An insulation fault edge self-healing method and apparatus for a power conversion system

Abstract

The application discloses an insulation fault edge self-healing method and device of a power conversion system, belongs to the technical field of fault diagnosis and automatic maintenance of power electronic equipment, and comprises the following steps: acquiring multi-modal sensing data of the power conversion system collected synchronously through a multi-modal sensor array; processing the multi-modal sensing data by using a preset light-weight CNN-Transformer hybrid model to obtain characteristic vectors of each mode; fusing the characteristic vectors of each mode by using a dynamic weight attention mechanism to obtain a fused fault characteristic vector; classifying the fused fault characteristic vector based on a fault classifier to determine a fault grade; and sending a corresponding control instruction to an executing mechanism according to the determined fault grade to trigger a graded self-healing action. The application can quickly complete fault identification and self-healing decision, and effectively improves the accuracy of fault detection.

Description

Technical Field

[0001] This invention relates to a method and apparatus for self-healing of insulation fault edges in a power conversion system, belonging to the field of fault diagnosis and automatic maintenance technology for power electronic equipment. Background Technology

[0002] As a core component of new energy power generation and energy storage systems, the insulation performance of the power conversion system (PCS) directly affects system safety. Existing technologies have the following shortcomings: 1. Traditional threshold detection methods are not sensitive to early insulation degradation, with a false alarm rate as high as 30% or more; 2. Cloud-based intelligent diagnostics suffer from communication delays exceeding 200ms, failing to meet real-time requirements; 3. Single-mode detection methods (such as leakage current monitoring) are difficult to handle multi-factor coupled faults under complex operating conditions; 4. Existing edge computing devices are limited by memory capacity of less than 10MB, making it impossible to deploy conventional AI models. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and device for self-healing insulation fault edges in a power conversion system, which can quickly complete fault identification and self-healing decision-making, and effectively improve the accuracy of fault detection.

[0004] To achieve the above objectives, the present invention is implemented using the following technical solution: In a first aspect, the present invention provides a method for self-healing insulation fault edges in a power conversion system, comprising: Acquire multimodal sensing data of the power conversion system synchronously acquired through a multimodal sensor array; The multimodal sensing data is processed using a preset lightweight CNN-Transformer hybrid model to obtain feature vectors for each modality. A dynamic weighted attention mechanism is used to fuse the feature vectors of each modality to obtain a fused fault feature vector; The fused fault feature vector is classified based on the fault classifier to determine the fault level; Based on the determined fault level, the corresponding control command is sent to the actuator to trigger the graded self-healing action.

[0005] Furthermore, the multimodal sensing data includes leakage current time-frequency characteristics, dielectric loss angle, partial discharge ultrasonic signals, and thermal distribution images.

[0006] Furthermore, the dynamic weighted attention mechanism is used to fuse the feature vectors of each modality to obtain the fused fault feature vector, as shown in the following formula: ; in, For the normalized first Dynamic weights of each modality This is a fault state vector driven by the rate of change of insulation resistance. , These are the feature vectors for each modality. Specifically, it is the key vector. Specifically, it is a value vector. The fused fault feature vector It is time series data. For feature dimension, This represents the total number of modes.

[0007] Furthermore, the graded self-healing action includes: When the decision is a Level 1 fault, the partial discharge suppression circuit is triggered. When the fault is classified as a level 2 fault, the IGBT drive pulse width is dynamically adjusted to limit the fault current. When the fault is classified as Level 3, a command is sent to disconnect the DC bus within a set time.

[0008] Furthermore, the lightweight CNN-Transformer hybrid model uses MobileNetV3 as the CNN backbone network to process the time-frequency features of leakage current and thermal distribution images to extract local spatial features; it uses a lightweight Transformer encoder to process dielectric loss angle and partial discharge ultrasonic signals to capture their temporal correlation features; the lightweight CNN-Transformer hybrid model introduces knowledge distillation technology during the model training stage, injects adversarial examples into the training set for robust training, and adopts a dual-objective optimization algorithm that combines floating-point computation and inference latency to search and optimize the model structure.

[0009] Furthermore, the method also includes: deploying a lightweight CNN-Transformer hybrid model on the edge node, storing historical fault data on the cloud server, and periodically updating the model parameters. The edge node and the cloud server transmit model parameter increments through a bidirectional HTTPS encrypted channel. When the edge node detects a level 3 fault, it automatically triggers a cloud warning and uploads complete fault waveform data.

[0010] Furthermore, the specific implementation of the dynamic weighted attention mechanism includes: The feature vectors of each modality are input into a pre-constructed cross-modal feature interaction matrix; Based on the insulation resistance change rate and fault development trend prediction values ​​obtained from real-time monitoring of the power conversion system, the weight coefficients of the feature vectors of each mode in the fusion formula are dynamically calculated and adjusted. The calculated modal weight coefficients are normalized using the Softmax normalization function to achieve adaptive fusion at the feature level. When the insulation resistance value is detected to be less than the set danger threshold, the fusion weight corresponding to the partial discharge ultrasonic signal is actively increased in the weight calculation.

[0011] Furthermore, the determination of the fault level is based on an insulation resistance threshold-based triggering mechanism, including: A preset third-level insulation resistance threshold is used as the decision boundary: Normal operating threshold R1: ≥20MΩ; Warning threshold R2: 10MΩ≤R<20MΩ; Danger threshold R3: R < 10 MΩ; When the insulation resistance R < R3 is detected, the preparatory commands corresponding to the first, second and third level faults are triggered in parallel.

[0012] Furthermore, the method also includes anti-interference protection measures, specifically: high-frequency electromagnetic noise suppression is performed on the collected multimodal sensing data, and a sliding window impulse noise filtering algorithm is used for real-time filtering; and before executing the control command, the legality of the control command and the operation permission are double-verified by setting a timeout threshold.

[0013] In a second aspect, the present invention provides an insulation fault edge self-healing device for a power conversion system, comprising: The data acquisition module is used to acquire multimodal sensing data of the power conversion system synchronously collected by the multimodal sensor array; The model inference module is used to process the multimodal sensing data using a preset lightweight CNN-Transformer hybrid model to obtain feature vectors for each modality. The feature fusion module is used to fuse the feature vectors of each modality using a dynamic weight attention mechanism to obtain a fused fault feature vector; The feature classification module is used to classify the fused fault feature vector based on the fault classifier to determine the fault level; The graded self-healing module is used to send corresponding control commands to the actuator based on the determined fault level, thereby triggering graded self-healing actions.

[0014] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods described above.

[0015] Fourthly, the present invention provides an electronic device, comprising: Memory, used to store computer programs / instructions; A processor for executing the computer program / instructions to implement the steps of any of the methods described above.

[0016] Fifthly, the present invention provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of any of the methods described above.

[0017] Compared with the prior art, the beneficial effects achieved by the present invention are as follows: This invention provides a method and device for edge self-healing of insulation faults in power conversion systems. Through a three-stage processing approach—multimodal sensor data acquisition, fusion analysis using a lightweight CNN-Transformer hybrid model, and hierarchical rapid self-healing control—microsecond-level fault diagnosis and millisecond-level autonomous recovery are achieved on resource-constrained edge computing devices. An innovative dynamic weighted attention fusion mechanism is proposed, solving the problems of high latency and numerous false alarms associated with traditional methods. Experimental results show that this method can shorten the insulation fault recovery time to less than 50ms, ensuring the safe and reliable operation of power electronic equipment. Attached Figure Description

[0018] Figure 1 This is a flowchart of the self-healing method for insulation fault edges in a power conversion system provided in an embodiment of the present invention; Figure 2 This is a system architecture diagram of the insulation fault edge self-healing method for power conversion systems provided in this embodiment of the invention. Detailed Implementation

[0019] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0020] like Figure 1 As shown in Example 1, this example introduces a self-healing method for insulation fault edges in a power conversion system, including: Acquire multimodal sensing data of the power conversion system synchronously acquired through a multimodal sensor array; The multimodal sensing data is processed using a preset lightweight CNN-Transformer hybrid model to obtain feature vectors for each modality. A dynamic weighted attention mechanism is used to fuse the feature vectors of each modality to obtain a fused fault feature vector; The fused fault feature vector is classified based on the fault classifier to determine the fault level; Based on the determined fault level, the corresponding control command is sent to the actuator to trigger the graded self-healing action.

[0021] like Figure 1 and Figure 2 As shown in the figure, the self-healing method for insulation fault edges in the power conversion system provided in this embodiment involves the following steps in its application process: Step 1: Simultaneously acquire the time-frequency characteristics of leakage current, dielectric loss angle, partial discharge ultrasonic signal and thermal distribution image of the power conversion system (PCS) through a multimodal sensor array to form an initial multimodal sensing dataset; The synchronous acquisition step also includes adaptive adjustment of the sampling frequency: under normal operating conditions determined based on historical data, a base sampling frequency of 1kHz is used; when the lightweight hybrid machine learning model or insulation resistance monitoring module identifies an abnormal trend, the sampling frequency is automatically increased to 100kHz, and the high-frequency data acquired after this switch is used as the transmission object in step b.

[0022] Step 2: The initial multimodal sensing dataset obtained in Step 1 is transmitted to a lightweight hybrid machine learning model deployed on a local edge computing SoC chip via a high-speed SPI bus. Step 3: On the edge computing SoC chip, the initial multimodal sensing dataset transmitted in Step 2 is processed in parallel using a lightweight hybrid machine learning model, and local spatial feature vectors and temporal correlation feature vectors corresponding to the electrical, acoustic and thermal modes are output respectively. The lightweight hybrid machine learning model is a CNN-Transformer hybrid model, which is adapted to edge computing resources through the following compression process: Depthwise separable convolutions are used to replace standard convolutional layers to reduce model parameters and computational cost; Implement structured parameter pruning to make the model weight sparsity greater than 70%; By employing hybrid 8-bit and 4-bit fixed-point quantization techniques, the model storage space and inference power consumption are reduced.

[0023] The construction and training of the lightweight CNN-Transformer hybrid model includes: MobileNetV3 is used as the backbone network of the CNN, specifically for processing the received leakage current signal and thermal distribution image to extract its local spatial features. A lightweight Transformer encoder with a head count compressed to 2 is used to process the received dielectric loss angle sequence data and partial discharge ultrasonic signals to capture their temporal correlation characteristics. During the model training phase, a knowledge distillation technique using a full-precision model as the teacher model is introduced, and adversarial examples are injected into the training set for robust training to improve the generalization ability of the lightweight model to extract features from diverse data. A dual-objective optimization algorithm combining FLOPs (floating-point operations) and inference latency is employed to search and optimize the model structure in order to balance the processing complexity after receiving data with the computational efficiency on edge devices.

[0024] Step 4: Input the modal feature vectors output in Step 3 into a dynamic weighted attention fusion module. This module generates dynamic weights for each modal feature based on the real-time calculated insulation resistance change rate, and performs weighted fusion according to the following formula to output a unified, cross-modal fault feature vector Z: ; in, The normalized first Dynamic weights of each modality This is a fault state vector driven by the rate of change of insulation resistance. , These are the feature vectors for each modality. Specifically, it is the key vector. Specifically, it is a value vector. The fused fault feature vector It is time series data. For feature dimension, The total number of modes; The specific implementation of the dynamic weighted attention mechanism includes: Construct a cross-modal feature interaction matrix, the input of which comes from the feature vectors of all modalities output; Based on the insulation resistance change rate and fault development trend prediction values ​​obtained from real-time monitoring of the PCS system, the weighting coefficients of each modal feature in the fusion formula are dynamically calculated and adjusted. The calculated modal weight coefficients are normalized using the Softmax normalization function to achieve adaptive fusion at the feature level. By setting specific trigger conditions, when a sudden change in insulation resistance value R < 10MΩ is detected, the fusion weight corresponding to the partial discharge ultrasonic signal is actively increased in the weight calculation to enhance the sensitivity to potential discharge faults.

[0025] Step 5: Input the unified fault feature vector Z obtained in Step 4 into a fault classifier, which maps the feature vector Z to a specific fault level. The fault level decision is based on an insulation resistance threshold-based triggering mechanism, which includes: A preset third-level insulation resistance threshold is used as the decision boundary: R1 (normal operating threshold): ≥20MΩ; R2 (warning threshold): 10MΩ≤R<20MΩ; R3 (hazard threshold): R < 10 MΩ; When the insulation resistance R is calculated to be lower than R3 based on the sensor data, the decision-making step will trigger the preparatory instructions corresponding to the first, second and third level faults in parallel, and finally execute the highest level action according to the classifier output. The thresholds at each level can be dynamically calibrated according to the real-time operating conditions of the PCS, including a load rate greater than 80% or an ambient temperature higher than 60°C, with a calibration cycle of no more than 1 hour.

[0026] Step 6: Based on the fault level determined in Step 5, send the corresponding control command to the actuator via the CAN bus to trigger the graded self-healing action; the graded self-healing action includes: When the decision is a Level 1 fault, the partial discharge suppression circuit is triggered. When the fault is classified as a level 2 fault, the IGBT drive pulse width is dynamically adjusted to limit the fault current. When the decision is a level 3 fault, a command is sent to disconnect the DC bus within 0.5 milliseconds.

[0027] In a further embodiment, the method further includes a collaborative data processing step between the edge and the cloud: Deploy lightweight inference models at edge nodes to achieve millisecond-level real-time self-healing control; Historical fault data (including multimodal features and self-healing results) is stored on a cloud server, and model parameters are updated regularly. The incremental model parameters are transmitted between edge nodes and the cloud via a two-way HTTPS encrypted channel to ensure data security; When an edge node detects a Level 3 fault, it automatically triggers a cloud-based alert and uploads complete fault waveform data (including timestamps and multimodal features).

[0028] In a further embodiment, the method also includes anti-interference protection measures: At the signal acquisition end, an EMC filter circuit with a cutoff frequency of 10kHz is integrated to suppress high-frequency electromagnetic noise in the raw signal acquired step by step. In the model inference module, the data transmitted to the edge computing SoC via the SPI bus is filtered in real time using a sliding window impulse noise filtering algorithm with a window length of 50ms to eliminate data anomalies caused by transient electromagnetic interference. At the control end, before sending the command to cut off the DC bus, a hardware watchdog circuit performs dual verification of the command's legality and operation authorization to prevent false triggering due to data interference.

[0029] In a further embodiment, an apparatus is provided according to any one of the foregoing methods, comprising: Multimodal sensor array: includes Rogowski coil (for measuring leakage current), dielectric loss tester (for detecting dielectric loss angle), ultrasonic microphone array (for acquiring partial discharge acoustic signals), and infrared thermal imager (for acquiring thermal distribution images); Edge computing SoC chip: integrates TinyML inference engine (supports INT8 / INT4 quantization operation), communicates with sensor array via SPI bus (communication rate ≥10Mbps), and controls DC bus relay group via CAN bus (response time ≤1ms). Protection circuit module: includes IGBT drive protection circuit (overcurrent / overvoltage protection threshold is configurable) and power management module (supports 3.3V-5V wide range industrial power supply, voltage fluctuation tolerance ±10%). Communication module: integrates 5G / Ethernet interface, supports bidirectional encrypted communication with cloud server (encryption algorithm AES-256).

[0030] This embodiment provides a method and device for edge self-healing of insulation faults in a power conversion system. Through a three-stage processing approach—multimodal sensor data acquisition, fusion analysis using a lightweight CNN-Transformer hybrid model, and hierarchical fast self-healing control—microsecond-level fault diagnosis and millisecond-level autonomous recovery are achieved on resource-constrained edge computing devices. An innovative dynamic weighted attention fusion mechanism is proposed, solving the problems of high latency and numerous false alarms associated with traditional methods. Experimental results show that this method can shorten the insulation fault recovery time to less than 50ms, ensuring the safe and reliable operation of power electronic equipment.

[0031] Example 2: This example provides a self-healing device for insulation fault edges in a power conversion system, comprising: The data acquisition module is used to acquire multimodal sensing data of the power conversion system synchronously collected by the multimodal sensor array; The model inference module is used to process the multimodal sensing data using a preset lightweight CNN-Transformer hybrid model to obtain feature vectors for each modality. The feature fusion module is used to fuse the feature vectors of each modality using a dynamic weight attention mechanism to obtain a fused fault feature vector; The feature classification module is used to classify the fused fault feature vector based on the fault classifier to determine the fault level; The graded self-healing module is used to send corresponding control commands to the actuator based on the determined fault level, thereby triggering graded self-healing actions.

[0032] The specific functions of each module described above are explained in the relevant content of the method in Embodiment 1, and will not be repeated here.

[0033] Example 3: This example provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods described in Example 1.

[0034] Example 4: This example provides an electronic device, including: Memory, used to store computer programs / instructions; A processor for executing the computer program / instructions to implement the steps of any of the methods described in Embodiment 1.

[0035] Example 5: This example provides a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of the method described in any one of Examples 1.

[0036] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

[0037] Those skilled in the art will understand that embodiments of this disclosure can be provided as methods, systems, or computer program products. Therefore, this disclosure can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this disclosure can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0038] This disclosure is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0039] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0040] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0041] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this disclosure and not to limit its protection scope. Although this disclosure has been described in detail with reference to the above embodiments, those skilled in the art should understand that after reading this disclosure, they can still make various changes, modifications or equivalent substitutions to the specific implementation of the invention, but these changes, modifications or equivalent substitutions are all within the protection scope of the pending claims.

Claims

1. A method for self-healing insulation fault edges in a power conversion system, characterized in that, include: Acquire multimodal sensing data of the power conversion system synchronously acquired through a multimodal sensor array; The multimodal sensing data is processed using a preset lightweight CNN-Transformer hybrid model to obtain feature vectors for each modality. A dynamic weighted attention mechanism is used to fuse the feature vectors of each modality to obtain the fused fault feature vector, as shown in the following formula: in, The normalized first Dynamic weights of each modality This is a fault state vector driven by the rate of change of insulation resistance. , These are the feature vectors for each modality. Specifically, it is the key vector. Specifically, it is a value vector. The fused fault feature vector It is time series data. For feature dimension, The total number of modes; The fused fault feature vector is classified based on the fault classifier to determine the fault level; Based on the determined fault level, a corresponding control command is sent to the actuator to trigger a graded self-healing action; The lightweight CNN-Transformer hybrid model uses MobileNetV3 as the CNN backbone network to process the time-frequency features of leakage current and thermal distribution images to extract local spatial features. A lightweight Transformer encoder is used to process the dielectric loss angle and partial discharge ultrasonic signals to capture their temporal correlation features. During the model training phase, the lightweight CNN-Transformer hybrid model introduces knowledge distillation technology, injecting adversarial examples into the training set for robust training. A dual-objective optimization algorithm combining floating-point computation and inference latency is employed to search and optimize the model structure. The determination of the fault level is based on an insulation resistance threshold-based triggering mechanism, including: A preset third-level insulation resistance threshold is used as the decision boundary: Normal operating threshold R1: ≥20MΩ; Warning threshold R2: 10MΩ≤R<20MΩ; Danger threshold R3: R < 10 MΩ; When the insulation resistance R < R3 is detected, the preparatory commands corresponding to the first, second and third level faults are triggered in parallel.

2. The self-healing method for insulation fault edges in a power conversion system according to claim 1, characterized in that, The multimodal sensing data includes leakage current time-frequency characteristics, dielectric loss angle, partial discharge ultrasonic signals, and thermal distribution images.

3. The self-healing method for insulation fault edges in a power conversion system according to claim 1, characterized in that, The graded self-healing process includes: When the decision is a Level 1 fault, the partial discharge suppression circuit is triggered. When the fault is classified as a level 2 fault, the IGBT drive pulse width is dynamically adjusted to limit the fault current. When the fault is classified as Level 3, a command is sent to disconnect the DC bus within a set time.

4. The self-healing method for insulation fault edges in a power conversion system according to claim 1, characterized in that, The method further includes: deploying a lightweight CNN-Transformer hybrid model on edge nodes, storing historical fault data on a cloud server, and periodically updating model parameters. The edge nodes and the cloud server transmit model parameter increments through a bidirectional HTTPS encrypted channel. When an edge node detects a level 3 fault, it automatically triggers a cloud warning and uploads complete fault waveform data.

5. The self-healing method for insulation fault edges in a power conversion system according to claim 1, characterized in that, The specific implementation of the dynamic weighted attention mechanism includes: The feature vectors of each modality are input into a pre-constructed cross-modal feature interaction matrix; Based on the insulation resistance change rate and fault development trend prediction values ​​obtained from real-time monitoring of the power conversion system, the weight coefficients of the feature vectors of each mode in the fusion formula are dynamically calculated and adjusted. The calculated modal weight coefficients are normalized using the Softmax normalization function to achieve adaptive fusion at the feature level. When the insulation resistance value is detected to be less than the set danger threshold, the fusion weight corresponding to the partial discharge ultrasonic signal is actively increased in the weight calculation.

6. The self-healing method for insulation fault edges in a power conversion system according to claim 1, characterized in that, The method also includes anti-interference protection measures, specifically: high-frequency electromagnetic noise suppression is performed on the collected multimodal sensing data, and real-time filtering is performed using a sliding window pulse noise filtering algorithm; In addition, before executing control commands, a timeout threshold is set to perform dual verification of the legality of the control commands and the operation permissions.

7. An insulation fault edge self-healing device for a power conversion system, used to implement the insulation fault edge self-healing method for the power conversion system according to any one of claims 1-6, characterized in that, include: The data acquisition module is used to acquire multimodal sensing data of the power conversion system synchronously collected by the multimodal sensor array; The model inference module is used to process the multimodal sensing data using a preset lightweight CNN-Transformer hybrid model to obtain feature vectors for each modality. The feature fusion module is used to fuse the feature vectors of each modality using a dynamic weight attention mechanism to obtain a fused fault feature vector; The feature classification module is used to classify the fused fault feature vector based on the fault classifier to determine the fault level; The graded self-healing module is used to send corresponding control commands to the actuator based on the determined fault level, thereby triggering graded self-healing actions.

Images