Cable defect identification method and device based on multiple physical quantities

Abstract

The invention discloses a cable defect identification method and device based on multiple physical quantities, and belongs to the technical field of cable maintenance, and the method comprises the steps: obtaining the multiple physical quantity time sequence data of a cable; converting the time sequence data, inputting the converted time sequence data into a neural network, and extracting a first composite vector; inputting the time sequence data into a DNet network to obtain a time sequence feature vector, inputting the first composite vector into an ENet network to obtain a second composite vector, and splicing the second composite vector to obtain a Hybrid feature vector; the Hybrid is input into the FNet network, and a third composite vector is obtained; inputting the third composite vector into a GNet network to obtain a fourth composite vector; positive and negative samples are selected for comparative learning, and network weights of DNet, ENet and FNet are finely adjusted; and replacing the GNet network with a multi-layer perceptron, carrying out cable defect identification supervision training, and outputting a cable defect type. According to the invention, the performance of reflecting cable operation defects is enhanced, and the accuracy of cable defect identification is improved.

Description

Technical Field

[0001] The invention relates to a cable defect identification method and device based on multiple physical quantities, belonging to the technical field of cable maintenance. Background Art

[0002] With the development of power transmission system, the safety and reliability of cables as an important power transmission equipment are increasingly valued. However, due to various reasons, such as defects in the manufacturing process, material aging, etc., cables may fail during use, leading to safety accidents. Therefore, regular inspection and fault diagnosis of cables are particularly important.

[0003] Cable defect identification aims to ensure the safe operation of cables in power systems and reduce power outages caused by cable faults. Currently, most methods use a single cable parameter to extract and identify features. For example, ultrasonic detection of partial discharge uses the ultrasonic probe to attach to the cable surface and uses the reflection and refraction principle of ultrasonic waves to detect partial discharge and then determine defects inside the cable, such as cracks. At the same time, there are also methods that use voiceprint detection technology and convert voiceprints into MFCC spectrograms and input them into deep learning models to identify defects.

[0004] However, there has been no in-depth research on the full fusion of the characteristics of multiple physical quantities of cables. At the same time, how to take the fusion of the characteristics of multiple physical quantities into consideration for the visualization of time series data also needs further discussion. Summary of the invention

[0005] In order to solve the above problems, the present invention proposes a cable defect identification method and device based on multiple physical quantities, which can enhance the performance of reflecting cable operation defects and improve the accuracy of cable defect identification.

[0006] The technical solution adopted by the present invention to solve the technical problem is: In a first aspect, an embodiment of the present invention provides a cable defect identification method based on multiple physical quantities, comprising the following steps: Step S1, obtaining multi-physical quantity time series data of the cable, wherein the time series data includes temperature data, partial discharge pulse data and voiceprint data; Step S2, converting the time series data into a JG-DCT image, and inputting the JG-DCT image into a neural network to extract a first feature synthesis mapping vector; Step S3, inputting the multi-physical quantity time series data into the DNet network to obtain a time series feature vector, inputting the first feature synthesis mapping vector into the ENet network to obtain a second feature synthesis mapping vector, and concatenating the second feature synthesis mapping vector and the time series feature vector to obtain a Hybrid feature vector; Step S4, inputting the Hybrid feature vector into the FNet network to obtain a third feature synthesis mapping vector; Step S5, inputting the third feature synthesis mapping vector into the GNet network to obtain a fourth feature synthesis mapping vector; Step S6, selecting positive and negative samples for comparative learning and fine-tuning the weights of the DNet network, the ENet network, and the FNet network to obtain pre-trained network weights; Step S7, replacing the GNet network with a multi-layer perceptron model, and performing supervised training for cable defect recognition based on the pre-trained network weights and the fourth feature synthesis mapping vector, and outputting the cable defect type.

[0007] As a possible implementation of this embodiment, converting the time series data into a JG-DCT image includes the following steps: Extract a window sequence L(n) from the time series S(x) of multiple physical quantities according to the interval N, where 0 <n<N=w 2 , w represents the width of the generated image; The time series of multiple physical quantities are constructed into multiple G images through the following formula: , , Where P is the G image, (l,c) is the horizontal and vertical coordinate index value of the G image, min(L) and max(L) are the maximum and minimum values ​​of the window sequence respectively; The JG image is obtained by averaging multiple G image elements corresponding to multiple physical quantities: , Among them, (x, y) is the horizontal and vertical coordinate index of the JG image, and m is the number of physical quantities.

[0008] As a possible implementation of this embodiment, the step of converting the time series data into a JG-DCT image further includes the following steps: After obtaining the JG image matrix, perform two-dimensional DCT transformation according to the following formula to obtain the JG-DCT image: , , Wherein, JG_D(u,v) represents the pixel on the JG-DCT image after the JG image is transformed by two-dimensional DCT, and C(u,v) is the conversion kernel function; The JG-DCT image is input into a neural network, and a first feature synthesis mapping vector is output, where the dimension of the first feature synthesis mapping vector is 1024.

[0009] As a possible implementation method of this embodiment, the DNet network includes a first residual network module, a second residual network module, a third residual network module, a fourth residual network module, a first pooling module, a second pooling module, a third pooling module and a first fully connected layer. The first residual network module, the second residual network module and the first pooling module constitute a first mapping module, the third residual network module and the second pooling module constitute a second mapping module, and the fourth residual network module and the third pooling module constitute a third mapping module. The multi-physical quantity time series data is used as the input of the first mapping module of the DNet network, the output of the first mapping module is used as the input of the second mapping module, the output of the second mapping module is used as the input of the third mapping module, and the output of the third mapping module is used as the input of the first fully connected layer. The first fully connected layer outputs a time series feature vector with a dimension of 512.

[0010] As a possible implementation of this embodiment, the ENet network includes a second fully connected layer, the first feature synthesis mapping vector is used as an input of the second fully connected layer, and the second fully connected layer outputs a second feature synthesis mapping vector with a dimension of 512.

[0011] As a possible implementation of this embodiment, the dimension of the Hybrid feature vector is 1024, which is obtained by cross-joining the time series feature vector and the second feature synthesis mapping vector: , Among them, c(2n) is the Hybrid eigenvector, is the time series feature vector, Synthesize the mapping vector for the second feature.

[0012] As a possible implementation of this embodiment, the FNet network includes a third fully connected layer, the Hybrid feature vector is used as the input of the third fully connected layer, and the third fully connected layer outputs a third feature synthesis mapping vector with a dimension of 512.

[0013] As a possible implementation of this embodiment, the GNet network includes a fourth fully connected layer, the third feature synthesis mapping vector is used as an input of the fourth fully connected layer, and the fourth fully connected layer outputs a fourth feature synthesis mapping vector with a dimension of 32.

[0014] As a possible implementation of this embodiment, step S6 includes the following steps: Select N positive and negative samples of the same category based on the collected multi-physical quantity time series data; The selected positive and negative samples are processed in sequence using the methods of step S2 to step S5 to obtain a fourth feature synthesis mapping vector; Construct the following loss function and perform contrastive learning: , , Among them, sim(*) is the feature similarity function , are the 2k-1th and 2kth fourth feature synthesis mapping vectors respectively, is the weight parameter; The weights of the DNet network, ENet network and FNet network are fine-tuned by the gradient descent method.

[0015] As a possible implementation of this embodiment, in the process of comparative learning, when N positive samples, i.e., the fourth feature synthesis mapping vector generated by multi-physical quantity time series data of the same category are selected for comparative learning, minimizing the loss function is used as the basis for fine-tuning; When N negative samples, i.e., the fourth feature synthetic mapping vector generated by multi-physical quantity time series data of the same category, are selected for comparative learning, maximizing the loss function is used as the basis for fine-tuning.

[0016] In a second aspect, an embodiment of the present invention provides a cable defect identification device based on multiple physical quantities, comprising: A data acquisition module is used to obtain time series data of multiple physical quantities of the cable, wherein the time series data includes temperature data, partial discharge pulse data and voiceprint data; A neural network module, used for converting the time series data into a JG-DCT image, and inputting the JG-DCT image into a neural network to extract a first feature synthesis mapping vector; The ENet network module is used to input the multi-physical quantity time series data into the DNet network to obtain a time series feature vector, input the first feature synthesis mapping vector into the ENet network to obtain a second feature synthesis mapping vector, and concatenate the second feature synthesis mapping vector and the time series feature vector to obtain a Hybrid feature vector; FNet network module, used to input the Hybrid feature vector into the FNet network to obtain the third feature synthesis mapping vector; A GNet network module, used for inputting the third feature synthesis mapping vector into the GNet network to obtain a fourth feature synthesis mapping vector; The weight fine-tuning module is used to select positive and negative samples for comparative learning and fine-tune the weights of the DNet network, ENet network, and FNet network to obtain the pre-trained network weights; The multi-layer perceptron module is used to replace the GNet network with a multi-layer perceptron model, and perform supervised training for cable defect recognition based on the pre-trained network weights and the fourth feature synthesis mapping vector, and output the cable defect type.

[0017] As a possible implementation of this embodiment, the specific process of the neural network module converting the time series data into a JG-DCT image is as follows: Extract a window sequence L(n) from the time series S(x) of multiple physical quantities according to the interval N, where 0 <n<N=w 2 , w represents the width of the generated image; The time series of multiple physical quantities are constructed into multiple G images through the following formula: , , Where P is the G image, (l,c) is the horizontal and vertical coordinate index value of the G image, min(L) and max(L) are the maximum and minimum values ​​of the window sequence respectively; The JG image is obtained by averaging multiple G image elements corresponding to multiple physical quantities: , Among them, (x, y) is the horizontal and vertical coordinate index of the JG image, and m is the number of physical quantities; After obtaining the JG image matrix, perform two-dimensional DCT transformation according to the following formula to obtain the JG-DCT image: , , Wherein, JG_D(u,v) represents the pixel on the JG-DCT image after the JG image is transformed by two-dimensional DCT, and C(u,v) is the conversion kernel function; The JG-DCT image is input into a neural network, and a first feature synthesis mapping vector is output, where the dimension of the first feature synthesis mapping vector is 1024.

[0018] As a possible implementation method of this embodiment, the DNet network includes a first residual network module, a second residual network module, a third residual network module, a fourth residual network module, a first pooling module, a second pooling module, a third pooling module and a first fully connected layer. The first residual network module, the second residual network module and the first pooling module constitute a first mapping module, the third residual network module and the second pooling module constitute a second mapping module, and the fourth residual network module and the third pooling module constitute a third mapping module. The multi-physical quantity time series data is used as the input of the first mapping module of the DNet network, the output of the first mapping module is used as the input of the second mapping module, the output of the second mapping module is used as the input of the third mapping module, and the output of the third mapping module is used as the input of the first fully connected layer. The first fully connected layer outputs a time series feature vector with a dimension of 512.

[0019] As a possible implementation of this embodiment, the ENet network includes a second fully connected layer, the first feature synthesis mapping vector is used as an input of the second fully connected layer, and the second fully connected layer outputs a second feature synthesis mapping vector with a dimension of 512.

[0020] As a possible implementation of this embodiment, the dimension of the Hybrid feature vector is 1024, which is obtained by cross-joining the time series feature vector and the second feature synthesis mapping vector: , Among them, c(2n) is the Hybrid eigenvector, is the time series feature vector, Synthesize the mapping vector for the second feature.

[0021] As a possible implementation of this embodiment, the FNet network includes a third fully connected layer, the Hybrid feature vector is used as the input of the third fully connected layer, and the third fully connected layer outputs a third feature synthesis mapping vector with a dimension of 512.

[0022] As a possible implementation of this embodiment, the GNet network includes a fourth fully connected layer, the third feature synthesis mapping vector is used as an input of the fourth fully connected layer, and the fourth fully connected layer outputs a fourth feature synthesis mapping vector with a dimension of 32.

[0023] As a possible implementation of this embodiment, the specific process of the weight fine-tuning module obtaining the pre-trained network weights is as follows: Select N positive and negative samples of the same category based on the collected multi-physical quantity time series data; The selected positive and negative samples are processed in sequence using the methods of step S2 to step S5 to obtain a fourth feature synthesis mapping vector; Construct the following loss function and perform contrastive learning: , , Among them, sim(*) is the feature similarity function , are the 2k-1th and 2kth fourth feature synthesis mapping vectors respectively, is the weight parameter; The weights of the DNet network, ENet network and FNet network are fine-tuned by the gradient descent method.

[0024] As a possible implementation of this embodiment, in the process of comparative learning, when N positive samples, i.e., the fourth feature synthesis mapping vector generated by multi-physical quantity time series data of the same category are selected for comparative learning, minimizing the loss function is used as the basis for fine-tuning; When N negative samples, i.e., the fourth feature synthetic mapping vector generated by multi-physical quantity time series data of the same category, are selected for comparative learning, maximizing the loss function is used as the basis for fine-tuning.

[0025] The beneficial effects of the technical solution of the embodiment of the present invention are as follows: A cable defect recognition method based on multi-physical quantities in the technical solution of an embodiment of the present invention comprises the following steps: step S1, obtaining multi-physical quantity time series data of the cable, wherein the time series data comprises temperature data, partial discharge pulse data and voiceprint data; step S2, converting the time series data into a JG-DCT image, and inputting the JG-DCT image into a neural network to extract a first feature synthesis mapping vector; step S3, inputting the multi-physical quantity time series data into a DNet network to obtain a time series feature vector, inputting the first feature synthesis mapping vector into an ENet network to obtain a second feature synthesis mapping vector, and performing the second feature synthesis mapping vector and the time series feature vector into a neural network. Step S4, input the Hybrid feature vector into the FNet network to obtain the third feature synthesis mapping vector; Step S5, input the third feature synthesis mapping vector into the GNet network to obtain the fourth feature synthesis mapping vector; Step S6, select positive and negative samples for comparative learning and fine-tune the weights of the DNet network, the ENet network and the FNet network to obtain the pre-trained network weights; Step S7, replace the GNet network with a multi-layer perceptron model, and perform supervised training for cable defect recognition based on the pre-trained network weights and the fourth feature synthesis mapping vector, and output the cable defect type. The present invention has high sensitivity and high resolution in cable defect recognition, enhances the performance of reflecting cable operation defects through the fusion of multi-physical quantities and multi-modal features, and uses contrast learning to learn meaningful representations from multi-modal fusion data, and pre-trains the model by generating positive and negative samples, and then fine-tunes it for downstream tasks to achieve recognition accuracy equivalent to supervised learning. The present invention enhances the performance of reflecting cable operation defects and improves the accuracy of cable defect recognition.

[0026] The present invention converts the time series data of the time series signal into an image, especially by combining G coding and DCT transformation methods, so that different defect data features are more visualized; the present invention realizes the effective fusion of signals and key features and maintains the balance between modes by designing a unique contrast learning architecture; using the contrast learning architecture, the present invention learns meaningful representations from multimodal fusion data, which helps to enhance the representation ability of the time series signal and improves the accuracy of cable defect recognition. Contrastive learning generates positive and negative samples for pre-training, and then fine-tunes the network to perform downstream tasks. This method of the present invention is close to or even reaches the level of supervised learning in terms of recognition accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 is a flow chart of a cable defect identification method based on multiple physical quantities according to an exemplary embodiment; Figure 2 is a structural schematic diagram of a cable defect identification device based on multiple physical quantities according to an exemplary embodiment; Figure 3 The present invention is a specific implementation flow chart of cable defect identification using the method of the present invention. DETAILED DESCRIPTION

[0028] In order to more clearly illustrate the technical features of the solution of the present invention, the present invention is described in detail below through specific implementation methods and in conjunction with the accompanying drawings.

[0029] like Figure 1 As shown, a cable defect identification method based on multiple physical quantities provided by an embodiment of the present invention includes the following steps: Step S1, obtaining multi-physical quantity time series data of the cable, wherein the time series data includes temperature data, partial discharge pulse data and voiceprint data; Step S2, converting the time series data into a JG-DCT image, and inputting the JG-DCT image into a neural network to extract a first feature synthesis mapping vector; Step S3, inputting the multi-physical quantity time series data into the DNet network to obtain a time series feature vector, inputting the first feature synthesis mapping vector into the ENet network to obtain a second feature synthesis mapping vector, and concatenating the second feature synthesis mapping vector and the time series feature vector to obtain a Hybrid feature vector; Step S4, inputting the Hybrid feature vector into the FNet network to obtain a third feature synthesis mapping vector; Step S5, inputting the third feature synthesis mapping vector into the GNet network to obtain a fourth feature synthesis mapping vector; Step S6, selecting positive and negative samples for comparative learning and fine-tuning the weights of the DNet network, the ENet network, and the FNet network to obtain pre-trained network weights; Step S7, replacing the GNet network with a multi-layer perceptron model, and performing supervised training for cable defect recognition based on the pre-trained network weights and the fourth feature synthesis mapping vector, and outputting the cable defect type.

[0030] As a possible implementation of this embodiment, converting the time series data into a JG-DCT image includes the following steps: Extract a window sequence L(n) from the time series S(x) of multiple physical quantities according to the interval N, where 0 <n<N=w 2 , w represents the width of the generated image; The time series of multiple physical quantities are constructed into multiple G images through the following formula: , , Where P is the G image, (l,c) is the horizontal and vertical coordinate index value of the G image, min(L) and max(L) are the maximum and minimum values ​​of the window sequence respectively; The JG image is obtained by averaging multiple G image elements corresponding to multiple physical quantities: , Among them, (x, y) is the horizontal and vertical coordinate index of the JG image, and m is the number of physical quantities.

[0031] As a possible implementation of this embodiment, the step of converting the time series data into a JG-DCT image further includes the following steps: After obtaining the JG image matrix, perform two-dimensional DCT transformation according to the following formula to obtain the JG-DCT image: , , Wherein, JG_D(u,v) represents the pixel on the JG-DCT image after the JG image is transformed by two-dimensional DCT, and C(u,v) is the conversion kernel function; The JG-DCT image is input into a neural network, and a first feature synthesis mapping vector is output, where the dimension of the first feature synthesis mapping vector is 1024.

[0032] As a possible implementation method of this embodiment, the DNet network includes a first residual network module, a second residual network module, a third residual network module, a fourth residual network module, a first pooling module, a second pooling module, a third pooling module and a first fully connected layer. The first residual network module, the second residual network module and the first pooling module constitute a first mapping module, the third residual network module and the second pooling module constitute a second mapping module, and the fourth residual network module and the third pooling module constitute a third mapping module. The multi-physical quantity time series data is used as the input of the first mapping module of the DNet network, the output of the first mapping module is used as the input of the second mapping module, the output of the second mapping module is used as the input of the third mapping module, and the output of the third mapping module is used as the input of the first fully connected layer. The first fully connected layer outputs a time series feature vector with a dimension of 512.

[0033] As a possible implementation of this embodiment, the ENet network includes a second fully connected layer, the first feature synthesis mapping vector is used as an input of the second fully connected layer, and the second fully connected layer outputs a second feature synthesis mapping vector with a dimension of 512.

[0034] As a possible implementation of this embodiment, the dimension of the Hybrid feature vector is 1024, which is obtained by cross-joining the time series feature vector and the second feature synthesis mapping vector: , Among them, c(2n) is the Hybrid eigenvector, is the time series feature vector, Synthesize the mapping vector for the second feature.

[0035] As a possible implementation of this embodiment, the FNet network includes a third fully connected layer, the Hybrid feature vector is used as the input of the third fully connected layer, and the third fully connected layer outputs a third feature synthesis mapping vector with a dimension of 512.

[0036] As a possible implementation of this embodiment, the GNet network includes a fourth fully connected layer, the third feature synthesis mapping vector is used as an input of the fourth fully connected layer, and the fourth fully connected layer outputs a fourth feature synthesis mapping vector with a dimension of 32.

[0037] As a possible implementation of this embodiment, step S6 includes the following steps: Select N positive and negative samples of the same category based on the collected multi-physical quantity time series data; The selected positive and negative samples are processed in sequence using the methods of step S2 to step S5 to obtain a fourth feature synthesis mapping vector; Construct the following loss function and perform contrastive learning: , , Among them, sim(*) is the feature similarity function , are the 2k-1th and 2kth fourth feature synthesis mapping vectors respectively, is the weight parameter; The weights of the DNet network, ENet network and FNet network are fine-tuned by the gradient descent method.

[0038] As a possible implementation of this embodiment, in the process of comparative learning, when N positive samples, i.e., the fourth feature synthesis mapping vector generated by multi-physical quantity time series data of the same category are selected for comparative learning, minimizing the loss function is used as the basis for fine-tuning; When N negative samples, i.e., the fourth feature synthetic mapping vector generated by multi-physical quantity time series data of the same category, are selected for comparative learning, maximizing the loss function is used as the basis for fine-tuning.

[0039] like Figure 2 As shown, an embodiment of the present invention provides a cable defect identification device based on multiple physical quantities, comprising: A data acquisition module is used to obtain time series data of multiple physical quantities of the cable, wherein the time series data includes temperature data, partial discharge pulse data and voiceprint data; A neural network module, used for converting the time series data into a JG-DCT image, and inputting the JG-DCT image into a neural network to extract a first feature synthesis mapping vector; The ENet network module is used to input the multi-physical quantity time series data into the DNet network to obtain a time series feature vector, input the first feature synthesis mapping vector into the ENet network to obtain a second feature synthesis mapping vector, and concatenate the second feature synthesis mapping vector and the time series feature vector to obtain a Hybrid feature vector; FNet network module, used to input the Hybrid feature vector into the FNet network to obtain the third feature synthesis mapping vector; A GNet network module, used for inputting the third feature synthesis mapping vector into the GNet network to obtain a fourth feature synthesis mapping vector; The weight fine-tuning module is used to select positive and negative samples for comparative learning and fine-tune the weights of the DNet network, ENet network, and FNet network to obtain the pre-trained network weights; The multi-layer perceptron module is used to replace the GNet network with a multi-layer perceptron model, and perform supervised training for cable defect recognition based on the pre-trained network weights and the fourth feature synthesis mapping vector, and output the cable defect type.

[0040] As a possible implementation of this embodiment, the specific process of the neural network module converting the time series data into a JG-DCT image is as follows: Extract a window sequence L(n) from the time series S(x) of multiple physical quantities according to the interval N, where 0 <n<N=w 2 , w represents the width of the generated image; The time series of multiple physical quantities are constructed into multiple G images through the following formula: , , Where P is the G image, (l,c) is the horizontal and vertical coordinate index value of the G image, min(L) and max(L) are the maximum and minimum values ​​of the window sequence respectively; The JG image is obtained by averaging multiple G image elements corresponding to multiple physical quantities: , Among them, (x, y) is the horizontal and vertical coordinate index of the JG image, and m is the number of physical quantities; After obtaining the JG image matrix, perform two-dimensional DCT transformation according to the following formula to obtain the JG-DCT image: , , Wherein, JG_D(u,v) represents the pixel on the JG-DCT image after the JG image is transformed by two-dimensional DCT, and C(u,v) is the conversion kernel function; The JG-DCT image is input into a neural network, and a first feature synthesis mapping vector is output, where the dimension of the first feature synthesis mapping vector is 1024.

[0041] As a possible implementation method of this embodiment, the DNet network includes a first residual network module, a second residual network module, a third residual network module, a fourth residual network module, a first pooling module, a second pooling module, a third pooling module and a first fully connected layer. The first residual network module, the second residual network module and the first pooling module constitute a first mapping module, the third residual network module and the second pooling module constitute a second mapping module, and the fourth residual network module and the third pooling module constitute a third mapping module. The multi-physical quantity time series data is used as the input of the first mapping module of the DNet network, the output of the first mapping module is used as the input of the second mapping module, the output of the second mapping module is used as the input of the third mapping module, and the output of the third mapping module is used as the input of the first fully connected layer. The first fully connected layer outputs a time series feature vector with a dimension of 512.

[0042] As a possible implementation of this embodiment, the ENet network includes a second fully connected layer, the first feature synthesis mapping vector is used as an input of the second fully connected layer, and the second fully connected layer outputs a second feature synthesis mapping vector with a dimension of 512.

[0043] As a possible implementation of this embodiment, the dimension of the Hybrid feature vector is 1024, which is obtained by cross-joining the time series feature vector and the second feature synthesis mapping vector: , Among them, c(2n) is the Hybrid eigenvector, is the time series feature vector, Synthesize the mapping vector for the second feature.

[0044] As a possible implementation of this embodiment, the FNet network includes a third fully connected layer, the Hybrid feature vector is used as the input of the third fully connected layer, and the third fully connected layer outputs a third feature synthesis mapping vector with a dimension of 512.

[0045] As a possible implementation of this embodiment, the GNet network includes a fourth fully connected layer, the third feature synthesis mapping vector is used as an input of the fourth fully connected layer, and the fourth fully connected layer outputs a fourth feature synthesis mapping vector with a dimension of 32.

[0046] As a possible implementation of this embodiment, the specific process of the weight fine-tuning module obtaining the pre-trained network weights is as follows: Select N positive and negative samples of the same category based on the collected multi-physical quantity time series data; The selected positive and negative samples are processed in sequence using the methods of step S2 to step S5 to obtain a fourth feature synthesis mapping vector; Construct the following loss function and perform contrastive learning: , , Among them, sim(*) is the feature similarity function , are the 2k-1th and 2kth fourth feature synthesis mapping vectors respectively, is the weight parameter; The weights of the DNet network, ENet network and FNet network are fine-tuned by the gradient descent method.

[0047] As a possible implementation of this embodiment, in the process of comparative learning, when N positive samples, i.e., the fourth feature synthesis mapping vector generated by multi-physical quantity time series data of the same category are selected for comparative learning, minimizing the loss function is used as the basis for fine-tuning; When N negative samples, i.e., the fourth feature synthetic mapping vector generated by multi-physical quantity time series data of the same category, are selected for comparative learning, maximizing the loss function is used as the basis for fine-tuning.

[0048] like Figure 3 As shown, the specific implementation process of using the method of the present invention to identify cable defects is as follows.

[0049] Step 1: Obtain multi-physics timing data of the cable.

[0050] Multi-physical quantity time series data includes temperature data, partial discharge pulse data and voiceprint data.

[0051] The collection of cable partial discharge (PD) pulse data, voiceprint data and temperature data are all important components of cable status monitoring. They correspond to different monitoring methods and technologies to evaluate the health status of the cable system.

[0052] By using high-frequency current transformers (HFCT), voltage sensors or special coupling capacitors, these sensors are installed on the ground wire, cross-connection wire or directly coupled to the cable body to capture the high-frequency pulse current or voltage signal generated by partial discharge. The signal captured by the sensor is transmitted to the data collector through a coaxial cable or a dedicated connection line. The data collector can be portable or fixed, and is responsible for the preliminary amplification and analog-to-digital conversion (ADC) of the signal.

[0053] Ultrasonic sensors (such as ultrasonic probes) are attached to the surface of the cable or built-in near the cable terminal. These sensors are sensitive to the ultrasonic signals generated by partial discharge of the cable. After the ultrasonic signals are received by the sensors, they are converted into electrical signals, which are also amplified and processed by analog-to-digital conversion.

[0054] Temperature sensors, such as thermocouples, infrared cameras or fiber optic temperature sensors, are installed at key locations on the cable. These sensors can continuously monitor temperature changes on or inside the cable. Temperature data is sent to the monitoring system via wired or wireless communication technology. Infrared cameras can instantly display temperature distribution images for quick on-site inspections.

[0055] Step 2: Convert the time series data into JG-DCT images and extract the first feature synthesis mapping vector through a neural network.

[0056] Specifically, the following steps are included: Extract a window sequence L(n) from the time series S(x) of multiple physical quantities according to the interval N, where 0 <n<N=w 2 , w represents the width of the generated image; The time series of multiple physical quantities are constructed into multiple G images through the following formula: , , Where P is the G image, (l,c) is the horizontal and vertical coordinate index value of the G image, min(L) and max(L) are the maximum and minimum values ​​of the window sequence respectively; The JG image is obtained by averaging multiple G image elements corresponding to multiple physical quantities: , Among them, (x, y) is the horizontal and vertical coordinate index of the JG image, and m is the number of physical quantities; After obtaining the JG image matrix, the JG-DCT image is obtained according to the following formula to complete the conversion of time series data: , , Wherein, JG_D(u,v) represents the pixel on the JG-DCT image after the JG image is transformed by two-dimensional DCT, and C(u,v) is the conversion kernel function; After conversion, the dimension of the first feature synthesis mapping vector obtained by the neural network output is 1024.

[0057] Step 3: Use the multi-physical quantity time series data as the input of the DNet network to obtain the time series feature vector, use the first feature synthesis mapping vector as the input of the ENet network to obtain the second feature synthesis mapping vector, and then concatenate the second feature synthesis mapping vector and the time series feature vector to obtain the Hybrid feature vector.

[0058] DNet includes four residual network modules Resnet0, Resnet1, Resnet2, and Resnet3, three pooling modules pool1, pool2, and pool3, and a fully connected layer FC1. The Resnet0, Resnet1, and pool1 form a first mapping module, the Resnet2 and pool2 form a second mapping module, and the Resnet3 and pool3 form a third mapping module. The multi-physical quantity data is used as the input of the first mapping module of the DNet module, the output of the first mapping module is used as the input of the second mapping module, the output of the second mapping module is used as the input of the third mapping module, and the output of the third mapping module is used as the input of FC1. Finally, DNet outputs a time series feature vector with a dimension of 512.

[0059] ENet includes a fully connected layer FC2, the first feature synthesis mapping vector is used as the input of ENet, and the second feature synthesis mapping vector with a dimension of 512 is output.

[0060] Specifically, the structure of Resnet0 is [256*3*512], the structure of Resnet1 is [256*3*512], the structure of pool1 is [128*3*256], the structure of Resnet2 is [128*3*128], the structure of pool2 is [128*3*128], the structure of Resnet3 is [128*3*128], the structure of pool3 is [128*3*256], the structure of FC1 is [2048,1024,512], and the structure of FC2 is [256,128,512].

[0061] Step 4: Use the Hybrid feature vector as the input of the FNet network to obtain the third feature synthesis mapping vector.

[0062] The dimension of the Hybrid feature vector is 1024, that is, the time series feature vector and the second feature synthesis mapping vector are built-in and cross-joined to obtain the Hybrid feature vector. Definition is the time series feature vector, is the second feature synthesis mapping vector, and c(2n) is the Hybrid feature vector.

[0063] , Wherein FNet includes a fully connected layer FC3, the Hybrid feature vector is used as the input of the fully connected layer FC3, and the third feature synthesis mapping vector with a dimension of 512 is output.

[0064] The FC3 structure is [1024,512,512].

[0065] Step 5: Use the third feature synthesis mapping vector as the input of GNet to obtain the fourth feature synthesis mapping vector.

[0066] Specifically, GNet includes a fully connected layer FC4, which takes the third feature synthesis mapping vector as the input of the fully connected layer FC4 and outputs the fourth feature synthesis mapping vector with a dimension of 32. The structure of FC4 is [512, 256, 32].

[0067] Step 6: Then select positive and negative samples for comparative learning and fine-tune the weights of DNet, ENet, and FNet to obtain the pre-trained network weights.

[0068] Specifically, the specific process of selecting positive and negative samples for comparative learning and fine-tuning the weights of DNet, ENet, and FNet to obtain the pre-trained network weights is: based on the collected multi-physical quantity time series data, N positive and negative samples of the same category are selected, and the fourth feature synthesis mapping vector is obtained through steps 2 to 5 respectively, and the following loss function is constructed and comparative learning is performed, and then the weights of DNet, ENet, and FNet are fine-tuned by the gradient descent method: , , Among them, sim(*) is the feature similarity function , are the 2k-1th and 2kth fourth feature synthesis mapping vectors respectively, is the weight parameter.

[0069] Specifically, in the process of comparative learning, when N positive samples, i.e., the fourth feature synthetic mapping vector generated by multi-physical quantity time series data of the same category are selected for comparative learning, minimizing the loss function is used as the basis for fine-tuning; when N negative samples, i.e., the fourth feature synthetic mapping vector generated by multi-physical quantity time series data of the same category are selected for comparative learning, maximizing the loss function is used as the basis for fine-tuning.

[0070] Step 7: While fixing the network weights, discard GNet and replace it with the multi-layer perceptron model to perform supervised training on the fourth feature synthesis mapping vector for cable defect recognition and then output the cable defect type.

[0071] The MLP structure is [512, 1024, 1024, 256, 32, 4].

[0072] In order to ensure that the test data fits the actual engineering background, a high-voltage cable of model APJ30G-25kV is used. According to the statistical results of on-site maintenance of cable accessories, four typical cable defects are designed: semi-conductive residue, ring cutting, air discharge defect, and metal debris. Keep it in line with the actual situation during production. Semi-conductive residue is a triangular barb caused by the failure to strip flat at the end of the outer semi-conductive layer; ring cutting scratches are caused by excessive manual cutting force, resulting in scratches on the main insulation surface caused by the tool penetrating the outer semi-conductive layer, which is set to a 1mm wide and 0.5mm deep ring-shaped knife mark; air gap discharge defects are caused by the air gap caused by the lack of strict fit between the stress tube and the main insulation, which is set to a rectangular groove with a length of 50mm, a width of 0.5mm, and a depth of 1mm; metal defects are caused by metal particles randomly adhering to the main insulation due to the complex on-site environment.

[0073] In the supervised training process, the loss function is the cross entropy loss function: , Among them, pi is the probability that the model predicts that the sample belongs to the i-th class, and yi is an indicator variable, 𝑦𝑖=1 when the sample belongs to the i-th class, otherwise 𝑦𝑖=0.

[0074] After the supervised training is completed, the best network weight model is saved and then tested on the test set to output the cable defect type or normal type.

[0075] Compared with the prior art, the present invention has the following advantages: 1) By converting the time series data of the timing signal into images, especially by combining G coding and DCT transformation methods, different defect data features can be more visualized; 2) Traditional multimodal fusion methods often focus on information integration, but may ignore the balance between different modalities. This invention achieves effective fusion of signals and key features and maintains the balance between modalities by designing a unique contrastive learning architecture; 3) Using contrastive learning architecture, the present invention learns meaningful representations from multimodal fusion data, which helps to enhance the representation ability of time series signals and improve the accuracy of cable defect recognition. Contrastive learning generates positive and negative samples for pre-training, and then fine-tunes the network to perform downstream tasks. This method is close to or even reaches the level of supervised learning in terms of recognition accuracy.

[0076] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the specific implementation modes of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be included in the scope of protection of the claims of the present invention.

Claims

1. A cable defect identification method based on multiple physical quantities, characterized in that: The steps include: Step S1, obtaining multi-physical quantity time series data of the cable, wherein the time series data includes temperature data, partial discharge pulse data and voiceprint data; Step S2, converting the time series data into a JG-DCT image, and inputting the JG-DCT image into a neural network to extract a first feature synthesis mapping vector; Step S3, inputting the multi-physical quantity time series data into the DNet network to obtain a time series feature vector, inputting the first feature synthesis mapping vector into the ENet network to obtain a second feature synthesis mapping vector, and concatenating the second feature synthesis mapping vector and the time series feature vector to obtain a Hybrid feature vector; Step S4, inputting the Hybrid feature vector into the FNet network to obtain a third feature synthesis mapping vector; Step S5, inputting the third feature synthesis mapping vector into the GNet network to obtain a fourth feature synthesis mapping vector; Step S6, selecting positive and negative samples for comparative learning and fine-tuning the weights of the DNet network, the ENet network, and the FNet network to obtain pre-trained network weights; Step S7, replacing the GNet network with a multi-layer perceptron model, and performing supervised training for cable defect recognition based on the pre-trained network weights and the fourth feature synthesis mapping vector, and outputting the cable defect type.

2. The cable defect identification method based on multiple physical quantities according to claim 1 is characterized in that: The step of converting the time series data into a JG-DCT image comprises the following steps: Extract a window sequence L(n) from the time series S(x) of multiple physical quantities according to the interval N, where 0 <n<N=w 2 , w represents the width of the generated image; The time series of multiple physical quantities are constructed into multiple G images through the following formula: ， ， Where P is the G image, (l,c) is the horizontal and vertical coordinate index value of the G image, min(L) and max(L) are the maximum and minimum values ​​of the window sequence respectively; The JG image is obtained by averaging multiple G image elements corresponding to multiple physical quantities: ， Among them, (x, y) is the horizontal and vertical coordinate index of the JG image, and m is the number of physical quantities.

3. The cable defect identification method based on multiple physical quantities according to claim 2 is characterized in that: The step of converting the time series data into a JG-DCT image further includes the following steps: After obtaining the JG image matrix, perform two-dimensional DCT transformation according to the following formula to obtain the JG-DCT image: ， ， Wherein, JG_D(u,v) represents the pixel on the JG-DCT image after the JG image is transformed by two-dimensional DCT, and C(u,v) is the conversion kernel function; The JG-DCT image is input into a neural network, and a first feature synthesis mapping vector is output, where the dimension of the first feature synthesis mapping vector is 1024.

4. The cable defect identification method based on multiple physical quantities according to claim 1 is characterized in that: The DNet network includes a first residual network module, a second residual network module, a third residual network module, a fourth residual network module, a first pooling module, a second pooling module, a third pooling module and a first fully connected layer. The first residual network module, the second residual network module and the first pooling module constitute a first mapping module, the third residual network module and the second pooling module constitute a second mapping module, and the fourth residual network module and the third pooling module constitute a third mapping module. The multi-physical quantity time series data is used as the input of the first mapping module of the DNet network, the output of the first mapping module is used as the input of the second mapping module, the output of the second mapping module is used as the input of the third mapping module, and the output of the third mapping module is used as the input of the first fully connected layer. The first fully connected layer outputs a time series feature vector with a dimension of 512.

5. The cable defect identification method based on multiple physical quantities according to claim 1 is characterized in that: The ENet network includes a second fully connected layer, the first feature synthesis mapping vector is used as an input of the second fully connected layer, and the second fully connected layer outputs a second feature synthesis mapping vector with a dimension of 512.

6. The cable defect identification method based on multiple physical quantities according to claim 1 is characterized in that: The dimension of the Hybrid feature vector is 1024, which is obtained by cross-joining the time series feature vector and the second feature synthesis mapping vector: ， Among them, c(2n) is the Hybrid eigenvector, is the time series feature vector, Synthesize the mapping vector for the second feature.

7. The cable defect identification method based on multiple physical quantities according to claim 1 is characterized in that: The FNet network includes a third fully connected layer, the Hybrid feature vector is used as the input of the third fully connected layer, and the third fully connected layer outputs a third feature synthesis mapping vector with a dimension of 512.

8. The cable defect identification method based on multiple physical quantities according to claim 1 is characterized in that: The GNet network includes a fourth fully connected layer, the third feature synthesis mapping vector is used as an input of the fourth fully connected layer, and the fourth fully connected layer outputs a fourth feature synthesis mapping vector with a dimension of 32.

9. The cable defect identification method based on multiple physical quantities according to any one of claims 1 to 8, characterized in that: The step S6 comprises the following steps: Select N positive and negative samples of the same category based on the collected multi-physical quantity time series data; The selected positive and negative samples are processed in sequence using the methods of step S2 to step S5 to obtain a fourth feature synthesis mapping vector; Construct the following loss function and perform contrastive learning: ， ， Among them, sim(*) is the feature similarity function , are the 2k-1th and 2kth fourth feature synthesis mapping vectors respectively, is the weight parameter; The weights of the DNet network, ENet network and FNet network are fine-tuned by the gradient descent method.

10. The cable defect identification method based on multiple physical quantities according to claim 9 is characterized in that: In the process of contrastive learning, when N positive samples, i.e., the fourth feature synthetic mapping vector generated by multi-physical quantity time series data of the same category, are selected for contrastive learning, minimizing the loss function is used as the basis for fine-tuning; When N negative samples, i.e., the fourth feature synthetic mapping vector generated by multi-physical quantity time series data of the same category, are selected for comparative learning, maximizing the loss function is used as the basis for fine-tuning.

11. A cable defect identification device based on multiple physical quantities, characterized in that: include: A data acquisition module is used to obtain time series data of multiple physical quantities of the cable, wherein the time series data includes temperature data, partial discharge pulse data and voiceprint data; A neural network module, used for converting the time series data into a JG-DCT image, and inputting the JG-DCT image into a neural network to extract a first feature synthesis mapping vector; The ENet network module is used to input the multi-physical quantity time series data into the DNet network to obtain a time series feature vector, input the first feature synthesis mapping vector into the ENet network to obtain a second feature synthesis mapping vector, and concatenate the second feature synthesis mapping vector and the time series feature vector to obtain a Hybrid feature vector; FNet network module, used to input the Hybrid feature vector into the FNet network to obtain the third feature synthesis mapping vector; A GNet network module, used for inputting the third feature synthesis mapping vector into the GNet network to obtain a fourth feature synthesis mapping vector; The weight fine-tuning module is used to select positive and negative samples for comparative learning and fine-tune the weights of the DNet network, ENet network, and FNet network to obtain the pre-trained network weights; The multi-layer perceptron module is used to replace the GNet network with a multi-layer perceptron model, and perform supervised training for cable defect recognition based on the pre-trained network weights and the fourth feature synthesis mapping vector, and output the cable defect type.

12. The cable defect identification device based on multiple physical quantities according to claim 11, characterized in that: The specific process of the neural network module converting time series data into JG-DCT images is as follows: Extract a window sequence L(n) from the time series S(x) of multiple physical quantities according to the interval N, where 0 <n<N=w 2 , w represents the width of the generated image; The time series of multiple physical quantities are constructed into multiple G images through the following formula: ， ， Where P is the G image, (l,c) is the horizontal and vertical coordinate index value of the G image, min(L) and max(L) are the maximum and minimum values ​​of the window sequence respectively; The JG image is obtained by averaging multiple G image elements corresponding to multiple physical quantities: ， Among them, (x, y) is the horizontal and vertical coordinate index of the JG image, and m is the number of physical quantities; After obtaining the JG image matrix, perform two-dimensional DCT transformation according to the following formula to obtain the JG-DCT image: ， ， Wherein, JG_D(u,v) represents the pixel on the JG-DCT image after the JG image is transformed by two-dimensional DCT, and C(u,v) is the conversion kernel function; The JG-DCT image is input into a neural network, and a first feature synthesis mapping vector is output, where the dimension of the first feature synthesis mapping vector is 1024.

13. The cable defect identification device based on multiple physical quantities according to claim 11, characterized in that: The DNet network includes a first residual network module, a second residual network module, a third residual network module, a fourth residual network module, a first pooling module, a second pooling module, a third pooling module and a first fully connected layer. The first residual network module, the second residual network module and the first pooling module constitute a first mapping module, the third residual network module and the second pooling module constitute a second mapping module, and the fourth residual network module and the third pooling module constitute a third mapping module. The multi-physical quantity time series data is used as the input of the first mapping module of the DNet network, the output of the first mapping module is used as the input of the second mapping module, the output of the second mapping module is used as the input of the third mapping module, and the output of the third mapping module is used as the input of the first fully connected layer. The first fully connected layer outputs a time series feature vector with a dimension of 512.

14. The cable defect identification device based on multiple physical quantities according to claim 11, characterized in that: The dimension of the Hybrid feature vector is 1024, which is obtained by cross-joining the time series feature vector and the second feature synthesis mapping vector: ， Among them, c(2n) is the Hybrid eigenvector, is the time series feature vector, Synthesize the mapping vector for the second feature.

15. The cable defect identification device based on multiple physical quantities according to claim 11, characterized in that: The specific process of the weight fine-tuning module obtaining the pre-trained network weights is as follows: Select N positive and negative samples of the same category based on the collected multi-physical quantity time series data; The selected positive and negative samples are processed in sequence using the methods of step S2 to step S5 to obtain a fourth feature synthesis mapping vector; Construct the following loss function and perform contrastive learning: ， ， Among them, sim(*) is the feature similarity function , are the 2k-1th and 2kth fourth feature synthesis mapping vectors respectively, is the weight parameter; The weights of the DNet network, ENet network and FNet network are fine-tuned by the gradient descent method.

16. The cable defect identification device based on multiple physical quantities according to claim 15, characterized in that: In the process of contrastive learning, when N positive samples, i.e., the fourth feature synthetic mapping vector generated by multi-physical quantity time series data of the same category, are selected for contrastive learning, minimizing the loss function is used as the basis for fine-tuning; When N negative samples, i.e., the fourth feature synthetic mapping vector generated by multi-physical quantity time series data of the same category, are selected for comparative learning, maximizing the loss function is used as the basis for fine-tuning.

Images