Power equipment monitoring method based on multivariate data

Abstract

The invention provides a power equipment monitoring method based on multivariate data, and belongs to the technical field of power monitoring. According to the technical scheme, firstly, intelligent sensing technologies of key state parameters of equipment such as partial discharge ultrahigh-frequency signal acquisition, infrared registration and comprehensive monitoring are researched on a sensing layer, on the basis, state early warning technologies such as insulation defect diagnosis and overheating diagnosis are researched, finally, an intelligent sensor is designed and developed, development of a monitoring system is completed, and the monitoring system is applied to the field of monitoring. The wide interconnection of equipment and personnel is realized, and the monitoring and inspection information depth acquisition and equipment operation and maintenance efficiency are improved. Meanwhile, the invention provides a systematic method for testing the electrical life of the sulfur hexafluoride circuit breaker for capacitor bank switching, and the method plays a guiding role in state evaluation of the sulfur hexafluoride circuit breaker of the capacitor bank. On the basis of the test method, a capacitor bank sulfur hexafluoride circuit breaker state evaluation model can be established from two aspects of switching data and dynamic resistance.

Description

Technical Field

[0001] The present invention belongs to the technical field of electric power monitoring, and in particular relates to an electric power equipment monitoring method based on multivariate data. Background Art

[0002] Understanding the operating status and operating environment of power transmission and transformation equipment and timely discovering hidden dangers in the operation of power grid equipment are key issues in the operation and maintenance of power equipment. Operation and maintenance inspections are of great significance for timely discovering equipment hidden dangers and ensuring the safe operation of the power grid. However, with the rapid increase in the number of equipment and the continuous improvement of the requirements for power supply reliability, the existing inspection model has gradually shown the following disadvantages:

[0003] (1) Inspection work is stressful. The contradiction between the rapid growth of the power grid and the relatively stable number of operation and maintenance personnel is becoming increasingly prominent. Inspection personnel have a heavy workload and high costs such as labor and vehicles, which limits the efficiency and benefits of operation inspections.

[0004] (2) Manual inspections are not real-time. Operation and maintenance personnel have time intervals between inspections. In bad weather and abnormal working conditions, it is difficult to complete routine inspections. It is impossible to continuously monitor the status of equipment 24 hours a day. Defects that occur during this period cannot be discovered and handled in a timely manner.

[0005] (3) Manual inspections have blind spots. Due to factors such as equipment structure, installation conditions, and personnel positions, operators are unable to observe abnormal conditions at the top of the main transformer and other locations, resulting in defects in some special locations not being discovered.

[0006] (4) Low level of intelligent inspection. Operation and maintenance personnel mainly rely on human eyes and some simple live detection instruments to conduct routine inspections and manually enter defect information. The level of informatization and intelligence is low, and the types of defects that can be found are small and the accuracy is insufficient.

[0007] In the operation and maintenance of power transmission and transformation, due to the large number of equipment, complex operating environment and limitations of status detection technology, how to timely and accurately control the operating status of equipment is still an urgent problem to be solved. The main problems of the commonly used detection methods are as follows:

[0008] 1) The existing online monitoring system for power equipment has low efficiency and poor economy, and has low operational stability and reliability, and often has false alarms or missed alarms. Therefore, the existing online monitoring system has low recognition and low use value.

[0009] 2) Traditional inspection and live-line detection methods cannot monitor the entire process of equipment operation due to their fixed inspection and test cycles. In addition, the number of inspection personnel, work experience, skill and technical level, and physical and psychological qualities and sense of responsibility of the personnel also have a great impact on the measurement results. The detection efficiency is low, and equipment hidden dangers may not be discovered in time.

[0010] 3) The current substation online monitoring system mainly collects on-site images and infrared thermal imaging data, but it is currently mainly at the stage of on-site data collection and simple threshold judgment, lacking intelligent analysis technology to automatically and accurately process and diagnose the detection data. In addition, due to the lack of partial discharge sensors, it is impossible to detect partial discharge of power equipment.

[0011] With the gradual promotion and application of online monitoring systems and inspection robots, the above problems have been alleviated to a certain extent. However, the online monitoring system is a single-point measurement and is expensive, with a low input-output ratio, making it difficult to apply on a large scale. The main problems of the commonly used detection methods are:

[0012] 1) The online monitoring system for power equipment is costly, has a single function, and has low operational stability and reliability. False alarms or missed alarms often occur. Therefore, the existing online monitoring system is not highly recognized and has low use value.

[0013] 2) Live detection has a fixed test cycle and cannot monitor the entire process of equipment operation. The number of testers, work experience, skill and technical level, and physical and psychological qualities and sense of responsibility also have a great impact on the measurement results. The detection efficiency is low and equipment hidden dangers may not be discovered in time.

[0014] 3) The inspection route and field of view of the inspection robot are limited, and it cannot achieve full coverage. There are relatively few inspection items. It mainly collects on-site images and infrared thermal imaging data, and mainly stays at the stage of on-site data collection and simple threshold judgment. It lacks intelligent analysis technology to automatically and accurately process and diagnose the detection data.

[0015] Therefore, current equipment status detection methods are insufficient in their ability to detect equipment defects, and there is an urgent need to improve the level of informationization and intelligence of substation operation inspections. Summary of the invention

[0016] The present invention aims to provide a method for monitoring electric power equipment based on multivariate data in order to address the technical defects of the prior art and solve the technical problems of low efficiency, high cost, poor timeliness, unsatisfactory stability and reliability of conventional inspection and detection methods.

[0017] In order to achieve the above technical objectives, the present invention adopts the following technical solutions:

[0018] A power equipment monitoring method based on multivariate data, including a partial discharge monitoring part, an infrared thermal imaging monitoring part, and a state diagnosis and early warning part;

[0019] Among them, the partial discharge monitoring part includes: conditioning the output signal of the partial discharge sensor, and then inputting it into the acquisition unit, the acquisition unit simultaneously acquires the synchronization signal, realizes the acquisition of the partial discharge signal and the synchronization signal, and then uploads the acquired data to the detection background;

[0020] The infrared thermal imaging monitoring part includes an image registration module and a fault diagnosis module. The image registration module uses a multi-source image registration method to register the infrared thermal image of the device and the visible light image of the device, and outputs the registered infrared thermal image to the fault diagnosis module. The fault diagnosis module extracts characteristic values ​​including temperature, texture, and shape, and the BP neural network diagnosis model outputs the infrared thermal imaging fault diagnosis result of the device according to the characteristic values;

[0021] The status diagnosis and early warning part includes: equipment identification and fault diagnosis through image processing methods; establishing a partial discharge data set and using a deep convolutional network model to identify the type of partial discharge; and identifying partial discharge insulation defects based on a convolutional neural network.

[0022] Preferably, the local discharge monitoring part is performed by a local discharge monitoring device, which includes a UHF sensor and a power phase synchronization module. A signal conditioning module, an MCU and a communication module are provided in the UHF sensor. The signal conditioning module, the MCU and the communication module are communicatively connected in sequence. The communication module is wirelessly connected to the power phase synchronization module via an RF component.

[0023] Preferably, when configuring the partial discharge monitoring device, the UHF sensing coupling unit is designed and calibrated, the design parameters include reflection coefficient, directivity coefficient, sensitivity, and the calibration methods include time threshold method and frequency sweep method.

[0024] Preferably, the amplification circuit of the UHF amplifier of the partial discharge monitoring device adopts a multi-stage cascade structure, and a passive protection circuit and a solid-state protection circuit are provided in the circuit of the UHF amplifier.

[0025] Preferably, in the partial discharge monitoring part, the analog signal conditioning unit converts the UHF signal into two signal outputs, LF and RF, wherein LF is the envelope of the UHF signal obtained by a broadband envelope detector, and RF is the output after the UHF signal is amplified; the LF signal output by the analog signal conditioning unit is buffered, discretely and digitally processed by the acquisition unit, and then compressed by ARM before being sent to the server through the optical fiber interface for analysis and display.

[0026] Preferably, in the partial discharge monitoring part, the partial discharge signal is preprocessed, the characteristic value is extracted, and then the data structure is expressed; the extraction of the characteristic value is realized by a pulse signal extraction module, and the pulse signal extraction module includes a high-speed ADC, FPGA and a main control ARM chip. The envelope detection waveform of the partial discharge discharge pulse is conditioned and input into the high-speed ADC, and the output continuous digital signal stream is subjected to real-time feature extraction by the FPGA to obtain the waveform characteristic parameters of the discharge pulse.

[0027] Preferably, in the infrared thermal imaging monitoring part, the SIFT local feature descriptor is used to align multi-source images; the SIFT feature descriptor is reflected by the gradient direction histogram in the neighborhood on the scale image: the center point is taken as the position of the feature point, and the coordinate axis is rotated to the direction of the feature point to maintain the rotation invariance of the feature descriptor, and then a 16×16 neighborhood is taken with the feature point as the center, and the neighborhood is evenly divided into 4×4 block sub-regions, and then the gradient direction histogram in 8 directions is calculated on each 4×4 sub-region, and the accumulated value of each gradient direction is plotted, and each feature point forms a 128-dimensional SIFT feature vector.

[0028] Preferably, in the infrared thermal imaging monitoring part, the registration method is calibrated in the following steps:

[0029] Step 1: Read the infrared thermal image and the corresponding visible light image of the power equipment in the substation;

[0030] Step 2: Set the threshold for selecting SIFT feature points to 0.8, and obtain N1 SIFT feature points in the infrared thermal image and N2 SIFT feature points in the visible light image;

[0031] Step 3: Randomly select 4 pairs of matching feature points from the candidate matching feature point pairs, and establish the corresponding equation group to calculate the geometric transformation matrix H of the visible light image and infrared thermal image;

[0032] Step 4: After geometric transformation of the remaining feature points of the image to be matched, calculate the distance from the mapping point to the candidate matching point;

[0033] Step 5: Set the RANSAC threshold to 1. If the calculated distance is less than the allowable error, the candidate point is the correctly matched feature point, i.e., the inlier point. The others are marked as outliers, and the number of inliers is counted.

[0034] Step 6: Select another 4 pairs of candidate matching points from the candidate points, and repeat steps 3 to 5. When the number of inliers meets the conditions, select the set with the largest number of inliers as the best inlier set, and the iteration ends.

[0035] Preferably, in the infrared thermal imaging monitoring part, the infrared and visible light images are registered by edge point detection, and the infrared and visible light images are fused by FDST scaling transformation.

[0036] Preferably, in the status diagnosis and warning part, the device identification and fault diagnosis are performed by image processing method, including: image superpixel segmentation based on HSV space; infrared image feature extraction based on Faster-RCNN neural network.

[0037] Preferably, in the status diagnosis and warning part, the identification of partial discharge insulation defects based on convolutional neural networks includes: establishing a fingerprint library of typical insulation defect maps of power equipment including conductor protrusions, suspended discharges, free particles, and surface discharges, and performing identification and classification based on a pattern recognition algorithm based on classifier ensemble learning.

[0038] Preferably, the method further comprises a sulfur hexafluoride circuit breaker monitoring section; the sulfur hexafluoride circuit breaker monitoring section comprises: performing a degradation simulation test on the electrical life of the sulfur hexafluoride circuit breaker, and monitoring the electrical life of the sulfur hexafluoride circuit breaker based on the test results.

[0039] Preferably, the degradation simulation test comprises the following steps:

[0040] Step 1: Perform a rated capacitive current breaking test on the sulfur hexafluoride circuit breaker to be tested;

[0041] Step 2: Conduct an electrical life degradation simulation test on the sulfur hexafluoride circuit breaker to be tested;

[0042] Step 3: Perform a dynamic resistance test on the sulfur hexafluoride circuit breaker to be tested;

[0043] Step 4: Perform a rated capacitive current breaking test, an electrical life degradation simulation test, and a dynamic resistance test on the sulfur hexafluoride circuit breaker to be tested.

[0044] The present invention provides a method for monitoring electric power equipment based on multivariate data. The method integrates partial discharge, infrared, and video. It monitors the heating conditions of electric power equipment and surrounding equipment in real time through infrared spectra to avoid accidents caused by local overheating. At the same time, it adopts broadband electromagnetic wave sensors to receive partial discharge signals generated by electric power equipment, and uses dynamic frequency scanning technology to find frequency bands with less background noise, thereby improving the sensitivity of partial discharge detection and timely warning of insulation defects of electric power equipment. The multi-source video data fusion processing of visible light + infrared is adopted to greatly improve the recognition rate of abnormal detection of the appearance and environment of electric power equipment. NB-IoT and 4G network technologies are adopted to achieve high-reliability access to monitoring terminals and improve the reliability of the detection system.

[0045] The present invention first studies the intelligent perception technology of key equipment status parameters such as partial discharge ultra-high frequency signal acquisition, infrared alignment, and comprehensive monitoring at the perception layer. On this basis, it studies the status warning technologies such as insulation defect diagnosis and overheating diagnosis. Finally, it designs and develops intelligent sensors to complete the development of the monitoring system, realizes the extensive interconnection of equipment and personnel, and at the same time improves the in-depth acquisition of monitoring and inspection information and the efficiency of equipment operation and maintenance. At the same time, the present invention proposes a systematic capacitor bank switching sulfur hexafluoride circuit breaker electrical life test method, which plays a guiding role in the status evaluation of capacitor bank sulfur hexafluoride circuit breakers. Based on the test method, a capacitor bank sulfur hexafluoride circuit breaker status evaluation model can be established from the two aspects of switching data and dynamic resistance.

[0046] The present invention adopts a wide-band electromagnetic wave sensor, and through dynamic frequency scanning, it searches for frequency bands with less background noise, improves the sensitivity of partial discharge monitoring, and realizes real-time online monitoring of spatial partial discharge signals. The present invention is based on a multi-source image registration method based on an improved SIFT, and adopts registration and analysis technology of infrared images and visible light images to realize three-dimensional reconstruction of substation equipment and eliminate interference from distant background equipment. The present invention integrates artificial intelligence algorithms and deep learning to develop a comprehensive state assessment model for multi-source data fusion. The present invention has developed an electrical life test method for sulfur hexafluoride circuit breakers for switching on and off 35kV capacitor banks, which can guide the research on electrical life tests of circuit breakers and serve as a technical indicator for circuit breaker monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a schematic diagram of UHF-based partial discharge intelligent sensing monitoring technology.

[0048] Figure 2 This is a schematic diagram of the amplifier cascade.

[0049] Figure 3 This is the protection schematic diagram of the amplifier (SA).

[0050] Figure 4 is a circuit diagram of an analog signal processing unit (ASPU).

[0051] Figure 5 This is the internal structure diagram of the data acquisition unit.

[0052] Figure 6 It is a data structure diagram of partial discharge pulse characteristic data stream.

[0053] Figure 7 It is the technical roadmap for infrared thermal imaging intelligent monitoring.

[0054] Figure 8 It is a schematic diagram of generating SIFT feature point descriptors from feature point domain gradient information.

[0055] Fig. 9It is the image registration and fusion technology roadmap.

[0056] Fig.10 It is a roadmap for automatic diagnosis of partial discharge insulation defects based on multi-classifier ensemble learning.

[0057] Fig.11 It is the schematic diagram of the rated capacitive current breaking test in the synthetic experiment.

[0058] Fig.12 It is a current waveform diagram of rated capacitive current breaking test in synthetic experiment.

[0059] Fig.13 This is a circuit diagram of an electrical life degradation simulation test.

[0060] Fig.14 This is the schematic diagram of the four-wire method for measuring dynamic resistance. DETAILED DESCRIPTION

[0061] The specific embodiments of the present invention will be described in detail below. In order to avoid too much unnecessary details, the well-known structures or functions will not be described in detail in the following examples. The approximate language used in the following examples can be used for quantitative expression, indicating that the quantity can be allowed to have a certain change without changing the basic function. Unless otherwise defined, the technical and scientific terms used in the following examples have the same meaning as those generally understood by those skilled in the art to which the present invention belongs.

[0062] 1. Partial discharge intelligent monitoring technology based on space ultra-high frequency electromagnetic wave (UHF) measurement

[0063] Figure 1 This is an implementation method of the UHF partial discharge measurement module. The output signal of the partial discharge sensor is input into the acquisition unit after signal conditioning. The acquisition unit simultaneously acquires the synchronization signal to realize the acquisition of the partial discharge signal and the synchronization signal, and finally uploads the acquired data to the detection background through the 4G network.

[0064] 1.1 Design and calibration of UHF sensor coupling unit

[0065] The key to detecting the UHF signal of partial discharge of power equipment is the performance of the receiving sensor. Since the frequency component of the partial discharge signal is as high as several GHz and the signal is attenuated during the propagation process, the detection sensor is required to have sufficient bandwidth and high sensitivity. At the same time, since the location of the discharge source is unknown, the sensor needs to have omnidirectional detection performance. Therefore, in order to realize the detection and positioning of the substation discharge source, the design of the UHF coupling unit needs to consider the following requirements:

[0066] Reflection coefficient: The reflection coefficient is an indicator to measure the matching between the antenna and the feeder, which is defined as:

[0067]

[0068] Where: Z in is the antenna input impedance; Z c is the feeder characteristic impedance.

[0069] When the antenna reflection coefficient is less than -10dB, it indicates that when the antenna is used as a radiator, the power loss is about 10% and the impedance matching is good. According to the reciprocity theorem of the antenna, when the antenna is used as a sensor, it also has good matching characteristics.

[0070] Directivity coefficient: The directivity coefficient D is used to describe the concentration of the electromagnetic wave radiation energy of the antenna. It is defined as: under the conditions of the same total radiation power and the same distance, the square of the field strength E in the maximum radiation direction of the antenna max The field strength E at the same location as the non-directional point source 0 The ratio of the squares of is denoted by D:

[0071]

[0072] Where: P r and P r0 is the radiated power of the antenna and the non-directional point source

[0073] The definition of directivity coefficient is based on the radiation characteristics of the antenna. According to the reciprocity theorem of the antenna, when the antenna is used for spatial detection of partial discharge in a substation, due to the ambiguity of the location of the discharge source, the directivity of the antenna is required to be omnidirectional or nearly omnidirectional.

[0074] Sensitivity: Sensitivity (i.e. effective height) is an important indicator of the UHF coupling unit, reflecting the ability of the sensor to convert the incident electric field into voltage output:

[0075]

[0076] In the formula, is the antenna induced potential, is the incident electric field strength.

[0077] The National Grid Company of the United Kingdom has made clear provisions for the sensitivity of the UHF coupling unit in the range of 500MHz-1500MHz. The minimum effective height over 80% of the frequency range is 2mm, and the average effective height over the full frequency range is at least greater than 6mm. The UHF coupling unit designed for this project is expected to have an average effective height of at least 8mm over the full frequency range.

[0078] The electromagnetic simulation software to be used in the design of the UHF coupling unit is CST electromagnetic wave simulation software and Ansoft's HFSS software, which can directly obtain various parameters of microwave coupling components through simulation calculations, such as characteristic impedance, propagation parameters, S parameters, various field distributions, antenna patterns and other results. When designing the structure of the UHF antenna for partial discharge, the actual installation position structure of the UHF detection sensor of the power equipment and the broadband characteristics of the UHF signal for partial discharge will be taken into account. The traditional test of the antenna is based on the frequency domain point of view, discussing the basic characteristics of the antenna, including the bandwidth, waveform dispersion, directional pattern, directivity coefficient, gain, aperture, field shape, polarization and matching of the antenna. This method is suitable for analyzing point-frequency working or relatively narrow bandwidth steady-state working systems. The UHF partial discharge pulse is essentially an ultra-wideband signal, so a broadband transient working system should be used for analysis. The following uses the time domain method and the frequency sweep method to analyze the pulse waveform.

[0079] 1) Time domain calibration

[0080] Electromagnetic pulse waveform; pulse source parameters: rise time 1.8ns, half-height width 23ns; sensor output voltage; sensor equivalent height.

[0081] 2) Frequency sweep calibration

[0082]

[0083] Where: He(f): effective height; U0(f): sensor output level; Ei(f): incident electric field intensity.

[0084] Sweep mode:

[0085] Frequency range: 300MHz-1.5GHz;

[0086] Sweep mode: linear sweep;

[0087] The frequency sweep step is 10MHz.

[0088] The average effective height of the sensor within the design frequency band (300MHz~1.5GHz) is 10.32mm.

[0089] 1.2UHF amplifier and its protection design

[0090] The UHF amplifier is one of the key components of UHF partial discharge signal detection. Its input protection performance directly affects the reliability of the system; the performance of the bandpass filter will also affect the authenticity of the detection signal; the quality of the amplifier characteristics affects the signal-to-noise ratio and detection sensitivity; the performance of the envelope detector will also affect the stability of the envelope signal. The UHF amplifier consists of an input protection circuit, a bandpass filter, a low-noise amplifier, an envelope detection circuit, etc. The input protection circuit prevents external instantaneous high voltage from breaking down the input bandpass filter electronic components. The bandpass filter filters out the out-of-band noise outside 300MHz to 1500MHz and the GSM 900MHz communication frequency. The low-noise amplifier amplifies the input UHF signal to effectively utilize the dynamic range of the diode-based envelope detector. The envelope detection circuit is used to extract the envelope of the UHF signal and convert the signal with a frequency of up to 1.5GHz into a lower signal frequency, which is convenient for digital quantity acquisition using a conventional embedded data acquisition unit.

[0091] The power of the partial discharge signal received by the UHF coupler is about -85dBm to -15dBm (decibel milliwatts). The sensitivity of the envelope detection circuit using a diode is no higher than -40dBm. Therefore, a low noise amplifier (LNA) must be used to amplify the UHF signal before envelope detection. Circuits such as bandpass filters have a certain power loss. The maximum gain of the UHF amplifier is designed to be 50dB. Obviously, a single-stage amplifier cannot meet the requirements; therefore, the amplifier circuit adopts a multi-stage cascade and is designed to be controllable. Figure 2 .

[0092] In terms of protection circuit design, the switching transients of circuit breaker switches cause more problems than the flashover transients in high voltage testing. The amplifiers with additional passive protection optimize the high frequency characteristics of the self-shutdown circuits. With these protection circuits integrated, the amplifiers can be directly connected to the sensors in the monitoring system.

[0093] 1.3 Analog signal conditioning unit and acquisition unit

[0094] The function of the analog signal conditioning unit is to convert the UHF signal into two signal outputs, LF and RF. Among them, LF is the envelope of the UHF signal, obtained by the broadband envelope detector, and RF is directly output after the UHF signal is amplified (10dB). The principle block diagram of the analog signal conditioning unit is as follows Figure 4 The broadband envelope detector mainly includes a buffer stage (40dB), two detection diodes and a 500KHz low-pass filter amplifier.

[0095] The LF signal output by the analog signal conditioning unit is buffered, discretely and digitally processed by the acquisition unit, and finally compressed by ARM before being sent to the server through the optical fiber interface for analysis and display.

[0096] The signal conditioning unit includes a multi-stage amplifier and an envelope detector. The gain of the first stage amplifier is 30dB, and the detection bandwidth is 1500MHz. In most cases, the spectrum of partial discharge is within 500 to 1300MHz, and the peak is within the range of 700 to 1100MHz. The dynamic range of the signal amplitude is very wide, the discharge generated by the corona is less than 1pC, while the discharge generated by the floating electrode is greater than 1000pC. By limiting at each stage, the multi-stage amplifier can be protected from overload. The envelope detector can use AD8313, which has a bandwidth of 0.1 to 2.5GHz and an input dynamic range of up to 70dB±3dB.

[0097] The DSP acquisition processing unit digitizes the pulse envelope and stores the amplitude. The acquisition unit also measures the power frequency phase information of the pulse to identify the type of discharge. In order to effectively identify the type of partial discharge, the system needs to continuously collect signals of 50 power frequency cycles each time it samples. The sampling rate of the acquisition system is set to 2.5MS / s, the default sampling interval is 30 minutes, and 48 sets of data are collected every day.

[0098] In order to realize the display, collection, storage and analysis of local discharge signals on site, the analog signal output by the sensor of the detection device needs to be sampled and converted into a digital signal for subsequent processing and analysis. However, combined with the characteristics of the local discharge signal, the signal has a higher frequency parameter, and in order to facilitate the spectrum analysis of the discharge signal, the phase of a single discharge waveform in the power frequency cycle must be detected in real time. Calculated at a sampling rate of 100MS / s, the signal of a complete power frequency cycle requires a size of 20MB, so the acquisition unit used in this system needs to have both a high sampling rate and a large storage depth, so as to be able to save the original discharge signal of the complete power frequency cycle.

[0099] Ultra-high frequency partial discharge detection requires that data acquisition must meet both high speed and trigger synchronization, and requires the acquisition card to have a high sampling frequency, wide input signal bandwidth, multiple trigger modes, suitable input impedance, and standard bus specifications. This project uses a self-developed FPGA high-speed acquisition card based on network transmission, realizing the system's distributed structure, front-end acquisition, and back-end analysis mode to simplify the system structure and improve stability.

[0100] The FPGA network data acquisition module based on network transmission includes a control part, an AD acquisition part, a DDR storage part, a transmission part and a watchdog part. The module structure is shown in Figure 5 :

[0101] The FPGA control chip is mainly responsible for collecting data from the 100MSPS AD (AD9433) and then storing it in SDRAM. When the host computer needs to read the data, the data is sent to the host computer through 100M Ethernet. The acquisition module is dual-channel and can complete the acquisition of two channels of data at the same time (AD*2, SDRAM*2). In order to ensure the stability of the module, the corresponding communication protocol is added to the FPGA internal program. At the same time, external controllable functions such as acquisition mode, frequency selection, and watchdog start are added according to actual requirements, making it more flexible and convenient to use.

[0102] 1.4 Partial discharge signal preprocessing, feature value extraction and data structure expression

[0103] The pulse signal extraction module mainly includes high-speed ADC, FPGA and main control ARM chip. The envelope detection waveform of the partial discharge pulse is conditioned and input into the high-speed ADC. The output continuous digital signal stream is subjected to real-time feature extraction by FPGA to obtain the waveform characteristic parameters of the discharge pulse. For UHF detection, the three basic characteristic parameters of the partial discharge pulse signal are pulse amplitude, discharge time interval (Δt) and discharge interval (ΔT). These basic characteristic parameters can be expressed in the form of two-dimensional or three-dimensional spectrograms of statistical parameters after statistical processing.

[0104] The smallest unit of the discharge pulse characteristic data stream is the pulse characteristic unit structure, including pulse number, pulse time, pulse peak value, pulse time-frequency characteristic parameters, trigger level and other basic characteristic parameters. The pulse characteristic unit data structure is organized in units of cycles in the time domain to form a data stream in the time domain, such as Figure 6 shown.

[0105] 2. Infrared thermal imaging preprocessing technology

[0106] The technical route of infrared thermal image preprocessing and intelligent diagnosis of power equipment is as follows: Figure 7 According to the technical route, the main content of infrared image preprocessing is to achieve the registration of infrared thermal image and visible light image.

[0107] 2.1 Research on multi-source infrared image registration method

[0108] This project uses SIFT (Scale Invariant Feature Transform) local feature descriptor to perform multi-source image registration. The basis of SIFT algorithm is Gaussian scale space theory. It is necessary to use the difference operator (DoG) to find the differential image scale space, extract the extreme points of the scale space, and then add the rotation invariant characteristics to finally obtain the scale invariant feature point rotation characteristic algorithm.

[0109] 2.2 Constructing SIFT feature descriptors

[0110] The SIFT feature descriptor is represented by the gradient direction histogram in the neighborhood on the scale image. The center point is the position of the feature point, and the coordinate axis is rotated to the direction of the feature point to maintain the rotation invariance of the feature descriptor. Then, a 16×16 neighborhood is taken with the feature point as the center, and the neighborhood is evenly divided into 4×4 blocks. Among them, the length of the arrow represents the gradient modulus value, the direction of the arrow represents the gradient direction of the pixel, and each small grid is a pixel in the scale space where the neighborhood is located. Figure 8 The large circle in (a) is the range of Gaussian weighting. Then, the gradient direction histogram of 8 directions is calculated on each 4×4 sub-region, and the accumulated value of each gradient direction is plotted, as shown in Figure 8 shown.

[0111] Each feature point will eventually form a 128-dimensional SIFT feature vector. The similarity criterion of feature point matching in this paper is based on the Euclidean distance of two SIFT feature description subspaces. If the feature descriptors of feature point pairs p and q are Des p and Des q , then the Euclidean distance is as follows:

[0112]

[0113] A feature point in the reference visible light image is randomly selected, and the first two feature points with the closest Euclidean distance to the infrared thermal image are found. If the ratio of the Euclidean distances of the two feature points (the Euclidean distance of the nearest feature point is the numerator, and the Euclidean distance of the second nearest feature point is the denominator) is less than the set threshold, the pair of matching points is recorded.

[0114] 2.3 Main steps to improve the registration method

[0115] Step 1: Read the infrared thermal image and the corresponding visible light image of the power equipment in the substation;

[0116] Step 2: Set the threshold for selecting SIFT feature points to 0.8, and obtain N1 SIFT feature points in the infrared thermal image and N2 SIFT feature points in the visible light image;

[0117] Step 3: Randomly select 4 pairs of matching feature points from the candidate matching feature point pairs, and establish the corresponding equation group to calculate the geometric transformation matrix H of the visible light image and infrared thermal image;

[0118] Step 4: After geometric transformation of the remaining feature points of the image to be matched, calculate the distance from the mapping point to the candidate matching point;

[0119] Step 5: Set the RANSAC threshold to 1. If the calculated distance is less than the allowable error, the candidate point is the correctly matched feature point, i.e., the inlier point. The others are marked as outliers, and the number of inliers is counted.

[0120] Step 6: Select another 4 pairs of candidate matching points from the candidate points, and repeat steps 3 to 5. When the number of inliers meets the conditions, select the set with the largest number of inliers as the best inlier set, and the iteration ends.

[0121] 2.4 Registration and fusion technology of infrared and visible light images

[0122] The matching accuracy of infrared and visible light images is much lower than that of homologous images. The main reasons can be summarized as follows: (1) Large resolution difference: The resolution of infrared images is mostly within 300×400, which is much lower than the resolution of ordinary visible light images of 1080×1920. This results in serious loss of grayscale details in infrared images and image blur, which is quite different from the clear texture in visible light grayscale images. (2) Large color difference: Since the imaging mechanism of infrared sensors is different from that of visible light sensors, the color of the same target in heterologous images is very different. For example, an object with a smaller (darker) grayscale value in a visible light image may have a higher (or lower) temperature, which results in a higher (or lower) grayscale value of the object in the red image, which is completely different from the color imaged in the visible light image, increasing the difficulty of registration.

[0123] Based on the analysis of the above two types of problems, the FDST transformation and iterative registration methods are used to optimize the problem of large resolution differences, and the problem of large color differences is solved by the sub-pixel optimization method of the registration point neighborhood. The initial resolution of infrared and visible light images is very different, but after interpolation and affine transformation, the resolution of the infrared image is increased to the size of the visible light image resolution. Therefore, re-registering the image at this time will reduce the impact of the large resolution difference. Due to the large color difference between infrared and visible light images, the gradient direction of the extracted PIIFD operator will be affected, resulting in inaccurate registration. Both infrared and visible light images have obviously the same contours. Using this feature, the results of the initial matching are made as follows. Fig. 9 Processing shown.

[0124] The Canny edge detection algorithm is used to extract the contour curve of the image, and the concept of edge correlation is introduced. Three pairs of reference registration points are established in the infrared image, and a sub-pixel traversal area is established for the remaining precise matching points. The sub-pixel units of 0.2 to 0.8 are selected for traversal calculation to make the sub-pixel point with the highest edge correlation as the precise registration point of the point until all candidate registration points are optimized. The optimized registration image is iterated to reduce the error caused by resolution difference and color difference.

[0125] 3. Equipment status diagnosis and abnormal warning technology based on multi-parameter fusion analysis

[0126] The typical defects of major power equipment and its components are analyzed to analyze the relationship between the severity of various defects and the detection quantities such as partial discharge, temperature, sound, and operating conditions, environment, etc., and to extract effective parameters and their combinations that can characterize internal defects and faults. Information fusion technologies such as multivariate statistical analysis and correlation analysis are used to perform information fusion processing and analysis on various data, and mathematical statistics, trend analysis and other technologies are used to mine the characteristic information that characterizes the operating status of different power equipment. Then, artificial intelligence technologies such as neural networks, fuzzy mathematics, and deep learning are used for collaborative analysis to establish a power equipment status assessment and diagnosis model. The main method is to combine the multi-dimensional analysis and machine learning of the horizontal (equipment of the same type) and vertical (historical data) data of the equipment status quantity, evaluate the key status quantities that affect the health status of the equipment item by item, and give the criterion for quantitative evaluation of the abnormal key status of the equipment. For equipment whose status evaluation results show that the key status has obviously declined (reliability decline status, defect status, critical status), abnormal warning information is given.

[0127] 3.1 Equipment identification and fault diagnosis technology based on image processing

[0128] The detection and identification of power equipment and components mainly include the extraction of targets of interest based on image over-segmentation, feature extraction of regions of interest, and classification and identification of targets. In this project, we plan to use traditional methods of feature extraction and classification, while exploring methods based on deep learning.

[0129] (1) Image superpixel segmentation based on HSV space

[0130] The visible light image or infrared thermal image of the inspection equipment is usually quite different from the background. Taking the infrared fault image of the current transformer as an example, a typical infrared fault image usually has the following three characteristics: 1) The imaging noise is large in the entire image range, with the faulty equipment as the main body; 2) The regional characteristics of the fault area are quite different from those of the background area, and the internal characteristics of different areas are similar; 3) The temperature of the fault area is high, showing yellow or close to white, and the temperature gradually diverges with the fault point as the center. Therefore, it is proposed to use the superpixel segmentation method of HSV space.

[0131] (2) Infrared image feature extraction based on Faster-RCNN neural network

[0132] This project is based on the Faster-RCNN algorithm to identify abnormal status of equipment.

[0133] The input image will first pass through a convolutional neural network to extract features from the image. The convolutional neural network contains 13 convolutional layers, 13 activation function layers, and 4 pooling layers. The convolutional layer is used to extract features from the image. The convolution kernel size used in this algorithm is 3×3, and the convolution kernel step size is 1. The activation function layer provides nonlinear characteristics for the network. The activation function uses the ReLU function. The reason why the classic sigmoid function is not used is that the sigmoid function has the gradient vanishing phenomenon, which is not conducive to training. The pooling layer is used to reduce the dimension of the feature vector output by the convolution layer. In this algorithm, the pooling kernel size of the pooling layer is 2×2, which means that every 2×2 pixel area in the input will correspond to 1 pixel in the output. Since the convolutional layer will reduce the edge of the image by one pixel, the algorithm also includes the operation of expanding the edge of the input sample by one pixel.

[0134] The feature map output by the convolutional neural network will be passed to the RPN network. The RPN network is called the region recommendation network, which is also the most important improvement of this algorithm compared to its previous generation algorithm. The region recommendation operation is implemented through "anchor points" and "anchor boxes". The algorithm will first generate a certain number of anchor points on the image, and the number of anchor points is the same as the number of pixels in the feature map output by the convolutional neural network. Considering that the convolutional neural network has passed through four pooling layers with a pooling kernel of 2×2, each anchor point corresponds to a 16×16 area in the original image. Each anchor point will generate 9 anchor boxes with different kernel sizes and aspect ratios.

[0135] The algorithm will compare different anchor boxes with the labels in the training samples, and then divide the anchor boxes into positive samples, negative samples, and irrelevant samples according to the degree of overlap with the correct area. It will also calculate the offset between each anchor box and the labeled box of the training sample. The network will learn through the offset, so that the RPN network has the ability to predict the real object box after training.

[0136] The last part of the algorithm is the region-based convolutional neural network, or R-CNN in the algorithm's name, which, after obtaining the convolutional feature map from the image, passes it through the RPN network to obtain possible object regions and finally extracts features for each possible region, and finally uses these features for classification. R-CNN tries to mimic the final stage of classification CNNs, where a fully connected layer outputs a score for each possible object class. The goal of R-CNN is calculated in almost the same way as the goal of RPN, but different possible categories are considered. R-CNN has two different goals: to classify the proposal into one class, plus a background class, and to better adjust the bounding box of the proposal region based on the predicted class.

[0137] 3.2 Partial discharge mode diagnosis method under complex data sources

[0138] Based on CNN as the basic model, the autoencoder network is used to perform unsupervised pre-training on the sample data to obtain the initial parameters of the convolution layer. Through convolution, pooling and back-propagation operations, the recognition parameters are optimized. By extracting feature maps of multi-source partial discharge big data, the pattern recognition accuracy of massive partial discharge data in complex scenes is effectively improved.

[0139] The impact of partial discharge detection system performance on detection data mainly includes the following aspects:

[0140] (1) Sensor performance. UHF sensors are the key to UHF partial discharge detection. Currently, the UHF sensors in the partial discharge detection systems used in substations are basically developed by various detection manufacturers, and their performance varies. The key parameters such as the effective height and spectrum response of UHF antennas from different manufacturers are very different, resulting in different detection amplitudes for the same discharge signal from different detection instruments.

[0141] (2) Amplitude measurement method. Different detection instruments have different amplitude measurement methods. For example, some instruments directly measure the voltage amplitude output by the sensor, that is, measure it in mV, while some instruments convert the voltage amplitude output by the sensor and measure it in dB or percentage. Different measurement methods also lead to some differences in the detection amplitude.

[0142] (3) Signal-to-noise ratio. There are some differences in the interference signal processing methods of different detection instruments. Some instruments use low-pass filtering, some instruments use band-pass filtering, some instruments use window filtering, and some instruments use typical signal comparison. The parameters applied under the same interference signal processing method are also different, resulting in differences in signal amplitudes. At the same time, there are also differences in the phase distribution of partial discharge signals.

[0143] (4) Whether there is a power frequency phase synchronization function. Since the voltage frequency at the substation site drifts and cannot guarantee strict power frequency, the detection data of the instrument without power frequency phase synchronization function will have phase offset after a certain period of time, the PRPS spectrum of partial discharge data will be distorted, and the PRPD spectrum will be scattered in phase, resulting in phase differences from the detection data with power frequency phase synchronization.

[0144] 3.3 Partial discharge insulation defect identification based on convolutional neural network

[0145] Establish a fingerprint library of typical insulation defects of power equipment such as conductor protrusions, suspended discharge, free particles, and surface discharge, and automatically identify and classify them based on the pattern recognition algorithm of classifier ensemble learning. The automatic diagnosis process of partial discharge insulation defects based on multi-classifier ensemble learning is as follows: Fig.10 shown.

[0146] (1) Basic principles of convolutional neural networks

[0147] Convolutional Neural Networks (CNN) is a feedforward neural network. Different from traditional pattern recognition models, convolutional neural networks do not extract features manually, but directly input the original two-dimensional image, automatically extract features during the training and learning process, and finally output the classification results.

[0148] A convolutional neural network is a multi-layer neural network, where each layer is composed of multiple two-dimensional planes, and each two-dimensional plane is composed of multiple neurons. The general network structure is that convolutional layers and pooling layers appear alternately. Through convolution and pooling of input data, data features are extracted during training and learning, and finally a fully connected layer is added to obtain the output.

[0149] (2) Partial discharge pattern recognition based on convolutional neural network

[0150] The specific implementation steps are as follows:

[0151] 1) Normalize the training sample set data;

[0152] 2) Construct an autoencoder model and use the training sample set data to perform unsupervised training on the autoencoder;

[0153] 3) Build a deep convolutional network model and use the model parameters obtained by the autoencoder to initialize the convolutional layer parameters of the convolutional network.

[0154] 4) Use the training sample set data to train the deep convolutional network. Calculate the output of the sample data, calculate the error between the output and the sample label, and use the back propagation (BP) algorithm and stochastic gradient descent method to iteratively update the network parameters to obtain the optimal parameters of the recognition model.

[0155] 5) Normalize the test data and input the trained deep convolutional network model to obtain the pattern recognition results.

[0156] 4. Sulfur hexafluoride circuit breaker electrical life monitoring

[0157] Electrical life test method for sulfur hexafluoride circuit breakers used for switching capacitor banks.

[0158] Step 1: Perform a rated capacitive current breaking test on the sulfur hexafluoride circuit breaker to be tested;

[0159] Step 2: Conduct an electrical life degradation simulation test on the sulfur hexafluoride circuit breaker to be tested;

[0160] Step 3: Perform a dynamic resistance test on the sulfur hexafluoride circuit breaker to be tested;

[0161] Step 4: Perform a rated capacitive current breaking test, an electrical life degradation simulation test, and a dynamic resistance test on the sulfur hexafluoride circuit breaker to be tested.

[0162] In step 1, in order to better simulate the arcing during normal breaking of the circuit breaker and the heavy breakdown after arc extinction, a synthetic circuit composed of a current source and a voltage source is used to perform a rated capacitive current breaking test. The current source provides a large current before the current passes through zero, and the voltage source is used on the circuit breaker under test before the current passes through zero, and the voltage source also provides part of the current. When the current passes through zero, the recovery voltage provided by the voltage source immediately and automatically acts on the circuit breaker under test to meet the large current during arcing and the high voltage after arc extinction during the circuit breaker breaking process.

[0163] Principle wiring diagram Fig.11 As shown. The left side of the circuit breaker CB3 under test is a current source, and its power supply device is an oscillation circuit composed of capacitor C1 and reactor L1. Closing circuit breaker CB1 is used to close the test circuit of the current source part, and the size of the test current is adjusted by reactor L1. LJ is a current zero-crossing detector, which is used to detect the current zero-crossing moment and send out a voltage pulse. The right side of CB3 is a voltage source, which is composed of an oscillation circuit composed of capacitor group C2 and reactor L2. C0 and R0 are the capacitor and resistor for adjusting the inherent transient recovery voltage parameters, and HS is the spark gap.

[0164] The operating procedure for this line test is as follows: Before the test, adjust the parameters of the reactors L1 and L2, charge the capacitor groups C1 and C2 to the specified values, and the values ​​of the capacitor group and the inductor meet the requirements of generating the rated capacitive current and recovery voltage. The auxiliary circuit breaker CB2 and the tested circuit breaker CB3 are in the closed state. During the test, the closing circuit breaker CB1 is closed, so that the current i1 flows through the current source circuit, and then CB2 and CB3 are commanded to be disconnected at the same time. Within 200 to 300us before the current i1 passes through zero, the current zero-crossing detector LJ sends out a voltage pulse to ignite the spark gap HS in the voltage source circuit, so that the capacitor group C2 discharges the reactor L2 to generate current i2. i2 is much smaller than i1 (1 / 5 to 1 / 20). At this time, the current flowing through CB2 is i1, and the current flowing through CB3 is i1+i2. The current waveform is as follows Fig.12 As shown, HS ignites at time t1, and t2 and t3 are the moments when i1 and i1+i2 cross zero, respectively. It can be seen that i1 crosses zero earlier than i2, so the arc in CB2 is extinguished first. After CB2 extinguishes the arc, the current source circuit is isolated from CB3, and then CB3 is completely transferred to the voltage source circuit. After i2 crosses zero, a transient recovery voltage determined by C2, L2, R0 and C0 appears on the arc gap of circuit breaker CB3, and its steady-state component is the residual voltage on capacitor bank C0.

[0165] The parameter selection method for each device during this synthetic loop test is as follows:

[0166] (1) In order to prevent the waveform of the test current from being excessively distorted by the arc voltage, the voltage of the current source should be no less than 10 times the arc voltage in the circuit breaker under test.

[0167] (2) The oscillation rate f2 of the voltage source circuit should be selected between 250 and 1000 Hz

[0168] (3) To ensure that the rate of decrease of the current through zero in the circuit breaker CB3 under test is the same as that in the direct test, the effective value I2 of the voltage source current should be selected as follows:

[0169]

[0170] Where I1 is the effective value of the current source.

[0171] In step 2, an ablation test circuit is used to perform an electrical life degradation simulation test. The purpose of this experiment is to use a large current to ablate the circuit breaker contacts, so only Fig.11 The current source circuit on the left and the circuit breaker CB3 under test do not require a voltage source circuit. The test circuit is as follows Fig.13 shown.

[0172] The circuit mainly includes capacitor bank C1, inductor L1, closing circuit breaker CB1, auxiliary circuit breaker CB2 and tested circuit breaker CB3. Before the test, adjust the parameters of reactor L1 to charge C1 to the specified value. The values ​​of capacitor bank and inductor meet the requirements of generating power frequency short-circuit current. Auxiliary circuit breaker CB2 and tested circuit breaker CB3 are in the closed state. During the test, close the closing circuit breaker CB1, and then command CB3 to open. Auxiliary circuit breaker CB2 does not undertake the opening task. The circuit current frequency calculation formula is as follows:

[0173]

[0174] In step 3, dynamic resistance measurement (DRM) is to measure the dynamic contact resistance of the moving and static contacts during the opening or closing process of the high-voltage circuit breaker, and analyze the arc contact erosion state through the dynamic contact resistance. Since the dynamic contact resistance during opening is more stable than that during closing, and the contact resistance of the main and arc contacts is clearly distinguished, the dynamic contact resistance during opening is selected to analyze the contact state of the arc contact. Specifically, the contact resistance of the circuit breaker under test is measured using the four-wire method, such as Fig.14As shown. The four-wire method eliminates the influence of the test line resistance and contact resistance, so it is generally used to accurately measure the resistance below 10Ω. In order to eliminate the inductance effect between contacts, a battery is used as a power source, and a test current of more than 1000A is provided. The change curve of the contact resistance between the moving and static contacts is obtained by measuring the voltage drop generated by the test current on the contacts during the opening process. In combination with the contact stroke test, the dynamic resistance-stroke curve is obtained.

[0175] Generally speaking, the resistance of the main and arc contacts increases gradually with the increase of the number of switching times of the circuit breaker. Studies have shown that the resistance of the main contact is generally normal when it is below 200μΩ, and the resistance of the arc contact is normal when it is below 400μΩ. Exceeding this range indicates that the main contact or arc contact is excessively eroded, and the electrical life of the circuit breaker is in the warning or end stage.

[0176] In step 4, the rated capacitive current breaking test, the electrical life degradation simulation test and the dynamic resistance test are performed cyclically, wherein the above cycle refers to completing at least two groups of tests to obtain the current value and the moving and arcing contact resistance values ​​of each group of circuit breaker switching action arcing stage;

[0177] When the sulfur hexafluoride circuit breaker to be tested cannot be successfully disconnected, the current values ​​of all arcing stages finally obtained are integrated to obtain the final total contact wear of the moving and arcing contacts of the circuit breaker, and the final total contact wear of the moving and arcing contacts of the circuit breaker is used as the life reference value of the circuit breaker. The specific calculation formula is:

[0178]

[0179] In the formula, K is a constant, which is related to the properties of the contact material, T is the arcing time, and a is between 1 and 2, which is a constant within a certain current variation range and is related to the properties of the contact material.

[0180] Then, the obtained moving contact and arc contact resistance values ​​are compared with the preset reference moving contact resistance value and reference arc contact resistance value. When it is found that one of the obtained moving contact and arc contact resistance values ​​is greater than the corresponding reference moving contact resistance value and reference arc contact resistance value, an integral calculation is performed based on the current values ​​of all arcing stages currently obtained and (4-3) to obtain the total reference contact wear of the moving and arc contacts of the circuit breaker; the total reference contact wear of the moving and arc contacts of the circuit breaker and the total final contact wear of the moving and arc contacts of the circuit breaker are averaged to obtain the total average contact wear of the moving and arc contacts of the circuit breaker, and the total average contact wear of the moving and arc contacts of the circuit breaker is used as the life reference value of the circuit breaker.

[0181] The embodiments of the present invention are described in detail above, but the contents are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the scope of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for monitoring electric power equipment based on multivariate data, characterized in that: Including partial discharge monitoring part, infrared thermal imaging monitoring part, status diagnosis and early warning part; Among them, the partial discharge monitoring part includes: conditioning the output signal of the partial discharge sensor, and then inputting it into the acquisition unit, the acquisition unit simultaneously acquires the synchronization signal, realizes the acquisition of the partial discharge signal and the synchronization signal, and then uploads the acquired data to the detection background; The infrared thermal imaging monitoring part includes an image registration module and a fault diagnosis module. The image registration module uses a multi-source image registration method to register the infrared thermal image of the device and the visible light image of the device, and outputs the registered infrared thermal image to the fault diagnosis module. The fault diagnosis module extracts characteristic values ​​including temperature, texture, and shape, and the BP neural network diagnosis model outputs the infrared thermal imaging fault diagnosis result of the device according to the characteristic values; The status diagnosis and early warning part includes: equipment identification and fault diagnosis through image processing methods; establishing a partial discharge data set and using a deep convolutional network model to identify the type of partial discharge; and identifying partial discharge insulation defects based on a convolutional neural network.

2. The method for monitoring electric power equipment based on multivariate data according to claim 1, characterized in that: The partial discharge monitoring part is performed by a partial discharge monitoring device, which includes a UHF sensor and a power phase synchronization module. A signal conditioning module, an MCU and a communication module are arranged in the UHF sensor. The signal conditioning module, the MCU and the communication module are sequentially connected in communication. The communication module is wirelessly connected to the power phase synchronization module via an RF component. When configuring the partial discharge monitoring device, the UHF sensing coupling unit is designed and calibrated. The design parameters include reflection coefficient, directivity coefficient and sensitivity. The calibration methods include time threshold method and frequency sweeping method. The UHF amplifier of the partial discharge monitoring device has an amplification circuit that adopts a multi-stage cascade structure. A passive protection circuit and a solid-state protection circuit are arranged in the circuit of the UHF amplifier.

3. The method for monitoring electric power equipment based on multivariate data according to claim 1, characterized in that: In the partial discharge monitoring part, the analog signal conditioning unit converts the UHF signal into two signal outputs, LF and RF, where LF is the envelope of the UHF signal obtained by a broadband envelope detector, and RF is the output after the UHF signal is amplified; the LF signal output by the analog signal conditioning unit is buffered, discretely and digitally processed by the acquisition unit, and then compressed by ARM before being sent to the server through the optical fiber interface for analysis and display.

4. The method for monitoring electric power equipment based on multivariate data according to claim 1, characterized in that: In the partial discharge monitoring part, the partial discharge signal is preprocessed, the characteristic value is extracted, and then the data structure is expressed; the extraction of the characteristic value is realized by a pulse signal extraction module, and the pulse signal extraction module includes a high-speed ADC, an FPGA and a main control ARM chip. The envelope detection waveform of the partial discharge discharge pulse is conditioned and input into the high-speed ADC, and the output continuous digital signal stream is subjected to real-time feature extraction by the FPGA to obtain the waveform characteristic parameters of the discharge pulse.

5. The method for monitoring electric power equipment based on multivariate data according to claim 1, characterized in that: In the infrared thermal imaging monitoring part, the SIFT local feature descriptor is used to align multi-source images; the SIFT feature descriptor is reflected by the gradient direction histogram in the neighborhood on the scale image: the center point is taken as the position of the feature point, and the coordinate axis is rotated to the direction of the feature point to maintain the rotation invariance of the feature descriptor, and then a 16×16 neighborhood is taken with the feature point as the center, and the neighborhood is evenly divided into 4×4 block sub-regions, and then the gradient direction histogram in 8 directions is calculated on each 4×4 sub-region, and the accumulated value of each gradient direction is plotted, and each feature point forms a 128-dimensional SIFT feature vector.

6. The method for monitoring electric power equipment based on multivariate data according to claim 1, characterized in that: In the infrared thermal imaging monitoring part, the registration method is calibrated in the following steps: Step 1: Read the infrared thermal image and the corresponding visible light image of the power equipment in the substation; Step 2: Set the threshold for selecting SIFT feature points to 0.8, and obtain N1 SIFT feature points in the infrared thermal image and N2 SIFT feature points in the visible light image; Step 3: Randomly select 4 pairs of matching feature points from the candidate matching feature point pairs, and establish the corresponding equation group to calculate the geometric transformation matrix H of the visible light image and infrared thermal image; Step 4: After geometric transformation of the remaining feature points of the image to be matched, calculate the distance from the mapping point to the candidate matching point; Step 5: Set the RANSAC threshold to 1. If the calculated distance is less than the allowable error, the candidate point is the correctly matched feature point, i.e., the inlier point. The others are marked as outliers, and the number of inliers is counted. Step 6: Select another 4 pairs of candidate matching points from the candidate points, and repeat steps 3 to 5. When the number of inliers meets the conditions, select the set with the largest number of inliers as the best inlier set, and the iteration ends.

7. The method for monitoring electric power equipment based on multivariate data according to claim 1, characterized in that: In the infrared thermal imaging monitoring part, the infrared and visible light images are registered by edge point detection, and the infrared and visible light images are fused by FDST scale transformation.

8. The method for monitoring electric power equipment based on multivariate data according to claim 1, characterized in that: In the status diagnosis and warning part, the device identification and fault diagnosis are performed through image processing methods, including: image superpixel segmentation based on HSV space; infrared image feature extraction based on Faster-RCNN neural network.

9. The method for monitoring electric power equipment based on multivariate data according to claim 1, characterized in that: In the status diagnosis and warning part, the identification of partial discharge insulation defects based on convolutional neural networks includes: establishing a fingerprint library of typical insulation defect maps of power equipment including conductor protrusions, suspended discharges, free particles, and surface discharges, and performing identification and classification based on a pattern recognition algorithm based on classifier ensemble learning.

10. The method for monitoring electric power equipment based on multivariate data according to claim 1, characterized in that: The method also includes a sulfur hexafluoride circuit breaker monitoring part; the sulfur hexafluoride circuit breaker monitoring part includes: performing a degradation simulation test on the electrical life of the sulfur hexafluoride circuit breaker, and monitoring the electrical life of the sulfur hexafluoride circuit breaker based on the test results.

Images