Partial discharge on-line detection method based on multi-sensor data fusion and deep learning

Abstract

The application discloses a kind of based on multi-sensor data fusion and deep learning's partial discharge on-line detection method, it is related to partial discharge technical field, solve the problem of insufficient discharge detection of existing detection method.The device is by mobile inspection node, distributed fixed monitoring node, network aggregation gateway, edge computing unit and cloud diagnosis center is constituted, through multi-source heterogeneous sensor synchronous acquisition high frequency, ultrasonic wave, current and transient ground voltage signal, after pre-processing and feature extraction, realize adaptive weighted fusion and local decision based on information geometry on edge side, cloud end uses multimodal Transformer and contrast learning to carry out deep feature fusion and pattern recognition, finally output diagnostic report.The application greatly improves the all-round perception of partial discharge, multi-level intelligent diagnosis capability.

Description

Technical Field

[0001] This invention relates to the field of partial discharge technology, and more specifically to an online partial discharge detection method based on multi-sensor data fusion and deep learning. Background Technology

[0002] Partial discharge is a significant early sign of insulation degradation in high-voltage electrical equipment, and its effective monitoring is crucial for preventing equipment failures and ensuring the safe operation of the power grid. Traditional partial discharge detection methods often rely on single-type sensors, such as ultra-high frequency (UHF), ultrasonic, or pulsed current methods. These methods are susceptible to interference in the complex electromagnetic environment of substations, and the diversity of discharge types and signal propagation paths makes it difficult for a single sensor to achieve comprehensive and accurate fault identification and location. Furthermore, existing fixed monitoring systems have limited coverage, making it difficult to monitor all critical equipment in a large substation without blind spots; while portable inspections suffer from poor data continuity, reliance on manual experience, and low efficiency. At the data processing level, traditional threshold alarms or simple feature classification methods are insufficient to effectively extract weak discharge characteristics from massive, high-dimensional, and non-stationary monitoring data.

[0003] Therefore, there is an urgent need to develop a comprehensive online detection technology that integrates flexible mobile inspection, wide-area fixed deployment, multimodal data collaboration, and edge-cloud collaborative intelligent diagnosis to overcome the limitations of existing methods in terms of monitoring flexibility, data comprehensiveness, diagnostic accuracy, and system adaptability. Summary of the Invention

[0004] To address the problems of large monitoring blind spots, weak anti-interference capabilities, and insufficient diagnostic intelligence in existing partial discharge detection technologies, this invention discloses an online detection method based on multi-sensor data fusion and deep learning. This method utilizes multiple source sensors—high-frequency, ultrasonic, current, and transient ground voltage—to simultaneously acquire data. It employs FPGA and ASIC hardware acceleration technologies to achieve signal preprocessing and Riemannian-based adaptive feature fusion. Furthermore, it leverages a multimodal Transformer model and a contrastive learning mechanism for deep feature-level fusion and pattern recognition, thereby improving discharge monitoring capabilities.

[0005] To achieve the above-mentioned technical effects, the present invention adopts the following technical solution:

[0006] A partial discharge online detection device based on multi-sensor data fusion and deep learning includes a mobile inspection node, a distributed fixed monitoring node, a network aggregation gateway, an edge computing unit, and a cloud diagnostic center;

[0007] The mobile inspection node includes a traveling chassis, a liftable and rotatable support mechanism mounted on the traveling chassis, and a multi-sensor integrated cabin mounted on the top of the support mechanism; the multi-sensor integrated cabin encapsulates a high-frequency sensor and an ultrasonic sensor.

[0008] The distributed fixed monitoring node includes a protective housing, a current sensor and a transient ground voltage sensor built into the protective housing;

[0009] The network aggregation gateway is connected to the mobile inspection node through a first wireless communication link and to the distributed fixed monitoring node through a second wired communication link, thereby aggregating multi-source monitoring data collected by the mobile inspection node and the distributed fixed monitoring node.

[0010] The edge computing unit is connected to the network aggregation gateway via a local area network. The edge computing unit includes a data preprocessing module, a feature extraction module, and a local fusion decision module.

[0011] The cloud-based diagnostic center is connected to the edge computing unit via a wide area network. The cloud-based diagnostic center includes a deep learning diagnostic model, a feature-level fusion engine, and a diagnostic result generation module.

[0012] The traveling chassis includes a pair of rubber track wheel sets driven by a waterproof motor, the waterproof motor being connected to the track drive wheels via a reducer; the chassis frame of the traveling chassis is welded from aluminum alloy profiles, and a battery compartment and motor controller are fixed inside.

[0013] The support mechanism includes a screw lifting mechanism and a rotating mechanism located at the top. The screw lifting mechanism includes a vertically mounted precision ball screw, which is driven by a first servo motor mounted on the chassis frame via a synchronous belt pulley mechanism. The nut seat that cooperates with the ball screw and the linear bearing sleeved on the parallel optical axis together form a lifting guide structure.

[0014] The rotating mechanism includes a slewing bearing, with the inner ring fixed to the nut seat via a flange and the outer ring connected to the shock-absorbing bracket of the multi-sensor integrated cabin via a flange, and driven by a second servo motor to rotate continuously 360 degrees.

[0015] The multi-sensor integrated cabin has a metal sealed cabin with electromagnetic shielding. The high-frequency sensor is fixed inside by a rigid mounting plate, and a damping vibration isolator is provided between the mounting plate and the cabin wall. The ultrasonic sensor is mounted on a sound-transmitting hole opened on the top of the cabin via a gimbal, which is driven by a micro stepper motor. The cabin also contains an environmental temperature and humidity sensor and a Beidou / GPS timing module. The circuit boards of the environmental temperature and humidity sensor and the timing module are fixed to the rigid mounting plate.

[0016] The base and top cover of the protective shell in the distributed fixed monitoring node are sealed and connected by stainless steel bolts, and an annular silicone sealing ring is embedded at the joint; the back of the base is provided with an arc-shaped groove for adapting to the mounting surface.

[0017] The current sensor is a detachable Rogowski coil structure. The main body of the Rogowski coil structure is encapsulated in a high-temperature resistant silicone sheath, and metal plug-in connectors are connected to both ends of the coil. A socket matching the metal plug-in connector is fixed inside the base of the protective housing, and the socket is equipped with a locking buckle mechanism.

[0018] The transient ground voltage sensor is equipped with an engineering plastic probe encapsulating a silver-plated copper sensing electrode. The probe is connected to the signal conditioning unit via a coaxial cable. The signal conditioning unit and the control board of the node are housed within the protective housing. The protective housing is divided into a main control board compartment and a sensor interface compartment by a metal partition. The socket of the current sensor and the coaxial cable interface of the transient ground voltage sensor are both located in the sensor interface compartment. A metal mounting plate with threaded holes is welded to the back of the protective housing.

[0019] The network aggregation gateway has a wall-mounted steel chassis with status indicator lights and an Ethernet interface on the front panel and a power input interface and grounding bolt on the back. The motherboard, power module and various communication interface cards are installed inside the wall-mounted steel chassis via guide rails.

[0020] The first wireless communication link has a first waterproof omnidirectional antenna installed outside the multi-sensor integration cabin of the mobile inspection node. The antenna is connected to the wireless communication module inside the cabin through a low-loss coaxial cable. A second waterproof omnidirectional antenna or directional antenna is installed on the top of the chassis of the network aggregation gateway. It is connected to the internal wireless communication module through the N-type interface on the top of the chassis. The wireless communication module is a 5G module and is soldered onto an expansion board inserted into the PCIe slot of the motherboard.

[0021] The second wired communication link is provided by an M12 industrial Ethernet interface integrated into the protective housing sidewall of the distributed fixed monitoring node, which is directly connected to the Ethernet interface on the front of the network aggregation gateway chassis via a shielded twisted pair cable.

[0022] The industrial computer chassis of the edge computing unit adopts a compact fanless design, and the outer shell is an aluminum casting with heat dissipation fins; the motherboard of the edge computing unit is fixed to the base plate by copper pillars, and the motherboard is soldered with an ARM processor and DDR4 memory chips.

[0023] The hardware carrier of the data preprocessing module is an FPGA acceleration card inserted into the PCIe x4 slot of the motherboard. The acceleration card is connected to the motherboard through gold fingers and integrates a high-speed ADC chip and a high-capacity FPGA chip.

[0024] The feature extraction module is a hardware carrier consisting of a main processor and a coprocessor acceleration card. The acceleration card can be an ASIC chip, connected to the motherboard via a PCIe x8 slot, and has independent video memory or cache and a parallel computing core.

[0025] The process of implementing edge computing functionality through the collaborative design of the above hardware architecture and software algorithms includes the following steps:

[0026] Step 1: Hardware Platform Construction and Thermal Management Design

[0027] The network aggregation gateway (3) adopts a compact fanless industrial computer chassis. The aluminum cast shell is formed into an integrated heat dissipation fin (411) structure through precision casting. The motherboard is mechanically fixed to the base plate through copper pillars (412). The copper pillars also serve as a heat conduction path to conduct the heat of the motherboard to the aluminum cast shell. The ARM processor and DDR4 memory chips are directly integrated into the motherboard using BGA soldering technology to form a minimized circuit path.

[0028] Step 2: Hardware Acceleration for Data Acquisition and Preprocessing

[0029] The FPGA acceleration card establishes a high-speed data transmission channel with the motherboard through the gold finger interface of the PCIe x4 slot; the high-speed ADC chip samples and quantizes the analog sensor signal, and then transmits it directly to the FPGA chip through the onboard bus; the FPGA chip is internally programmed to implement a parallel pipeline architecture, and synchronously performs filtering, calibration and format conversion operations of multi-channel data through configurable logic units;

[0030] Step 3: Accelerating Feature Extraction through Coprocessing

[0031] An ASIC coprocessor accelerator card is used to obtain a high-bandwidth connection through the PCIe x8 slot, and the independent video memory uses GDDR6 chips to implement data caching; when the main processor sends a calculation instruction, the parallel computing core in the ASIC chip starts feature extraction operation according to the hardened circuit structure, and implements convolution operation and feature dimensionality reduction through fixed logic circuit;

[0032] Step 4: Edge Intelligent Decision Making and Output:

[0033] The ARM processor coordinates the computation results of the FPGA and ASIC, performs data fusion in DDR4 memory, and then executes a lightweight decision algorithm. Finally, it outputs control commands through the industrial communication interface integrated on the motherboard, while continuously dissipating heat through the heat sink fins of the aluminum cast shell to ensure long-term stable operation of the equipment in industrial environments.

[0034] The server rack of the cloud diagnostic center is a standard 42U rack, which contains multiple blade servers and storage devices.

[0035] The deep learning diagnostic model runs on a GPU server equipped with multiple GPU accelerator cards. The GPU servers are interconnected via a high-speed InfiniBand network to form a computing cluster. Each GPU accelerator card is vertically inserted into the backplane of the server via a PCIe Gen4 slot and is forced to be cooled by air through a heat dissipation duct.

[0036] The feature-level fusion engine is deployed on another set of high-performance GPU servers. These servers are connected to the GPU server cluster and storage devices via a 10 Gigabit Ethernet switch at the top of the rack.

[0037] The diagnostic result generation module is deployed on the application server responsible for web services and database management;

[0038] The GPU server cluster, GPU server group, application server, and storage devices are all connected to the network via a 10 Gigabit Ethernet switch located at the top of the rack.

[0039] A partial discharge online detection method based on multi-sensor data fusion and deep learning, using any one of the above-mentioned partial discharge online detection devices based on multi-sensor data fusion and deep learning, includes the following steps:

[0040] Step 1: Synchronous Acquisition and Transmission of Multi-Source Heterogeneous Data

[0041] The mobile inspection node moves and positions itself according to a preset path, the support mechanism adjusts the position and orientation of the multi-sensor integrated cabin, the high-frequency sensor and the ultrasonic sensor synchronously collect signals, and at the same time, the current sensor and the transient ground voltage sensor of the distributed fixed monitoring node continuously collect data, and all data are embedded with a unified time scale.

[0042] Step 2: Monitoring Data Aggregation and Preliminary Preprocessing

[0043] The mobile inspection node aggregates data to the network aggregation gateway via the first wireless communication link, and the distributed fixed monitoring node aggregates data via the second wired communication link. The data is then transmitted to the edge computing unit via the local area network. The data preprocessing module performs synchronous sampling, quantization, and filtering on the signal.

[0044] Step 3: Edge-side feature extraction and local decision-making:

[0045] The feature extraction module extracts time-domain and frequency-domain features from the preprocessed multi-source signal, and the local fusion decision module combines environmental parameters to perform feature association and threshold comparison to generate local early warning decisions.

[0046] Step 4: Cloud-based deep feature fusion and model diagnosis:

[0047] The edge computing unit uploads the feature data and the original data fragments to the cloud diagnostic center through the wide area network. The feature-level fusion engine performs multi-dimensional feature fusion, and the deep learning diagnostic model performs pattern recognition and risk level assessment on the fused features.

[0048] Step 5: Generation and Visualization of Diagnostic Results

[0049] The diagnostic result generation module generates a structured diagnostic report and pushes it to the user terminal via a Web service interface for graphical display and risk point marking;

[0050] Step Six: Feedback Optimization and Model Update

[0051] The cloud-based diagnostic center collects new sample data to retrain the deep learning diagnostic model and sends the optimized model parameters to the edge computing unit.

[0052] In step three, the edge-side feature fusion decision algorithm based on information geometry and adaptive weighting includes:

[0053] The manifold space mapping and representation of multi-source features treats the time-frequency domain feature vectors extracted by each sensor as points on a Riemannian manifold, and calculates the covariance matrix to represent the deep statistical properties of the signal. For a data window of sensor s containing M feature points, the covariance matrix is ​​defined as: In formula (1), It is the feature vector at the m-th time point. It is the mean of the feature vectors within the window. D is the feature dimension;

[0054] Riemann distance-based adaptive decision fusion calculates the geometric distance between the current multi-sensor feature manifold and the preset health state baseline manifold, and then performs adaptive weighted fusion.

[0055] Comprehensive abnormal indicators Calculated based on log-Euclidean-Riemann distance: In formula (2), the adaptive weights , It is the gradient norm of the current window covariance matrix. It is a sensitivity adjustment hyperparameter; It is the baseline covariance matrix of sensor s under healthy conditions, and Logm(·) is the matrix logarithm operation. It is the square of the Frobenius norm. In step four, the cloud-based deep diagnostic algorithm based on multimodal Transformer and cross-modal contrastive learning includes:

[0056] The spatiotemporal alignment and embedding of multimodal features utilizes the Transformer encoder to jointly encode heterogeneous spectral data and capture long-range dependencies;

[0057] The input to the cross-modal Transformer is a stitch of features from all sensors. , It is the feature vector of the ultra-high frequency sensor. It is the feature vector of the ultrasonic sensor. It is the characteristic vector of a high-frequency current transformer. The feature vector of the transient voltage sensor is output by the self-attention mechanism as follows: In formula (3), For the query, key, and value matrix, M is a specially designed modal mask matrix. This represents the dimension of each feature vector in the query matrix and the key matrix. This indicates matrix multiplication between the query matrix and the transpose of the key matrix; Softmax() is the activation function.

[0058] Decoupling representation and diagnosis based on contrastive learning introduces decoupling contrastive loss into model training to separate discharge type features from environmental interference features.

[0059] Decoupling contrast loss function Defined as: In formula (4), It is the feature vector of the target sample. It is a set of positive samples. It is a set of negative samples. It is a similarity measurement function. It's a temperature over-parameter. It is a logarithmic function.

[0060] In step six, the incremental model update algorithm based on meta-learning and Bayesian optimization includes:

[0061] The meta-gradient fast adaptation of model parameters is achieved by using the Model Independent Meta-Learning (MAML) framework, which enables the model to adapt quickly based on a small number of new samples.

[0062] The inner and outer loop update processes of meta-learning are as follows: In formulas (5) and (6), These are the model's meta-parameters. It is the i-th task. The model in the task It's training loss. It is a loss function The gradient of the meta-parameters, It is the inner loop learning rate. The model in the task The parameters after rapid adaptation The model in the task Test loss, From task distribution Multiple tasks in mid-sampling Sum of losses;

[0063] The update strategy optimization based on Bayesian inference is to model the model update decision as a Bayesian optimization problem, balancing exploration and exploitation;

[0064] Update the payoff function In formula (7), the utility function , It is the KL divergence. It is an existing distributed model. It is to quantify the distribution of the new model. It is a compromise factor. These are the parameters of the candidate model to be evaluated. These are the current optimal parameters. It is the expectation operator. It is a non-negative truncation function.

[0065] The positive and beneficial technical effects of this invention are as follows:

[0066] This invention effectively addresses a series of problems described in the background art by constructing a heterogeneous sensor network combining mobile inspection and fixed monitoring, and employing an edge-cloud collaborative intelligent processing architecture. Through an adaptive fusion algorithm based on information geometry and hardware acceleration processing at the edge, it significantly improves the extraction capability of weak discharge characteristics under strong noise backgrounds and the real-time performance of local decision-making. The multimodal Transformer model and meta-learning update mechanism used in the cloud achieve accurate pattern recognition and risk level assessment. Attached Figure Description

[0067] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein:

[0068] Figure 1 This is a diagram illustrating the overall structure of the online partial discharge detection method based on multi-sensor data fusion and deep learning of this invention.

[0069] Figure 2 This is a schematic diagram of a mobile inspection node for the online partial discharge detection method based on multi-sensor data fusion and deep learning of the present invention.

[0070] Figure 3 This is a schematic diagram of a distributed fixed monitoring node for the online partial discharge detection method based on multi-sensor data fusion and deep learning according to the present invention.

[0071] Figure 4 This is a schematic diagram of the network aggregation gateway for the online partial discharge detection method based on multi-sensor data fusion and deep learning of the present invention;

[0072] Figure 5 This is a schematic diagram of the edge computing unit of the online partial discharge detection method based on multi-sensor data fusion and deep learning of the present invention;

[0073] Figure 6 This is a flowchart of the online partial discharge detection method based on multi-sensor data fusion and deep learning of the present invention;

[0074] Figure 7 This is a flowchart of the edge computing process of the online partial discharge detection method based on multi-sensor data fusion and deep learning of the present invention.

[0075] Figure 8 This is a flowchart of the edge-side adaptive fusion decision-making process of the partial discharge online detection method based on multi-sensor data fusion and deep learning in this invention.

[0076] Figure 9 This is a flowchart illustrating the incremental update process of the cloud model for the online partial discharge detection method based on multi-sensor data fusion and deep learning, as described in this invention.

[0077] In the diagram: Mobile inspection node 1, Distributed fixed monitoring node 2, Network aggregation gateway 3, Edge computing unit 4, Cloud diagnostic center 5, Traveling chassis 11, Support mechanism 12, Multi-sensor integrated cabin 13, High-frequency sensor 131, Ultrasonic sensor 132, Protective housing 21, Current sensor 22, Transient ground voltage sensor 23, First wireless communication link L1, Second wired communication link L2, Local area network L3, Data preprocessing module 41, Feature extraction module 42, Local fusion decision module 43, Wide area network L4, Deep learning diagnostic model 51, Feature-level fusion engine 52, Diagnostic result generation module 53, Waterproof motor 111, Screw lifting mechanism 121, Rotation mechanism 122, Micro stepper motor 133, Stainless steel bolt 211, Engineering plastic probe head 231, Status indicator light 31, Ethernet interface 32, Motherboard 33, Power module 34, Communication interface board 35, Heat sink 411, Copper pillar 412. Detailed Implementation

[0078] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the embodiments described herein are for illustration and explanation only and are not intended to limit the present invention. Figures 1-9 As shown: Example 1:

[0079] A partial discharge online detection device based on multi-sensor data fusion and deep learning includes a mobile inspection node 1, a distributed fixed monitoring node 2, a network aggregation gateway 3, an edge computing unit 4, and a cloud diagnostic center 5;

[0080] The mobile inspection node 1 includes a traveling chassis 11, a liftable and rotatable support mechanism 12 mounted on the traveling chassis 11, and a multi-sensor integrated cabin 13 mounted on the top of the support mechanism 12; the multi-sensor integrated cabin 13 encapsulates a high-frequency sensor 131 and an ultrasonic sensor 132.

[0081] The distributed fixed monitoring node 2 includes a protective housing 21, a current sensor 22 and a transient ground voltage sensor 23 built into the protective housing 21;

[0082] The network aggregation gateway 3 is connected to the mobile inspection node 1 through the first wireless communication link L1 and to the distributed fixed monitoring node 2 through the second wired communication link L2, thereby aggregating multi-source monitoring data collected by the mobile inspection node 1 and the distributed fixed monitoring node 2.

[0083] The edge computing unit 4 is connected to the network aggregation gateway 3 via a local area network L3. The edge computing unit 4 includes a data preprocessing module 41, a feature extraction module 42, and a local fusion decision module 43.

[0084] The cloud-based diagnostic center 5 is connected to the edge computing unit 4 via a wide area network L4. The cloud-based diagnostic center 5 includes a deep learning diagnostic model 51, a feature-level fusion engine 52, and a diagnostic result generation module 53.

[0085] In a specific embodiment, the mobile inspection node 1, via its traveling chassis 11 and a liftable and rotatable support mechanism 12, drives the ultra-high frequency sensor 131 and ultrasonic sensor 132 within the multi-sensor integrated cabin 13 to achieve spatial adaptive positioning and detection. The network aggregation gateway 3 aggregates the aforementioned multi-source heterogeneous data via wireless link L1 and wired link L2. The edge computing unit 4 uses a data preprocessing module 41 to perform real-time signal normalization, a feature extraction module 42 to extract time-frequency features, and a local fusion decision module 43 to perform rapid fusion decision-making based on information geometry algorithms, ensuring real-time early warning. The cloud-based diagnostic center 5 relies on a deep learning diagnostic model 51 and a feature-level fusion engine 52 to perform deep feature fusion and pattern recognition, and a diagnostic result generation module 53 generates an accurate diagnostic report. Example 2:

[0086] The traveling chassis 11 includes a pair of rubber track wheels driven by a waterproof motor 111. The waterproof motor 111 is connected to the track drive wheels through a reducer. The chassis frame of the traveling chassis 11 is welded from aluminum alloy profiles, and a battery compartment and a motor controller are fixed inside.

[0087] The support mechanism 12 includes a screw lifting mechanism 121 and a rotating mechanism 122 located at the top. The screw lifting mechanism 121 includes a vertically mounted precision ball screw, which is driven by a first servo motor mounted on the chassis frame through a synchronous belt pulley mechanism. The nut seat that cooperates with the ball screw and the linear bearing sleeved on the parallel optical axis together form a lifting guide structure.

[0088] The rotating mechanism 122 includes a slewing bearing, with the inner ring fixed to the nut seat via a flange and the outer ring connected to the shock-absorbing bracket of the multi-sensor integrated cabin 13 via a flange, and driven by a second servo motor to rotate continuously in 360 degrees.

[0089] The shell of the multi-sensor integrated cabin 13 is a metal sealed cabin with electromagnetic shielding function. The high-frequency sensor 131 is fixed inside by a rigid mounting plate. A damping vibration isolator is provided between the mounting plate and the cabin wall. The ultrasonic sensor 132 is mounted on a sound-transmitting hole opened on the top of the cabin via a gimbal. The gimbal is driven by a micro stepper motor 133. The cabin also contains an environmental temperature and humidity sensor and a Beidou / GPS timing module. The circuit boards of the environmental temperature and humidity sensor and the timing module are fixed to the rigid mounting plate.

[0090] In a further specific embodiment, the chassis 11 of the present invention uses an aluminum alloy profile welded frame as its load-bearing foundation. Its rubber track wheel set is driven by a waterproof motor 111 via a reducer. The track structure improves the adaptability to complex terrain. The chassis is powered by a battery compartment and the motor controller realizes power regulation. In the support mechanism 12, the screw lifting mechanism 121 is based on a vertical precision ball screw. It is driven by a first servo motor via a synchronous belt pulley mechanism and a guide structure composed of a nut seat and a parallel optical axis outer linear bearing to achieve stable lifting. The rotating mechanism 122 relies on a slewing bearing. The inner flange is fixed to the nut seat and the outer flange is connected to the shock-absorbing bracket of the multi-sensor integrated compartment 13. It is driven by a second servo motor to complete 360-degree continuous rotation. The multi-sensor integrated cabin 13 employs a metal-sealed cabin body for electromagnetic shielding. The high-frequency sensor 131 is fixed via a rigid mounting plate, with a damping vibration isolator between the mounting plate and the cabin wall to reduce vibration interference. The ultrasonic sensor 132 is mounted on a gimbal via a sound-permeable opening in the cabin body, with its detection angle adjusted by a micro stepper motor 133. The circuit boards of the cabin's environmental temperature and humidity sensors and the BeiDou / GPS timing module are also fixed to the rigid mounting plate, ensuring data acquisition synchronization and environmental adaptability. Example 3:

[0091] The base and top cover of the protective housing 21 in the distributed fixed monitoring node 2 are sealed together by stainless steel bolts 211, and an annular silicone sealing ring is embedded at the joint; the back of the base is provided with an arc-shaped groove for adapting to the mounting surface.

[0092] The current sensor 22 is a separable Rogowski coil structure. The main body of the Rogowski coil structure is encapsulated in a high-temperature resistant silicone sheath, and the two ends of the coil are connected to metal plug-in connectors. The base of the protective housing 21 has a socket that matches the metal plug-in connector, and the socket is equipped with a locking buckle mechanism.

[0093] The transient ground voltage sensor 23 is equipped with an engineering plastic probe 231 encapsulating a silver-plated copper sensing electrode. The probe 231 is connected to the signal conditioning unit via a coaxial cable. The signal conditioning unit and the control board of the node are housed within the protective housing 21, and the protective housing 21 is divided into a main control board compartment and a sensor interface compartment by a metal partition. The socket of the current sensor 22 and the coaxial cable interface of the transient ground voltage sensor 23 are both located in the sensor interface compartment. A metal mounting plate with threaded holes is welded to the back of the protective housing 21.

[0094] In application, the protective housing 21 adopts a structure in which the base and the top cover are sealed together by stainless steel bolts 211. The annular silicone sealing ring at the joint ensures the IP protection level. The arc-shaped groove on the back of the base adapts to different mounting surface curvatures, and the metal mounting plate with threaded holes on the back ensures reliable fixation. The current sensor 22 adopts a separable Rogowski coil structure. The coil body is encapsulated in a high-temperature resistant silicone sheath to resist the high ambient temperature. The metal plug-in connectors at both ends mate with the matching sockets inside the base of the protective housing 21. The socket locking buckle mechanism prevents the connection from loosening. The transient ground voltage sensor 23 collects signals with an engineering plastic probe 231 encapsulating silver-plated copper sensing electrodes. The signals are transmitted to the signal conditioning unit via a coaxial cable. The metal partition inside the protective housing 21 separates the main control board compartment and the sensor interface compartment. The socket of the current sensor 22 and the coaxial cable interface of the transient ground voltage sensor 23 are both located in the sensor interface compartment. Example 4:

[0095] The network aggregation gateway 3 has a wall-mounted steel chassis with a status indicator light 31 and an Ethernet interface 32 on the front panel and a power input interface and a grounding bolt on the back. Inside the wall-mounted steel chassis, a motherboard 33, a power module 34 and various communication interface cards 35 are installed via guide rails.

[0096] The first wireless communication link L1 has a first waterproof omnidirectional antenna installed outside the multi-sensor integration compartment 13 of the mobile inspection node 1. The antenna is connected to the wireless communication module inside the compartment through a low-loss coaxial cable. A second waterproof omnidirectional antenna or directional antenna is installed on the top of the chassis of the network aggregation gateway 3. It is connected to the internal wireless communication module through the N-type interface on the top of the chassis. The wireless communication module is a 5G module and is soldered onto an expansion board inserted into the PCIe slot of the motherboard.

[0097] The second wired communication link L2 has an M12 industrial Ethernet interface integrated on the side wall of the protective housing 21 of the distributed fixed monitoring node 2, which is directly connected to the Ethernet interface on the front of the network aggregation gateway 3 chassis via a shielded twisted pair cable.

[0098] In application, the network aggregation gateway 3 uses a wall-mounted steel chassis as its carrier. The front panel features status indicator lights 31 for visualizing the operating status and an Ethernet interface 32 for data interaction. The rear power input interface ensures power supply, and grounding bolts provide lightning protection and interference suppression. Inside the chassis, a motherboard 33, a power module 34, and various communication interface cards 35 are mounted via guide rails, facilitating module maintenance and expansion. In the first wireless communication link L1, the multi-sensor integrated compartment 13 of the mobile inspection node 1 is externally equipped with a first waterproof omnidirectional antenna, connected to the internal wireless communication module via a low-loss coaxial cable. The top of the network aggregation gateway 3 chassis is equipped with a second waterproof omnidirectional or directional antenna, which interfaces with the internal wireless communication module via an N-type interface. This wireless communication module is a 5G module, soldered onto an expansion board inserted into the motherboard's PCIe slot, enabling high-speed wireless data transmission. In the second wired communication link L2, the protective housing 21 of the distributed fixed monitoring node 2 integrates an M12 industrial Ethernet interface on its side wall. Example 5:

[0099] The industrial computer chassis of the edge computing unit 4 adopts a compact fanless design, and the outer shell is an aluminum casting with heat dissipation fins 411; the motherboard of the edge computing unit 4 is fixed to the base plate by copper pillars 412, and the motherboard is soldered with an ARM processor and DDR4 memory chips.

[0100] The hardware carrier of the data preprocessing module 41 is an FPGA acceleration card inserted into the PCIe x4 slot of the motherboard. The acceleration card is connected to the motherboard through gold fingers and integrates a high-speed ADC chip and a high-capacity FPGA chip.

[0101] The feature extraction module 42 is a hardware carrier of a main processor and a coprocessor acceleration card. The acceleration card can be an ASIC chip, which is connected to the motherboard through a PCIe x8 slot. The card has independent video memory or cache and a parallel computing core.

[0102] The process of implementing edge computing functionality includes the following steps:

[0103] Step 1: Hardware Platform Construction and Thermal Management Design

[0104] The network aggregation gateway (3) adopts a compact fanless industrial computer chassis. The aluminum cast shell is formed into an integrated heat dissipation fin (411) structure through precision casting. The motherboard is mechanically fixed to the base plate through copper pillars (412). The copper pillars also serve as a heat conduction path to conduct the heat of the motherboard to the aluminum cast shell. The ARM processor and DDR4 memory chips are directly integrated into the motherboard using BGA soldering technology to form a minimized circuit path.

[0105] Step 2: Hardware Acceleration for Data Acquisition and Preprocessing

[0106] The FPGA acceleration card establishes a high-speed data transmission channel with the motherboard through the gold finger interface of the PCIe x4 slot; the high-speed ADC chip samples and quantizes the analog sensor signal, and then transmits it directly to the FPGA chip through the onboard bus; the FPGA chip is internally programmed to implement a parallel pipeline architecture, and synchronously performs filtering, calibration and format conversion operations of multi-channel data through configurable logic units;

[0107] Step 3: Accelerating Feature Extraction through Coprocessing

[0108] An ASIC coprocessor accelerator card is used to obtain a high-bandwidth connection through the PCIe x8 slot, and the independent video memory uses GDDR6 chips to implement data caching; when the main processor sends a calculation instruction, the parallel computing core in the ASIC chip starts feature extraction operation according to the hardened circuit structure, and implements convolution operation and feature dimensionality reduction through fixed logic circuit;

[0109] Step 4: Edge Intelligent Decision Making and Output:

[0110] The ARM processor coordinates the computation results of the FPGA and ASIC, performs data fusion in DDR4 memory, and then executes a lightweight decision algorithm. Finally, it outputs control commands through the industrial communication interface integrated on the motherboard, while continuously dissipating heat through the heat sink fins of the aluminum cast shell to ensure long-term stable operation of the equipment in industrial environments.

[0111] In application, the edge computing unit 4 adopts a compact fanless aluminum cast chassis with heat sink fins 411. The motherboard is fixed by copper pillars 412 and conducts heat. The ARM processor and DDR4 memory chips are integrated on the motherboard using BGA technology. The data preprocessing module 41 relies on the FPGA acceleration card in the PCIe x4 slot. It samples and quantizes analog signals through a high-speed ADC chip and transmits them to the FPGA chip via the onboard bus. Its parallel pipeline architecture synchronously performs multi-channel data filtering, calibration, and format conversion. In the feature extraction module 42, the ASIC coprocessing acceleration card is connected through the PCIe x8 slot. It uses GDDR6 video memory to cache data. The parallel computing core implements convolution operations and feature dimensionality reduction based on hardened circuitry. The ARM processor coordinates the FPGA and ASIC calculation results, completes data fusion in DDR4 memory, and executes a lightweight decision algorithm. It outputs control commands through an industrial communication interface. The aluminum cast heat sink ensures continuous heat dissipation. Example 6:

[0112] The server rack of the cloud diagnostic center 5 is a standard 42U rack, which contains multiple blade servers and storage devices.

[0113] The deep learning diagnostic model 51 runs on a GPU server equipped with multiple GPU accelerator cards. The GPU servers are interconnected through a high-speed InfiniBand network to form a computing cluster. Each GPU accelerator card is vertically inserted into the backplane of the server through a PCIeGen4 slot and is forced to be cooled by air through a heat dissipation duct.

[0114] The feature-level fusion engine 52 is deployed on another set of high-performance CPU servers. These servers are connected to the GPU server cluster and storage devices via a 10 Gigabit Ethernet switch at the top of the rack.

[0115] The diagnostic result generation module 53 is deployed on the application server responsible for web services and database management;

[0116] The GPU server cluster, CPU server group, application server, and storage devices are all connected to the network via a 10 Gigabit Ethernet switch located at the top of the rack.

[0117] In application, the cloud-based diagnostic center 5 uses a standard 42U rack as its physical carrier, integrating blade servers and storage devices. The deep learning diagnostic model 51 is deployed on a GPU server configured with multiple GPU accelerator cards. Each GPU accelerator card is vertically mounted on the server backplane via a PCIe Gen4 slot, relying on forced air cooling through a heat dissipation duct to ensure stable operation. The GPU servers are interconnected via a high-speed InfiniBand network to form a computing cluster, enhancing parallel computing capabilities. The feature-level fusion engine 52 is deployed on a high-performance CPU server group, which establishes a data interaction link with the GPU server cluster and storage devices through a 10 Gigabit Ethernet switch at the top of the rack. The diagnostic result generation module 53 is deployed on the application server responsible for web services and database management. Example 7:

[0118] A method for online detection of partial discharge based on multi-sensor data fusion and deep learning includes the following steps:

[0119] Step 1: Synchronous Acquisition and Transmission of Multi-Source Heterogeneous Data

[0120] The mobile inspection node 1 moves and positions itself according to a preset path. The support mechanism 12 adjusts the position and orientation of the multi-sensor integrated cabin 13. The high-frequency sensor 131 and the ultrasonic sensor 132 synchronously collect signals. At the same time, the current sensor 22 and the transient ground voltage sensor 23 of the distributed fixed monitoring node 2 continuously collect data. All data are embedded with a unified time scale.

[0121] Step 2: Monitoring Data Aggregation and Preliminary Preprocessing

[0122] The mobile inspection node 1 aggregates data to the network aggregation gateway 3 via the first wireless communication link L1, and the distributed fixed monitoring node 2 aggregates data via the second wired communication link L2. The data is then transmitted to the edge computing unit 4 via the local area network L3. The data preprocessing module 41 performs synchronous sampling, quantization and filtering on the signal.

[0123] Step 3: Edge-side feature extraction and local decision-making:

[0124] The feature extraction module 42 extracts time-domain and frequency-domain features from the preprocessed multi-source signal, and the local fusion decision module 43 combines environmental parameters to perform feature association and threshold comparison to generate local early warning decisions.

[0125] Step 4: Cloud-based deep feature fusion and model diagnosis:

[0126] The edge computing unit 4 uploads the feature data and the original data fragment to the cloud diagnostic center 5 through the wide area network L4. The feature-level fusion engine 52 performs multi-dimensional feature fusion, and the deep learning diagnostic model 51 performs pattern recognition and risk level assessment on the fused features.

[0127] Step 5: Generation and Visualization of Diagnostic Results

[0128] The diagnostic result generation module 53 generates a structured diagnostic report and pushes it to the user terminal via a Web service interface for graphical display and risk point marking;

[0129] Step Six: Feedback Optimization and Model Update

[0130] The cloud-based diagnostic center 5 collects new sample data to retrain the deep learning diagnostic model 51, and sends the optimized model parameters to the edge computing unit 4.

[0131] This invention achieves online partial discharge detection through a six-step collaborative process: A mobile inspection node 1 moves and positions itself along a preset path; a support mechanism 12 adjusts the orientation of the multi-sensor integrated cabin 13; a high-frequency sensor 131 and an ultrasonic sensor 132 synchronously acquire signals; and current sensors 22 and transient ground voltage sensors 23 of the distributed fixed monitoring node 2 continuously acquire data. All data are embedded with a unified timescale. The mobile inspection node 1, via a first wireless communication link L1, and the distributed fixed monitoring node 2, via a second wired communication link L2, aggregate the data to a network aggregation gateway 3, and then transmit it to an edge computing unit 4 via a local area network L3. A data preprocessing module 41 synchronously samples, quantizes, and filters the signals. A feature extraction module 4... 2. Time-domain and frequency-domain features are extracted from the preprocessed multi-source signals. The local fusion decision module 43 combines environmental parameters to perform feature correlation and threshold comparison to generate local early warning decisions. The edge computing unit 4 uploads the feature data and original data fragments to the cloud diagnostic center 5 via the wide area network L4. The feature-level fusion engine 52 performs multi-dimensional feature fusion, and the deep learning diagnostic model 51 performs pattern recognition and risk level assessment on the fused features. The diagnostic result generation module 53 generates a structured diagnostic report, which is pushed to the user terminal for graphical display and risk point marking via the Web service interface. The cloud diagnostic center 5 collects new sample data to retrain the deep learning diagnostic model 51 and sends the optimized model parameters to the edge computing unit 4. Example 8:

[0132] In step three, the edge-side feature fusion decision algorithm based on information geometry and adaptive weighting includes:

[0133] The manifold space mapping and representation of multi-source features treats the time-frequency domain feature vectors extracted by each sensor as points on a Riemannian manifold, and calculates the covariance matrix to represent the deep statistical properties of the signal. For a data window of sensor s containing M feature points, the covariance matrix is ​​defined as: In formula (1), It is the feature vector at the m-th time point. It is the mean of the feature vectors within the window. D is the feature dimension;

[0134] Riemann distance-based adaptive decision fusion calculates the geometric distance between the current multi-sensor feature manifold and the preset health state baseline manifold, and then performs adaptive weighted fusion.

[0135] Comprehensive abnormal indicators Calculated based on log-Euclidean-Riemann distance: In formula (2), the adaptive weights , It is the gradient norm of the current window covariance matrix. It is a sensitivity adjustment hyperparameter; It is the baseline covariance matrix of sensor s under healthy conditions, and Logm(·) is the matrix logarithm operation. It is the square of the Frobenius norm. The above scheme is verified by the following experiment, which is carried out in a shielded laboratory simulating the electromagnetic interference environment of a substation. The hardware platform adopts the hardware architecture of the edge computing unit (4), the main processor is ARM Cortex-A72, the FPGA accelerator card is Xilinx Zynq UltraScale+, and the ASIC coprocessor accelerator card is a customized parallel computing unit. The data sources are four types of sensors: ultra-high frequency sensor (UHF), ultrasonic sensor (US), high frequency current transformer (HFCT) and transient ground voltage sensor (TEV), with sampling frequencies of 1 GHz, 500 kHz, 100 MHz and 200 MHz, respectively. A total of 500 healthy state samples and 800 typical partial discharge fault samples (including corona discharge, surface discharge, internal discharge, etc.) were collected in the experiment. The duration of each sample is 10 ms and the feature dimension D=16.

[0136] Method Comparison: The algorithm of this invention (denoted as IG-AW) is compared with two traditional methods: one is a weighted average fusion based on Euclidean distance (denoted as Euclid-Avg), and the other is a multi-sensor decision fusion based on fixed weights (denoted as Fixed-Weight). Evaluation Metrics: The area under the receiver operating characteristic curve (AUC), accuracy, and false positive rate (FPR) are used as the core evaluation metrics for anomaly detection.

[0137] Using 500 health status samples, the baseline covariance matrix of each sensor is calculated according to formula (1). For 800 fault samples and an additional 200 healthy test samples, the current covariance matrix is ​​calculated by sliding over a time window (M=50 feature points). The comprehensive anomaly index for each time window is calculated according to formula (2). The hyperparameter λ was optimized to 0.5 using a grid search. A dynamic threshold was set. The methods were categorized and their performance metrics were statistically analyzed. The experimental data processing results are shown in Table 1.

[0138] Table 1 Performance Comparison of Different Feature Fusion Algorithms

[0139]

[0140] Experimental data show that the algorithm of this invention (IG-AW) significantly outperforms the traditional Euclidean distance-weighted average method and fixed-weight method in key indicators such as AUC, accuracy, and false alarm rate. The fundamental reason is that this algorithm maps feature vectors to Riemannian manifold space, using the covariance matrix to characterize deeper statistical properties of the signal. Its ability to describe the intrinsic geometric structure of the data is superior to the Euclidean space assumption. Simultaneously, the adaptive weighting mechanism based on the gradient norm of the covariance matrix can dynamically respond to instantaneous changes in the signal quality of each sensor, thus more robustly fusing multi-source information in strong noise backgrounds. This effectively improves the sensitivity and decision-making accuracy for detecting partial discharge anomalies, meeting the dual requirements of real-time performance and reliability at the edge. Example 9:

[0141] In step four, the cloud-based deep diagnostic algorithm based on multimodal Transformer and cross-modal contrastive learning includes:

[0142] The spatiotemporal alignment and embedding of multimodal features utilizes a Transformer encoder to jointly encode heterogeneous spectral data, capturing long-range dependencies; the input to the cross-modal Transformer is the concatenation of features from all sensors. , It is the feature vector of the ultra-high frequency sensor. It is the feature vector of the ultrasonic sensor. It is the characteristic vector of a high-frequency current transformer. The feature vector of the transient voltage sensor is output by the self-attention mechanism as follows: In formula (3), For the query, key, and value matrix, M is a specially designed modal mask matrix. This represents the dimension of each feature vector in the query matrix and the key matrix. This indicates matrix multiplication between the query matrix and the transpose of the key matrix; Softmax() is the activation function.

[0143] Decoupling representation and diagnosis based on contrastive learning introduces decoupling contrastive loss into model training to separate discharge type features from environmental interference features.

[0144] Decoupling contrast loss function Defined as: In formula (4), It is the feature vector of the target sample. It is a set of positive samples. It is a set of negative samples. It is a similarity measurement function. It's a temperature over-parameter. It is a logarithmic function. The experiment was conducted in the hardware environment of the cloud diagnostic center (5), using two servers equipped with 4 NVIDIA A100 GPU accelerator cards, interconnected through the InfiniBand network. The software environment was PyTorch 1.12.0, the number of Transformer model layers was set to 6, the number of attention heads was 8, and the feature dimension d_k=64. The dataset contains a total of 15,000 samples from four types of sensors: UHF, US, HFCT, and TEV, covering three typical fault types: corona discharge, surface discharge, and internal discharge, as well as background electromagnetic interference of various intensities. All samples were labeled by experts. The algorithm of this invention (denoted as MMT-CCL, i.e., multimodal Transformer and cross-modal contrastive learning) was compared with three baseline methods: one is the feature fusion method based on convolutional neural network (CNN-Concat), the second is the ordinary multimodal Transformer without contrastive loss (MMT-Base), and the third is the support vector machine multi-core learning method (SVM-MKL).

[0145] Macro-F1 score and average accuracy for multi-class classification tasks were used as core evaluation metrics, while the classification stability (standard deviation) of the model on noisy samples was also recorded. All sensor data were normalized and time series of the same length were truncated and divided into training, validation, and test sets in an 8:1:1 ratio. The MMT-CCL model was trained according to formulas (3) and (4), with AdamW as the optimizer and the temperature hyperparameter τ set to 0.1. The baseline model was trained under the same data partition to obtain a fair comparison. The classification performance of each model was evaluated on the test set, and its performance on challenging samples with high-intensity noise was analyzed in particular. The performance comparison results of different diagnostic algorithms on the test set are shown in Table 2.

[0146] Table 2 Performance Comparison of Different Cloud-based Diagnostic Algorithms

[0147]

[0148] The experimental results demonstrate that the cloud-based deep diagnostic algorithm based on multimodal Transformer and cross-modal contrastive learning exhibits significant advantages in partial discharge type identification tasks. The algorithm of this invention (MMT-CCL) outperforms all baseline models in both macro-F1 score and average accuracy, particularly demonstrating superior classification stability (measured by the standard deviation of accuracy) in noisy environments, proving its strong robustness. This is mainly attributed to the Transformer encoder effectively capturing the long-range dependencies between multimodal signals, while the introduced decoupling contrastive loss function effectively separates the essential characteristics of discharge from environmental interference features within the representation space by bringing similar discharge samples closer together and distancing dissimilar discharge and interference samples further away, thereby improving the model's generalization ability and diagnostic accuracy under complex real-world conditions. Example 10:

[0149] In step six, the incremental model update algorithm based on meta-learning and Bayesian optimization includes:

[0150] The meta-gradient fast adaptation of model parameters is achieved by using the Model Independent Meta-Learning (MAML) framework, which enables the model to adapt quickly based on a small number of new samples.

[0151] The inner and outer loop update processes of meta-learning are as follows: In formulas (5) and (6), These are the model's meta-parameters. It is the i-th task. The model in the task It's training loss. It is a loss function The gradient of the meta-parameters, It is the inner loop learning rate. The model in the task The parameters after rapid adaptation The model in the task Test loss, From task distribution Multiple tasks in mid-sampling Sum of losses;

[0152] The update strategy optimization based on Bayesian inference is to model the model update decision as a Bayesian optimization problem, balancing exploration and exploitation;

[0153] Update the payoff function In formula (7), the utility function , It is the KL divergence. It is an existing distributed model. It is to quantify the distribution of the new model. It is a compromise factor. These are the parameters of the candidate model to be evaluated. These are the current optimal parameters. It is the expectation operator. It is a non-negative truncation function. Experiments were conducted on a GPU server cluster in a cloud-based diagnostic center, using the PyTorch 1.12.0 and GPyOpt libraries. The basic diagnostic model was a pre-trained multimodal Transformer model. The simulated data stream contained monitoring data from five consecutive time periods, each introducing a new partial discharge mode or operating condition change (such as surface discharge under different humidity conditions). Each new task provided only 20 labeled samples (small sample condition). The tradeoff coefficient β was set to 0.8. The algorithm of this invention (denoted as MAML-BO, i.e., meta-learning and Bayesian optimization) was compared with three methods: traditional full-data fine-tuning, naive fine-tuning using only new data, and random search-based model updates.

[0154] The main focus is on the model's average accuracy on new tasks (reflecting its ability to adapt quickly), average accuracy on old tasks (reflecting its ability to resist forgetting), and the average number of updates required to achieve the target performance.

[0155] Experimental implementation process:

[0156] Meta-training phase: The MAML framework is meta-trained using a large number of historical tasks to obtain a meta-model with good initialization parameters.

[0157] Incremental update phase: When a new task arrives, the MAML-BO algorithm performs fast gradient adaptation according to formulas (5) and (6) to generate candidate models; then, based on the expectation boosting criterion of formula (7), the optimal update strategy is selected through Bayesian optimization. Other comparative methods use their respective strategies for updating.

[0158] Performance evaluation: After each update cycle, evaluate the model performance on an independent test set containing both new and old tasks, repeat the experiment 5 times and take the average value.

[0159] Table 3 shows the performance comparison results of different model update algorithms after 5 incremental learning stages.

[0160] Table 3 Performance Comparison of Different Incremental Model Update Algorithms

[0161]

[0162] The algorithm of this invention (MAML-BO) achieves the highest recognition accuracy on new tasks while minimizing forgetting of old tasks, demonstrating its efficiency in rapidly adapting to new knowledge and maintaining existing knowledge under small sample conditions. Furthermore, the algorithm requires significantly fewer updates to reach the target performance compared to comparative methods, showcasing the efficiency of its optimization process. This is primarily due to the MAML framework's use of meta-gradient updates to achieve good parameter initialization, enabling the model to quickly adapt to new tasks with a small gradient step size. Meanwhile, Bayesian optimization intelligently finds update strategies that simultaneously improve performance on new tasks without severely impairing performance on old tasks by balancing exploration and utilization. This effectively solves the stability-plasticity dilemma in incremental learning, ensuring the diagnostic model maintains high performance throughout its long-term evolution.

[0163] While specific embodiments of the present invention have been described above, those skilled in the art should understand that these specific embodiments are merely illustrative. Those skilled in the art can omit, substitute, and modify the details of the above methods and systems in various ways without departing from the principles and essence of the present invention. For example, combining the above method steps to perform substantially the same function in substantially the same way to achieve substantially the same result falls within the scope of the present invention. Therefore, the scope of the present invention is defined only by the appended claims.

Claims

1. A partial discharge on-line detection device based on multi-sensor data fusion and deep learning, characterized in that, The mobile inspection node (1), the distributed fixed monitoring node (2), the network aggregation gateway (3), the edge computing unit (4) and the cloud diagnosis center (5) are included. The mobile inspection node (1) includes a traveling chassis (11), a liftable and rotatable supporting mechanism (12) arranged on the traveling chassis (11), and a multi-sensor integrated cabin (13) carried on the top end of the supporting mechanism (12); the multi-sensor integrated cabin (13) is encapsulated with a high-frequency sensor (131) and an ultrasonic sensor (132); The distributed fixed monitoring node (2) includes a protective shell (21), a current sensor (22) and a transient ground voltage sensor (23) arranged in the protective shell (21); The network aggregation gateway (3) is connected with the mobile inspection node (1) through a first wireless communication link (L1), and is connected with the distributed fixed monitoring node (2) through a second wired communication link (L2), and aggregates multi-source monitoring data collected by the mobile inspection node (1) and the distributed fixed monitoring node (2); The edge computing unit (4) is connected with the network aggregation gateway (3) through a local area network (L3), and the edge computing unit (4) includes a data preprocessing module (41), a feature extraction module (42) and a local fusion decision module (43); The cloud diagnosis center (5) is connected with the edge computing unit (4) through a wide area network (L4), and the cloud diagnosis center (5) includes a deep learning diagnosis model (51), a feature level fusion engine (52) and a diagnosis result generation module (53); The traveling chassis (11) includes a pair of rubber track wheel groups driven by waterproof motors (111), and the waterproof motors (111) are connected with track drive wheels through reducers; the chassis frame of the traveling chassis (11) is welded by aluminum alloy profiles, and a battery cabin and a motor controller are fixed inside; The supporting mechanism (12) includes a lead screw lifting mechanism (121) and a rotating mechanism (122) arranged at the top end; the lead screw lifting mechanism (121) includes a vertical precision ball screw, and the ball screw is driven by a first servo motor through a synchronous belt wheel mechanism and is installed on the chassis frame; a nut seat matched with the ball screw and a linear bearing sleeved outside the parallel optical axis jointly constitute a lifting guide structure; The rotating mechanism (122) includes a rotary bearing, the inner ring is fixed with the nut seat through a flange, the outer ring is connected with the damping support of the multi-sensor integrated cabin (13) through a flange, and is driven by a second servo motor to rotate continuously by 360 degrees; The shell of the multi-sensor integrated cabin (13) is a metal sealed cabin body with electromagnetic shielding function, the high-frequency sensor (131) is fixed inside through a rigid mounting plate, and a damping vibration isolator is arranged between the mounting plate and the cabin wall; the ultrasonic sensor (132) is installed at a sound transmission hole opened at the top of the cabin body through a universal gimbal, and the universal gimbal is driven by a micro stepping motor (133); an environmental temperature and humidity sensor and a Beidou / GPS time service module are further arranged in the cabin body, and the circuit boards of the environmental temperature and humidity sensor and the time service module are fixed on the rigid mounting plate.

2. The apparatus of claim 1, wherein, The base and the upper cover of the protective shell (21) in the distributed fixed monitoring node (2) are sealingly connected through stainless steel bolts (211), and an annular silica gel sealing ring is embedded at the joint; The current sensor (22) is a separable Rogowski coil structure, the Rogowski coil structure body is packaged in a high-temperature-resistant silica gel sheath, and metal plug-in connectors are connected at both ends of the coil; a socket matched with the metal plug-in connector is fixed in the base of the protective shell (21); and the socket is provided with a locking buckle mechanism. The transient ground voltage sensor (23) is provided with an engineering plastic probe head (231) packaging a silver-plated copper induction electrode, the probe head (231) is connected to a signal conditioning unit through a coaxial cable; the signal conditioning unit and the control mainboard of the node are accommodated in the protective shell (21), and the protective shell (21) is divided into a main control board cabin and a sensor interface cabin by a metal partition plate; the socket of the current sensor (22) and the coaxial cable interface of the transient ground voltage sensor (23) are located in the sensor interface cabin; and a metal mounting plate with threaded holes is welded on the back of the protective shell (21).

3. The apparatus of claim 1, wherein, The network aggregation gateway (3) is a wall-mounted steel case, the front panel is provided with a state indicating lamp (31) and an Ethernet interface (32), the back is provided with a power input interface and a grounding bolt; the wall-mounted steel case is internally provided with a mainboard (33), a power module (34) and multiple communication interface board cards (35) through guide rails; The first wireless communication link (L1) is provided with a first waterproof omnidirectional antenna outside the multi-sensor integrated cabin (13) of the mobile inspection node (1), the antenna is connected with the wireless communication module in the cabin through a low-loss coaxial cable; a second waterproof omnidirectional antenna or a directional antenna is installed on the top of the case of the network aggregation gateway (3), and is connected with the wireless communication module inside through an N-type interface on the top of the case; the wireless communication module adopts a 5G module and is welded on an expansion board plugged into a PCIe slot of the mainboard; The second wired communication link (L2), the protective shell (21) of the distributed fixed monitoring node (2) is integrated with an M12 specification industrial Ethernet interface on the side wall, and is directly connected to the Ethernet interface in front of the network aggregation gateway (3) through a shielded twisted pair line.

4. The apparatus of claim 1, wherein, The industrial computer case of the edge computing unit (4) adopts a compact fanless design, and the shell is an aluminum casting body with heat dissipation fins (411); the mainboard of the edge computing unit (4) is fixed on the bottom plate through a copper column (412), and an ARM processor and a DDR4 memory particle are welded on the mainboard; The hardware carrier of the data preprocessing module (41) is an FPGA acceleration card inserted into a PCIe x4 slot of the mainboard, the acceleration card is connected with the mainboard through a gold finger, and the card is integrated with a high-speed ADC chip and a large-capacity FPGA chip; The hardware carrier of the feature extraction module (42) is a main processor and a coprocessor acceleration card, the acceleration card can be an ASIC chip, connected with the mainboard through a PCIe x8 slot, and the card has independent video memory or cache and parallel computing cores; The process of the edge computing function includes the following steps: Step one, hardware platform construction and heat management design: The network aggregation gateway (3) adopts a compact fanless design of an industrial computer case, and the aluminum casting body shell is formed into an integrated heat dissipation fin (411) structure through precision casting; the mainboard is mechanically fixed with the bottom plate through a copper column (412), and the copper column simultaneously serves as a heat conduction path to conduct the heat of the mainboard to the aluminum casting body shell; an ARM processor and a DDR4 memory particle are directly integrated on the mainboard through a BGA welding process, forming a minimized circuit path; Step two, data acquisition and preprocessing hardware acceleration: The FPGA acceleration card establishes a high-speed data transmission channel with the mainboard through the gold finger interface of the PCIe x4 slot; after the analog sensor signal is sampled and quantized by the high-speed ADC chip, it is directly transmitted to the FPGA chip through the on-board bus; the FPGA chip is internally programmed to realize a parallel pipeline architecture, and performs filtering, calibration and format conversion operations on multi-channel data through configurable logic units; Step three, feature extraction coprocessing acceleration: The ASIC coprocessing acceleration card obtains high-bandwidth connection through the PCIe x8 slot, and the independent video memory adopts GDDR6 particles to realize data caching; after the main processor sends a computing instruction, the parallel computing cores in the ASIC chip start feature extraction operations according to the hardened circuit structure, and realize convolution operations and feature dimension reduction through fixed logic circuits; Step four, edge intelligent decision and output: The ARM processor coordinates the operation results of the FPGA and the ASIC, executes a lightweight decision algorithm after completing data fusion in the DDR4 memory; finally, the control instruction is output through the industrial communication interface integrated on the mainboard, and the heat is continuously dissipated through the heat dissipation fins of the aluminum casting body shell, ensuring the long-term stable operation of the device in an industrial environment.

5. The apparatus of claim 1, wherein, The server cabinet of the cloud diagnosis center (5) is a standard 42U cabinet, and multiple blade servers and storage devices are installed in the cabinet; The deep learning diagnosis model (51) runs on a GPU server equipped with multiple GPU acceleration cards, which are interconnected through a high-speed InfiniBand network to form a computing cluster; each GPU acceleration card is vertically inserted into the backplane of the server through a PCIe Gen4 slot and is forcibly cooled through a cooling air duct; The feature-level fusion engine (52) is deployed on another group of high-performance CPU servers, which are connected to the GPU server cluster and storage devices through a 10 G Ethernet switch on the top of the cabinet; The diagnosis result generation module (53) is deployed on an application server responsible for Web service and database management; The GPU server cluster, CPU server group, application server, and storage device are all connected through a 10 G Ethernet switch arranged on the top of the cabinet.

6. A partial discharge on-line detection method based on multi-sensor data fusion and deep learning, characterized in that, The multi-sensor data fusion and deep learning-based partial discharge online detection device according to any one of claims 1-5 comprises the following steps: Step one, multi-source heterogeneous data synchronous acquisition and transmission: The mobile inspection node (1) travels along a preset path and locates, the support mechanism (12) adjusts the pose of the multi-sensor integrated cabin (13), the high-frequency sensor (131) and the ultrasonic sensor (132) synchronously acquire signals, and the current sensor (22) and the transient ground voltage sensor (23) of the distributed fixed monitoring node (2) continuously acquire data, all data are embedded with a unified time scale; Step two, monitoring data aggregation and preliminary preprocessing: The mobile inspection node (1) transmits data to the network aggregation gateway (3) through the first wireless communication link (L1), and the distributed fixed monitoring node (2) transmits data to the network aggregation gateway (3) through the second wired communication link (L2), and then transmits data to the edge computing unit (4) through the local area network (L3), and the data preprocessing module (41) performs synchronous sampling, quantization and filtering on the signals; Step three, edge-side feature extraction and local decision: The feature extraction module (42) extracts time-domain and frequency-domain features from the preprocessed multi-source signals, and the local fusion decision module (43) combines environmental parameters to perform feature correlation and threshold comparison to generate a local warning decision; Step four, cloud-side deep feature fusion and model diagnosis: The edge computing unit (4) uploads feature data and original data segments to the cloud-side diagnosis center (5) through the wide area network (L4), the feature-level fusion engine (52) performs multi-dimensional feature fusion, and the deep learning diagnosis model (51) performs pattern recognition and risk level evaluation on the fused features; Step five, diagnosis result generation and visualization presentation: The diagnosis result generation module (53) generates a structured diagnosis report and pushes it to a user terminal through a Web service interface for graphical display and risk point labeling; Step six, feedback optimization and model updating: The cloud diagnosis center (5) collects new sample data to retrain the deep learning diagnosis model (51), and issues the optimized model parameters to the edge computing unit (4).

7. The partial discharge on-line detection method based on multi-sensor data fusion and deep learning according to claim 6, characterized in that, In the third step, the edge side feature fusion decision algorithm based on information geometry and adaptive weighting includes: The manifold space mapping and representation of multi-source features is to regard the time-frequency domain feature vectors extracted by each sensor as points on a Riemann manifold, and to calculate a covariance matrix to represent the deep statistical characteristics of the signals. For a data window of a sensor s containing M feature points, the covariance matrix is defined as: ; in formula (1), is the feature vector at the mth time point, is the mean of the feature vectors within the window, , and D is the feature dimension. The adaptive decision fusion based on Riemann distance is to calculate the geometric distance between the current multi-sensor feature manifold and the preset health state reference manifold, and to perform adaptive weighted fusion; Comprehensive anomaly indicator Based on the log-Euclidean Riemannian distance calculation: ; in formula (2), the adaptive weight , is the gradient norm of the current window covariance matrix, is the sensitivity adjustment hyperparameter; is the reference covariance matrix of the sensor s in the healthy state, Logm(·) is the matrix logarithm operation, is the square of the Frobenius norm. 8.The partial discharge on-line detection method based on multi-sensor data fusion and deep learning according to claim 6, characterized in that, In the fourth step, the cloud deep diagnosis algorithm based on multi-modal Transformer and cross-modal contrast learning includes: The spatio-temporal alignment and embedding of multi-modal features is to use the Transformer encoder to jointly encode the heterogeneous atlas data and capture long-range dependencies; The input of the cross-modal Transformer is the splicing of all sensor features , is the feature vector of the ultra-high frequency sensor, is the feature vector of the ultrasonic sensor, is the feature vector of the high-frequency current transformer, is the feature vector of the transient earth voltage sensor, and the output of the self-attention mechanism is: ; in formula (3), is the query, key, and value matrix, and M is a specially designed modal mask matrix, indicates the dimension of each feature vector in the query matrix and the key matrix, indicates that the query matrix is multiplied by the transpose matrix of the key matrix, and Softmax(·) is an activation function; The decoupled representation and diagnosis based on contrast learning is to introduce a decoupled contrast loss in model training to separate discharge type features and environmental interference features; Decoupled contrastive loss function is defined as: ; in equation (4), is the target sample feature vector, is the positive sample set, is the negative sample set, is the similarity measure function, is the temperature hyperparameter, is the logarithm function. 9.The partial discharge on-line detection method based on multi-sensor data fusion and deep learning of claim 6, wherein, In the sixth step, the incremental model updating algorithm based on meta-learning and Bayesian optimization includes: The meta-gradient fast adaptation of model parameters is to use the model-agnostic meta-learning (MAML) framework to enable the model to quickly adapt based on a small number of new samples; The inner loop update and the outer loop update of meta-learning are as follows: In the formula (5) and the formula (6), is a meta parameter of the model, is the i-th task, is the parameter of the model on the task is a training loss, is a loss function is a gradient of the meta parameter, is an inner loop learning rate, is the parameter of the model after fast adaptation on the task is a test loss of the model on the task is a loss sum of a plurality of tasks sampled from a task distribution ​​​ The update strategy optimization based on Bayesian inference is to model the model update decision as a Bayesian optimization problem to balance exploration and utilization. updating the reward function ; in equation (7), the utility function , is the KL divergence, is the deployed old model distribution, is the quantized new model distribution, is a trade-off coefficient, is the candidate model parameter to be evaluated, is the current optimal parameter, is the expectation operator, is a non-negative truncation function.

Images