A Smart Diagnostic Method for Switchgear Based on Multi-Source Data Fusion

Abstract

This application relates to a multi-source data fusion-based intelligent diagnostic method for switchgear, comprising: real-time acquisition of multi-source operating data from the switchgear; unified time synchronization and timestamp annotation based on a precision clock protocol; multi-channel synchronous latching via a hardware trigger bus when abnormal triggering conditions are detected; preprocessing and feature extraction of the latched data; time axis alignment and dynamic weight allocation based on contribution of different modal features for fusion calculation; execution of rule logic judgment for rapid risk screening; if no alarm is triggered, inputting the fused features into a pre-trained deep learning model to output fault category probability and risk score; generating a health status assessment conclusion and uploading it in a structured package. This application enables high-speed parallel acquisition of multi-source data, hardware-level synchronous alignment, and edge fusion inference at the equipment site, effectively improving the accuracy and real-time response capability of complex fault diagnosis, and reducing dependence on cloud computing and communication bandwidth.

Description

Technical Field

[0001] This application relates to the field of power equipment condition monitoring and intelligent operation and maintenance technology, specifically to an intelligent diagnostic system and method for switchgear based on multi-source data fusion and edge computing architecture. Background Technology

[0002] Switchgear is a crucial power distribution and control device in power systems, widely used in substations, industrial enterprises, and various power distribution locations. Its operational safety directly impacts power supply reliability. During long-term operation, switchgear may experience potential faults such as partial discharge, abnormal temperature rise, electrical fluctuations, and abnormal vibration due to insulation aging, poor contact, mechanical loosening, overload, or environmental factors. Failure to detect and warn of these issues in a timely manner could lead to insulation breakdown, short circuits, or even power outages.

[0003] Existing switchgear condition monitoring technologies mostly use single or a small number of sensors for detection, resulting in limited information dimensions and difficulty in accurately identifying complex faults, easily leading to false alarms or missed alarms. Meanwhile, high-speed signals such as partial discharge and ultrasound require high real-time acquisition and processing capabilities, which traditional single-processor-based software architectures cannot meet on-site real-time requirements; while some centralized cloud-based analysis solutions require uploading large amounts of raw data, consuming communication bandwidth and suffering from network latency and reliability issues.

[0004] Therefore, there is an urgent need for an intelligent diagnostic system for switchgear that can achieve parallel acquisition of multi-source data, hardware-level time synchronization, real-time fusion reasoning at the edge, and output diagnostic results, and to develop a complete set of multi-source data collaborative detection and intelligent diagnostic methods. Summary of the Invention

[0005] To address the issues of limited information dimensions, insufficient real-time performance, and over-reliance on cloud computing in existing technologies, this application provides a multi-source data fusion-based intelligent diagnostic edge computing system and method for switchgear, which can complete the entire detection process from data acquisition, time synchronization, feature extraction, fusion analysis to risk assessment and result output on-site.

[0006] This application provides a method for intelligent diagnosis of switchgear using multi-source data fusion, including the following steps:

[0007] Step S100: Real-time acquisition of multi-source operating data during the operation of the switchgear. The multi-source operating data includes at least partial discharge signals, temperature information, electrical quantity data, vibration signals, and environmental parameters to form a multi-modal raw data set. Step S200: Based on the precision clock protocol, the collected multi-source operation data is uniformly synchronized and timestamped to ensure that the data of different modes are under the same time reference. Step S300: When any modal data is detected to meet the preset abnormal triggering condition, a synchronous latching instruction is broadcast to all acquisition channels through the hardware trigger bus, so that each acquisition channel latches data within the same time window, forming a detection sample under a unified time window. Step S400: Preprocess and extract features from the latched data. The preprocessing and feature extraction include filtering and noise reduction, frequency domain feature extraction, time-frequency energy feature extraction, and pulse statistical feature extraction to obtain key feature information reflecting the device status. Step S500: Perform time axis alignment and scale normalization on the key feature information of different modalities, and dynamically allocate fusion weights according to the contribution of each modal feature to the current fault type to construct a unified fusion feature vector; Step S600: Based on the unified fusion feature vector, perform preset rule logic judgment to determine the trend and statistics of key features by threshold judgment in order to conduct rapid risk screening. Step S700: If no immediate alarm is triggered, the unified fusion feature vector is input into the pre-trained deep learning diagnostic model, and the fault category probability distribution and comprehensive risk score are output through forward inference operation. Step S800: Determine the fault category and risk level based on the fault category probability distribution and comprehensive risk score, and generate the corresponding health status assessment conclusion; Step S900: The health status assessment conclusion, risk level and related data fragments are structured and packaged, and uploaded to the monitoring platform to realize the closed loop of online detection and diagnosis process.

[0008] In one possible implementation, in step S100, the multi-source operating data is acquired through six parallel field-programmable gate arrays (FPGAs), with the following specific functions: the first FPGA is used for partial discharge high-frequency channel acquisition and pulse detection; the second FPGA is used for ultrasonic or acoustic emission broadband channel acquisition; the third FPGA is used for current and voltage synchronous sampling and phase information extraction; the fourth FPGA is used for infrared array temperature data acquisition and frame buffering; the fifth FPGA is used for point-based temperature and environmental parameter acquisition; and the sixth FPGA serves as the main synchronization unit for global clock synchronization, triggering broadcasts, and system-level acquisition control. The raw high-frequency waveform data is not transmitted between the FPGAs. The raw data is preprocessed and converted into structured feature data at the corresponding FPGA side, and then sent to the processor cluster for fusion inference via a shared storage unit.

[0009] In one possible implementation, in step S100, the sensor is connected through a standardized interface, which includes a DC power supply line, a high-speed data line, a low-speed configuration line, and a hardware trigger synchronization line. The interface end is equipped with an electronic tag that stores logical addresses and interface capability descriptions. The sensor end has an electronic data sheet that records identity information, range, accuracy, calibration parameters, and default sampling rate. The system automatically identifies, registers, and configures the sensor by automatically reading the electronic tag and the electronic data sheet, and allocates corresponding field-programmable gate array resources according to the parsed bandwidth and synchronization requirements.

[0010] In one possible implementation, in step S200, a master-slave synchronization topology is constructed using a master synchronous field-programmable gate array (FPGA) as a global reference clock source and other FPGAs as slave synchronization units to continuously correct local clock deviations; the timestamp is directly generated and marked on the hardware side of the FPGA to avoid time uncertainty introduced by software scheduling delays.

[0011] In one possible implementation, in step S400, the parallel multiplier array of the field-programmable gate array is used to perform finite-length unit impulse response bandpass filtering for noise reduction on the latched data of the streaming input; the spectral amplitude and phase characteristics of the partial discharge signal are extracted by fast Fourier transform; multi-scale time-frequency energy characteristics are extracted by discrete wavelet decomposition to characterize the discharge pulse intensity and vibration anomaly degree; and envelope demodulation is performed by Hilbert transform, and pulse peak detection is performed in combination with a preset threshold to output the partial discharge pulse count, peak amplitude and discharge frequency.

[0012] In one possible implementation, in step S500, considering the differences in sampling frequencies of different modal sensors, the modal feature data are mapped to a unified time axis, and missing value compensation and outlier removal are performed on each modal feature; the mean and standard deviation of the modal features are used to normalize each modal feature to eliminate dimensional differences; based on the differences in sensitivity of partial discharge faults, overheating faults, or mechanical loosening faults to different modal features, the modal contribution of each modal feature is calculated, and learnable weight parameters are assigned based on the modal contribution, and the weighted modal features are concatenated to generate the unified fused feature vector.

[0013] In one possible implementation, in step S700, the deep learning diagnostic model is deployed on the artificial intelligence inference acceleration unit, and the comprehensive risk score is a weighted sum of the predicted probability of each category and the risk weight coefficient corresponding to each category; the system uses a real-time scheduling management unit to uniformly orchestrate the field-programmable gate array acquisition link, the processor cluster fusion link, and the artificial intelligence inference acceleration link, and dynamically enables or bypasses the corresponding data processing link according to the operating conditions, trigger events, and resource usage; the artificial intelligence inference acceleration unit, the processor cluster, and the field-programmable gate array exchange high-speed data through the unified address space of the shared storage unit.

[0014] In one possible implementation, step S600 includes the following rule logic judgment: extracting partial discharge pulse amplitude features and temperature rise rate features from the fusion features; when the partial discharge pulse amplitude feature exceeds the partial discharge threshold or the temperature rise rate feature exceeds the temperature threshold, determining that the warning conditions are met and directly outputting an alarm signal; at the same time, recording the corresponding data segment as evidence for archiving; and bypassing the reasoning process of the deep learning diagnostic model when triggering an immediate alarm.

[0015] In one possible implementation, in step S800, the health status assessment conclusion is mapped to a preset set of graded alarm levels based on the trend characteristics of the comprehensive risk score or the temperature rise rate and the partial discharge count growth rate. The set of graded alarm levels includes normal status, minor alarm, moderate alarm and severe alarm.

[0016] In one possible implementation, in step S900, the structured encapsulated data object includes fault category, probability distribution, risk index, hotspot location, and evidence set; the system uploads the structured encapsulated data object and operation log to the monitoring platform via industrial Ethernet using link encryption and authentication mechanisms; the system adopts a centralized redundant power supply and health monitoring strategy to monitor key indicators such as power supply voltage, current, and temperature in real time, and sets up a hardware and software watchdog mechanism to trigger an alarm and execute a protection reset strategy when a power supply abnormality is detected or there is no response for multiple consecutive watchdog feeding cycles.

[0017] The embodiments of this application have the following beneficial effects: By closely coordinating nine steps—multi-source data acquisition, hardware-level unified time synchronization based on a precision clock protocol, hardware bus synchronous latching under abnormal triggering, low-latency preprocessing and feature extraction, dynamic weight fusion based on modal contribution, rule-based logic judgment for rapid risk interception, and edge-side deep model inference—a complete closed-loop detection process for equipment on-site is formed. This not only overcomes the problem of false alarms and missed alarms caused by the limited information dimension of a single sensor, significantly improving the diagnostic accuracy of complex faults and the spatiotemporal consistency of data from different modalities, but also effectively avoids the communication bandwidth bottleneck and network latency caused by traditional software scheduling delays and the uploading of large amounts of raw data to the cloud. This significantly enhances the system's real-time response capability to sudden anomalies and the reliability of on-site operation. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings required in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart of a multi-source data fusion intelligent diagnostic method for switchgear, as described in an embodiment of this application. Detailed Implementation

[0020] The detailed description of the accompanying drawings is intended to illustrate the present embodiments of this application and is not intended to represent only the forms in which this application can be implemented. It should be understood that the same or equivalent functions can be accomplished by different embodiments intended to be included within the spirit and scope of this application.

[0021] See Figure 1 This application provides a method for intelligent diagnosis of switchgear based on multi-source data fusion, including the following steps: Step S100: Real-time acquisition of multi-source operating data during the operation of the switchgear. The multi-source operating data includes at least partial discharge signals, temperature information, electrical quantity data, vibration signals, and environmental parameters to form a multi-modal raw data set. Specifically, this step lays the data foundation for the entire diagnostic method. The introduction of dimensions such as partial discharge, temperature, electrical quantities, vibration, and environmental parameters comprehensively characterizes the physical manifestations of the switchgear at different stages of fault evolution. For example, in actual 10kV switchgear operation cases, a single temperature and humidity sensor can only detect overheating at joints but cannot detect initial partial discharge caused by moisture on the surface of insulation components; however, with the addition of a UHF partial discharge sensor, the system can simultaneously capture anomalies in both thermal and electrical dimensions. This multi-dimensional raw data set effectively overcomes the limitations of information dimensions in traditional single-sensor detection, avoiding false alarms or missed alarms due to missing features at the source, and significantly improving the data completeness for subsequent complex fault identification.

[0022] Step S200: Based on the precision clock protocol, the collected multi-source operation data is uniformly synchronized and timestamped to ensure that the data of different modes are under the same time reference. Specifically, due to the significant differences in sampling frequencies among various sensors, severe timing misalignment will occur in subsequent fusion analysis without strict time alignment. For example, in actual monitoring of a power distribution room in an industrial park, the sampling rate of the partial discharge sensor is as high as 100MHz, while the sampling rate of the environmental temperature and humidity sensor is only 1Hz. When a transient discharge caused by mechanical vibration occurs, without unified time synchronization, the system may mismatch environmental data from one second ago with the current discharge pulse, leading to misjudgment. By introducing a Precision Clock Protocol (PTP) for hardware-level unified time synchronization, clock drift and communication delay errors between multimodal data are fundamentally eliminated, ensuring absolute consistency of features across different dimensions on the time axis and providing a reliable time reference for subsequent accurate causal relationship analysis and feature fusion.

[0023] Step S300: When any modal data is detected to meet the preset abnormal triggering condition, a synchronous latching instruction is broadcast to all acquisition channels through the hardware trigger bus, so that each acquisition channel latches data within the same time window, forming a detection sample under a unified time window. Specifically, sudden faults in switchgear are often fleeting, and this step establishes a globally linked snapshot mechanism. For example, in a lightning overvoltage simulation test of a substation switchgear, the transient discharge pulse lasted only a few microseconds, and traditional systems often only record messy and incomplete waveforms. However, using the hardware trigger bus in this embodiment, once the UHF channel detects a spike exceeding the threshold, it broadcasts latching signals to all channels such as vibration and temperature within nanoseconds, instantly freezing 10 milliseconds of full-dimensional data before and after the event. This mechanism can accurately capture a full-dimensional state slice at the moment the anomaly occurs, avoiding the loss of crucial transient evidence and greatly improving the accuracy of transient fault tracing.

[0024] Step S400: Preprocess and extract features from the latched data. The preprocessing and feature extraction include filtering and noise reduction, frequency domain feature extraction, time-frequency energy feature extraction, and pulse statistical feature extraction to obtain key feature information reflecting the device status. Specifically, this step achieves a dimensionality reduction transformation from massive amounts of raw waveforms to high-value structured indicators. For example, in monitoring switchgear in an old residential community, there was extremely strong background white noise and carrier communication interference. Through FPGA-based FIR bandpass filtering and frequency domain feature extraction, the system can accurately extract the 50Hz power frequency harmonic anomaly and the partial discharge characteristic frequency band of several hundred megahertz submerged in noise. Combined with time-frequency energy analysis, the waveform, which lasts for several seconds, is compressed into dozens of core feature values ​​that can reflect the degree of insulation degradation. This not only significantly reduces the data size but also preserves diagnostically sensitive information to the greatest extent, effectively alleviating the processing pressure on subsequent computing units.

[0025] Step S500: Perform time axis alignment and scale normalization on the key feature information of different modalities, and dynamically allocate fusion weights according to the contribution of each modal feature to the current fault type to construct a unified fusion feature vector; Specifically, different modal features differ significantly in their dimensions and numerical distributions, making direct concatenation difficult for model training. For example, the amplitude of a partial discharge pulse may reach several thousand millivolts, while the temperature rise at the same moment may only be a few tenths of a degree Celsius. By eliminating the influence of dimensions through normalization and introducing a dynamic weight allocation mechanism based on contribution, the dominant mode's influence can be adaptively amplified according to the current suspected fault type. When facing a "poor contact" fault, the system automatically assigns a very high weight to temperature features; while when facing a "floating discharge" fault, the weight of partial discharge features is significantly increased. This intelligent fusion strategy significantly improves the ability of the fused feature vector to focus on specific faults, enhancing the model's diagnostic robustness.

[0026] Step S600: Based on the unified fusion feature vector, perform preset rule logic judgment to determine the trend and statistics of key features by threshold judgment in order to conduct rapid risk screening. Specifically, this step constructs a lightweight security defense before the computationally intensive deep learning inference. For example, in a case where a switchgear was operating at full load under extreme high temperatures, the infrared sensor reported a temperature rise rate of 5°C / minute, far exceeding the safety threshold. At this point, the system directly triggered a rule-based alarm, skipping the subsequent time-consuming neural network inference process, and immediately issued a circuit breaker command to the monitoring center. For extreme dangerous characteristics with clear physical boundaries, such as sudden temperature increases or bursts of partial discharge, microsecond-level risk interception can be achieved through simple threshold comparison. This mechanism not only buys the most valuable response time when the system faces serious threats but also provides basic security redundancy, ensuring the bottom-line security of the edge diagnostic system.

[0027] Step S700: If no immediate alarm is triggered, the unified fusion feature vector is input into the pre-trained deep learning diagnostic model, and the fault category probability distribution and comprehensive risk score are output through forward inference operation. Specifically, for cases where the extreme alarm threshold has not been reached but early or complex coupling defects may exist, this step relies on the powerful nonlinear mapping capabilities of deep learning models for refined diagnosis. For example, in one test, both temperature rise and partial discharge were in a "sub-healthy" critical state, and no single indicator exceeded the standard. However, the AI ​​model, by fusing feature vectors, keenly captured the composite pattern of "temperature rise accompanied by low-frequency vibration and partial discharge phase concentrated in a specific range," accurately determining it as "moisture discharge of the insulating rod caused by mechanical jamming." By performing forward inference directly at the edge, the system can output quantification results without relying on cloud computing power, effectively avoiding the communication bandwidth bottleneck and network latency caused by uploading massive amounts of raw data, and achieving true edge intelligent decision-making.

[0028] Step S800: Determine the fault category and risk level based on the fault category probability distribution and comprehensive risk score, and generate the corresponding health status assessment conclusion; Specifically, this step transforms abstract mathematical output into concrete engineering semantics. For example, a deep model might output a probability distribution (70% insulation degradation, 20% mechanical loosening, 10% normal), which is difficult for maintenance personnel to directly use as a basis for action. This embodiment precisely maps this to the specific conclusion and health level of "moderate alarm - suspected C-phase busbar insulation degradation," directly clarifying the current health baseline of the equipment. This provides on-site maintenance personnel with a clear and concise decision-making basis that requires no further interpretation, significantly reducing the practical implementation threshold of intelligent systems.

[0029] Step S900: The health status assessment conclusion, risk level and related data fragments are structured and packaged, and uploaded to the monitoring platform to realize the closed loop of online detection and diagnosis process; Specifically, isolated diagnostic results lack engineering traceability value. This step uses structured encapsulation to strongly bind the conclusion to the original chain of evidence. For example, the issued alarm message not only includes the fault category but also packages key waveform segments before and after the alarm is triggered, infrared thermal imaging screenshots, and corresponding precise timestamps. Uploading data using a standardized format not only ensures the integrity and security of information transmission but also allows remote experts to review and assess damage without being physically present on-site by retrieving the evidence package, truly establishing a fully automated closed loop from on-site perception and edge analysis to remote collaborative handling.

[0030] In some embodiments, in step S100, the multi-source operating data is acquired through six parallel field-programmable gate arrays (FPGAs), with the following specific functions: the first FPGA is used for partial discharge high-frequency channel acquisition and pulse detection; the second FPGA is used for ultrasonic or acoustic emission broadband channel acquisition; the third FPGA is used for current and voltage synchronous sampling and phase information extraction; the fourth FPGA is used for infrared array temperature data acquisition and frame buffering; the fifth FPGA is used for point-based temperature and environmental parameter acquisition; and the sixth FPGA serves as the main synchronization unit for global clock synchronization, triggering broadcasts, and system-level acquisition control. The raw high-frequency waveform data is not transmitted between the FPGAs. The raw data is preprocessed and converted into structured feature data at the corresponding FPGA side, and then sent to the processor cluster for fusion inference via a shared storage unit.

[0031] Specifically, due to the extremely high frequency and massive data volume of signals such as partial discharge and ultrasonic waves from switchgear, traditional single-core serial processing architectures simply cannot meet real-time requirements. By employing six FPGAs to construct a distributed parallel acquisition architecture, the sorting and preprocessing tasks of high-frequency, medium-frequency, and low-speed signals are completely offloaded to the hardware layer. The core advantage of this architecture design lies in "simultaneous acquisition and computation," with the overall system input being a multi-source state data set: (1) Where X(t) represents the set of multi-source state data at time t; x i (t) represents the data output by the i-th type of sensor at time t; m represents the total number of sensor types.

[0032] After underlying hardware preprocessing, the overall system output is the switch cabinet health status assessment and fault diagnosis results: (2) in, (t) represents the system diagnostic output at time t; f() represents the system's fusion inference function.

[0033] Inside the FPGA, data is pipelined in the form of discrete sampling sequences. Let the discrete sampling sequence of the i-th acquisition channel be: (3) in, N represents the value of the nth sample point in the i-th channel; i This indicates the sampling window length of the channel.

[0034] To implement the above parallel processing mechanism, this application adopts a parallel acquisition architecture of 6 FPGAs, forming an FPGA set: (4) in, This indicates a collection of FPGAs; F1 to F6 represent the first to sixth field-programmable gate arrays, respectively.

[0035] Furthermore, to support this multi-path parallel hardware architecture and avoid messy wiring and interface incompatibility issues in the field, the system defines a unified and standardized hardware interface at the physical layer. The interface set is defined as follows: (5) in, For the interface set, PWR is the unified DC power supply line; HS is the high-speed data line; LS is the low-speed configuration line; and TRIG is the trigger / synchronization line.

[0036] At the hardware interface level, each physical port is configured with an electronic tag for self-identification. The interface end sets the electronic tag to store the logical address and interface capability description: (6) in, The interface identifier is represented by `addr`; the physical address is represented by `addr`; and the interface capability description is represented by `cap`.

[0037] At the same time, the sensor also has a built-in standardized description file, and the sensor has a fixed TEDS (Electronic Data Sheet) that records identification information, range, accuracy, calibration parameters and default sampling rate, etc. (7) Wherein, TEDS represents the sensor electronic data sheet; UID represents the unique identifier; Type represents the sensor type; Range represents the measurement range; Acc represents the accuracy; Cal represents the calibration parameters; Fs represents the default sampling rate; and Proto represents the communication protocol.

[0038] Once the hardware connection is established, the system automatically packages the sensor's self-description information with the acquired raw data. The multi-source acquisition module outputs data in a unified encapsulation format, forming a data acquisition package. (8) Among them, D i This indicates that the data packet for the i-th channel is being acquired. This represents the sampled data for the i-th channel; is the sampling timestamp of the sampling data of the i-th channel; TEDS is the sensor electronic data sheet; Indicates the interface identifier.

[0039] In the complex electromagnetic interference application scenarios of wind farm booster stations, this distributed parallel architecture and local preprocessing strategy completely avoids bus congestion caused by the transmission of high-frequency raw waveforms between chips, realizes the local resolution of massive data, and multiplies the overall data throughput and real-time response limit of the system.

[0040] In some embodiments, in step S100, the sensor is connected through a standardized interface, which includes a DC power supply line, a high-speed data line, a low-speed configuration line, and a hardware trigger synchronization line. The interface end is equipped with an electronic tag that stores the logical address and interface capability description, and the sensor end is equipped with an electronic data sheet that records the identity information, range, accuracy, calibration parameters, and default sampling rate. The system automatically identifies, registers, and configures the sensor by automatically reading the electronic tag and the electronic data sheet, and allocates corresponding field-programmable gate array resources according to the parsed bandwidth and synchronization requirements.

[0041] Specifically, this embodiment further illustrates the significant value of the above formulas (5) to (8) in actual engineering deployment. When replacing or adding sensors, traditional industrial monitoring systems often require professionals to modify the underlying register configuration and software drivers, which is time-consuming and prone to errors. For example, in a chemical plant power distribution room renovation project, maintenance personnel needed to upgrade the original low-frequency partial discharge sensor to a high-frequency UHF sensor. Thanks to the electronic tag and TEDS mechanism of this solution, the system can automatically read the bandwidth and sampling rate parameters of the new sensor after powering on, automatically bind it to FPGA1 with high-frequency processing capabilities, and release the original low-speed channel resources after plugging and unplugging the connector once. This plug-and-play design eliminates the cumbersome manual dialing and underlying code configuration, which not only greatly reduces the difficulty and error rate of on-site construction, but also intelligently matches the most suitable FPGA hardware resources according to the parsed bandwidth attributes, realizing efficient on-demand allocation of acquisition computing power and extreme flexible expansion of the system.

[0042] In some embodiments, in step S200, a master-slave synchronization topology is constructed using a master synchronous field-programmable gate array (FPGA) as a global reference clock source and other FPGAs as slave synchronization units to continuously correct local clock deviations; the timestamp is directly generated and marked on the hardware side of the FPGA to avoid time uncertainty introduced by software scheduling delays.

[0043] Specifically, the lifeline of multimodal data fusion lies in time consistency. Even a microsecond-level clock skew can lead to completely erroneous conclusions when analyzing the correlation between high-frequency partial discharge and low-frequency temperature rise. To avoid the random jitter caused by the operating system's task scheduling in traditional software time synchronization, this embodiment constructs a pure hardware master-slave clock tree. Let the global reference clock be Tg(t), and the i-th FPGA local clock be Ti(t), with the following deviation: (9) in, This represents the deviation between the i-th FPGA local clock and the global reference clock at time t; Indicates the global reference clock; This represents the local clock of the i-th FPGA.

[0044] The system's internal clock synchronization service monitors and dynamically smooths out this deviation in real time. It estimates and corrects the deviation in real time through the PTP synchronization mechanism to obtain an aligned clock. (10) in, This represents the aligned clock of the i-th FPGA after deviation correction.

[0045] After obtaining a precise clock, the stamping action must be perfectly synchronized with the data sampling rhythm. Therefore, software stamping was completely abandoned, and hardware timestamp generation was implemented on the FPGA side. For the nth sampling point of the i-th channel, its timestamp is: (11) in, This represents the hardware timestamp of the nth sampling point in the i-th channel; This represents the sampling period of the i-th channel.

[0046] Furthermore, in order to force all channels to align in the event of transient anomalies, this application sets up a global trigger broadcast mechanism, and the trigger synchronization condition is defined as follows: (12) in, Indicates the start time of synchronous latching for each channel; Indicates the initial time when the trigger event occurred; i represents the FPGA channel index.

[0047] In a joint diagnostic case targeting surface discharge faults inside a switchgear, the deep integration of master-slave topology and a pure hardware stamping mechanism completely eliminated microsecond-level or even millisecond-level jitter caused by operating system task scheduling. The system accurately reconstructed the phase relationship between the occurrence of the partial discharge pulse and the zero-crossing point of the three-phase current, ensuring the strict determinism of multi-channel timestamps and laying a perfect timing foundation for high-precision cross-modal joint analysis.

[0048] In some embodiments, in step S400, the parallel multiplier array of the field-programmable gate array is used to perform finite-length unit impulse response bandpass filtering for noise reduction on the latched data of the streaming input; the spectral amplitude and phase characteristics of the partial discharge signal are extracted by fast Fourier transform; multi-scale time-frequency energy characteristics are extracted by discrete wavelet decomposition to characterize the discharge pulse intensity and vibration anomaly; and envelope demodulation is performed by Hilbert transform, and pulse peak detection is performed in combination with a preset threshold to output the partial discharge pulse count, peak amplitude and discharge frequency.

[0049] Specifically, this embodiment details how an FPGA utilizes the parallel characteristics of its underlying logic gate array to transform traditional serial time-consuming algorithms into pipelined real-time hardware computation. Firstly, addressing the background white noise and narrowband interference in the environment, the FIR filtering denoising mechanism outputs the following: (20) Where y(n) represents the output signal after filtering at the nth sampling point; h(i) represents the coefficient of the i-th filter; M represents the filter order; and x(ni) represents the input signal delayed by i periods.

[0050] The filtered clean signal is then fed into the FFT kernel for spectrum analysis. The Fast Fourier Transform (FFT) extracts frequency domain features, and the Discrete Fourier Transform (DFT) is: (twenty one) Where X(k) represents the frequency domain complex spectrum value with frequency index k; N represents the sampling window length; and j represents the imaginary unit.

[0051] To facilitate subsequent pattern recognition, the modulus and phase angle of the complex spectrum are extracted to construct a feature vector. The output frequency domain feature vector is defined as follows: (twenty two) Where F represents the frequency domain eigenvector; |X(k)| represents the amplitude spectrum; and ∠X(k) represents the phase spectrum.

[0052] Considering the strong non-stationary transient characteristics of partial discharge signals and mechanical vibration signals, simple frequency domain analysis would lose time information. Therefore, wavelet transform is introduced. In the discrete wavelet transform for extracting multi-scale time-frequency features, the expression for the continuous wavelet transform is: (twenty three) in, denoted by ; a is the continuous wavelet transform coefficient; b is the scaling parameter; and is the translation parameter. This is the complex conjugate mother wavelet function after scaling and translation.

[0053] Due to limited computing resources in FPGA implementation, discrete wavelet decomposition is used to extract energy features at various scales. (twenty four) Where E(a) represents the wavelet energy feature at scale a.

[0054] Based on the envelope characteristics of partial discharge pulses, an analytic signal is constructed using the Hilbert transform. In envelope demodulation and pulse detection, the envelope signal is calculated as follows: (25) Where e(n) represents the demodulated envelope signal; x(n) is the original signal; (n) is the Hilbert transform result of x(n).

[0055] After extracting the envelope, the pulse peak value is captured by setting a fixed threshold. The pulse detection rule is as follows: (26) Where p(n) represents the pulse detection result; γ is the preset pulse detection threshold.

[0056] The design of all the above hardware algorithms must meet a strict physical constraint: the processing time cannot exceed the interval between two adjacent sampling points. In FPGA real-time constraints and delay analysis, the computational delay satisfies: (27) Among them, L FPGA Ts represents the total FPGA processing delay; Ts represents the sampling period.

[0057] Finally, the FPGA transforms the complex waveforms into a set of compact numerical indicators, and the FPGA outputs a structured feature set after processing: (28) Among them, f FPGA F represents the set of structured features output by the FPGA. FFT Indicates spectral characteristics; E wavelet Represents multi-scale energy characteristics; Ppulse This indicates the statistical characteristics of the pulse.

[0058] In a practical investigation of corona discharge at a substation busbar connection, the advantages of the FPGA parallel pipeline architecture were fully utilized to complete complex convolution, Fourier and wavelet matrix operations within an extremely short sampling period. In real time, the ultra-high frequency spectrum energy characterizing the corona was extracted from strong electromagnetic interference, realizing lossless feature capture and near-zero delay conversion of high-speed transient signals, which greatly improved the real-time performance and fidelity of signal processing.

[0059] In some embodiments, in step S500, for the sampling frequency differences of different modal sensors, the modal feature data are mapped to a unified time axis, and missing value compensation and outlier removal are performed on each modal feature; the mean and standard deviation of the modal features are used to normalize each modal feature to eliminate dimensional differences; based on the sensitivity differences of partial discharge faults, overheating faults or mechanical loosening faults to different modal features, the modal contribution of each modal feature is calculated, and learnable weight parameters are assigned based on the modal contribution, and the weighted modal features are concatenated to generate the unified fused feature vector.

[0060] Specifically, due to the different sampling rates of each sensor (e.g., partial discharge reaching hundreds of megahertz, while temperature is only one hertz), the CPU's primary task after receiving the FPGA's features is to solve the "spatiotemporal alignment" problem. Let the time mapping function of the i-th mode be τ. - ¹ i (t), then the aligned signal is: (13) in, τ represents the time-aligned i-th mode signal; - ¹ i (t) represents the timestamp reverse mapping function.

[0061] After mapping all modalities to a unified time scale, a complete multi-source data frame is formed: (14) in, This represents the aligned multi-source data frame with time index k.

[0062] To eliminate the significant gap in numerical scale between different physical quantities (such as voltage amplitude and temperature), and to prevent large numerical features from overshadowing small numerical features, the synchronized data pairs feature quantities f. i Adopting standardization: (15) in, f represents the standardized eigenvalues; iRepresents the original eigenvalues; μ i σ represents the mean of this modal characteristic. i The standard deviation represents the modal characteristic.

[0063] The standardized data is concatenated into a high-dimensional vector, ultimately forming the normalized input vector: (16) Where z represents the normalized multimodal feature concatenation vector.

[0064] It is worth noting that the starting point for the data flow of the entire edge system is a multi-source sensor data set: (17) Where X(t) represents the input multi-source sensor data set, x1(t) represents the partial discharge signal, x2(t) represents the thermal imaging data, x3(t) represents the vibration signal, and x4(t) represents the environmental parameters.

[0065] After a series of underlying synchronization, preprocessing, and fusion processes, the ultimate goal is to output high-precision equipment health status estimation results: (18) in, The output represents the device health status estimation result; f() represents the fusion inference function.

[0066] Before entering the deep model, the CPU fusion computing unit receives a set of structured features uploaded by the underlying hardware: (29) Wherein, F(t) represents the set of structured features of the input; f1(t) represents the partial discharge spectrum and pulse statistics features; f2(t) represents the thermal imaging temperature distribution features; f3(t) represents the vibration time-frequency energy features; and f4(t) represents the environmental state features.

[0067] As mentioned earlier, these features also require precise time alignment, which the CPU performs using a unified clock mapping: (30) in, This represents time-aligned modal data; This represents the reverse timestamp mapping function.

[0068] After alignment, normalization and concatenation are performed again. After normalization, a unified feature concatenation vector is constructed. (32) Where z represents the normalized unified feature concatenation vector.

[0069] The core innovation of this embodiment lies in breaking away from the traditional equal-weight concatenation and introducing the idea of ​​an attention mechanism. Let the i-th modal feature be f. i The CPU calculates its importance weight: (33) Among them, f i R represents the eigenvector of the i-th modality; d Let represent the d-dimensional real space.

[0070] The dynamic weighting of a sample is calculated using the Softmax function: (34) Where, α i w represents the contribution weight of the i-th modal feature; i T represents the learnable weight parameter vector; T represents the vector transpose; m represents the total number of modes.

[0071] The calculated weights are used to perform a weighted summation of the original features, resulting in the following fused features: (35) Among them, z This represents the final unified fusion feature vector.

[0072] In handling a complex fault case involving "insulation aging accompanied by mechanical loosening," this series of refined alignments and attention-based feature reorganizations not only eliminated the physical gap between heterogeneous data but also enabled the fused vector to adaptively focus on key fault representations. The model automatically assigned high weights to vibration features, accurately distinguishing between simple insulation dampness and complex defects with mechanical risks, significantly improving the model's diagnostic accuracy when facing complex faults.

[0073] In some embodiments, in step S700, the deep learning diagnostic model is deployed on the artificial intelligence inference acceleration unit, and the comprehensive risk score is the weighted sum of the predicted probability of each category and the risk weight coefficient corresponding to each category; the system uniformly orchestrates the field-programmable gate array acquisition link, the processor cluster fusion link, and the artificial intelligence inference acceleration link through the real-time scheduling management unit, and dynamically enables or bypasses the corresponding data processing link according to the operating conditions, trigger events, and resource usage; the artificial intelligence inference acceleration unit, the processor cluster, and the field-programmable gate array exchange high-speed data through the unified address space of the shared storage unit.

[0074] Specifically, while deep learning models offer high accuracy, their network structures are complex, and the basic form of their inference models is as follows: (38) in, This represents the model's prediction result; f θ This represents the inference structure of a neural network with parameter set θ; z This represents the input fused feature vector.

[0075] In classification tasks, the last layer of a network typically outputs a logical value, which needs to be converted into a probability distribution. The probability output (Softmax function) for classification tasks is as follows: (39) Where P(y=k) represents the predicted probability that the sample belongs to class k; This represents the score corresponding to category k of the input feature; This represents the score for category j.

[0076] Because different fault types pose vastly different risks to the power grid (e.g., short-circuit risk is much greater than minor overheating), the system introduces a risk weighting coefficient, and the risk scoring mechanism is as follows: (40) Where R represents the overall risk score result; This represents the risk weight coefficient corresponding to category k.

[0077] To support the aforementioned AI inference and the fusion of FPGA preprocessing with the CPU, the system constructs a high-speed data channel at the hardware bus level and defines a shared memory buffer: (41) Where B represents the set of shared storage buffers; B FPGA Indicates the FPGA-side buffer; B CPU Indicates the CPU-side cache; B AI This represents the cache area for the AI ​​inference unit.

[0078] Data flow between different computing units is achieved through memory mapping, and the data writing strategy is as follows: (42) Among them, B CPU (t) represents the data state of the CPU cache at time t; Δt represents the data synchronization delay; Φ(·) represents the data mapping or preprocessing function.

[0079] At the system software level, the real-time scheduling and management unit acts like an intelligent traffic controller, abstracting all computational tasks into a set of defined tasks: (43) Where T represents the set of tasks to be processed by the system; T iLet represent the i-th task; n represents the total number of tasks.

[0080] The underlying hardware is abstracted as a collection of resources: (44) Where R represents the system's computing resource set; R FPGA Represents FPGA resources; R CPU Represents CPU resources; R AI This refers to AI inference resources.

[0081] The scheduler determines the flow of tasks using a state matrix and defines allocation variables: (45) in, This indicates a task assignment indicator variable; if the task... Allocate resources If it is 1, then it is 1; otherwise, it is 0.

[0082] The ultimate goal of scheduling is to minimize system power consumption while ensuring real-time performance. Optimization objective: (46) in, Indicates task In resources Execution delay; Indicates task In resources Energy consumption; α and β represent weighting coefficients.

[0083] Of course, this scheduling must satisfy basic physical rules, constraint 1: (47) The formula stipulates that each task must be allocated to one and only one computing resource.

[0084] Constraint 2: (48) Among them, C i Indicates task The computational load requirements; Representing resources The maximum computing power limit.

[0085] In real-world operation during peak summer load periods, this heterogeneous collaboration and dynamic orchestration mechanism played a crucial role. When the system detects consecutive minor anomalies, the scheduler dynamically increases the priority and computing power quota of the AI ​​inference link; while when the devices are functioning normally, it actively bypasses the AI ​​units to reduce power consumption within the chassis. This mechanism breaks the rigidity of traditional hardware and software being isolated from each other, achieving global optimization of system-level latency and energy consumption while ensuring deep learning inference computing power, thus ensuring the long-term stable and efficient operation of edge devices under harsh conditions.

[0086] In some embodiments, in step S600, the rule logic judgment includes: extracting the partial discharge pulse amplitude feature and the temperature rise rate feature from the fusion features; when the partial discharge pulse amplitude feature exceeds the partial discharge threshold, or the temperature rise rate feature exceeds the temperature threshold, determining that the warning condition is met and directly outputting an alarm signal; at the same time, recording the corresponding data segment as evidence for archiving; and bypassing the reasoning process of the deep learning diagnostic model when triggering an immediate alarm.

[0087] Specifically, in the safety guidelines of power systems, exceeding the limits of certain physical quantities signifies an impending disaster. If, at this point, we still have to wait for the deep learning model to complete dozens of layers of its network before issuing instructions, irreparable damage will occur. Therefore, the system incorporates a "short-circuit" logic during the CPU fusion phase, and the alarm rule can be expressed as: (36) Where Alarm(t) represents the alarm flag at time t; p(t) represents the amplitude characteristics of the partial discharge pulse; γ p Indicates the partial discharge pulse threshold; T(t) represents the temperature rise characteristic; γ T This indicates the temperature threshold.

[0088] Once this rule is triggered, the CPU will directly output the current result and terminate the subsequent process. The CPU fusion computing unit will output the fusion characteristics and preliminary health assessment. (37) Where Output(t) represents the set of outputs of the CPU unit; z (t) represents the fused feature vector; Alarm(t) represents the alarm flag.

[0089] In a case where an arcing accident was detected due to oxidation of the disconnector contacts, this "rule-based interception + bypass reasoning" design played a decisive role. Faced with a sudden temperature rise, the system, upon encountering a clear high-risk characteristic, eliminated the time-consuming neural network forward propagation process, achieving an extremely fast safety response with minimal computational cost, triggering the protection trip several seconds in advance, and establishing an impenetrable safety baseline for the system.

[0090] In some embodiments, in step S800, the health status assessment conclusion is mapped to a preset set of graded alarm levels based on the trend characteristics of the comprehensive risk score or the temperature rise rate and the partial discharge count growth rate. The set of graded alarm levels includes normal status, minor alarm, moderate alarm and severe alarm.

[0091] Specifically, a single numerical score is not conducive to maintenance personnel quickly determining the priority of handling. The system refers to the traditional grading standards of the power industry and sets up a set of graded alarm levels: (56) Where L represents the set of alarm levels; L0 represents the normal state; L1 represents a minor alarm; L2 represents a moderate alarm; and L3 represents a severe alarm.

[0092] In terms of mapping logic, the system does not rigidly rely on probability scores, but instead introduces a rate of change indicator over time. The alarm level mapping function is as follows: (57) Where Level(t) represents the alarm level output at time t; R(t) represents the risk assessment index; (t) represents the diagnostic conclusion; Trend(t) represents the trend characteristics such as the rate of temperature rise and the rate of increase in partial discharge count; h() represents a multi-parameter mapping function.

[0093] In a progressive fault tracking case involving moisture in cable terminal joints, the introduction of trend features into the level mapping enabled the system to keenly capture the degradation trajectory of the equipment. Although the absolute temperature had not yet reached the moderate alarm threshold, the system accurately predicted the approaching insulation breakdown critical point by calculating the temperature rise rate and partial discharge frequency growth rate over three consecutive days, triggering a high-level warning in advance. This represented a significant leap from "post-event alarm" to "pre-event warning."

[0094] In some embodiments, in step S900, the structured encapsulated data object includes fault category, probability distribution, risk index, hotspot location, and evidence set; the system uploads the structured encapsulated data object and operation log to the monitoring platform via industrial Ethernet using link encryption and authentication mechanisms; the system adopts a centralized redundant power supply and health monitoring strategy to monitor key indicators of power supply voltage, current, and temperature in real time, and sets up a hardware and software watchdog mechanism to trigger an alarm and execute a protection reset strategy when a power supply abnormality is detected or there is no response for multiple consecutive watchdog feeding cycles.

[0095] Specifically, before data is sent out, the system's internal software bus needs to be aware of the operating status of each link. The system uniformly orchestrates the FPGA acquisition links and defines a set of data flow states. (51) Where S represents the set of data stream states; S FPGA S CPU S AI These represent the status of the corresponding links; =1 indicates that the link is enabled. =0 indicates that the link is bypassed or disabled.

[0096] When the data stream successfully arrives at the network interface, the system will extract the most crucial diagnostic conclusions and send them up. The data to be sent is: (52) Where U(t) represents the data object sent to the platform; Alarm(t) is the alarm flag; Meta represents metadata such as device number and model version.

[0097] To prevent edge devices from malfunctioning, the system monitors its own physiological indicators and defines power supply health indicators: (53) in, V(t) represents the power supply health index at time t; V(t) represents the power supply voltage; I(t) represents the power supply current; Temp(t) represents the temperature of key components; g() represents the health assessment function.

[0098] In extreme cases (such as when strong electromagnetic interference causes the program to crash), the system relies on underlying hardware logic to save itself by setting up a hardware and software watchdog timer. If the watchdog timer is not fed for K consecutive cycles, a reset is triggered. (54) Where Reset represents the reset trigger signal; T represents the dog-feeding state in the k-th cycle; w This indicates the dog feeding cycle; K represents the maximum allowed number of unfed dog feeding cycles.

[0099] The final complete data packet sent to the external network is defined in a standard interface format, and the system outputs a structured output object: (55) Where O(t) represents the structured output object; P(t) represents the fault category probability distribution; Indicates the hot spot location or fault location; Evd(t) represents the evidence set.

[0100] Form an external output interface: (58) API(t) represents the interface data output to the outside world; Log(t) represents the running log, event record and device status information.

[0101] In an extreme scenario where a thunderstorm caused a brief network outage at a substation, the combination of structured encapsulation and a watchdog mechanism demonstrated exceptional survivability. During the outage, edge devices maintained stable operation thanks to local redundant power supply and the watchdog mechanism. Upon network restoration, they immediately uploaded the accumulated structured evidence packets in batches. This not only ensured the confidentiality and tamper resistance of diagnostic results during network transmission but also endowed edge hardware with self-healing capabilities under extreme conditions or software crashes, significantly extending the mean time between failures (MTBF) in unattended environments.

[0102] The various embodiments of this application have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used in these embodiments is chosen to best explain the principles, practical applications, or technological improvements to the embodiments in the market, or to enable other those skilled in the art to understand the embodiments disclosed in this application.

Claims

1. A method for intelligent diagnosis of switchgear using multi-source data fusion, characterized in that, Includes the following steps: Step S100: Real-time acquisition of multi-source operating data during the operation of the switchgear. The multi-source operating data includes at least partial discharge signals, temperature information, electrical quantity data, vibration signals, and environmental parameters to form a multi-modal raw data set. Step S200: Based on the precision clock protocol, the collected multi-source operation data is uniformly synchronized and timestamped to ensure that the data of different modes are under the same time reference. Step S300: When any modal data is detected to meet the preset abnormal triggering condition, a synchronous latching instruction is broadcast to all acquisition channels through the hardware trigger bus, so that each acquisition channel latches data within the same time window, forming a detection sample under a unified time window. Step S400: Preprocess and extract features from the latched data. The preprocessing and feature extraction include filtering and noise reduction, frequency domain feature extraction, time-frequency energy feature extraction, and pulse statistical feature extraction to obtain key feature information reflecting the device status. Step S500: Perform time axis alignment and scale normalization on the key feature information of different modalities, and dynamically allocate fusion weights according to the contribution of each modal feature to the current fault type to construct a unified fusion feature vector; Step S600: Based on the unified fusion feature vector, perform preset rule logic judgment to determine the trend and statistics of key features by threshold judgment in order to conduct rapid risk screening. Step S700: If no immediate alarm is triggered, the unified fusion feature vector is input into the pre-trained deep learning diagnostic model, and the fault category probability distribution and comprehensive risk score are output through forward inference operation. Step S800: Determine the fault category and risk level based on the fault category probability distribution and comprehensive risk score, and generate the corresponding health status assessment conclusion; Step S900: The health status assessment conclusion, risk level and related data fragments are structured and packaged, and uploaded to the monitoring platform to realize the closed loop of online detection and diagnosis process.

2. The method according to claim 1, characterized in that, In step S100, the multi-source operating data is acquired through six parallel field-programmable gate arrays (FPGAs). Specifically, the first FPGA is used for partial discharge high-frequency channel acquisition and pulse detection; the second FPGA is used for ultrasonic or acoustic emission broadband channel acquisition; the third FPGA is used for current and voltage synchronous sampling and phase information extraction; the fourth FPGA is used for infrared array temperature data acquisition and frame buffering; the fifth FPGA is used for point-based temperature and environmental parameter acquisition; and the sixth FPGA serves as the main synchronization unit for global clock synchronization, triggering broadcasts, and system-level acquisition control. The raw high-frequency waveform data is not transmitted between the FPGAs. The raw data is preprocessed and converted into structured feature data at the corresponding FPGA side, and then sent to the processor cluster for fusion inference via a shared storage unit.

3. The method according to claim 2, characterized in that, In step S100, the sensor is connected through a standardized interface, which includes a DC power supply line, a high-speed data line, a low-speed configuration line, and a hardware trigger synchronization line. The interface end is equipped with an electronic tag that stores the logical address and interface capability description. The sensor end has an electronic data sheet that records the identity information, range, accuracy, calibration parameters, and default sampling rate. The system automatically identifies, registers, and configures the sensor by automatically reading the electronic tag and the electronic data sheet, and allocates corresponding field-programmable gate array resources according to the parsed bandwidth and synchronization requirements.

4. The method according to claim 1, characterized in that, In step S200, a master-slave synchronization topology is constructed using a master synchronous field-programmable gate array (FPGA) as the global reference clock source and other FPGAs as slave synchronization units to continuously correct local clock deviations. The timestamp is directly generated and marked on the hardware side of the FPGA to avoid time uncertainty introduced by software scheduling delays.

5. The method according to claim 1, characterized in that, In step S400, the parallel multiplier array of the field-programmable gate array is used to perform finite-length unit impulse response bandpass filtering to denoise the latched data of the streaming input; the spectral amplitude and phase characteristics of the partial discharge signal are extracted by fast Fourier transform. Multi-scale time-frequency energy features are extracted by discrete wavelet decomposition to characterize the discharge pulse intensity and vibration anomaly degree; envelope demodulation is performed by Hilbert transform, and pulse peak detection is performed by combining a preset threshold, outputting partial discharge pulse count, peak amplitude and discharge frequency.

6. The method according to claim 1, characterized in that, In step S500, considering the differences in sampling frequencies of different modal sensors, the modal feature data are mapped to a unified time axis, and missing value compensation and outlier removal are performed on each modal feature. The mean and standard deviation of the modal features are used to normalize each modal feature to eliminate dimensional differences. Based on the differences in sensitivity of partial discharge faults, overheating faults, or mechanical loosening faults to different modal features, the modal contribution of each modal feature is calculated, and learnable weight parameters are assigned based on the modal contribution. The weighted modal features are then concatenated to generate the unified fused feature vector.

7. The method according to claim 1, characterized in that, In step S700, the deep learning diagnostic model is deployed on the artificial intelligence inference acceleration unit, and the comprehensive risk score is the weighted sum of the predicted probability of each category and the risk weight coefficient corresponding to each category. The system uses a real-time scheduling management unit to uniformly orchestrate the field-programmable gate array acquisition link, the processor cluster fusion link, and the artificial intelligence inference acceleration link, and dynamically enables or bypasses the corresponding data processing link according to the operating conditions, trigger events, and resource usage. The artificial intelligence inference acceleration unit, the processor cluster, and the field-programmable gate array exchange high-speed data through the unified address space of the shared storage unit.

8. The method according to claim 1, characterized in that, In step S600, the rule logic judgment includes: extracting the partial discharge pulse amplitude feature and the temperature rise rate feature from the fusion features; when the partial discharge pulse amplitude feature exceeds the partial discharge threshold or the temperature rise rate feature exceeds the temperature threshold, it is determined that the warning condition is met and an alarm signal is directly output; at the same time, the corresponding data segment is recorded as evidence for archiving; and when an immediate alarm is triggered, the reasoning process of the deep learning diagnostic model is bypassed.

9. The method according to any one of claims 1 to 8, characterized in that, In step S800, the health status assessment conclusion is mapped to a preset set of graded alarm levels based on the trend characteristics of the comprehensive risk score or the temperature rise rate and the partial discharge count growth rate. The set of graded alarm levels includes normal status, minor alarm, moderate alarm and severe alarm.

10. The method according to any one of claims 1 to 8, characterized in that, In step S900, the structured encapsulated data object includes fault category, probability distribution, risk index, hotspot location, and evidence set; the system uploads the structured encapsulated data object and operation log to the monitoring platform via industrial Ethernet using link encryption and authentication mechanisms; the system adopts a centralized redundant power supply and health monitoring strategy to monitor key indicators of power supply voltage, current, and temperature in real time, and sets up a hardware and software watchdog mechanism to trigger an alarm and execute a protection reset strategy when an abnormal power supply is detected or when there is no response for multiple consecutive watchdog feeding cycles.

Images