Online defect monitoring method and device for distribution network power equipment based on multi-feature fusion

Abstract

The invention discloses an online defect monitoring method and device for distribution network power equipment based on multi-feature fusion. The method comprises the following steps: S1, collecting multi-source data; s2, signal filtering and enhancement processing; s3, multi-feature fusion study and judgment; s4, online early warning and data storage; s5, on-site diagnosis verification and defect elimination; according to the method, core characteristics of different types of defects are covered through multi-modal data complementation, and meanwhile, a differential processing technology is adopted for different signal characteristics, so that the limitation of a single signal is effectively avoided; three-level feature fusion study and judgment logic is constructed, and compared with a single feature study and judgment mode or a simple fusion mode, the accuracy of defect type recognition, accurate position positioning and severity degree judgment is effectively improved; a framework of cooperative work of the online monitoring system and the portable diagnosis device is constructed, efficient cooperation of online large-range preliminary screening and field accurate verification is achieved, and the accuracy and timeliness of defect disposal are guaranteed.

Description

Technical Field

[0001] This invention relates to the field of power distribution network equipment condition monitoring technology, specifically to a method and device for online defect monitoring of power distribution network equipment based on multi-feature fusion. Background Technology

[0002] The distribution network is the final link connecting users in the power system, encompassing a large number of distributed devices such as transformers, switchgear, cable heads, and surge arresters. Its operating status directly affects the reliability of power supply. Currently, defect monitoring of distribution network equipment mainly relies on technologies such as sensor acquisition, signal processing, and feature analysis. However, distribution network sites generally face complex conditions such as power frequency electromagnetic interference, environmental noise, and dispersed equipment distribution. Traditional monitoring technologies need to adapt to the needs of scenarios involving multiple types of defects, complex interference, and distributed equipment to improve the accuracy of defect identification and operation and maintenance efficiency. Therefore, they have certain shortcomings in practical applications.

[0003] Firstly, current defect monitoring of power distribution equipment often uses single-mode sensors to collect data. For example, it may only use ultra-high frequency sensors to monitor partial discharge or only use infrared thermal imaging to detect heat generation. Although some technologies use multiple sensors, they only perform simple filtering on the signals. However, the environment of power distribution networks is complex. There is 50Hz power frequency interference and environmental noise around switchgear, cable heads, and other equipment. A single ultra-high frequency signal is prone to characteristic distortion due to interference, and a single infrared thermal imaging may misjudge temperature fluctuations caused by normal heat dissipation of equipment as defects. Even with multi-sensor acquisition, if only general filtering is used, it is difficult to specifically retain the characteristic frequency band of partial discharge and improve the signal-to-noise ratio of the partial discharge source. This results in insufficient basic data quality for subsequent analysis and is prone to missed or misjudged defects.

[0004] Secondly, existing defect diagnosis methods for distribution network equipment often rely on single-feature analysis or simple decision-level splicing and fusion of multi-source features. For example, defects are determined solely based on the temperature characteristics of infrared thermal imaging, or the results of ultrasonic signals and infrared images are displayed side by side. However, distribution network equipment defects are diverse. For instance, partial discharge is accompanied by both ultra-high frequency and ultrasonic signal features, while insulation aging is characterized by abnormal temperature. Single-feature analysis cannot cover the multi-dimensional characteristics of different defects. Simple feature splicing also fails to consider the differences in the contribution of each feature to defect diagnosis. For example, the core feature of overheating cable head joints is abnormal temperature, while the core feature of partial discharge is both ultra-high frequency and ultrasonic signals. This indiscriminate fusion easily amplifies the interference of irrelevant features, leading to inaccurate defect type identification and large location deviations.

[0005] Third, current distribution network equipment defect monitoring mostly uses a single online monitoring system or a single portable field testing device. Although online monitoring systems can achieve wide-area equipment coverage, they are prone to misjudgment when some defect signals are weak due to deployment distance and on-site interference. Portable field testing devices can collect data at close range, but they mostly operate independently and lack linkage with online monitoring data. Maintenance personnel need to recollect all data after arriving on-site, and cannot conduct targeted recollection based on the preliminary results of online monitoring. At the same time, the diagnostic models of existing monitoring systems are mostly deployed after fixed training, lacking continuous iteration with actual on-site defect data. When new defect characteristics appear in distribution network equipment (such as the aging characteristics of new insulation materials), the model is difficult to adapt, resulting in a long-term decline in monitoring accuracy.

[0006] Therefore, it is necessary to design an online monitoring method and device for defects in power distribution network equipment based on multi-feature fusion. Summary of the Invention

[0007] The purpose of this invention is to provide an online defect monitoring method and device for power distribution network equipment based on multi-feature fusion, in order to solve the problems mentioned in the background art, such as insufficient basic data quality and easy omission or misjudgment of defects caused by single-modal data acquisition or simple signal processing in the defect monitoring process of power distribution network equipment; inaccurate defect type identification and large location deviation caused by single feature analysis or simple feature splicing and fusion; inefficient operation and maintenance due to lack of data linkage between single online monitoring system and independent portable detection device; and the inability of diagnostic model to continuously iterate and adapt to new defect features based on actual field data, resulting in long-term decline in monitoring accuracy.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] Firstly, a method for online defect monitoring of power distribution network equipment based on multi-feature fusion is provided, including the following steps:

[0010] S1: Multi-source data acquisition: Sensor units containing multiple types of sensor components are deployed in key monitoring areas of power distribution network equipment to simultaneously acquire multi-source data related to equipment defects: UHF electrical signals, infrared thermal imaging, partial discharge ultrasonic signals, and visible light images. The selection of multiple types of sensor components, including UHF, infrared, ultrasonic, and visible light, is based on the signal characteristics of different defects in power distribution network equipment: UHF signals correspond to the electrical characteristics of partial discharge, infrared thermal imaging corresponds to heating defects, ultrasonic signals correspond to the mechanical wave characteristics of partial discharge, and visible light images provide a reference for the appearance of the equipment. Multi-source data can achieve complementarity of defect characteristics and avoid misjudgment based on a single signal.

[0011] S2: Signal Filtering and Enhancement Processing: Noise suppression algorithms are used to remove interference noise from the UHF electrical signals and partial discharge ultrasonic signals acquired by S1, while retaining the characteristic frequency bands of partial discharge. The ultrasonic signals are further enhanced by spatial domain enhancement algorithms to improve the signal-to-noise ratio in the direction of the partial discharge sound source. Denoising and contrast enhancement processing are performed on the infrared thermal imaging and visible light images acquired by S1 to improve the clarity of visual features. There is electromagnetic and environmental noise interference in the distribution network site, and the signal-to-noise ratio of the directly acquired signals is low. Targeted filtering and enhancement processing can purify the signals, completely retain the defect features, and provide a reliable data foundation for subsequent fusion analysis.

[0012] S3: Multi-feature fusion analysis: Based on the multi-source data processed by S2, multi-dimensional features are fused through a three-level feature fusion algorithm to generate an integrated diagnostic map. The integrated diagnostic map clarifies the defect information. The three-level feature fusion adopts a progressive logic of preliminary positioning, precise positioning, and comprehensive diagnosis. Compared with single feature analysis, it can avoid the limitations of a single dimension and improve the accuracy and reliability of defect diagnosis.

[0013] S4: Online Early Warning and Data Storage: Transmits the defect information and integrated diagnostic map output by S3 to the online monitoring host, triggers graded early warnings according to the severity of defects, and stores the relevant data from S1 to S3 together in the database. Graded early warning enables differentiated responses to defects, prioritizing the handling of high-risk hazards; data storage links and archives information throughout the entire process, supporting defect tracing and accumulating data for intelligent model iteration.

[0014] S5: On-site Diagnosis Verification and Defect Elimination: Maintenance personnel carry portable diagnostic devices to the site to recollect multi-source data on defect locations. The data is then processed by the device's built-in fusion algorithm to generate on-site diagnostic results. The on-site diagnostic results are compared with the warning results from S4 to verify the accuracy of the defect information and generate a defect elimination plan. Online monitoring enables large-scale initial screening, while the portable device completes precise on-site verification. The collaboration of the two ensures monitoring coverage and avoids misjudgments from online monitoring, thereby improving the accuracy of defect handling.

[0015] As a further technical solution of the present invention, in S1, the multi-type sensing components include ultra-high frequency sensors, infrared thermal imaging sensors, ultrasonic sensor arrays, and visible light cameras. The sensing components are deployed one-to-one with the key monitoring areas of the power distribution network equipment. Among them, the ultrasonic sensor array for online monitoring is a 128-channel sensor array used to collect partial discharge ultrasonic signals. The deployment of the sensing components with the key monitoring areas of the equipment is designed based on the locations of typical defects in the power distribution network equipment, which can improve the targeting of signal acquisition and ensure effective capture of defect characteristics.

[0016] As a further technical solution of the present invention, the ultra-high frequency electrical signal is acquired by an ultra-high frequency sensor, with an acquisition frequency covering 300MHz-3GHz, retaining the power frequency repeatability characteristics of the signal peak and phase; the infrared thermal image is acquired by an infrared thermal imaging sensor, with a temperature measurement range of -20℃-150℃ and a resolution ≥384×288; the partial discharge ultrasonic signal is acquired by the 128-channel ultrasonic sensor array, with a frequency covering 2kHz-35kHz; the visible light image is acquired by a visible light camera, with a resolution ≥1080P; the acquisition process is carried out synchronously, and the data is transmitted to the data preprocessing module via Ethernet, with a transmission delay of no more than 1 second.

[0017] As a further technical solution of the present invention, in S2, the noise suppression algorithm is an FIR frequency domain filtering algorithm, whose frequency response function satisfies:

[0018]

[0019] in, These are the coefficients of the FIR filter. Let the filter order be . The FIR filter is used to specifically suppress 50Hz power frequency interference and low-frequency noise below 1kHz in the environment. It has linear phase characteristics, which can suppress interference while avoiding signal phase distortion and fully preserving the power frequency repeatability characteristics of the UHF electrical signal, thus ensuring the effectiveness of subsequent feature extraction.

[0020] The spatial enhancement algorithm is a delay-summation beamforming algorithm, and its pattern function satisfies:

[0021]

[0022] in, This represents the number of elements in the ultrasonic sensor array. For the first The received signal of each array element The frequency of the ultrasound signal. For the signal to arrive at the The delay time of each array element For the first The position coordinates of each array element The unit vector is the direction of the sound source. For the azimuth and elevation angles of the sound source, To improve the signal-to-noise ratio by focusing on the direction of the partial discharge source, the algorithm enhances the signal strength in the direction of the partial discharge source through delay compensation and signal summation, suppresses environmental noise, realizes the directional localization of the partial discharge source, and improves the defect directivity of the ultrasonic signal.

[0023] The Gaussian spatial domain denoising algorithm is used for denoising infrared thermal imaging and visible light images, and its filtering formula satisfies:

[0024]

[0025] in, The coordinates of the Gaussian kernel center are: Standard deviation;

[0026] Contrast enhancement processing employs a histogram equalization algorithm, whose transformation function satisfies:

[0027]

[0028] in, The original image grayscale levels, To enhance the gray levels of the post-image, The total number of gray levels in the image. , The number of rows and columns of pixels in the image. grayscale The number of pixels is increased, Gaussian denoising can reduce interference from ambient light and device dirt, and histogram equalization can improve the contrast of weak visual features, making subtle anomalies on the device surface easier to identify.

[0029] As a further technical solution of the present invention, the specific process of the three-level feature fusion algorithm in S3 is as follows:

[0030] The first step is to extract the temperature features from the infrared thermal image after S2 processing, and then identify abnormally high temperature areas using a threshold judgment formula:

[0031]

[0032] in, For pixels in infrared thermal imaging Temperature value, This represents the normal operating temperature of the device corresponding to that pixel. A temperature anomaly threshold is set, a temperature map is generated, and parameters such as the area of ​​the anomaly region, temperature gradient, and highest temperature value are extracted from the map as features. Initially, potential defect areas were identified;

[0033] The second step is to convert the partial discharge ultrasonic signal processed by S2 into an acoustic spectrum using Fourier transform, and the transform formula satisfies:

[0034]

[0035] in, For spatiotemporal ultrasonic signals, For frequency domain acoustic signals, the acoustic spectrum is superimposed on the visible light image processed by S2 according to pixel coordinates to locate the partial discharge source. Parameters such as the energy peak, source location coordinates, and signal duration of the acoustic spectrum are extracted as features of the ultrasound-visible light superimposed spectrum. ;

[0036] The third step is to extract the peak value, phase, and spectral distribution of the UHF electrical signal after S2 processing as characteristics of the UHF electrical signal. ,Will , and The pre-trained intelligent diagnostic model is input together, and multi-dimensional features are fused through an attention mechanism feature fusion algorithm. The fusion formula satisfies:

[0037]

[0038] in, For attention weight coefficients, and The weighting coefficients are obtained through adaptive learning of the intelligent diagnostic model and dynamically allocated according to the contribution of each feature to defect diagnosis, generating an integrated diagnostic map. The three-level fusion is progressive, from initial narrowing of thermal anomalies to ultrasonic visible light location of sound sources, and then to comprehensive diagnosis of multi-dimensional features, which can comprehensively cover the characteristics of different types of defects in distribution network equipment and improve the comprehensiveness of diagnosis.

[0039] As a further technical solution of the present invention, the intelligent diagnostic model is a CNN-LSTM deep learning model. The model structure includes a CNN feature extraction layer and an LSTM temporal analysis layer, and the training process uses the cross-entropy loss function.

[0040]

[0041] in, For the number of defect categories, For the first The true label of class defects For the model to predict the first The probability of a class of defects;

[0042] The training samples cover typical defect types of power distribution network equipment, with no fewer than 500 samples. The defect information includes defect type, precise location, and severity. CNN layers are good at extracting spatial features of multimodal data, while LSTM layers can capture temporal changes in signals. The combination of the two can effectively mine the spatiotemporal correlation of defect signals and adapt to the complex characteristics of defects in power distribution network equipment.

[0043] As a further technical solution of the present invention, in step S4, the data stored in the database includes the original multi-source data of S1, the signal processing results of S2, and the integrated diagnostic map and defect information of S3, and the data is synchronously associated with timestamps when stored. Equipment identification The data storage format adopts JSON structured format, and supports SQL query, historical data backtracking and batch export.

[0044] As a further technical solution of the present invention, in step S5, the portable diagnostic device integrates a miniature infrared sensing module, a 32-channel portable ultrasonic sensing array, and a visible light acquisition module; wherein, the miniature infrared sensing module has a temperature measurement range of -20℃ to 120℃ and a resolution ≥256×192; the portable ultrasonic sensing array has a frequency coverage of 2kHz to 35kHz; and the visible light acquisition module has a resolution ≥720P.

[0045] The device's built-in fusion algorithm is derived from the S3's attention mechanism feature fusion algorithm, with similar weight coefficients. The calculation logic is consistent, ensuring the standardization of on-site diagnosis and online monitoring, and the processing delay of re-collected data does not exceed 30 seconds.

[0046] As a further technical solution of the present invention, it also includes S6: data closure and model optimization: the on-site diagnostic results, defect elimination plan and the status data of the equipment running continuously for 72 hours after defect elimination in S5 are transmitted back to the database to update the defect sample dataset; the CNN-LSTM deep learning model in S3 is iteratively trained based on the new samples to continuously optimize the monitoring accuracy. The data closure realizes the accumulation of information on the entire life cycle of defects. The model is iterated based on actual scenario data, which can make the diagnostic capability of the model continuously improve with application and adapt to the diverse changes of defects in distribution network equipment.

[0047] Secondly, a defect online monitoring device for power distribution network equipment based on multi-feature fusion is provided, including an online monitoring system and a portable diagnostic device. The online monitoring system and the portable diagnostic device achieve bidirectional data communication through a 4G or 5G wireless communication module, and work together to complete the online monitoring and on-site diagnosis of defects in power distribution network equipment.

[0048] The online monitoring system includes:

[0049] The sensing unit, deployed in the key monitoring area of ​​power distribution equipment, includes a UHF sensor, an infrared thermal imaging sensor, a 128-channel ultrasonic sensor array and a visible light camera. It is used to simultaneously collect multi-source data and transmit it to the data preprocessing module in real time. The sensing unit adopts a distributed deployment, which can cover the key areas of multiple devices in the power distribution network, realize the synchronous monitoring of multiple devices, and improve the overall coverage of power distribution equipment monitoring.

[0050] The data preprocessing module is connected to the sensing unit via an Ethernet signal. It has a built-in FIR frequency domain filtering module, a delay-summing beamforming module, and an image enhancement module. It executes the filtering, spatial domain enhancement, and image processing algorithms described in S2, and outputs a clean UHF electrical signal, an enhanced ultrasound signal, a denoised and enhanced infrared thermal image, and a visible light image. The data preprocessing module is the core of signal purification. It converts the original non-standard data into effective data that meets the requirements of fusion analysis, ensuring the reliability of subsequent diagnosis.

[0051] The fusion analysis module is connected to the data preprocessing module via a PCIe 3.0 interface. It incorporates the three-level feature fusion algorithm and CNN-LSTM deep learning model described in S3. It is used to perform multi-dimensional feature fusion on clean signals and clear imaging data, generate an integrated diagnostic map, and output defect information. The fusion analysis module is the core computing unit of the device, integrating algorithms and models to realize the conversion of multi-source data into defect information, supporting online early warning and operation and maintenance decisions.

[0052] The online monitoring host is connected to the fusion analysis module via an industrial bus. It has a built-in hierarchical early warning module for receiving defect information and integrated diagnostic maps. It triggers red and yellow level two early warnings according to the severity of the defects. The red light corresponds to severe defects and the yellow light corresponds to mild to moderate defects. It also realizes data forwarding and command issuance. The online monitoring host is the scheduling core of the system, realizing the reception of diagnostic results, early warning triggering and command flow, thereby improving the response efficiency of power distribution network operation and maintenance.

[0053] The database, connected to the online monitoring host via fiber optic data, has a storage capacity of ≥1TB and adopts a relational database management system. It is used for structured storage of raw multi-source data, signal processing results, integrated diagnostic maps, defect information and associated timestamps and device identifiers. It supports concurrent queries and historical data backtracking. The database is a basic component supporting data, providing a basis for defect tracing and data analysis, and accumulating data for intelligent model iteration to achieve data-driven operation and maintenance.

[0054] The portable diagnostic device includes:

[0055] The miniature sensing module integrates a miniature infrared sensing module, a 32-channel portable ultrasonic sensing array, and a visible light acquisition module. It is used for on-site re-collection of multi-source data on defect locations. The miniature sensing module adopts a miniaturized design, which ensures data validity while adapting to complex on-site operating environments such as poles and narrow switch cabinets, improving operational convenience.

[0056] The embedded processing module is connected to the micro-sensing module via an SPI interface signal. It is equipped with an ARM Cortex-A9 processor and the attention mechanism feature fusion algorithm described in S3. It is used to process the re-collected data in real time and generate on-site diagnostic results. The embedded processing module has on-site computing capabilities and can complete data processing and diagnosis offline. It is suitable for on-site scenarios without network and improves the applicability of the device.

[0057] The interaction and communication module, electrically connected to the embedded processing module, includes a 7-inch touch screen and a 4G or 5G communication unit. It is used to display on-site diagnostic results and defect maps, and supports data interaction with the online monitoring host to realize the comparison and verification of on-site diagnostic results and early warning results. The interaction module realizes the intuitive display of diagnostic results, and the communication unit ensures data synchronization between the on-site and online systems, supporting on-site decision-making and closed-loop verification.

[0058] The power module supplies power to all modules of the portable diagnostic device. It uses a 10000mAh rechargeable lithium battery to meet the needs of long-term field operation and ensure the device's continuous operation in outdoor environments without external power supply.

[0059] Compared with existing technologies, the beneficial effects of this online defect monitoring method and device for power distribution network equipment based on multi-feature fusion are:

[0060] Sensing units, including UHF sensors, infrared thermal imaging sensors, ultrasonic sensor arrays, and visible light cameras, are deployed in key monitoring areas of power distribution network equipment. These units simultaneously acquire UHF electrical signals, infrared thermal imaging, partial discharge ultrasonic signals, and visible light images. Multimodal data complementarity covers the core characteristics of different types of defects. Furthermore, differentiated processing techniques are employed for different signal characteristics. An FIR frequency domain filtering algorithm suppresses interference noise in UHF electrical signals and partial discharge ultrasonic signals while preserving the characteristic frequency bands of partial discharge. A delay-summation beamforming algorithm improves the signal-to-noise ratio in the direction of the partial discharge source. Gaussian spatial domain denoising and histogram equalization algorithms optimize the visual feature clarity of infrared thermal imaging and visible light images. Compared to single-modal acquisition and simple processing methods, this approach effectively avoids the limitations of single signals, providing high-quality and highly reliable basic data support for subsequent fusion analysis.

[0061] A three-level feature fusion analysis logic is constructed. First, the temperature features of the infrared thermal image after S2 processing are extracted. Then, abnormally high temperature areas are identified and temperature spectral features are generated through a threshold judgment formula. The potential defect area was initially identified; then the processed partial discharge ultrasonic signal was converted into an acoustic spectrum through Fourier transform, and the ultrasonic-visible superimposed spectrum features were extracted after being superimposed with the visible light image. This enables precise localization of the partial discharge source; finally, , and extracted UHF electrical signal characteristics By inputting a pre-trained CNN-LSTM deep learning model, multi-dimensional features are fused through an attention mechanism feature fusion algorithm to generate an integrated diagnostic map. Compared with single feature judgment or simple fusion methods, the progressive judgment logic can comprehensively cover the feature differences of typical defects such as partial discharge, insulation aging, and joint overheating. Through the adaptive allocation of the weights of each feature, multi-dimensional information is deeply mined, which effectively improves the accuracy of defect type identification, precise location positioning, and severity determination.

[0062] An architecture was constructed to facilitate collaborative operation between an online monitoring system and a portable diagnostic device. The online monitoring system, through sensing units, a data preprocessing module, a fusion analysis module, an online monitoring host, and a database, achieves large-scale, real-time defect monitoring and graded early warning. The portable diagnostic device integrates a miniature infrared sensing module, a portable ultrasonic sensing array, and a visible light acquisition module. It incorporates an attention mechanism feature fusion algorithm derived from the fusion analysis module, enabling on-site verification of early warning defects to ensure the accuracy of defect information and generate targeted defect elimination plans. Simultaneously, through an S6 data closed-loop design, on-site diagnostic results, defect elimination plans, and post-defect equipment operation data are fed back to the database to update the defect sample dataset. Based on the newly added samples, the CNN-LSTM deep learning model is iteratively trained. Compared to a single monitoring or detection mode, this approach achieves efficient collaboration between large-scale online initial screening and precise on-site verification, ensuring the accuracy and timeliness of defect handling. Furthermore, continuous model optimization dynamically improves the system's monitoring capabilities, adapting to the diverse changes in the defect characteristics of distribution network equipment. Attached Figure Description

[0063] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0064] Figure 2 This is a schematic diagram of the fusion and visualization results of partial discharge ultrasound and visible light multi-source imaging.

[0065] Figure 3 This is a schematic diagram illustrating the principle of infrared thermal imaging acquisition.

[0066] Figure 4 A comparison of waveforms before and after frequency domain filtering of ultra-high frequency electrical signals and partial discharge ultrasonic signals;

[0067] Figure 5 The waveform of the partial discharge ultrasonic signal after spatial enhancement and the comparison of acoustic patterns before and after enhancement are shown in the diagram. Detailed Implementation

[0068] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0069] Please see the appendix Figure 1 -Appendix Figure 5 The present invention provides an embodiment 1: an online defect monitoring method for power distribution network equipment based on multi-feature fusion, comprising the following steps:

[0070] S1: Multi-source data acquisition: Sensor units containing multiple types of sensor components are deployed in key monitoring areas of power distribution network equipment to simultaneously acquire multi-source data related to equipment defects: UHF electrical signals, infrared thermal imaging, partial discharge ultrasonic signals, and visible light images. The selection of multiple types of sensor components, including UHF, infrared, ultrasonic, and visible light, is based on the signal characteristics of different defects in power distribution network equipment: UHF signals correspond to the electrical characteristics of partial discharge, infrared thermal imaging corresponds to heating defects, ultrasonic signals correspond to the mechanical wave characteristics of partial discharge, and visible light images provide a reference for the appearance of the equipment. Multi-source data can achieve complementarity of defect characteristics and avoid misjudgment based on a single signal.

[0071] The system incorporates various sensing components, including UHF sensors, infrared thermal imaging sensors, ultrasonic sensor arrays, and visible light cameras. These components are deployed one-to-one with the key monitoring areas of the power distribution network equipment. The online monitoring ultrasonic sensor array is a 128-channel array used to acquire partial discharge ultrasonic signals. This deployment, corresponding to the key monitoring areas of the equipment, is designed based on the locations of typical defects in the power distribution network equipment, enhancing the targeting of signal acquisition and ensuring effective capture of defect characteristics. UHF electrical signals are acquired by UHF sensors, covering a frequency range of 300MHz-3GHz, preserving the power frequency repeatability characteristics of signal peaks and phases. Infrared thermal imaging is acquired by infrared thermal imaging sensors, with a temperature measurement range of -20℃ to 150℃ and a resolution ≥384×288. Partial discharge ultrasonic signals are acquired by a 128-channel ultrasonic sensor array, covering a frequency range of 2kHz-35kHz. Visible light images are acquired by a visible light camera with a resolution ≥1080P. The acquisition process is synchronized, with data transmitted via Ethernet to the data preprocessing module, with a transmission delay of no more than 1 second.

[0072] S2: Signal Filtering and Enhancement Processing: Noise suppression algorithms are used to remove interference noise from the UHF electrical signals and partial discharge ultrasonic signals acquired by S1, while retaining the characteristic frequency bands of partial discharge. The ultrasonic signals are further enhanced by spatial domain enhancement algorithms to improve the signal-to-noise ratio in the direction of the partial discharge sound source. Denoising and contrast enhancement processing are performed on the infrared thermal imaging and visible light images acquired by S1 to improve the clarity of visual features. There is electromagnetic and environmental noise interference in the distribution network site, and the signal-to-noise ratio of the directly acquired signals is low. Targeted filtering and enhancement processing can purify the signals, completely retain the defect features, and provide a reliable data foundation for subsequent fusion analysis.

[0073] The noise suppression algorithm is an FIR frequency domain filtering algorithm, and its frequency response function satisfies:

[0074]

[0075] in, These are the coefficients of the FIR filter. Let the filter order be . The FIR filter is used to specifically suppress 50Hz power frequency interference and low-frequency noise below 1kHz in the environment. It has linear phase characteristics, which can suppress interference while avoiding signal phase distortion and fully preserving the power frequency repeatability characteristics of the UHF electrical signal, thus ensuring the effectiveness of subsequent feature extraction.

[0076] The spatial enhancement algorithm is a delay-summation beamforming algorithm, and its pattern function satisfies:

[0077]

[0078] in, This represents the number of elements in the ultrasonic sensor array. For the first The received signal of each array element The frequency of the ultrasound signal. For the signal to arrive at the The delay time of each array element For the first The position coordinates of each array element The unit vector is the direction of the sound source. For the azimuth and elevation angles of the sound source, To improve the signal-to-noise ratio by focusing on the direction of the partial discharge source, the algorithm enhances the signal strength in the direction of the partial discharge source through delay compensation and signal summation, suppresses environmental noise, realizes the directional localization of the partial discharge source, and improves the defect directivity of the ultrasonic signal.

[0079] The Gaussian spatial domain denoising algorithm is used for denoising infrared thermal imaging and visible light images, and its filtering formula satisfies:

[0080]

[0081] in, The coordinates of the Gaussian kernel center are: Standard deviation;

[0082] Contrast enhancement processing employs a histogram equalization algorithm, whose transformation function satisfies:

[0083]

[0084] in, The original image grayscale levels, To enhance the gray levels of the post-image, The total number of gray levels in the image. , The number of rows and columns of pixels in the image. grayscale The number of pixels, Gaussian noise reduction can reduce the interference of ambient light and equipment dirt, and histogram equalization can improve the contrast of weak visual features, making it easier to identify subtle abnormalities on the surface of the equipment.

[0085] S3: Multi-feature fusion analysis: Based on the multi-source data processed by S2, multi-dimensional features are fused through a three-level feature fusion algorithm to generate an integrated diagnostic map. The integrated diagnostic map clarifies the defect information. The three-level feature fusion adopts a progressive logic of preliminary positioning, precise positioning, and comprehensive diagnosis. Compared with single feature analysis, it can avoid the limitations of a single dimension and improve the accuracy and reliability of defect diagnosis.

[0086] The specific process of the three-level feature fusion algorithm is as follows:

[0087] The first step is to extract the temperature features from the infrared thermal image after S2 processing, and then identify abnormally high temperature areas using a threshold judgment formula:

[0088]

[0089] in, For pixels in infrared thermal imaging Temperature value, This represents the normal operating temperature of the device corresponding to that pixel. A temperature anomaly threshold is set, a temperature map is generated, and parameters such as the area of ​​the anomaly region, temperature gradient, and highest temperature value are extracted from the map as features. Initially, potential defect areas were identified;

[0090] The second step is to convert the partial discharge ultrasonic signal processed by S2 into an acoustic spectrum using Fourier transform, and the transform formula satisfies:

[0091]

[0092] in, For spatiotemporal ultrasonic signals, For frequency domain acoustic signals, the acoustic spectrum is superimposed on the visible light image processed by S2 according to pixel coordinates to locate the partial discharge source. Parameters such as the energy peak, source location coordinates, and signal duration of the acoustic spectrum are extracted as features of the ultrasound-visible light superimposed spectrum. ;

[0093] The third step is to extract the peak value, phase, and spectral distribution of the UHF electrical signal after S2 processing as characteristics of the UHF electrical signal. ,Will , and The pre-trained intelligent diagnostic model is input together, and multi-dimensional features are fused through an attention mechanism feature fusion algorithm. The fusion formula satisfies:

[0094]

[0095] in, For attention weight coefficients, and The weighting coefficients are obtained through adaptive learning of the intelligent diagnostic model and dynamically allocated according to the contribution of each feature to defect diagnosis. An integrated diagnostic map is generated, and the three-level fusion is progressive, from the initial narrowing of thermal anomalies to the location of sound sources by ultrasound and visible light, and then to the comprehensive diagnosis of multi-dimensional features. It can fully cover the features of different types of defects in distribution network equipment and improve the comprehensiveness of diagnosis.

[0096] The intelligent diagnostic model is a CNN-LSTM deep learning model. The model structure includes a CNN feature extraction layer and an LSTM temporal analysis layer. The training process uses the cross-entropy loss function.

[0097]

[0098] in, For the number of defect categories, For the first The true label of class defects For the model to predict the first The probability of a class of defects;

[0099] The training samples cover typical defect types of power distribution network equipment, with no fewer than 500 samples. The defect information includes defect type, precise location, and severity. CNN layers are good at extracting spatial features of multimodal data, while LSTM layers can capture temporal changes in signals. The combination of the two can effectively mine the spatiotemporal correlation of defect signals and adapt to the complex characteristics of defects in power distribution network equipment.

[0100] S4: Online Early Warning and Data Storage: Transmits the defect information and integrated diagnostic map output by S3 to the online monitoring host, triggers graded early warnings according to the severity of defects, and stores the relevant data from S1 to S3 together in the database. Graded early warning enables differentiated responses to defects, prioritizing the handling of high-risk hazards; data storage links and archives information throughout the entire process, supporting defect tracing and accumulating data for intelligent model iteration.

[0101] The data stored in the database includes the raw multi-source data of S1, the signal processing results of S2, and the integrated diagnostic map and defect information of S3, and the data is stored synchronously with timestamps. Equipment identification The data storage format adopts JSON structured format, which supports SQL query, historical data backtracking and batch export;

[0102] S5: On-site Diagnosis Verification and Defect Removal: Maintenance personnel carry portable diagnostic devices to the site to collect multi-source data on defect locations. The data is then processed by the device's built-in fusion algorithm to generate on-site diagnostic results. The on-site diagnostic results are compared with the warning results from S4 to verify the accuracy of the defect information and generate a defect removal plan. Online monitoring enables large-scale initial screening, while the portable device completes precise on-site verification. The collaboration of the two ensures monitoring coverage and avoids misjudgments from online monitoring, thereby improving the accuracy of defect handling.

[0103] The portable diagnostic device integrates a miniature infrared sensing module, a 32-channel portable ultrasound sensing array, and a visible light acquisition module. The miniature infrared sensing module has a temperature measurement range of -20℃ to 120℃ and a resolution of ≥256×192. The portable ultrasound sensing array covers a frequency range of 2kHz to 35kHz. The visible light acquisition module has a resolution of ≥720P.

[0104] The device's built-in fusion algorithm is derived from the S3's attention mechanism feature fusion algorithm, with similar weight coefficients. The calculation logic is consistent to ensure the standardization of on-site diagnosis and online monitoring, and the processing delay of re-collected data does not exceed 30 seconds;

[0105] It also includes S6: Data closed loop and model optimization: The on-site diagnostic results, defect elimination plan and the status data of the equipment running continuously for 72 hours after defect elimination in S5 are sent back to the database to update the defect sample dataset; Based on the new samples, the CNN-LSTM deep learning model of S3 is iteratively trained to continuously optimize the monitoring accuracy. The data closed loop realizes the accumulation of information throughout the entire defect life cycle. Based on the actual scenario data, the model is iterated, which can make the diagnostic capability of the model continuously improve with the application and adapt to the diverse changes of defects in distribution network equipment.

[0106] An embodiment 2 of the present invention provides an online monitoring device for defects in power distribution network equipment based on multi-feature fusion, comprising an online monitoring system and a portable diagnostic device. The online monitoring system and the portable diagnostic device achieve bidirectional data communication through a 4G or 5G wireless communication module, and collaboratively complete the online monitoring and on-site diagnosis of defects in power distribution network equipment.

[0107] The online monitoring system includes:

[0108] The sensing unit, deployed in the key monitoring area of ​​power distribution equipment, includes a UHF sensor, an infrared thermal imaging sensor, a 128-channel ultrasonic sensor array and a visible light camera. It is used to simultaneously collect multi-source data and transmit it to the data preprocessing module in real time. The sensing unit adopts a distributed deployment, which can cover the key areas of multiple devices in the power distribution network, realize the synchronous monitoring of multiple devices, and improve the overall coverage of power distribution equipment monitoring.

[0109] The data preprocessing module is connected to the sensing unit via Ethernet signal. It has built-in FIR frequency domain filtering module, delay-summing beamforming module and image enhancement module, which respectively execute S2 filtering, spatial domain enhancement and image processing algorithms, and output clean UHF electrical signal, enhanced ultrasound signal, denoised and enhanced infrared thermal image and visible light image. The data preprocessing module is the core link of signal purification, which converts the original non-standard data into effective data that meets the requirements of fusion analysis, and ensures the reliability of subsequent diagnosis.

[0110] The fusion analysis module is connected to the data preprocessing module via a PCIe 3.0 interface. It incorporates the S3 three-level feature fusion algorithm and the CNN-LSTM deep learning model to perform multi-dimensional feature fusion on clean signals and clear imaging data, generate an integrated diagnostic map and output defect information. The fusion analysis module is the core computing unit of the device, integrating algorithms and models to realize the conversion of multi-source data into defect information, supporting online early warning and operation and maintenance decisions.

[0111] The online monitoring host is connected to the integrated analysis module via an industrial bus. It has a built-in hierarchical early warning module to receive defect information and integrated diagnostic maps. It triggers red and yellow level warnings according to the severity of the defects. Red lights correspond to severe defects and yellow lights correspond to mild to moderate defects. It also realizes data forwarding and command issuance. The online monitoring host is the scheduling core of the system, realizing the reception of diagnostic results, early warning triggering and command flow, and improving the response efficiency of power distribution network operation and maintenance.

[0112] The database, connected to the online monitoring host via fiber optic data, has a storage capacity of ≥1TB and adopts a relational database management system. It is used for structured storage of raw multi-source data, signal processing results, integrated diagnostic maps, defect information and associated timestamps and device identifiers. It supports concurrent queries and historical data backtracking. The database is the basic component supporting the data, providing a basis for defect tracing and data analysis, and accumulating data for intelligent model iteration to achieve data-driven operation and maintenance.

[0113] Portable diagnostic devices include:

[0114] The miniature sensing module integrates a miniature infrared sensing module, a 32-channel portable ultrasonic sensing array, and a visible light acquisition module. It is used for on-site re-collection of multi-source data on defect locations. The miniature sensing module adopts a miniaturized design, which ensures data validity while adapting to complex on-site operating environments such as poles and narrow switch cabinets, improving operational convenience.

[0115] The embedded processing module is connected to the micro-sensing module via an SPI interface signal. It is equipped with an ARM Cortex-A9 processor and an S3 attention mechanism feature fusion algorithm to process the re-collected data in real time and generate on-site diagnostic results. The embedded processing module has on-site computing capabilities and can complete data processing and diagnosis offline, adapting to on-site scenarios without network coverage and improving the applicability of the device.

[0116] The interaction and communication module, electrically connected to the embedded processing module, includes a 7-inch touch screen and a 4G or 5G communication unit. It is used to display on-site diagnostic results and defect maps, and supports data interaction with the online monitoring host to realize the comparison and verification of on-site diagnostic results and early warning results. The interaction module realizes the intuitive display of diagnostic results, and the communication unit ensures data synchronization between the on-site and online systems, supporting on-site decision-making and closed-loop verification.

[0117] The power module supplies power to all modules of the portable diagnostic device. It uses a 10000mAh rechargeable lithium battery to meet the needs of long-term field operation and ensure the device's continuous operation in outdoor environments without external power supply.

[0118] In summary, this invention deploys sensing units containing UHF sensors, infrared thermal imaging sensors, ultrasonic sensor arrays, and visible light cameras in key monitoring areas of power distribution network equipment. These units simultaneously acquire UHF electrical signals, infrared thermal imaging, partial discharge ultrasonic signals, and visible light images, providing complementary multimodal data to cover the core features of different types of defects. Furthermore, differentiated processing techniques are employed for different signal characteristics. An FIR frequency domain filtering algorithm suppresses interference noise in UHF electrical signals and partial discharge ultrasonic signals while preserving the characteristic frequency bands of partial discharge. A delay-summation beamforming algorithm improves the signal-to-noise ratio in the direction of the partial discharge source. Gaussian spatial domain denoising and histogram equalization algorithms optimize the visual feature clarity of infrared thermal imaging and visible light images. Compared to single-modal acquisition and simple processing methods, this invention effectively avoids the limitations of single signals, providing high-quality and highly reliable basic data support for subsequent fusion analysis.

[0119] A three-level feature fusion analysis logic is constructed. First, the temperature features of the infrared thermal image after S2 processing are extracted. Then, abnormally high temperature areas are identified and temperature spectral features are generated through a threshold judgment formula. The potential defect area was initially identified; then the processed partial discharge ultrasonic signal was converted into an acoustic spectrum through Fourier transform, and the ultrasonic-visible superimposed spectrum features were extracted after being superimposed with the visible light image. This enables precise localization of the partial discharge source; finally, , and extracted UHF electrical signal characteristics By inputting a pre-trained CNN-LSTM deep learning model, multi-dimensional features are fused through an attention mechanism feature fusion algorithm to generate an integrated diagnostic map. Compared with single feature judgment or simple fusion methods, the progressive judgment logic can comprehensively cover the feature differences of typical defects such as partial discharge, insulation aging, and joint overheating. Through the adaptive allocation of the weights of each feature, multi-dimensional information is deeply mined, which effectively improves the accuracy of defect type identification, precise location positioning, and severity determination.

[0120] An architecture was constructed to facilitate collaborative operation between an online monitoring system and a portable diagnostic device. The online monitoring system, through sensing units, a data preprocessing module, a fusion analysis module, an online monitoring host, and a database, achieves large-scale, real-time defect monitoring and graded early warning. The portable diagnostic device integrates a miniature infrared sensing module, a portable ultrasonic sensing array, and a visible light acquisition module. It incorporates an attention mechanism feature fusion algorithm derived from the fusion analysis module, enabling on-site verification of early warning defects to ensure the accuracy of defect information and generate targeted defect elimination plans. Simultaneously, through an S6 data closed-loop design, on-site diagnostic results, defect elimination plans, and post-defect equipment operation data are fed back to the database to update the defect sample dataset. Based on the newly added samples, the CNN-LSTM deep learning model is iteratively trained. Compared to a single monitoring or detection mode, this approach achieves efficient collaboration between large-scale online initial screening and precise on-site verification, ensuring the accuracy and timeliness of defect handling. Furthermore, continuous model optimization dynamically improves the system's monitoring capabilities, adapting to the diverse changes in the defect characteristics of distribution network equipment.

[0121] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A method for online defect monitoring of power distribution network equipment based on multi-feature fusion, characterized in that: Includes the following steps: S1: Multi-source data acquisition: Deploy sensing units containing multiple types of sensing components in key monitoring areas of power distribution network equipment to simultaneously acquire multi-source data related to equipment defects: ultra-high frequency electrical signals, infrared thermal imaging, partial discharge ultrasonic signals, and visible light images; S2: Signal filtering and enhancement processing: The UHF electrical signal and partial discharge ultrasonic signal acquired by S1 are processed by noise suppression algorithm to remove interference noise and retain the characteristic frequency band of partial discharge. The signal-to-noise ratio of the ultrasonic signal in the direction of the partial discharge source is further improved by the spatial enhancement algorithm; the infrared thermal imaging and visible light images acquired by S1 are subjected to denoising and contrast enhancement processing to improve the clarity of visual features. S3: Multi-feature fusion analysis: Based on the multi-source data processed by S2, multi-dimensional features are fused through a three-level feature fusion algorithm to generate an integrated diagnostic map, and the defect information is clarified through the integrated diagnostic map. S4: Online Early Warning and Data Storage: Transmits the defect information and integrated diagnostic map output by S3 to the online monitoring host, triggers graded early warnings according to the severity of defects, and stores the relevant data from S1 to S3 together in the database; S5: On-site diagnosis, verification and defect elimination: Maintenance personnel carry portable diagnostic devices to the site, recollect multi-source data of defect locations, and generate on-site diagnostic results through the device's built-in fusion algorithm. Compare the on-site diagnostic results with the S4 warning results to verify the accuracy of the defect information and generate a defect elimination plan.

2. The online defect monitoring method for distribution network power equipment based on multi-feature fusion according to claim 1, characterized in that: In S1, the various types of sensing components include ultra-high frequency sensors, infrared thermal imaging sensors, ultrasonic sensor arrays, and visible light cameras. The sensing components are deployed one-to-one with the key monitoring areas of the power distribution network equipment. Among them, the ultrasonic sensor array for online monitoring is a 128-channel sensor array used to collect partial discharge ultrasonic signals.

3. The online defect monitoring method for distribution network power equipment based on multi-feature fusion according to claim 2, characterized in that: The ultra-high frequency electrical signal is acquired by an ultra-high frequency sensor, with a acquisition frequency covering 300MHz-3GHz, retaining the power frequency repeatability characteristics of the signal peak and phase; the infrared thermal image is acquired by an infrared thermal imaging sensor, with a temperature measurement range of -20℃-150℃ and a resolution ≥384×288; the partial discharge ultrasonic signal is acquired by the 128-channel ultrasonic sensor array, with a frequency covering 2kHz-35kHz; the visible light image is acquired by a visible light camera, with a resolution ≥1080P; the acquisition process is performed synchronously, and the data is transmitted to the data preprocessing module via Ethernet, with a transmission delay of no more than 1 second.

4. The online defect monitoring method for distribution network power equipment based on multi-feature fusion according to claim 1, characterized in that: In step S2, the noise suppression algorithm is an FIR frequency domain filtering algorithm, and its frequency response function satisfies: in, These are the coefficients of the FIR filter. Let the filter order be . It is the angular frequency, used to specifically suppress 50Hz power frequency interference and low-frequency environmental noise below 1kHz. The spatial enhancement algorithm is a delay-summation beamforming algorithm, and its pattern function satisfies: in, This represents the number of elements in the ultrasonic sensor array. For the first The received signal of each array element The frequency of the ultrasound signal. For the signal to arrive at the The delay time of each array element For the first The position coordinates of each array element The unit vector is the direction of the sound source. For the azimuth and elevation angles of the sound source, The signal-to-noise ratio is improved by focusing the direction of the partial discharge sound source, which is the speed of sound. The Gaussian spatial domain denoising algorithm is used for denoising infrared thermal imaging and visible light images, and its filtering formula satisfies: in, The coordinates of the Gaussian kernel center are: Standard deviation; Contrast enhancement processing employs a histogram equalization algorithm, whose transformation function satisfies: in, The original image grayscale levels, To enhance the gray levels of the post-image, The total number of gray levels in the image. , The number of rows and columns of pixels in the image. grayscale The number of pixels.

5. The online defect monitoring method for distribution network power equipment based on multi-feature fusion according to claim 1, characterized in that: In S3, the specific process of the three-level feature fusion algorithm is as follows: The first step is to extract the temperature features from the infrared thermal image after S2 processing, and then identify abnormally high temperature areas using a threshold judgment formula: in, For pixels in infrared thermal imaging Temperature value, This represents the normal operating temperature of the device corresponding to that pixel. A temperature anomaly threshold is set, a temperature map is generated, and parameters such as the area of ​​the anomaly region, temperature gradient, and highest temperature value are extracted from the map as features. Initially, potential defect areas were identified; The second step is to convert the partial discharge ultrasonic signal processed by S2 into an acoustic spectrum using Fourier transform, and the transform formula satisfies: in, For spatiotemporal ultrasonic signals, For frequency domain acoustic signals, the acoustic spectrum is superimposed on the visible light image processed by S2 according to pixel coordinates to locate the partial discharge source. Parameters such as the energy peak, source location coordinates, and signal duration of the acoustic spectrum are extracted as features of the ultrasound-visible light superimposed spectrum. ; The third step is to extract the peak value, phase, and spectral distribution of the UHF electrical signal after S2 processing as characteristics of the UHF electrical signal. ,Will , and The pre-trained intelligent diagnostic model is input together, and multi-dimensional features are fused through an attention mechanism feature fusion algorithm. The fusion formula satisfies: in, Here are the attention weight coefficients, and The weighting coefficients are obtained through adaptive learning of the intelligent diagnostic model and dynamically allocated according to the contribution of each feature to defect diagnosis, generating an integrated diagnostic map.

6. The online defect monitoring method for distribution network power equipment based on multi-feature fusion according to claim 5, characterized in that: The intelligent diagnostic model is a CNN-LSTM deep learning model. The model structure includes a CNN feature extraction layer and an LSTM temporal analysis layer. The training process uses the cross-entropy loss function. in, For the number of defect categories, For the first The true label of class defects For the model to predict the first The probability of a class of defects; The training samples cover typical defect types of power distribution network equipment, with no fewer than 500 samples; the defect information includes defect type, precise location, and severity.

7. The online defect monitoring method for distribution network power equipment based on multi-feature fusion according to claim 1, characterized in that: In step S4, the data stored in the database includes the raw multi-source data from S1, the signal processing results from S2, and the integrated diagnostic map and defect information from S3, and the data is stored synchronously with a timestamp. Equipment Identification The data storage format adopts JSON structured format, and supports SQL query, historical data backtracking and batch export.

8. The online defect monitoring method for distribution network power equipment based on multi-feature fusion according to claim 1, characterized in that: In step S5, the portable diagnostic device integrates a miniature infrared sensing module, a 32-channel portable ultrasonic sensing array, and a visible light acquisition module; wherein, the miniature infrared sensing module has a temperature measurement range of -20℃ to 120℃ and a resolution of ≥256×192; the portable ultrasonic sensing array has a frequency coverage of 2kHz to 35kHz; and the visible light acquisition module has a resolution of ≥720P. The device's built-in fusion algorithm is derived from the S3's attention mechanism feature fusion algorithm, with similar weight coefficients. The calculation logic is consistent, ensuring the standardization of on-site diagnosis and online monitoring, and the processing delay of re-collected data does not exceed 30 seconds.

9. The online defect monitoring method for distribution network power equipment based on multi-feature fusion according to claim 1, characterized in that: It also includes S6: Data closure and model optimization: The on-site diagnostic results, defect elimination plan and the status data of the equipment running continuously for 72 hours after defect elimination in S5 are sent back to the database to update the defect sample dataset; the CNN-LSTM deep learning model of S3 is iteratively trained based on the new samples to continuously optimize the monitoring accuracy.

10. An online defect monitoring device for power distribution network equipment based on multi-feature fusion, characterized in that: It includes an online monitoring system and a portable diagnostic device. The online monitoring system and the portable diagnostic device achieve bidirectional data communication through a 4G or 5G wireless communication module, and work together to complete the online monitoring and on-site diagnosis of defects in power distribution network equipment. The online monitoring system includes: The sensing unit, deployed in the key monitoring area of ​​the power distribution network equipment, includes an ultra-high frequency sensor, an infrared thermal imaging sensor, a 128-channel ultrasonic sensor array, and a visible light camera, which are used to simultaneously collect multi-source data and transmit it to the data preprocessing module in real time. The data preprocessing module is connected to the sensing unit via an Ethernet signal. It has a built-in FIR frequency domain filtering module, a delay-summing beamforming module and an image enhancement module. It executes the filtering, spatial domain enhancement and image processing algorithms described in S2 respectively, and outputs a clean ultra-high frequency electrical signal, an enhanced ultrasonic signal, a denoised and enhanced infrared thermal image and a visible light image. The fusion analysis module is connected to the data preprocessing module via a PCIe 3.0 interface. It incorporates the three-level feature fusion algorithm and CNN-LSTM deep learning model described in S3, which are used to perform multi-dimensional feature fusion on clean signals and clear imaging data, generate an integrated diagnostic map, and output defect information. The online monitoring host is connected to the fusion analysis module via an industrial bus. It has a built-in hierarchical early warning module for receiving defect information and integrated diagnostic maps. It triggers red and yellow level two early warnings according to the severity of the defects. The red light corresponds to severe defects and the yellow light corresponds to mild to moderate defects. It can also forward data and issue instructions. The database is connected to the online monitoring host via fiber optic data connection, with a storage capacity of ≥1TB. It adopts a relational database management system for structured storage of raw multi-source data, signal processing results, integrated diagnostic maps, defect information and associated timestamps and device identifiers, and supports concurrent queries and historical data backtracking. The portable diagnostic device includes: The miniature sensing module integrates a miniature infrared sensing module, a 32-channel portable ultrasonic sensing array, and a visible light acquisition module for on-site re-collection of multi-source data on defect locations. An embedded processing module is connected to the micro-sensing module via an SPI interface signal. It is equipped with an ARM Cortex-A9 processor and the attention mechanism feature fusion algorithm described in S3, and is used to process the re-collected data in real time and generate on-site diagnostic results. The interaction and communication module, electrically connected to the embedded processing module, includes a 7-inch touch screen and a 4G or 5G communication unit, used to display on-site diagnostic results and defect maps, and supports data interaction with the online monitoring host to realize the comparison and verification of on-site diagnostic results and early warning results; The power module supplies power to all modules of the portable diagnostic device and uses a 10000mAh rechargeable lithium battery.

Images