter circuit. These signals reflect the dynamic changes of current and voltage in the charging device, which helps to deeply analyze the problems in the power conversion process and quickly identify potential faults.

[0052] In one embodiment, referring to Figure 3 , Figure 3 is a schematic structural diagram of a fault detection model provided by an embodiment of the present invention.

[0053] The fault detection model includes a data fusion module, a feature extraction and enhancement module, a feature fusion module, and a classification and prediction module; where:

[0054] The data fusion module is used to perform feature fusion on the input data through a bidirectional gated recurrent unit (Bi-GRU) to obtain a hidden state vector.

[0055] The feature extraction and enhancement module is used to perform multi-scale feature extraction on the hidden state vector through multiple convolutional branches and perform attention weighting to obtain multiple feature tensors.

[0056] The feature fusion module is used to perform global average pooling and concatenation on the multiple feature tensors to obtain a target feature vector.

[0057] The classification and prediction module is used to classify the target feature vector through a fully connected layer to obtain a fault detection result.

[0058] In one implementation, during the operation of the tram charging device, its faults are often reflected by complex time series signals, and the front and back states of these signals may be crucial for fault diagnosis. The bidirectional gated recurrent unit can capture the dependency relationships at different time points in the sequence by processing data in both forward and backward directions simultaneously. Specifically, forward propagation can analyze the timing law of the data from the start to the end of the sequence, while backward propagation can reverse-mine potential fault information from the end to the start. This capture of bidirectional time dependencies improves the model's ability to understand complex signal patterns, enabling the model to comprehensively consider the changes of signals in different time dimensions and thus make more accurate fault predictions.

[0059] In one implementation, the hidden state vector output at the last time step is taken as the final output.

[0060] In one implementation, multi-scale feature extraction is performed on the hidden state vector through multiple convolutional branches, which can capture the key features of the signal from different scales and angles. This multi-dimensional feature extraction can effectively improve the model's ability to identify complex fault patterns. In addition, attention weighting can automatically adjust the weights according to the importance of different features, enabling the model to focus on key features and further improving the accuracy of fault diagnosis.

[0061] In one embodiment, the feature extraction and enhancement module includes a multi-scale differentiation layer and multiple convolutional branches, namely the first convolutional branch, the second convolutional branch, and the third convolutional branch.

[0062] The multi-scale differentiation layer is used to perform max pooling and average pooling on the hidden state vector to obtain a first vector and a second vector.

[0063] The target convolutional branch is used to perform downsampling on the input of this branch through one-dimensional convolution and max pooling to obtain a first feature tensor; and perform feature enhancement on the first feature tensor through an attention module to obtain a second feature tensor.

[0064] Among them, the first convolutional branch receives the first vector as input; the second convolutional branch receives the second vector as input; the third convolutional branch receives the hidden state vector as input; the target convolutional branch is any one of the multiple convolutional branches.

[0065] In one implementation, the pooling window sizes of max pooling and average pooling in the multi-scale differentiation layer can both be 5×1, and the stride is 2.

[0066] In one implementation, the downsampling structures of each convolutional branch are the same; ×n in the convolutional branch represents n consecutive convolutional pooling structures Conv+MaxPool (Conv represents one-dimensional convolution, and MaxPool represents max pooling). Among them, n can be equal to 3, and the kernel sizes of the convolutional kernels in each convolutional branch are 5×1, 3×1, and 3×1 in sequence; the pooling window sizes are all 2×1. The attention module (Attention) can adopt channel attention and calculate the attention weights through two fully connected layers.

[0067] The calculation process of the target convolutional branch is as follows: ; where F2 is the output of the target convolutional branch; X is the input of the target convolutional branch; F1 is the first feature tensor; f is an operation operator, and the subscripts Conv and MP represent one-dimensional convolution and max pooling respectively, and the superscripts represent the kernel size or pooling window size, and the stride is 2; W is the channel attention weight; GAP represents global average pooling; S1 and S2 are the weight parameters of the two fully connected layers respectively; ReLu and sigmoid are activation functions; the operation symbol represents element-wise multiplication; the operation symbol Denotes channel weighted multiplication. In one embodiment, the operation process of the feature fusion module includes:

[0068] Perform global average pooling on each channel of the target second feature tensor to obtain a temporary vector.

[0069] Concatenate multiple temporary vectors to obtain a target feature vector.

[0070] Wherein, the target second feature tensor is any one of the multiple second feature tensors obtained by multiple convolutional branches.

[0071] In one implementation, global average pooling can not only extract global features, but also effectively reduce the dimension of data, lower the computational complexity, and relieve the burden of subsequent processing. Concatenating multiple temporary vectors can integrate the feature information extracted by different convolutional branches. This concatenation method fuses the information of multiple feature channels together to form a unified target feature vector, enhancing the feature representation ability and enabling the model to consider signal features at multiple levels simultaneously.

[0072] In one embodiment, the classification and prediction module uses a multi-layer perceptron (MLP) and a Softmax function for classification and prediction.

[0073] In one implementation, the hidden layer of the MLP can be two layers, and the activation function uses ReLu. As a classic feedforward neural network, the multi-layer perceptron (MLP) processes information through multiple layers of neuron nodes and can effectively capture the non-linear relationship between input features and the output. In the fault detection of tram charging equipment, the change of fault features is usually non-linear. The MLP can effectively map the relationship between complex input features and fault types through its deep structure, thereby improving the accuracy of fault classification.

[0074] The above has described an embodiment of the present invention in detail, but the content described is only a preferred embodiment of the present invention and cannot be considered as limiting the scope of implementation of the present invention. All equivalent changes and improvements made according to the scope of the application of the present invention should still fall within the scope covered by the patent of the present invention.

Claims

1. An intelligent detection method for electric vehicle charging equipment, characterized in that: The method comprises: Acquire real-time operation data of the target charging device, and intercept a data sequence of a preset time length to obtain a real-time operation signal; the real-time operation signal includes a plurality of voltage signals and current signals; Filtering and aligning the multiple signals to obtain a first feature matrix; Normalizing the first characteristic matrix to obtain a second characteristic matrix; The second feature matrix is ​​used as an input of a pre-trained fault detection model to obtain a fault detection result; the fault detection model includes a data fusion module, a feature extraction and enhancement module, a feature fusion module and a classification prediction module; wherein: The data fusion module is used to perform feature fusion on the input data through a bidirectional gated recurrent unit to obtain a hidden state vector; The feature extraction and enhancement module is used to perform multi-scale feature extraction on the hidden state vector through multiple convolution branches, and perform attention weighting to obtain multiple feature tensors; specifically, the feature extraction and enhancement module includes a multi-scale differentiation layer and multiple convolution branches, which are respectively a first convolution branch, a second convolution branch and a third convolution branch; The multi-scale differentiation layer is used to perform maximum pooling and average pooling on the hidden state vector to obtain a first vector and a second vector; the first convolution branch receives the first vector as input; the second convolution branch receives the second vector as input; the third convolution branch receives the hidden state vector as input; The target convolution branch is used to downsample the input of the branch through one-dimensional convolution and maximum pooling to obtain a first feature tensor; and to enhance the first feature tensor through an attention module to obtain a second feature tensor; the target convolution branch is any one of the multiple convolution branches; the downsampling structure of each convolution branch is consistent; the attention module calculates the attention weight through two fully connected layers; the calculation process of the target convolution branch is: ； Among them, F2 is the output of the target convolution branch; X is the input of the target convolution branch; F1 is the first feature tensor; f is the operation operator, the subscripts Conv and MP represent one-dimensional convolution and maximum pooling respectively, the superscript represents the convolution kernel size or the pooling window size, and the stride is 2; W is the channel attention weight; GAP represents global average pooling; S1 and S2 are the weight parameters of the two fully connected layers respectively; ReLu and sigmoid are activation functions; The feature fusion module is used to perform global average pooling and concatenation on multiple second feature tensors to obtain a target feature vector; The classification prediction module is used to classify the target feature vector through a fully connected layer to obtain a fault detection result.

2. The intelligent detection method for electric vehicle charging equipment according to claim 1 is characterized in that: The charging module of the target charging device includes a front-stage circuit and a rear-stage circuit; the front-stage circuit adopts a three-phase VIENNA rectifier structure; and the rear-stage circuit adopts a full-bridge LLC resonant converter structure.

3. The intelligent detection method for electric vehicle charging equipment according to claim 2 is characterized in that: The post-stage circuit includes a full-bridge inverter circuit, a resonant cavity circuit, a transformer module and a rectifier and filter circuit; The operation signal includes a current signal of a three-phase input in the front-stage circuit, a voltage signal of the full-bridge inverter circuit, a current signal of the resonant cavity circuit, and a voltage signal of the rectifier and filter circuit.

4. The intelligent detection method for electric vehicle charging equipment according to claim 1, characterized in that: The operation process of the feature fusion module includes: Performing global average pooling on each channel of a target second feature tensor to obtain a temporary vector; the target second feature tensor is any one of the multiple second feature tensors; Multiple temporary vectors are concatenated and spliced ​​to obtain the target feature vector.

5. The intelligent detection method for electric vehicle charging equipment according to claim 1, characterized in that: The classification prediction module uses a multi-layer perceptron and a Softmax function to perform classification prediction.

Images