Electric energy quality disturbance classification method for tunneling roadway power supply system

Abstract

The invention belongs to the field of electric energy quality analysis of an electric power system, and provides an electric energy quality disturbance classification method for a tunneling roadway power supply system in order to solve the problem that various electric energy quality disturbances cannot be accurately identified and classified in a tunneling roadway at present. According to the method, the time domain information and the frequency domain information of the electric energy quality disturbance signals are fused, the model has high noise immunity, high training speed and good classification accuracy, the electric energy quality disturbance signal data collected in real time are classified, and therefore the electric energy quality disturbance of the coal mine driving working face power supply system is rapidly and accurately classified.

Description

Technical Field

[0001] The invention belongs to the field of power system power quality analysis, and in particular relates to a method for classifying power quality disturbances in a tunneling tunnel power supply system. Background Art

[0002] With the continuous advancement of intelligent coal mines, more and more high-power coal mining equipment is being used in underground coal mines. Taking tunneling equipment as an example, the electrical power of the whole set of equipment can reach the megawatt level. These equipment have the characteristics of frequent start and stop, which makes the coal mine power supply system often have power quality problems such as large voltage fluctuations, large harmonic distortion, voltage sag, and voltage interruption. These power quality problems affect the performance and function of tunneling equipment, causing overheating, abnormal vibration, motor stalling, etc. of cutting motors and climbing claw motors, causing motor damage to affect the continuity of tunneling work; at the same time, it will also threaten the stability and reliability of the coal mine power supply system. Monitoring and analyzing power quality is a prerequisite for discovering power quality problems and conducting governance and improvement. Among them, accurate and timely identification and classification of various power quality disturbances is one of the key issues.

[0003] The noise pollution of the underground power supply system in coal mines is quite serious, and the power quality disturbance is mainly a composite disturbance superimposed by steady-state and transient disturbances. Therefore, it is necessary to study a classification method that can combine the advantages of different models to improve the accuracy of identifying composite disturbances. Summary of the invention

[0004] In order to solve at least one of the above technical problems existing in the prior art, the present invention provides a method for classifying power quality disturbances in a tunneling tunnel power supply system.

[0005] The present invention is implemented by the following technical solution: a method for classifying power quality disturbances in a tunneling power supply system, comprising:

[0006] Acquire power quality disturbance signal data;

[0007] Constructing a power quality disturbance classification model with time-frequency feature fusion, and training the model based on power quality disturbance signal data; the power quality disturbance classification model with time-frequency feature fusion includes at least a convolution layer, a BiGRU layer, a cross attention layer, a pooling layer, a fully connected layer, and a softmax layer;

[0008] The power quality disturbance signal data of the tunneling tunnel power supply system is collected and input into the time-frequency feature fusion power quality disturbance classification model; first, feature extraction is performed in the convolution layer to obtain primary time domain features and primary frequency domain features, the BiGRU layer extracts corresponding time series features according to the primary time domain features and primary frequency domain features, and generates a time domain feature set and a frequency domain feature set, the cross attention layer calculates the attention weight according to the time domain feature set and the frequency domain feature set, the pooling layer performs adaptive average pooling according to the attention weight calculation result, the fully connected layer performs linear transformation on the result after adaptive evaluation pooling, and finally the softmax layer performs recognition and classification of the power quality disturbance signal according to the transformation result to obtain the power quality disturbance classification result.

[0009] Preferably, after acquiring the power quality disturbance signal data, the method further includes:

[0010] The power quality disturbance signal data is subjected to data outlier elimination and difference compensation, and then normalization processing is performed.

[0011] Preferably, model training is performed based on power quality disturbance signal data, including:

[0012] Dividing the power quality disturbance signal data into a training set, a validation set and a test set according to a preset ratio;

[0013] The parameters of the power quality disturbance classification model fused with time-frequency features are initialized, and then the parameters are trained using the training set. The trained power quality disturbance classification model fused with time-frequency features is verified using the verification set to determine whether the classification accuracy of the power quality disturbance classification model fused with time-frequency features is the highest accuracy. When it is the highest accuracy, the corresponding model parameters are saved to obtain the trained power quality disturbance classification model fused with time-frequency features.

[0014] Preferably, using the verification set to verify the trained time-frequency feature fusion power quality disturbance classification model also includes:

[0015] The loss value is calculated by a cross entropy loss function, and the model parameters of the time-frequency feature fusion power quality disturbance classification model are updated by the loss value.

[0016] Preferably, the convolution layer includes a first convolution structure and a second convolution structure;

[0017] Inputting the power quality disturbance signal data into the first convolution structure for convolution operation, and summing the result after the convolution operation and the first offset value to obtain a primary time domain feature;

[0018] The power quality disturbance signal data is subjected to fast Fourier transform to obtain a frequency domain discrete signal, and then the frequency domain discrete signal is input into the second convolution structure for convolution operation, and the result after the convolution operation and the second bias value are summed to obtain a primary frequency domain feature.

[0019] Preferably, the BiGRU layer includes a first bidirectional gated loop structure and a second bidirectional gated loop structure; the first bidirectional gated loop structure and the second bidirectional gated loop structure are both composed of two independent gated recurrent neural network units;

[0020] Inputting the primary time domain features into the first bidirectional gated loop structure for calculation to obtain a final hidden state in the time domain, and generating a time domain feature set according to the final hidden state in the time domain;

[0021] The primary frequency domain features are input into the second bidirectional gated loop structure for calculation to obtain a final hidden state in the frequency domain, and a frequency domain feature set is generated according to the final hidden state in the frequency domain.

[0022] Preferably, the cross attention layer includes a query vector, a key vector, a value vector and a softmax function;

[0023] The time domain feature set is input into the key vector and the value vector, and the frequency domain feature set is input into the query vector, and finally the attention weight is calculated by the softmax function.

[0024] Compared with the prior art, the present invention has the following beneficial effects:

[0025] By constructing a time-frequency feature fusion power quality disturbance classification model, the time domain and frequency domain information of the power quality disturbance signal are fused. The model has strong noise resistance, high training speed and good classification accuracy. It can quickly and accurately classify the power quality disturbance of the power supply system of the coal mine excavation working face. Through this model, real-time power quality disturbance identification and classification can be performed, and the power quality problems existing in the power supply system of the excavation working face can be clarified, so as to provide early warning information for operating personnel and take measures in advance to avoid or reduce losses. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0027] Figure 1 A schematic flow chart of a method for classifying power quality disturbances in a tunnel power supply system provided by an embodiment of the present invention;

[0028] Figure 2 A network structure diagram of a time-frequency feature fusion power quality disturbance classification model provided in an embodiment of the present invention;

[0029] Figure 3 A flow chart of training and testing a time-frequency feature fusion power quality disturbance classification model provided in an embodiment of the present invention;

[0030] Figure 4 A graph showing the loss and accuracy change trend during the training process of the power quality disturbance classification method based on measured data provided by an embodiment of the present invention;

[0031] Figure 5 A confusion matrix of classification results of a test set of a power quality disturbance classification method based on measured data provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0032] In conjunction with the drawings in the embodiments of the present invention, the technical solutions in the embodiments of the present invention are clearly and completely described. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other implementation methods obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0033] It should be noted that the structures, proportions, sizes, etc. illustrated in the drawings of this specification are only used to match the contents disclosed in the specification so that people familiar with this technology can understand and read them. They are not used to limit the conditions under which the present invention can be implemented, and therefore have no substantive technical significance. Any structural modification, change in proportional relationship or adjustment of size should fall within the scope of the technical contents disclosed in the present invention without affecting the effects and purposes that can be achieved by the present invention. It should be noted that in this specification, relational terms such as first and second are only used to distinguish one entity from several other entities, and do not necessarily require or imply any actual relationship or order between these entities.

[0034] Power quality disturbance (PQD) is a significant problem that affects the safe operation of power grids. The traditional PQD classification method is the "feature extraction + pattern classification" method. Feature extraction requires manual participation and there is no unified method guidance. Often after a lot of work, the classification performance of the extracted features is not ideal. In recent years, convolutional neural networks have been used to classify power quality disturbances, but most of them only use the original time domain signal for multiple learning and extraction of features, and do not make full use of other features such as frequency domain. Therefore, this application provides a method for classifying power quality disturbances in a tunneling power supply system. By constructing a power quality disturbance classification model that fuses time-frequency features, the features of the time domain and frequency domain are combined to more accurately classify power quality disturbances.

[0035] In the embodiment of the present invention, a high-power oil pump motor and a cutting motor in the excavation equipment of the coal mine excavation working face are used as an example for explanation, and the working process includes: shutdown, no-load operation, and loaded cutting. By continuously collecting its three-phase voltage and comparing it with the data of the power quality analyzer, the collected voltage data is segmented and each segment of the data is marked with a corresponding power quality disturbance category label, the power quality disturbance signal data is obtained, and the original data set X is generated. The present invention is not limited to this, and it only needs to realize the corresponding function.

[0036] like Figure 1 As shown, an embodiment of the present invention provides a method for classifying power quality disturbances in a tunnel power supply system, comprising the following steps:

[0037] S1: Acquire power quality disturbance signal data;

[0038] S2: construct a time-frequency feature fusion power quality disturbance classification model, and perform model training based on power quality disturbance signal data; the time-frequency feature fusion power quality disturbance classification model at least includes a convolution layer, a BiGRU layer, a cross attention layer, a pooling layer, a fully connected layer, and a softmax layer;

[0039] S3: Collect the power quality disturbance signal data of the tunnel power supply system and input it into the time-frequency feature fusion power quality disturbance classification model; firstly, perform feature extraction through the convolution layer to obtain primary time domain features and primary frequency domain features, the BiGRU layer extracts the corresponding time series features according to the primary time domain features and primary frequency domain features, generates a time domain feature set and a frequency domain feature set, the cross attention layer calculates the attention weight according to the time domain feature set and the frequency domain feature set, the pooling layer performs adaptive average pooling according to the attention weight calculation result, the fully connected layer performs linear transformation on the result after adaptive evaluation pooling, and finally, the softmax layer performs recognition and classification of the power quality disturbance signal according to the transformation result to obtain the power quality disturbance classification result.

[0040] Optionally, after acquiring the power quality disturbance signal data, the method further includes: removing data outliers and performing difference compensation on the power quality disturbance signal data, and then performing normalization processing.

[0041] In this embodiment, the normalization processing method is:

[0042]

[0043] Where, X new is the normalized power quality disturbance signal data set, X i is the i-th voltage value in the power quality disturbance signal data set, X min is the minimum voltage value in the power quality disturbance signal data set, X max is the maximum voltage of the power quality disturbance signal data set.

[0044] In this embodiment, a time-frequency feature fusion power quality disturbance classification model based on FFT-CNN-BiGRU-CrossAttention is constructed. The model structure to be constructed is: two parallel first convolution structures and second convolution structures - bidirectional gated recurrent unit (BiGRU) combined with cross attention mechanism (CrossAttention) as the main framework of the model, and finally through the pooling layer, full connection layer and softmax layer, the classification result of power quality disturbance is obtained. The model structure is as follows Figure 2 shown.

[0045] Optionally, model training is performed based on power quality disturbance signal data, including: dividing the power quality disturbance signal data into a training set, a verification set and a test set according to a preset ratio; initializing parameters of a time-frequency feature fusion power quality disturbance classification model, and then performing parameter training using the training set, and verifying the trained time-frequency feature fusion power quality disturbance classification model using the verification set, and determining whether the classification accuracy of the time-frequency feature fusion power quality disturbance classification model is the highest accuracy. When it is the highest accuracy, saving the corresponding model parameters to obtain a trained time-frequency feature fusion power quality disturbance classification model.

[0046] In this embodiment, the training process of the time-frequency feature fusion power quality disturbance classification model is as follows: Figure 3 As shown, according to Figure 3 The training process shown iteratively trains and verifies the time-frequency feature fusion power quality disturbance classification model until the set number of training times is reached and the optimal model is output.

[0047] Optionally, after verifying the trained time-frequency feature fusion power quality disturbance classification model using the verification set, it also includes: calculating a loss value through a cross entropy loss function, and updating model parameters of the time-frequency feature fusion power quality disturbance classification model through the loss value.

[0048] In this embodiment, the cross entropy loss function is defined as follows:

[0049] L=-Σ(y i log(p i ))

[0050] Where: y i The power quality disturbance label i The value of the class, which is 0 or 1; p i The model predicts i The probability of the class; Σ is the sum of all categories.

[0051] In this embodiment, RMSProp is used to optimize the time-frequency feature fusion power quality disturbance classification model. The model optimization algorithm RMSProp includes two parts: calculating the moving average of the square of the gradient and updating the parameters.

[0052] The moving average formula for calculating the square of the gradient is as follows:

[0053]

[0054] Where: E[g 2 ] t is the moving average of the squared gradient of the tth iteration; β is the decay rate, which usually takes a value between 0 and 1 and controls the weight of the moving average; is the square of the gradient at the tth iteration.

[0055] The update parameter formula is as follows:

[0056]

[0057] Where: θ t is the parameter of the tth iteration; α is the learning rate; ∈ is a small constant used to prevent division by zero errors, usually 10 -8 , g t is the gradient of the tth iteration.

[0058] Optionally, the convolution layer includes a first convolution structure and a second convolution structure;

[0059] Inputting the power quality disturbance signal data into the first convolution structure for convolution operation, and summing the result after the convolution operation and the first offset value to obtain a primary time domain feature;

[0060] The power quality disturbance signal data is subjected to fast Fourier transform to obtain a frequency domain discrete signal, and then the frequency domain discrete signal is input into the second convolution structure for convolution operation, and the result after the convolution operation and the second bias value are summed to obtain a primary frequency domain feature.

[0061] In this embodiment, the time domain data set and the frequency domain data set of the power quality disturbance signal data are convolved with the corresponding convolution kernels respectively, and the result after the convolution operation and the corresponding offset value are summed to obtain the primary feature Z 时域 With Z 频域 , the convolution operation formula is as follows:

[0062]

[0063] Where: Input data X∈R L (where L is the length of the input data), convolution kernel k∈R F (where F is the size of the convolution kernel), the output Z∈R of the one-dimensional convolution layer L′ (Where L' is the output data length).

[0064] In this embodiment, the one-dimensional power quality disturbance signal data is input into the first convolution structure for convolution to extract the global features in the time domain. After the one-dimensional power quality disturbance signal data is subjected to fast Fourier transform to generate frequency domain features, it is input into the second convolution structure for convolution operation to capture the local pattern of the frequency domain features.

[0065] In this embodiment, the Fast Fourier Transformation (FFT) is used to extract frequency domain features and generate a frequency domain data set discrete signal defined as:

[0066]

[0067] Where: j is the imaginary number symbol, X(K) is the relative amplitude of N discrete frequency points (integer multiples of 50 Hz) corresponding to X(n), and K is the frequency index.

[0068] Optionally, the BiGRU layer includes a first bidirectional gated loop structure and a second bidirectional gated loop structure; the first bidirectional gated loop structure and the second bidirectional gated loop structure are both composed of two independent gated recurrent neural network units;

[0069] Inputting the primary time domain features into the first bidirectional gated loop structure for calculation to obtain a final hidden state in the time domain, and generating a time domain feature set according to the final hidden state in the time domain;

[0070] The primary frequency domain features are input into the second bidirectional gated loop structure for calculation to obtain a final hidden state in the frequency domain, and a frequency domain feature set is generated according to the final hidden state in the frequency domain.

[0071] In this embodiment, the BiGRU network structure is composed of two independent GRUs (Gated Recurrent Neural Network Units). The first bidirectional gated recurrent structure processes data in the forward direction of the time series, and the second bidirectional gated recurrent structure processes data in the reverse direction of the time series. Through this bidirectional structure, the BiGRU model can capture both the forward and backward information of the sequence data and can adapt to more complex sequence patterns. 时域 With Z 频域 Input them into the first bidirectional gated recurrent structure and the second bidirectional gated recurrent structure of BiGRU respectively to calculate h′ of the time domain feature t时域 and h′ of frequency domain characteristics t频域 .

[0072] The calculation process of the GRU unit is as follows:

[0073] Update Gate:

[0074] z t =σ(W z ·[h t-1 ,x t ]+b z )

[0075] Where: x t Enter information for the current moment; t is the update gate; W z is the update gate weight matrix; h t-1 is the hidden state of the previous moment; σ is the sigmoid function, through which the data can be converted into a value in the range of 0-1; b z is the bias corresponding to the update gate.

[0076] Reset the gate:

[0077] r t =σ(W r ·[h t-1 ,x t ]+b r )

[0078] Where: r t To reset the gate, W r To reset the gate weight matrix, b r is the bias corresponding to the reset gate.

[0079] New candidate status:

[0080]

[0081] Where: W is the weight matrix, is the candidate hidden state, b h is the corresponding bias, and tanh is the hyperbolic tangent function.

[0082] Update hidden state:

[0083]

[0084] Where: h t is the new hidden state of GRU.

[0085] The first bidirectional gated recurrent structure and the second bidirectional gated recurrent structure generate the final output by combining the new hidden states in the two directions of the forward GRU and the reverse GRU, respectively. The forward hidden state is calculated in the two directions of the forward GRU and the reverse GRU. With the reverse hidden state The forward hidden state and the reverse hidden state are combined to obtain the final hidden state, thereby determining h′ of the first bidirectional gated recurrent structure. t时域 and h′ of the second bidirectional gated loop structure t频域 , the final hidden state h′ t as follows:

[0086]

[0087] Where: h′ t is the final hidden state of BiGRU, and || represents vector concatenation.

[0088] Optionally, the criss-cross attention layer includes a query vector, a key vector, a value vector, and a criss-cross attention layer softmax function;

[0089] The time domain feature set is input into the key vector and the value vector, and the frequency domain feature set is input into the query vector, and finally the attention weight is calculated by the cross attention layer softmax function.

[0090] In this embodiment, let x F =h′ t频域 、x T =h′ t时域 , the query vector Q, key vector K, and value vector V are defined as follows:

[0091]

[0092] Where: W Q , W V , W K Represents the learnable parameter matrix of the corresponding vector in the criss-cross attention mechanism.

[0093] In this embodiment, the softmax function is:

[0094]

[0095] Where: Z i is the primary feature of the output of the i-th node, and m is the number of output nodes.

[0096] In this embodiment, the attention score is calculated by the query vector Q and the key vector K, and the weight W of each combined feature vector is obtained by the softmax function. A :

[0097] W A =softmax(Q·K)

[0098] Then combine the value vector to output the final output V′ of the cross attention layer:

[0099] V′=W A ·V

[0100] In this embodiment, the pooling layer performs adaptive average pooling on the attention weight results of the cross attention layer to obtain an output vector Y, which contains the number of elements corresponding to a number of power quality disturbance feature data, and each element is defined as:

[0101]

[0102] Where: k t is the number of power quality disturbance features in the pooling window corresponding to the t-th output element, s t and e t are the starting and ending indexes of the tth pooling window on the input power quality disturbance feature data V′; V′ i is the i-th element in the power quality disturbance feature.

[0103] In this embodiment, the output vector Y after adaptive pooling is passed through a fully connected layer and a softmax layer to obtain the power quality disturbance signal category P, and the power quality disturbance signal category P is:

[0104] p=softmax(W z ·Y)

[0105] Where: W z is the weight vector of the fully connected layer.

[0106] To further verify the present invention, the present application pre-classifies the acquired power quality disturbance signal data to obtain standard signals and common power quality disturbance categories in the excavation working face, including: C0: standard sinusoidal signal, C1: voltage swell, C2: voltage sag, C3: voltage interruption, C4: voltage flicker, C5: harmonics, C6: transient oscillation, C7: transient pulse, C8: voltage notch, C9: voltage spike, C10: harmonics + swell, C11: harmonics + sag, C12: harmonics + interruption, C13: harmonics + flicker, C14: harmonics + pulse, C15: flicker + oscillation, C16: flicker + pulse, C17: swell + oscillation, C18: sag + oscillation.

[0107] The collected power quality disturbance signal data is preprocessed to obtain the original data set X, and then the original data set is divided into training sets (X) according to the ratio of [70%, 20%, 10%] 时训 ), validation set (X 时验 ), test set (X 时测 ), and at the same time, the time-frequency feature fusion power quality disturbance classification model is constructed for parameter initialization, and the initial learning rate is set to: 0.0002; batch_size (number of samples per iteration) is: 64; the number of training times is: 50; the first convolution structure is ((2,32), (2,64), (2,128), (2,256)), the second convolution structure is ((2,16), (2,32), (2,64), (2,128)); the number of BiGRU layers and dimensions is: [128,256,64].

[0108] The original data set X is input into the first convolution structure for convolution to extract primary time domain features. At the same time, the original data set X is transformed by FFT to generate frequency domain features and input into the second convolution structure for convolution to extract primary frequency domain features.

[0109] The extracted primary time domain features and primary frequency domain features are respectively input into the BiGRU layer including the first bidirectional gated loop structure and the second bidirectional gated loop structure to extract the timing features, capture the long-term dependency of the signal data, and generate h′ t时域 and h′ t频域 .

[0110] The frequency domain feature set h′ t频域 Input the query vector (query) to the CrossAttention mechanism and convert the time domain feature set h′ t时域At the same time, the key vector (keys) and value vector (values) of the cross attention mechanism (CrossAttention) are input to calculate the corresponding attention weights. The calculated attention weights are used to enhance the attention to the important features of the power quality disturbance signal. Finally, the recognition and classification result P of the power quality disturbance signal is output through the fully connected layer and the softmax activation function, and the classification accuracy pre is calculated.

[0111] According to the above steps, the constructed time-frequency feature fusion power quality disturbance classification model is trained using the training set, and the model parameters are reversely updated through the loss value calculated by the cross entropy loss function, and the validation set is used to verify the quality of the model parameters. The specific training process is as follows: Figure 3 As shown, the model is iteratively trained and verified until the set number of training times is reached and the optimal model is output. The loss value and accuracy of the training set and verification set during the training process are shown in Figure 4 shown.

[0112] Finally, the test set is input into the trained model to verify the generalization ability of the model. The test set confusion matrix is ​​as follows: Figure 5 As shown, the darker the color, the more correct the classification; Table 1 shows the classification accuracy of each power quality disturbance in the test set.

[0113] Table 1

[0114]

[0115] The method for classifying power quality disturbances in the power supply system of a tunneling roadway proposed in the present application fuses the time domain and frequency domain information of the power quality disturbance signal, has strong noise resistance, high training speed and good classification accuracy, and can quickly and accurately classify the power quality disturbances in the power supply system of a coal mine tunneling working face; at the same time, the frequency of occurrence of various types of power quality disturbances is calculated through real-time power quality disturbance classification, and early warning information can be provided based on the frequency of occurrence of power quality disturbances, thereby predicting possible serious failures or accidents, so that operating personnel can take measures in advance to avoid or reduce losses.

[0116] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A method for classifying power quality disturbances in a tunnel power supply system, characterized in that: include: Obtain power quality disturbance signal data; Constructing a power quality disturbance classification model with time-frequency feature fusion, and training the model based on power quality disturbance signal data; the power quality disturbance classification model with time-frequency feature fusion includes at least a convolution layer, a BiGRU layer, a cross attention layer, a pooling layer, a fully connected layer, and a softmax layer; The power quality disturbance signal data of the tunneling tunnel power supply system is collected and input into the time-frequency feature fusion power quality disturbance classification model; first, feature extraction is performed in the convolution layer to obtain primary time domain features and primary frequency domain features, the BiGRU layer extracts corresponding time series features according to the primary time domain features and primary frequency domain features, and generates a time domain feature set and a frequency domain feature set, the cross attention layer calculates the attention weight according to the time domain feature set and the frequency domain feature set, the pooling layer performs adaptive average pooling according to the attention weight calculation result, the fully connected layer performs linear transformation on the result after adaptive evaluation pooling, and finally the softmax layer performs recognition and classification of the power quality disturbance signal according to the transformation result to obtain the power quality disturbance classification result.

2. The method for classifying power quality disturbances in a tunneling power supply system according to claim 1 is characterized in that: After obtaining the power quality disturbance signal data, it also includes: The power quality disturbance signal data is subjected to data outlier elimination and difference compensation, and then normalization processing is performed.

3. The method for classifying power quality disturbances in a tunneling power supply system according to claim 1 is characterized in that: Model training based on power quality disturbance signal data, including: Dividing the power quality disturbance signal data into a training set, a validation set and a test set according to a preset ratio; The parameters of the power quality disturbance classification model fused with time-frequency features are initialized, and then the parameters are trained using the training set. The trained power quality disturbance classification model fused with time-frequency features is verified using the verification set to determine whether the classification accuracy of the power quality disturbance classification model fused with time-frequency features is the highest accuracy. When it is the highest accuracy, the corresponding model parameters are saved to obtain the trained power quality disturbance classification model fused with time-frequency features.

4. The method for classifying power quality disturbances in a tunneling power supply system according to claim 3 is characterized in that: Using the verification set to verify the trained time-frequency feature fusion power quality disturbance classification model also includes: The loss value is calculated by a cross entropy loss function, and the model parameters of the time-frequency feature fusion power quality disturbance classification model are updated by the loss value.

5. The method for classifying power quality disturbances in a tunneling power supply system according to claim 1 is characterized in that: The convolution layer includes a first convolution structure and a second convolution structure; Inputting the power quality disturbance signal data into the first convolution structure for convolution operation, and summing the result after the convolution operation and the first offset value to obtain a primary time domain feature; The power quality disturbance signal data is subjected to fast Fourier transform to obtain a frequency domain discrete signal, and then the frequency domain discrete signal is input into the second convolution structure for convolution operation, and the result after the convolution operation and the second bias value are summed to obtain a primary frequency domain feature.

6. The method for classifying power quality disturbances in a tunneling power supply system according to claim 1, characterized in that: The BiGRU layer includes a first bidirectional gated loop structure and a second bidirectional gated loop structure; the first bidirectional gated loop structure and the second bidirectional gated loop structure are both composed of two independent gated recurrent neural network units; Inputting the primary time domain features into the first bidirectional gated loop structure for calculation to obtain a final hidden state in the time domain, and generating a time domain feature set according to the final hidden state in the time domain; The primary frequency domain features are input into the second bidirectional gated loop structure for calculation to obtain a final hidden state in the frequency domain, and a frequency domain feature set is generated according to the final hidden state in the frequency domain.

7. The method for classifying power quality disturbances in a tunneling power supply system according to claim 1, characterized in that: The cross attention layer includes a query vector, a key vector, a value vector and a softmax function; The time domain feature set is input into the key vector and the value vector, and the frequency domain feature set is input into the query vector, and finally the attention weight is calculated by the softmax function.

Images