Partial discharge ultrahigh frequency storage and calculation integrated device and method

Abstract

The application discloses a partial discharge ultrahigh frequency storage and calculation integrated device and method, and belongs to the technical field of power system monitoring and relay protection. The device comprises a high-speed acquisition board, a main control board and a calculation power board. The method comprises the following steps: extracting characteristic value messages of ultrahigh frequency partial discharge pulses through a lossy compression channel, outputting complete waveform data packets of the ultrahigh frequency partial discharge pulses through a lossless data channel to form an online sample data set, calculating physical trend characteristics of the online sample data set in a first period, executing multi-modal large model inference on the accumulated characteristic value messages and the waveform data packets in a second period, and obtaining an inference result. The first period is shorter than the second period. A preliminary monitoring result is output based on the physical trend characteristics, the preliminary monitoring result and the inference result are subjected to confidence weighted fusion, and a final monitoring result is output. The application reduces transmission delay and high-speed data loss risk under high sampling rate.

Description

Technical Field

[0001] This invention belongs to the field of power system monitoring and relay protection technology, and more specifically, relates to a partial discharge ultra-high frequency storage and calculation integrated device and method. Background Technology

[0002] Converter transformers simultaneously endure the combined effects of AC voltage, DC bias, and high-frequency harmonics during operation, resulting in fundamental differences in their partial discharge characteristics compared to conventional transformers. This uniqueness manifests in three main aspects: First, the DC electric field induces space charge accumulation, significantly altering the discharge initiation characteristics; second, the electro-thermal-mechanical multi-physics coupling effect induces novel discharge modes; and third, MHz-level switching noise on the valve side severely interferes with signal acquisition. These complex operating conditions present three major technical challenges for partial discharge monitoring: the need for wide-frequency signal acquisition capabilities covering the kHz-GHz range, the acquisition of ultra-high-frequency signal extraction techniques under strong electromagnetic noise environments, and analytical methods capable of resolving time-varying nonlinear discharge characteristics.

[0003] Existing technical document 1 (CN120336957A) discloses an intelligent recognition system and method based on time-frequency feature fusion and adaptive classification. Its shortcomings include relatively simple functional implementation, extremely complex model building and full-dimensional feature algorithms, theoretically inability to achieve real-time extraction of fault information, long fault discrimination latency, and applicability only to status monitoring applications, unable to support millisecond-level actions.

[0004] Prior art document 2 (CN119622461A) discloses a method for partial discharge pattern recognition of GIS equipment based on deep learning. Its shortcomings are that the sampling frequency is low, only 50Msps. For GHz bandwidth signals, according to Shannon's theorem, this method can only realize the perception of significant features of waveform information through detection and cannot collect frequency domain information exceeding 25MHz. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an integrated storage and computing device and method for ultra-high frequency partial discharge (UHF) monitoring. This invention utilizes a three-in-one architecture comprising a high-speed acquisition board, a main control board, and a computing power board. The high-speed acquisition board performs real-time pulse triggering on the sampled data stream of UHF partial discharge signals and separates it into two types of data: feature value messages and waveform data packets. The main control board's onboard storage continuously accumulates waveform data to form an online sample dataset. Simultaneously, the onboard processor reads the online sample dataset in real-time while writing new waveform data, calculates physical trend characteristics, and forwards the data to the computing power board. The computing power board performs multimodal large-model inference and then returns the results. The main control board fuses the AI ​​inference results with the physical trend characteristics to output the monitoring results. This achieves a closed-loop integrated storage and computing system, integrating acquisition, storage, computation, and fusion within the same device. This improves the real-time performance and fault identification accuracy of UHF partial discharge online monitoring and reduces reliance on external computing resources, network bandwidth, and background scheduling.

[0006] The present invention adopts the following technical solution.

[0007] The first aspect of the present invention provides a partial discharge ultra-high frequency storage and computing integrated device, including a high-speed acquisition board, a main control board, and a computing power board; The high-speed acquisition board is used to perform real-time pulse triggering on the sampled data stream of UHF partial discharge signals. It separates the triggered UHF partial discharge pulses into characteristic value messages and waveform data packets, which are then output to the main control board. The high-speed acquisition board has built-in lossless and lossy compression paths. The lossless data path outputs the complete waveform data packets of the triggered UHF partial discharge pulses to the main control board. The lossy compression path extracts characteristic value messages from the triggered UHF partial discharge pulses and outputs them to the main control board. The real-time priority of the characteristic value messages is higher than that of the waveform data packets. The main control board includes onboard storage and an onboard processor, which are connected to the onboard storage. The onboard storage is used to store the received waveform data packets sequentially according to the acquisition time, accumulating them to form an online sample dataset. The onboard processor continuously calculates the physical trend characteristics of the online sample dataset in a first cycle, and forwards the accumulated feature value messages and waveform data packets to the computing power board in a second cycle. The first cycle is shorter than the second cycle. Before the computing power board sends back the inference results, the onboard processor independently outputs preliminary monitoring results based on the physical trend characteristics. After receiving the inference results, the onboard processor performs confidence-weighted fusion of the preliminary monitoring results and the inference results, using the rate of change of the physical trend characteristics as the dynamic adjustment coefficient of the confidence of the inference results, and outputs the final monitoring results. The computing power board is used to receive feature value messages and waveform data packets forwarded by the main control board, perform multimodal large model inference, and send the inference results back to the main control board.

[0008] Preferably, the high-speed acquisition board includes: a signal conditioning module, an analog-to-digital converter module, and an FPGA module; the signal conditioning module is connected to the analog-to-digital converter module, and the FPGA module is connected to the analog-to-digital converter module. The signal conditioning module is used to receive and condition the ultra-high frequency partial discharge signal, and the resulting conditioned analog signal is output to the analog-to-digital converter module. The analog-to-digital conversion module is used to perform analog-to-digital conversion on the conditioned analog signal, and the resulting UHF partial discharge signal sampling data stream is output to the FPGA module; The FPGA module is used to identify valid discharge pulses in the sampling data stream of UHF partial discharge signals, extract the amplitude, phase, and time stamp of the valid discharge pulses, and first cache the complete waveform sampling data of the valid discharge pulses in the RAM within the FPGA module, and then import them into the onboard DDR temporary storage connected to the FPGA module. The FPGA has lossy compression path and lossless compression path. The lossy compression path encapsulates the amplitude, phase, and time stamp of the valid discharge pulses into feature value messages according to a preset message format and outputs them to the main control board. The lossless compression path reads the complete waveform sampling data temporarily stored in the onboard DDR, encapsulates it into waveform data packets according to a preset data packet format, and outputs them to the main control board. The output priority of the feature value messages is higher than that of the waveform data packets.

[0009] Preferably, the onboard processor continuously calculates the physical trend characteristics of the online sample dataset in a first cycle, including: The onboard processor is connected to the onboard storage and is used to read the online sample dataset in the onboard storage at intervals of the first cycle to calculate spectral features. The onboard processor connects to the high-speed acquisition board and is used to extract the time stamps of each channel from the feature value messages output by the high-speed acquisition board. Based on the time stamps of each channel, it calculates multi-channel synchronization features. Using the first period as an interval, it statistically analyzes the discharge repetition rate, average amplitude, and phase distribution entropy of the amplitude, phase, and time stamp in the online sample dataset to obtain multi-timescale trend features. Using the first period as an interval, it constructs a three-dimensional point cloud spectrum based on the amplitude, phase, and time stamp in the online sample dataset. From the three-dimensional point cloud spectrum, it extracts the point cloud density, cluster number, and average curvature to obtain point cloud spectrum features. The onboard processor uses the spectrum features, multi-channel synchronization features, multi-timescale trend features, and point cloud spectrum features as physical trend features. When any two of the physical trend features meet the anomaly judgment conditions, it directly outputs the corresponding preliminary monitoring results without waiting for the inference results from the computing board. The physical trend features are also retained and used to fuse with the inference results returned by the computing board to generate the final monitoring results. The first cycle is shorter than the second cycle in which the main control board forwards data to the computing power board. Within the duration of each second cycle, the physical trend characteristics are updated multiple times, and preliminary monitoring results are continuously output without waiting for the computing power board to return inference results, thus maintaining a low-latency early warning response.

[0010] Preferably, the inference results returned by the computing board include defect probability distribution, fault mechanism explanation information, and defect characterization vector; The main control board also integrates an AI management module. The AI ​​management module connects to the onboard processor and the computing power board. It receives the preliminary monitoring results output by the onboard processor and the inference results returned by the computing power board. It dynamically allocates the confidence weights of the defect probability distribution corresponding to the preliminary monitoring results and the defect probability distribution in the inference results using the rate of change of physical trend characteristics as a dynamic adjustment coefficient. After the defect probability distribution corresponding to the preliminary monitoring results and the defect probability distribution in the inference results are fused with confidence weights, the fused result is concatenated with physical trend characteristics, fault mechanism explanation information, and defect characterization vector to form an initial fusion vector. After transformation by the fully connected layer built into the AI ​​management module, a multi-source feature fusion criterion is output. Multi-level decisions are performed on the multi-source feature fusion criterion. Based on the decision results, the partial discharge type and fault mechanism conclusion are determined and output as the monitoring result.

[0011] Preferably, the AI ​​management module also includes a first-level judgment unit, a second-level judgment unit, a third-level judgment unit, and an arbitration unit; The first-level decision unit is connected to the fully connected layer and is used to receive the multi-source feature fusion criteria output by the fully connected layer. It calculates the voting scores of various defects through softmax. When the highest score is greater than the set threshold, the defect category with the highest score is taken as the first candidate category. When the voting score of the first candidate category is greater than the set threshold, the partial discharge type and fault mechanism conclusion corresponding to the first candidate category are directly output. When the voting score of the first candidate category is not greater than the set threshold, the first-level decision unit sends the multi-source feature fusion criterion to the second-level decision unit. The second-level decision unit calculates the posterior probability based on the pre-statistical prior probability and the conditional probability of the multivariate Gaussian distribution, and takes the category with the largest posterior probability as the second candidate category. When the second candidate category and the first candidate category are the same category, the partial discharge type and fault mechanism conclusion corresponding to the same category are output. When the second candidate category is different from the first candidate category, the second-level decision unit sends the multi-source feature fusion criterion to the third-level decision unit. The third-level decision unit extracts the waveform, spectrum and three-dimensional point cloud spectrum of the pulse to be judged from the multi-source feature fusion criterion, matches them with the preset templates, and takes the category with the highest confidence weighted vote as the third candidate category. When the weighted vote of the third candidate category is greater than the set threshold, the partial discharge type and fault mechanism conclusion corresponding to the third candidate category are output. When the weighted votes of the third candidate category are not greater than the set threshold, the third-level decision unit sends the multi-source feature fusion criterion to the arbitration unit. The arbitration unit marks the pulse as a sample to be manually reviewed, outputs an alarm prompt, and saves the multi-source feature fusion criterion for manual review to determine the partial discharge type and fault mechanism conclusion.

[0012] Preferably, the third-level decision unit includes a time-domain matching subunit, a spectrum matching subunit, a point cloud matching subunit, and a weighted voting unit; The time-domain matching subunit is connected to the second-level decision unit and is used to extract the waveform of the pulse to be judged from the multi-source feature fusion criterion sent by the second-level decision unit, calculate the normalized cross-correlation function between the waveform and the average waveform template of various defects, take the defect category corresponding to the maximum value of the normalized cross-correlation as the time-domain voting category, take the maximum value as the time-domain confidence, and output it to the weighted voting unit. The spectrum matching subunit is connected to the second-level decision unit and is used to extract the spectrum of the pulse to be judged from the multi-source feature fusion criterion. It extracts the peak frequency, center frequency and bandwidth triplet from the spectrum, calculates the Euclidean distance between the triplet and the triplet of various defect templates, takes the defect category corresponding to the minimum distance as the spectrum voting category, and takes the reciprocal of the minimum distance as the spectrum confidence and outputs it to the weighted voting unit. The point cloud matching subunit is connected to the second-level decision unit. It is used to extract the three-dimensional point cloud spectrum of the pulse to be discriminated from the multi-source feature fusion criteria, perform persistent homology analysis on the three-dimensional point cloud to generate a Betti number persistent spectrum, calculate the cosine similarity between the Betti number persistent spectrum and various defect spectrum template libraries, take the defect category corresponding to the maximum cosine similarity as the point cloud voting category, take the maximum value as the point cloud confidence, and output the point cloud voting category and point cloud confidence to the weighted voting unit. The weighted voting unit is connected to the time-domain matching subunit, the spectrum matching subunit, and the point cloud matching subunit. It is used to receive the voting categories and confidence levels output by the three subunits, accumulate the confidence levels of the same defect category in the three subunits, obtain the weighted votes for each defect category, and take the defect category with the highest weighted votes as the third candidate category and output the third candidate category.

[0013] Preferably, the computing power board is used to deploy a multimodal large model; the multimodal large model includes a 1D-CNN module, a 2D-CNN module, a Point-Transformer module, and a cross-modal fusion module; The input terminals of the 1D-CNN module, 2D-CNN module, and Point-Transformer module are all connected to the main control board, and are used to jointly receive the feature value messages and waveform data packets forwarded by the main control board; The 1D-CNN module is used to extract the original time series consisting of amplitude, phase, and time stamp from feature value messages and waveform data packets, and outputs a time-series encoded vector through convolutional encoding. The 2D-CNN module is used to construct PRPD spectrograms, PRPS spectrograms, and time-frequency spectrograms from feature value messages and waveform data packets, and output spectrogram encoding vectors through convolutional coding; The Point-Transformer module is used to construct a 3D point cloud spectral map from feature value messages and waveform data packets, and output the point cloud encoded vector through self-attention encoding. The cross-modal fusion module is connected to the 1D-CNN module, 2D-CNN module and Point-Transformer module respectively. It is used to receive temporal coding vectors, graph coding vectors and point cloud coding vectors, obtain the defect characterization vector by cross-modal attention weighted summation, map it into the probability distribution of various defects by the fully connected classification head, and send the defect probability distribution, fault mechanism explanation information and defect characterization vector as inference results back to the main control board.

[0014] A second aspect provides a partial discharge ultra-high frequency storage and computing integrated method, operating within the partial discharge ultra-high frequency storage and computing integrated device described in the first aspect, comprising: The sampling data stream of the UHF partial discharge signal is pulse-triggered in real time, and the triggered UHF partial discharge pulse is separated into feature value messages and waveform data packets. The feature value messages of the UHF partial discharge pulse are extracted through a lossy compression path, and the complete waveform data packets of the UHF partial discharge pulse are output through a lossless data path. The real-time priority of the feature value messages is higher than that of the waveform data packets. The received waveform data packets are stored sequentially according to the acquisition time, accumulating to form an online sample dataset; the physical trend characteristics of the online sample dataset are continuously calculated in the first cycle, and the accumulated feature value messages and waveform data packets are forwarded in the second cycle to perform multimodal large model inference to obtain the inference results; the first cycle is shorter than the second cycle; Before the inference results are sent back, preliminary monitoring results are output based on physical trend characteristics. After the inference results are received, the preliminary monitoring results and the inference results are weighted and fused with confidence. The rate of change of physical trend characteristics is used as the dynamic adjustment coefficient of the confidence of the inference results, and the final monitoring results are output.

[0015] A third aspect of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded onto the processor, it implements the partial discharge ultra-high frequency memory-computer integrated method described in the first aspect.

[0016] The fourth aspect of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the partial discharge ultra-high frequency storage and computing integrated method described in the first aspect.

[0017] Compared with the prior art, the beneficial effects of the present invention include at least the following: Based on the high-speed acquisition board's built-in lossless data path, lossy compression path, and higher real-time priority of feature value messages, the real-time control feature data and offline analysis waveform data are separated at the hardware physical level at the acquisition front end, rather than through temporary scheduling at the software level. By using a physical diversion mechanism that prioritizes high-priority features to occupy bus transmission resources, the technical challenge of the trade-off between high sampling rate and low latency under a single bus architecture is solved. This improves the real-time transmission performance of high-priority features and the efficiency of bus resource utilization, while reducing transmission latency and the risk of high-speed data loss at high sampling rates.

[0018] Based on the technical approach of the onboard processor calculating physical trend characteristics in a first short cycle, forwarding data to the computing power board in a second long cycle, and independently outputting preliminary monitoring results before the computing power board returns the inference results, since physical trend calculation and AI inference are two independent links that are asynchronous and parallel, the physical monitoring link does not depend on the operating status of the computing power board at all. Through the asynchronous inference mechanism of independent output of physical trends in a fast cycle and enhanced verification in AI inference in a slow cycle, the technical problem of traditional simple three-board superposition schemes being heavily dependent on computing power boards and having monitoring interrupted due to computing power failure or inference timeout is solved. This improves the architectural reliability and early warning response speed of the device, and reduces the risk of monitoring interruption caused by computing power board failure and early warning lag caused by inference delay.

[0019] The technical approach of using the rate of change of physical trend characteristics as a dynamic adjustment coefficient for the confidence of inference results and performing confidence-weighted fusion of preliminary monitoring results and inference results solves the technical problems of fixed-weight fusion being unable to adapt to different discharge development rates and difficult to balance real-time performance and accuracy. This is because the fusion weight can be dynamically adjusted according to the actual rate of change of the discharge state, rather than a fixed preset weight. The dynamic fusion mechanism of adaptively allocating confidence weights on the physical side and the AI ​​side through the rate of change solves the technical problems of fixed-weight fusion being unable to adapt to different discharge development rates and difficult to balance real-time performance and accuracy. This improves the scene adaptability of the fusion criteria and the accuracy of the diagnostic results, and reduces the probability of scene deviation and misjudgment of boundary discharge states caused by fixed weights. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall architecture of a partial discharge ultra-high frequency in-memory computing integrated device provided according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the functional structure of the high-speed acquisition board provided according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the process of the partial discharge ultra-high frequency in-memory computing integrated method provided in accordance with the embodiments of the present invention; Figure 4 This is a schematic diagram of the circuit principle of the signal conditioning module provided in accordance with an embodiment of the present invention; Figure 5 This is a schematic diagram of the data flow of an FPGA module provided according to an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. The described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the spirit of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of this invention.

[0022] Combination Figure 1 As shown, this embodiment provides a partial discharge ultra-high frequency in-memory computing integrated device, which adopts a VPX bus architecture and consists of a high-speed acquisition board, a main control board, a computing power board, a power module, and an integrated chassis. The main control board is connected to both the high-speed acquisition board and the computing power board. The overall device adopts an in-memory computing collaborative architecture, and solves the technical problems of the incompatibility between high sampling rate and low latency, and the high dependence on computing power and poor reliability through two core designs: dual-path physical power splitting at the acquisition end and asynchronous dual-cycle inference on the main control side.

[0023] The output of the high-speed acquisition board is connected to the main control board and is used to perform real-time pulse triggering on the sampled data stream of the UHF partial discharge signal. The triggered UHF partial discharge pulse is separated into two types of data, feature value messages and waveform data packets, and output to the main control board. The high-speed acquisition board has built-in lossless data path and lossy compression path. The lossless data path is used to output the complete waveform data packet of the triggered UHF partial discharge pulse to the main control board, and the lossy compression path is used to extract feature value messages from the triggered UHF partial discharge pulse and output them to the main control board. The real-time priority of the feature value messages is higher than that of the waveform data packets. Hardware priority scheduling ensures low-latency transmission of high-priority features.

[0024] The computing power board connects to the main control board and is used to receive feature value messages and waveform data packets forwarded by the main control board, perform multimodal large model inference, and send the inference results back to the main control board.

[0025] Figure 1 Solid arrows indicate the data bus, and dashed arrows indicate the power bus. The three boards are interconnected via the PCIe bus, powered by a power module, and installed in an integrated chassis.

[0026] like Figure 2 As shown in the figure (solid arrows represent analog signals, and dashed arrows represent digital signals), the high-speed acquisition board includes a signal conditioning module, an analog-to-digital conversion module, and an FPGA module.

[0027] The signal conditioning module is connected to the analog-to-digital converter module. It receives the UHF partial discharge signal, conditions it, and outputs the conditioned analog signal to the analog-to-digital converter module. The analog-to-digital converter module performs analog-to-digital conversion on the conditioned analog signal and outputs the UHF partial discharge signal sampling data stream to the FPGA module.

[0028] The FPGA module is connected to the analog-to-digital converter module to identify valid discharge pulses in the sampling data stream of UHF partial discharge signals. It extracts the amplitude, phase, and time stamp from the valid discharge pulses. The complete waveform sampling data of the valid discharge pulses is first buffered in the RAM within the FPGA module and then imported into the onboard DDR temporary storage connected to the FPGA module. The FPGA has lossy compression path and lossless compression path. The lossy compression path encapsulates the amplitude, phase, and time stamp of the valid discharge pulses into feature value messages according to a preset message format and outputs them to the main control board. The lossless compression path reads the complete waveform sampling data temporarily stored in the onboard DDR, encapsulates it into waveform data packets according to a preset data packet format, and outputs them to the main control board. The output priority of the feature value messages is higher than that of the waveform data packets.

[0029] It is worth noting that, in response to the technical problem that high sampling rate and low latency cannot be achieved simultaneously under the single bus architecture in the prior art, this invention sets up parallel dual data paths in the high-speed acquisition board to distinguish real-time control feature data from offline analysis waveform data at the hardware level. It achieves microsecond-level low-latency transmission of high-priority features through priority scheduling, while retaining the complete waveform for subsequent in-depth analysis. This balances real-time performance and data integrity, and reduces bus bandwidth pressure and the risk of high-speed data loss.

[0030] Figure 2 In this process, ultra-high frequency signals output from multiple sensors are fed into a signal conditioner. The signal conditioner performs filtering and gain control at two stages before outputting the signal to an ADC. The ADC converts the analog signal into a digital signal, which is then fed into the FPGA. Driven by its onboard clock, the FPGA extracts pulses, performs amplitude, phase, and time-scale calculations, and extracts the waveform. Simultaneously, it outputs gain feedback to the signal conditioner for dynamic gain adjustment. External synchronization signals and B-code signals are fed into the FPGA for time-scale and phase calculations. The FPGA encapsulates the calculation results into feature value messages and network waveform messages, which are then output via Ethernet ports 1, 2, and 3. Finally, it encapsulates these results into PCIe waveform messages, which are output via the PCIe bus. The feature value messages are characteristic value messages, and the PCIe waveform messages are waveform data packets.

[0031] like Figure 4As shown, the specific circuit of the signal conditioning module is as follows: The signal is input through the RFIN port, first limited by a limiter for amplitude protection, then attenuated by a programmable attenuator, then amplified by a first-stage amplifier, and finally filtered by a band-stop filter to suppress narrowband interference. The controller controls a switching switch via an RS232 bus to select the signal to the full-pass, low-pass, or high-pass channel of the band-pass filter. Finally, after amplification by a second-stage amplifier, the conditioned analog signal is output from the RFOUT port to the analog-to-digital converter module. The signal conditioning module has a bandpass range of 300MHz~1500MHz, a sensitivity of -80dBm, an RFIN input power range of -80dBm~0dBm, and an RFOUT output amplitude range of 0mV~3500mV. It has programmable selection of five gain levels from 0~40dB and five attenuation levels from -10dB to -40dB, ensuring flat gain within the detection frequency band and preventing distortion of the UHF partial discharge pulse signal.

[0032] like Figure 5 As shown, the data flow of the FPGA module is as follows: The FPGA receives the sampled data stream output by the ADC. First, it identifies pulses that exceed the threshold through threshold judgment (triggered by register configuration). Then, it extracts the valid discharge pulses. Driven by the onboard clock, the FPGA, in conjunction with the externally input B-code signal and synchronization signal, performs characteristic calculations on the extracted pulses. The calculations include V (amplitude), φ (phase), and t (timescale): amplitude is the peak value of the pulse waveform; phase is the power frequency phase (0°~360°) corresponding to the peak pulse moment, calculated by the synchronization signal and the onboard clock; timescale is the absolute timestamp of the peak pulse moment, calculated by the B-code signal and the onboard clock. The FPGA can also calculate the pulse's rise time, decay time, and spectrum, and output gain feedback to the signal conditioning module based on the current amplitude to dynamically adjust the gain level. For multi-channel synchronization, the FPGA outputs GHz-level synchronization pulses, supporting optical port cascading to achieve multi-board multi-channel synchronization and external phase synchronization. The feature calculation results are encapsulated into feature messages, buffered in RAM, and then output by the network driver to network ports 1, 2, and 3. The pulse waveform obtained from pulse extraction is divided into two paths: one path is encapsulated into a network waveform recording message, buffered in RAM, and then output by the network driver to the network port; the other path is encapsulated into a PCIe waveform recording message, buffered in RAM, written to DDR temporary storage, and then output to the main control board via the PCIe bus. The analog-to-digital conversion module uses a 12-bit resolution, a 5GSPS sampling rate, and an input bandwidth of 30MHz~2500MHz.

[0033] The main control board connects to the high-speed acquisition board and the computing power board, and includes onboard storage and an onboard processor. The onboard storage receives waveform data packets and stores them sequentially according to the acquisition time, continuously accumulating them to form an online sample dataset. The onboard processor continuously calculates the physical trend characteristics of the online sample dataset in a first cycle, and forwards the accumulated feature value messages and waveform data packets to the computing power board in a second cycle. The first cycle is shorter than the second cycle. Before the computing power board sends back the inference results, the onboard processor independently outputs preliminary monitoring results based on the physical trend characteristics. After receiving the inference results, the onboard processor performs confidence-weighted fusion of the preliminary monitoring results and the inference results, using the rate of change of the physical trend characteristics as the dynamic adjustment coefficient of the confidence of the inference results, and outputs the final monitoring results.

[0034] Further, the onboard processor continuously calculates the physical trend characteristics of the online sample dataset in a first cycle, including: the onboard processor is connected to onboard storage and used to read the online sample dataset in the onboard storage at a first cycle interval to calculate spectral characteristics; the onboard processor is connected to a high-speed acquisition board and used to extract the time stamp of each channel from the feature value message output by the high-speed acquisition board, and calculate multi-channel synchronization characteristics based on the time stamp of each channel; at a first cycle interval, the discharge repetition rate, average amplitude, and phase distribution entropy of the amplitude, phase, and time stamp in the online sample dataset are statistically analyzed to obtain multi-timescale trend characteristics; at a first cycle interval, a three-dimensional point cloud spectrum is constructed based on the amplitude, phase, and time stamp in the online sample dataset, and the three-dimensional point cloud spectrum is analyzed... Point cloud density, cluster number, and average curvature are extracted to obtain point cloud spectral features. The onboard processor uses spectral features, multi-channel synchronization features, multi-timescale trend features, and point cloud spectral features as physical trend features. When any two of the physical trend features meet the anomaly judgment conditions, the corresponding preliminary monitoring results are directly output without waiting for the computing power board's inference results. The physical trend features are also retained and used to fuse with the inference results returned by the computing power board to generate the final monitoring results. The first cycle is shorter than the second cycle of the main control board forwarding data to the computing power board. Within the duration of each second cycle, the physical trend features are updated and calculated multiple times, and the preliminary monitoring results are continuously output without waiting for the computing power board to return the inference results, maintaining a low-latency early warning response.

[0035] Furthermore, the onboard processor is connected to the onboard storage to read the written online sample dataset from the onboard storage and calculate the spectral characteristics of the waveform data packets in the online sample dataset. The onboard processor is also connected to the high-speed acquisition board to extract the time stamp of each channel from the feature value message output by the high-speed acquisition board, and calculate the multi-channel synchronization characteristics based on the time stamp of each channel. It obtains multi-timescale trend characteristics by statistically analyzing the discharge repetition rate, average amplitude, and phase distribution entropy of the amplitude, phase, and time stamp in the online sample dataset. A three-dimensional point cloud spectrum is constructed based on the amplitude, phase, and time stamp in the online sample dataset, and point cloud density, cluster number, and average curvature are extracted from the three-dimensional point cloud spectrum to obtain point cloud spectrum features. The onboard processor uses the spectral characteristics, multi-channel synchronization characteristics, multi-timescale trend features, and point cloud spectrum features as physical trend features for fusion with the inference results returned by the computing board. In this embodiment, the onboard processor uses a CPU running a Linux kernel-based operating system, the onboard storage uses a 4TB SSD, and the main control board is also equipped with DDR memory and cache, providing optical ports, network ports, PCIe interfaces, and debugging interfaces.

[0036] It is worth noting that, in response to the technical problem in existing technologies where monitoring devices heavily rely on computing power units and warnings are interrupted when computing power fails, this invention adopts an asynchronous dual-cycle inference design. Physical trends are calculated quickly in a short cycle and output preliminary warnings independently, while AI inference is performed at a long cycle with a slow pace to enhance verification. Even if the computing power board fails or inference times out, the system can still maintain basic warning functions, thereby improving the reliability of the device from the architectural level and reducing the risk of monitoring interruption caused by computing power dependence.

[0037] Furthermore, the main control board also integrates an AI management module; the AI ​​management module is connected to the onboard processor and the computing power board; the inference results returned by the computing power board include defect probability distribution, fault mechanism explanation information, and defect characterization vector; the AI ​​management module is used to receive the preliminary monitoring results output by the onboard processor and the inference results returned by the computing power board, and dynamically allocates the confidence weights of the defect probability distribution corresponding to the preliminary monitoring results and the defect probability distribution in the inference results with the change rate of physical trend characteristics as a dynamic adjustment coefficient. After performing confidence weighted fusion of the defect probability distribution corresponding to the preliminary monitoring results and the defect probability distribution in the inference results, the fusion result is concatenated with the physical trend characteristics, fault mechanism explanation information, and defect characterization vector to form an initial fusion vector. After transformation by the fully connected layer built into the AI ​​management module, a multi-source feature fusion criterion is output. Multi-level decisions are performed on the multi-source feature fusion criterion, and the partial discharge type and fault mechanism conclusion are determined based on the decision results as the monitoring results output.

[0038] It is worth noting that, in response to the technical problem that the fusion of physical features and AI features in the existing technology adopts fixed weights and cannot adapt to different discharge development rates, this invention dynamically adjusts the fusion weights by changing the physical trend rate. The faster the discharge state changes, the higher the weight of the real-time physical result; the more stable the discharge state, the higher the weight of the AI ​​accuracy. This adaptive approach balances real-time performance and accuracy, improves the scene adaptability of the fusion criteria, and reduces the scene deviation caused by fixed weights.

[0039] The AI ​​management module also includes a first-level judgment unit, a second-level judgment unit, a third-level judgment unit, and an arbitration unit.

[0040] The first-level decision unit is connected to the fully connected layer and is used to receive the multi-source feature fusion criteria output by the fully connected layer. It calculates the voting scores of various defects through softmax. When the highest score is greater than the set threshold, the defect category with the highest score is taken as the first candidate category. When the voting score of the first candidate category is greater than the set threshold, the partial discharge type and fault mechanism conclusion corresponding to the first candidate category are directly output. When the voting score of the first candidate category is not greater than the set threshold, the first-level decision unit sends the multi-source feature fusion criterion to the second-level decision unit. The second-level decision unit calculates the posterior probability based on the pre-statistical prior probability and the conditional probability of the multivariate Gaussian distribution, and takes the category with the largest posterior probability as the second candidate category. When the second candidate category and the first candidate category are the same category, the partial discharge type and fault mechanism conclusion corresponding to the same category are output. When the second candidate category is different from the first candidate category, the second-level decision unit sends the multi-source feature fusion criterion to the third-level decision unit. The third-level decision unit extracts the waveform, spectrum and three-dimensional point cloud spectrum of the pulse to be judged from the multi-source feature fusion criterion, matches them with the preset templates, and takes the category with the highest confidence weighted vote as the third candidate category. When the weighted vote of the third candidate category is greater than the set threshold, the partial discharge type and fault mechanism conclusion corresponding to the third candidate category are output. When the weighted votes of the third candidate category are not greater than the set threshold, the third-level decision unit sends the multi-source feature fusion criterion to the arbitration unit. The arbitration unit marks the pulse as a sample to be manually reviewed, outputs an alarm prompt, and saves the multi-source feature fusion criterion for manual review to determine the partial discharge type and fault mechanism conclusion.

[0041] It is worth noting that, in response to the technical problem in the existing technology that the single-level decision-making process applies to samples of different difficulty in a one-size-fits-all manner, resulting in wasted computation of high-confidence samples and inaccurate identification of low-confidence samples, this invention sets up a three-level progressive decision-making architecture. The first two levels quickly output simple samples, and the third level provides a backup for difficult samples through micro-matching. This not only improves the decision-making efficiency but also reduces the misjudgment rate of low-confidence samples.

[0042] The third-level decision unit includes a time-domain matching subunit, a spectrum matching subunit, a point cloud matching subunit, and a weighted voting unit.

[0043] The time-domain matching subunit is connected to the second-level decision unit and is used to extract the waveform of the pulse to be judged from the multi-source feature fusion criterion sent by the second-level decision unit, calculate the normalized cross-correlation function between the waveform and the average waveform template of various defects, take the defect category corresponding to the maximum value of the normalized cross-correlation as the time-domain voting category, take the maximum value as the time-domain confidence, and output it to the weighted voting unit. The spectrum matching subunit is connected to the second-level decision unit and is used to extract the spectrum of the pulse to be judged from the multi-source feature fusion criterion. It extracts the peak frequency, center frequency and bandwidth triplet from the spectrum, calculates the Euclidean distance between the triplet and the triplet of various defect templates, takes the defect category corresponding to the minimum distance as the spectrum voting category, and takes the reciprocal of the minimum distance as the spectrum confidence and outputs it to the weighted voting unit. The point cloud matching subunit is connected to the second-level decision unit. It is used to extract the three-dimensional point cloud spectrum of the pulse to be discriminated from the multi-source feature fusion criteria, perform persistent homology analysis on the three-dimensional point cloud to generate a Betti number persistent spectrum, calculate the cosine similarity between the Betti number persistent spectrum and various defect spectrum template libraries, take the defect category corresponding to the maximum cosine similarity as the point cloud voting category, take the maximum value as the point cloud confidence, and output the point cloud voting category and point cloud confidence to the weighted voting unit. The weighted voting unit is connected to the time-domain matching subunit, the spectrum matching subunit, and the point cloud matching subunit. It is used to receive the voting categories and confidence levels output by the three subunits, accumulate the confidence levels of the same defect category in the three subunits, obtain the weighted votes for each defect category, and take the defect category with the highest weighted votes as the third candidate category and output the third candidate category.

[0044] It is worth noting that, in response to the technical problems of unstable single-modal template matching on boundary samples and susceptibility to amplitude and sampling changes in the prior art, this invention improves the accuracy and robustness of the third-level decision by weighted fusion of matching results from three independent dimensions: time domain, spectrum, and point cloud, and multi-dimensional cross-validation, thereby reducing the impact of random errors in single-modal matching.

[0045] Furthermore, the computing power board is used to deploy multimodal large models; the multimodal large models include 1D-CNN modules, 2D-CNN modules, Point-Transformer modules, and cross-modal fusion modules; The input terminals of the 1D-CNN module, 2D-CNN module, and Point-Transformer module are all connected to the main control board, and are used to jointly receive the feature value messages and waveform data packets forwarded by the main control board; The 1D-CNN module is used to extract the original time series consisting of amplitude, phase, and time stamp from feature value messages and waveform data packets, and outputs a time-series encoded vector through convolutional encoding. The 2D-CNN module is used to construct PRPD spectrograms, PRPS spectrograms, and time-frequency spectrograms from feature value messages and waveform data packets, and output spectrogram encoding vectors through convolutional coding; The Point-Transformer module is used to construct a 3D point cloud spectral map from feature value messages and waveform data packets, and output the point cloud encoded vector through self-attention encoding. The cross-modal fusion module is connected to the 1D-CNN module, 2D-CNN module and Point-Transformer module respectively. It is used to receive temporal coding vectors, graph coding vectors and point cloud coding vectors, obtain the defect characterization vector by cross-modal attention weighted summation, map it into the probability distribution of various defects by the fully connected classification head, and send the defect probability distribution, fault mechanism explanation information and defect characterization vector as inference results back to the main control board.

[0046] It is worth noting that, in response to the technical problems of incomplete information in single-modal models and the inability of a unified network structure to adapt to the characteristics of different modal data in existing technologies, this invention adapts to the characteristics of three types of data—time series, graph, and point cloud—through a multimodal branch architecture, and integrates multi-dimensional features through cross-modal attention fusion, thereby improving the comprehensiveness and recognition accuracy of defect feature extraction and reducing recognition errors caused by the loss of single-modal information.

[0047] The computing board provides a peak computing power of no less than 280 TOPS at INT8 precision, supports the Transformer and PyTorch frameworks, is compatible with ARM and x86 architecture main controllers, and supports multi-card collaboration and single-card independent computing. The 1D-CNN module contains 3 convolutional layers (kernel sizes 3, 5, and 7) and 2 pooling layers; the 2D-CNN module contains 4 convolutional layers (3×3) and 3 pooling layers. The PRPD spectrum divides the power frequency cycle into 360 1° phase intervals for statistical discharge distribution, and the time-frequency spectrum uses a Hanning window with 256 points and a stride of 64 points for short-time Fourier transform; the Point-Transformer module contains 4 self-attention layers and 2 downsampling layers; the cross-modal fusion module also outputs an attention heatmap for interpretability analysis. The training data for the multimodal large model includes four categories: real-world field data, laboratory simulation data, standard source data, and simulation data. Before training, the time reference is unified using B-code timestamps, linear correction is performed using standard sources, and phase is aligned using power frequency zero-crossing points. Preprocessing uses db4 wavelet 5-layer soft thresholding to remove white noise and adaptive notch filtering to remove power frequency doubling. Sample augmentation uses temporal scaling (0.8~1.2 times), temporal translation (±10 points), noise addition (10~30dB), spectrum rotation (±5°), spectrum cropping (90%), brightness adjustment (±10%), point cloud random sampling (95%), point cloud jitter (5%), and GAN generation. Feature extraction introduces domain adversarial training, physical constraint autoencoders, and point cloud density clustering. The training process is to first perform self-supervised pre-training through mask reconstruction, and then fine-tune it with cross-entropy loss on labeled samples.

[0048] Furthermore, the power supply module adopts a dual-redundant DC power supply design, featuring overvoltage, overcurrent, and short-circuit protection, and supports wide-voltage input and hot-swapping. The integrated chassis is a VPX architecture all-metal shielded chassis with an IP40 protection rating. All boards are interconnected via the VPX backplane PCIe bus, providing unified thermal management. In this embodiment, two high-speed acquisition boards, two computing boards, two power supply modules, and one main control board are configured. The integrated chassis is an all-metal shielded structure with an IP40 protection rating, integrating all boards and modules and providing thermal management. The power supply module adopts a dual-redundant DC power supply design, featuring overvoltage, overcurrent, and short-circuit protection, and supports wide-voltage input and hot-swapping. All boards are interconnected via the PCIe bus, and the backplane power bus supplies power to each board. Figure 1 Solid arrows indicate the data bus, and dashed arrows indicate the power bus.

[0049] like Figure 3 As shown, this embodiment 2 provides a partial discharge ultra-high frequency storage and computing integrated method, which operates on the device described in embodiment 1, and includes the following steps: S1: Real-time pulse triggering is performed on the sampled data stream of the UHF partial discharge signal to separate the triggered UHF partial discharge pulse into feature value messages and waveform data packets; the feature value messages of the UHF partial discharge pulse are extracted through a lossy compression path, and the complete waveform data packets of the UHF partial discharge pulse are output through a lossless data path; the real-time priority of the feature value messages is higher than that of the waveform data packets.

[0050] Specifically, the UHF partial discharge signal is acquired by an UHF sensor and then processed by an RFI input signal conditioning module. It then sequentially passes through a limiter for amplitude limiting, a digitally controlled attenuator for programmable attenuation, a first-stage amplification, a band-stop filter, a band-pass filter, and a second-stage amplification before being output as a conditioned analog signal to an analog-to-digital converter (ADC). The ADC performs 12-bit, 5GSPS analog-to-digital conversion on the conditioned analog signal, outputting the UHF partial discharge signal sampling data stream to the FPGA module. The FPGA module identifies valid discharge pulses based on over-threshold triggering rules, and after background noise filtering, extracts the amplitude, phase, and time scale (time). The amplitude is calculated by an external B-code synchronization signal in conjunction with an onboard crystal oscillator, and the front-end gain level is dynamically adjusted according to the amplitude. The complete waveform sampling data is first buffered in the FPGA's internal RAM and then written to the onboard DDR temporary storage. The lossy compression path encapsulates the amplitude, phase, and time stamp into a feature value message (network frame) according to a preset message format, which is then buffered in RAM and output through the network port. The lossless compression path reads the complete waveform sampling data from the onboard DDR, encapsulates it into a waveform data packet (PCIe frame) according to a preset data packet format, and then outputs it to the main control board via the PCIe bus according to priority after being buffered in RAM and temporarily stored in DDR. In multi-channel scenarios, the FPGA achieves multi-channel synchronization through GHz synchronization pulses and optical port cascading.

[0051] S2: The received waveform data packets are stored sequentially according to the acquisition time to form an online sample dataset; the physical trend characteristics of the online sample dataset are continuously calculated in the first cycle, and the accumulated feature value messages and waveform data packets are forwarded in the second cycle to perform multimodal large model inference to obtain the inference results; the first cycle is shorter than the second cycle.

[0052] Specifically, the main control board writes waveform data packets into the onboard storage and accumulates them to build an online sample dataset in the order of acquisition time. The onboard processor reads the sample set at the first short cycle interval, calculates the spectral features, multi-channel synchronization features, multi-timescale trend features, and point cloud spectral features to obtain physical trend features. At the same time, at the second long cycle interval, the accumulated feature value messages and waveform data packets are forwarded to the computing power board to trigger multimodal large model inference and obtain inference results containing defect probability distribution, fault mechanism explanation information, and defect characterization vectors. The physical trend features are updated multiple times within each second cycle without waiting for the inference results to return.

[0053] S3: Before the inference results are returned, preliminary monitoring results are output based on physical trend characteristics; after the inference results are received, the preliminary monitoring results and the inference results are weighted and fused with confidence, and the rate of change of physical trend characteristics is used as the dynamic adjustment coefficient of the confidence of the inference results to output the final monitoring results.

[0054] Specifically, the 1D-CNN module (3 convolutional layers with 3 / 5 / 7 kernels + 2 pooling layers) extracts temporal encoding vectors from the original time series; the 2D-CNN module (4 convolutional layers with 3×3+3 pooling layers) extracts spectral encoding vectors from the PRPD spectrum (360 1° phase intervals), PRPS spectrum, and time-frequency spectrum (Hanning window 256-point stride 64-point STFT); the Point-Transformer module (4 self-attention layers, 2 downsampling layers) extracts point cloud encoding vectors from the 3D point cloud spectrum; the cross-modal fusion module performs cross-modal attention weighted summation on the three encoding vectors to obtain the defect representation vector, outputs the probability distribution of various defects and attention heatmaps through the fully connected classification head, and sends the defect probability distribution, fault mechanism explanation information, and defect representation vectors back to the main control board as inference results. The 1D-CNN module extracts the temporal encoding vector, the 2D-CNN module extracts the graph encoding vector, the Point-Transformer module extracts the point cloud encoding vector, the cross-modal fusion module obtains the defect representation vector by cross-modal attention weighted summation, and outputs the probability distribution of various defects and attention heatmaps through the fully connected classification head, which are then fed back as inference results.

[0055] When any two of the physical trend characteristics meet the anomaly judgment criteria, the preliminary monitoring results are directly output to achieve low-latency early warning; after receiving the inference results, the dynamic adjustment coefficient of the confidence level of the inference results is obtained by weighting the relative change rate of the discharge repetition rate and the average amplitude within the first cycle, as expressed by the following formula:

[0056] In the formula, This indicates the preset weighting coefficient. The configuration is based on the on-site working conditions; the larger the absolute value of k, the faster the rate of change of the discharge state. The relative rate of change of the discharge repetition rate is expressed by the following formula:

[0057] In the formula, , These are the discharge repetition rates for the nth and (n-1th)th first cycles, respectively. The relative rate of change of the average amplitude is expressed by the following formula:

[0058] In the formula, , These are the average amplitudes of the nth and (n-1)th first cycles, respectively.

[0059] Based on the dynamic adjustment coefficient of the confidence level of the inference results, the confidence weights of the preliminary monitoring results and the inference results are calculated separately, as expressed by the following formula:

[0060]

[0061] In the formula, The confidence weights representing the inference results The confidence weight of the preliminary monitoring results is indicated by the faster the physical trend characteristics change (the larger the absolute value of the dynamic adjustment coefficient).

[0062] The lower the confidence weight of the inference result, the higher the confidence weight of the preliminary monitoring result. The preliminary monitoring result and the inference result are then fused with confidence weights to obtain the fused confidence result, expressed by the following formula:

[0063] In the formula, This represents the fused defect probability distribution vector, with dimensions matching the total number of defect categories. It serves as the confidence score result after fusion. This represents the defect probability distribution vector corresponding to the preliminary monitoring results. This represents the probability distribution vector of defects in the reasoning result.

[0064] After fusion, the fused confidence result is concatenated with physical trend features, fault mechanism explanation information, and defect characterization vector to form an initial fusion vector. After transformation by a fully connected layer, the multi-source feature fusion criterion is output and sent to the subsequent multi-level decision process to finally output the monitoring result.

[0065] The returned inference results (including defect probability distribution, fault mechanism explanation information, and defect representation vector) and the physical trend features calculated by the main control board can be deeply integrated in two ways: First, feature-level fusion, which involves concatenating the high-dimensional defect representation vector output by the computing board with physical trend features (such as amplitude change rate, pulse repetition frequency, etc.) or merging them through a gating mechanism to form a joint feature vector; Second, physical-semantic alignment fusion, which involves establishing a mapping relationship between AI defect probability and changes in specific physical quantities to make the inference results interpretable. This fusion mechanism significantly improves subsequent multi-level discrimination. It upgrades the traditional single-level AI discrimination to a multi-level architecture of "preliminary physical trend discrimination - AI multimodal deep inference - comprehensive feature / decision-level fusion adjudication," making the discrimination process more rigorous and hierarchical. At the same time, physical trend features can output preliminary monitoring results in real time, avoiding the delay caused by waiting for computing board inference. Even if the computing board fails or the inference takes too long, the system can still maintain basic early warning functions, greatly improving overall reliability. In addition, physical trend features are naturally good at describing the state evolution process, complementing the pattern recognition capabilities of AI. This allows the system to not only identify the current defect type, but also predict future state evolution based on the long-term trend of physical quantities, thus possessing predictive maintenance capabilities. Finally, the fused comprehensive judgment results can be further fed back to the sample management strategy of the main control board. For example, when the physical trend is stable and the AI ​​inference confidence is high, low-value samples are prioritized for elimination, while when the two conflict or both point to high risk, high-value samples are locked for storage, thereby optimizing the utilization efficiency of storage and computing resources.

[0066] Figure 3In this process, UHF signals are acquired by UHF sensors and sent to a signal conditioning module. After conditioning, they are converted into digital signals by an ADC and sent to an FPGA. The FPGA performs noise filtering on the sampled data, and the filtered data is split into two paths: one path performs original waveform recording, and the other path performs eigenvalue calculation to obtain amplitude, phase, and time-stamped eigenvalues. The eigenvalue calculation results are encapsulated into network frame packets (eigenvalue packets), buffered in the FPGA's onboard RAM, and then sent via the network protocol. The original waveform recording data is encapsulated into PCIe frame packets (waveform data packets), buffered in the FPGA's onboard RAM, written to the DDR cache, and then sent to the main unit via the PCIe bus. The process involves several steps: Step 3 involves sending PCIe frame packets to the main control board via the PCIe bus, caching them in the main control board's RAM and DDR, and then writing them to the SSD storage to continuously accumulate and form a sample dataset. The main control board also forwards network frame packets and PCIe frame packets to the computing power board. Step 4 involves sending the data to the computing power board, caching it in the computing power board's DDR, and then having the multimodal large model perform AI inference. The inference results are then sent back to the main control board via the PCIe bus. Step 5 involves the main control board fusing the returned inference results with the physical trend characteristics calculated in real time from the onboard storage, and then sending the partial discharge pulse waveform and monitoring results via the network protocol.

[0067] Furthermore, the signal conditioning module sequentially passes through a limiter, a programmable attenuator, a first-stage amplification, a band-stop filter, a band-pass filter, and a second-stage amplification to output the conditioned analog signal; the ADC samples at 12 bits and 5 GSPS; the FPGA obtains the effective discharge pulse through threshold judgment and pulse extraction, and calculates V, φ, and t under the drive of the onboard clock, combined with the B code signal and the synchronization signal, and adjusts the front-end gain according to the amplitude output gain feedback.

[0068] Furthermore, the onboard processor calculates spectral features, multi-channel synchronization features, multi-timescale trend features, and point cloud spectral features as physical trend features for the online sample dataset.

[0069] Furthermore, the AI ​​management module concatenates the physical trend features and inference results into an initial fusion vector, which is then transformed by a fully connected layer and subjected to multi-level decision-making to determine the partial discharge type and fault mechanism before outputting the monitoring results.

[0070] Taking a converter transformer partial discharge scenario as an example, after a partial discharge occurs in the converter transformer, the UHF sensor collects the signal and sends it to the high-speed acquisition board. After being conditioned by the signal conditioning module, the ADC samples the signal into a UHF partial discharge signal sampling data stream, which is then sent to the FPGA. When the sampled value exceeds the trigger threshold, the FPGA extracts the effective discharge pulse and calculates the amplitude, phase, and time scale. The complete waveform is cached in RAM and written to DDR, and then encapsulated into feature value messages and waveform data packets, which are output to the main control board. The main control board writes the waveform data packets to the onboard storage to accumulate and form an online sample dataset. While writing new waveform data packets, the onboard processor reads the already written online sample dataset to calculate the physical trend characteristics and forwards the feature value messages and waveform data packets to the computing board. The computing board performs multimodal large model inference and then sends the inference results back to the main control board. The main control board fuses the inference results with the physical trend characteristics, determines the partial discharge type and fault mechanism through multi-level decision-making, and outputs the monitoring results.

[0071] It is worth noting that, in response to the technical problems of poor real-time performance caused by the separation of acquisition and calculation and the external transmission of data in existing partial discharge monitoring methods, this invention uses an acquisition board to perform pulse triggering on the sampled data stream in real time and separate feature value messages and waveform data packets to send to the main control board. The main control board forwards the data to the computing board and simultaneously writes the waveform into the onboard storage to accumulate the online sample set. The onboard processor then calculates the physical trend characteristics in real time. The computing board performs multimodal large model inference and then sends back the results. The main control board fuses the inference results and physical trend characteristics to output the monitoring results, thereby improving the execution efficiency of the online partial discharge monitoring method and reducing the method's dependence on the external computing environment.

[0072] Compared with the prior art, the beneficial effects of the present invention include at least the following: (1) This invention achieves the integration of storage and computing in terms of hardware structure. Storage is manifested in the onboard storage on the main control board, which continuously receives waveform data packets output by the high-speed acquisition board and accumulates them in the order of acquisition time to form an online sample dataset; computing is manifested in two parts—the onboard processor on the main control board calculates the physical trend characteristics of the online sample dataset in the onboard storage in real time, and the AI ​​acceleration chip on the computing board performs multimodal large model inference on the feature value messages and waveform data packets forwarded by the main control board. Storage and computing are not distributed in different devices or different systems, but are integrated on three boards of the same device: the onboard storage and onboard processor are directly connected inside the main control board through the onboard high-speed bus, and the computing board is directly connected to the main control board through the PCIe bus. The storage resources and computing resources belong to the same chassis, are connected through the same backplane bus, and are powered by the same power supply in terms of hardware structure, forming a board-level integrated storage and computing hardware platform. This structure improves the system integration and reduces the data transmission overhead between storage and computing.

[0073] (2) The storage and computation processes are structurally interconnected, enabling simultaneous storage and computation. The onboard storage's write port continuously receives new waveform data packets, while the onboard processor's read port simultaneously accesses the online sample dataset already written to the onboard storage for computation. The write and read computation operations are physically concurrent and do not block each other on the onboard bus. Each time a waveform data packet is sent from the high-speed acquisition board, the data is written to the onboard storage to expand the online sample set, and simultaneously, the onboard processor immediately performs incremental computation of physical trend features from the stored samples. The storage and computation processes overlap on the same timeline. This structure improves the parallelism of sample accumulation and feature extraction, and reduces the processing latency caused by the "store first, compute later" approach.

[0074] (3) The three cards form a closed-loop data association in structure. The output of the high-speed acquisition card is connected to the main control card. The lossy compression path and lossless compression path inside the FPGA are structurally output in two ways: the feature value message and waveform data packet are forwarded to the computing power card via the main control card to drive AI calculation. The waveform data packet is simultaneously written to the onboard storage to drive the physical trend calculation of the onboard processor. The output of the computing power card is connected back to the main control card via the PCIe bus to send the AI ​​inference result back to the main control card. The AI ​​management module on the main control card is structurally connected to both the onboard processor and the computing power card. After receiving the two calculation results, it is fused and output. The five links of acquisition, storage, AI calculation, physical calculation and fusion form a data loop that is connected end to end through the physical connection between the cards. The data flows between the three cards along a fixed path without the need for external scheduling. This structure improves the determinism of data flow and reduces the system's dependence on external networks and background scheduling.

[0075] (4) The two computation paths are structurally complementary and interconnected. The multimodal large model on the computing board is responsible for extracting deep defect characterization from the feature value messages and waveform data packets of single pulses, while the onboard processor is responsible for extracting the discharge development trend from the long-term accumulated samples stored on the board. The two computation paths run on different computing units of different boards, process data at different time scales, and output results at different levels of abstraction. However, at the AI ​​management module of the main control board, they are combined into a single multi-source feature fusion criterion through a fully connected layer structure, and the monitoring results are output after multi-level decision-making. This structure enables AI inference and physical trend analysis to be independently parallel in hardware and structurally integrated, improving the accuracy of partial discharge type identification and fault mechanism analysis, and reducing the risk of misjudgment from a single computation path.

[0076] Embodiment 3 of the present invention proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded onto the processor, it implements the partial discharge ultra-high frequency memory-computer integration method described in Embodiment 1.

[0077] Embodiment 4 of the present invention proposes a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the partial discharge ultra-high frequency storage and computing integrated method described in Embodiment 1.

[0078] This disclosure can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of this disclosure.

[0079] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A partial discharge ultra-high frequency storage and computing integrated device, characterized in that, include: High-speed acquisition boards, main control boards, and computing power boards; The main control board is connected to the high-speed acquisition board and the computing power board respectively; The high-speed acquisition board is used to perform real-time pulse triggering on the sampled data stream of UHF partial discharge signals. It separates the triggered UHF partial discharge pulses into characteristic value messages and waveform data packets, which are then output to the main control board. The high-speed acquisition board has built-in lossless and lossy compression paths. The lossless data path outputs the complete waveform data packets of the triggered UHF partial discharge pulses to the main control board; the lossy compression path extracts characteristic value messages from the triggered UHF partial discharge pulses and outputs them to the main control board. The real-time priority of feature value messages is higher than that of waveform data packets; The main control board includes onboard storage and an onboard processor, with the onboard processor connected to the onboard storage; The onboard storage is used to store the received waveform data packets sequentially according to the acquisition time, accumulating them to form an online sample dataset; the onboard processor continuously calculates the physical trend characteristics of the online sample dataset in the first cycle, and forwards the accumulated feature value messages and waveform data packets to the computing board in the second cycle; the first cycle is shorter than the second cycle; Before the computing board transmits the inference results, the onboard processor independently outputs preliminary monitoring results based on physical trend characteristics. After receiving the inference results, the onboard processor performs confidence-weighted fusion of the preliminary monitoring results and the inference results, uses the rate of change of physical trend characteristics as the dynamic adjustment coefficient of the confidence of the inference results, and outputs the final monitoring results. The onboard processor continuously calculates the physical trend characteristics of the online sample dataset in a first cycle, including: The onboard processor is connected to the onboard storage and is used to read the online sample dataset in the onboard storage at intervals of the first cycle to calculate spectral features. The onboard processor connects to the high-speed acquisition board and is used to extract the time stamps of each channel from the feature value messages output by the high-speed acquisition board. Based on the time stamps of each channel, it calculates multi-channel synchronization features. Using the first period as an interval, it statistically analyzes the discharge repetition rate, average amplitude, and phase distribution entropy of the amplitude, phase, and time stamp in the online sample dataset to obtain multi-timescale trend features. Using the first period as an interval, it constructs a three-dimensional point cloud spectrum based on the amplitude, phase, and time stamp in the online sample dataset. From the three-dimensional point cloud spectrum, it extracts the point cloud density, cluster number, and average curvature to obtain point cloud spectrum features. The onboard processor uses the spectrum features, multi-channel synchronization features, multi-timescale trend features, and point cloud spectrum features as physical trend features. When any two of the physical trend features meet the anomaly judgment conditions, it directly outputs the corresponding preliminary monitoring results without waiting for the inference results from the computing board. The physical trend features are also retained and used to fuse with the inference results returned by the computing board to generate the final monitoring results. The first cycle is shorter than the second cycle in which the main control board forwards data to the computing power board; within the duration of each second cycle, the physical trend characteristics are updated multiple times and preliminary monitoring results are continuously output without waiting for the computing power board to return inference results, thus maintaining a low-latency early warning response. The computing power board is used to receive feature value messages and waveform data packets forwarded by the main control board, perform multimodal large model inference, and send the inference results back to the main control board.

2. The partial discharge ultra-high frequency storage and computing integrated device according to claim 1, characterized in that: The high-speed acquisition board includes: a signal conditioning module, an analog-to-digital converter module, and an FPGA module; the signal conditioning module is connected to the analog-to-digital converter module, and the FPGA module is also connected to the analog-to-digital converter module. The signal conditioning module is used to receive and condition the ultra-high frequency partial discharge signal, and the resulting conditioned analog signal is output to the analog-to-digital converter module. The analog-to-digital conversion module is used to perform analog-to-digital conversion on the conditioned analog signal, and the resulting UHF partial discharge signal sampling data stream is output to the FPGA module; The FPGA module is used to identify valid discharge pulses in the sampling data stream of UHF partial discharge signals, extract the amplitude, phase, and time stamp of the valid discharge pulses, and first cache the complete waveform sampling data of the valid discharge pulses in the RAM within the FPGA module, and then import them into the onboard DDR temporary storage connected to the FPGA module. The FPGA has lossy compression path and lossless compression path. The lossy compression path encapsulates the amplitude, phase, and time stamp of the valid discharge pulses into feature value messages according to a preset message format and outputs them to the main control board. The lossless compression path reads the complete waveform sampling data temporarily stored in the onboard DDR, encapsulates it into waveform data packets according to a preset data packet format, and outputs them to the main control board. The output priority of the feature value messages is higher than that of the waveform data packets.

3. The partial discharge ultra-high frequency storage and computing integrated device according to claim 1, characterized in that, The inference results returned by the computing board include defect probability distribution, fault mechanism explanation information, and defect characterization vector; The main control board also integrates an AI management module. The AI ​​management module connects to the onboard processor and the computing power board. It receives the preliminary monitoring results output by the onboard processor and the inference results returned by the computing power board. It dynamically allocates the confidence weights of the defect probability distribution corresponding to the preliminary monitoring results and the defect probability distribution in the inference results using the rate of change of physical trend characteristics as a dynamic adjustment coefficient. After the defect probability distribution corresponding to the preliminary monitoring results and the defect probability distribution in the inference results are fused with confidence weights, the fused result is concatenated with physical trend characteristics, fault mechanism explanation information, and defect characterization vector to form an initial fusion vector. After transformation by the fully connected layer built into the AI ​​management module, a multi-source feature fusion criterion is output. Multi-level decisions are performed on the multi-source feature fusion criterion. Based on the decision results, the partial discharge type and fault mechanism conclusion are determined and output as the monitoring result.

4. The partial discharge ultra-high frequency storage and computing integrated device according to claim 3, characterized in that, include: The AI ​​management module also includes a first-level judgment unit, a second-level judgment unit, a third-level judgment unit, and an arbitration unit; The first-level decision unit is connected to the fully connected layer and is used to receive the multi-source feature fusion criteria output by the fully connected layer. It calculates the voting scores of various defects through softmax. When the highest score is greater than the set threshold, the defect category with the highest score is taken as the first candidate category. When the voting score of the first candidate category is greater than the set threshold, the partial discharge type and fault mechanism conclusion corresponding to the first candidate category are directly output. When the voting score of the first candidate category is not greater than the set threshold, the first-level decision unit sends the multi-source feature fusion criterion to the second-level decision unit. The second-level decision unit calculates the posterior probability based on the pre-statistical prior probability and the conditional probability of the multivariate Gaussian distribution, and takes the category with the largest posterior probability as the second candidate category. When the second candidate category and the first candidate category are the same category, the partial discharge type and fault mechanism conclusion corresponding to the same category are output. When the second candidate category is different from the first candidate category, the second-level decision unit sends the multi-source feature fusion criterion to the third-level decision unit. The third-level decision unit extracts the waveform, spectrum and three-dimensional point cloud spectrum of the pulse to be judged from the multi-source feature fusion criterion, matches them with the preset templates, and takes the category with the highest confidence weighted vote as the third candidate category. When the weighted vote of the third candidate category is greater than the set threshold, the partial discharge type and fault mechanism conclusion corresponding to the third candidate category are output. When the weighted votes of the third candidate category are not greater than the set threshold, the third-level decision unit sends the multi-source feature fusion criterion to the arbitration unit. The arbitration unit marks the pulse as a sample to be manually reviewed, outputs an alarm prompt, and saves the multi-source feature fusion criterion for manual review to determine the partial discharge type and fault mechanism conclusion.

5. The partial discharge ultra-high frequency storage and computing integrated device according to claim 4, characterized in that, include: The third-level decision unit includes a time-domain matching subunit, a spectrum matching subunit, a point cloud matching subunit, and a weighted voting unit. The time-domain matching subunit is connected to the second-level decision unit and is used to extract the waveform of the pulse to be judged from the multi-source feature fusion criterion sent by the second-level decision unit, calculate the normalized cross-correlation function between the waveform and the average waveform template of various defects, take the defect category corresponding to the maximum value of the normalized cross-correlation as the time-domain voting category, take the maximum value as the time-domain confidence, and output it to the weighted voting unit. The spectrum matching subunit is connected to the second-level decision unit and is used to extract the spectrum of the pulse to be judged from the multi-source feature fusion criterion. It extracts the peak frequency, center frequency and bandwidth triplet from the spectrum, calculates the Euclidean distance between the triplet and the triplet of various defect templates, takes the defect category corresponding to the minimum distance as the spectrum voting category, and takes the reciprocal of the minimum distance as the spectrum confidence and outputs it to the weighted voting unit. The point cloud matching subunit is connected to the second-level decision unit. It is used to extract the three-dimensional point cloud spectrum of the pulse to be discriminated from the multi-source feature fusion criteria, perform persistent homology analysis on the three-dimensional point cloud to generate a Betti number persistent spectrum, calculate the cosine similarity between the Betti number persistent spectrum and various defect spectrum template libraries, take the defect category corresponding to the maximum cosine similarity as the point cloud voting category, take the maximum value as the point cloud confidence, and output the point cloud voting category and point cloud confidence to the weighted voting unit. The weighted voting unit is connected to the time-domain matching subunit, the spectrum matching subunit, and the point cloud matching subunit. It is used to receive the voting categories and confidence levels output by the three subunits, accumulate the confidence levels of the same defect category in the three subunits, obtain the weighted votes for each defect category, and take the defect category with the highest weighted votes as the third candidate category and output the third candidate category.

6. The partial discharge ultra-high frequency storage and computing integrated device according to claim 1, characterized in that, include: The computing power board is used to deploy multimodal large models; the multimodal large models include 1D-CNN modules, 2D-CNN modules, Point-Transformer modules, and cross-modal fusion modules; The input terminals of the 1D-CNN module, 2D-CNN module, and Point-Transformer module are all connected to the main control board, and are used to jointly receive the feature value messages and waveform data packets forwarded by the main control board; The 1D-CNN module is used to extract the original time series consisting of amplitude, phase, and time stamp from feature value messages and waveform data packets, and outputs a time-series encoded vector through convolutional encoding. The 2D-CNN module is used to construct PRPD spectrograms, PRPS spectrograms, and time-frequency spectrograms from feature value messages and waveform data packets, and output spectrogram encoding vectors through convolutional coding; The Point-Transformer module is used to construct a 3D point cloud spectral map from feature value messages and waveform data packets, and output the point cloud encoded vector through self-attention encoding. The cross-modal fusion module is connected to the 1D-CNN module, 2D-CNN module and Point-Transformer module respectively. It is used to receive temporal coding vectors, graph coding vectors and point cloud coding vectors, obtain the defect characterization vector by cross-modal attention weighted summation, map it into the probability distribution of various defects by the fully connected classification head, and send the defect probability distribution, fault mechanism explanation information and defect characterization vector as inference results back to the main control board.

7. A partial discharge ultra-high frequency storage and computing integrated method, operating in the partial discharge ultra-high frequency storage and computing integrated device according to any one of claims 1-6, characterized in that, include: The sampled data stream of the UHF partial discharge signal is pulse-triggered in real time, and the triggered UHF partial discharge pulse is separated into feature value messages and waveform data packets. The feature value messages of the UHF partial discharge pulse are extracted through a lossy compression path, and the complete waveform data packets of the UHF partial discharge pulse are output through a lossless data path. The real-time priority of feature value messages is higher than that of waveform data packets; The received waveform data packets are stored sequentially according to the acquisition time, and accumulated to form an online sample dataset. The physical trend characteristics of the online sample dataset are continuously calculated in the first cycle. In the second cycle, the accumulated feature value messages and waveform data packets are forwarded to perform multimodal large model inference to obtain the inference results. The first cycle is shorter than the second cycle; Before the inference results are sent back, preliminary monitoring results are output based on physical trend characteristics. After the inference results are received, the preliminary monitoring results and the inference results are weighted and fused with confidence. The rate of change of physical trend characteristics is used as the dynamic adjustment coefficient of the confidence of the inference results, and the final monitoring results are output.

8. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded onto the processor, it implements the partial discharge ultra-high frequency memory-computer integrated method as described in claim 7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the UHF partial discharge storage and computing method as described in claim 7.

Images