Electrical equipment defect identification method and system based on acoustic infrared feature fusion

Abstract

The invention discloses a power equipment defect identification method and system based on acoustic infrared feature fusion, and the method comprises the steps: 1), obtaining an audio signal and infrared gray data, and obtaining PRPD atlas data and an audio signal power spectrum according to the audio signal; 2) carrying out dimension reduction feature extraction on the PRPD atlas data, the audio signal power spectrum and the infrared grayscale data to obtain corresponding dimension-reduced PRPD atlas features, audio signal power spectrum features and infrared features; and 3) fusing the dimension-reduced PRPD atlas features, the audio signal power spectrum features and the infrared features, and inputting the fused features into a multi-scale perception multi-input network to carry out defect identification and classification on the power equipment. The method has the advantages of low training cost, high recognition precision and the like.

Description

Technical Field

[0001] The present invention mainly relates to the technical field of power equipment, and particularly relates to a method and system for identifying defects of power equipment by fusing acoustic and infrared features. Background Art

[0002] There may be certain defects or aging phenomena in the insulating materials of power equipment, such as bubbles, cracks, moisture, and impurities. These defects will cause the electrical performance of the insulating materials to decline, thereby triggering partial discharge. During operation, power equipment will be affected by mechanical stresses, such as vibration, shock, and temperature changes. These mechanical stresses may cause the insulating materials to be damaged or loosened, thereby triggering partial discharge. The environment around power equipment will also affect the performance of insulating materials. For example, too high contents of dust, salt spray, and moisture will enhance the surface conductivity of insulating materials, thereby causing partial discharge. There may be problems in the design and manufacturing process of power equipment, such as inappropriate electrode spacing and improper selection of insulating materials, resulting in uneven electric field distribution and thus triggering partial discharge. In addition, overvoltage and current surges in the power system will also trigger partial discharge. When an overvoltage or current surge occurs in the system, it will cause the electric field strength to exceed the tolerance of the insulating material, thereby triggering partial discharge.

[0003] The acoustic fingerprint detection technology is based on a high-performance ultrasonic microphone array and beamforming technology to collect ultrasonic signals emitted by partial discharge, uses spatial filtering technology to pick up the ultrasonic signals of the partial discharge point, extracts the acoustic fingerprint features of the partial discharge point, and determines the type of partial discharge at the defect point. There are still certain problems with the traditional acoustic fingerprint technology, mainly including: environmental interference: The structure of power equipment is relatively complex. For example, in a switch cabinet, there are complex path propagations of sound wave reflection and refraction, which will cause the deviation of the positioning point and a sharp drop in the signal-to-noise ratio of the extracted acoustic fingerprint, resulting in misjudgment. The background noise at the working condition site will also affect the process of acoustic fingerprint detection, leading to the deviation of the partial discharge positioning point and misjudgment of the predicted type.

[0004] Infrared detection is a conventional and effective detection method in the live detection of power grid power equipment, with advantages such as long distance, non-contact, and real-time imaging. During the entire detection process, infrared thermal imaging technology is used to detect power equipment, and possible thermal defects of the equipment are found by analyzing and comparing the infrared thermal images of the equipment, and potential fault problems of power equipment are discovered predictively to avoid serious damage to power equipment.

[0005] Challenges of complex parameters in traditional pattern recognition: Traditional pattern recognition algorithms cannot effectively identify and classify multi-modal parameter inputs, and often require a network with complex parameters to achieve high accuracy, which will increase the hardware computing power burden and power consumption in actual use. Summary of the Invention

[0006] Aiming at the technical problems existing in the prior art, the present invention provides a method and system for identifying power equipment defects by fusing acoustic and infrared features with high recognition accuracy.

[0007] To solve the above technical problems, the technical solution proposed by the present invention is as follows:

[0008] A method for identifying power equipment defects by fusing acoustic and infrared features, comprising the steps of:

[0009] 1) Obtain audio signals and infrared grayscale data, and obtain PRPD map data and the power spectrum of the audio signal according to the audio signal;

[0010] 2) Perform dimensionality reduction feature extraction on the PRPD map data, the power spectrum of the audio signal, and the infrared grayscale data to obtain the corresponding dimensionality-reduced PRPD map features, the power spectrum features of the audio signal, and the infrared features;

[0011] 3) Fuse the dimensionality-reduced PRPD map features, the power spectrum features of the audio signal, and the infrared features, and input them into a multi-scale perception multi-input network for power equipment defect identification and classification.

[0012] Preferably, in step 1), the audio signal is collected by a multi-channel microphone array, the defect point is located according to the beamforming algorithm, the spatial filtering is performed on the defect point to obtain the defect point audio signal, and the PRPD map data and the power spectrum of the audio signal are obtained according to the defect point audio signal; wherein the PRPD map reflects the relationship between the discharge phase, the discharge amount, and the discharge times; and the power spectral density reflects the energy distribution characteristics of the defect ultrasonic signal.

[0013] Preferably, in step 1), after obtaining the infrared grayscale data, image denoising preprocessing is performed, and median filtering and Gaussian filtering are jointly applied to remove internal noise and external noise, wherein the internal noise type is salt-and-pepper noise, and the external noise is Gaussian noise.

[0014] Preferably, in step 2), the PRPD map data and the power spectrum of the audio signal are used as the input of the cascaded stacked autoencoder SAE for dimensionality reduction, and the voiceprint data is reduced from the dimension of 2*1*4096 to 2*1*100; wherein each SAE includes 4 autoencoding processes AE, thereby generating a feature extraction network with 8 hidden layers.

[0015] Preferably, in step 2), the infrared grayscale data is used as the input of the cascaded stacked autoencoder SAE for dimensionality reduction, and the infrared grayscale data is reduced from the dimension of 1*81920 to 1*200; wherein each SAE includes 8 autoencoding processes AE, thereby generating a feature extraction network with 16 hidden layers.

[0016] Preferably, the multi-scale perception multi-input network adopts the "Adam" optimizer, consists of 4 hidden layers and an output layer. The activation function of the output layer is soft-max, and the loss function is "Categorical_cross-entropy", with 5 dimensions output; the activation functions of other layers are all "Relu", and the loss function is "MSE".

[0017] Preferably, the 5 dimensions correspond to five classifications, corresponding to intact, heating defect, corona discharge, floating discharge and surface discharge respectively.

[0018] The present invention also discloses a computer program product, including a computer program, and the steps of the above-mentioned method are executed when the computer program is run by a processor.

[0019] The present invention further discloses a computer-readable storage medium, on which a computer program is stored, and the steps of the above-mentioned method are executed when the computer program is run by a processor.

[0020] The present invention also discloses a power equipment defect recognition system for acoustic infrared feature fusion, including a memory and a processor connected to each other. A computer program is stored on the memory, and the steps of the above-mentioned method are executed when the computer program is run by the processor.

[0021] Compared with the prior art, the advantages of the present invention are as follows:

[0022] Compared with traditional acoustic defect detection, the present invention introduces infrared data for joint determination. On the one hand, it can reduce the misjudgment of the defect point position caused by sound wave reflection and noise, and on the other hand, it can improve the accuracy of partial discharge type determination. The present invention picks up the acoustic fingerprint information of the defect point through the microphone array beamforming technology, which has the characteristics of non-contact, remote collection, and shielding of sound sources that are not interested in the space compared with traditional acoustic fingerprint collection methods.

[0023] The present invention introduces a multi-scale perception autoencoder network model. The layer-by-layer training method can reduce the computing power requirement, prevent gradient dissipation, achieve fast convergence, share weights in different application scenarios, eliminate the pre-training process, reduce the calculation amount, and reduce the pressure of the model on the hardware platform.

[0024] The multi-scale perception multi-input network based on stacked autoencoders of the present invention can synthesize the information contained in infrared grayscale information, ultrasonic PRPD spectrograms, and power spectral densities, and perform feature extraction and dimensionality reduction separately at different scales. Finally, transfer learning is used to combine the output neurons of the trained encoding network, and then the final classification training is performed through the terminal network. Among them, the stacked autoencoder's layer-by-layer training method can reduce the computing power requirements, prevent gradient dissipation, and achieve rapid convergence of the results in subsequent training, thereby realizing data dimensionality reduction through unsupervised training. This network can be used as an algorithm for transfer learning, sharing weights in different application scenarios, eliminating the pre-training process, and reducing the amount of calculation. Description of the Drawings

[0025] Figure 1 It is an embodiment diagram when the power equipment defect recognition method based on acoustic infrared feature fusion of the present invention is specifically applied. Detailed Embodiment

[0026] The present invention will be further described below in conjunction with the specification drawings and specific embodiments.

[0027] As Figure 1 shown, the power equipment defect recognition method based on acoustic infrared feature fusion provided by the embodiment of the present invention specifically includes the steps:

[0028] 1) Obtain voiceprint data: Collect audio signals through a multi-channel microphone array, locate the defect point according to the beamforming algorithm, perform spatial filtering on the defect point to obtain the defect point audio signal, and obtain the PRPD spectrogram data and the audio signal power spectrum according to the audio signal. Among them, the PRPD spectrogram reflects the relationship between the discharge phase, discharge amount, and discharge times; the power spectral density reflects the energy distribution characteristics of the defect ultrasonic signal.

[0029] 2) Obtain normalized infrared grayscale data, and perform image denoising preprocessing, and jointly apply median filtering and Gaussian filtering to remove internal noise and external noise. Among them, the infrared grayscale data reflects the distribution trend of the temperature of the defect point.

[0030] 3) Voiceprint infrared feature extraction: The PRPD data and power spectrum data obtained from the audio signal are 1*4096 matrices, which are used as the input of the cascaded stacked autoencoder (SAE) for dimensionality reduction, and the voiceprint data is reduced from a dimension of 2*1*4096 to 2*1*100. Among them, each SAE includes 4 autoencoding processes (AE), thereby generating a feature extraction network with 8 hidden layers.

[0031] The grayscale data obtained through median filtering and Gaussian filtering preprocessing is 1*81920. After being used as the input of the cascaded stacked autoencoder (SAE) for dimensionality reduction, the infrared grayscale data is reduced from a dimension of 1*81920 to 1*200. Each SAE includes 8 autoencoding processes (AE), thus generating a feature extraction network with 16 hidden layers.

[0032] 4) Feature fusion and classification

[0033] The multi-modal input data are respectively fused after passing through their respective SAR feature extraction layers and then input into the multi-scale perception multi-input network for power equipment defect identification and classification. The multi-scale perception multi-input network uses the "Adam" optimizer, consists of 4 hidden layers and an output layer. The activation function of the output layer is soft-max, the loss function is "Categorical_cross-entropy", and it outputs 5 dimensions. The activation functions of other layers are all "Relu", and the loss function is "MSE".

[0034] Among them, the intact power equipment and defects include the following five types:

[0035] Number Type 0 In good condition 1 Overheating defect 2 Corona discharge 3 Floating discharge 4 Surface discharge

[0036] Compared with the traditional acoustic defect detection, the present invention reduces the misjudgment of the defect point position caused by sound wave reflection and noise on the one hand by introducing infrared data for joint determination, and improves the accuracy of partial discharge type determination on the other hand. The present invention picks up the acoustic defect information through the microphone array beamforming technology, which has the characteristics of non-contact, remote acquisition, and shielding of sound sources not interested in the space compared with the traditional acoustic acquisition method.

[0037] The present invention introduces a multi-scale perception autoencoder network model. The layer-by-layer training method can reduce the computing power requirement, prevent gradient dissipation, achieve fast convergence, share weights in different application scenarios, eliminate the pre-training process, reduce the amount of calculation, and reduce the pressure on the hardware platform of the model.

[0038] The multi-scale perception multi-input network based on stacked autoencoder of the present invention can comprehensively extract and reduce the dimensions of features from infrared grayscale information, ultrasonic PRPD spectrogram, and power spectral density at different scales. Finally, the output end neurons of the trained encoding network are combined using transfer learning, and then the final classification training is carried out through the end network. The layer-by-layer training method of the stacked autoencoder can reduce the computing power requirement, prevent gradient dissipation, and achieve fast convergence of the results in subsequent training, thereby realizing data dimensionality reduction through unsupervised training. This network can be used as an algorithm for transfer learning, sharing weights in different application scenarios, eliminating the pre-training process, and reducing the amount of calculation.

[0039] To better understand the above technical solution, the above technical solution will be described in detail below in conjunction with the accompanying drawings of the specification and specific embodiments.

[0040] As Figure 1 shown, the method for identifying power equipment defects by acoustic-infrared feature fusion according to the embodiment of the present invention specifically includes the following steps:

[0041] S110. Collect multi-channel signals through a microphone array, and remotely pick up power defect point signals through beam forming processing;

[0042] S120. Extract PRPD signals from the defect point signals, that is, the relationships between the pulse amplitude, pulse phase, and pulse number in the ultrasonic signals, corresponding to the relationships between the partial discharge quantity, discharge phase, and discharge times;

[0043] S130. Input the PRPD data for 4-layer SAE dimensionality reduction feature extraction;

[0044] S140. Extract power spectral density features from the defect point signals;

[0045] S150. Input the power spectral density for 4-layer SAE dimensionality reduction feature extraction;

[0046] S160. Collect infrared grayscale data;

[0047] S170. Remove internal noise (the noise type is salt-and-pepper noise) and external noise (Gaussian noise) in the grayscale data through Gaussian filtering and median filtering;

[0048] S180. Input the filtered grayscale data for 8-layer SAE dimensionality reduction feature extraction;

[0049] S190. The connection layer fuses the voiceprint features (PRPD and power spectrum features) and the infrared features. Specifically, the audio signal and the infrared grayscale data are input into the cross-modal Transformer module, and the self-attention mechanism is applied to learn the correspondence between the specific frequency components in the audio signal and the hot spots in the infrared image. For example, when local overheating occurs in the device, it may cause the vibration of the surrounding air, thus generating specific frequency components in the audio signal. Through the self-attention mechanism, the correlation between these frequency components and the hot spots in the infrared image is captured. By establishing this correspondence, the model can more comprehensively understand the mutual influence between different modal features, thereby improving the accuracy of defect recognition. For example, when identifying device defects, the model can simultaneously consider the specific frequency components in the audio signal and the hot spots in the infrared image, so as to more accurately judge the type and severity of the device failure. At the same time, through adaptive weight allocation and attention weight sharing, the contributions of different modal features are dynamically adjusted, and model regularization techniques are applied to prevent over-reliance on the features of a single modality.

[0050] S200 - S220. The voiceprint infrared features are input into the cascaded hidden layer for dimensionality reduction.

[0051] S230. The output layer outputs 5 classification results, corresponding to intact, heat defect, corona discharge, floating discharge, and surface discharge.

[0052] The present invention fuses the voiceprint and infrared features for determination, reducing the misjudgment probability caused by sound wave reflection and refraction in the traditional voiceprint method; introducing a cascaded autoencoder network for feature extraction and dimensionality reduction, reducing the computing power requirement, preventing gradient dissipation, and achieving rapid convergence of the results; remotely obtaining the power defect signal through the microphone array beam forming, having the advantages of non-contact, remote acquisition, and shielding the sound sources not interested in the space.

[0053] The present invention also discloses a computer program product, including a computer program, and the steps of the above-mentioned method are executed when the computer program is run by a processor. The present invention further discloses a computer-readable storage medium, on which a computer program is stored, and the steps of the above-mentioned method are executed when the computer program is run by a processor. The present invention also discloses a power equipment defect recognition system with acoustic infrared feature fusion, including a memory and a processor connected to each other, a computer program is stored on the memory, and the steps of the above-mentioned method are executed when the computer program is run by the processor. The products, media, and systems of the present invention, corresponding to the above-mentioned method, also have the advantages as described in the above-mentioned method.

[0054] The present invention can implement all or part of the processes in the above-described embodiment methods through hardware related to computer program instructions. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, the steps of the above-described method embodiments can be implemented. Among them, the computer program includes computer program code, and the computer program code can be in the form of source code, object code, executable file, or some intermediate form, etc. The computer-readable storage medium includes: any entity or device capable of carrying computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disc, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electrical carrier signal, telecommunication signal, and software distribution medium, etc. The memory is used to store computer programs and / or modules. The processor realizes various functions by running or executing the computer programs and / or modules stored in the memory, and by calling the data stored in the memory. The memory can include high-speed random access memory, and can also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one magnetic disk storage device, flash device, or other volatile solid-state storage devices, etc.

[0055] The above are only the preferred embodiments of the present invention. The protection scope of the present invention is not limited to the above embodiments. All technical solutions falling within the idea of the present invention belong to the protection scope of the present invention. It should be noted that for those of ordinary skill in the art in this technical field, several improvements and refinements made without departing from the principle of the present invention should be regarded as within the protection scope of the present invention.

Claims

1. A method for identifying defects in electric power equipment by fusion of acoustic and infrared features, characterized in that: Includes steps: 1) Obtain audio signal and infrared grayscale data, and obtain PRPD spectrum data and audio signal power spectrum according to the audio signal; 2) performing dimension reduction feature extraction on the PRPD spectrum data, audio signal power spectrum and infrared grayscale data to obtain corresponding dimension-reduced PRPD spectrum features, audio signal power spectrum features and infrared features; 3) The PRPD spectrum features, audio signal power spectrum features and infrared features after dimension reduction are fused and input into the multi-scale perception multi-input network for power equipment defect identification and classification.

2. The method for identifying defects in electric power equipment by combining acoustic and infrared features according to claim 1 is characterized in that: In step 1), audio signals are collected through a multi-channel microphone array, the defect point is located according to the beamforming algorithm, the defect point is spatially filtered to obtain the defect point audio signal, and the PRPD spectrum data and the audio signal power spectrum are obtained according to the defect point audio signal; the PRPD spectrum reflects the relationship between the discharge phase, discharge amount, and discharge number; the power spectrum density reflects the energy distribution characteristics of the defect ultrasonic signal.

3. The method for identifying defects in electric power equipment by combining acoustic and infrared features according to claim 1 is characterized in that: In step 1), after acquiring the infrared grayscale data, image denoising preprocessing is performed, and median filtering and Gaussian filtering are jointly applied to remove internal noise and external noise, wherein the noise type of the internal noise is salt and pepper noise, and the external noise is Gaussian noise.

4. The method for identifying defects in electric power equipment by combining acoustic and infrared features according to claim 1, 2 or 3, characterized in that: In step 2), the PRPD spectrum data and the audio signal power spectrum are used as the input of the cascade stacked autoencoder SAE for dimensionality reduction, and the dimension of the voiceprint data is reduced from 2*1*4096 to 2*1*100; each SAE includes 4 autoencoder processes AE, thereby generating a feature extraction network with 8 hidden layers.

5. The method for identifying defects in electric power equipment by combining acoustic and infrared features according to claim 1, 2 or 3, characterized in that: In step 2), the infrared grayscale data is used as the input of the cascade stacked autoencoder SAE for dimensionality reduction, and the dimension of the infrared grayscale data is reduced from 1*81920 to 1*200; each SAE includes 8 autoencoder processes AE, thereby generating a feature extraction network with 16 hidden layers.

6. The method for identifying defects in electric power equipment by combining acoustic and infrared features according to claim 1, 2 or 3, characterized in that: The multi-scale perception multi-input network uses the "Adam" optimizer and consists of 4 hidden layers and an output layer. The output layer activation function is soft-max, the loss function is "Categorical_cross-entropy", and the output is 5 dimensions; the activation function of other layers is "Relu", and the loss function is "MSE".

7. The method for identifying defects in electric power equipment by combining acoustic and infrared features according to claim 6 is characterized in that: The five dimensions correspond to five categories, namely intact, thermal defects, corona discharge, suspended discharge and surface discharge.

8. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are performed.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the computer program performs the steps of the method according to any one of claims 1 to 7.

10. An acoustic infrared feature fusion power equipment defect identification system, comprising a memory and a processor connected to each other, wherein a computer program is stored in the memory, characterized in that: When the computer program is executed by a processor, the computer program performs the steps of the method according to any one of claims 1 to 7.

Images