Discharge detection device and method for mine general high-low voltage switch cabinet

Abstract

The application discloses a discharge detection device and method for a general high-low voltage switch cabinet for mines, and belongs to the technical field of switch cabinet detection. The device comprises a modularized sensor assembly, a signal acquisition unit, an edge computing node unit and an upper monitoring platform, and adopts a UHF sensor, an ultrasonic sensor and a transient ground voltage sensor for cooperative detection. The method comprises multi-source signal synchronous acquisition, adaptive denoising based on a sparrow search algorithm optimization, double-domain feature extraction, discharge type identification based on an attention-enhanced convolutional neural network, and multi-sensor information fusion and hierarchical early warning based on evidence theory. The application compensates for the blind area of a single detection method through cooperative work of multiple types of sensors, improves the denoising effect through adaptive parameter optimization, enhances the recognition accuracy through double-domain features and an attention mechanism, and improves the diagnosis reliability through multi-sensor fusion, so that accurate detection and intelligent diagnosis of local discharge of a mine switch cabinet are realized.

Description

Technical Field

[0001] This invention relates to the field of switchgear testing technology, specifically to a discharge detection device and method for general-purpose high and low voltage switchgear used in mining. Background Technology

[0002] Mining-grade high and low voltage switchgear is a core piece of equipment in underground coal mine power distribution systems, undertaking crucial functions of power distribution and line protection. Its operational status directly impacts the safety and reliability of the underground power supply system. During long-term operation, the insulation materials inside the switchgear are subjected to the combined effects of electric field stress, thermal stress, mechanical stress, and environmental factors, gradually leading to aging and deterioration. When insulation defects develop to a certain extent, partial discharge occurs. Partial discharge is a significant indicator of insulation degradation; if not detected and addressed promptly, it can lead to further insulation deterioration and ultimately insulation breakdown, causing serious consequences such as equipment damage, power outages, and even personal injury or death. Therefore, effective partial discharge detection and diagnosis of mining switchgear is of paramount importance for ensuring the safe and stable operation of underground coal mine power supply systems.

[0003] Currently, methods for partial discharge detection in switchgear mainly include ultra-high frequency (UHF) methods, ultrasonic methods, and transient ground voltage methods. Existing technologies typically employ only a single type of sensor for detection. Because single detection methods have blind spots for specific discharge types or installation locations, they struggle to comprehensively perceive various discharge phenomena within the switchgear, resulting in insufficient comprehensiveness and reliability. The underground environment in coal mines is complex and harsh, with strong electromagnetic interference, high humidity, and abundant dust. The anti-interference capability and environmental adaptability of existing sensors need improvement, and safety protection measures for sensors in high-voltage environments also require further refinement. In terms of signal processing, discharge signals often contain various noise components such as power frequency interference, white noise, and random pulse interference, resulting in a low signal-to-noise ratio that directly affects the effectiveness of subsequent feature extraction and pattern recognition. Existing signal denoising methods typically employ fixed-parameter signal decomposition and thresholding, making it difficult to adaptively select optimal processing parameters for discharge signals with different characteristics. This leads to unstable denoising effects and the potential for noise residue or loss of useful signals. In terms of feature extraction, existing methods often extract discharge features from only a single dimension, resulting in limited feature representation capabilities and insufficient separability between different types of discharge in the feature space, affecting the accuracy of discharge type identification. Regarding intelligent recognition, traditional machine learning classification methods rely on manually designed features, leading to complex feature engineering and limited generalization ability. While deep learning methods possess powerful automatic feature learning capabilities, existing network structures still need improvement in their ability to extract features from discharge signals, lacking attention mechanisms tailored to the characteristics of discharge signals, and exhibiting unstable recognition performance under different operating conditions. In terms of diagnostic decision-making, existing methods typically rely on the detection results of a single sensor, failing to effectively utilize the complementary information between multiple sensor types. When a single sensor signal is interfered with or exhibits anomalies, the reliability of the diagnosis decreases significantly.

[0004] Therefore, there is an urgent need for a discharge detection system and method that comprehensively utilizes multi-source detection information and has adaptive signal processing and diagnostic capabilities to improve the sensitivity, accuracy and reliability of discharge detection in mining switchgear. Summary of the Invention

[0005] To address the problems existing in the background technology, this invention provides a discharge detection device for a general-purpose high and low voltage switchgear used in mines, comprising a modular sensor assembly, a signal acquisition unit, an edge computing node unit, and a host monitoring platform. The modular sensor assembly is installed in the switchgear and includes a sensor body, which is an ultra-high frequency sensor, an ultrasonic sensor, or a transient ground voltage sensor. The signal acquisition unit includes a preamplifier, a bandpass filter, a programmable gain amplifier, an anti-aliasing filter, and an analog-to-digital converter connected in sequence. The preamplifier is connected to the modular sensor assembly, and the analog-to-digital converter is connected to the edge computing node unit. The edge computing node unit includes an embedded processor, a data storage module, a communication interface module, and a synchronization trigger module. The embedded processor is connected to the analog-to-digital converter, and the synchronization trigger module is connected to the trigger input terminals of each analog-to-digital converter. The host monitoring platform is connected to the edge computing node unit through the communication interface module.

[0006] In a preferred embodiment, the modular sensor assembly further includes a mounting base and a quick-release interface; the mounting base is located at the bottom of the sensor body and is connected to a magnetically attached base with an embedded permanent magnet array; the quick-release interface is located on the side of the sensor body; the sensor body has a multi-layer shielding structure, consisting of an equal-voltage shielding layer, an insulating isolation layer, and an electromagnetic shielding layer from the outside to the inside.

[0007] In a preferred embodiment, the ultra-high frequency sensor is installed at the ventilation hole on the top of the cabinet or at the joint of the cabinet body; the ultrasonic sensor is attached to the outer wall of the busbar compartment, circuit breaker compartment or cable compartment; and the transient ground voltage sensor is attached to the inner surface of the cabinet door or near the grounding busbar.

[0008] In a preferred embodiment, the control terminal of the programmable gain amplifier is connected to the embedded processor; the synchronization trigger module is provided with a power frequency voltage sampling input port, and generates a synchronization trigger pulse by detecting the zero-crossing point of the power frequency voltage; the communication interface module is connected to the embedded processor.

[0009] In a preferred embodiment, the modular sensor assembly is connected to the signal acquisition unit via a shielded signal cable; the shielded signal cable has a double-layer shielding structure, including an inner aluminum foil shielding layer and an outer metal braided shielding layer; both ends of the shielded signal cable are provided with aviation plugs that mate with the quick-release interface.

[0010] This invention also provides a discharge detection method, including: S1, multi-source signal acquisition: using a UHF sensor, an ultrasonic sensor, and a transient ground voltage sensor, with the zero-crossing point of the power frequency voltage as the synchronization reference, discharge signals are synchronously acquired, and multi-channel raw discharge signals are obtained after signal conditioning and analog-to-digital conversion; S2, adaptive signal denoising: using a sparrow search algorithm to determine the number of decomposition modes and the penalty factor of variational mode decomposition, the original discharge signal is decomposed to obtain intrinsic mode function components, and an improved wavelet threshold function is used for denoising, and the denoised discharge signal is obtained by superposition and reconstruction; S3, dual-domain feature extraction: the denoised discharge signal is obtained by... S4. Discharge type identification: The denoised discharge signal is input into an attention-enhanced convolutional neural network model. Features are extracted through residual convolutional blocks and SE attention modules. The discharge type identification result is output after global average pooling and fully connected layers. S5. Information fusion and early warning: A basic probability allocation is constructed based on the identification results of each sensor channel. Evidence from multiple sensors is fused to output hierarchical early warning information.

[0011] Furthermore, S2 includes: S21, constructing the objective function. The optimization objective is to minimize the sum of the envelope entropies of all intrinsic mode function components; where, To decompose the mode number; As a penalty factor; For the first The envelope entropy of each component; S22, initialize the sparrow population, encoding the positions as... and The combination of these parameters updates the position based on three roles: discoverer, follower, and vigilant, and iterates to obtain the optimal parameters. and S23, Use and The original discharge signal is subjected to variational mode decomposition, and the correlation coefficient between each component and the original signal is calculated. Components with correlation coefficients below a threshold are removed. S24. Wavelet decomposition is performed on the effective components, and the wavelet coefficients are processed using an improved threshold function: when... hour, ;when hour, ;in, These are the processed wavelet coefficients; These are the original wavelet coefficients; For the threshold; To adjust the parameters; S25, perform inverse transform reconstruction on the denoised wavelet coefficients, and superimpose the components to obtain the denoised discharge signal.

[0012] Further, S3 includes: S31, mapping the discharge pulses onto a two-dimensional plane according to the power frequency phase and amplitude, dividing the phase window and amplitude level, accumulating multiple power frequency cycle data, counting the number of pulses in each grid, and obtaining a phase-resolved partial discharge spectrum; S32, counting the number of discharges in the positive and negative half cycles. and Calculate the ratio of discharge times ; Calculate the phase mean ;in, For the first The center phase value of each phase window; The number of discharges within this phase window is calculated; the phase standard deviation, skewness, and kurtosis are calculated to obtain the phase domain statistical characteristics; S33, the maximum discharge quantity and average discharge quantity are calculated. and the standard deviation of discharge quantity; among which, For the first Each pulse amplitude; S34. Calculate the ratio of positive to negative half-cycle discharge to obtain the amplitude domain statistical characteristics; S35. Coarse-grain the denoised discharge signal according to different scales, calculate the sample entropy at each scale, and obtain the multi-scale entropy characteristics; S36. Combine the phase domain characteristics, amplitude domain characteristics, and multi-scale entropy characteristics to form the discharge feature vector.

[0013] Furthermore, the neural network model in S4 includes an input layer, multiple residual convolutional blocks, a global average pooling layer, a fully connected layer, and an output layer, with an SE attention module embedded in each residual convolutional block; S4 includes: S41, truncating or interpolating the denoised discharge signal into a fixed-length sequence, and then... Normalization; among which, The original signal value; , These are the minimum and maximum values, respectively; S42, the signal passes through each residual convolutional block sequentially, and global average pooling is performed on the feature map within the block to obtain the channel description vector. Channel weights are generated through a fully connected layer, and the feature map is weighted channel by channel; S43, global average pooling is performed on the output of the last residual convolutional block, and after passing through a fully connected layer, the Softmax function is used to calculate the probability of each class. ;in, For the first The output class is selected; the class with the highest probability is chosen as the recognition result.

[0014] Furthermore, S5 includes: S51, outputting probabilities according to each sensor channel. Constructing basic probability assignments ;in, For the first Discharge type propositions; The uncertainty coefficient is used; the remaining probability is assigned to the recognition frame. S52. Multi-sensor evidence is fused using the Dempster synthesis rule, and the fusion formula is as follows: The summation condition is ;in, , Each of the two pieces of evidence supports the proposition. , The assigned value; express and The intersection is ; As the normalization factor, The condition for summation is: , S53. Select the discharge type with the highest probability after fusion as the diagnostic conclusion. If the maximum probability exceeds the decision threshold, the diagnosis is valid. Output graded warnings according to discharge type and intensity.

[0015] The beneficial effects achieved by this invention are as follows:

[0016] This invention employs a three-sensor system—ultra-high frequency (UHF), ultrasonic, and transient ground voltage (TPD)—to detect partial discharge phenomena within switchgear from three physical dimensions: electromagnetic radiation, acoustic vibration, and surface potential changes. The detection frequency bands of each sensor complement each other, effectively compensating for blind spots inherent in single detection methods for specific discharge types or installation locations, significantly improving the comprehensiveness and reliability of discharge detection. The modular sensor assembly utilizes a multi-layered shielding structure, consisting of an equal-voltage shielding layer, an insulating layer, and an electromagnetic shielding layer, arranged sequentially from the outside in. The equal-voltage shielding layer, connected at the same potential as the cabinet, homogenizes the electric field distribution on the sensor surface, preventing the sensor itself from becoming a new source of discharge. The insulating layer provides reliable electrical isolation and mechanical buffering, while the electromagnetic shielding layer effectively shields external interference signals. This three-layered structure works synergistically to ensure the safety of the detection equipment in high-voltage environments while improving the anti-interference capability and signal-to-noise ratio of signal detection. The magnetic base and quick-release interface design allow for rapid installation and removal of the sensors without any modification to the switchgear, making it particularly suitable for portable inspections and temporary testing scenarios, enhancing the flexibility and convenience of system deployment.

[0017] This invention employs a sparrow search algorithm to adaptively determine the number of decomposition modes and the penalty factor in variational mode decomposition (DMD). The optimization objective is to minimize the sum of the envelope entropies of all intrinsic mode function components. This allows DMD to automatically find the optimal parameter combination for discharge signals with different characteristics, overcoming the limitations of traditional fixed-parameter methods that struggle to adapt to diverse signal features and avoiding under- or over-decomposition due to improper parameter settings. By calculating the correlation coefficients between each component and the original signal and removing noise-dominant components with correlation coefficients below a threshold, noise components are effectively removed while retaining useful signal components. An improved wavelet threshold function uses a continuously differentiable exponential transition form near the threshold, avoiding the oscillations caused by discontinuities in hard threshold functions and reducing the constant bias generated by soft threshold functions, thus improving the fidelity of the denoised signal. This adaptive denoising method significantly improves the signal-to-noise ratio of the discharge signal, providing high-quality input data for subsequent feature extraction and pattern recognition.

[0018] This invention extracts statistical features of discharge signals from both the phase and amplitude domains, calculates multi-scale sample entropy features, and combines them to form a high-dimensional feature vector that comprehensively characterizes the discharge signal's properties. Phase domain statistical features include the discharge frequency ratio, phase mean, phase standard deviation, phase skewness, and phase kurtosis, depicting the temporal distribution and symmetry of the discharge pulse within the power frequency cycle. Amplitude domain statistical features include the maximum discharge quantity, average discharge quantity, discharge quantity standard deviation, and discharge quantity ratio, describing the distribution and polarity of the discharge intensity. Multi-scale entropy features, by calculating sample entropy values ​​at different time scales, reveal the complexity and irregularity of the discharge signal across multiple time scales. These three types of features characterize the discharge signal from different perspectives and complement each other, allowing different types of discharges to occupy clearly separated regions in the feature space. This significantly enhances the separability and discriminative power of the features, laying a solid feature foundation for the accurate identification of discharge types.

[0019] This invention employs an attention-enhanced convolutional neural network (CNN) for discharge type identification. The network structure includes multiple residual convolutional blocks. The residual structure, by introducing cross-layer connections, allows gradients to propagate directly to previous layers, mitigating the gradient vanishing problem in deep networks and enabling deeper network training to learn higher-level abstract features. An attention module (SE) is embedded within each residual convolutional block. This module acquires global information for each feature channel through global average pooling, and then generates channel weight vectors through fully connected layers to adaptively weight the feature map channel by channel. It automatically learns the relative importance of each feature channel, strengthens the expression of useful feature channels, and suppresses redundant or noisy feature channels, thus improving the discriminative power of the feature representation. This attention-enhanced mechanism allows the network to adaptively adjust its feature extraction strategy based on the input signal, enabling the extraction of effective discriminative features in different application scenarios. Compared to traditional convolutional neural networks, it has stronger feature learning and generalization capabilities, significantly improving the accuracy of discharge type identification.

[0020] This invention fuses the identification results from multiple sensor channels, constructs a basic probability allocation function based on the neural network output probabilities of each sensor channel, and introduces an uncertainty coefficient to reflect the inherent uncertainty of the model prediction. Through evidence fusion, mutually supporting evidence increases the probability allocation value of the corresponding category, while contradictory evidence is processed through conflict terms. The final fusion result is more reliable and accurate than the judgment of a single sensor. Multi-sensor information fusion fully utilizes the complementarity and redundancy between different types of sensors. When the signal of one sensor is interfered with, the information from other sensors plays a supplementary and verification role, significantly improving the robustness of diagnosis and reducing the false diagnosis rate and false negative rate. Based on an edge computing architecture, signal processing and intelligent diagnosis are performed locally, reducing the amount of data that needs to be transmitted remotely, reducing the demand for communication bandwidth, and also reducing the delay time of diagnostic decisions. This gives the system better real-time performance, enabling timely issuance of graded early warning information to maintenance personnel, and providing effective technical support for the status monitoring and preventive maintenance of mine switchgear. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the composition and structure of the discharge detection device for a general-purpose high and low voltage switchgear for mining applications according to the present invention.

[0022] Figure 2 This is a schematic diagram of the modular sensor assembly structure of the present invention.

[0023] Figure 3 yes Figure 2 A cross-sectional structural diagram.

[0024] Figure 4 This is a comparison chart of the signal-to-noise ratio curves after denoising for Examples 1, 2, and 3 and Comparative Example 1 under different initial signal-to-noise ratios.

[0025] Figure 5 The following is a comparison of the confusion matrix heatmaps of Example 1, Comparative Example 2 and Comparative Example 3, where (a) is the confusion matrix heatmap of Example 1, (b) is the confusion matrix heatmap of Comparative Example 2 and (c) is the confusion matrix heatmap of Comparative Example 3.

[0026] Figure 6 This is a scatter plot showing the recognition accuracy of Examples 2, 3, and Comparative Example 3 at different voltage levels.

[0027] Figure 7 The diagram shows a comparison of the diagnostic performance of Examples 1, 2, and 3 with Comparative Example 4 under different signal qualities. (a) is a comparison of diagnostic accuracy curves, and (b) is a comparison of diagnostic confidence box plots.

[0028] Figure 8 This is a schematic diagram of the structure of an ultra-high frequency sensor, an ultrasonic sensor, and a transient ground voltage sensor.

[0029] Numbering on the map:

[0030] 1. Modular sensor assembly; 11. UHF sensor; 12. Ultrasonic sensor; 13. Transient ground voltage sensor; 101. Sensor body; 102. Mounting base; 1021. Magnetic base; 103. Quick-release interface; 104. Overvoltage protection circuit; 1011. Voltage equalization shielding layer; 1012. Insulation isolation layer; 1013. Electromagnetic shielding layer; 2. Signal acquisition unit; 21. Preamplifier; 22. Bandpass filter; 23. Programmable gain amplifier; 24. Anti-aliasing filter; 25. Analog-to-digital converter; 3. Edge computing node unit; 31. Embedded processor; 32. Data storage module; 33. Communication interface module; 34. Synchronization triggering module; 4. Upper-level monitoring platform; 5. Shielded signal cable. Detailed Implementation

[0031] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. In addition, the forms of the various structures described in the following embodiments are merely illustrative. The present invention is not limited to the structures described in the following embodiments. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] Reference Figures 1-3 ,as well as Figure 8The discharge detection device for general-type high and low voltage switchgear used in mining, as described in this invention, adopts a layered architecture design, consisting of four levels: a modular sensor assembly 1, a signal acquisition unit 2, an edge computing node unit 3, and a host monitoring platform 4. The modular sensor assembly 1 is installed on or inside the cabinet of the general-type high and low voltage switchgear used in mining, responsible for sensing and acquiring various physical signals generated during the discharge process. The signal acquisition unit 2 is electrically connected to the modular sensor assembly 1 via a shielded signal cable 5, responsible for conditioning and digitizing the weak analog signals output by the sensors. The edge computing node unit 3 is connected to the signal acquisition unit 2 via a data bus, receives the digitized discharge signal data, and performs signal processing, feature extraction, and intelligent recognition operations locally. The host monitoring platform 4 is remotely connected to the edge computing node unit 3 via a wired or wireless communication network, used to receive diagnostic results, store historical data, perform trend analysis, and display early warning information. This layered architecture design allows signal acquisition, data processing, and information display to each perform their respective functions, meeting the needs of real-time on-site detection while facilitating distributed deployment and centralized management of the system.

[0033] The modular sensor assembly 1 includes a sensor body 101, which is the signal sensing front end of the system. It can be one of three types of sensors: a UHF sensor 11, an ultrasonic sensor 12, and a transient ground voltage sensor 13. These three sensors are based on different physical detection principles, sensing partial discharge phenomena inside the switchgear from three dimensions: electromagnetic radiation, acoustic vibration, and surface potential changes. The UHF sensor 11 is installed at the ventilation holes on the top of the switchgear or at the joints of the cabinet panels. These locations allow the high-frequency electromagnetic waves generated by the discharge inside the cabinet to effectively leak to the sensor's receiving antenna. The ultrasonic sensor 12 is attached to the outer wall surface of the switchgear busbar compartment, circuit breaker compartment, and cable compartment using a coupling agent or magnetic attraction. Multiple ultrasonic sensors 12 work together, utilizing the time difference of sound wave arrival to achieve spatial positioning of the discharge source. The transient ground voltage sensor 13 is attached to the inner surface of the switchgear door or the cabinet surface near the grounding busbar. When partial discharge occurs inside the cabinet, the transient voltage change generated by the discharge current on the metal surface of the cabinet is detected by this sensor through capacitive coupling. The detection frequency bands of the three types of sensors each have their own focus and complement each other. The collaborative work of multiple types of sensors makes up for the blind spots of a single detection method and improves the comprehensiveness and reliability of discharge detection.

[0034] The sensors in the modular sensor assembly 1 all structurally include a sensor body 101, a multi-layer shielding structure, a mounting base 102, and a quick-release interface 103, among other main components. The sensor body 101 adopts a flattened rectangular shell design, preferably with a very thin shell to accommodate the narrow space inside the switchgear. This ultra-thin design allows the sensor to be installed in a compact location within the switchgear without affecting the layout of existing electrical components or heat dissipation channels. The outer shell of the sensor body 101 is made of high-temperature resistant and flame-retardant engineering plastic, preferably ABS or PC materials with a flame-retardant rating of V-0. These materials possess excellent electrical insulation properties, mechanical strength, and resistance to environmental aging, making them suitable for the harsh environment of high humidity and dust in underground coal mines.

[0035] The internal space of the sensor body 101 houses three functional layers from bottom to top: a sensing element layer, a front circuit board layer, and an interface adapter layer. The sensing element layer, located on the inner bottom surface of the lower housing, is used to mount the sensitive elements required for different types of sensors. Depending on the sensor type, these can be ultra-high frequency antennas, piezoelectric ceramic transducers, or planar capacitive coupling probes. The front circuit board layer is located above the sensing element layer and is fixedly connected to the lower housing via support pillars. The circuit board integrates signal conditioning circuitry and overvoltage protection circuitry 104. The front circuit board is electrically connected to the sensitive elements of the sensing element layer via flexible ribbon cables or soldered wires. The interface adapter layer is located above the front circuit board layer and near the side wall opening of the upper housing, used to connect the signal output terminals of the circuit board to the quick-release interface 103. Appropriate spacing is maintained between the functional layers to achieve electrical isolation and heat dissipation.

[0036] The sensor body 101 also features a multi-layered shielding structure, comprising, from the outside in, a voltage equalization shielding layer 1011, an insulating isolation layer 1012, and an electromagnetic shielding layer 1013. The voltage equalization shielding layer 1011 covers the outer surface of the sensor body 101 and is made of conductive metal materials such as aluminum alloy sheet or galvanized steel sheet. The edges and corners of the voltage equalization shielding layer 1011 are rounded, effectively eliminating the sharp-point effect and preventing corona discharge due to electric field concentration in a high-voltage electric field. A grounding terminal is located on one side of the voltage equalization shielding layer 1011, which is reliably connected to the metal frame of the switchgear cabinet via a short grounding wire. This ensures that the voltage equalization shielding layer 1011 and the cabinet are at the same potential, thereby homogenizing the electric field distribution on the sensor surface and preventing the sensor itself from becoming a new source of discharge.

[0037] An insulating layer 1012 is disposed in the interlayer space between the equalizing shielding layer 1011 and the sensor body 101 housing, and is made of elastic insulating materials such as silicone rubber or polyurethane. The insulating layer 1012 is sheet-like or sleeve-like and covers the surface of the housing. Its main function is to form reliable electrical isolation between the equalizing shielding layer 1011 and the internal circuitry of the sensor. Simultaneously, the elastic material also provides mechanical buffering and vibration damping, protecting the internal precision circuitry from external vibrations and impacts. The insulation strength of the insulating layer 1012 should meet the requirements of the rated voltage level of the switchgear. An electromagnetic shielding layer 1013 is disposed on the inner surface of the sensor body 101 housing, and is implemented using a conductive coating or conductive tape, preferably a nickel-based or copper-based conductive coating spraying process. The electromagnetic shielding layer 1013 is tightly fitted to the inner wall of the housing, covering the entire inner surface of the housing and forming a continuous conductive path at the joints. The function of the electromagnetic shielding layer 1013 is to shield the influence of external electromagnetic interference signals on the internal detection circuitry of the sensor, improving the signal-to-noise ratio of signal detection. The three-layer shielding structure provides triple protection functions from the outside in, including electric field equalization, electrical isolation, and electromagnetic shielding. This ensures the safety of the sensor in high-voltage environments and improves the anti-interference capability of signal detection.

[0038] Mounting base 102 is located at the bottom of sensor body 101 and is used to fix the sensor to the switch cabinet. Mounting base 102 is connected to magnetic base 1021.

[0039] The magnetic base 1021 is fixedly connected to the lower housing of the sensor body 101. The magnetic base 1021 has a flat, circular or rectangular disc-shaped structure with an array of countersunk holes inside to accommodate permanent magnets. Multiple high-performance neodymium iron boron permanent magnets are embedded in the countersunk hole array, arranged in a matrix. The magnetic poles of the permanent magnets are all perpendicular to the bottom surface of the base and have the same polarity. The magnetic force generated by the permanent magnet array firmly attaches the sensor to the steel frame or the inside of the cabinet door, or other ferromagnetic metal surfaces. The bottom surface of the magnetic base 1021 has an anti-slip pad made of elastic material such as silicone or rubber, which is adhered to the bottom surface of the base to increase friction and protect the cabinet surface. The advantage of the magnetic installation method is that no modification to the switchgear is required, and the installation and disassembly process is quick and convenient, making it particularly suitable for portable inspection or temporary testing scenarios.

[0040] A quick-release interface 103 is located on the side of the sensor body 101 and is used to connect the signal transmission cable between the sensor and the signal acquisition unit 2. The quick-release interface 103 adopts an aviation plug structure, preferably using a standard-sized circular metal-shell connector with a multi-pin contact configuration for transmitting positive, negative, shielding ground, and power signals. The socket of the quick-release interface 103 is fixedly installed at an opening on the side wall of the sensor body 101. A sealing gasket is provided between the socket flange and the housing, and the inner cavity of the socket communicates with the inside of the housing to allow signal extraction. The housing of the quick-release interface 103 is made of conductive metal materials such as nickel-plated brass, providing good conductivity and corrosion resistance. The internal contacts of the interface are made of gold-plated copper alloy, ensuring reliable electrical contact and low contact resistance. The quick-release interface 103 has a high protection level, with a sealing structure using fluororubber O-rings, effectively preventing moisture and dust intrusion in humid and dusty underground environments. The quick-release interface 103 supports blind insertion; a guide structure is provided between the plug and socket, allowing operators to perform insertion and removal actions without visual alignment. For aviation connectors, threaded or bayonet locking methods are preferred, providing good vibration and pull-out resistance after connection.

[0041] An overvoltage protection circuit 104 is installed at the input of the preamplifier circuit inside the sensor body 101 to protect the subsequent circuits from transient overvoltage damage caused by high-voltage discharge environments. The overvoltage protection circuit 104 is located at the signal input terminal of the preamplifier circuit board and adopts a two-stage series protection architecture. The first stage uses a gas discharge tube connected between the signal input terminal and signal ground. When a high-amplitude transient overvoltage occurs at the input terminal, the gas discharge tube breaks down and conducts within a very short time, dissipating most of the energy to ground, providing coarse protection. The second stage uses a TVS transient suppression diode connected between the signal path after the gas discharge tube and ground. This diode clamps the residual overvoltage after dissipation by the gas discharge tube to a safe level, providing fine protection. The gas discharge tube and the TVS transient suppression diode are isolated by a current-limiting resistor, the value of which is selected according to the signal frequency band and protection current. This two-stage protection architecture balances high energy dissipation capability and fast response speed, effectively protecting the precision preamplifier circuit inside the sensor.

[0042] The sensor body 101 can be an ultra-high frequency (UHF) sensor 11, an ultrasonic sensor 12, or a transient ground voltage sensor 13. The UHF sensor 11 is a sensor in the modular sensor assembly 1 used to detect UHF electromagnetic radiation signals generated by discharge. The UHF sensor 11 integrates an UHF antenna and a matching preamplifier circuit. The UHF antenna adopts a planar printed antenna form, preferably a wideband antenna form such as a Hilbert fractal structure, spiral structure, or butterfly structure. These antenna structures have the advantages of wideband characteristics and miniaturization, achieving effective reception of UHF signals within a limited housing size. The antenna substrate uses a low-loss high-frequency dielectric material, preferably a high-frequency PCB-specific substrate, to ensure high-frequency signal transmission performance. The preamplifier circuit adopts a low-noise amplifier structure to perform primary amplification of the weak signal received by the antenna. The UHF sensor 11 is preferably installed at the metal louvered ventilation holes on the top of the switch cabinet, or near the joint between the side panel and the top panel of the cabinet. These locations allow the electromagnetic waves generated by the discharge inside the cabinet to be effectively radiated to the sensor through the openings or gaps. The advantages of ultra-high frequency detection are strong resistance to power frequency corona interference, detection of internal discharge in the insulating medium inside the cabinet, and high detection sensitivity and positioning accuracy.

[0043] The ultrasonic sensor 12 is a sensor in the modular sensor assembly 1 used to detect ultrasonic signals generated by discharge. The sensor body 101 of the ultrasonic sensor 12 integrates a piezoelectric ceramic transducer and a matching charge amplification circuit. The piezoelectric ceramic transducer is made of PZT piezoelectric ceramic material and is generally disk-shaped, with its sensitive surface facing the bottom of the sensor body 101. When mechanical vibration is transmitted from the cabinet surface, the piezoelectric ceramic sheet deforms and generates a charge signal proportional to the vibration amplitude. The charge amplification circuit converts the high-impedance charge signal output by the piezoelectric ceramic into a low-impedance voltage signal, facilitating subsequent transmission and processing. The installation of the ultrasonic sensor 12 requires ensuring good acoustic coupling between the transducer's sensitive surface and the switch cabinet surface. When using a magnetic base 1021 for installation, it is preferable to apply an ultrasonic-specific coupling agent between the transducer's sensitive surface and the cabinet surface to eliminate air gaps and reduce sound wave reflection loss. The ultrasonic sensors 12 are preferably distributed and installed in multiple locations, such as the outer wall of the busbar compartment of the switchgear, the outer wall of the circuit breaker compartment, and the outer wall of the cable compartment. The time difference between multiple sensors receiving the same discharge pulse can be used to calculate the spatial location of the discharge source. The advantages of ultrasonic detection are strong anti-electromagnetic interference capability, easy location of discharge source, and high detection sensitivity for surface discharge and surface discharge.

[0044] The transient ground voltage sensor 13 is a sensor in the modular sensor assembly 1 used to detect transient voltage changes on the cabinet surface caused by discharge. The sensor body 101 of the transient ground voltage sensor 13 integrates a planar capacitive coupling probe and a matching high input impedance amplifier circuit. The planar capacitive coupling probe consists of a metal disc, which is installed inside the sensor body 101 and parallel to the bottom surface. The metal disc is separated from the metal surface of the switchgear cabinet by an insulating gasket, forming a capacitive coupling relationship between the metal disc and the cabinet surface. When a partial discharge occurs inside the cabinet, the discharge pulse current generates a transient voltage change on the cabinet surface. This voltage change is transmitted to the probe's metal disc through capacitive coupling and then converted into a measurable voltage signal by the high input impedance amplifier circuit. The insulating gasket is made of a material with good high-frequency dielectric properties, such as polytetrafluoroethylene or polyimide. The transient ground voltage sensor 13 is preferably mounted on the inner surface of the switchgear cabinet door in the middle or on the inner wall surface of the cabinet near the grounding busbar. During installation, it is necessary to ensure that the probe surface is parallel to the metal surface of the cabinet. The advantages of transient ground voltage detection are non-contact detection and easy installation, making it particularly suitable for rapid inspection and preliminary screening scenarios.

[0045] The shielded signal cable 5 is used to connect the modular sensor assembly 1 and the signal acquisition unit 2, transmitting analog signals between them. The shielded signal cable 5 adopts a double-layer shielding structure design. From the inside out, the cable structure includes a signal core wire, an insulation layer, an inner shielding layer, an isolation layer, an outer shielding layer, and a sheath layer. The signal core wire is preferably made of multi-strand tinned copper conductor, which has good flexibility and resistance to bending fatigue. The inner shielding layer uses an aluminum-plastic composite film longitudinally wrapped with the aluminum foil facing outwards, effectively shielding high-frequency electromagnetic interference. The outer shielding layer uses a tinned copper wire braided mesh, effectively shielding low-frequency electromagnetic interference and providing mechanical protection. The isolation layer is located between the inner and outer shielding layers, using polyester film or paper tape wrapping to prevent direct contact between the inner and outer shielding layers. The sheath layer is made of flame-retardant polyvinyl chloride or polyurethane, which has good wear resistance, oil resistance, and flame-retardant properties. Both ends of the shielded signal cable 5 are equipped with aviation plug connectors that mate with the quick-release interface 103. The plug housing and the braided shielding layer of the cable form a reliable electrical connection through crimping or welding. The tail of the aviation plug is equipped with a dedicated grounding terminal. During installation, this grounding terminal must be reliably connected to the switch cabinet or the housing of the signal acquisition unit 2 through a short grounding wire to achieve multi-point grounding of the shielding layer and further improve the shielding effectiveness.

[0046] Signal acquisition unit 2 is the analog signal conditioning and digitization module of this system, responsible for converting the weak analog signals output by the sensors into a standard signal format suitable for digital processing. The signal processing path of signal acquisition unit 2 consists of a preamplifier 21, a bandpass filter 22, a programmable gain amplifier 23, an anti-aliasing filter 24, and an analog-to-digital converter 25 cascaded in sequence.

[0047] The input of preamplifier 21 is connected to the quick-release interface 103 of modular sensor assembly 1 via shielded signal cable 5, and its output is connected to the input of bandpass filter 22. Preamplifier 21 is constructed using a low-noise operational amplifier or a low-noise dedicated integrated circuit to perform primary amplification of the weak signal output from the sensor. Its noise figure is preferably as low as possible to ensure the minimum detectable signal level of the system. The input of bandpass filter 22 is connected to the output of preamplifier 21, and its output is connected to the input of programmable gain amplifier 23. The passband range of bandpass filter 22 is set according to the corresponding sensor type and is used to filter out out-of-band noise and interference components outside the signal frequency band. Bandpass filter 22 preferably adopts a multi-order active filter or LC filter structure.

[0048] The input terminal of the programmable gain amplifier 23 is connected to the output terminal of the bandpass filter 22, and its output terminal is connected to the input terminal of the anti-aliasing filter 24. Its control terminal is connected to the control bus of the edge computing node unit 3. The gain value of the programmable gain amplifier 23 can be digitally set by the edge computing node unit 3 via the control bus. The main function of the programmable gain amplifier 23 is to dynamically adjust the amplification factor according to the actual amplitude of the input signal, ensuring that discharge signals of different intensities fall within the optimal input range of the analog-to-digital converter 25 after amplification. This avoids weak signals being submerged in quantization noise and prevents clipping distortion of strong signals. The input terminal of the anti-aliasing filter 24 is connected to the output terminal of the programmable gain amplifier 23, and its output terminal is connected to the analog input terminal of the analog-to-digital converter 25. The anti-aliasing filter 24 adopts a low-pass filter structure to filter out high-frequency components above the Nyquist frequency, preventing spectral aliasing distortion after sampling. The analog input of the analog-to-digital converter 25 is connected to the output of the anti-aliasing filter 24, its digital output is connected to the data bus of the edge computing node unit 3, and its trigger input is connected to the trigger output of the synchronization trigger module 34. The analog-to-digital converter 25 is responsible for converting the conditioned analog voltage signal into a digital code stream. Its sampling clock and trigger signal come from the synchronization trigger module 34 of the edge computing node unit 3, thereby realizing phase synchronization acquisition of multi-channel signals.

[0049] Edge computing node unit 3 performs on-site data processing and intelligent diagnosis. It is responsible for receiving the digital discharge signal output by signal acquisition unit 2, performing signal denoising, feature extraction, pattern recognition, and information fusion locally, and uploading the diagnostic results to the upper-level monitoring platform 4. Edge computing node unit 3 consists of four functional modules: embedded processor 31, data storage module 32, communication interface module 33, and synchronization triggering module 34.

[0050] The embedded processor 31 executes the operations of the edge computing node unit 3. Its data port is connected to the digital output of the analog-to-digital converter 25 of the signal acquisition unit 2 via a data bus, and its control port is connected to the control terminal of the programmable gain amplifier 23 via a control bus. The embedded processor 31 preferably uses an ARM architecture or a similar performance embedded processor and is equipped with sufficient memory resources to support the operation of the computation program. The embedded processor 31 runs an embedded operating system, which internally stores signal denoising programs, feature extraction programs, pre-trained neural network model parameters, and multi-sensor information fusion programs. Upon receiving discharge signal data from the signal acquisition unit 2, the embedded processor 31 sequentially executes adaptive denoising, dual-domain feature extraction, discharge type identification, and evidence fusion diagnosis processes, ultimately outputting diagnostic information such as the discharge type identification result, severity level, and treatment recommendations. The advantage of the edge computing architecture is that most data processing is completed on-site, reducing the amount of data that needs to be transmitted remotely, lowering the demand for communication bandwidth, and reducing the latency of diagnostic decisions, thus giving the system better real-time performance.

[0051] The data storage module 32 is connected to the embedded processor 31 via a high-speed data bus, used for temporary caching of acquired raw waveform data and long-term storage of diagnostic results and statistical data. The data storage module 32 preferably uses industrial-grade solid-state storage media, which has high read / write speeds and a long service life. The storage capacity configuration should meet the local storage requirements of historical data within a certain period. When the storage space is close to saturation, a first-in-first-out (FIFO) strategy is used to automatically overwrite the oldest historical data. The communication interface module 33 is connected to the embedded processor 31 via an internal bus, responsible for enabling data communication between the edge computing node unit 3 and the upper-level monitoring platform 4. The communication interface module 33 preferably supports both wired and wireless communication methods, allowing users to select the appropriate communication method based on site conditions. The communication protocol preferably adopts the industrial Internet of Things (IoT) standard protocol, facilitating integration with various upper-level systems.

[0052] The synchronous trigger module 34 is connected to the embedded processor 31 via an internal bus, and its trigger output is connected to the trigger input of the analog-to-digital converter 25 of each signal acquisition unit 2. The synchronous trigger module 34 generates the clock signal and trigger signal required for multi-channel synchronous acquisition. The synchronous trigger module 34 has a power frequency voltage sampling input port, which generates a trigger pulse that is phase-locked with the power frequency by detecting the zero-crossing point of the power frequency voltage signal. This trigger pulse is simultaneously sent to the analog-to-digital converter 25 of each signal acquisition unit 2 through a trigger signal distribution circuit, ensuring that each channel starts sampling at the same time and guaranteeing the phase synchronization relationship between the multi-channel signals. Power frequency synchronous acquisition is of great significance for the subsequent construction of phase-resolved partial discharge maps.

[0053] The supervisory control platform 4 is the remote monitoring and data management terminal of this system, deployed at the ground monitoring center or cloud server, and interacts with the edge computing node units 3 distributed in various locations via a communication network. The communication port of the supervisory control platform 4 is connected to the communication interface module 33 of the edge computing node unit 3 via a wired or wireless communication network. The main functions of the supervisory control platform 4 include real-time reception and display of diagnostic results, storage and query of historical data, analysis and prediction of discharge trends, hierarchical push of early warning information, and automatic generation of reports. The human-machine interface of the supervisory control platform 4 adopts a graphical design, intuitively displaying discharge detection results in various forms such as switch cabinet plan, waveform curves, phase resolution spectra, and trend curves. When the edge computing node unit 3 reports discharge early warning information, the supervisory control platform 4 promptly reminds maintenance personnel through various methods such as audible and visual alarms, pop-up prompts, SMS push notifications, or email notifications. The supervisory control platform 4 also has user access control functions, with different levels of users having different operating permissions to ensure the security of system operation. The upper-level monitoring platform 4 can also send parameter configuration instructions and model update data to the edge computing node unit 3 to realize remote parameter adjustment and algorithm upgrade.

[0054] The discharge detection device for general-purpose high and low voltage switchgear in mining, as described in this invention, ensures the safety of the detection equipment in high-voltage environments through a multi-layer shielding structure. It achieves multi-dimensional perception of discharge phenomena through the collaborative operation of multiple types of sensors and realizes localized and real-time diagnostic decision-making through an edge computing architecture. The signal flow between the components is as follows: the modular sensor component 1 collects discharge signals and outputs analog electrical signals, which are transmitted to the signal acquisition unit 2 via shielded signal cable 5 for conditioning and digital conversion. The digital signals are transmitted to the edge computing node unit 3 via a data bus for computation and diagnostic decision-making. The diagnostic results are uploaded to the upper-level monitoring platform 4 via a communication network for storage, display, and early warning. The control signal flow is as follows: the edge computing node unit 3 outputs a programmable gain control signal to the programmable gain amplifier 23 of the signal acquisition unit 2; the synchronous trigger module 34 outputs a synchronous trigger signal to each analog-to-digital converter 25; and the upper-level monitoring platform 4 can issue parameter configuration commands to the edge computing node unit 3. This design of signal and control flows enables the entire system to work in a coordinated and orderly manner, achieving effective monitoring and intelligent diagnosis of the partial discharge state of general-purpose high and low voltage switchgear in mining.

[0055] This invention also designs a discharge detection method for general-type high and low voltage switchgear used in mines. It combines multi-sensor collaborative detection with edge intelligent diagnosis, employing multi-source signal synchronous acquisition, adaptive noise reduction processing, dual-domain feature extraction, deep learning recognition, and multi-sensor information fusion to achieve accurate detection and intelligent diagnosis of partial discharges inside the mine switchgear. This improves the sensitivity, accuracy, and intelligence level of discharge detection. The specific steps are as follows:

[0056] Step S1 is the multi-source signal acquisition step. This step utilizes multiple types of sensors installed at different locations in a general-purpose high and low voltage switchgear for mining to synchronously acquire various physical signals generated during the discharge process under a unified time reference at the zero-crossing point of the power frequency voltage. After signal conditioning and analog-to-digital conversion, the digitized multi-channel raw discharge signal is obtained. In specific implementation, the zero-crossing point of the power frequency voltage signal is first detected by the synchronous trigger module 34. The zero-crossing point of the power frequency voltage signal refers to the instant when the power frequency voltage changes from negative to positive or from positive to negative. At this moment, the voltage amplitude is zero and the rate of change is the largest, which is the natural boundary point of the power frequency cycle. The synchronous trigger module 34 detects the power frequency voltage signal in real time through a voltage comparator or digital signal processing algorithm. When a zero-crossing point is detected, a pulse trigger signal is immediately generated. This trigger signal is simultaneously sent to the trigger input terminal of the analog-to-digital converter 25 of each signal acquisition unit 2 through the trigger signal distribution circuit, ensuring that the analog-to-digital converter 25 of each channel starts the sampling process at the same time. This synchronous triggering mechanism based on the zero-crossing point of the power frequency ensures that the discharge pulses acquired by each sensor channel maintain a strict phase correspondence on the time axis, laying the foundation for subsequent construction of phase-resolved partial discharge maps. The UHF sensor 11 receives the UHF electromagnetic radiation signal generated by the discharge within the cabinet, while the ultrasonic sensor 12 receives the ultrasonic signal generated by the discharge. Since ultrasonic waves attenuate less when propagating in a solid medium, the spatial location of the discharge source can be achieved by analyzing the arrival time difference of the ultrasonic signals received by sensors at different locations. The transient ground voltage sensor 13 detects the transient voltage change generated by the discharge current on the metal surface of the cabinet; this signal reflects the amplitude and waveform characteristics of the discharge pulse. Based on different physical detection principles, these three sensors observe the same discharge event from multiple angles—electromagnetic, acoustic, and electrical—providing a data foundation for improving diagnostic accuracy and reliability.

[0057] The weak analog signals output by each sensor are first amplified by preamplifier 21. Preamplifier 21 employs a low-noise operational amplifier design, and its noise figure directly affects the minimum detectable signal level of the system. Preferably, a dedicated amplifier chip with a low noise figure and high input impedance is used. The amplified signal is then fed into bandpass filter 22. The passband range of bandpass filter 22 is set according to the sensor type and is used to filter out noise and interference components outside the signal frequency band. For ultra-high frequency sensor channels, the passband range of bandpass filter 22 is preferably set to cover the main energy distribution frequency band of ultra-high frequency signals; for ultrasonic sensor channels, the passband range of bandpass filter 22 is preferably set near the resonant frequency of the ultrasonic sensor; for transient ground voltage sensor channels, the passband range of bandpass filter 22 is preferably set to include the main frequency components of transient pulse signals. The signal after bandpass filtering enters the programmable gain amplifier 23. The gain value of the programmable gain amplifier 23 is dynamically adjusted by the edge computing node unit 3 according to the actual amplitude of the input signal. When the input signal amplitude is detected to be weak, the gain is automatically increased to improve the signal-to-noise ratio. When the input signal amplitude is detected to be strong, the gain is automatically decreased to prevent saturation distortion in the subsequent circuit. This adaptive gain adjustment mechanism ensures that discharge signals of different intensities fall within the optimal input dynamic range of the analog-to-digital converter 25 after amplification. The signal amplified by the programmable gain amplifier 23 undergoes final filtering processing through the anti-aliasing filter 24. The anti-aliasing filter 24 adopts a low-pass filter structure, and its cutoff frequency is set slightly lower than half of the sampling frequency to filter out high-frequency components higher than the Nyquist frequency, preventing spectral aliasing distortion during digital sampling. Finally, the analog signal after processing through the complete signal conditioning link is converted into a digital code stream by the analog-to-digital converter 25. The sampling frequency of the analog-to-digital converter 25 is preferably set to meet the requirements of signal bandwidth and waveform fidelity, and the resolution is preferably set high enough to ensure the accuracy of signal amplitude quantization. The analog-to-digital converters 25 of each channel simultaneously start sampling under the control of the synchronous trigger signal. The sampling process lasts for one or more power frequency cycles. The acquired digital data is transmitted in real time to the embedded processor 31 of the edge computing node unit 3 via the data bus for subsequent processing. Through this synchronous acquisition mechanism, the system acquires multi-channel raw discharge signals containing complete phase information, providing a high-quality data source for subsequent signal processing and feature extraction.

[0058] Step S2 is the adaptive signal denoising step. This step targets various noises and interferences mixed in the original discharge signal, employing an adaptive denoising method combining variational mode decomposition optimized based on the sparrow search algorithm and an improved wavelet threshold function. This effectively suppresses noise components while retaining useful discharge feature information. This step includes five sub-steps, S21 to S25. Step S21 constructs an objective function with the optimization objective of minimizing the sum of the envelope entropies of each intrinsic mode function component after variational mode decomposition. Envelope entropy is an indicator of signal complexity and uncertainty. For intrinsic mode function components containing discharge information, their envelopes are relatively regular, and their envelope entropy values ​​are small; while for noise-dominated intrinsic mode function components, their envelopes exhibit random fluctuation characteristics, and their envelope entropy values ​​are large. Therefore, by minimizing the sum of the envelope entropies of each intrinsic mode function component, variational mode decomposition can suppress noise while maximizing the retention of useful signal components. The objective function expression is:

[0059] In the formula, The objective function value is dimensionless. The decomposition mode number indicates that the original signal is decomposed into... Each intrinsic mode function component is dimensionless. This is a penalty factor used to balance data fidelity and modal smoothness; it is dimensionless. For the first The envelope entropy of each intrinsic mode function component is dimensionless. Indicates to From 1 to Summation. The decomposition quality and denoising effect of variational mode decomposition are related to the number of decomposed modes. and penalty factor The value of is closely related to the decomposed mode number. Too small a value will lead to under-decomposition, with different frequency components overlapping in the same mode; the number of decomposed modes... An excessively large value can lead to over-decomposition, breaking down a meaningful signal component into multiple fragmented modes; penalty factor Too small a value will result in blurred modal boundaries and insufficient frequency resolution; penalty factor Excessive values ​​can lead to over-constraint and loss of signal details. Traditional methods typically use manual experience or grid search to determine these two parameters. However, manually set parameters are often not optimal, and grid search is computationally intensive and difficult to adapt to different signal characteristics. This invention uses a sparrow search algorithm to adaptively determine these two parameters, automatically finding the optimal parameter combination for discharge signals with different characteristics, thus improving the adaptability and effectiveness of denoising processing.

[0060] Step S22 initializes the sparrow population, encoding the position of each sparrow as a combination of decomposed modes and a penalty factor. Positions are updated according to three roles: discoverer, follower, and vigilant. The algorithm iteratively searches for the minimum value of the objective function to obtain the optimal decomposed modes and the optimal penalty factor. The sparrow search algorithm is a swarm intelligence optimization algorithm inspired by sparrow foraging behavior. The algorithm divides individuals in the sparrow population into three roles: discoverer, follower, and vigilant. Discoverers are responsible for exploring the search space to find potential food sources; followers update their positions based on the discoverer's information; and vigilants issue warnings and guide the population to relocate when the population is threatened. In specific implementation, the sparrow population is first initialized, with the population size preferably set large enough to ensure the coverage of the search space. The position of each sparrow is represented by a two-dimensional vector, with the first dimension encoding the decomposed modes. The second dimension encodes the penalty factor. Decomposition of mode numbers The range of values ​​for the penalty factor should preferably be set to include possible optimal values. The optimal value range is set to cover a numerical interval that encompasses typical application scenarios. For each sparrow's position, the objective function value is calculated under the corresponding parameter combination. This involves performing variational mode decomposition on the original signal based on the decomposed mode number encoded at that position and the penalty factor, calculating the envelope entropy of each intrinsic mode function component, and summing the entropies to obtain the objective function value. Individuals in the population are sorted according to the objective function value. Several individuals with the smallest objective function value are selected as discoverers, the remaining individuals as followers, and a small number of individuals are randomly selected as vigilants. During the iterative search, discoverers adjust their positions based on their fitness and vigilance values. When there is no threat, discoverers perform a local search near their current position; when a threat appears, they conduct a global exploration to find a safer area. Followers update their positions based on the discoverers' position information. Followers with lower fitness move closer to the optimal discoverer, while followers with higher fitness randomly search globally. Vigilants monitor the safety status of the population. When the population is at the boundary or trapped in a local optimum, vigilants guide the population out of the local optimum area through large-scale movement. After a predetermined number of iterations or when the convergence condition is met, the position of the individual with the smallest objective function value in the population is the optimal decomposition mode number. and optimal penalty factor The Sparrow Search algorithm has the advantages of fast convergence speed and strong global search capability. It finds an approximate optimal solution in a relatively small number of iterations, avoiding the large computational cost and the problem of getting trapped in local optima in grid search.

[0061] Step S23 uses optimal parameters and Variational mode decomposition was performed on the original discharge signal to obtain For each intrinsic mode function (EMF) component, the correlation coefficient between each component and the original signal is calculated. Components with correlation coefficients below a preset threshold are removed to obtain the set of effective EMF components. Variational mode decomposition (VMD) is an adaptive signal decomposition method that decomposes a signal into several amplitude-modulated (AM) and frequency-modulated (FM) components with specific center frequencies and finite bandwidths. Each component is called an EMF. VMD achieves adaptive frequency band division of the signal by constructing and solving variational problems, and has a better mathematical theoretical foundation and stronger anti-mode aliasing ability compared to traditional empirical mode decomposition. The optimal number of decomposed modes obtained in step S22 is used. and optimal penalty factor Variational mode decomposition was performed on the original discharge signal to obtain There are 10 intrinsic mode function (IMF) components. These components include useful components related to the discharge signal and irrelevant components mainly composed of noise. To distinguish between useful and irrelevant components, the correlation coefficient between each IMF component and the original signal is calculated. The correlation coefficient characterizes the degree of linear correlation between the two signals, and its value ranges from -1 to +1, with a larger absolute value indicating a stronger correlation. For IMF components related to the discharge signal, their waveform characteristics are highly similar to the discharge pulse portion of the original signal, and the absolute value of the correlation coefficient is large. For IMF components mainly composed of noise, their waveforms exhibit randomness, and the absolute value of the correlation coefficient with the original signal is small. A correlation coefficient threshold is set, preferably determined experimentally based on the specific application scenario. IMF components with an absolute correlation coefficient value below the threshold are identified as noise-dominated components and removed, while IMF components with an absolute correlation coefficient value above the threshold are retained as valid components. This correlation coefficient-based screening mechanism removes most of the noise components while retaining the useful signal components, reducing the burden on subsequent wavelet threshold denoising and improving denoising efficiency and effectiveness.

[0062] Step S24 performs wavelet decomposition on each effective intrinsic mode function component to obtain wavelet coefficients at each level, and then processes the wavelet coefficients using an improved threshold function. Wavelet transform decomposes the signal into wavelet coefficients of different scales and frequencies. Noise and useful signals have different distribution characteristics in the wavelet domain. The energy of the useful signal is mainly concentrated on wavelet coefficients with larger amplitudes, while the energy of noise is dispersed among a large number of wavelet coefficients with smaller amplitudes. Traditional wavelet thresholding denoising methods use hard or soft threshold functions. Hard threshold functions are discontinuous at the threshold, causing oscillations, while soft threshold functions are continuous at the threshold but produce a constant deviation, affecting the signal reconstruction accuracy. This invention uses an improved threshold function, which employs a continuous and differentiable transition form near the threshold, ensuring both function continuity and reducing reconstruction deviation. The expression of the improved threshold function has two cases: when the absolute value of the wavelet coefficient is less than the adaptive threshold, the processed wavelet coefficients... The processed wavelet coefficients are equal to zero; when the absolute value of the wavelet coefficients is greater than or equal to the adaptive threshold, the processed wavelet coefficients are equal to zero. equal:

[0063] In the formula, These are the wavelet coefficients after thresholding, and are dimensionless. These are the original wavelet coefficients, which are dimensionless. for The absolute value of is dimensionless; An adaptive threshold, dimensionless; For a sign function, when hour ,when hour ,when hour Dimensionless; It is a natural exponential function, expressed in terms of the natural constant. Exponential operations with base 0; This is a dimensionless parameter used to adjust the smoothness of the threshold function's transition. The characteristic of this threshold function is that when the absolute value of the wavelet coefficients just exceeds the threshold, a smooth transition is achieved through an exponential function. As the absolute value of the wavelet coefficients increases, the exponential term approaches zero, and the processed wavelet coefficients tend to have a constant offset from the original coefficients. This design avoids the discontinuities of hard thresholding and reduces the constant deviation of soft thresholding, thus improving the fidelity of the denoised signal. Adaptive Thresholding A general threshold estimation method is used to calculate the threshold, which is related to the signal's noise standard deviation and signal length, and is adaptively adjusted according to the actual noise level of the signal. Adjustment parameters... The optimal value should be selected based on signal characteristics and noise reduction requirements. A larger value results in a smoother transition. Smaller values ​​approximate the hard threshold characteristic. Wavelet decomposition is performed on each effective intrinsic mode function component. The number of wavelet decomposition levels is preferably determined based on the signal's sampling rate and frequency characteristics, typically selecting a number that adequately represents the signal's multi-scale characteristics. After wavelet decomposition, wavelet coefficients for each level are obtained, including detail coefficients and approximation coefficients. An improved threshold function is applied to each level's wavelet coefficients, setting coefficients with amplitudes less than the threshold to zero to remove noise, and shrinking coefficients with amplitudes greater than the threshold according to the rules of the improved threshold function.

[0064] Step S25 involves performing an inverse wavelet transform on the denoised wavelet coefficients to reconstruct each intrinsic mode function (EMF) component. The reconstructed components are then superimposed to obtain the denoised discharge signal. The inverse wavelet transform is the reverse process of wavelet decomposition, reconstructing the time-domain signal based on the thresholded wavelet coefficients. For each EMF component, an inverse wavelet transform is performed on the thresholded wavelet coefficients at each level to reconstruct the denoised EMF component. Since the noise-dominant wavelet coefficients have been zeroed or shrunk during the thresholding process, the noise component in the reconstructed EMF component is effectively suppressed, while useful discharge signal features are preserved. All reconstructed effective EMF components are superimposed and summed according to the time series to obtain the final denoised discharge signal. This denoised discharge signal retains the main discharge features of the original signal, including key information such as the amplitude, phase, and waveform of the discharge pulse, while significantly reducing the noise level and improving the signal-to-noise ratio. The adaptive denoising process in steps S21 to S25 effectively solves the problem of severe noise interference in the original discharge signal, providing high-quality input data for subsequent feature extraction and pattern recognition, and improving the accuracy and robustness of discharge type identification.

[0065] Step S3 is the dual-domain feature extraction step. This step extracts statistical features of the discharge signal from both the phase and amplitude domains, calculates multi-scale entropy features, and combines them to form a feature vector that comprehensively characterizes the discharge signal. It includes five sub-steps, S31 to S35. Step S31 uses the power frequency cycle as the time reference, mapping each discharge pulse to a two-dimensional plane according to its corresponding power frequency phase and discharge amplitude at the time of occurrence. The phase axis is divided into several phase windows, and the amplitude axis is quantized into several amplitude levels. Data from multiple power frequency cycles is accumulated, and the number of discharge pulses falling into each phase-amplitude grid is counted to obtain a phase-resolved partial discharge spectrum. The phase-resolved partial discharge spectrum is a classic characterization method in the field of partial discharge detection. This spectrum uses the power frequency phase as the horizontal axis and the discharge amplitude as the vertical axis. By statistically analyzing the distribution of a large number of discharge pulses in the phase-amplitude plane, a two-dimensional spectrum that intuitively reflects the discharge characteristics is formed. Different types of partial discharge exhibit different distribution patterns in phase-resolved partial discharge maps. For example, corona discharge typically presents dense discharge pulses near the peak of the power frequency voltage, surface discharge usually shows discharge pulse distribution during both the voltage rise and fall phases, and internal discharge typically presents symmetrically distributed discharge pulse clusters before and after the voltage peak. In practical implementation, the power frequency period is first determined. For a 50Hz power frequency system, one power frequency period is 20ms; for a 60Hz power frequency system, one power frequency period is approximately 16.67ms. Using the zero-crossing point of the power frequency voltage as the phase zero point, the 360-degree phase range of one power frequency period is divided into several phase windows. The number of phase windows is preferably determined based on the phase resolution requirements. A larger number of phase windows results in higher phase resolution but also a greater computational burden; a value balancing resolution and computational efficiency is usually chosen. The discharge amplitude range is quantized into several amplitude levels, the number of which is preferably determined based on the amplitude resolution requirements. For each detected discharge pulse, the corresponding power frequency phase is calculated based on its occurrence time, and its amplitude level is determined based on its amplitude. The pulse is then assigned to a two-dimensional grid defined by the corresponding phase window and amplitude level. Discharge data from multiple power frequency cycles is accumulated, typically tens to hundreds of cycles, to ensure the stability and reliability of the statistical characteristics. The number of discharge pulses falling within each phase-amplitude grid is counted, and this number is used as the pixel value for that grid location, forming a phase-resolved partial discharge map. This map visually displays the temporal and amplitude distribution characteristics of the discharge pulses within the power frequency cycle, serving as the data foundation for subsequent extraction of phase and amplitude domain statistical features.

[0066] Step S32: Count the number of discharge pulses during the positive and negative half-cycles respectively. and Calculate the ratio of discharge times The discharge count ratio reflects the symmetry of the discharge in the positive and negative half-cycles. For discharge types with good symmetry, the discharge counts in the positive and negative half-cycles are close, and the discharge count ratio is close to 1; for asymmetrical discharges, the discharge count ratio deviates significantly from 1. The calculation formula is:

[0067] In the formula, The ratio of discharge times is dimensionless. The number of positive half-cycle discharges, i.e., the total number of discharge pulses detected in the power frequency phase range of 0 to 180 degrees, is dimensionless; This represents the number of negative half-cycle discharges, i.e., the total number of discharge pulses detected within the power frequency phase range of 180 degrees to 360 degrees, and is dimensionless. The average discharge phase is calculated. The discharge phase mean reflects the concentration trend of the discharge pulse within the power frequency cycle. The calculation formula is:

[0068] In the formula, This is the average phase value, in degrees. For the first The center phase value of each phase window, in degrees; For the first The number of discharges within a phase window is dimensionless. This represents the summation over all phase windows. In actual calculations, all phase windows are traversed, and the center phase value of each phase window is multiplied by the number of discharges within that window. Then, the products of all phase windows are summed, and finally, the result is divided by the total number of discharges to obtain the weighted average phase mean. Besides the discharge count ratio and phase mean, higher-order statistical features such as phase standard deviation, phase skewness, and phase kurtosis are also calculated. Phase standard deviation measures the dispersion of the discharge pulse phase distribution; a larger standard deviation indicates a more dispersed discharge phase distribution, while a smaller standard deviation indicates a more concentrated discharge phase distribution. Phase skewness measures the symmetry of the discharge pulse phase distribution; zero skewness indicates a symmetrical distribution, a positive skewness indicates a skew towards higher phases, and a negative skewness indicates a skew towards lower phases. Phase kurtosis measures the sharpness of the discharge pulse phase distribution; a larger kurtosis indicates a sharper distribution, and a smaller kurtosis indicates a flatter distribution. These phase domain statistical features characterize the distribution pattern of discharge pulses within the power frequency cycle from different perspectives. Different types of discharges exhibit significant differences in these features, providing a basis for discharge type identification.

[0069] Step S33: Statistically analyze the amplitude of all discharge pulses, and calculate the maximum discharge quantity, average discharge quantity, and standard deviation of discharge quantity. The maximum discharge quantity reflects the peak level of the discharge intensity, the average discharge quantity reflects the overall level of the discharge intensity, and the standard deviation of discharge quantity reflects the dispersion of the discharge amplitude. Average discharge quantity... The calculation formula is:

[0070] In the formula, This is the average discharge quantity, and the unit is the same as the unit for measuring discharge quantity. For the first The amplitude of each discharge pulse is measured in the same unit as the discharge quantity measurement unit. The total number of discharge pulses is dimensionless. Indicates to From 1 to Summation. Calculate the sum of discharge quantities for the positive and negative half-cycles separately. The sum of discharge quantities in the positive half-cycle is the sum of the amplitudes of all discharge pulses in the positive half-cycle, and the sum of discharge quantities in the negative half-cycle is the sum of the amplitudes of all discharge pulses in the negative half-cycle. Then calculate the discharge quantity ratio, which is equal to the sum of discharge quantities in the positive half-cycle divided by the sum of discharge quantities in the negative half-cycle. The discharge quantity ratio reflects the symmetry of the discharge energy distribution in the positive and negative half-cycles. For discharges with good symmetry, the discharge quantity ratio is close to 1; for discharges with significant polarity effects, the discharge quantity ratio deviates significantly from 1. In addition to the above basic statistics, characteristics such as the quantiles of the discharge amplitude, the skewness of the amplitude distribution, and the kurtosis can also be calculated. These amplitude domain statistical characteristics comprehensively describe the distribution law and statistical characteristics of the discharge pulse amplitude. Different types of discharges differ in amplitude distribution, and extracting these characteristics provides important information for discharge type identification.

[0071] Step S34 coarse-grained the denoised discharge signal according to different scale factors, calculating the sample entropy for the coarse-grained sequence at each scale to obtain the multi-scale entropy feature vector. Multi-scale entropy analysis is a method for evaluating the complexity of time series. By calculating the sample entropy at different time scales, it reveals the irregularity and complexity of the signal at multiple time scales. Sample entropy is an improved approximate entropy used to quantify the complexity and unpredictability of time series. The larger the sample entropy value, the more irregular the sequence; the smaller the sample entropy value, the more regular the sequence. Different types of discharge signals exhibit different patterns in multi-scale entropy characteristics. Stable periodic discharge signals usually have smaller multi-scale entropy values, while highly random discharge signals usually have larger multi-scale entropy values. Coarse-graining involves segmenting the original sequence according to different scale factors and averaging the values. For a scale factor of 1... Coarsening, dividing the original sequence into... The data points are divided into groups, and the average value of the data points within each group is taken as the coarsening value for that group. This results in the length being shortened to the original sequence length divided by [the original length]. The coarse-grained sequences are calculated. Sample entropy is calculated for coarse-grained sequences at different scale factors. The calculation of sample entropy involves pattern matching and conditional probability estimation. Specific calculation steps include setting the embedding dimension and tolerance threshold, constructing embedding vectors, counting the number of similar vector pairs, calculating the conditional probability and taking its logarithm, and finally obtaining the sample entropy value. The scale factor is preferably selected from 1 to a certain upper limit value. Scale 1 corresponds to the original sequence, scale 2 corresponds to the sequence after averaging every two points, and so on. The sample entropy values ​​calculated at different scales are arranged in scale order to form a multi-scale entropy feature vector. This vector characterizes the complexity variation of the signal at different time scales, providing temporal complexity-level feature information for discharge type identification.

[0072] Step S35 arranges the phase domain statistical features, amplitude domain statistical features, and multi-scale entropy features in sequence to form a discharge feature vector. The phase domain statistical features include several feature components such as the discharge frequency ratio, phase mean, phase standard deviation, phase skewness, and phase kurtosis. These features describe the temporal distribution of the discharge from the perspective of power frequency phase. The amplitude domain statistical features include several feature components such as maximum discharge quantity, average discharge quantity, discharge quantity standard deviation, and discharge quantity ratio. These features describe the intensity distribution of the discharge from the perspective of discharge amplitude. The multi-scale entropy features include sample entropy values ​​at different scales. These features describe the irregularity and multi-scale characteristics of the discharge signal from the perspective of time domain complexity. These three types of features are concatenated in a certain order to form a high-dimensional discharge feature vector. This feature vector integrates information from the phase domain, amplitude domain, and time domain complexity, comprehensively characterizing the multifaceted characteristics of the discharge signal. Different types of discharge occupy different regions in this feature space, providing input features with good discriminative capabilities for subsequent intelligent recognition algorithms. Through the dual-domain feature extraction in steps S31 to S35, high-level statistical and complexity features are extracted from the original time-domain waveform signal, realizing the transformation from data to information and laying the feature foundation for intelligent identification of discharge types.

[0073] Step S4 is the intelligent discharge type identification step. In this step, the denoised discharge signal is input into a pre-trained attention-enhanced convolutional neural network model. Deep learning methods are used to automatically extract deep features of the signal and perform classification and identification, outputting the discharge type identification result. This step includes three sub-steps: steps S41 to S43.

[0074] Step S41 involves truncating or interpolating the denoised discharge signal into a fixed-length one-dimensional sequence and then normalizing it. Convolutional neural networks require input data to have a fixed dimension, but the actual length of the collected discharge signals may vary; therefore, signal length normalization is necessary. For signals exceeding the fixed length, truncation is used to retain the main part of the signal; for signals shorter than the fixed length, interpolation is used to extend the signal to the fixed length using methods such as linear interpolation or spline interpolation. The fixed length is preferably determined based on the typical duration of the discharge signal and the network input requirements. The purpose of normalization is to eliminate the influence of different dimensions and numerical ranges on model training, improving the model's convergence speed and generalization ability. The normalization formula is:

[0075] In the formula, The normalized signal value is dimensionless. This is the original signal value, and the unit is the same as the original signal. This is the minimum value of the signal sequence, with the same unit as the original signal; The maximum value of the signal sequence is represented by the same unit as the original signal. After normalization, the signal value is mapped to the range of 0 to 1, eliminating the influence of dimensions. Signals with different amplitude ranges are comparable after normalization. The normalized signal is then used as input to a convolutional neural network for feature extraction and classification.

[0076] Step S42 passes the normalized signal sequentially through multiple residual convolutional blocks. Each residual convolutional block contains a convolutional layer, a batch normalization layer, and an activation function. Within the residual convolutional block, global average pooling is performed on the feature map output by the convolution to obtain a channel description vector. Two fully connected layers then generate channel weight vectors, which are used to scale the feature map channel by channel. Convolutional neural networks (CNNs) are deep learning models specifically designed to process data with a grid-like topological structure. They automatically learn local features and hierarchical representations of data through convolutional operations. The residual convolutional block is the basic building block in a CNN. The residual structure introduces cross-layer connections, allowing gradients to propagate directly to previous layers, alleviating the gradient vanishing problem in deep networks, enabling deeper network training and better performance. The main path of each residual convolutional block contains several convolutional layers. These layers perform convolution operations on the input feature map using convolution kernels to extract local feature patterns. A batch normalization layer follows the convolutional layers, normalizing the data in each batch, accelerating model convergence and improving training stability. After the batch normalization layer, an activation function is connected, introducing a non-linear transformation. Commonly used activation functions include ReLU and its variants. The bypass connection of the residual convolutional block directly adds the input to the output of the main path, forming a residual learning structure. Inside the residual convolutional block, an SE attention module is introduced to adaptively weight the feature channels. The SE attention module works by first performing global average pooling on the feature map of the convolution output, compressing the two-dimensional feature map of each feature channel into a scalar value, forming a channel description vector. This vector contains global information for each feature channel. Then, the channel description vector is input into two fully connected layers. The first fully connected layer performs dimensionality reduction and introduces non-linearity through an activation function; the second fully connected layer performs dimensionality increase and maps the output value to between 0 and 1 using a sigmoid activation function, obtaining the channel weight vector. Each element of the channel weight vector corresponds to the importance weight of a feature channel; a weight value close to 1 indicates high channel importance, and a weight value close to 0 indicates low channel importance. Channel-wise scaling of the original feature map is performed using channel weight vectors, which involves multiplying the feature map of each feature channel by its corresponding weight value to achieve adaptive weighting for different feature channels. Through the SE attention module, the network automatically learns the relative importance of each feature channel, strengthening the representation of useful feature channels, suppressing redundant or noisy feature channels, and improving the discriminative power of the feature representation. The signal is processed sequentially through multiple residual convolutional blocks, gradually increasing the number of network layers and the level of feature abstraction, progressively extracting semantic features from lower-level local waveform features to higher-level semantic features.

[0077] Step S43 performs global average pooling on the output of the last residual convolutional block to obtain a global feature vector. After mapping through a fully connected layer, the Softmax function is used to calculate the probability of each category, and the category with the highest probability is selected as the recognition result. Global average pooling averages all positions of the feature map, compressing the two-dimensional feature map of each feature channel into a scalar value, thus converting the multi-dimensional feature map into a one-dimensional feature vector. Global average pooling has the advantages of fewer parameters and preventing overfitting. The global feature vector obtained after global average pooling contains high-level abstract features of the signal, and this vector serves as the input to the fully connected layer. The fully connected layer maps the high-dimensional feature vector to the category space. The number of output nodes of the fully connected layer is equal to the number of discharge types, and each output node corresponds to a discharge type. The Softmax function normalizes the output of the fully connected layer, converting the output value into a probability distribution. The calculation formula for the Softmax function is: In the formula, For the first The predicted probability of a class is dimensionless. For the fully connected layer One output, dimensionless; It is a natural exponential function, expressed in terms of the natural constant. Exponential operations with base 0; The total number of categories, dimensionless; Indicates to From 1 to Summation. The Softmax function ensures that the sum of the probabilities of each category equals 1, and the probability value of each category is between 0 and 1. The category with the highest probability value is selected as the model's recognition result, which is the discharge type determined by the model. Through the processing of steps S41 to S43, the attention-enhanced convolutional neural network model automatically learns discriminative features and performs classification from the original one-dimensional discharge signal, without the need for manual design of feature extraction rules, and has powerful feature learning capabilities and classification performance. Before use, the model needs to be trained offline on a large amount of labeled data. During the training process, the model parameters are continuously updated through backpropagation and optimization algorithms until the model's performance on the validation set meets the requirements. The trained model parameters are stored in the data storage module 32 of the edge computing node unit 3. During online detection, the parameters are directly loaded for forward inference to achieve rapid discharge type recognition.

[0078] Step S5 is the multi-sensor information fusion and early warning step. Based on the identification results of each sensor channel, this step uses evidence theory to fuse information, obtain a comprehensive diagnostic conclusion, and output tiered early warning information based on the diagnostic conclusion. It includes three sub-steps: S51 to S53.

[0079] Step S51 constructs a basic probability assignment based on the neural network output probability values ​​of each sensor channel. Evidence theory is a mathematical framework for handling uncertain information, using a basic probability assignment function to represent the degree of support of evidence for a proposition. For each sensor channel, the neural network model outputs predicted probabilities for each category; these probability values ​​reflect the model's confidence level in each category. To convert the neural network output into a basic probability assignment within the evidence theory framework, the uncertainty of the model's predictions needs to be considered. The formula for constructing the basic probability assignment is:

[0080] In the formula, For the proposition The basic probability allocation value, Indicates the discharge type is number 1 Propositions of a class are dimensionless; For the first The predicted probability of a class, output by the neural network model, is dimensionless; The uncertainty coefficient characterizes the inherent uncertainty in the model's predictions, ranging from 0 to 1, and is dimensionless. The residual probability is assigned to the entire recognition framework, representing the uncertainty support for all possible categories. The formula for calculating the residual probability is:

[0081] In the formula, For identification framework The basic probability allocation value, This represents the set of all possible discharge types and is dimensionless. The total number of categories, dimensionless; This represents the summation over all categories. In this way, the deterministic output of the neural network is transformed into evidence containing uncertainty information, providing input for subsequent evidence fusion. Uncertainty coefficient. The value of can be set according to the model's performance on the validation set. The higher the model accuracy, the smaller the uncertainty coefficient can be set; the lower the model accuracy, the larger the uncertainty coefficient should be set to reflect the reliability of the model's prediction.

[0082] Step S52 uses the Dempster synthesis rule to fuse evidence from multiple sensor channels. The Dempster synthesis rule is a fundamental rule in evidence theory used to fuse multiple independent pieces of evidence. By combining information from multiple evidence sources, a more reliable comprehensive judgment is obtained than from a single evidence source. The basic probability allocation calculation formula after fusion is as follows:

[0083] In the formula, For the integrated proposition The basic probability distribution value is dimensionless. The first piece of evidence against the proposition The assigned value is dimensionless; The second piece of evidence supports the proposition. The assigned value is dimensionless; Expressing a proposition With proposition The intersection equals the proposition That is, summation is performed only on combinations of propositions that satisfy the intersection condition; As the normalization factor, ,in Represents the empty set. It reflects the degree of conflict between pieces of evidence and is dimensionless. It means that for all satisfying Summation of propositional combinations with given conditions. Normalization factor. The purpose of fusion is to normalize the fused probability distributions, ensuring that the sum of the basic probability distributions for each proposition is 1. When multiple sensor channels exist, a sequential fusion strategy can be employed. First, the first two pieces of evidence are fused to obtain an intermediate result. Then, this intermediate result is fused with the third piece of evidence, and so on, until all evidence is fused. Through evidence fusion, information from different sensors is integrated. Mutually supporting evidence increases the probability distribution value of the corresponding category, while contradictory evidence is handled through conflict terms. The final fused result is theoretically more reliable and accurate than the judgment from a single sensor. Multi-sensor information fusion utilizes the complementarity and redundancy between sensors, improving diagnostic robustness and reducing the false positive rate.

[0084] Step S53 selects the discharge type with the highest probability value from the fused basic probability distribution as the diagnostic conclusion. If the maximum probability value is greater than the preset decision threshold, the diagnosis is confirmed as valid, and warning information is output according to the preset grading standard based on the discharge type and discharge intensity. All propositions corresponding to discharge types are traversed, and the proposition with the highest basic probability distribution value is found. The discharge type corresponding to this proposition is the comprehensive diagnostic conclusion determined by the system. To avoid making unreliable judgments when evidence is highly uncertain or conflicting, a decision threshold is set. The diagnostic conclusion is considered valid only when the maximum probability value exceeds this threshold. The value of the decision threshold is preferably set according to the tolerance for false positives and false negatives in the application scenario. Setting the threshold too high will cause the system to be too conservative and miss reports, while setting it too low will cause the system to be too aggressive and false positives. If the maximum probability value does not exceed the decision threshold, the system outputs a diagnostic uncertainty prompt, suggesting further manual verification or re-judgment after increasing the amount of detection data. Warning information is output according to the determined discharge type and discharge intensity, based on the preset grading standard. The grading standard for the warning information can be set according to the severity of the discharge type and the magnitude of the discharge intensity, for example, dividing the warning into several levels such as normal, attention, warning, and alarm. For discharge types with minor hazards and low discharge intensity, a normal or warning level alert is output; for discharge types with moderate hazards or moderate discharge intensity, a warning level alert is output and enhanced monitoring is recommended; for discharge types with severe hazards or very high discharge intensity, an alarm level alert is output and immediate shutdown for maintenance is recommended. The warning information is uploaded to the upper-level monitoring platform 4 via communication interface module 33. The upper-level monitoring platform 4 takes corresponding response measures based on the warning level, including audible and visual alarms, information push notifications, and work order dispatch, ensuring that maintenance personnel are promptly informed of equipment anomalies and can take appropriate measures. Through multi-sensor information fusion and early warning in steps S51 to S53, the system comprehensively utilizes the detection information from multiple sensors to obtain more reliable diagnostic conclusions than a single sensor, and promptly issues warnings to maintenance personnel. This provides decision support for preventative maintenance and fault prevention, effectively reducing equipment failure risks and ensuring the safe and stable operation of the power supply system.

[0085] Example 1: This example applies the discharge detection system and method of the present invention to a 10kV high-voltage switchgear in a central substation of a coal mine. The substation has 12 high-voltage switchgears, and three switchgears that have been in operation for more than 8 years were selected for online monitoring. The system configuration is as follows: Each switchgear is equipped with one set of modular sensor assembly 1, including a UHF sensor 11 installed at the ventilation hole on the top of the cabinet, with a center frequency of 500MHz and a bandwidth of 300MHz; three ultrasonic sensors 12, respectively attached to the outer walls of the busbar compartment, circuit breaker compartment, and cable compartment, with a resonant frequency of 40kHz; and a transient ground voltage sensor 13 installed on the inner surface of the cabinet door. The signal acquisition unit 2 has a preamplifier 21 with a gain of 40dB, a bandpass filter 22 with a passband range corresponding to the frequency band of each sensor, a programmable gain amplifier 23 with an adjustable gain range of 0 to 60dB, and an analog-to-digital converter 25 with a sampling frequency of 100MHz and a resolution of 12bit. Edge computing node unit 3 uses an ARM Cortex-A53 processor with a main frequency of 1.5GHz and 2GB of memory.

[0086] The specific detection process follows the method of this invention, with the following specific settings: In step S1, the synchronous trigger module 34 detects the zero-crossing point of the 100V voltage signal on the secondary side of the 10kV bus voltage transformer, generating a synchronous trigger pulse. Each channel analog-to-digital converter 25 starts sampling at the same time, with each sampling lasting for 4 power frequency cycles (80ms), and the number of sampling points is 8000. In step S2, the population size of the sparrow search algorithm is set to 30, the maximum number of iterations is set to 50, and the decomposition mode number is... The search range is 3 to 10, with a penalty factor. The search range was 1000 to 5000, which was optimized to obtain... It is 6. The value is set to 2500; the correlation coefficient threshold is set to 0.3; the wavelet decomposition uses the db4 wavelet with 5 decomposition levels, and the parameters are adjusted. Set to 1.5. In step S3, the phase axis is divided into 36 phase windows, each 10 degrees; the amplitude axis is quantized into 50 levels; 200 power frequency cycle data are accumulated to construct a phase-resolved partial discharge map; the scale range of the multi-scale entropy calculation is 1 to 20, the embedding dimension is 2, and the tolerance threshold is 0.2 times the standard deviation. In step S4, the attention-enhanced convolutional neural network contains 4 residual convolutional blocks, each containing 2 convolutional layers, with a kernel size of 3, and the number of feature channels are 32, 64, 128, and 256 respectively; the compression ratio of the SE attention module is 16; the input signal length is normalized to 2048 points; the output layer is divided into 4 categories, corresponding to corona discharge, surface discharge, internal discharge, and normal state. In step S5, the uncertainty coefficient... Set it to 0.05, and the decision threshold to 0.6.

[0087] Example 2: This example is applied to the monitoring of switchgear in a 35kV substation of a coal mine. The difference from Example 1 is that the ultra-high frequency sensor 11 is installed at the joint of the cabinet, with a center frequency of 800MHz and a bandwidth of 500MHz; the ultrasonic sensor 12 uses a 60kHz resonant frequency; and the analog-to-digital converter 25 has a sampling frequency increased to 200MHz with a resolution of 14bit. Regarding the method parameters, the sparrow search algorithm was optimized to obtain... It is 7. The value is set to 3200; the correlation coefficient threshold is set to 0.25; wavelet decomposition uses the sym8 wavelet with 6 decomposition levels, and the parameters are adjusted. Set to 2.0; the number of phase windows is 72, each window is 5 degrees; the scale range for multi-scale entropy calculation is extended to 1 to 30.

[0088] Example 3: This example is applied to the monitoring of a 6kV switchgear in a mine hoisting room. The main difference from the previous examples is that the center frequency of the ultra-high frequency sensor 11 is adjusted to 300MHz, with a bandwidth of 200MHz, to adapt to lower voltage levels; the analog-to-digital converter 25 has a sampling frequency of 50MHz; and the sparrow search algorithm is optimized. It is 5. The value is set to 1800; the correlation coefficient threshold is set to 0.35; the wavelet decomposition uses the COIF5 wavelet with 4 decomposition levels, and the parameters are adjusted. Set to 1.2; the number of phase windows is 60, each window is 6 degrees; the multi-scale entropy scale range is 1 to 15.

[0089] Comparative Example 1 employs a traditional fixed-parameter variational mode decomposition and soft-threshold denoising method. The system configuration is the same as in Example 1, but instead of using the sparrow search algorithm for optimization in step S2, the number of decomposed modes is manually fixed. The penalty factor is 5. The threshold value is 2000; the denoising process uses a traditional soft threshold function instead of the improved threshold function, and the soft threshold is determined using a general threshold estimation method. The remaining steps are the same as in Example 1.

[0090] Comparative Example 2 employs a single-domain feature extraction method. The system configuration and denoising method are the same as in Example 1, but in step S3, only phase domain statistical features are extracted; amplitude domain statistical features and multi-scale entropy features are not calculated, significantly reducing the feature vector dimension. The remaining steps are the same as in Example 1.

[0091] Comparative Example 3 uses a traditional convolutional neural network. The steps are the same as those in Example 1, but in step S4, a standard convolutional neural network is used, without residual structures and SE attention modules. The network contains four ordinary convolutional layers, with the same kernel size and number of feature channels as in Example 1, but lacks cross-layer connections and channel attention mechanisms. The remaining steps are the same as in Example 1.

[0092] Comparative Example 4 employs a single-sensor detection method. Only the ultra-high frequency sensor 11 is used for detection; the ultrasonic sensor 12 and transient ground voltage sensor 13 are not installed. In step S5, multi-sensor information fusion is not performed; the output result of the single-channel neural network is directly used as the diagnostic conclusion. The remaining steps are the same as in Example 1.

[0093] Experiment 1: Comparative Performance of Examples at Different Voltage Levels. This experiment, based on GB / T 7354-2018 "Partial Discharge Measurement" and DL / T 664-2016 "Application Specification for Infrared Diagnosis of Live Equipment," compares the denoising performance and overall detection performance of Examples 1, 2, and 3 with Comparative Example 1 at different voltage levels and noise levels. Tests were conducted at three voltage levels: 6kV, 10kV, and 35kV. Initial signal-to-noise ratios (SNRs) of 5dB, 10dB, 15dB, and 20dB were set for each voltage level. A needle-plate electrode structure was used to simulate internal discharge, with corresponding voltage amplitudes applied at a frequency of 50Hz. Evaluation metrics included the denoised signal-to-noise ratio (SNR) and discharge detection sensitivity.

[0094] Experimental results are as follows Figure 4 As shown. Figure 4 This is a comparison chart of the denoised signal-to-noise ratio curves of different embodiments and comparative examples under different initial signal-to-noise ratios. From Figure 4As can be seen, in Example 1, with initial signal-to-noise ratios (SNRs) of 5dB, 10dB, 15dB, and 20dB, the SNRs after denoising reached 22.5dB, 28.3dB, 32.7dB, and 35.8dB, respectively. Example 2, due to its higher sampling frequency and more refined parameter optimization, achieved SNRs of 23.8dB, 29.5dB, 33.9dB, and 37.2dB, respectively. Example 3, optimized for lower voltage levels, achieved SNRs of 21.2dB, 27.1dB, 31.5dB, and 34.3dB, respectively. Comparative Example 1, using fixed parameters, only achieved SNRs of 18.2dB, 24.1dB, 28.5dB, and 31.2dB, respectively. The average improvement of the three examples compared to Comparative Example 1 was 22.8%, 26.5%, and 18.3%, respectively, with Example 2 showing the best performance. It can be seen that the denoising method using the sparrow search algorithm to adaptively optimize VMD parameters and improve the wavelet threshold function can achieve stable and excellent denoising effects under different voltage levels and noise conditions. The sparrow search algorithm of this invention adaptively finds the optimal decomposition parameters for different signal characteristics, and the improved threshold function effectively suppresses noise while retaining useful signals through a continuous transition mechanism. The parameter optimization strategy under different voltage levels enables each embodiment to adapt to the corresponding application scenarios.

[0095] Experiment 2: Impact of Feature Extraction Methods on Classification Performance. This experiment, based on the IEC 60270-2015 international standard "Voltage Testing Techniques - Partial Discharge Measurement," compares the performance differences in feature extraction and classification between Example 1, Comparative Example 2, and Comparative Example 3. This experiment uses a standard discharge model to generate four types of discharge: corona discharge, surface discharge, internal discharge, and floating potential discharge, collecting 500 samples for each type. Example 1 uses dual-domain feature extraction and attention-enhanced CNN; Comparative Example 2 extracts only phase domain features but uses attention-enhanced CNN; and Comparative Example 3 uses dual-domain feature extraction but employs a traditional CNN. The evaluation metric is the confusion matrix for the four discharge types, which visually displays the recognition accuracy and misclassification rates for each category.

[0096] Experimental results are as follows Figure 5 As shown. Figure 5 The confusion matrix heatmaps of Example 1, Comparative Example 2, and Comparative Example 3 are compared. Figure 5 In Figure (a), the confusion matrix of Example 1 is shown. The diagonal elements are 97.0%, 96.5%, 98.0%, and 95.5%, respectively, indicating that the recognition accuracy of the four discharge types is over 95%. The off-diagonal elements are all less than 3%, indicating that the false positive rate is very low. Figure 5In Figure (b), the confusion matrix of Comparative Example 2 is shown. The diagonal elements are 89.0%, 87.5%, 91.0%, and 86.5%, respectively, with an average accuracy of 88.5%, which is 8.3 percentage points lower than that of Example 1. The cross-misclassification rate of corona discharge and surface discharge reaches 6.5%, indicating that the single-domain feature separability is insufficient. Figure 5 In Figure (c), the confusion matrix of Comparative Example 3 is shown, with diagonal elements of 92.0%, 91.5%, 94.0%, and 90.5%, respectively. The average accuracy is 92.0%, a decrease of 4.8 percentage points compared to Example 1. The cross-misclassification rate between internal and floating potential discharges reaches 4.2%, indicating that the traditional CNN feature extraction capability is limited. These results show that the dual-domain feature extraction method comprehensively characterizes discharge characteristics from three dimensions: phase domain, amplitude domain, and temporal complexity, significantly improving class separability. Attention-enhanced CNN learns deeper discriminative features through residual structure and channel attention mechanism; the combination of the two achieves optimal recognition performance. Its technical principle lies in the fact that phase domain features reflect the temporal distribution law of discharge, amplitude domain features reflect the intensity distribution law of discharge, and multi-scale entropy features reflect the complexity characteristics of the signal. The complementarity of the three types of features makes different types of discharge occupy clearly separated regions in the feature space; the SE attention module enhances useful feature channels through adaptive weighting, suppresses redundant features, and improves the discriminative ability of feature representation.

[0097] Experiment 3: Stability of Recognition Performance at Different Voltage Levels; verifying the recognition performance stability of Examples 2, 3, and Comparative Example 3 in different application scenarios. This experiment tested the discharge recognition accuracy at four voltage levels: 6kV, 10kV, 20kV, and 35kV. Example 2 was optimized for 35kV, Example 3 for 6kV, and Comparative Example 3 used a fixed network structure. 100 samples were tested at each voltage level, including four discharge types. The evaluation metrics were the average recognition accuracy and the standard deviation of accuracy at different voltage levels; a smaller standard deviation indicates more stable performance.

[0098] Experimental results are as follows Figure 6 As shown. Figure 6 The image shows a scatter plot of the recognition accuracy at different voltage levels for Examples 2, 3, and Comparative Example 3. Figure 6As can be seen, the recognition accuracy rates of Example 2 at 6kV, 10kV, 20kV, and 35kV were 95.2%, 96.8%, 97.1%, and 97.5%, respectively, with an average accuracy of 96.65% and a standard deviation of 0.93%. The recognition accuracy rates of Example 3 at the four voltage levels were 95.2%, 94.8%, 93.5%, and 92.0%, respectively, with an average accuracy of 93.88% and a standard deviation of 1.35%. The recognition accuracy rates of Comparative Example 3 at the four voltage levels were 90.5%, 92.0%, 91.5%, and 91.0%, respectively, with an average accuracy of 91.25% and a standard deviation of 0.62%. Although Example 2 has a slightly larger standard deviation than Comparative Example 3, its average accuracy is 5.4 percentage points higher, and it performs best at the 35kV target voltage level. Example 3 performs best at the 6kV target voltage level, but its accuracy decreases as the voltage level increases, highlighting the importance of targeted optimization. Comparative Example 3 has relatively stable accuracy, but its overall level is low, and the lack of an attention mechanism results in insufficient adaptability to different scenarios. It can be seen that the examples with parameter optimization for specific voltage levels achieve optimal performance in the target scenario and maintain good performance in other scenarios as well. The attention-enhanced CNN has stronger feature learning and generalization capabilities than traditional CNNs. The spectral characteristics and amplitude distribution of discharge signals differ at different voltage levels. This invention can better adapt to the target scenario by adjusting sensor parameters, sampling frequency, and algorithm parameters. The SE attention module adaptively adjusts the feature channel weights according to the input signal, enabling the model to extract effective discriminative features in different scenarios.

[0099] Experiment 4: Multi-sensor fusion reliability comparison experiment. This experiment, based on the power industry standard DL / T 596-2021 "Preventive Testing Procedures for Power Equipment," compares the diagnostic reliability of Examples 1, 2, and 3 with that of Comparative Example 4 under different signal quality conditions. This experiment selected 300 discharge samples collected on-site, divided into three groups according to signal quality: Group 1 consisted of 100 high-quality signals, Group 2 of 100 medium-quality signals, and Group 3 of 100 low-quality signals. Low-quality signals included situations such as sensor signal interference, low signal-to-noise ratio, and ambiguous judgments from a single sensor. Examples 1, 2, and 3 all employed multi-sensor fusion diagnosis, while Comparative Example 4 used only single-sensor diagnosis. Evaluation indicators included diagnostic accuracy and confidence distribution.

[0100] Experimental results are as follows Figure 7 As shown. Figure 7 The diagram shows a combination of diagnostic accuracy curves and confidence box plots for different embodiments and comparative examples under different signal qualities. Figure 7In Figure (a), the diagnostic accuracy curves are shown. Example 1 achieved accuracy rates of 98.0%, 96.5%, and 93.8% on high-quality, medium-quality, and low-quality samples, respectively. Example 2 achieved 98.5%, 97.2%, and 94.5%, respectively. Example 3 achieved 97.5%, 95.8%, and 92.5%, respectively. Comparative Example 4 achieved 95.0%, 89.5%, and 82.5%, respectively. It can be seen that when signal quality deteriorates, the accuracy of the three examples decreases by 4.2%, 4.0%, and 5.0%, respectively, while the decrease in Comparative Example 4 reaches 12.5%, indicating that multi-sensor fusion diagnosis has stronger robustness. Figure 7 (b) is a box plot of confidence levels. The median diagnostic confidence levels for Examples 1, 2, and 3 are 0.92, 0.94, and 0.90, respectively. The narrow box range and high position indicate stable and high diagnostic confidence. Comparative Example 4 has a median diagnostic confidence level of 0.78, a wider box range, and many low-confidence outliers, indicating that single-sensor diagnosis suffers a significant drop in confidence when signal quality is poor. These results demonstrate that multi-sensor information fusion significantly improves the accuracy and reliability of diagnosis by integrating detection information from different sensors, especially under poor signal quality conditions. Examples optimized for different voltage levels all perform excellently in fusion diagnosis. This invention uses different sensors based on different physical detection principles to observe the same discharge event from multiple angles. When the signal of one sensor is interfered with, the information from other sensors can supplement and verify it. The Dempster synthesis rule of evidence theory is used for fusion. Mutually supporting evidence increases the probability of correct judgment, while contradictory evidence is processed through conflict terms. The final fusion result is more reliable and has higher diagnostic confidence than single-sensor judgment.

[0101] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A discharge detection device for a general-purpose high and low voltage switchgear used in mining, characterized in that, It includes modular sensor components, signal acquisition units, edge computing node units, and upper-level monitoring platforms; The modular sensor assembly is installed in the switch cabinet and includes a sensor body, which is an ultra-high frequency sensor, an ultrasonic sensor, or a transient ground voltage sensor. The signal acquisition unit includes a preamplifier, a bandpass filter, a programmable gain amplifier, an anti-aliasing filter, and an analog-to-digital converter connected in sequence. The preamplifier is connected to the modular sensor assembly, and the analog-to-digital converter is connected to the edge computing node unit. The edge computing node unit includes an embedded processor, a data storage module, a communication interface module, and a synchronization triggering module. The embedded processor is connected to the analog-to-digital converter, and the synchronization triggering module is connected to the trigger input terminal of each analog-to-digital converter. The upper-level monitoring platform is connected to the edge computing node unit through the communication interface module.

2. The discharge detection device according to claim 1, characterized in that, The modular sensor assembly also includes a mounting base and a quick-release interface; The mounting base is located at the bottom of the sensor body and is connected to a magnetic base with an embedded permanent magnet array. The quick-release interface is located on the side of the sensor body; The sensor body has a multi-layer shielding structure, consisting of an equalizing shielding layer, an insulating isolation layer, and an electromagnetic shielding layer from the outside to the inside.

3. The discharge detection device according to claim 1, characterized in that: The ultra-high frequency sensor is installed in the ventilation hole on the top of the cabinet or in the gap between the cabinet panels. The ultrasonic sensor is attached to the outer wall of the busbar compartment, circuit breaker compartment, or cable compartment. The transient ground voltage sensor is attached to the inner surface of the cabinet door or near the grounding busbar.

4. The discharge detection device according to claim 1, characterized in that: The control terminal of the programmable gain amplifier is connected to the embedded processor; The synchronization triggering module is equipped with a power frequency voltage sampling input port, which generates a synchronization trigger pulse by detecting the zero crossing point of the power frequency voltage. The communication interface module is connected to the embedded processor.

5. The discharge detection device according to claim 2, characterized in that, The modular sensor assembly is connected to the signal acquisition unit via a shielded signal cable; The shielded signal cable adopts a double-layer shielding structure, including an inner aluminum foil shielding layer and an outer metal braided shielding layer; The shielded signal cable is equipped with aviation plugs at both ends that mate with the quick-release interface.

6. A discharge detection method using the apparatus according to any one of claims 1-5, characterized in that, include: S1. Multi-source signal acquisition: Using ultra-high frequency sensors, ultrasonic sensors and transient ground voltage sensors, with the zero-crossing point of the power frequency voltage as the synchronization reference, discharge signals are acquired synchronously, and after signal conditioning and analog-to-digital conversion, multi-channel raw discharge signals are obtained. S2. Adaptive signal denoising: The sparrow search algorithm is used to determine the number of decomposition modes and the penalty factor of variational mode decomposition. The original discharge signal is decomposed to obtain the intrinsic mode function components. The improved wavelet threshold function is used for denoising. The denoised discharge signal is obtained by superposition and reconstruction. S3. Dual-domain feature extraction: The denoised discharge signal is mapped to power frequency phase and discharge amplitude to construct a phase-resolved partial discharge spectrum, extract phase domain statistical features and amplitude domain statistical features, calculate multi-scale sample entropy, and combine them to form a discharge feature vector. S4. Discharge type identification: The denoised discharge signal is input into the attention-enhanced convolutional neural network model. Features are extracted through residual convolutional blocks and SE attention modules. The discharge type identification result is output after global average pooling and fully connected layers. S5. Information Fusion and Early Warning: Construct a basic probability allocation based on the identification results of each sensor channel, fuse evidence from multiple sensors, and output hierarchical early warning information.

7. The discharge detection method according to claim 6, characterized in that, S2 includes: S21. Constructing the objective function The optimization objective is to minimize the sum of the envelope entropies of all intrinsic mode function components; where, To decompose the mode number; As a penalty factor; For the first Envelope entropy of each component; S22. Initialize the sparrow population and encode their positions as follows: and The combination of these parameters updates the position based on three roles: discoverer, follower, and vigilant, and iterates to obtain the optimal parameters. and ; S23, Use and The original discharge signal is subjected to variational mode decomposition, the correlation coefficient between each component and the original signal is calculated, and the components with correlation coefficients below the threshold are removed. S24. Perform wavelet decomposition on the effective components and process the wavelet coefficients using an improved threshold function: when hour, ;when hour, ;in, These are the processed wavelet coefficients; These are the original wavelet coefficients; For the threshold; To adjust the parameters; S25. Perform inverse transform on the denoised wavelet coefficients to reconstruct them, and superimpose the components to obtain the denoised discharge signal.

8. The discharge detection method according to claim 6, characterized in that, S3 includes: S31. Map the discharge pulses to a two-dimensional plane according to the power frequency phase and amplitude, divide the phase window and amplitude level, accumulate multiple power frequency cycle data, count the number of pulses in each grid, and obtain the phase-resolved partial discharge spectrum. S32. Count the number of discharges during the positive and negative half-cycles. and Calculate the ratio of discharge times ; Calculate the phase mean ;in, For the first The center phase value of each phase window; The number of discharges within the phase window is calculated; the phase standard deviation, skewness, and kurtosis are calculated to obtain the statistical characteristics of the phase domain. S33. Calculate the maximum discharge quantity and the average discharge quantity. and the standard deviation of discharge quantity; among which, For the first Each pulse amplitude; The total number of pulses is calculated; the ratio of positive to negative half-cycle discharge is calculated to obtain the statistical characteristics of the amplitude domain. S34. Coarse-grain the denoised discharge signal at different scales, calculate the sample entropy at each scale, and obtain multi-scale entropy features. S35. Combine phase domain features, amplitude domain features and multi-scale entropy features to form a discharge feature vector.

9. The discharge detection method according to claim 6, characterized in that, The neural network model in S4 includes an input layer, multiple residual convolutional blocks, a global average pooling layer, a fully connected layer, and an output layer, with an SE attention module embedded within each residual convolutional block; S4 includes: S41. Extract or interpolate the denoised discharge signal into a fixed-length sequence, according to... Normalization; among which, The original signal value; , These are the minimum and maximum values, respectively. S42. The signal passes through each residual convolution block in sequence. Within the block, the feature map is subjected to global average pooling to obtain the channel description vector. Channel weights are generated through the fully connected layer and the feature map is weighted channel by channel. S43. Perform global average pooling on the output of the last residual convolutional block, and then use the Softmax function to calculate the probabilities of each class after passing through a fully connected layer. ;in, For the first The output class is selected; the class with the highest probability is chosen as the recognition result.

10. The discharge detection method according to claim 6, characterized in that, S5 includes: S51, based on the output probability of each sensor channel Constructing basic probability assignments ;in, For the first Discharge type propositions; The uncertainty coefficient is used; the remaining probability is assigned to the recognition frame. ; S52. Multi-sensor evidence is fused using the Dempster synthesis rule. The fusion formula is as follows: The summation condition is ;in, , Each of the two pieces of evidence supports the proposition. , The assigned value; express and The intersection is ; As the normalization factor, The condition for summation is: , It is an empty set; S53. Select the discharge type with the highest probability after fusion as the diagnostic conclusion. If the highest probability exceeds the decision threshold, the diagnosis is valid. Output graded warnings according to discharge type and intensity.

Images