Cable terminal discharge warning method, device and system based on dual-channel waveform identification

Abstract

The present invention relates to the field of data signal processing, and specifically to a network algorithm cable terminal discharge early warning method, device and system based on dual-channel waveform identification. First, the method collects the periodic ultrasonic signals and ultra-high frequency electromagnetic field signals generated by the cable terminal discharge, uses the short-time Fourier transform to convert these two signals into corresponding time-frequency diagrams, and performs normalization processing as the input of the subsequent network. A dual-channel high-order feature extraction network composed of a cascaded ConvNeXt backbone network and a convolutional block attention module is constructed to respectively extract the time-frequency diagram detail fault features of the ultrasonic signal and the high-frequency electromagnetic field signal. Finally, by constructing an adaptive feature fusion network, the fault features of the two channels are fused, and the cross-channel fault joint features are obtained using the adaptive weight coefficient vector, so as to output the diagnosis result and realize the accurate assessment of the defect state near the cable terminal.

Description

Technical Field

[0001] The present invention relates to the field of data signal processing, and specifically to a network algorithm cable terminal discharge warning method, device and system based on dual-channel waveform identification. Background Art

[0002] The development of urban economy has made the urban land resources increasingly tense. To improve the utilization rate of urban land, a large number of newly built lines or existing operating lines have been transformed to operate in the form of overhead-cable hybrid, and the cable rate has been increasing year by year. As an important medium connecting the cable line and the overhead transmission wire lead wire, the cable terminal plays the roles of shielding, insulation and stress control, and is essential when the hybrid transmission line is put into operation.

[0003] The cable terminal joint has a complex structure and serious electric field distortion. Its insulation performance is lower than that of the cable body, and it is a weak part of the cable system. According to the statistics of the State Grid Corporation, currently among the operating faults of cable lines, the faults of cable accessories account for more than 70%; among the faults of cable accessories, the insulation faults account for about 40%. Therefore, it is of great significance to timely detect the defects of the cable terminal joint and carry out repairs to improve its operation reliability.

[0004] At present, the partial discharge detection technology has become an ideal detection method for detecting insulation defects of cable terminals. The more widely used methods mainly include the pulse current method, the ultra-high frequency method, the ultrasonic detection method, the infrared detection method, etc. Among them, the pulse current method obtains discharge information such as apparent discharge amount and discharge phase by measuring the pulse current on the grounding wire of the cable terminal, and has high measurement sensitivity, but is mostly used for off-line measurement; the ultra-high frequency method and the ultrasonic method detect the discharge ability by measuring the UHF and ultrasonic signals generated during the discharge of the cable terminal, and can realize on-line monitoring, but it is difficult to distinguish the discharge strength during single diagnosis, there are a large number of false alarms, and it is impossible to locate the specific phase of the specific loop, and there may be misjudgments; the infrared detection method measures the abnormal temperature rise generated by the discharge by receiving the infrared rays radiated by the cable terminal, and has strong anti-interference ability, but there are monitoring blind spots and poor economy.

[0005] When there are insulation defects in the cable terminal, a long-term developing partial discharge phenomenon will be formed, accompanied by significant changes in ultrasonic, UHF, and temperature signals. Among them, the ultrasonic signal and the UHF signal will diffuse into the space with air as the medium, and the temperature signal can cause abnormal temperature rise on the terminal surface. The existing cable terminal partial discharge detection means generally adopt single parameter monitoring respectively, there are problems of multi-system islands, a large number of false alarms are difficult to eliminate, and it is difficult to locate the specific terminal. Summary of the Invention

[0006] This patent provides a network algorithm cable terminal discharge early warning method, device and system based on dual-channel waveform identification, which can realize synchronous monitoring of ultrasonic and UHF signals, realize the evaluation of the defect state near the cable terminal, and combine the surface temperature measurement of the cable terminal to accurately locate the specific loop, specific point and specific phase discharge point of the cable terminal.

[0007] The technical solution of the present invention is as follows:

[0008] A cable terminal discharge early warning method based on dual-channel waveform identification, characterized by comprising the following steps:

[0009] S1: Collect the periodic ultrasonic signal and UHF signal generated by the cable terminal discharge through the hybrid line panoramic perception and rapid inspection system, then perform STFT transformation on the collected ultrasonic signal and UHF signal respectively to obtain their corresponding time-frequency diagrams, and normalize them as the input of the overall network;

[0010] S2: Construct a dual-channel high-order feature extraction network for ultrasonic signals and UHF signals formed by cascading the ConvNeXt backbone network and the convolutional block attention module, and each channel respectively obtains the detailed fault features of the corresponding time-frequency diagram and

[0011] S3: Construct an adaptive feature fusion network to obtain the fault features and of the adaptive weight coefficient vector, fuse to obtain the cross-channel fault joint feature, output the diagnosis result, and realize the evaluation of the defect state near the cable terminal;

[0012]

[0013] where F is the cross-channel fault joint feature after fusion; α and β are the adaptive weight coefficient vectors, determined by the adaptive weight network composed of the global pooling layer, the fully connected layer and the Sigmoid activation function, satisfying α i ∈[0,1], β i ∈[0,1], α i +β i =1;

[0014] The process of obtaining the adaptive weight coefficient vectors α and β is as follows: First, perform global pooling on the fault features and to compress the fault features to the depth dimension to obtain the fault global features G I and G U , as shown in Equation (4);

[0015]

[0016] Then, the global feature G I and G U are merged, and then through the fully connected layer and the Sigmoid activation function, the fault global features G I and G U are converted into weights, as shown in Equation (5).

[0017]

[0018] In the formula, H represents the number of column vectors of the extracted features, is the depth of d , d = 1, 2,..., D; G I , G U ∈R 1×1×D ; FC represents the fully connected layer function for realizing feature weighting.

[0019] For the cable terminal discharge warning method based on dual-channel waveform identification, measure the abnormal temperature rise generated by the cable terminal discharge, and cooperate with the evaluation result of the adjacent cable terminal defect state output in step S3 to judge the accurate positioning of the defective discharge position.

[0020] For the cable terminal discharge warning method based on dual-channel waveform identification, for the ultrasonic signal, the detection frequency band range is 20 kHz to 100 MHz, and it is collected regularly; for the UHF signal, the detection frequency band range is 300 MHz to 3 GHz, and it is collected synchronously and regularly with the ultrasonic sensor; the temperature measurement range is -40°C to 125°C, and it is collected synchronously and regularly with the ultrasonic sensor.

[0021] For the cable terminal discharge warning method based on dual-channel waveform identification, in step S1, the implementation process of the Fourier STFT transform is as follows:

[0022] ① Select appropriate window functions such as Hamming window, Hanning window, rectangular window, etc. to intercept short time segments of the ultrasonic

[0023] signal and the high-frequency electromagnetic field signal waveform;

[0024] ② Frame the input signal waveform according to the window length and overlap rate;

[0025] ③ Perform the fast Fourier transform FFT on each frame of the signal waveform and generate a time-frequency diagram,

[0026] and display the intensity of the signal waveform at different times and frequencies through the power spectrum;

[0027] The expression of the time-domain signal s(t) after the STFT transform is:

[0028]

[0029] Where f is the frequency, S(t, f) is the output result of the STFT transform, s(τ) is the original input signal, and w(t) is a window function with a fixed length.

[0030] In the cable terminal discharge warning method based on dual-channel waveform identification, in step S1, the normalization process is to adjust the time-frequency map data to a distribution with a mean of 0 and a standard deviation of 1 through the Z-score standardization method. The implementation process is as follows:

[0031]

[0032] Where X is the original time-frequency map data, X’ is the normalized data, μ is the mean, and σ is the standard deviation.

[0033] In the cable terminal discharge warning method based on dual-channel waveform identification, in step S2,

[0034] The ConvNeXt backbone network includes the following architectures:

[0035] ① Input layer. It is used to receive the time-frequency image data after the STFT transform and normalization of ultrasonic and UHF signals. Usually, it is an RGB image of 224×224×3, and the number of channels is 3;

[0036] ② Initial processing block. It performs initial convolution operations to extract low-level features. Downsampling is performed using convolution with a stride of 2, and the GELU activation function is used to convert the feature map of 224×224×3 into a feature map with a shape of 56×56×96, and the number of channels is 96;

[0037] ③ Convolution block. The convolution block is the core component of the network, consisting of multiple convolutional layers and optimized modules. The convolution operation is used for feature extraction, introducing non-linearity through the GELU activation function, and then normalizing the feature map to alleviate the problem of gradient disappearance during training. A skip connection similar to ResNet is used to add the input and output to promote gradient flow, and partial paths are randomly discarded to enhance the robustness of the model. After convolution and activation, the shape of the feature map is transformed into 7×7×768, and the number of channels is 768;

[0038] ④ Bottleneck layer. 1*1 convolution is used for channel compression to reduce the number of channels and spatial dimensions to effectively reduce the computational amount, thereby improving the computational efficiency, and the number of channels of the feature map remains unchanged.

[0039] The convolutional block attention module is an attention mechanism used to enhance the performance of convolutional neural networks. First, channel attention is applied to the time-frequency map to obtain key channel information; then spatial attention is applied to mark the key regions of the time-frequency map. Based on this, the detailed features of the time-frequency map corresponding to each channel can be accurately obtained.

[0040] In the cable terminal discharge warning method based on dual-channel waveform identification, in step S2, the implementation method of the convolutional block attention module is as follows:

[0041] ① Channel attention module: Perform global average pooling and global maximum pooling on the input time-frequency diagram to generate two channel descriptors. Pass these two descriptors through a shared multi-layer perceptron to generate a channel attention map, and apply the attention weight through element-wise multiplication with the original feature map. After passing through the channel attention module, the shape of the feature map remains at 7×7×768, but the features of each channel have been weighted according to the channel attention;

[0042] ② Spatial attention module: Perform global average pooling and global maximum pooling respectively on the outputs of each channel attention module to generate spatial descriptors. Combine these two descriptors to generate a spatial attention map, and also multiply it with the feature map through element-wise multiplication to apply spatial attention. After passing through the spatial attention module, the shape of the feature map remains at 7×7×768, but the features at each spatial position have been weighted according to the spatial attention.

[0043] A computer device includes a memory, a processor, and a computer program stored on the memory and executable on the processor. It is characterized in that when the processor executes the program, it implements the steps in the cable terminal discharge warning method based on the dual-channel synchronous waveform identification network algorithm as described in any one of the foregoing items.

[0044] A cable terminal discharge warning system based on dual-channel waveform identification is installed on the cable terminal and includes a central control system. It is characterized in that it includes a sensing module, a short-distance communication module, a data preprocessing module, a data analysis module, a data fusion module, a fault diagnosis module, and an integrated communication module;

[0045] The sensing module includes an ultrasonic sensor, a UHF sensor, and a wireless temperature measurement sensor. The ultrasonic sensor and the UHF sensor are integrated inside the system. The ultrasonic sensor is used to measure the periodic ultrasonic signals generated by the cable terminal discharge, and the UHF sensor is used to measure the periodic UHF signals generated by the cable terminal discharge. The wireless temperature measurement sensor adopts a strap-type infrared temperature measurement technology and integrates an optical system and a photoelectric converter. It is installed on the surface of the cable terminal and is used to measure the abnormal temperature rise generated by the cable terminal discharge;

[0046] The short-distance communication module uses a wireless transmission and reception module to achieve short-distance communication between the wireless temperature measurement sensor and the system;

[0047] The data preprocessing module includes a filtering circuit and a signal amplification circuit, which respectively perform basic noise filtering and signal amplification on the collected ultrasonic, UHF, and temperature signals for subsequent data analysis and data fusion;

[0048] The described data analysis module preliminarily classifies the collected data according to the time-frequency characteristics of the collected signals, distinguishes effective and invalid signals for the collected ultrasonic, UHF, and temperature signals respectively, discards the invalid signals, and stores and uploads the effective signals.

[0049] The described data fusion module integrates the data from ultrasonic sensors, UHF sensors, and temperature sensors, ensures the consistency of data from different sources, processes data format and timestamp differences, enhances the fault tolerance of the system to data loss or sensor failures, and ensures the stable operation of the system.

[0050] The described data storage module stores the effective data collected at regular intervals and supports real-time retrieval and viewing.

[0051] The described integrated communication module summarizes the data of each module inside the device and realizes external communication with the monitoring center, cloud platform, or other terminal devices for more complex data processing and artificial intelligence training to achieve accurate early warning of cable terminal defect discharges.

[0052] The described cable terminal discharge early warning system based on dual-channel waveform identification is characterized in that

[0053] The detection frequency band range of the described ultrasonic sensor is 20 kHz to 100 MHz.

[0054] The detection frequency band range of the described UHF sensor is 300 MHz to 3 GHz.

[0055] The temperature measurement range of the described wireless temperature measurement sensor is -40°C to 125°C.

[0056] The described short-distance communication module uses a 433 MHz wireless transmitting and receiving module.

[0057] UHF signal refers to the ultra-high frequency electromagnetic field signal.

[0058] The present invention first collects the periodic ultrasonic signals and UHF signals generated by cable terminal discharges. These two signals respectively reflect different physical characteristics of cable terminal discharges. The ultrasonic signal mainly reflects the mechanical vibration effect of the discharge, while the UHF signal reflects the electromagnetic radiation generated by the discharge.

[0059] The collected signals are transformed by STFT to convert the time-domain signals into time-frequency diagrams. The STFT transformation can provide information about the signal in both the time and frequency domains, which is helpful for subsequent feature extraction. The time-frequency diagrams are normalized to ensure the consistency of the input data and facilitate the processing of subsequent network models.

[0060] The present invention constructs a dual-channel high-order feature extraction network formed by cascading a ConvNeXt backbone network and a convolutional block attention module. The ConvNeXt network is based on its powerful feature extraction ability, while the convolutional block attention module further enhances the network's ability to extract key features by introducing an attention mechanism.

[0061] The ultrasonic signal channel and the UHF signal channel are respectively input into the corresponding network branches. After a series of convolutional operations and the processing of the attention mechanism, the time-frequency map detail fault features of their respective channels are extracted.

[0062] The present invention constructs an adaptive feature fusion network for fusing the fault features of the ultrasonic signal channel and the UHF signal channel. This network can dynamically adjust the weights of the features of the two channels to obtain more accurate cross-channel fault joint features.

[0063] The present invention outputs the discharge warning result and the defect state evaluation of the cable terminal. By comparing with a preset threshold or the model prediction result, it can be judged whether there is a discharge phenomenon in the cable terminal and the severity of the discharge, so as to take corresponding preventive measures.

[0064] The cable terminal discharge warning method based on dual-channel waveform identification combines the dual detection means of ultrasonic signals and UHF signals, and uses an advanced deep learning network for feature extraction and fusion, realizing accurate warning of cable terminal discharge phenomena and defect state evaluation. This technical solution provides a strong technical guarantee for the safe operation of the power system.

[0065] The present invention collects ultrasonic signals and UHF signals generated by cable terminal discharge, and uses a ConvNeXt backbone network and a convolutional block attention module to extract high-order features. Through an adaptive feature fusion network, the features of the two channels are fused to obtain cross-channel fault joint features. Finally, based on this feature, the discharge warning result and the defect state evaluation of the cable terminal are output. The present invention improves the accuracy and reliability of cable terminal discharge warning and provides a strong guarantee for the safe operation of the power system.

[0066] Aiming at problems such as hidden cable terminal defects and multi-monitoring system islands, the present invention provides a cable terminal discharge warning method based on dual-channel waveform identification. It is installed on a cable terminal tower and has the functions of measuring ultrasonic signals, UHF signals, and the surface temperature signal of the cable terminal. Through processes such as data preprocessing, dual-channel high-order feature extraction, and adaptive feature fusion, accurate identification of the healthy state, defect state, and fault state of the cable terminal is realized. Combining the abnormal temperature rise signal monitored on its surface, accurate positioning of the defect location is realized. Description of the Drawings

[0067] Figure 1This is the flowchart of the present invention.

[0068] Figure 2 This is the working principle diagram of the present invention.

[0069] Figure 3 This is the architecture of the dual-channel synchronous waveform identification network algorithm model.

[0070] Figure 4 This is an example of constructing the time-frequency diagram of UHF signals.

[0071] (a) Original UHF signal waveform - left; (b) Power spectrum after STFT transformation.

[0072] Figure 5 This is an example of constructing the time-frequency diagram of ultrasonic signals.

[0073] (a) Original ultrasonic signal waveform - left; (b) Power spectrum after STFT transformation.

[0074] Figure 6 This is the basic architecture of the dual-channel high-order feature extraction network.

[0075] Figure 7 This is the adaptive weight network. Detailed implementation

[0076] Example 1

[0077] See Figure 2 . The cable terminal discharge warning system based on dual-channel waveform identification is installed near the tower of the cable terminal and includes a central control system, which includes a sensing module, a short-distance communication module, a data preprocessing module, a data analysis module, a data fusion module, a fault diagnosis module, and an integrated communication module;

[0078] The sensing module includes an ultrasonic sensor, a UHF sensor, and a wireless temperature measurement sensor. The ultrasonic sensor and the UHF sensor are integrated inside the system.

[0079] (1) Ultrasonic sensor. It uses piezoelectric ceramic materials and is integrated inside the device to measure the periodic ultrasonic signals generated by the cable terminal discharge. The detection frequency band range is 20 kHz to 100 MHz, and it is collected regularly at a frequency of 1 time / h.

[0080] (2) UHF sensor. It uses ceramics, polymers, and metals as the substrate materials and is integrated inside the device to measure the periodic UHF signals generated by the cable terminal discharge. The detection frequency band range is 300 MHz to 3 GHz, and it is collected synchronously and regularly with the ultrasonic sensor at a frequency of 1 time / h.

[0081] (3) Wireless temperature measurement sensor. Adopting the strap-type infrared temperature measurement technology, with an optical system and a photoelectric converter integrated inside, it is installed on the surface of the cable terminal to measure the abnormal temperature rise generated by the discharge of the cable terminal. Combining with the diagnostic results of ultrasonic and UHF signal fusion, it determines the specific terminal points with defective discharge phenomena. The temperature measurement range is -40°C to 125°C, and it synchronously collects data at regular intervals with the ultrasonic sensor, and the collection frequency is 1 time / h.

[0082] (4) Short-distance communication module. Adopting a 433MHz wireless transmitting and receiving module to achieve short-distance communication between the wireless temperature measurement sensor and the device main body within a range of 200m.

[0083] (5) Data preprocessing module. It includes a filtering circuit and a signal amplification circuit, which respectively perform basic noise filtering and signal amplification on the collected ultrasonic, UHF, and temperature signals for subsequent data analysis and data fusion.

[0084] (6) Data analysis module. According to the time-frequency characteristics of the collected signals, it preliminarily classifies the collected data, distinguishes the effective and invalid signals of the collected ultrasonic, UHF, and temperature signals respectively, discards the invalid signals, and stores and uploads the effective signals.

[0085] (7) Data fusion module. It integrates the data from multiple sensors (ultrasonic sensor, UHF sensor, temperature sensor), ensures the consistency of data from different sources, processes differences in data formats, timestamps, etc., enhances the fault tolerance of the system to data loss or sensor failures, and ensures the stable operation of the system.

[0086] (8) Data storage module. It stores the effective data collected at regular intervals, and the storage time is not less than 3 months, supporting real-time retrieval and viewing.

[0087] (9) Integrated communication module. It summarizes the data of each module inside the device and realizes external communication with the monitoring center, cloud platform or other terminal devices for more complex data processing and artificial intelligence training to achieve accurate early warning of cable terminal defective discharge.

[0088] Example 2

[0089] See Figures 1 - 7 . A cable terminal discharge early warning method based on a dual-channel synchronous waveform identification network algorithm includes the following steps:

[0090] S1: Collect the periodic ultrasonic signals generated by the cable terminal discharge, measure the UHF signals generated by the cable terminal inlet discharge, perform STFT transformation on the collected waveforms of the ultrasonic signals and high-frequency electromagnetic field signals to obtain their corresponding time-frequency diagrams; and perform normalization processing on them as the input of the overall network;

[0091] Construction of Time-Frequency Diagram of the Present Invention

[0092] A large number of theoretical studies and engineering practices have shown that when there are insulation defects in the cable terminal and continuous discharge occurs, the waveform data of the ultrasonic and UHF signals generated have different characteristics in both the time domain and the frequency domain. However, the actual operating state of the power system is complex and variable, and it is difficult to effectively extract the characteristics of cable terminal defect discharges in the time domain and the frequency domain; and considering that the discharges of the line cable terminals are such that both the ultrasonic and UHF signals have obvious changing characteristics, and there is a certain similarity and complementarity in their defect information, this patent uses the STFT transform to convert the waveform of the ultrasonic signal and the UHF signal generated during the cable terminal discharge into a time-frequency diagram (time-frequency image) to express its characteristics in a higher dimension.

[0093] The implementation idea of the STFT transform is as follows:

[0094] ① Select a suitable window function (such as Hamming window, Hanning window, rectangular window, etc.) to intercept a short time segment of the signal;

[0095] ② Frame the input signal according to the window length and overlap rate;

[0096] ③ Perform a fast Fourier transform (FFT) on each frame of the signal and generate a time-frequency diagram, and the intensity of the signal at different times and frequencies can be displayed through the power spectrum.

[0097] The expression of the time-domain signal s(t) after the STFT transform is:

[0098]

[0099] In the formula, f is the frequency, S(t,f) is the output result of the STFT transform, s(τ) is the original input signal, and w(t) is a window function with a fixed length.

[0100] Demonstration of implementation effect:

[0101] Then, normalize the time-frequency diagram data. Through the Z-score standardization method, adjust the data to a distribution with a mean of 0 and a standard deviation of 1. The implementation process is as follows:

[0102]

[0103] In the formula, X is the original time-frequency diagram data, X’ is the normalized data, μ is the mean, and σ is the standard deviation.

[0104] S2: Construct a dual-channel high-order feature extraction network for ultrasonic signals and high-frequency electromagnetic field signals formed by cascading the ConvNeXt backbone network and the convolutional block attention module, and each channel obtains the detailed fault characteristics of its corresponding time-frequency diagram and

[0105] The basic architecture of the dual-channel high-order feature extraction network is as Figure 6 shown. ConvNeXt is an architecture based on the convolutional neural network (CNN), aiming to combine the design of modern convolutional networks with the self-attention mechanism to improve the performance of tasks such as image classification. It mainly consists of the following architectures:

[0106] ① Input layer. It is used to receive the time-frequency image data after the STFT transformation and normalization of ultrasonic and UHF signals. Usually, it is an RGB image of 224×224×3, and the number of channels is 3;

[0107] ② Initial processing block. It performs initial convolution operations to extract low-level features. It uses convolution with a stride of 2 for downsampling and the GELU activation function to convert the feature map of 224×224×3 into a feature map with a shape of 56×56×96, and the number of channels is 96;

[0108] ③ Convolution block. The convolution block is the core component of the network, consisting of multiple convolutional layers and optimized modules. The convolution operation is used for feature extraction, introducing non-linearity through the GELU activation function, and then normalizing the feature map to alleviate the vanishing gradient problem during training. It uses skip connections similar to ResNet to add the input and output to promote gradient flow, and randomly drops some paths to enhance the robustness of the model. After convolution and activation, the shape of the feature map is transformed into 7×7×768, and the number of channels is 768;

[0109] ④ Bottleneck layer. It uses 1*1 convolution for channel compression, reducing the number of channels and spatial dimensions to effectively reduce the computational amount, thereby improving the computational efficiency, and the number of channels of the feature map remains unchanged.

[0110] The convolutional block attention module is an attention mechanism used to enhance the performance of convolutional neural networks. First, it applies channel attention to the feature map to obtain key channel information; then it applies spatial attention to mark the key regions of the feature map. Based on this, the detailed features of each channel corresponding to the time-frequency map can be accurately obtained.

[0111] The implementation key points are as follows:

[0112] ① Channel attention module: It performs global average pooling and global maximum pooling on the input time-frequency map to generate two channel descriptors, passes these two descriptors through a shared multi-layer perceptron to generate a channel attention map, and applies the attention weight through element-wise multiplication with the original feature map. After passing through the channel attention module, the shape of the feature map remains at 7×7×768, but the features of each channel have been weighted according to the channel attention;

[0113] ② Spatial attention module: Global average pooling and global max pooling are respectively performed on the outputs of each channel attention module to generate spatial descriptors. These two descriptors are combined to generate a spatial attention map, which is also multiplied with the feature map through element-wise multiplication to apply spatial attention. After passing through the spatial attention module, the shape of the feature map remains at 7×7×768, but the features at each spatial position have been weighted according to the spatial attention.

[0114] S3: Construct an adaptive feature fusion network to obtain the fault features of ultrasonic and UHF signals respectively through a dual-channel high-order feature extraction network and When diagnosing their individual fault features, there are certain limitations, resulting in a large number of false alarms.

[0115] As Figure 7 shown, to better represent the fault features, a method for constructing an adaptive weight network is proposed to obtain the adaptive weight coefficient vector, integrate to obtain the cross-channel fault joint feature, output the diagnostic result, and realize the evaluation of the defect state of the adjacent cable terminal;

[0116]

[0117] where F is the fused cross-channel fault joint feature; α and β are the adaptive weight coefficient vectors, determined by the adaptive weight network composed of a global pooling layer, a fully connected layer, and a Sigmoid activation function, satisfying α i ∈[0,1], β i ∈[0,1], α i +β i =1;

[0118] The process of obtaining the adaptive weight coefficient vectors α and β is as follows: First, perform global pooling on the fault features and to compress the fault features to the depth dimension to obtain the fault global features G I and G U , as shown in Equation (4);

[0119]

[0120] Then, merge the global features G I and G U , and then pass the merged result through a fully connected layer and a Sigmoid activation function to convert the fault global features G I and G U into weights, as shown in Equation (5).

[0121]

[0122] Wherein, H represents the number of feature column vectors extracted, is the depth of d , where d = 1, 2,..., D; G I , G U ∈R 1×1×D ; FC represents the fully connected layer function for realizing feature weighting.

[0123] Measure the abnormal temperature rise generated by the discharge of the cable terminal, combine the abnormal temperature rise signal monitored on the surface of the cable terminal, and cooperate with the evaluation result of the defect state of the adjacent cable terminal output in step S3 to determine the accurate positioning of the defective discharge location.

Claims

1. A cable terminal discharge warning method based on dual-channel waveform identification, characterized in that It includes the following steps: S1: Collect the periodic ultrasonic signals and ultra-high frequency electromagnetic field UHF signals generated by the cable terminal discharge, then perform short-time Fourier STFT transforms on the collected ultrasonic signals and ultra-high frequency electromagnetic field UHF signals respectively to obtain their corresponding time-frequency diagrams, and normalize them to be used as the input of the overall network; S2: Construct a dual-channel high-order feature extraction network for ultrasonic signals and ultra-high frequency electromagnetic field (UHF) signals by cascading the ConvNeXt backbone network and the convolutional block attention module, and obtain the detailed fault features of the corresponding time-frequency diagrams for each channel respectively. and ; S3: Construct an adaptive feature fusion network to obtain fault features and the adaptive weight coefficient vector, fuse to obtain cross-channel fault joint features, output the diagnostic result, and realize the defect state evaluation of the adjacent cable terminal; （3） In the formula F is the combined cross-channel fault joint feature after fusion; α and β is an adaptive weight coefficient vector, determined by an adaptive weight network composed of a global pooling layer, a fully connected layer, and a Sigmoid activation function, satisfying α i ∈[0,1], β i ∈[0,1], α i + β i = 1; Adaptive weight coefficient vector α and β The acquisition process is as follows: First, perform global pooling on the fault features of the ultrasonic and ultra-high frequency electromagnetic field UHF signals and to compress the fault features into the depth dimension to obtain the fault global features G I and G U , as shown in Equation (4); （4） Then the global features G I and G U are merged, and then the fault global features G I and G U are converted into weights as shown in Equation (5). （5）， Wherein, H represents the number of feature column vectors extracted, and are d the depths of d = 1, 2, ..., D ; G I , G U ∈ R 1×1×D ; FC represents the fully connected layer function for implementing feature weighting.

2. The cable terminal discharge warning method based on dual-channel waveform identification according to claim 1, wherein Measure the abnormal temperature rise generated by the cable terminal discharge, and cooperate with the evaluation result of the adjacent cable terminal defect state output in step S3 to judge the accurate positioning of the defective discharge position.

3. The cable terminal discharge warning method based on dual-channel waveform identification according to claim 2, wherein For the ultrasonic signals, the detection frequency band range is 20kHz~100MHz, and they are collected regularly; for the UHF signals, the detection frequency band range is 300MHz~3GHz, and they are collected synchronously and regularly with the ultrasonic sensor; for the abnormal temperature rise, the temperature measurement range is -40°C~125°C, and it is collected synchronously and regularly with the ultrasonic sensor.

4. The cable terminal discharge warning method based on dual-channel waveform identification according to claim 1, characterized in that In step S1, the implementation process of the STFT transform is as follows: ① Select appropriate window functions such as Hamming window, Hanning window, and rectangular window to intercept short time segments of the ultrasonic signal and UHF signal waveforms; ② Frame the input signal waveforms according to the window length and overlap rate; ③ Perform fast Fourier transform on each frame of the signal waveform and generate a time-frequency diagram, and display the intensity of the signal waveform at different times and frequencies through the power spectrum; Time-domain signal s ( t ) The expression after STFT transformation is: （1） wherein, f is the frequency, S ( t , f ) is the output result of the STFT transform, is the original input signal, w ( t ) is a window function with a fixed length.

5. The method for early warning of cable terminal discharge based on dual-channel waveform identification according to claim 1, characterized in that In step S1, the normalization process is to adjust the time-frequency diagram data to a distribution with a mean of 0 and a standard deviation of 1 through the Z-score normalization method; its implementation process is as follows: （2）， In the formula, X is the original time-frequency diagram data, X’ is the data after normalization, μ is the mean value, is the standard deviation.

6. The method for early warning of cable terminal discharge based on dual-channel waveform identification according to claim 1, characterized in that, In step S2, The ConvNeXt backbone network includes the following architectures: ① Input layer, used to receive the time-frequency image data of ultrasonic and UHF signals after STFT transform and normalization, usually an RGB image of 224×224×3, and the number of channels is 3; ② Preliminary processing block, perform preliminary convolution operations to extract low-level features; Use convolution with a stride of 2 for downsampling, use the GELU activation function, and convert the feature map of 224×224×3 into a feature map with a shape of 56×56×96, and the number of channels is 96; ③ Convolution block, the convolution block is the core component of the network, composed of multiple convolutional layers and optimized modules. The convolution operation is used for feature extraction, introduces non-linearity through the GELU activation function, then normalizes the feature map to alleviate the problem of gradient disappearance during training, uses skip connections similar to ResNet to add the input and output to promote gradient flow, and randomly discards some paths to enhance the robustness of the model. After convolution and activation, the shape of the feature map is transformed into 7×7×768, and the number of channels is 768; ④ Bottleneck layer, use 1*1 convolution for channel compression, reduce the number of channels and spatial dimensions to effectively reduce the calculation amount, thereby improving the calculation efficiency, and the number of channels of the feature map remains unchanged; The convolutional block attention module is an attention mechanism used to enhance the performance of convolutional neural networks. First, apply channel attention to the time-frequency diagram to obtain key channel information; then apply spatial attention to mark the key areas of the time-frequency diagram. Based on this, the detailed features of the time-frequency diagrams corresponding to each channel can be accurately obtained.

7. The method for cable terminal discharge warning based on dual-channel waveform identification according to claim 6, characterized in that, In step S2, the implementation method of the convolutional block attention module is as follows: ① Channel attention module: Perform global average pooling and global max pooling on the input time-frequency map to generate two channel descriptors. Pass these two descriptors through a shared multi-layer perceptron to generate a channel attention map, and apply the attention weights through element-wise multiplication with the original feature map. After passing through the channel attention module, the shape of the feature map remains 7×7×768, but the features of each channel have been weighted according to the channel attention; ② Spatial attention module: Perform global average pooling and global max pooling respectively on the outputs of each channel attention module to generate spatial descriptors. Combine these two descriptors to generate a spatial attention map, and also multiply it with the feature map through element-wise multiplication to apply spatial attention. After passing through the spatial attention module, the shape of the feature map remains 7×7×768, but the features at each spatial position have been weighted according to the spatial attention.

8. A computer device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the cable terminal discharge warning method based on the dual-channel synchronous waveform identification network algorithm as described in any one of claims 1-7.

9. A cable terminal discharge warning system based on dual-channel waveform identification is used to implement the cable terminal discharge warning method based on dual-channel waveform identification according to any one of claims 1-7, and is characterized in that, Installed at the cable terminal, the central control system includes a sensing module, a short-range communication module, a data preprocessing module, a data analysis module, a data fusion module, a fault diagnosis module, and an integrated communication module; The sensing module includes an ultrasonic sensor, a UHF sensor, and a wireless temperature sensor. The ultrasonic sensor and the UHF sensor are integrated inside the system. The ultrasonic sensor is used to measure the periodic ultrasonic signals generated by the cable terminal discharge, and the UHF sensor is used to measure the periodic UHF signals generated by the cable terminal discharge. The wireless temperature sensor uses a strap-type infrared temperature measurement technology and integrates an optical system and a photoelectric converter. It is installed on the surface of the cable terminal and is used to measure the abnormal temperature rise generated by the cable terminal discharge; The short-range communication module uses a wireless transmission and reception module to achieve short-range communication between the wireless temperature sensor and the system; The data preprocessing module includes a filter circuit and a signal amplification circuit, which respectively perform basic noise filtering and signal amplification on the collected ultrasonic, UHF, and temperature signals for subsequent data analysis and data fusion; The data analysis module preliminarily classifies the collected data according to the time-frequency characteristics of the collected signals, distinguishes the valid and invalid signals of the collected ultrasonic, UHF, and temperature signals respectively, discards the invalid signals, and stores and uploads the valid signals; The data fusion module integrates the data from the ultrasonic sensor, the UHF sensor, and the temperature sensor to ensure the consistency of data from different sources, processes the differences in data formats and timestamps, enhances the fault tolerance of the system to data loss or sensor failures, and ensures the stable operation of the system; The data storage module stores the valid data collected at regular intervals and supports real-time retrieval and viewing; The integrated communication module aggregates the data of each module inside the device and realizes external communication with the monitoring center and the cloud platform for more complex data processing and artificial intelligence training to achieve accurate early warning of cable terminal defect discharge.

10. The cable terminal discharge early warning system based on dual-channel waveform identification according to claim 9, characterized in that the detection frequency band range of the ultrasonic sensor is 20 kHz to 100 MHz; the detection frequency band range of the ultra-high frequency sensor is 300 MHz to 3 GHz; the temperature measurement range of the wireless temperature measurement sensor is -40°C to 125°C; the short-distance communication module uses a 433 MHz wireless transmitting and receiving module.

Images