Power cable defect identification method and device based on incremental learning

Abstract

The invention provides a power cable defect identification method and device based on incremental learning. The method comprises the following steps: constructing partial discharge multi-parameter characteristic sample data of different defect types of power cables; the multi-parameter features corresponding to the partial discharge signals of the power cable under the laboratory condition and the defect types of the partial discharge signals serve as marked data to be used for training a CNN defect recognition model, and the multi-parameter features corresponding to the partial discharge signals of the power cable on the industrial site serve as unmarked data with unknown defect types to carry out a self-training process. A CNNL-based semi-supervised incremental learning defect identification model is obtained; and performing defect identification based on semi-supervised CNNL incremental learning, updating the CNNL-based semi-supervised incremental learning defect identification model, and determining a defect type corresponding to the power cable partial discharge signal of the industrial field, which is acquired in real time. Therefore, the defect type detection efficiency is high.

Description

Technical Field

[0001] The present invention belongs to the technical field of power cable monitoring, and in particular relates to a power cable defect identification method and device based on incremental learning. Background Art

[0002] With the development of power systems, power cables, as an important part of the power transmission and distribution network, are receiving more and more attention for their operational safety and reliability.

[0003] Partial discharge is an early sign of aging and insulation failure of power cables. Timely and effective detection and identification of partial discharge signals and defect types are crucial to preventing cable failures.

[0004] The existing methods for detecting defect types of aging power cables rely on manual experience analysis, and have problems such as high misjudgment rate and low detection efficiency, making it difficult to meet the needs of analyzing and evaluating the status of a large number of old cables. Summary of the invention

[0005] In view of this, the present invention proposes a power cable defect identification method and device based on incremental learning, aiming to solve the problem of high misjudgment rate or low detection efficiency in defect type detection work in the prior art.

[0006] In a first aspect, the present invention proposes a method for identifying power cable defects based on incremental learning, comprising:

[0007] Construct multi-parameter feature sample data of partial discharge of power cables with different defect types as input for network model training; wherein the multi-parameter feature sample data includes: multi-parameter features and defect types corresponding to partial discharge signals of power cables under laboratory conditions, and multi-parameter features corresponding to partial discharge signals of power cables at industrial sites;

[0008] The multi-parameter features and defect types corresponding to the partial discharge signals of power cables under laboratory conditions are used as labeled data for the training of the CNN defect recognition model. The multi-parameter features corresponding to the partial discharge signals of power cables at industrial sites are used as unlabeled data with unknown defect types for self-training, and a semi-supervised incremental learning defect recognition model based on CNNL is obtained.

[0009] The defect recognition model based on semi-supervised incremental learning of CNNL is used to perform defect recognition based on semi-supervised incremental learning of CNNL, update the defect recognition model based on semi-supervised incremental learning of CNNL, and determine the defect type corresponding to the partial discharge signal of the power cable collected in real time at the industrial site.

[0010] Furthermore, the construction of multi-parameter characteristic sample data of partial discharge of power cables with different defect types includes:

[0011] Create defective sections of power cables with different defect types;

[0012] Conduct partial discharge tests on power cables based on typical defects under laboratory conditions and collect partial discharge signals of different defect types under laboratory conditions;

[0013] A real power cable operation platform is built at the industrial site to collect partial discharge signals of power cables in the industrial site environment.

[0014] Furthermore, the construction of multi-parameter characteristic sample data of partial discharge of power cables with different defect types includes:

[0015] Based on the partial discharge signals collected under laboratory conditions and industrial sites, the signal denoising preprocessing is performed; the normalized phase spectrum is drawn and its characteristic parameters are extracted;

[0016] Annular cutting mark defect samples, grinding irregularity samples, stress cone displacement defect samples and metal spike defect samples are constructed respectively, and a multi-parameter feature sample library of partial discharge of power cables based on typical defects is constructed based on different types of defect samples.

[0017] Furthermore, the drawing of the normalized phase spectrum and extracting its characteristic parameters include:

[0018] Based on the drawn normalized phase map of the power cable based on typical defects, the high-dimensional features of the phase map are extracted through VGG, ResNET, SENet and FPNet respectively, and the high-dimensional features of the VGG map, the high-dimensional features of the ResNET map, the high-dimensional features of the SENet map and the high-dimensional features of the FPNet map are obtained;

[0019] Based on the normalized phase map of power cables with typical defects, the phase map image features are extracted, which are map color features, map shape features, map LBP features and map geometric features.

[0020] Furthermore, the construction of multi-parameter characteristic sample data of partial discharge of power cables with different defect types includes:

[0021] Based on the partial discharge signals collected under laboratory conditions and industrial sites, the signal denoising preprocessing is performed; the pulse rise time, pulse repetition rate, pulse skewness and pulse kurtosis are selected as the characteristic parameters of partial discharge pulses;

[0022] The high-dimensional features of the partial discharge phase map under laboratory conditions and in industrial sites, the image features of the partial discharge phase map and the characteristic parameters of the partial discharge pulse are fused respectively to obtain the fused multi-parameter feature sample data as the input for model training.

[0023] Furthermore, the semi-supervised incremental learning defect recognition model based on CNNL includes:

[0024] The characteristic parameters of power cables based on typical defects under laboratory conditions are used as labeled data to train a CNN-based power cable defect recognition model.

[0025] The characteristic parameters of power cables based on typical defects in industrial field environments are taken as unlabeled data, and 30% of the data are extracted with replacement as unlabeled data set 1 and sent to the trained CNN-based power cable defect recognition model to obtain pseudo labels 1 for the unlabeled data in industrial fields;

[0026] 30% of the industrial field data are extracted again with replacement as unlabeled data set 2 and fed into the trained CNN-based power cable defect recognition model. The incremental learner is fine-tuned based on pseudo label 1 to obtain pseudo label 2.

[0027] The fine-tuning incremental learner is firstly trained by replaying real data to train the CNN-based power cable defect recognition model by using data feature replay, and then replaying the generated data based on unlabeled data and pseudo labels generated by it, fine-tuning the existing CNN defect recognition network, generating a new CNN-based power cable defect recognition model, and obtaining new pseudo labels generated for unlabeled data;

[0028] The above steps are repeated until the CNN-based power cable defect recognition network reaches the expected convergence effect, the self-training process is completed, and the semi-supervised incremental learning defect recognition model based on CNNL is obtained.

[0029] Furthermore, the determining of the defect type corresponding to the partial discharge signal of the power cable collected in real time at the industrial site includes:

[0030] A semi-supervised incremental learning defect recognition model based on CNNL is used to perform defect recognition based on semi-supervised CNNL incremental learning, wherein the multi-parameter features corresponding to the partial discharge signals of power cables collected in real time at industrial sites are fused and used as unlabeled data. The self-training is performed in combination with the trained semi-supervised incremental learning defect recognition model based on CNNL to update the semi-supervised incremental learning defect recognition model based on CNNL, and the defect type corresponding to the partial discharge signals of power cables collected in real time at industrial sites is determined.

[0031] In a second aspect, the present invention provides a power cable defect recognition device based on incremental learning, which executes the power cable defect recognition method based on incremental learning as described in the first aspect, including:

[0032] A sample data construction module is used to construct multi-parameter feature sample data of partial discharge of power cables with different defect types as input for network model training; wherein the multi-parameter feature sample data includes: multi-parameter features and defect types corresponding to partial discharge signals of power cables under laboratory conditions, and multi-parameter features corresponding to partial discharge signals of power cables at industrial sites;

[0033] The defect recognition model generation module is used to use the multi-parameter features and defect types corresponding to the partial discharge signals of power cables under laboratory conditions as labeled data for the training of the CNN defect recognition model, and to use the multi-parameter features corresponding to the partial discharge signals of power cables at industrial sites as unlabeled data with unknown defect types for self-training, thereby obtaining a semi-supervised incremental learning defect recognition model based on CNNL.

[0034] The industrial site defect type recognition module is used to utilize the semi-supervised incremental learning defect recognition model based on CNNL, perform defect recognition based on semi-supervised CNNL incremental learning, update the semi-supervised incremental learning defect recognition model based on CNNL, and determine the defect type corresponding to the partial discharge signal of the power cable at the industrial site collected in real time.

[0035] In a third aspect, the present invention provides a terminal, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method described in the first aspect.

[0036] In a fourth aspect, the present invention provides a computer storage medium storing computer controllable instructions, wherein the computer controllable instructions are used to execute the method described in the first aspect.

[0037] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present invention. Moreover, the same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings:

[0039] Figure 1 A schematic diagram of a flow chart of a power cable defect identification method based on incremental learning according to an embodiment of the present invention;

[0040] Figure 2 A schematic diagram of the principle flow of a power cable defect identification method based on incremental learning according to an embodiment of the present invention;

[0041] Figure 3 A schematic diagram of the composition of a power cable defect recognition system to which the power cable defect recognition method based on incremental learning according to an embodiment of the present invention is applied;

[0042] Figure 4 Schematic diagram of the composition of the experimental system used to carry out power cable partial discharge signal acquisition tests under laboratory conditions;

[0043] Figure 5 A schematic diagram of the composition of the experimental system used to carry out power cable partial discharge signal acquisition tests under field operating conditions;

[0044] Figure 6 It is the PRPD spectrum of each voltage level (10kV, 16kV, 18kV and 22kV) under stress cone displacement defect under laboratory conditions;

[0045] Figure 7 PRPD spectra of various voltage levels (13kV, 16kV, 20kV and 22kV) under metal spike defects under laboratory conditions;

[0046] Figure 8 The PRPD spectra of various voltage levels (8kV, 12kV, 18kV and 24kV) under laboratory conditions for grinding irregular defects;

[0047] Fig. 9 It is the PRPD spectrum of each voltage level (12kV, 15kV, 20kV and 24kV) under the condition of ring cutting knife mark defect under laboratory conditions;

[0048] Fig.10 This is the variation trend of partial discharge with voltage level when grinding irregular defects under laboratory conditions;

[0049] Fig.11 This is the variation trend of partial discharge of ring cutting mark defect with voltage level under laboratory conditions;

[0050] Fig.12 This is the variation trend of partial discharge of metal spike defects with voltage level under laboratory conditions;

[0051] Fig.13 The variation trend of partial discharge of stress cone displacement defect with voltage level under laboratory conditions;

[0052] Fig.14 It is a schematic diagram of the composition of a power cable defect identification device based on incremental learning according to an embodiment of the present invention;

[0053] Fig.15 The present invention is a schematic diagram of the composition of a terminal to which the power cable defect identification method based on incremental learning according to an embodiment of the present invention is applied. DETAILED DESCRIPTION

[0054] The exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided in order to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art. It should be noted that the embodiments of the present invention and the features in the embodiments can be combined with each other without conflict.

[0055] For the convenience of the following description, the terms are explained as follows.

[0056] Incremental learning, Incremental Learning; Continual Learning; Life-Long Learning.

[0057] Continual Neural Network Learning, Continual Neural Network Learning, CNNL.

[0058] Convolutional Neural Network, ConvolutionalNeuralNetwork, CNN.

[0059] Generative Adversarial network, GAN.

[0060] ResNet (Residual Neural Network) used ResNet Unit to train a 152-layer neural network and won the ILSVRC2015 competition with an error rate of 3.57% in the top 5.

[0061] SENet (Squeeze-and-Excitation Networks) won the ImageClassification task in the ImageNet 2017 competition, reducing the top5 error rate to 2.251% on the ImageNet dataset.

[0062] Feature Pyramid Network, Feature PyramidNetwork, FPNet.

[0063] The phase-resolved partial discharge (PRPD) spectrum is used to describe the amplitude (such as pico-coulomb pC) and phase angle (such as 360 degrees) of the partial discharge signal, and can intuitively display the partial discharge activity within 360 degrees of an AC cycle.

[0064] Existing power cable defect recognition methods based on partial discharge signals have multiple shortcomings. First, traditional recognition methods rely on fixed training data sets. Once the recognition model is deployed, it cannot adapt to new partial discharge patterns or defect types that appear during cable operation, resulting in a decrease in recognition rate. Secondly, traditional methods require frequent retraining of the entire model, which not only increases the computational cost, but also easily leads to the loss of existing recognition capabilities and reduces the robustness of the system. Finally, when dealing with complex and diverse partial discharge signals, existing detection methods often find it difficult to balance real-time and accuracy, resulting in missed detection or misjudgment of partial discharge signals and defects.

[0065] The present invention proposes a technical solution for power cable defect recognition based on incremental learning, including a power cable defect recognition method based on incremental learning and a power cable defect recognition device based on incremental learning.

[0066] Incremental learning is also called continuous learning or lifelong learning. It allows the model trained in the existing data set to continuously learn new data knowledge, constantly adjust the network structure, and adapt to changes in various complex environments. It can show good generalization performance in industrial sites. The power cable defect recognition method and device based on incremental learning proposed in the present invention, by introducing incremental learning technology and using a semi-supervised incremental learning defect recognition model, can dynamically update the model (such as adjusting the network structure) when identifying new defects, avoiding the impact on existing recognition results, thereby improving the accuracy and real-time performance of recognition. At the same time, by optimizing the model structure and algorithm, the computing efficiency of the system can be improved, the operation and maintenance costs can be reduced, and the partial discharge signals of old cables can be more reliably monitored and analyzed in actual operation.

[0067] like Figure 1 As shown, the power cable defect identification method based on incremental learning in an embodiment of the present invention includes:

[0068] S100: constructing multi-parameter feature sample data of partial discharge of power cables with different defect types as input for network model training; wherein the multi-parameter feature sample data includes: multi-parameter features and defect types corresponding to partial discharge signals of power cables under laboratory conditions, and multi-parameter features corresponding to partial discharge signals of power cables at industrial sites;

[0069] S200: The multi-parameter features and defect types corresponding to the partial discharge signals of power cables under laboratory conditions are used as labeled data for training the CNN defect recognition model. The multi-parameter features corresponding to the partial discharge signals of power cables at industrial sites are used as unlabeled data with unknown defect types for self-training, and a semi-supervised incremental learning defect recognition model based on CNNL is obtained.

[0070] S300: Utilize the semi-supervised incremental learning defect recognition model based on CNNL to perform defect recognition based on semi-supervised incremental learning, update the semi-supervised incremental learning defect recognition model based on CNNL, and determine the defect type corresponding to the partial discharge signal of the power cable collected in real time at the industrial site.

[0071] Furthermore, the construction of multi-parameter characteristic sample data of partial discharge of power cables with different defect types includes:

[0072] Create defective sections of power cables with different defect types;

[0073] Conduct partial discharge tests on power cables based on typical defects under laboratory conditions and collect partial discharge signals of different defect types under laboratory conditions;

[0074] A real power cable operation platform is built at the industrial site to collect partial discharge signals of power cables in the industrial site environment.

[0075] Furthermore, the construction of multi-parameter characteristic sample data of partial discharge of power cables with different defect types includes:

[0076] Based on the partial discharge signals collected under laboratory conditions and industrial sites, the signal denoising preprocessing is performed; the normalized phase spectrum is drawn and its characteristic parameters are extracted;

[0077] Annular cutting mark defect samples, grinding irregularity samples, stress cone displacement defect samples and metal spike defect samples are constructed respectively, and a multi-parameter feature sample library of partial discharge of power cables based on typical defects is constructed based on different types of defect samples.

[0078] Furthermore, the drawing of the normalized phase spectrum and extracting its characteristic parameters include:

[0079] Based on the drawn normalized phase map of the power cable based on typical defects, the high-dimensional features of the phase map are extracted through VGG, ResNET, SENet and FPNet respectively, and the high-dimensional features of the VGG map, the high-dimensional features of the ResNET map, the high-dimensional features of the SENet map and the high-dimensional features of the FPNet map are obtained;

[0080] Based on the normalized phase map of power cables with typical defects, the phase map image features are extracted, which are map color features, map shape features, map LBP features and map geometric features.

[0081] Furthermore, the construction of multi-parameter characteristic sample data of partial discharge of power cables with different defect types includes:

[0082] Based on the partial discharge signals collected under laboratory conditions and industrial sites, the signal denoising preprocessing is performed; the pulse rise time, pulse repetition rate, pulse skewness and pulse kurtosis are selected as the characteristic parameters of partial discharge pulses;

[0083] The high-dimensional features of the partial discharge phase map under laboratory conditions and in industrial sites, the image features of the partial discharge phase map and the characteristic parameters of the partial discharge pulse are fused respectively to obtain the fused multi-parameter feature sample data as the input for model training.

[0084] Furthermore, the semi-supervised incremental learning defect recognition model based on CNNL includes:

[0085] The characteristic parameters of power cables based on typical defects under laboratory conditions are used as labeled data to train a CNN-based power cable defect recognition model.

[0086] The characteristic parameters of power cables based on typical defects in industrial field environments are taken as unlabeled data, and 30% of the data are extracted with replacement as unlabeled data set 1 and sent to the trained CNN-based power cable defect recognition model to obtain pseudo labels 1 for the unlabeled data in industrial fields;

[0087] 30% of the industrial field data are extracted again with replacement as unlabeled data set 2 and fed into the trained CNN-based power cable defect recognition model. The incremental learner is fine-tuned based on pseudo label 1 to obtain pseudo label 2.

[0088] The fine-tuning incremental learner is firstly trained by replaying real data to train the CNN-based power cable defect recognition model by using data feature replay, and then replaying the generated data based on unlabeled data and pseudo labels generated by it, fine-tuning the existing CNN defect recognition network, generating a new CNN-based power cable defect recognition model, and obtaining new pseudo labels generated for unlabeled data;

[0089] The above steps are repeated until the CNN-based power cable defect recognition network reaches the expected convergence effect, the self-training process is completed, and the semi-supervised incremental learning defect recognition model based on CNNL is obtained.

[0090] Furthermore, the determining of the defect type corresponding to the partial discharge signal of the power cable collected in real time at the industrial site includes:

[0091] A semi-supervised incremental learning defect recognition model based on CNNL is used to perform defect recognition based on semi-supervised CNNL incremental learning, wherein the multi-parameter features corresponding to the partial discharge signals of power cables collected in real time at industrial sites are fused and used as unlabeled data. The self-training is performed in combination with the trained semi-supervised incremental learning defect recognition model based on CNNL to update the semi-supervised incremental learning defect recognition model based on CNNL, and the defect type corresponding to the partial discharge signals of power cables collected in real time at industrial sites is determined.

[0092] like Figure 2 As shown, the power cable defect identification method based on incremental learning in an embodiment of the present invention includes the following steps:

[0093] S1: Make defective sections of power cables with different defect types and carry out partial discharge tests of typical defects of power cables in laboratory environment, collect partial discharge signals of power cables under laboratory conditions; collect partial discharge signals of power cables at industrial sites; draw their normalized phase maps and extract characteristic parameters; construct multi-parameter characteristic sample data of partial discharge of power cables with different defect types as input for network model training;

[0094] S2: The multi-parameter features and defect types corresponding to the partial discharge signals of power cables under laboratory conditions are used as labeled data for the training of the CNN defect recognition model. The multi-parameter features corresponding to the partial discharge signals of power cables at industrial sites are used as unlabeled data with unknown defect types for self-training and finally a semi-supervised incremental learning defect recognition model based on CNNL (continuous neural network learning) is obtained.

[0095] S3: Utilize the semi-supervised incremental learning defect recognition model based on CNNL to perform defect recognition based on semi-supervised incremental learning, update the semi-supervised incremental learning defect recognition model based on CNNL, and determine the defect type corresponding to the partial discharge signal of the power cable collected in real time at the industrial site.

[0096] Specifically, step S1 includes the following S101 to S107:

[0097] S101: Make defective sections of power cables with different defect types, and carry out partial discharge tests on power cables based on typical defects in a shielded laboratory (such as Figure 4 As shown), collecting partial discharge signals of different defect types under laboratory conditions;

[0098] S102: Build a real power cable operation platform at the industrial site (such as Figure 5 As shown), collect partial discharge signals of power cables in industrial field environments;

[0099] S103: Based on the partial discharge signals collected under laboratory conditions and industrial sites, perform signal denoising preprocessing; draw a normalized phase spectrum and extract its characteristic parameters; construct ring cutting mark defect samples, grinding irregularity samples, stress cone shift defect samples and metal spike defect samples respectively; based on different types of defect samples, construct a multi-parameter feature sample library of partial discharge of power cables based on typical defects.

[0100] S104: Based on the normalized phase map of the power cable based on typical defects drawn in step S103, high-dimensional features of the phase map are extracted through VGG, ResNET, SENet and FPNet respectively to obtain high-dimensional features of VGG map, high-dimensional features of ResNET map, high-dimensional features of SENet map and high-dimensional features of FPNet map;

[0101] S105: Based on the normalized phase map of the power cable based on typical defects drawn in step S103, phase map image features are extracted, which are map color features, map shape features, map LBP features and map geometric features;

[0102] S106: based on the normalized characteristic parameters of the power cable based on typical defects extracted in step S103, the pulse rise time, pulse repetition rate, pulse skewness and pulse kurtosis are selected as the partial discharge pulse characteristic parameters;

[0103] S107: respectively fuse the high-dimensional features of the partial discharge phase map under laboratory conditions and at the industrial site, the image features of the partial discharge phase map, and the characteristic parameters of the partial discharge pulse to obtain the fused multi-parameter features as the input of the model training in the following steps.

[0104] like Figure 4 As shown in the figure, when conducting a partial discharge signal acquisition test for power cables with known defects under laboratory conditions, the partial discharge signal acquisition unit installed at the power cable monitoring station includes:

[0105] Polypropylene cable defective section (cable section with known defects set artificially), HFCT (high frequency current transformer), high-pass filter, acquisition card, PC or Laptop-based data acquisition system, high-voltage control test bench;

[0106] The high voltage control test bench is equipped with a desktop electrical display device, a current transformer, a transformer, a protective resistor, a capacitor voltage divider, and a low voltage matching unit; the AC power supply forms a loop through the transformer, the protective resistor, the capacitor voltage divider, the low voltage matching unit, and the desktop electrical display device;

[0107] The current transformer is used to detect the power supply current and is connected to a desktop electrical display device to display the power supply current in real time;

[0108] The high voltage is led out through the capacitive voltage divider and applied to the defective section of the polypropylene cable;

[0109] The HFCT installed on the power cable with the defective section of polypropylene cable collects the partial discharge signal in real time;

[0110] The current signal collected by HFCT in real time is transmitted to the data acquisition system through a high-speed communication interface (such as an acquisition card with a PCI or PCIe or USB interface) after passing through a high-pass filter;

[0111] The data acquisition system is provided with a signal processing module, which pre-processes the collected partial discharge signals, including denoising, feature extraction, etc. Specifically, the adaptive filtering technology is used to improve the signal processing efficiency on the one hand, and ensure the high fidelity of the data on the other hand.

[0112] In the above, the data acquisition system collects partial discharge signals in real time through the HFCT installed on the power cable. The collected data is transmitted to the signal processing module through a high-speed communication interface.

[0113] like Figure 5 As shown in the figure, when conducting a partial discharge signal acquisition test on an old power cable with unknown defects under field operation conditions, the partial discharge signal acquisition unit installed at the power cable monitoring site includes:

[0114] Polypropylene test cable segment (on-site running cable, no artificial setting of known defects), temperature sensor, high-frequency current transformer, multi-parameter monitoring system; the multi-parameter monitoring system includes high-pass filter, acquisition card, PC or Laptop-based data acquisition system;

[0115] Control console, transformer, protective resistor, voltage divider, short-circuit copper bus, through-core induction transformer;

[0116] The control console is equipped with a desktop electrical display device, a current transformer, and a low-voltage matching unit. The AC power supply forms a loop through the transformer, a protective resistor, a capacitor voltage divider, a short-circuited copper bus, a through-core induction transformer, and a polypropylene test cable segment, and applies voltage and current to the polypropylene test cable segment;

[0117] The high-frequency current transformer installed in the polypropylene test cable section collects the partial discharge signal in real time;

[0118] The temperature sensor installed in the polypropylene test cable section collects the temperature signal on the cable surface or in the environment in real time;

[0119] The real-time collected current signal or temperature signal is transmitted to the data acquisition system through a high-speed communication interface (such as an acquisition card with a PCI or PCIe or USB interface) after passing through a high-pass filter;

[0120] The data acquisition system is provided with a signal processing module, which pre-processes the collected partial discharge signals, including denoising, feature extraction, etc. Specifically, the adaptive filtering technology is used to improve the signal processing efficiency on the one hand, and ensure the high fidelity of the data on the other hand.

[0121] In the above, the data acquisition system collects partial discharge signals in real time through the HFCT installed on the power cable. The collected data is transmitted to the signal processing module through a high-speed communication interface.

[0122] above Figure 4 and Figure 5 In the partial discharge test, the test objects involved are different, and the experimental environment and the operating environment of the acquisition equipment are different. Therefore, the samples obtained under laboratory conditions and industrial sites are not only different in quantity, but also in the noise or interference in the partial discharge signal. Specifically, Figure 4 In the test, voltage is applied to the defective cable. Figure 5 In this test, voltage and current are applied simultaneously to the actual operating cable (without artificial defects) for testing. Figure 4 Only voltage is applied so the cable does not heat up; Figure 5 Current is also added, so the cable will heat up, so a temperature sensor is set up and the multi-parameter monitoring system measures the temperature of the cable.

[0123] Specifically, the normalized PRPD spectra of partial discharge monitored when the power cable has different types of defects are as follows: Figures 6 to 9 As shown, Figure 6 It is the PRPD spectrum of partial discharge of power cables under stress cone displacement defects under laboratory conditions at voltage levels of 10kV, 16kV, 18kV and 22kV; Figure 7 It is the PRPD spectrum of partial discharge of power cables at voltage levels of 13kV, 16kV, 20kV and 22kV under laboratory conditions with metal spike defects; Figure 8 It is the PRPD spectrum of partial discharge of power cables at voltage levels of 8kV, 12kV, 18kV and 24kV under laboratory conditions when grinding irregular defects; Fig. 9 It is the PRPD spectrum of partial discharge of power cables at voltage levels of 12kV, 15kV, 20kV and 24kV under laboratory conditions with ring cutting marks. Figures 6 to 9 The normalized PRPD atlases shown can respectively extract VGG atlas high-dimensional features, ResNET atlas high-dimensional features, SENet atlas high-dimensional features, FPNet atlas high-dimensional features; atlas color features, atlas shape features, atlas LBP features, and atlas geometric features.

[0124] The pulse parameter characteristics of the partial discharge signal extracted when the power cable has different types of defects are as follows: Figures 10 to 13 shown. Fig.10 This is the variation trend of the partial discharge signal with the voltage level when there are irregular grinding defects in the cable segment under laboratory conditions. The physical meaning of the left vertical axis is the discharge amount (in pC), and the physical meaning of the right vertical axis is the pulse repetition rate (in pieces / second). Fig.11 This is the variation trend of the partial discharge signal with the voltage level when there is a ring cutting mark defect in the cable segment under laboratory conditions, where the physical meaning of the left vertical axis is the discharge amount (in pC), and the physical meaning of the right vertical axis is the pulse repetition rate (in pieces / second); Fig.12 This is the variation trend of partial discharge signal with voltage level when there is a metal spike defect in the cable segment under laboratory conditions, where the physical meaning of the left vertical axis is the discharge amount (in pC), and the physical meaning of the right vertical axis is the pulse repetition rate (in pieces / second); Fig.13 This is the variation trend of partial discharge signal with voltage level when stress cone displacement defect exists in cable segment under laboratory conditions. The physical meaning of the left vertical axis is discharge amount (in pC), and the physical meaning of the right vertical axis is pulse repetition rate (in pieces / second). 。 in this way, Figures 10 to 13 The pulse parameter characteristics of the partial discharge signals of each power cable displayed include: discharge amount and pulse repetition rate.

[0125] Specifically, step S2 includes the following S201 to S205:

[0126] S201: Using the characteristic parameters of power cables based on typical defects under laboratory conditions as labeled data, a CNN-based power cable defect recognition model is trained;

[0127] S202: Taking the characteristic parameters of the power cable based on typical defects in the industrial field environment as unlabeled data, extracting 30% of the data with replacement as the unlabeled data set 1 and sending it to the trained CNN-based power cable defect recognition model described in step S201 to obtain the pseudo label 1 of the unlabeled data of the industrial field;

[0128] S203: extract 30% of the industrial field data again with replacement as the unlabeled data set 2 and input it into the trained CNN-based power cable defect recognition model described in step S201, and fine-tune the incremental learner based on the pseudo label 1 to obtain the pseudo label 2;

[0129] S204: For the fine-tuning incremental learner described in step S202 and step S203, by using data feature playback, firstly, the CNN-based power cable defect recognition model is trained by real data playback, and then the generated data is played back based on the unlabeled data and the pseudo labels generated therefrom, and the existing CNN defect recognition network is fine-tuned to generate a new CNN-based power cable defect recognition model, and obtain new pseudo labels generated for the unlabeled data;

[0130] S205: The above steps are repeated until the CNN-based power cable defect recognition network reaches the expected convergence effect, completes the self-training process, and finally obtains a semi-supervised incremental learning defect recognition model based on CNNL (continuous neural network learning).

[0131] Specifically, step S3 includes: using a semi-supervised incremental learning defect recognition model based on CNNL to perform defect recognition based on semi-supervised CNNL incremental learning, wherein the multi-parameter features corresponding to the partial discharge signals of power cables at industrial sites collected in real time are fused and used as unlabeled data, combined with the trained semi-supervised incremental learning defect recognition model based on CNNL for self-training, updating the semi-supervised incremental learning defect recognition model based on CNNL, and determining the defect type corresponding to the partial discharge signals of power cables at industrial sites collected in real time.

[0132] like Figure 2 As shown in the figure, the power cable defect recognition method based on incremental learning is divided into two stages. The first stage is data collection and construction of typical defect samples; the second stage is defect recognition based on semi-supervised CNNL incremental learning.

[0133] Specifically, in the first phase, data collection and construction of typical defect samples include:

[0134] Carry out partial discharge signal acquisition of power cables with typical defects under laboratory conditions;

[0135] Carry out partial discharge signal collection of power cables with unknown defect types at industrial sites;

[0136] The data preprocessing is performed on the partial discharge signals of power cables collected by the above two methods respectively;

[0137] respectively determining the normalized phase spectra of the partial discharge signals of each power cable after data preprocessing;

[0138] The ring cutting mark defect, grinding irregularity, stress cone defect and metal spike defect are used as sample labels to extract the multi-parameter features of the partial discharge signal of each power cable after data preprocessing.

[0139] The multi-parameter features include: VGG spectrum high-dimensional features, ResNET spectrum high-dimensional features, SENet spectrum high-dimensional features, FPNet spectrum high-dimensional features; spectrum color features, spectrum shape features, spectrum LBP features, spectrum geometry features; pulse rise time, pulse repetition rate, pulse skewness features, pulse kurtosis features.

[0140] The multi-parameter features corresponding to the partial discharge signals of power cables collected under laboratory conditions are fused;

[0141] The multi-parameter features corresponding to the partial discharge signals of power cables collected at industrial sites are fused.

[0142] In the second stage, defect recognition is performed based on the semi-supervised incremental learning defect recognition model of CNNL, including:

[0143] S301: using the multi-parameter features corresponding to the partial discharge signals of power cables collected under laboratory conditions and the corresponding defect types as labeled data to train a CNN-based defect recognition model;

[0144] S302: The multi-parameter features corresponding to the partial discharge signals of power cables collected at the industrial site are used as unlabeled data, and are combined with the trained CNN-based defect recognition model for self-training to generate a semi-supervised incremental learning defect recognition model based on CNNL.

[0145] Specifically, a self-training is performed in combination with a trained CNN-based defect recognition model to generate a CNNL-based semi-supervised incremental learning defect recognition model, including: taking characteristic parameters of power cables based on typical defects in an industrial field environment as unlabeled data, extracting 30% of the data with replacement as an unlabeled data set 1 and sending it to the trained CNN-based power cable defect recognition model described in step S301 to obtain a pseudo label 1 for the unlabeled data of the industrial field;

[0146] S303: extract 30% of the industrial field data again with replacement as the unlabeled data set 2 and input it into the trained CNN-based power cable defect recognition model described in step S201, and fine-tune the incremental learner based on the pseudo label 1 to obtain the pseudo label 2;

[0147] S304: For the fine-tuning incremental learner described in step S302 and step S303, by using data feature playback, firstly, the CNN-based power cable defect recognition network model is trained by real data playback, and then the generated data is played back based on the unlabeled data and the pseudo labels generated therefrom, the existing CNN defect recognition network is fine-tuned, a new CNN-based power cable defect recognition model is generated, and new pseudo labels generated for the unlabeled data are obtained;

[0148] S305: The above steps are repeated until the CNN-based power cable defect recognition network reaches the expected convergence effect, completes the self-training process, and finally obtains a semi-supervised incremental learning defect recognition model based on CNNL (continuous neural network learning).

[0149] Specifically, a semi-supervised incremental learning defect recognition model based on CNNL is used to perform defect recognition based on semi-supervised CNNL incremental learning, wherein the multi-parameter features corresponding to the partial discharge signals of power cables at industrial sites collected in real time are fused and used as unlabeled data. The self-training is carried out in combination with the trained semi-supervised incremental learning defect recognition model based on CNNL, and the semi-supervised incremental learning defect recognition model based on CNNL is updated to determine the defect type corresponding to the partial discharge signals of power cables at industrial sites collected in real time.

[0150] Specifically, a semi-supervised incremental learning defect recognition model based on CNNL is used to perform defect recognition based on semi-supervised CNNL incremental learning, wherein the multi-parameter features corresponding to the partial discharge signals of power cables at industrial sites collected in real time are fused and used as unlabeled data. The self-training is carried out in combination with the trained semi-supervised incremental learning defect recognition model based on CNNL, and the semi-supervised incremental learning defect recognition model based on CNNL is updated to determine the defect type corresponding to the partial discharge signals of power cables at industrial sites collected in real time.

[0151] Naturally, based on the process of obtaining the CNNL semi-supervised incremental learning defect recognition model and the obtained CNNL semi-supervised incremental learning defect recognition model, a power cable defect recognition system based on incremental learning can be further constructed.

[0152] In some embodiments, Figure 3 As shown, the power cable defect identification system based on incremental learning includes: a display unit 10, a partial discharge signal acquisition unit 20 and a data server unit 30.

[0153] Typically, the partial discharge signal acquisition unit 20 is disposed at each power cable monitoring site; the data server unit 30 is disposed in the cloud or remotely relative to each partial discharge signal acquisition unit 20, and the physical distance between the two may be thousands of kilometers.

[0154] The display unit 10 includes: a human-computer interaction channel 11, which is used to control the switch, function selection, and page selection of the entire device; a defect recognition result display module 12 based on incremental learning, which is used to display the defect recognition results or discharge pattern recognition results after real-time monitoring and analysis; a data integration display module 13, which is used to display the local discharge signal (such as a pulse waveform) collected in real time and the cable segment to which it belongs, such as displaying the geographic information of the power cable where the cable segment is located based on integrated GIS data.

[0155] The partial discharge signal acquisition unit 20 includes a cable status monitoring module 21 and a communication module 22. The cable status monitoring module 21 acquires partial discharge signals of each cable in real time and transmits them to the data server unit 30 through the communication module for data analysis, partial discharge pattern recognition or defect recognition.

[0156] like Figure 3 As shown, the cable status monitoring module includes a data acquisition and processing unit 1, a sensor device 2 and a cable segment 3; the communication module 22 includes a plurality of data transceivers, and the data transceivers are provided with a 5G communication chip. The power supply from the power cabinet supplies power to the data acquisition and processing unit 1.

[0157] The data server unit 30 includes a data interface 31, a partial discharge database 32, and a power cable defect recognition module 33 based on incremental learning. The power cable defect recognition module 33 based on incremental learning obtains the partial discharge signals of each cable collected in real time by the cable status monitoring module from the data interface 31, extracts database data from the partial discharge database 32, and processes the partial discharge signals of each cable in combination with the database data to generate a defect recognition result for the partial discharge signal of each cable, that is, a defect type, and the defect recognition result includes at least one of the following: circumferential cut defect, irregular grinding, stress cone displacement defect, and metal spike defect; the defect recognition result for the partial discharge signal of each cable is transmitted back to the display unit 10, and accordingly, the defect recognition result display module 12 displays the defect recognition result for the partial discharge signal of each cable, and the data integration display module 13 displays the waveform of the partial discharge signal of each cable and the cable segment to which it belongs.

[0158] Specifically, the partial discharge database 32 includes a typical defect discharge database, a partial discharge feature database, an equipment information database, and an offline detection database. The partial discharge database updates the received partial discharge signals in real time, such as storing the multi-parameter features corresponding to each received partial discharge signal in the partial discharge feature database as a data set for incremental learning defect recognition model training.

[0159] Thus, the power cable defect recognition system based on incremental learning includes a display unit, a partial discharge signal acquisition unit and a data server unit. The functions of the display unit include: (1) a human-computer interaction channel, which is used to control the switch, function selection, page selection and other functions of the entire device; (2) a defect recognition result display module based on incremental learning, which displays the defect type recognition results after real-time monitoring and analysis; (3) an integrated display of data, which is used to display the partial discharge signal collected in real time and the cable segment to which it belongs.

[0160] Thus, the partial discharge signal acquisition unit further includes a communication module, which transmits the cable status information acquired in real time to the data server unit through the communication part for data analysis and pattern recognition.

[0161] The data server unit includes a data interface 31, a partial discharge database 32 and a defect recognition module 33 based on incremental learning. The data interface receives the real-time data collected by the partial discharge signal acquisition unit for power cable defect recognition based on incremental learning, and transmits the results back to the display unit for result display. The partial discharge database updates the received partial discharge signal in real time and stores it in the database, which is used as a data set for incremental learning defect recognition model training. The defect recognition module 33 based on incremental learning adopts an incremental learning algorithm based on a neural network. The algorithm can dynamically update the model and maintain the stability of existing knowledge when introducing new defects; the incremental learning defect recognition model adopts a structure based on online learning, and only part of the weights need to be updated each time new data is recognized, avoiding the waste of computing resources caused by full retraining. The defect recognition module is used to perform real-time analysis and pattern recognition on the processed signal through the incremental learning defect recognition model, identify the possible defect type or fault of the cable, and generate an alarm signal.

[0162] The incremental learning-based defect recognition module 33 adopts an incremental learning algorithm based on a neural network. The algorithm can dynamically update the model and maintain the stability of existing knowledge when introducing new defects; the incremental learning defect recognition model adopts a structure based on online learning, and only needs to update some weights each time new data is recognized, avoiding the waste of computing resources caused by full retraining. The pattern recognition module is used to perform real-time analysis and pattern recognition on the processed signal through the incremental learning defect recognition model, identify the possible defect type or fault of the cable, and generate an alarm signal.

[0163] The identification system also provides a statistical analysis function based on historical data to help technicians conduct more in-depth fault diagnosis. The identification system also provides a monitoring and alarm module, which displays the cable operation status in real time through the display interface. When an abnormal defect type is detected, the system will notify the maintenance personnel through sound and light alarms, text messages and emails, and automatically generate a diagnostic report.

[0164] In summary, the power cable defect identification method based on incremental learning of the present invention has shown significant technical effects in practical applications and has the following advantages and improvements:

[0165] 1) Improved recognition accuracy. By adopting an incremental learning algorithm, it is possible to dynamically adapt to changes in new defect types. The existing fixed training set method often significantly reduces the recognition rate when faced with diverse and changing partial discharge signals. The embodiment of the present invention avoids the forgetting effect by gradually updating the model, ensuring continuous improvement in recognition accuracy.

[0166] 2) Enhanced real-time performance. An incremental model update strategy is adopted to avoid frequent full retraining processes, significantly reducing the time and computing resources required for model updates. Tests show that under the same hardware conditions, the present invention can reduce the model update time to 1 / 3 of the traditional method, ensuring the real-time response of the system.

[0167] 3) Improved system robustness. When introducing new data for training, traditional methods often affect the existing recognition capabilities, resulting in reduced system robustness. The embodiments of the present invention effectively maintain the integrity and accuracy of historical knowledge through an optimized incremental learning algorithm, and enhance adaptability to complex cable environments. Laboratory conditions and industrial field experiments have verified that when faced with emergencies in cable operation, the defect recognition accuracy and stability of the present invention are better than those of traditional methods.

[0168] 4) Reduced operation and maintenance costs. The embodiments of the present invention reduce the operation and maintenance costs of the system by reducing the frequency of model retraining and optimizing computing resources. In addition, the recognition system structure design of the embodiments of the present invention is simple, easy to integrate and deploy, and can effectively shorten the equipment installation and debugging time. In practical applications, the operating cost of the system is reduced by about 25% compared with traditional methods, which greatly improves the user experience.

[0169] 5) Energy saving and environmental protection. When processing complex partial discharge signals, the embodiments of the present invention reduce energy consumption and resource waste through efficient algorithm optimization. At the same time, since the system has strong adaptability and reliability, it reduces energy consumption and material waste caused by frequent maintenance, which is beneficial to environmental protection.

[0170] In this way, it has demonstrated significant advantages in recognition accuracy, real-time performance, system robustness, operation and maintenance costs, energy conservation and environmental protection, etc. It can effectively solve the shortcomings of existing technologies and has broad application prospects.

[0171] In one embodiment of the present invention, the aforementioned semi-supervised incremental learning defect recognition model based on CNNL is used to build an old cable defect recognition system for old cable application scenarios.

[0172] Specifically, for the application scenario of old cables, a more adaptable incremental learning defect recognition model is provided, which can handle the changes in signal characteristics caused by cable aging.

[0173] Specifically, the incremental learning defect recognition model imports initial data, including: firstly initializing the incremental learning defect recognition model through the historical maintenance data of the old cables and the cable insulation characteristic data to adapt to the characteristics of the old cables.

[0174] Specifically, the incremental learning defect recognition model performs adaptive learning, including: as the old cables monitored in real time further age, the incremental learning defect recognition model automatically adjusts the defect recognition strategy; when the intensity and frequency of the partial discharge signal of the old cables monitored in real time change significantly, the incremental learning defect recognition model is adaptively updated. In this way, the reliability of defect recognition can be ensured.

[0175] Specifically, in an application in an old substation, various complex defect types caused by cable aging were identified. Experimental results showed that the overall recognition accuracy rate reached more than 95%, greatly improving the early warning capability of old cable failures.

[0176] The technical solution provided by the present invention is based on a power cable defect recognition method based on incremental learning, and adopts the following technical means:

[0177] (1) Application and optimization of incremental learning algorithms:

[0178] Unlike the existing full retraining model, this optimized incremental learning algorithm can dynamically update the defect recognition model when new data arrives, avoiding the loss of existing knowledge and ensuring recognition accuracy and stability.

[0179] The optimized incremental learning algorithm includes: an online learning method based on a neural network and a model weight updating strategy.

[0180] (2) Real-time signal processing and dynamic model updating:

[0181] An efficient signal processing module is integrated into the system, which can process the collected partial discharge signals in real time and dynamically update the pattern recognition model based on the processing results. This signal processing module optimizes the signal denoising and feature extraction process, ensuring high signal fidelity.

[0182] Furthermore, the signal processing module and the defect recognition module 33 based on incremental learning have a collaborative working mechanism, and the data processed by the signal processing module drives the pattern recognition model to be dynamically updated, and can optimize the balance between real-time performance and computational efficiency.

[0183] (3) Adaptive model update mechanism:

[0184] The system can adaptively adjust the pattern recognition model according to the changes in partial discharge signals monitored during cable operation. It is particularly suitable for dealing with new defect types and variant forms of known defect types, thereby improving the adaptability and robustness of the system.

[0185] Specifically, the adaptive model update mechanism includes adjusting model parameters under data guidance and dynamic learning strategies to cope with signal changes.

[0186] (4) Adaptability to old cables:

[0187] The system is optimized for the special partial discharge signals that may appear in old cables, ensuring that the defect type can still be effectively identified under cable aging conditions and providing early warning. Adaptive design of incremental learning defect recognition model for old cables and recognition strategy when signal characteristics change.

[0188] In some embodiments, Fig.14 As shown, the power cable defect recognition device based on incremental learning according to the embodiment of the present invention executes the aforementioned power cable defect recognition method based on incremental learning, including:

[0189] The sample data construction module 1000 is used to construct multi-parameter feature sample data of partial discharge of power cables with different defect types as input for network model training; wherein the multi-parameter feature sample data includes: multi-parameter features and defect types corresponding to partial discharge signals of power cables under laboratory conditions, and multi-parameter features corresponding to partial discharge signals of power cables at industrial sites;

[0190] The defect recognition model generation module 2000 is used to use the multi-parameter features and defect types corresponding to the partial discharge signals of power cables under laboratory conditions as labeled data for training the CNN defect recognition model, and use the multi-parameter features corresponding to the partial discharge signals of power cables at industrial sites as unlabeled data with unknown defect types for self-training to obtain a semi-supervised incremental learning defect recognition model based on CNNL;

[0191] The industrial site defect type identification module 3000 is used to utilize the semi-supervised incremental learning defect identification model based on CNNL, perform defect identification based on semi-supervised CNNL incremental learning, update the semi-supervised incremental learning defect identification model based on CNNL, and determine the defect type corresponding to the partial discharge signal of the power cable collected in real time at the industrial site.

[0192] The embodiment of the present invention also provides a terminal to execute the method. Fig.15 It shows a schematic diagram of a terminal provided by some embodiments of the present invention. Fig.15 As shown, the terminal 8 includes: a processor 800, a memory 801, a bus 802 and a communication interface 803, wherein the processor 800, the communication interface 803 and the memory 801 are connected via the bus 802; the memory 801 stores a computer program that can be run on the processor 800, and the processor 800 executes the method provided by any of the aforementioned embodiments of the present invention when running the computer program.

[0193] The memory 801 may include a high-speed random access memory (RAM), and may also include a non-volatile memory (Non-volatile Memory), such as at least one disk storage. The communication connection between the device network element and at least one other network element is realized through at least one communication interface 803 (which may be wired or wireless), and the Internet, wide area network, local area network, metropolitan area network, etc. may be used.

[0194] The bus 802 may be an ISA bus, a PCI bus, or an EISA bus, etc. The bus may be divided into an address bus, a data bus, a control bus, etc. The memory 801 is used to store programs, and the processor 800 executes the programs after receiving the control instructions. The method disclosed in any implementation of the above embodiments of the present invention may be applied to the processor 800, or implemented by the processor 800.

[0195] The processor 800 may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the method can be completed by the hardware integrated logic circuit in the processor 800 or the instruction in the form of software. The processor 800 may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it may also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. The methods, steps and logic block diagrams disclosed in the embodiments of the present invention can be implemented or executed. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc. The steps of the method disclosed in the embodiments of the present invention may be directly embodied as being executed by a hardware decoding processor, or executed by a combination of hardware and software modules in a decoding processor. The software module may be located in a mature storage medium in the field such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory 801, and the processor 800 reads the information in the memory 801 and completes the steps of the method in combination with its hardware.

[0196] The terminal and the method are based on the same inventive concept and have the same beneficial effects as the method adopted, operated or implemented by them.

[0197] An embodiment of the present invention also provides a computer-readable storage medium corresponding to the method provided in the aforementioned embodiment. The computer-readable storage medium is a CD on which a computer program (i.e., a program product) is stored. When the computer program is run by a processor, it will execute the method provided in any of the aforementioned embodiments.

[0198] It should be noted that examples of the computer-readable storage medium may also include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other optical or magnetic storage media, which are not listed here one by one.

[0199] The computer-readable storage medium provided by the embodiment of the present invention is based on the same inventive concept as the present method and has the same beneficial effects as the method adopted, run or implemented by the application program stored therein.

[0200] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these adjustments and modifications of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these changes and modifications.

Claims

1. A power cable defect identification method based on incremental learning, characterized in that: include: Construct multi-parameter feature sample data of partial discharge of power cables with different defect types as input for network model training; wherein the multi-parameter feature sample data includes: multi-parameter features and defect types corresponding to partial discharge signals of power cables under laboratory conditions, and multi-parameter features corresponding to partial discharge signals of power cables at industrial sites; The multi-parameter features and defect types corresponding to the partial discharge signals of power cables under laboratory conditions are used as labeled data for the training of the CNN defect recognition model. The multi-parameter features corresponding to the partial discharge signals of power cables at industrial sites are used as unlabeled data with unknown defect types for self-training, and a semi-supervised incremental learning defect recognition model based on CNNL is obtained. The defect recognition model based on semi-supervised incremental learning of CNNL is used to perform defect recognition based on semi-supervised incremental learning of CNNL, update the defect recognition model based on semi-supervised incremental learning of CNNL, and determine the defect type corresponding to the partial discharge signal of the power cable collected in real time at the industrial site.

2. The power cable defect identification method based on incremental learning according to claim 1, characterized in that: The method of constructing multi-parameter characteristic sample data of partial discharge of power cables with different defect types includes: Create defective sections of power cables with different defect types; Conduct partial discharge tests on power cables based on typical defects under laboratory conditions and collect partial discharge signals of different defect types under laboratory conditions; A real power cable operation platform is built at the industrial site to collect partial discharge signals of power cables in the industrial site environment.

3. The power cable defect identification method based on incremental learning according to claim 2, characterized in that: The method of constructing multi-parameter characteristic sample data of partial discharge of power cables with different defect types includes: Based on the partial discharge signals collected under laboratory conditions and industrial sites, the signal denoising preprocessing is performed; the normalized phase spectrum is drawn and its characteristic parameters are extracted; Annular cutting mark defect samples, grinding irregularity samples, stress cone displacement defect samples and metal spike defect samples are constructed respectively, and a multi-parameter feature sample library of partial discharge of power cables based on typical defects is constructed based on different types of defect samples.

4. The power cable defect identification method based on incremental learning according to claim 3, characterized in that: The drawing of the normalized phase spectrum and extraction of its characteristic parameters include: Based on the drawn normalized phase map of the power cable based on typical defects, the high-dimensional features of the phase map are extracted through VGG, ResNET, SENet and FPNet respectively, and the high-dimensional features of the VGG map, the high-dimensional features of the ResNET map, the high-dimensional features of the SENet map and the high-dimensional features of the FPNet map are obtained; Based on the normalized phase map of power cables with typical defects, the phase map image features are extracted, which are map color features, map shape features, map LBP features and map geometric features.

5. The power cable defect identification method based on incremental learning according to claim 4, characterized in that: The method of constructing multi-parameter characteristic sample data of partial discharge of power cables with different defect types includes: Based on the partial discharge signals collected under laboratory conditions and industrial sites, the signal denoising preprocessing is performed; the pulse rise time, pulse repetition rate, pulse skewness and pulse kurtosis are selected as the characteristic parameters of partial discharge pulses; The high-dimensional features of the partial discharge phase map under laboratory conditions and in industrial sites, the image features of the partial discharge phase map and the characteristic parameters of the partial discharge pulse are fused respectively to obtain the fused multi-parameter feature sample data as the input for model training.

6. The power cable defect identification method based on incremental learning according to claim 1, characterized in that: The obtained semi-supervised incremental learning defect recognition model based on CNNL includes: The characteristic parameters of power cables based on typical defects under laboratory conditions are used as labeled data to train a CNN-based power cable defect recognition model. The characteristic parameters of power cables based on typical defects in industrial field environments are taken as unlabeled data, and 30% of the data are extracted with replacement as unlabeled data set 1 and sent to the trained CNN-based power cable defect recognition model to obtain pseudo labels 1 for the unlabeled data in industrial fields; 30% of the industrial field data are extracted again with replacement as unlabeled data set 2 and fed into the trained CNN-based power cable defect recognition model. The incremental learner is fine-tuned based on pseudo label 1 to obtain pseudo label 2. The fine-tuning incremental learner is firstly trained by replaying real data to train the CNN-based power cable defect recognition model by using data feature replay, and then replaying the generated data based on unlabeled data and pseudo labels generated by it, fine-tuning the existing CNN defect recognition network, generating a new CNN-based power cable defect recognition model, and obtaining new pseudo labels generated for unlabeled data; The above steps are repeated until the CNN-based power cable defect recognition network reaches the expected convergence effect, the self-training process is completed, and the semi-supervised incremental learning defect recognition model based on CNNL is obtained.

7. The power cable defect identification method based on incremental learning according to claim 1, characterized in that: The step of determining the defect type corresponding to the partial discharge signal of the power cable collected in real time at the industrial site includes: A semi-supervised incremental learning defect recognition model based on CNNL is used to perform defect recognition based on semi-supervised CNNL incremental learning, wherein the multi-parameter features corresponding to the partial discharge signals of power cables collected in real time at industrial sites are fused and used as unlabeled data. The self-training is performed in combination with the trained semi-supervised incremental learning defect recognition model based on CNNL to update the semi-supervised incremental learning defect recognition model based on CNNL, and the defect type corresponding to the partial discharge signals of power cables collected in real time at industrial sites is determined.

8. A power cable defect identification device based on incremental learning, characterized in that: The method for identifying power cable defects based on incremental learning as claimed in claim 1 comprises: A sample data construction module is used to construct multi-parameter feature sample data of partial discharge of power cables with different defect types as input for network model training; wherein the multi-parameter feature sample data includes: multi-parameter features and defect types corresponding to partial discharge signals of power cables under laboratory conditions, and multi-parameter features corresponding to partial discharge signals of power cables at industrial sites; The defect recognition model generation module is used to use the multi-parameter features and defect types corresponding to the partial discharge signals of power cables under laboratory conditions as labeled data for the training of the CNN defect recognition model, and to use the multi-parameter features corresponding to the partial discharge signals of power cables at industrial sites as unlabeled data with unknown defect types for self-training, thereby obtaining a semi-supervised incremental learning defect recognition model based on CNNL. The industrial site defect type recognition module is used to utilize the semi-supervised incremental learning defect recognition model based on CNNL, perform defect recognition based on semi-supervised CNNL incremental learning, update the semi-supervised incremental learning defect recognition model based on CNNL, and determine the defect type corresponding to the partial discharge signal of the power cable at the industrial site collected in real time.

9. A terminal, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method described in any one of claims 1 to 7.

10. A computer storage medium, characterized in that: Computer controllable instructions are stored, and the computer controllable instructions are used to execute the method according to any one of claims 1 to 7.

Images