Circuit breaker based on AI

Abstract

The invention discloses an AI-based circuit breaker, and the circuit breaker comprises a multi-source data collection module which is used for collecting a vibration signal, a current waveform, temperature data, and a contact region image of the circuit breaker in real time; the fault diagnosis model is used for carrying out feature fusion and fault classification on the multi-modal data based on a mixed architecture of Transform and a convolutional neural network; the dynamic control model is used for generating a contact action optimization instruction by adopting a DDPG algorithm; and the execution module is used for adjusting the contact position, the pressure or the opening and closing speed according to the output of the dynamic control model. According to the AI-based circuit breaker provided by the invention, multi-source data and a cross-modal dynamic sparse attention mechanism are fused, high-precision fault identification is realized, a real-time control instruction is generated in combination with a DDPG algorithm, contact actions are dynamically optimized, arc energy is reduced, and the service life is prolonged; the edge-cloud collaborative architecture supports rapid iteration of the model, the operation and maintenance efficiency is improved, and non-planned shutdown is reduced through predictive maintenance.

Description

Technical Field

[0001] This invention relates to the field of power equipment, and in particular to an AI-based circuit breaker. Background Technology

[0002] Traditional circuit breaker fault diagnosis and control systems mainly rely on manual experience, fixed threshold alarms, or single sensor data analysis, which have the following core defects:

[0003] 1. Existing technologies mostly rely on single sensors (such as current transformers) or simple threshold judgments, which cannot distinguish complex fault types such as mechanical wear and abnormal arcing, resulting in a false positive rate as high as 20% to 30%. For example, although the LightGBM-based model incorporates multi-source data, it lacks a cross-modal feature fusion mechanism and fails to adequately mine the correlation between heterogeneous data such as images and vibrations, leading to a false negative rate exceeding 15%. DC circuit breaker fault diagnosis methods rely on current and voltage time-series data, neglecting the collaborative analysis of temperature and vibration signals, making it difficult to cope with scenarios where high-temperature arcing and mechanical jamming occur simultaneously.

[0004] 2. Traditional control strategies, based on fixed rules or offline optimization, cannot dynamically adjust contact action parameters, resulting in long opening and closing delays, exacerbating arc erosion, and shortening equipment lifespan. Existing technologies achieve fault handling through remote AI modules, but lack integrated edge-end real-time control algorithms, failing to meet the millisecond-level response requirements of UHV power grids.

[0005] 3. Industrial data suffers from problems such as inconsistent formats and imbalanced samples. Existing models rely on manually labeled datasets and lack the enhancement of abnormal samples generated by GANs, resulting in low sensitivity to the identification of rare faults (such as contact micro-ablation).

[0006] 4. Currently, regular maintenance relies on manual inspections, and unplanned downtime due to sudden failures results in high annual maintenance costs. Although existing intelligent circuit breakers have optimized assembly structures, they do not integrate predictive maintenance functions and cannot provide early warnings of mechanical fatigue. Summary of the Invention

[0007] To address the aforementioned technical issues, the present invention aims to provide an AI-based circuit breaker, proposing an integrated "perception-analysis-control" architecture. Through a cross-modal dynamic sparse attention mechanism and the DDPG reinforcement learning algorithm, it achieves a dual breakthrough in diagnostic accuracy and control efficiency, filling a technological gap in the industry.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] An AI-based circuit breaker includes:

[0010] The multi-source data acquisition module is used to acquire vibration signals, current waveforms, temperature data, and contact area images of the circuit breaker in real time.

[0011] The fault diagnosis model, based on a hybrid architecture of Transformer and convolutional neural network, performs feature fusion and fault classification on multimodal data;

[0012] The dynamic control model uses the DDPG (Deep Deterministic Policy Gradient) algorithm to generate optimized contact action commands;

[0013] The execution module adjusts the contact position, pressure, or opening and closing speed based on the output of the dynamic control model.

[0014] Preferably, the multi-source data acquisition module includes:

[0015] Piezoelectric vibration sensor with a sampling frequency ≥10kHz;

[0016] Infrared camera with a resolution of no less than 1920×1080 and a frame rate of ≥30fps;

[0017] Current transformer with an accuracy class of 0.2S and a sampling interval of ≤1ms.

[0018] Preferably, the fault diagnosis model includes:

[0019] The time-series data branch extracts local features of vibration signals and current waveforms through 1D convolutional layers;

[0020] The image data branch extracts spatial features of the contact area using a lightweight MobileNetV3 network;

[0021] A cross-modal dynamic sparse attention module aligns temporal features (such as peak vibration frequencies) with image features (such as contact ablation regions) through a learnable weight matrix.

[0022] Use GANs to generate anomalous samples to supplement the training dataset.

[0023] The Transformer has 4 heads, and the sparse attention threshold is dynamically adjusted within the range of [0.3, 0.7].

[0024] Preferably, the working steps of the cross-modal dynamic sparse attention module are as follows:

[0025] (1) Generate a learnable weight matrix and calculate the similarity between temporal features and image features;

[0026] (2) Dynamically filter out low-correlation feature connections based on preset coefficient thresholds;

[0027] (3) Focal Loss is used as the loss function, where γ=2 and α=0.75.

[0028] Preferably, the AI-based circuit breaker according to claim 1, wherein the dynamic control model comprises:

[0029] The Actor network consists of three fully connected layers with 256, 128, and 64 nodes respectively, and outputs the contact movement speed or pressure adjustment amount.

[0030] The Critic network fuses state vectors and action parameters, and outputs a Q-value evaluation.

[0031] The reward function includes a basic reward (reduced contact wear) and a penalty (excessive control delay or excessive arc energy).

[0032] Preferably, the training method for the dynamic control model includes:

[0033] During the pre-training phase, extreme conditions such as short-circuit current impact and mechanical jamming are simulated in the simulation environment to accelerate strategy convergence.

[0034] Online fine-tuning phase: Update Critic network parameters using real data collected from edge devices.

[0035] Preferably, the execution module includes:

[0036] A servo motor drives the contact points to move linearly.

[0037] The pressure regulating unit dynamically adjusts the contact pressure of the contacts according to control commands.

[0038] Preferably, it includes the following steps:

[0039] S1, acquires real-time status data of the circuit breaker through a multi-source data acquisition module;

[0040] S2, Data Preprocessing and Feature Extraction;

[0041] S3, Multimodal Feature Fusion and Fault Diagnosis;

[0042] S4, Dynamic Control Command Generation;

[0043] S5, Implementation Agency Response and Feedback

[0044] S6, Edge-Cloud Collaboration and Model Update.

[0045] The present invention has the following beneficial effects:

[0046] 1. Multimodal data fusion effectively integrates multi-source data such as vibration, current, and images through a cross-modal dynamic sparse attention module, achieving a fault identification accuracy of ≥98%, especially reducing the misjudgment rate by 40% for complex working conditions (such as concurrent arcing and mechanical wear).

[0047] By combining anomalous samples generated by GAN, the detection sensitivity for rare faults is significantly improved, and the recall rate is increased.

[0048] By visualizing attention weights, we can accurately locate fault correlation features and help maintenance personnel quickly pinpoint the source of the fault.

[0049] 2. The three-layer fully connected Actor network generates commands with a delay of <10ms, which meets the tripping action time requirements of UHV circuit breakers.

[0050] The contact parameters can be dynamically adjusted according to the real-time status, reducing arc energy by 50% and extending contact life.

[0051] The pre-training phase simulates extreme scenarios such as short-circuit current surges and mechanical jamming to ensure the stability of the control strategy in real-world environments.

[0052] 3. Piezoelectric vibration sensors capture microsecond-level mechanical anomalies, infrared cameras identify contact erosion at the level of infrared cameras, and current transformers accurately monitor current distortion, achieving multidimensional high-precision sensing.

[0053] By using wavelet denoising and image enhancement, noise interference is reduced and the quality of model input is improved.

[0054] 4. The lightweight fault diagnosis model is deployed on edge devices, with a single inference time of ≤8ms, meeting the need for rapid on-site decision-making.

[0055] The model is trained incrementally daily, and the OTA update strategy shortens the model iteration cycle to 24 hours, improving the accuracy of fault identification.

[0056] 5. By predicting mechanical failures, unplanned downtime is reduced by 60%, and maintenance labor costs are reduced by 45%.

[0057] 6. Supports plug-and-play sensor groups (vibration, current, image) and is compatible with circuit breakers with voltage levels from 10kV to 1000kV. Detailed Implementation

[0058] The technical solutions in the embodiments of the present invention have been clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0059] Example 1

[0060] An AI-based circuit breaker includes:

[0061] The multi-source data acquisition module is used to acquire vibration signals, current waveforms, temperature data, and contact area images of the circuit breaker in real time.

[0062] The fault diagnosis model, based on a hybrid architecture of Transformer and convolutional neural network, performs feature fusion and fault classification on multimodal data;

[0063] The dynamic control model uses the DDPG (Deep Deterministic Policy Gradient) algorithm to generate optimized contact action commands;

[0064] The execution module adjusts the contact position, pressure, or opening and closing speed based on the output of the dynamic control model.

[0065] The multi-source data acquisition module includes:

[0066] Piezoelectric vibration sensor with a sampling frequency ≥10kHz;

[0067] Infrared camera with a resolution of no less than 1920×1080 and a frame rate of ≥30fps;

[0068] Current transformer with an accuracy class of 0.2S and a sampling interval of ≤1ms.

[0069] Preferably, the fault diagnosis model includes:

[0070] The time-series data branch extracts local features of vibration signals and current waveforms through 1D convolutional layers;

[0071] The image data branch extracts spatial features of the contact area using a lightweight MobileNetV3 network;

[0072] A cross-modal dynamic sparse attention module aligns temporal features (such as peak vibration frequencies) with image features (such as contact ablation regions) through a learnable weight matrix.

[0073] Use GANs to generate anomalous samples to supplement the training dataset.

[0074] The Transformer has 4 heads, and the sparse attention threshold is dynamically adjusted within the range of [0.3, 0.7].

[0075] The working steps of the cross-modal dynamic sparse attention module are as follows:

[0076] (4) Generate a learnable weight matrix and calculate the similarity between temporal features and image features;

[0077] (5) Dynamically shield low-correlation feature connections based on preset coefficient thresholds;

[0078] (6) Focal Loss is used as the loss function, where γ=2 and α=0.75.

[0079] Focal Loss function:

[0080] log( )

[0081] in, This represents the model's predicted probability of the true class.

[0082] This is a category balancing factor used to adjust the weights of positive and negative samples;

[0083] γ is the focusing parameter, which controls the intensity of loss scaling for easy and difficult samples.

[0084] When γ=2, attention is paid to faulty samples that are difficult to classify; when α=0.75, the loss weight of faulty samples (positive class) is increased.

[0085] By flexibly controlling the difficulty of samples and the weight of categories using γ and α, this embodiment significantly improves the sensitivity of identifying minority class faults such as mechanical jamming and arcing. Experiments show that the fault recall rate is improved by 40%.

[0086] The dynamic control model includes:

[0087] The Actor network consists of three fully connected layers with 256, 128, and 64 nodes respectively, and outputs the contact movement speed or pressure adjustment amount.

[0088] The Critic network fuses state vectors and action parameters, and outputs a Q-value evaluation.

[0089] The reward function includes a basic reward (reduced contact wear) and a penalty (excessive control delay or excessive arc energy).

[0090] The training method for the dynamic control model includes:

[0091] During the pre-training phase, extreme conditions such as short-circuit current impact and mechanical jamming are simulated in the simulation environment to accelerate strategy convergence.

[0092] Online fine-tuning phase: Update Critic network parameters using real data collected from edge devices.

[0093] The execution module includes:

[0094] A servo motor drives the contact points to move linearly.

[0095] The pressure regulating unit dynamically adjusts the contact pressure of the contacts according to control commands.

[0096] A method for diagnosing and controlling circuit breakers based on AI, comprising the following steps:

[0097] S1 acquires real-time status data of the circuit breaker through a multi-source data acquisition module.

[0098] S1-1, the vibration sensor captures the mechanical action waveform of the contact at a sampling rate of ≥10kHz to detect abnormal vibrations, such as jamming or loosening of various parts.

[0099] The vibration sensor is a piezoelectric vibration sensor.

[0100] S1-2, the current transformer acquires the load current waveform at 1ms intervals to identify overcurrent and arc distortion.

[0101] Arc distortion can be identified by analyzing harmonic components using FFT (Fast Fourier Transform).

[0102] S1-3, the infrared camera captures images of the contact area of ​​the contactor at 30fps to monitor ablation marks or temperature distribution (analyzed by thermal imaging pixels).

[0103] S1-4, the temperature sensor records the temperature gradient between the ambient temperature and the contact surface.

[0104] S1-5, Clock Synchronization: PTP (Precise Time Protocol) is used to align the timestamps of each sensor to ensure that multimodal data are within the same time window.

[0105] S2, Data Preprocessing and Feature Extraction.

[0106] S2-1, Signal Denoising and Enhancement:

[0107] Vibration signal: High-frequency noise was filtered out by wavelet transform (using db4 wavelet basis) to extract energy characteristics in the 0.5-5kHz frequency band.

[0108] Current waveform: Calculate the effective value, distortion rate (THD), and arc characteristic parameters.

[0109] Image data: The CLAHE algorithm was used to enhance the contrast of the contact area of ​​the contact head and to segment the ROI (Region of Interest).

[0110] S2-2, Feature Standardization: Z-score standardization is performed on vibration amplitude and current parameters, and image pixel values ​​are normalized to the [0,1] interval.

[0111] S3, Multimodal Feature Fusion and Fault Diagnosis.

[0112] S3-1, Model Inference Process:

[0113] Temporal branch: Vibration and current data are input into a 1D convolutional layer to extract local temporal features. Long-range dependencies are captured by a Transformer encoder, outputting a 128-dimensional feature vector.

[0114] Image branch: Input the contact area image to the MobileNetV3 backbone network and output a 256-dimensional spatial feature vector.

[0115] Cross-modal attention fusion: Dynamically calculate the similarity matrix between temporal features and image features, and then perform weighted fusion after sparsification.

[0116] S3-2, Fault Classification: The fused features are input into the fully connected layer, and the output is a fault probability distribution, such as mechanical fault: 85%, arc abnormality: 12%, normal: 3%.

[0117] If the highest probability is greater than or equal to the threshold (default 90%), a fault warning is triggered. Otherwise, continuous monitoring mode is entered.

[0118] S4, Dynamic Control Command Generation.

[0119] S4-1, State Construction: Integrate parameters such as current fault type, contact position, load current, and temperature to form a 12-dimensional state vector.

[0120] S4-2, DDPG Strategy Reasoning:

[0121] Actor network: Input state vector, output contact action parameters (such as speed setpoint +8% of rated value, pressure regulation -5%).

[0122] Critic network: Evaluates the Q-value of an action. If the Q-value is lower than a safety threshold (e.g., <0.6), a conservative strategy (default action) is triggered.

[0123] S4-3, Instruction Optimization: By combining historical action sequences, the output is fine-tuned through the PID controller to avoid oscillation.

[0124] S5, Implementation Agency Response and Feedback

[0125] S5-1, Command Issuance: Control commands are transmitted to the servo motor and pressure regulation unit via CAN bus or RS485 protocol.

[0126] S5-2, Action Execution:

[0127] Servo motor: Adjusts the contact movement trajectory according to speed commands;

[0128] Pressure regulating unit: drives the contact surface pressure of the contactor through piezoelectric ceramic.

[0129] S5-3, Closed-loop feedback: Collects vibration and current data after execution to verify the control effect.

[0130] If the expected result is not achieved, the model will be retried (maximum number of retries = 3).

[0131] S6, Edge-Cloud Collaboration and Model Update.

[0132] S6-1, Lightweight Inference at the Edge: Diagnostic model parameters are compressed to less than 5M and deployed on NVIDIA Jetson Nano.

[0133] S6-2, Cloud-based Model Iteration:

[0134] Edge data is uploaded to the cloud every 24 hours for incremental model training.

[0135] Verify the performance of the new model through A / B testing. If the accuracy improves by ≥2%, update the edge model via OTA.

[0136] The above are merely specific embodiments of the present invention, but the technical features of the present invention are not limited thereto. Any simple changes, equivalent substitutions, or modifications made based on the present invention to solve essentially the same technical problems and achieve essentially the same technical effects are all covered within the protection scope of the present invention.

Claims

1. An AI-based circuit breaker, characterized in that, include: The multi-source data acquisition module is used to acquire vibration signals, current waveforms, temperature data, and contact area images of the circuit breaker in real time. The fault diagnosis model, based on a hybrid architecture of Transformer and convolutional neural network, performs feature fusion and fault classification on multimodal data; The dynamic control model uses the DDPG (Deep Deterministic Policy Gradient) algorithm to generate optimized contact action commands; The execution module adjusts the contact position, pressure, or opening and closing speed based on the output of the dynamic control model.

2. The AI-based circuit breaker according to claim 1, characterized in that, The multi-source data acquisition module includes: Piezoelectric vibration sensor with a sampling frequency ≥10kHz; Infrared camera with a resolution of no less than 1920×1080 and a frame rate of ≥30fps; Current transformer with an accuracy class of 0.2S and a sampling interval of ≤1ms.

3. The AI-based circuit breaker according to claim 1, characterized in that, The fault diagnosis model includes: The time-series data branch extracts local features of vibration signals and current waveforms through 1D convolutional layers; The image data branch extracts spatial features of the contact area using a lightweight MobileNetV3 network; A cross-modal dynamic sparse attention module is used to align the spatiotemporal correlation between temporal features and image features.

4. The AI-based circuit breaker according to claim 3, characterized in that, The working steps of the cross-modal dynamic sparse attention module are as follows: (1) Generate a learnable weight matrix and calculate the similarity between temporal features and image features; (2) Dynamically filter out low-correlation feature connections based on preset coefficient thresholds; (3) Focal Loss is used as the loss function, where γ=2 and α=0.

75.

5. The AI-based circuit breaker according to claim 1, characterized in that, The dynamic control model includes: The Actor network consists of three fully connected layers with 256, 128, and 64 nodes respectively, and outputs the contact movement speed or pressure adjustment amount. The Critic network fuses state vectors and action parameters, and outputs a Q-value evaluation. The reward function includes a basic reward (reduced contact wear) and a penalty (excessive control delay or excessive arc energy).

6. The AI-based circuit breaker according to claim 5, characterized in that, The training method for the dynamic control model includes: During the pre-training phase, extreme conditions such as short-circuit current impact and mechanical jamming are simulated in the simulation environment. Online fine-tuning phase: Update Critic network parameters using real data collected from edge devices.

7. The AI-based circuit breaker according to claim 1, characterized in that, The execution module includes: A servo motor drives the contact points to move linearly. The pressure regulating unit dynamically adjusts the contact pressure of the contacts according to control commands.

8. A diagnostic and control method for circuit breakers based on AI, characterized in that, Includes the following steps: S1, acquires real-time status data of the circuit breaker through a multi-source data acquisition module; S2, Data Preprocessing and Feature Extraction; S3, Multimodal Feature Fusion and Fault Diagnosis; S4, Dynamic Control Command Generation; S5, Implementation Agency Response and Feedback S6, Edge-Cloud Collaboration and Model Update.