A partial discharge monitoring device and method using multi-modal sensing

Abstract

The application provides a partial discharge monitoring device and method based on multi-modal sensing, wherein a multi-dimensional sensing acquisition module captures signals generated by partial discharge from different physical dimensions; a signal conditioning and data conversion module performs time division multiplexing on the signals and converts the signals into multi-dimensional sensing acquisition data; a local processing module extracts a high-dimensional feature vector from the multi-dimensional sensing acquisition data through a multi-dimensional data model, and performs preliminary partial discharge diagnosis based on the high-dimensional feature vector and a global model obtained from a cloud federated learning module; the cloud federated learning module adjusts and updates the weight of the global model of the terminal device where the local processing module is located according to the high-dimensional feature vector of the different local processing modules, performs global model weighted aggregation on the high-dimensional feature vector uploaded by each local processing module according to the weight of the terminal device where the local processing module is located in the global model aggregation, and obtains an updated global model, thereby improving the reliability and adaptability of partial discharge identification.

Description

Technical Field

[0001] This application relates to the field of partial discharge monitoring technology, and in particular to a multimodal sensing partial discharge monitoring device and method. Background Technology

[0002] Partial discharge is a significant indicator of insulation degradation in power equipment, and its early detection and accurate diagnosis are crucial for ensuring the safe and stable operation of power systems. Traditional partial discharge monitoring methods mainly include pulsed current methods, ultrasonic methods, ultra-high frequency methods, photoelectric methods, and infrared thermography. Each of these methods has its advantages and disadvantages and is typically used independently, making it difficult to comprehensively capture the complex characteristics of partial discharge. For example, pulsed current methods are susceptible to electromagnetic interference in the field; ultrasonic methods are limited by the sound propagation medium and obstacles; ultra-high frequency methods are suitable for gas-insulated switchgear and transformers, but have high requirements for sensor installation location and sensitivity; photoelectric methods and infrared thermography are mainly used for auxiliary judgment of surface discharge and thermal faults.

[0003] Existing technologies for partial discharge monitoring face the following challenges: 1. Limited data dimensions: Most monitoring systems rely on single or limited sensing technologies, failing to acquire comprehensive information on partial discharge under multiple physical fields such as electricity, sound, light, and heat. This results in incomplete feature representation and difficulty in accurately distinguishing different types of partial discharge or assessing their development trends. 2. Poor data synchronization: When collecting data from multiple sensors, differences in sampling rates, triggering mechanisms, and transmission paths make it difficult to achieve strict spatiotemporal synchronization, limiting the depth and effectiveness of multi-source data fusion. 3. Data silos and privacy issues: With the development of smart grids, a large amount of monitoring data is scattered across various terminal devices, forming data silos. For example, patent CN119622461A, "Deep Learning-Based Partial Discharge Pattern Recognition Method for GIS Equipment," discloses a deep learning-based partial discharge pattern recognition method for GIS equipment, involving the field of GIS partial discharge recognition technology, including: using ultra-high frequency... Sensors collect partial discharge signals from GIS equipment and perform signal conditioning. The conditioned partial discharge signals are then digitally sampled and processed to generate a phase-resolved pulse sequence (PRPS) map. Intelligent noise reduction is applied to the PRPS map, which contains three-dimensional information: phase, amplitude, and frequency. The PRPS map is input into a pre-defined neural network model to extract map features. Based on these extracted features, the partial discharge type is identified and classified, and the identification results are output. This method enables automatic acquisition, intelligent processing, and accurate identification of partial discharge signals. However, the data dimension is limited, failing to capture comprehensive information about partial discharge under multiple physical fields such as electricity, sound, light, and heat. This results in incomplete feature representation, making it difficult to accurately distinguish different types of partial discharge or assess their development trends. Furthermore, relying solely on local terminal devices for neural network model identification prevents adjustments to the neural network model based on the real-time conditions of each terminal device, leading to low reliability and adaptability in partial discharge identification.

[0004] To address this problem, the present invention provides a multimodal sensing partial discharge monitoring device and method to solve the aforementioned issues. Summary of the Invention

[0005] In order to solve the problems existing in the prior art, this invention innovatively proposes a multimodal sensing partial discharge monitoring device and method, which effectively solves the problem of low reliability and adaptability of partial discharge identification caused by the prior art, and effectively improves the reliability and adaptability of partial discharge identification.

[0006] The first aspect of this invention provides a multimodal sensing partial discharge monitoring device, comprising: a multidimensional sensing acquisition module, a signal conditioning and data conversion module, a local processing module, and a cloud-based federated learning module. The multidimensional sensing acquisition module is used to capture signals generated by partial discharge from different physical dimensions. The signal conditioning and data conversion module is used to acquire and condition the signals generated by partial discharge captured by the multidimensional sensing acquisition module, and to perform time-division multiplexing on the signals generated by partial discharge captured from different physical dimensions, converting them into multidimensional sensing acquisition data. The local processing module is used to establish a multidimensional data model, and through the multidimensional data model, to extract high-dimensional feature vectors from the multidimensional sensing acquisition data, and based on... A preliminary partial discharge diagnosis is performed using high-dimensional feature vectors and a global model obtained from a cloud-based federated learning module, and the preliminary partial discharge diagnosis results are sent to the cloud-based federated learning module. The high-dimensional feature vectors include time-domain features, frequency-domain features, spatiotemporal correlation features, and statistical features. The cloud-based federated learning module receives the high-dimensional feature vectors and the partial discharge diagnosis results, adjusts and updates the weights of the terminal devices where the local processing modules are located in the global model based on the high-dimensional feature vectors of different local processing modules, and performs weighted aggregation of the high-dimensional feature vectors uploaded by each local processing module in the global model based on the weights of the terminal devices where the local processing modules are located in the global model, resulting in an updated global model.

[0007] A second aspect of the present invention provides a multimodal sensing method for monitoring partial discharge, implemented based on the multimodal sensing partial discharge monitoring device described in the first aspect of the present invention, comprising:

[0008] The multidimensional sensing acquisition module captures signals generated by partial discharge from different physical dimensions; the signal conditioning and data conversion module is used to acquire and condition the signals generated by partial discharge captured by the multidimensional sensing acquisition module, and to perform time-division multiplexing on the signals generated by partial discharge captured from different physical dimensions, converting them into multidimensional sensing acquisition data.

[0009] The local processing module establishes a multidimensional data model, extracts high-dimensional feature vectors from the multidimensional sensor data, and performs preliminary partial discharge diagnosis based on the high-dimensional feature vectors and the global model obtained from the cloud federated learning module. The preliminary partial discharge diagnosis results are then sent to the cloud federated learning module. The high-dimensional feature vectors include time-domain features, frequency-domain features, spatiotemporal correlation features, and statistical features.

[0010] The cloud-based federated learning module is used to receive high-dimensional feature vectors and partial discharge diagnosis results. It adjusts and updates the weights of the terminal devices where the local processing modules are located in the global model based on the high-dimensional feature vectors of different local processing modules. It also performs global model weighted aggregation on the high-dimensional feature vectors uploaded by each local processing module based on the weights of the terminal devices where the local processing modules are located in the global model, and obtains the updated global model.

[0011] The local processing module retrieves the updated global model from the cloud-based federated learning module.

[0012] The technical solution adopted in this invention has the following technical effects:

[0013] 1. In the technical solution of this invention, the multi-dimensional sensing acquisition module is used to capture signals generated by partial discharge from different physical dimensions; the signal conditioning and data conversion module is used to time-division multiplex the signals generated by partial discharge captured from different physical dimensions and convert them into multi-dimensional sensing acquisition data; the local processing module is used to establish a multi-dimensional data model, extract high-dimensional feature vectors from the multi-dimensional sensing acquisition data through the multi-dimensional data model, and perform preliminary partial discharge diagnosis based on the high-dimensional feature vectors and the global model obtained from the cloud federated learning module, and send the preliminary partial discharge diagnosis results to the cloud federated learning module; the cloud federated learning module is used to receive the high-dimensional feature vectors and the partial discharge diagnosis results, adjust and update the weights of the terminal devices where the local processing modules are located in the global model according to the high-dimensional feature vectors of different local processing modules, and perform global model weighted aggregation on the high-dimensional feature vectors uploaded by each local processing module according to the weights of the terminal devices where the local processing modules are located in the global model to obtain the updated global model, effectively solving the problem of low reliability and adaptability of partial discharge identification caused by existing technologies, and effectively improving the reliability and adaptability of partial discharge identification.

[0014] 2. The technical solution of the present invention also includes a self-calibration module, which has a built-in programmable pulse generator and digital signal processor. It can automatically compensate for the sensitivity drift and gain attenuation of the sensor link caused by environmental changes or component aging, significantly improve the measurement accuracy and reliability of the device during long-term operation, reduce the frequency of manual calibration and maintenance costs, and ensure the continuous accuracy of monitoring data.

[0015] 3. In the technical solution of this invention, the multi-dimensional data model is a four-dimensional data model. By synchronously collecting multi-dimensional sensor data, the correlation between partial discharge events in three-dimensional space and time is established as a spatiotemporal correlation feature in the four-dimensional data model. This solves the problem of spatiotemporal synchronization of multi-sensor data acquisition. This ensures that the responses of different sensors to the same partial discharge event at the same time can be accurately aligned, which greatly improves the effectiveness of data fusion and the accuracy of subsequent analysis, laying the foundation for the accurate location and type identification of partial discharge.

[0016] 4. This invention innovatively introduces a feature-layer federated learning mechanism, completing the preprocessing of raw data and high-dimensional feature extraction on local devices. Only the high-dimensional feature vectors are uploaded to the cloud for aggregation and model training. This approach avoids the direct transmission of raw sensitive data, effectively protecting the privacy and security of power equipment operation data. Furthermore, compared to transmitting raw data, the amount of feature vector data is significantly reduced, substantially lowering communication bandwidth requirements and transmission overhead.

[0017] 5. In the technical solution of this invention, the cloud-based federated learning module dynamically adjusts the weight of the terminal device where each local processing module is located in the global model aggregation based on the quality, amount of data, and contribution to the global model of the high-dimensional feature vectors uploaded by each local processing module's terminal device. This means that clients with larger data volumes and better local model performance contribute more to the global model, making the global model more robust and more adaptable to non-independent identically distributed (Non-IID) data. This significantly improves the generalization ability and diagnostic accuracy of the partial discharge diagnosis model under different power equipment and operating environments. The local processing module has edge intelligence capabilities, enabling preliminary partial discharge diagnosis and early warning on the terminal device. This allows the device to respond to abnormal situations in real time without waiting for data to be uploaded to the cloud for processing, greatly shortening the early warning response time and improving the timeliness of fault handling.

[0018] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0019] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0020] Figure 1 This is a schematic diagram of the overall architecture of the device in Embodiment 1 of the present invention;

[0021] Figure 2 This is a schematic diagram of the hardware architecture of the partial discharge monitoring terminal in the device of Embodiment 1 of the present invention;

[0022] Figure 3 This is a schematic diagram of the broadband signal conditioning circuit in the device of Embodiment 1 of the present invention;

[0023] Figure 4 This is a schematic diagram of the calibration process of the self-calibration module in the device of Embodiment 1 of the present invention;

[0024] Figure 5 This is a schematic diagram of the software architecture of the device in Embodiment 1 of the present invention;

[0025] Figure 6 This is a schematic diagram of the federated learning process in the cloud-based federated learning module of the device in Embodiment 1 of the present invention;

[0026] Figure 7 This is a flowchart illustrating the method of Embodiment 2 in the present invention. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0028] Example 1

[0029] like Figure 1 As shown, this invention provides a multimodal sensing partial discharge monitoring device, comprising: a multi-dimensional sensing acquisition module (multiplexed four-dimensional sensing acquisition module), a signal conditioning and data conversion module, a local processing module (local processing unit), and a cloud-based federated learning module (cloud-based federated learning client layer). The multi-dimensional sensing acquisition module is used to capture signals generated by partial discharge from different physical dimensions; the signal conditioning and data conversion module is used to acquire and condition the signals generated by partial discharge captured by the multi-dimensional sensing acquisition module, and to perform time-division multiplexing on the signals generated by partial discharge captured from different physical dimensions, converting them into multi-dimensional sensing acquisition data; the local processing module is used to establish a multi-dimensional data model, and to process the multi-dimensional sensing acquisition data through the multi-dimensional data model. The system extracts high-dimensional feature vectors and performs preliminary partial discharge diagnosis based on these vectors and the global model obtained from the cloud-based federated learning module. The preliminary partial discharge diagnosis results are then sent to the cloud-based federated learning module. The high-dimensional feature vectors include time-domain features, frequency-domain features, spatiotemporal correlation features, and statistical features. The cloud-based federated learning module receives the high-dimensional feature vectors and the partial discharge diagnosis results. It adjusts and updates the weights of the terminal devices where the local processing modules reside in the global model based on the high-dimensional feature vectors of different local processing modules. Based on the weights of the terminal devices where the local processing modules reside in the global model, it performs weighted aggregation of the high-dimensional feature vectors uploaded by each local processing module to obtain the updated global model.

[0030] The overall system architecture of this device is as follows: Figure 1As shown, it mainly consists of front-end partial discharge monitoring terminals (edge ​​devices) and back-end cloud-based federated learning centers (cloud-based federated learning client layers). Multiple monitoring terminals are distributed and deployed at the power equipment site, responsible for data acquisition, preprocessing, and local feature extraction; the cloud-based federated learning center is responsible for receiving high-dimensional feature vectors uploaded by each monitoring terminal, performing feature aggregation and global model training, and then distributing the updated global model to the monitoring terminals.

[0031] The hardware architecture of the partial discharge monitoring terminal mainly includes a multiplexed four-dimensional sensing acquisition module (multi-dimensional sensing acquisition module), a signal conditioning and data conversion module, a self-calibration module, a local processing unit, and a communication module. Its detailed hardware architecture block diagram is shown below. Figure 2 As shown.

[0032] The multiplexed four-dimensional sensing acquisition module comprises four types of sensors: a UHF sensor, an ultrasonic array, a photoelectric sensor, and an infrared thermal imager, forming a multi-dimensional sensing acquisition module. These sensors can capture signals generated by partial discharge from different physical dimensions, including electromagnetic waves, sound waves, light radiation, and thermal effects, comprehensively sensing partial discharge signals from four dimensions: electrical, acoustic, optical, and thermal. To achieve efficient and synchronous data acquisition, different sensors share a signal conditioning circuit (operating frequency range 1MHz-3GHz), and time-division multiplexing technology is used to switch sensor channels for real-time acquisition. This multiplexing mechanism not only reduces hardware costs and system complexity but also ensures the synchronization of data from different sensors in the time dimension, thereby achieving synchronous acquisition of spatial-temporal four-dimensional data (three-dimensional spatial coordinates + time).

[0033] UHF sensor: Employs a broadband antenna design covering the 300MHz-3GHz frequency band to capture electromagnetic wave signals generated by partial discharge. Its output signal is pre-amplified by a low-noise amplifier (LNA).

[0034] Ultrasonic array: Composed of multiple piezoelectric ultrasonic transducers, precisely arranged to achieve three-dimensional localization of the partial discharge source. The acoustic signal received by each transducer is amplified and filtered.

[0035] Photoelectric sensor: Employs a high-sensitivity photodiode to detect transient optical radiation generated by partial discharge, particularly ultraviolet and visible light. Its output is a pulse signal, reflecting the flash frequency and pulse width of the discharge.

[0036] Infrared thermal imager: Employs an uncooled focal plane array detector to acquire real-time images of the temperature distribution on the surface of the equipment, used to identify hot spots and calculate temperature rise gradients.

[0037] The signal conditioning and data conversion module is crucial for achieving multi-sensor data multiplexing and synchronization. It includes:

[0038] Wideband signal conditioning circuit (including wideband programmable amplifier circuit and adjustable filter circuit): Designed for the characteristics of UHF, ultrasonic, and photoelectric signals, it uses a shared wideband programmable amplifier circuit and adjustable filter circuit to ensure good linearity and low noise characteristics within a wide frequency band of 1MHz-3GHz. Its working principle can be expressed as follows:

[0039] ;

[0040] Where H(f) represents the system's frequency response function, f0 is the cutoff frequency, G is the gain, and f is the frequency. In practical circuits, a multi-stage amplification and filtering structure is used, and the transfer function is the product of the transfer functions of each stage.

[0041] The broadband signal conditioning circuit is composed of a multi-stage cascaded topology, specifically as follows: Figure 3 As shown,

[0042] The input protection circuit consists of a bidirectional TVS diode, a resistor, and a ferrite bead, providing overcurrent and ESD protection, and impedance matching up to 50Ω.

[0043] The broadband low-noise amplifier circuit uses a chip with a gain-bandwidth product of 0.5-30GHz, a gain of 15dB, and a noise figure of 1.4dB to form a basic operational amplifier circuit. The signal line is connected to the ADL90XX chip (low-noise amplifier chip) through a series inductor and a 50Ω impedance matching. The output of the low-noise amplifier is then output with a 50Ω impedance matching.

[0044] The programmable gain amplifier circuit uses the HMCXX digital attenuator. The output signal of the low-noise amplifier is input to the input terminal of the digital attenuator through a series DC blocking capacitor. The input terminal of the local FPGA (local processing module or local processing unit) communicates and controls the frequency band analysis through the SPI interface.

[0045] The adjustable bandpass filter circuit consists of a high-speed switch and a bandpass filter circuit.

[0046] Specifically, the high-speed switch uses a dual-channel or higher switch, and the local FPGA selects which channel to turn on based on feedback from the programmable gain amplifier circuit.

[0047] The bandpass filter circuit consists of two frequency bands: 1-50MHz and 50MHz-3GHz. The 1-50MHz band uses a 7th-order elliptic low-pass filter (cutoff frequency 50MHz) or a Butterworth filter, composed of passive LC filters; the 50MHz-3GHz bandpass filter is formed by cascading a high-pass filter (cutoff frequency 50MHz) and a low-pass filter (cutoff frequency 3GHz).

[0048] The gain buffer driver has a high-speed switching dual-channel output signal with 50Ω impedance matching at the input. It is connected to the broadband amplifier THSXX chip (requiring high bandwidth and low distortion) to form a gain buffer circuit.

[0049] In addition to the broadband signal conditioning circuit, the signal conditioning and data conversion module also includes a time-division multiplexing switching module. This module controls the access of different sensor channels via a high-speed switching array. Within each sampling period, signals from UHF, ultrasonic, photoelectric, and infrared sensors are sequentially acquired according to a preset timing sequence. This mechanism ensures that the system can acquire instantaneous data from different sensors in turn within the same time window, thus logically achieving synchronous acquisition of four-dimensional data. At any given time, only one channel is active.

[0050] The local processing module (local processing unit) typically uses a high-performance embedded processor (such as an ARM Cortex-M series microcontroller or FPGA) and is responsible for the following tasks:

[0051] Data preprocessing: Performing operations such as denoising, filtering, and baseline correction on the raw digital signals acquired by the ADC.

[0052] Feature extraction: Run a pre-trained feature extraction algorithm to extract high-dimensional feature vectors from the pre-processed multi-dimensional data, including time-domain features (peak value, rising edge, pulse width), frequency-domain features (spectral entropy, main frequency, bandwidth), spatiotemporal correlation features, and statistical features.

[0053] Feature extraction formula (taking CNN as an example): ;

[0054] in, This represents the convolution operation. and The first The layer's weights and biases, where σ is the activation function.

[0055] Local model inference: Run the global model downloaded from the cloud federated learning module to perform preliminary partial discharge diagnosis and early warning locally, realizing edge intelligence.

[0056] Preferably, it further includes: a self-calibration module, which includes a programmable pulse generator and a digital signal processor running a closed-loop calibration system. The programmable pulse generator is used to automatically trigger the pulse generator to generate simulated partial discharge signals at a preset time interval or according to system instructions. The closed-loop calibration system of the digital signal processor is used to synchronously acquire the response signals of each sensor in the multi-dimensional sensing acquisition module to the simulated partial discharge signals; compare the response signals with the reference template in the ideal state or during factory calibration, calculate the sensitivity offset; and adjust and compensate the conditioning parameters in the signal conditioning circuit according to the calculated offset.

[0057] Built-in reference source: The terminal integrates a high-precision, programmable pulse generator capable of producing analog partial discharge signals with nanosecond-level rising edges. The output signal of this reference source is precisely injected into each sensor channel through an injection network to simulate real partial discharge signals as a calibration benchmark.

[0058] like Figure 4 As shown, the closed-loop calibration process is as follows:

[0059] Step 1: Periodically trigger the reference source to acquire the response signals of each sensor. At preset time intervals or according to system commands, automatically trigger the built-in pulse generator and simultaneously acquire the responses of the UHF, ultrasonic, photoelectric, and infrared sensors to the analog signal. These response signals contain the current characteristics of the sensor chain (including the sensors themselves, cables, conditioning circuits, etc.).

[0060] Step 2: Compare the response signal with the reference template to calculate the sensitivity offset. The acquired response signal is compared with a pre-stored reference template obtained under ideal conditions (or during factory calibration). Using signal processing algorithms, the sensitivity drift, gain attenuation, phase delay, and other offsets of the current sensor link are calculated. The sensitivity offset can be represented using the MSE function or the SD function (standard deviation).

[0061] Step 3: Automatically adjust amplifier gain / filter parameters to compensate for drift errors. Based on the calculated offset, the device's control unit automatically adjusts the programmable amplifier gain and tunable filter parameters in the signal conditioning circuit. For example, if a decrease in sensitivity is detected, the amplifier gain is increased; if a change in frequency response is detected, the filter parameters are adjusted. This process forms a closed-loop control, ensuring that the sensor link always operates in optimal condition and effectively compensating for measurement errors caused by environmental changes, component aging, and other factors.

[0062] Compensation can be provided in the following ways:

[0063] 1. Adjust the PGA gain (compensate for amplitude across the entire frequency band) to the target gain (current gain plus...). However, it cannot exceed the PGA's allowable range.

[0064] 2. Adjust the parameters of the adjustable filter (such as center frequency, bandwidth, etc.) to compensate for changes in frequency response. Adjust to a new center frequency and bandwidth, but do not exceed the adjustable calibration range.

[0065] Preferably, the system also includes a communication module. This module establishes a communication connection between the local processing module and the cloud-based federated learning module, encrypts and transmits the high-dimensional feature vectors extracted by the local processing module and the partial discharge diagnostic results, and receives global model parameters from the cloud-based federated learning module. To ensure data security and privacy, encrypted communication protocols (such as TLS / SSL) and secure transmission channels (such as 4G / 5G cellular networks, LoRaWAN, or industrial Ethernet) are used. This module primarily transmits the locally extracted high-dimensional feature vectors and receives global model updates.

[0066] The software architecture of this device adopts a layered design, mainly including system initialization, data acquisition layer, self-calibration, data preprocessing layer, feature extraction layer, local inference layer, communication management layer, and federated learning client layer. Its software architecture diagram is shown below. Figure 5 As shown.

[0067] System initialization: After the device is powered on, it performs hardware self-test, sensor configuration, and communication module initialization.

[0068] Data Acquisition Layer: Following a pre-defined time-division multiplexing strategy, it polls and acquires raw data from UHF, ultrasonic, photoelectric, and infrared sensors. It is responsible for controlling the sensors and ADC to achieve synchronous acquisition and buffering of raw data.

[0069] Self-calibration: The self-calibration process is triggered periodically or as needed, adjusting sensor link parameters through a built-in reference source and closed-loop calibration algorithm to ensure measurement accuracy.

[0070] Data preprocessing layer: Performs noise reduction, filtering, baseline correction and other operations on the raw data to improve data quality.

[0071] Feature extraction layer: This layer constructs a four-dimensional data model and extracts the amplitude, phase, frequency domain, and spatiotemporal correlation features of partial discharge. This part can integrate machine learning models, such as convolutional neural networks (CNNs) for temporal and frequency domain feature extraction, or graph neural networks (GNNs) for spatiotemporal correlation feature learning.

[0072] Local Inference Layer: Uses the latest global model stored locally to infer the extracted features, determine whether partial discharge exists and its type and severity, and generate local diagnostic results and early warning information.

[0073] Communication Management Layer: Responsible for managing communication links with the cloud-based federated learning center, data encryption and decryption, feature vector uploading, and model parameter downloading.

[0074] Federated learning client layer: coordinates local feature extraction, local model updates and feature vector uploads, participates in the aggregation process of federated learning, and displays local diagnostic results and early warning information through human-computer interface or remote monitoring platform.

[0075] Specifically, the multidimensional data model in the local processing module is a four-dimensional data model. This model integrates features from multiple dimensions, including amplitude, phase, frequency domain, and spatiotemporal correlation, significantly improving the representation capability of partial discharge characteristics. Specific feature quantities are shown in Table 1 below:

[0076] Table 1: Sensor Characteristic Table

[0077]

[0078] Specifically, the extraction of high-dimensional feature vectors from multi-dimensional sensor data involves:

[0079] S101 maps the information (amplitude and phase information) of the UHF sensor and photoelectric sensor to the PRPD / PRPS spectrum respectively, as time-domain features in the four-dimensional data model;

[0080] Amplitude-Phase (PRPD / PRPS): Traditional partial discharge characteristic analysis methods identify different partial discharge types by statistically analyzing the relationship between the amplitude of the discharge pulse and the phase of the power frequency voltage (Phase Resolved Partial Discharge, PRPD) or the number of pulses and the phase of the power frequency voltage (Phase Resolved Partial Sum, PRPS). This invention maps the amplitude and pulse information of UHF and photoelectric sensors onto the PRPD / PRPS spectrum, using it as the time-domain feature of a four-dimensional data model.

[0081] PRPD: The phase of the power frequency voltage is used as the horizontal axis and the amplitude of the discharge pulse is used as the vertical axis. Each point represents a discharge pulse, and a two-dimensional or three-dimensional spectrum is formed through statistics. PRPS: In addition to phase and amplitude, time series (usually continuous power frequency cycles) are also considered to form a three-dimensional spectrum (phase, cycle number, amplitude) or the amplitude is represented by color.

[0082] The data from UHF and photoelectric sensors are mapped to PRPD / PRPS spectra as follows:

[0083] (1) Synchronous acquisition: Ensure that the signal acquisition of all sensors is synchronized with the phase of the power frequency voltage. Use the signal of the voltage transformer (PT) as the power frequency phase reference, and generate the phase signal through the phase-locked loop (PLL).

[0084] (2) Pulse detection: Extract the discharge pulse from the UHF and photoelectric sensor signals. For the UHF signal, since it is a high-frequency signal, we obtain the pulse signal by envelope detection after bandpass filtering. The photoelectric sensor directly outputs the optical pulse signal.

[0085] (3) Pulse feature extraction: For each detected pulse, record the following information:

[0086] Pulse amplitude: The peak voltage (or relative value) of the pulse.

[0087] Pulse phase: The phase angle of the power frequency voltage corresponding to the pulse (0°~360°).

[0088] Pulse timestamp: The precise time of pulse occurrence (used for cycle number calculation in PRPS).

[0089] (4) Constructing the PRPD map:

[0090] The power frequency cycle is divided into several phase windows (360 windows, one per degree).

[0091] For each pulse from each sensor, based on its phase, its amplitude is accumulated into the corresponding phase window (forming a two-dimensional scatter plot, with phase on the horizontal axis and amplitude on the vertical axis), or the number of pulses and amplitude distribution within each window are statistically analyzed (common statistical quantities such as maximum amplitude, average amplitude, etc.).

[0092] (5) Constructing the PRPS map:

[0093] It is necessary to record pulse information for multiple consecutive power frequency cycles (e.g., 256 cycles).

[0094] The X-axis of the graph represents the phase (0°~360°), the Y-axis represents the sequence number of consecutive power frequency cycles, and the Z-axis (or color) represents the position of the phase and the pulse amplitude (or number) within the cycle.

[0095] (6) Multi-sensor data fusion: Since there are multiple sensors, you can choose to: plot separately: plot PRPD / PRPS maps for UHF and photoelectric separately. Fusion plot: display the pulses of the two sensors in the same coordinate system and use different colors to distinguish the sensor sources.

[0096] S102 performs Fourier transform on the signals from the UHF sensor and the ultrasonic array respectively to extract spectral features, which are used as frequency domain features in the four-dimensional data model. The spectral features include the main frequency, bandwidth, and spectral entropy. Spectral entropy can measure the complexity and randomness of a signal and is of great significance for distinguishing different types of partial discharges.

[0097] S103 establishes the correlation between partial discharge events in three-dimensional space and time dimensions by synchronously collecting multi-dimensional sensor data, which serves as the spatiotemporal correlation feature in the four-dimensional data model;

[0098] Specifically, the spatiotemporal correlation characteristics are a key innovation of the four-dimensional data model in this invention. By simultaneously acquiring data from four types of sensors, the correlation between partial discharge events in three-dimensional space (determined by ultrasonic TDOA and infrared hotspot coordinates) and time can be established. For example, a partial discharge event may simultaneously manifest as a high-frequency pulse in a UHF signal, as a sound wave at a specific location in an ultrasonic signal, as a momentary flash in a photoelectric signal, and cause a localized temperature rise in an infrared image. By analyzing the synchronicity and correlation of these multi-source heterogeneous data in time and space, the source of the partial discharge can be located more accurately, and its development trend can be assessed. Step S103 specifically includes:

[0099] S1031, through an array of multiple ultrasonic sensors, calculate the time difference of partial discharge sound waves arriving at each sensor, establish a hyperboloid equation system, solve the equation system, and obtain the coordinates of the partial discharge source of the ultrasonic positioning point. The coordinates of each ultrasonic array represent the physical position coordinates of the ultrasonic sensor in space.

[0100] The three-dimensional spatial position calibration method is as follows:

[0101] (1) Ultrasonic TDOA localization: Using an array of at least four ultrasonic sensors, the time difference (TDOA) of partial discharge sound waves arriving at each sensor is calculated, and a hyperboloid equation system is established:

[0102]

[0103] Solving this system of equations yields the partial discharge source coordinates (x, y, z) of the ultrasonic positioning point (where c is the speed of sound). The time difference between ultrasonic sensor 1 and ultrasonic sensor 2. (Time difference between ultrasonic sensor 1 and ultrasonic sensor 3), coordinates of the four ultrasonic arrays (x) a , y a , z a ) represents the physical position coordinates of the a-th ultrasonic sensor in space (a = 1, 2, 3, 4).

[0104] S1032, Infrared thermal imager detects the surface temperature distribution of equipment, locates hotspot areas through gradient segmentation, and projects them onto a 3D equipment model to obtain the infrared hotspot coordinates. , , );

[0105] S1033, when the Euclidean distance between the ultrasonic positioning point and the infrared hotspot is not greater than the preset distance threshold, it is determined to be the same physical event;

[0106] Specific spatial association rules: When the Euclidean distance between the ultrasonic positioning point and the infrared hotspot... They were determined to be the same physical event.

[0107] S1034 acquires the current time in real time and synchronizes the time of all sensors in the multi-dimensional sensing acquisition module, sharing a unified time and taking turns to connect the sensor channels; all data points carry a timestamp assigned by a unified clock source, realizing the alignment of the four-channel data on the time axis;

[0108] Time base establishment:

[0109] (1) The current time is obtained in real time through the high-precision clock chip of the local FPGA unit, and all sensors are synchronized to use the same time. Time jitter is required. <1ns.

[0110] (2) Sensor channels are switched on alternately under clock control via a high-speed electronic switch (multiplexer). Assume the number of channels is n1=4, and the switching cycle is... Then the sampling time window for each channel is:

[0111] Absolute timestamp of each sampling point Determined by the following formula:

[0112] in, Absolute time of data collection start. Sampling point number ( =1,2...N q，1 ), N q，1 This represents the total number of sampling point numbers. ADC sampling interval Sensor channel number ( =1,2,3,4).

[0113] (3) All data points carry a nanosecond-level timestamp assigned by a unified clock source. This achieves strict alignment of the four-channel data on the time axis.

[0114] S1035, compensate and calibrate the acquisition time of different physical signal propagation to obtain the compensation time after compensation and calibration of different sensor signal propagation in different multi-dimensional sensing acquisition modules;

[0115] Time delay compensation calibration:

[0116] The difference in the propagation speed of different physical signals leads to different actual signal arrival times, which needs to be compensated for.

[0117] Because different signals have different time delays, for example, the delay time 'd' of electromagnetic waves is the distance from the power source to the sensor, and 'c' is the velocity of the medium, i.e., air. Therefore, the formula is the compensation time, so the final actual acquisition time is the acquisition timestamp minus the compensation time. The medium for sound waves is oil; light signals are actually air, with no medium. The next formula is the standard formula for the thermal diffusion time constant. The specific delay compensation formulas for different signals are shown in Table 2 below.

[0118] Table 2: Formulas for Compensation of Delay for Different Signals

[0119]

[0120] Actual data collection time after compensation: ,

[0121] UHF, US The four signals IR represent four types of sensors: UHF sensor, ultrasonic array, photoelectric sensor, and infrared thermal imager, respectively. This is the timestamp for data acquisition from the UHF sensor. For delay compensation of UHF sensors, This is the acquisition timestamp for the ultrasound array. For delay compensation of ultrasonic array sensors, This is the timestamp for data collection by the photoelectric sensor. This is the timestamp for data acquisition by the infrared thermal imager. For delay compensation of infrared thermal imagers (same) ).

[0122] S1036, based on the timestamps of the data collected by different sensors in the multidimensional sensing acquisition module and the compensation duration after compensation, calculate the covariance, mutual information and maximum time delay mutual information between any two sensors in the multidimensional sensing acquisition module.

[0123] Specifically, S1036 includes:

[0124] Based on the timestamps of data collected by different sensors and the compensation duration, the actual time t of the data collected by different sensors is determined, where the actual time t is the timestamp T. q The difference between the compensated duration and the compensated duration;

[0125] The covariance, mutual information, and time delay mutual information between any two sensors in the multi-dimensional sensing acquisition module are calculated based on the actual time of data acquisition from different sensors. The method for calculating the covariance is as follows:

[0126] ,

[0127] in, Let U be the covariance between sensor u and sensor v, where U = {u1, u2, ..., u}. D} and V={v1,v2...v D These are two continuous sensor signals from different acquisition time points during time alignment. For the signal acquired by a certain sensor at the e-th acquisition time point during time alignment, For the signal acquired by another sensor at the e-th acquisition time point during time alignment, This represents the average of all signals from a single sensor at different acquisition time points during time alignment. This represents the average of all signals from another sensor at different acquisition time points during time alignment, where D is the total number of acquisition time points; it is necessary to ensure... and The sampled values ​​of e at the same time point correspond strictly. Covariance mainly quantifies the strength of the linear correlation between two signals in the time dimension, for example: (S UHF ,S US )>0 indicates an electromagnetic pulse signal (S) UHF ) and ultrasound signal (S US The amplitude changes have the same trend.

[0128] The specific method for calculating mutual information is as follows:

[0129] ,

[0130] in, For mutual information between sensor u and sensor v, Let be the joint probability of the i-th first event classification interval of the continuous signal U acquired by sensor u and the j-th second event classification interval of the continuous signal V acquired by sensor v. Let the edge probability of the i-th first event classification interval be the continuous signal U collected by sensor u. M1 represents the marginal probability of the j-th second event category interval for the continuous signal V acquired by sensor v, M2 represents the total number of first event category intervals for the continuous signal U acquired by sensor u, and M3 represents the total number of second event category intervals for the continuous signal V acquired by sensor v.

[0131] Mutual information primarily measures the amount of information shared under arbitrary statistical dependencies, for example:

[0132] MI( , )≈max indicates that the optical signal ( ) and electromagnetic wave signals (SUHF Completely synchronized changes,

[0133] MI( , ) <MI( , ) indicates thermal signal ( The correlation between electromagnetic waves and sound waves is higher. ).

[0134] First, amplitude discretization is performed, dividing the continuous signal U into M1 first event classification intervals (bins). This means that the sampled values ​​of sensor u are divided into multiple sampled value ranges according to the event category range, each corresponding to a different event category. For example, taking an infrared thermal imager as an example, the event categories can include high temperature, medium temperature, and low temperature. The sampled values ​​of the infrared thermal imager u are divided (custom division) into three range intervals, corresponding to high temperature, medium temperature, and low temperature respectively. Each sample... Mapped to the category range of the first event , .

[0135] The continuous signal V is divided into M2 second event classification intervals (bins). This means that the sampled values ​​of sensor V are divided into multiple sampled value ranges according to the event category range, each corresponding to a different event category. For example, if V is a UHF sensor, the event categories could include high temperature, medium temperature, and low temperature. The sampled values ​​of infrared thermal imager V are divided (custom division) into three range intervals, corresponding to high intensity, medium intensity, and low intensity, respectively. Each sample... Mapping to discrete values , The classification of other types of sensors (existing types or categories) is consistent with the classification principles for infrared thermal imagers and UHF sensors, and this embodiment does not impose any restrictions on them.

[0136] Next, probability distribution estimation is performed:

[0137] Marginal probability: ,

[0138] Joint probability: ,

[0139] Where N is the total number of collected values ​​across all intervals in U. For example, if the temperature is sampled 1000 times, then N is 1000. iNi represents the number of events corresponding to the data collected in the i-th interval of the first event category in U. For example, i=1 represents "low temperature", i=2 represents "medium temperature", i=3 represents "high temperature", etc. For instance, in 1000 temperature samples, if the "high temperature" category (i=3) appears 120 times, then N3=120. It is the estimated probability that the collected value in U belongs to the i-th first event category interval. This represents the number or frequency of sample values ​​belonging to the i-th interval of the first event category among all sampled values. This represents the number or frequency of sample values ​​belonging to the j-th second event category interval among all sampled values; for example, in 1000 observations (1000 sampled values): the temperature "high temperature" (i=3) and the discharge "moderate intensity" (j=2) occur 15 times, then count((u disc =i)∩(v disc =j)) is 15, u disc and v disc represents the discretized random variable, i.e., the corresponding classification level, which can be given as temperature or discharge level, and i and j are the categories of the corresponding parameters.

[0140] The method for calculating the maximum delay mutual information is as follows:

[0141] =arg max ,

[0142] in, The maximum time delay mutual information between the continuous signal U acquired by sensor u and the continuous signal V acquired by sensor v. Let τ be the time delay mutual information between the continuous signal U acquired by sensor u and the continuous signal V acquired by sensor v, and let τ be the time delay between the continuous signal U acquired by sensor u and the continuous signal V acquired by sensor v.

[0143] The method for calculating time delay mutual information is as follows:

[0144] ,

[0145] in, The continuous signal U acquired by sensor u at the actual acquisition time t and the signal U acquired by sensor v at the actual acquisition time t and the time delay are given by the two signals. The mutual information between the continuously acquired signals V.

[0146] To capture the causal relationship of asynchronous signals, a time delay τ is introduced. By scanning τ, the delay time that maximizes mutual information can be found. This value reflects the causal delay in the physical mechanism (such as the time of temperature rise lag discharge).

[0147] Spatiotemporal correlation feature fusion: Feature vector construction (output of each partial discharge event), including the covariance, mutual information, and maximum delay mutual information of six pairs of sensor combinations (any four sensor combinations).

[0148] ;

[0149] in, This refers to the spatiotemporal correlation features after fusion.

[0150] Statistical analysis is performed on time-domain features, frequency-domain features, and spatiotemporal correlation features to obtain statistical feature vectors, which serve as statistical features in the four-dimensional data model. Further analysis of the statistical properties of the aforementioned time-domain features, frequency-domain features, and spatiotemporal correlation features yields features that statistically describe the regularity and distribution characteristics of discharge behavior, such as mean, median, mode, range, variance, and standard deviation.

[0151] The extracted high-dimensional feature vectors are input into the global model for computation to achieve partial discharge type classification (internal discharge / surface discharge / corona discharge), fault severity assessment, and reliability analysis of multi-sensor data fusion.

[0152] The high-dimensional feature vectors extracted by the local processing module are encrypted and uploaded to the cloud-based federated learning center. The high-dimensional feature vectors uploaded by each terminal device where the local processing module is located consist of E types of feature vectors, forming a feature matrix: ,in: The feature matrix uploaded by the terminal device where the k-th local processing module is located. This is the set of temporal feature vectors (dimension dt) uploaded by the terminal device where the k-th local processing module is located. This is the set of frequency domain feature vectors (dimension df) uploaded by the terminal device where the k-th local processing module is located. The set of spatiotemporal feature vectors (dimension ds) uploaded by the terminal device where the k-th local processing module is located, i.e. , This is the set of statistical feature vectors (dimension dstat) uploaded by the terminal device where the k-th local processing module is located.

[0153] The cloud-based federated learning module dynamically adjusts the weight of the terminal device where each local processing module is located in the global model aggregation based on the quality, amount of data, and contribution of the high-dimensional feature vector uploaded by the terminal device where the local processing module is located. It also performs global model weighted aggregation on the high-dimensional feature vector uploaded by each local processing module based on the weight of the terminal device where the local processing module is located in the global model aggregation.

[0154] The software operation process of this device is as follows: Figure 6As shown, the main components include feature uploading, dynamic weight aggregation algorithm, global model update, and early warning display.

[0155] This invention innovates on the federated learning mechanism, proposing and implementing feature-layer federated learning for the first time. Unlike traditional model parameter federation or raw data federation, the core idea of ​​feature-layer federated learning is that the terminal device (partial discharge monitoring device) preprocesses and extracts features from the raw sensor data locally, and then uploads the extracted high-dimensional feature vectors to the central server for aggregation and model training, instead of uploading the raw data or complete model parameters.

[0156] Feature Upload: The high-dimensional feature vectors extracted by the local processing module are encrypted and uploaded to the cloud-based federated learning module. The cloud-based federated learning module dynamically adjusts the weights of the features uploaded by each terminal device in the global model aggregation based on a dynamic weight aggregation algorithm (factors such as the quality and amount of data of each feature uploaded by the terminal device and its contribution to the global model). This solves the problem that simple average weights in traditional federated learning may lead to a decrease in model performance, especially when the data distribution is uneven or the data quality of some devices is poor.

[0157] The core idea of ​​the dynamic weight aggregation algorithm is: for each client participating in federated learning... The weight of the uploaded features in the global model aggregation It is no longer fixed, but dynamically calculated based on the performance of its local model on the validation set, the amount of data, or feature quality evaluation metrics.

[0158] Global model aggregation: The cloud-based federated learning module (cloud aggregator) receives the encryption signatures of the terminal devices where each local processing module is located. After decryption, a weighted aggregation is performed: ,

[0159] Note: Here This refers to the feature vector group uploaded by the terminal device where the k-th local processing module is located. The actual aggregation is performed according to the feature type. For example, the aggregation of the time-domain feature part is as follows: .

[0160] The weights of the local processing module's terminal device in the global model aggregation are dynamically adjusted based on the quality, volume, and contribution of the high-dimensional feature vectors uploaded by each local processing module's terminal device to the global model.

[0161] ,

[0162] in, The total weight of the terminal device where the k-th local processing module is located. The weight of the data volume of the terminal device where the k-th local processing module is located. The data quality weight of the terminal device where the k-th local processing module is located. This represents the performance weight of the terminal device where the k-th local processing module is located.

[0163] in, The calculation method is as follows: ,

[0164] in, This represents the number of data points on the terminal device where the k-th local processing module is located.

[0165] in, The calculation method is as follows: ,

[0166] Where b is the high-dimensional feature vector of the b-th class, and B is the total number of high-dimensional feature vectors. The data quality of the b-th type high-dimensional feature vector uploaded to the terminal device where the k-th local processing module is located; The method for determining it is as follows: ,

[0167] in, The signal-to-noise ratio (the ratio of measured power to noise power, dB unit needs to be linearized) of the high-dimensional feature vector of class b uploaded to the terminal device where the k-th local processing module is located. Upload the high-dimensional feature vector of class b to the terminal device where the k-th local processing module is located. The relative fluctuation (the rate of change of characteristics in adjacent time periods, reflecting stability); M Upload the high-dimensional feature vector of class b to the terminal device where the k-th local processing module is located. With tag r k The mutual information reflects the discriminativeness; α, β, γ are the weighting coefficients; (satisfying α+β+γ=1), usually α=0.5, β=0.3, γ=0.2 (α+β+γ=1, signal-to-noise ratio + stability + discriminativeness).

[0168] in, The calculation method is as follows: ,

[0169] in, It is the AUC value of the classifier trained separately on the validation set using the b-th class of features on the terminal device where the k-th local processing module is located. First, for each local processing module located on terminal device k, calculate the average AUC value corresponding to the terminal device k under the total number of high-dimensional feature vectors B. Then, select the largest average AUC value from the multiple average AUC values ​​corresponding to all local processing modules located on terminal device k.

[0170] It should be noted that the global model training in this embodiment of the invention can adopt a multi-branch structure, with each branch processing one type of feature and finally fusing them: Feature extraction branch: For each type of feature, train a sub-model, and finally perform feature fusion of the sub-models: Concatenate the outputs of each branch and then pass them through the fusion layer.

[0171] Specifically, for the b-th feature, train a sub-model:

[0172] ,

[0173] ,

[0174] Based on the set of time-domain feature vectors After time-domain feature transformation Extracted output vector, It is a time-domain feature transformation Trainable parameters; Based on the set of frequency domain feature vectors After frequency domain feature transformation Extracted output vector, It is a frequency domain feature transformation Trainable parameters; Based on the set of spatiotemporal feature vectors After spatiotemporal feature transformation Extracted output vector, It is a spatiotemporal feature transformation Trainable parameters; Based on the set of statistical feature vectors After statistical feature transformation Extracted output vector, It is a statistical characteristic transformation Trainable parameters.

[0175] Feature fusion layer: The outputs of each branch are concatenated and then fused through the fusion layer. Its specific form is as follows:

[0176] ,in, The output vector after feature fusion layer is used to obtain the fused label. .

[0177] Feature adaptive loss function during training: Weighted cross-entropy (considering client weights):

[0178] ;

[0179] in, The weighted cross-entropy value represents the weight of the b-th feature type. , represents the importance weight of the b-th feature.

[0180] Feature type independent evaluation: in quality weight and performance weight In this process, the quality / performance index of each type of feature is calculated independently, and then averaged.

[0181] Dimension of weight aggregation: final weight It is for the entire client, but the calculation takes into account the comprehensive performance of all feature types of the client.

[0182] Anomaly handling: If the quality of a certain type of feature... If the value is below a threshold (e.g., 0.4), it can be removed from the client's feature group (i.e., set to zero). At the same time, the weights are recalculated.

[0183] In this way, clients with larger datasets and better local model performance will receive higher weights, making the global model more robust and accurate. Preliminary experimental results show that the algorithm can reduce transmission overhead by 8 times while maintaining an accuracy loss of less than 2%.

[0184] Global model update: Receives and feeds back the latest global model parameters from the cloud; Results display and early warning: Local diagnostic results and early warning information are displayed through a human-machine interface or a remote monitoring platform.

[0185] In this invention, a multi-dimensional sensing acquisition module is used to capture signals generated by partial discharge from different physical dimensions; a signal conditioning and data conversion module is used to time-division multiplex the signals generated by partial discharge captured from different physical dimensions, converting them into multi-dimensional sensing acquisition data; a local processing module is used to establish a multi-dimensional data model, extract high-dimensional feature vectors from the multi-dimensional sensing acquisition data through the multi-dimensional data model, and perform preliminary partial discharge diagnosis based on the high-dimensional feature vectors and the global model obtained from the cloud federated learning module, and send the preliminary partial discharge diagnosis results to the cloud federated learning module; the cloud federated learning module is used to receive the high-dimensional feature vectors and the partial discharge diagnosis results, adjust and update the weights of the terminal devices where the local processing modules are located in the global model according to the high-dimensional feature vectors of different local processing modules, and perform global model weighted aggregation on the high-dimensional feature vectors uploaded by each local processing module according to the weights of the terminal devices where the local processing modules are located in the global model, to obtain an updated global model, effectively solving the problem of low reliability and adaptability of partial discharge identification caused by existing technologies, and effectively improving the reliability and adaptability of partial discharge identification.

[0186] The technical solution of this invention also includes a self-calibration module with a built-in programmable pulse generator and digital signal processor, which can automatically compensate for sensitivity drift and gain attenuation caused by environmental changes or component aging in the sensor link. This significantly improves the measurement accuracy and reliability of the device during long-term operation, reduces the frequency of manual calibration and maintenance costs, and ensures the continuous accuracy of monitoring data.

[0187] In this invention, the multi-dimensional data model is a four-dimensional data model. By synchronously acquiring multi-dimensional sensor data, the correlation between partial discharge events in three-dimensional space and time is established as a spatiotemporal correlation feature in the four-dimensional data model. This solves the problem of spatiotemporal synchronization of multi-sensor data acquisition, ensuring that the responses of different sensors to the same partial discharge event at the same time can be accurately aligned. This greatly improves the effectiveness of data fusion and the accuracy of subsequent analysis, laying the foundation for the accurate location and type identification of partial discharge.

[0188] This invention innovatively introduces a feature-layer federated learning mechanism, where preprocessing of raw data and high-dimensional feature extraction are performed on local devices. Only the high-dimensional feature vectors are uploaded to the cloud for aggregation and model training. This approach avoids the direct transmission of raw sensitive data, effectively protecting the privacy and security of power equipment operation data. Furthermore, compared to transmitting raw data, the amount of feature vector data is significantly reduced, substantially lowering communication bandwidth requirements and transmission overhead.

[0189] In this invention, the cloud-based federated learning module dynamically adjusts the weight of each local processing module's terminal device in the global model aggregation based on the quality, volume, and contribution of the high-dimensional feature vectors uploaded by each local processing module's terminal device. This means that clients with larger data volumes and better local model performance contribute more to the global model, making the global model more robust and adaptable to non-independent identically distributed (Non-IID) data. This significantly improves the generalization ability and diagnostic accuracy of the partial discharge diagnosis model under different power equipment and operating environments, with accuracy loss controlled within 2%. The local processing module possesses edge intelligence capabilities, enabling preliminary partial discharge diagnosis and early warning on the terminal device. This allows the device to respond to abnormal situations in real time without waiting for data to be uploaded to the cloud for processing, greatly shortening the early warning response time and improving the timeliness of fault handling.

[0190] The technical solution of this invention adopts a modular design, with each functional module (sensing acquisition, signal conditioning, self-calibration, local processing, and communication) being relatively independent, which facilitates system upgrades, maintenance, and functional expansion. For example, new types of sensors or more advanced feature extraction algorithms can be easily integrated in the future.

[0191] In summary, the technical solution of this invention has made groundbreaking progress in multi-dimensional perception, data synchronization, self-calibration, data privacy protection, communication efficiency, model generalization ability, and edge intelligence. It provides an efficient, safe, and reliable solution for the intelligent monitoring and diagnosis of partial discharge in power equipment, and has important theoretical significance and broad engineering application prospects.

[0192] Example 2

[0193] like Figure 7 As shown, this embodiment of the invention also provides a multimodal sensing partial discharge monitoring method, which is implemented based on a multimodal sensing partial discharge monitoring device in Embodiment 1, and includes:

[0194] S1, the multi-dimensional sensing acquisition module captures signals generated by partial discharge from different physical dimensions; the signal conditioning and data conversion module is used to acquire and condition the signals generated by partial discharge captured by the multi-dimensional sensing acquisition module, and to perform time-division multiplexing on the signals generated by partial discharge captured from different physical dimensions, converting them into multi-dimensional sensing acquisition data.

[0195] S2, the local processing module establishes a multidimensional data model, extracts high-dimensional feature vectors from the multidimensional sensor data through the multidimensional data model, and performs preliminary partial discharge diagnosis based on the high-dimensional feature vectors and the global model obtained from the cloud federated learning module, and sends the preliminary partial discharge diagnosis results to the cloud federated learning module; among them, the high-dimensional feature vectors include time domain features, frequency domain features, spatiotemporal correlation features, and statistical features;

[0196] S3, the cloud federated learning module is used to receive high-dimensional feature vectors and partial discharge diagnosis results. It adjusts and updates the weights of the terminal devices where the local processing modules are located in the global model according to the high-dimensional feature vectors of different local processing modules. It performs global model weighted aggregation on the high-dimensional feature vectors uploaded by each local processing module according to the weights of the terminal devices where the local processing modules are located in the global model, and obtains the updated global model.

[0197] S4 is the updated global model obtained by the local processing module from the cloud-based federated learning module.

[0198] In this invention, a multi-dimensional sensing acquisition module is used to capture signals generated by partial discharge from different physical dimensions; a signal conditioning and data conversion module is used to time-division multiplex the signals generated by partial discharge captured from different physical dimensions, converting them into multi-dimensional sensing acquisition data; a local processing module is used to establish a multi-dimensional data model, extract high-dimensional feature vectors from the multi-dimensional sensing acquisition data through the multi-dimensional data model, and perform preliminary partial discharge diagnosis based on the high-dimensional feature vectors and the global model obtained from the cloud federated learning module, and send the preliminary partial discharge diagnosis results to the cloud federated learning module; the cloud federated learning module is used to receive the high-dimensional feature vectors and the partial discharge diagnosis results, adjust and update the weights of the terminal devices where the local processing modules are located in the global model according to the high-dimensional feature vectors of different local processing modules, and perform global model weighted aggregation on the high-dimensional feature vectors uploaded by each local processing module according to the weights of the terminal devices where the local processing modules are located in the global model, to obtain an updated global model, effectively solving the problem of low reliability and adaptability of partial discharge identification caused by existing technologies, and effectively improving the reliability and adaptability of partial discharge identification.

[0199] The technical solution of this invention also includes a self-calibration module with a built-in programmable pulse generator and digital signal processor, which can automatically compensate for sensitivity drift and gain attenuation caused by environmental changes or component aging in the sensor link. This significantly improves the measurement accuracy and reliability of the device during long-term operation, reduces the frequency of manual calibration and maintenance costs, and ensures the continuous accuracy of monitoring data.

[0200] In this invention, the multi-dimensional data model is a four-dimensional data model. By synchronously acquiring multi-dimensional sensor data, the correlation between partial discharge events in three-dimensional space and time is established as a spatiotemporal correlation feature in the four-dimensional data model. This solves the problem of spatiotemporal synchronization of multi-sensor data acquisition, ensuring that the responses of different sensors to the same partial discharge event at the same time can be accurately aligned. This greatly improves the effectiveness of data fusion and the accuracy of subsequent analysis, laying the foundation for the accurate location and type identification of partial discharge.

[0201] This invention innovatively introduces a feature-layer federated learning mechanism, where preprocessing of raw data and high-dimensional feature extraction are performed on local devices. Only the high-dimensional feature vectors are uploaded to the cloud for aggregation and model training. This approach avoids the direct transmission of raw sensitive data, effectively protecting the privacy and security of power equipment operation data. Furthermore, compared to transmitting raw data, the amount of feature vector data is significantly reduced, substantially lowering communication bandwidth requirements and transmission overhead.

[0202] In this invention, the cloud-based federated learning module dynamically adjusts the weight of each local processing module's terminal device in the global model aggregation based on the quality, volume, and contribution of the high-dimensional feature vectors uploaded by each local processing module's terminal device. This means that clients with larger data volumes and better local model performance contribute more to the global model, making the global model more robust and adaptable to non-independent identically distributed (Non-IID) data. This significantly improves the generalization ability and diagnostic accuracy of the partial discharge diagnosis model under different power equipment and operating environments. The local processing module possesses edge intelligence capabilities, enabling preliminary partial discharge diagnosis and early warning on the terminal device. This allows the device to respond to abnormal situations in real time without waiting for data to be uploaded to the cloud for processing, greatly shortening the early warning response time and improving the timeliness of fault handling.

[0203] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A multimodal sensing partial discharge monitoring device, characterized in that, include: The system comprises a multi-dimensional sensing acquisition module, a signal conditioning and data conversion module, a local processing module, and a cloud-based federated learning module. The multi-dimensional sensing acquisition module includes a UHF sensor, an ultrasonic array, a photoelectric sensor, and an infrared thermal imager, used to capture signals generated by partial discharge from different physical dimensions. The signal conditioning and data conversion module acquires and conditions the signals captured by the multi-dimensional sensing acquisition module, and performs time-division multiplexing on the signals from different physical dimensions to convert them into multi-dimensional sensing acquisition data. The local processing module is used to establish a multi-dimensional data model and, through multi-dimensional... The data model extracts high-dimensional feature vectors from multi-dimensional sensor data and performs preliminary partial discharge diagnosis based on these high-dimensional feature vectors and a global model obtained from the cloud-based federated learning module. The preliminary partial discharge diagnosis results are then sent to the cloud-based federated learning module. The high-dimensional feature vectors include time-domain features, frequency-domain features, spatiotemporal correlation features, and statistical features. Specifically, extracting high-dimensional feature vectors from multi-dimensional sensor data involves establishing the correlation between partial discharge events in three-dimensional space and time dimensions using synchronously acquired multi-dimensional sensor data, serving as the spatiotemporal correlation features in the multi-dimensional data model. By using an array of multiple ultrasonic sensors, the time difference of partial discharge sound waves arriving at each sensor is calculated, a hyperboloid equation system is established, and the equation system is solved to obtain the coordinates of the partial discharge source of the ultrasonic positioning point; an infrared thermal imager detects the surface temperature distribution of the equipment, and the hot spot area is located by gradient segmentation and projected onto a three-dimensional equipment model to obtain the coordinates of the infrared hot spot. When the Euclidean distance between the ultrasonic positioning point and the infrared hotspot is not greater than a preset distance threshold, it is determined to be the same physical event; The system acquires the current time in real time and synchronizes the time of all sensors in the multi-dimensional sensing acquisition module, sharing a unified time and taking turns connecting the sensor channels; all data points carry a timestamp assigned by a unified clock source, achieving alignment of the four-channel data on the time axis; The acquisition time of different physical signal propagation is compensated and calibrated to obtain the compensation time after compensation and calibration of different sensor signal propagation in different multidimensional sensing acquisition modules; Based on the timestamps of the data collected by different sensors and the compensation duration after compensation, the actual time t of the data collected by different sensors is determined, where the actual time t is the difference between the timestamp and the compensation duration after compensation. Calculate the covariance, mutual information, and time delay mutual information between any two sensors in the multidimensional sensing acquisition module based on the actual time of data acquisition from different sensors. The cloud-based federated learning module receives high-dimensional feature vectors and partial discharge diagnostic results, and adjusts and updates the weights of the terminal devices where the local processing modules reside in the global model based on the high-dimensional feature vectors of different local processing modules. Specifically: ， in, The total weight of the terminal device where the k-th local processing module is located. The weight of the data volume of the terminal device where the k-th local processing module is located. The data quality weight of the terminal device where the k-th local processing module is located. This represents the performance weight of the terminal device where the k-th local processing module is located. in, The calculation method is as follows: , in, This represents the number of data points on the terminal device where the k-th local processing module is located. in, The calculation method is as follows: , Where b is the high-dimensional feature vector of the b-th class, and B is the total number of high-dimensional feature vectors. The data quality of the b-th type high-dimensional feature vector uploaded to the terminal device where the k-th local processing module is located; The method for determining it is as follows: , in, The signal-to-noise ratio of the b-th class of high-dimensional feature vectors uploaded to the terminal device where the k-th local processing module is located; Upload the high-dimensional feature vector of class b to the terminal device where the k-th local processing module is located. The relative fluctuations of M; Upload the high-dimensional feature vector of class b to the terminal device where the k-th local processing module is located. With the tag r k Mutual information; α, β, γ are weighting coefficients, respectively; in, The calculation method is as follows: , in, It is the AUC value of the classifier trained separately on the validation set using the b-th class of features on the terminal device where the k-th local processing module is located. First, for each local processing module located on terminal device k, calculate the average AUC value corresponding to the terminal device k under the total number of high-dimensional feature vectors B. Then, select the largest average AUC value from the multiple average AUC values ​​corresponding to all local processing modules located on terminal device k. Based on the weight of the terminal device where the local processing module is located in the global model, the high-dimensional feature vector uploaded by each local processing module is weighted and aggregated in the global model to obtain the updated global model.

2. The partial discharge monitoring device with multimodal sensing according to claim 1, characterized in that, The UHF sensor is used to acquire electromagnetic wave signals generated by partial discharge, the ultrasonic array is used to acquire acoustic wave signals generated by partial discharge, the photoelectric sensor is used to acquire instantaneous light radiation generated by partial discharge, and the infrared thermal imager is used to acquire temperature distribution images generated by partial discharge.

3. The partial discharge monitoring device with multimodal sensing according to claim 1, characterized in that, It also includes a self-calibration module, which comprises a programmable pulse generator and a digital signal processor. The programmable pulse generator is used to automatically trigger the pulse generator to generate simulated partial discharge signals at preset time intervals or according to system instructions. The digital signal processor is used to synchronously acquire the response signals of each sensor in the multi-dimensional sensing acquisition module to the simulated partial discharge signals; compare the response signals with the reference template under ideal conditions or during factory calibration, calculate the sensitivity offset; and adjust and compensate the conditioning parameters in the signal conditioning circuit according to the calculated offset.

4. The partial discharge monitoring device with multimodal sensing according to claim 1, characterized in that, It also includes a communication module, which is used to establish a communication connection between the local processing module and the cloud federated learning module, encrypt and transmit the high-dimensional feature vector and partial discharge diagnosis results extracted by the local processing module, and receive global model parameters issued by the cloud federated learning module.

5. A multimodal sensing partial discharge monitoring device according to claim 2, characterized in that, The multidimensional data model is a four-dimensional data model, wherein the extraction of high-dimensional feature vectors from multidimensional sensor data specifically includes: Information from UHF sensors and photoelectric sensors is mapped onto PRPD / PRPS maps, respectively, to serve as temporal features in the four-dimensional data model; Fourier transforms were performed on the signals from the UHF sensor and the ultrasonic array to extract spectral features, which were used as frequency domain features in the four-dimensional data model. The spectral features included the dominant frequency, bandwidth, and spectral entropy. Statistical analysis is performed on time-domain features, frequency-domain features, and spatiotemporal correlation features to obtain statistical feature vectors, which are used as statistical features in the four-dimensional data model.

6. The partial discharge monitoring device with multimodal sensing according to claim 1, characterized in that, The method for calculating the covariance is as follows: ， in, Let U be the covariance between sensor u and sensor v, where U = {u1, u2, ..., u}. D } and V={v1,v2...v D These are two continuous sensor signals from different acquisition time points during time alignment. For the signal acquired by a certain sensor at the e-th acquisition time point during time alignment, For the signal acquired by another sensor at the e-th acquisition time point during time alignment, This represents the average of all signals from a single sensor at different acquisition time points during time alignment. The mean of all signals from another sensor at different acquisition time points during time alignment is given, where D is the total number of acquisition time points. The specific method for calculating mutual information is as follows: ， in, For mutual information between sensor u and sensor v, Let be the joint probability of the i-th first event classification interval of the continuous signal U acquired by sensor u and the j-th second event classification interval of the continuous signal V acquired by sensor v. Let the edge probability of the i-th first event classification interval be the continuous signal U collected by sensor u. M1 represents the marginal probability of the j-th second event category interval for the continuous signal V acquired by sensor v, M2 represents the total number of first event category intervals for the continuous signal U acquired by sensor u, and M3 represents the total number of second event category intervals for the continuous signal V acquired by sensor v. The method for calculating maximum delay mutual information is as follows: =angry max , in, The maximum time delay mutual information between the continuous signal U acquired by sensor u and the continuous signal V acquired by sensor v. Let τ be the time delay mutual information between the continuous signal U acquired by sensor u and the continuous signal V acquired by sensor v, and let τ be the time delay between the continuous signal U acquired by sensor u and the continuous signal V acquired by sensor v. The method for calculating time delay mutual information is as follows: ， in, The continuous signal U acquired by sensor u at the actual acquisition time t and the signal U acquired by sensor v at the actual acquisition time t and the time delay are given by the two signals. The mutual information between the continuously acquired signals V.

7. A multimodal sensing method for monitoring partial discharge, characterized in that, Based on the multimodal sensing partial discharge monitoring device according to any one of claims 1-6, it includes: The multidimensional sensing acquisition module captures signals generated by partial discharge from different physical dimensions; the signal conditioning and data conversion module is used to acquire and condition the signals generated by partial discharge captured by the multidimensional sensing acquisition module, and to perform time-division multiplexing on the signals generated by partial discharge captured from different physical dimensions, converting them into multidimensional sensing acquisition data. The local processing module establishes a multidimensional data model, extracts high-dimensional feature vectors from the multidimensional sensor data, and performs preliminary partial discharge diagnosis based on the high-dimensional feature vectors and the global model obtained from the cloud federated learning module. The preliminary partial discharge diagnosis results are then sent to the cloud federated learning module. The high-dimensional feature vectors include time-domain features, frequency-domain features, spatiotemporal correlation features, and statistical features. The cloud-based federated learning module is used to receive high-dimensional feature vectors and partial discharge diagnosis results. It adjusts and updates the weights of the terminal devices where the local processing modules are located in the global model based on the high-dimensional feature vectors of different local processing modules. It also performs global model weighted aggregation on the high-dimensional feature vectors uploaded by each local processing module based on the weights of the terminal devices where the local processing modules are located in the global model, and obtains the updated global model. The local processing module retrieves the updated global model from the cloud-based federated learning module.

Images