Arc fault recognition method and device based on multi-physical field fusion

Abstract

The application discloses an arc fault recognition method and device based on multi-physical field fusion, comprising the steps of synchronously collecting current waveform signals, electromagnetic radiation signals and ultraviolet light signals; extracting multi-dimensional features, adaptively weighting fusion according to real-time signal quality, and obtaining a comprehensive feature vector; inputting the time sequence of the comprehensive feature vector into a CNN-LSTM hybrid neural network model to recognize the arc fault type; grading and breaking according to the recognition result, parallel arc zero arc breaking, series arc regular circuit breaker breaking, and normal working arc not breaking. By synchronously perceiving the current waveform, electromagnetic radiation and ultraviolet light three physical fields into adaptive weighted fusion of multi-dimensional features, and dynamically adjusting the contribution weight of each physical field based on the time-varying confidence coefficient of real-time signal quality, the inherent defects of the existing single current detection scheme, such as difficulty in recognizing parallel arcs, high misjudgment rate in variable frequency load scenes, are overcome, which is beneficial to improving the arc recognition accuracy and reducing the misjudgment rate.

Description

Technical Field

[0001] This invention relates to the field of low-voltage power distribution safety technology, and in particular to an arc fault identification method and device based on multi-physics field fusion. Background Technology

[0002] Arc faults are one of the leading causes of electrical fires. Statistics show that approximately 60% of electrical fires in my country each year are caused by arc faults. Arc faults are mainly classified into two types based on their morphology: series arcs and parallel arcs. Series arcs are typically caused by poor line contact, broken wires, etc. The current amplitude generally does not exceed the rated current, but the sustained arc temperature can reach over 3000℃, easily igniting surrounding insulation materials and combustibles. They are characterized by strong concealment, slow development, but significant and continuous damage. Parallel arcs are typically caused by insulation damage, short circuits between live and neutral wires, etc. The current amplitude can reach thousands of amperes, releasing enormous energy in a very short time. They are characterized by an extremely high current rise rate and concentrated energy release, placing extremely high demands on breaking speed.

[0003] However, existing arc fault circuit breakers are primarily designed based on UL1699 and IEC62606 standards, relying on a single current waveform detection principle. They determine the presence of an arc fault by identifying distortion characteristics in the current waveform (such as prolonged zero-point dwell time, increased harmonic content, and high-frequency noise components), making it difficult to balance accuracy, response speed, and the rationality of the protection strategy. This is because single current detection struggles to identify parallel arcs, whose current waveforms are highly similar to normal short-circuit currents. Furthermore, the arc plasma channel emits strong electromagnetic radiation and ultraviolet light, characteristics absent in pure metallic short circuits, leading to the inability to promptly interrupt parallel arc faults. Moreover, the current waveforms generated during normal operation of equipment such as electric drills, welding machines, vacuum cleaners, and variable frequency air conditioners are similar to those of fault arcs, both containing zero-point distortion and high-frequency harmonic components. This causes frequent malfunctions in existing arc fault circuit breakers, with some models experiencing false alarm rates exceeding 30% in scenarios with variable frequency loads, severely impacting product usability. Furthermore, the breaking time of existing arc fault circuit breakers is generally over 10ms. For high-energy, rapidly developing fault types such as parallel arcs, delayed breaking is sufficient to cause insulation carbonization, line melting, or even ignition of surrounding combustibles. Theoretically, the energy released by a parallel arc within 0.5ms can bring the PVC insulation layer to its thermal decomposition temperature. At the same time, it can only detect the existence of arc faults but cannot determine the specific location of the fault. In large-area low-voltage distribution networks, maintenance personnel need to inspect the lines section by section, which seriously increases maintenance costs and troubleshooting time. Using mechanical circuit breakers to break all arc faults in a uniform manner cannot implement differentiated breaking strategies based on the type and degree of danger of the arc. Using fast zero-arc breaking for series arcs with lower danger would increase the cost and complexity of the devices, while using conventional mechanical circuit breakers to break for parallel arcs with extremely high danger would not meet the speed requirements. Summary of the Invention

[0004] The technical problem solved by this invention is: to provide an arc fault protection method and device that can simultaneously and accurately identify series arcs and parallel arcs, has a low false alarm rate, fast breaking speed, fault location capability, and supports a graded breaking strategy, thereby improving the reliability and safety of arc fault protection.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: Firstly, an arc fault identification method based on multi-physics field fusion, comprising: Step S1: Synchronously acquire current waveform signals, electromagnetic radiation signals, and ultraviolet light signals; Step S2: Extract multi-dimensional features, and adaptively weight and fuse them according to real-time signal quality to obtain a comprehensive feature vector; Step S3: Input the time sequence of the comprehensive feature vector into the CNN-LSTM hybrid neural network model to identify the arc fault type; Step S4: According to the identification results, the circuit is divided into different levels: parallel arcs are broken by zero arc, series arcs are broken by conventional circuit breakers, and normal operating arcs are not broken. Step S5: After the fault is broken, the fault location is calculated by co-location of time-domain reflection and electromagnetic radiation attenuation.

[0006] As a preferred embodiment of the arc fault identification method based on multi-physics field fusion described in this invention, the adaptive weighted fusion adopts a time-varying confidence coefficient based on real-time signal quality. The feature vectors of the current, electromagnetic radiation, and ultraviolet light channels are weighted and summed according to the time-varying confidence coefficients of their respective channels to obtain the comprehensive feature vector. The time-varying confidence coefficient of each channel is dynamically calculated based on the real-time signal-to-noise ratio and signal integrity of the corresponding channel, and the sum of the time-varying confidence coefficients of each channel is 1.

[0007] As a preferred embodiment of the arc fault identification method based on multi-physics field fusion described in this invention, the method further includes: before extracting multi-dimensional features, performing adaptive digital preprocessing on the synchronously acquired current waveform signal, electromagnetic radiation signal and ultraviolet light signal respectively, wherein the adaptive digital preprocessing includes 50Hz notch filtering, adaptive baseline correction, wavelet threshold denoising and data normalization. The multidimensional features include current features, electromagnetic radiation features, and ultraviolet light features.

[0008] As a preferred embodiment of the arc fault identification method based on multi-physics field fusion described in this invention, the CNN-LSTM hybrid neural network model includes a one-dimensional convolutional neural network front-end, a two-layer LSTM temporal network, and a Softmax classification output layer. The one-dimensional convolutional neural network front-end is used to extract short-term local temporal features, the two-layer LSTM temporal network is used to model the long-range dependence of arc features in time, and the Softmax classification output layer is used to output three probability distributions: series arc, parallel arc, and normal operating arc.

[0009] As a preferred embodiment of the arc fault identification method based on multi-physics field fusion described in this invention, the time-domain reflection and electromagnetic radiation attenuation co-location includes a first-level location and a second-level location. The first-level positioning is coarse positioning using the time-domain reflection method, which calculates the coarse positioning distance based on the arrival time difference between the forward current wave and the reflected wave, combined with the line wave velocity. The second level of positioning is electromagnetic radiation attenuation correction positioning. Based on the characteristic that the electromagnetic radiation field strength decreases exponentially with increasing distance, the correction distance is calculated from the field strength at the reference distance. The final fault location is a weighted fusion of the coarse positioning distance and the correction distance, and the fusion weight is dynamically adjusted according to the positioning environment.

[0010] As a preferred embodiment of the arc fault identification method based on multi-physics field fusion described in this invention, the zero arc interruption includes two SiC MOSFETs connected back-to-back with common source driven by an isolation driver chip to complete bidirectional current turn-off within 500ns. The turn-off transient energy is absorbed by a parallel transient voltage suppressor and an RC absorption network. Subsequently, the mechanical contacts of the electromagnetic contactor close to provide continuous mechanical isolation and current carrying.

[0011] The synchronous data acquisition is performed using a current transformer, an electromagnetic radiation sensor, and an ultraviolet light sensor.

[0012] After calculating the fault location, the fault type, fault location, and disconnection action record are uploaded to the cloud platform, and an audible and visual alarm signal is issued locally. The device then enters a fault-locked state and awaits manual reset.

[0013] Secondly, an arc fault identification device based on multi-physics field fusion includes: A multi-physics sensing unit is used to simultaneously acquire current waveform signals, electromagnetic radiation signals, and ultraviolet light signals. The multimodal fusion processing unit is used to adaptively weighted fuse multi-dimensional features, identify the type of arc fault through a CNN-LSTM hybrid neural network model, and calculate the fault location through time-domain reflection and electromagnetic radiation attenuation. The graded disconnection execution unit is used to perform graded disconnection according to the identification results, including disconnection of parallel arc zero-arc circuit breaker and disconnection of series arc conventional circuit breaker.

[0014] Thirdly, a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the arc fault identification method based on multi-physics field fusion as described in claim 1.

[0015] The beneficial effects of this invention are as follows: This invention unifies the synchronous sensing of three physical fields—current waveform, electromagnetic radiation, and ultraviolet light—into an adaptive weighted fusion of multi-dimensional features. By dynamically adjusting the contribution weight of each physical field based on the time-varying confidence coefficient of real-time signal quality, it overcomes the inherent defects of existing single current detection schemes, such as difficulty in identifying parallel electric arcs and high misjudgment rate in variable frequency load scenarios. This is beneficial to improving the accuracy of electric arc identification and reducing the misjudgment rate. Furthermore, it maintains the identification function by relying on soft redundancy fault tolerance when a single sensor fails, thereby improving the reliability and availability of the device in real complex power distribution environments. The CNN-LSTM hybrid neural network model enables end-to-end three-class identification of series arcs, parallel arcs and normal operating arcs. It eliminates the need for manual design of feature threshold rules and the identification rules are not limited by human experience. After quantization and acceleration by the CNN-LSTM hybrid neural network model, sub-millisecond inference latency can be achieved on low-cost embedded processors. By employing a tiered fault-breaking strategy, zero-arc breaking resources are reserved only for the most dangerous parallel arc scenarios, while series arcs are broken using conventional circuit breakers. Normal arcs do not malfunction, achieving an optimal balance between protection performance and cost. The two-level positioning strategy, which combines time-domain reflection and electromagnetic radiation attenuation for coordinated positioning, improves fault location accuracy and solves the problem of existing solutions that only know there is a fault but not its location. Furthermore, by uploading fault type, fault location, and waveform data to the cloud, a data foundation is provided for smart power distribution and preventative maintenance. Attached Figure Description

[0016] Figure 1 This is a block diagram of the overall structure of an arc fault identification and hierarchical separation device based on multiphysics field fusion provided in one embodiment of this application.

[0017] Figure 2 This is a schematic diagram of the structure of a multiphysics sensing unit provided in one embodiment of this application.

[0018] Figure 3 This is a schematic diagram of the connection of a hierarchical segmentation execution unit provided in one embodiment of this application.

[0019] Figure 4 This is a complete flowchart of an arc fault identification and hierarchical segmentation method based on multiphysics field fusion provided in one embodiment of this application.

[0020] Figure 5 This is a signal processing flowchart of an adaptive weighted fusion algorithm provided in one embodiment of this application.

[0021] Figure 6 This is a schematic diagram of the structure of a CNN-LSTM hybrid neural network model provided in one embodiment of this application.

[0022] Figure 7 This is a schematic diagram illustrating the principle of time-domain reflection and electromagnetic radiation-assisted fault location in one embodiment of this application.

[0023] Figure 8 This is a schematic diagram of an embedded software layered architecture provided in one embodiment of this application.

[0024] Figure 9 This is a schematic diagram illustrating the time-frequency domain synergistic complementarity of three physical fields as provided in one embodiment of this application.

[0025] Figure 10 This is a flowchart illustrating the anti-false judgment process in a variable frequency load scenario provided by one embodiment of this application. Detailed Implementation

[0026] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0027] Example 1 - Figure 3 It provides an arc fault identification device based on multi-physics field fusion, including a multi-physics field sensing unit, a multi-modal fusion processing unit, and a hierarchical segmentation execution unit.

[0028] A multi-physics sensing unit is used to simultaneously acquire current waveform signals, electromagnetic radiation signals, and ultraviolet light signals. The multimodal fusion processing unit is used to adaptively weighted fuse multi-dimensional features, identify the type of arc fault through a CNN-LSTM hybrid neural network model, and calculate the fault location through time-domain reflection and electromagnetic radiation attenuation. The graded disconnection execution unit is used to perform graded disconnection according to the identification results, including disconnection of parallel arc zero-arc circuit breaker and disconnection of series arc conventional circuit breaker.

[0029] In this preferred embodiment, the comprehensive feature vector is obtained by weighting and summing the feature vectors of the three channels—current, electromagnetic radiation, and ultraviolet light—according to time-varying confidence coefficients. This vector is used as input to the CNN-LSTM hybrid neural network model for arc fault type identification. The time-varying confidence coefficient is a fusion weight dynamically calculated based on the real-time signal-to-noise ratio and signal integrity of the corresponding channel, reflecting the diagnostic confidence of the channel signal under the current operating conditions. Zero-arc interruption is a method of interruption in which, after being identified as a parallel arc, the fault current is cut off by semiconductor power devices within microseconds, preventing the arc energy from being injected into the line. Graded interruption is a differentiated interruption strategy executed according to the arc fault type, including three handling methods: zero-arc interruption for parallel arcs, conventional circuit breaker interruption for series arcs, and no interruption for normal operating arcs. Cooperative positioning is a weighted fusion of the two-level positioning results of coarse positioning using time-domain reflectometry and electromagnetic radiation attenuation correction positioning, used to calculate the distance from the fault point to the device installation point. By synchronously sensing and adaptively weighting and fusing the characteristic information of electric arcs in three physical fields—current, electromagnetic radiation, and ultraviolet light—end-to-end identification of the three types of electric arcs is achieved through a deep neural network. Based on the identification results, hierarchical segmentation and collaborative positioning are performed, thereby compressing the current interruption time of parallel electric arcs to the microsecond level while ensuring identification accuracy, and enabling the device to have fault location and fault-tolerant operation capabilities.

[0030] Reference Figure 2The multi-physics sensing unit includes a current transformer using a Rogowski coil, an electromagnetic radiation sensor using a broadband microstrip antenna, and an ultraviolet light sensor using a solar-blind ultraviolet photodiode. Data is simultaneously acquired via a 24-bit Δ-Σ ADC at a sampling rate of 2MHz, forming a three-channel synchronous data stream. The non-magnetic core design of the Rogowski coil ensures unsaturation under high current impact, enabling accurate measurement of thousands of amperes of fault current in parallel arcs. The selection of the 30MHz–300MHz frequency band maximizes the distinguishability of electromagnetic radiation characteristics between series arcs, parallel arcs, and normal operating arcs. The 240nm–280nm solar-blind band is unaffected by natural light, ensuring the detection specificity of the ultraviolet light channel in both indoor and outdoor environments.

[0031] Please refer to Figure 3 The graded interruption execution unit includes a zero-arc interruption module, a conventional interruption module, and an interruption control module. The zero-arc interruption module adopts a back-to-back common-source SiCMOSFET topology and is driven by an isolation driver chip. It completes bidirectional current turn-off within 500ns. After turn-off, continuous mechanical isolation and current carrying are provided by the mechanical contact closure of a series electromagnetic contactor. The conventional interruption module adopts an electromagnetic circuit breaker (typically with an interruption time of 5ms to 10ms). The interruption control module uses a programmable logic device to select the corresponding interruption path according to the type of arc fault.

[0032] In this preferred embodiment, the device is also used to: perform adaptive weighted fusion using time-varying confidence coefficients based on real-time signal quality; when any channel sensor fails or the signal is saturated, the time-varying confidence coefficient of that channel automatically drops to 0, relying on the remaining channels to maintain the arc fault identification function, thereby realizing the soft redundancy fault-tolerant operation of the sensor.

[0033] In this preferred embodiment, the device is also used to: perform adaptive digital preprocessing on the synchronously acquired current waveform signal, electromagnetic radiation signal and ultraviolet light signal respectively. The adaptive digital preprocessing includes 50Hz notch filtering, adaptive baseline correction, wavelet threshold denoising and data normalization. A total of 64 multi-dimensional features are extracted, including current features, electromagnetic radiation features and ultraviolet light features.

[0034] In this preferred embodiment, the device is also used to: when the maximum class probability is lower than the decision threshold of 0.85, expand the observation window from 128 frames to 256 frames to perform secondary inference.

[0035] In this preferred embodiment, the device is also used to: perform two-level positioning with time-domain reflection and electromagnetic radiation attenuation collaborative positioning. The first-level positioning calculates the coarse positioning distance based on the arrival time difference between the forward current wave and the reflected wave. The second-level positioning calculates the correction distance based on the exponential decay model of electromagnetic radiation field strength with distance. The final fault location is a weighted fusion of the two-level positioning results.

[0036] In this preferred embodiment, the device is also used to: when a parallel arc is detected, drive two SiC MOSFETs connected back-to-back with common source via an isolation driver chip to complete bidirectional current turn-off within 500ns; the turn-off transient energy is absorbed by a parallel transient voltage suppressor and an RC absorption network; subsequently, the electromagnetic contactor mechanical contacts close to provide continuous mechanical isolation and current carrying; when a series arc is detected, the electromagnetic circuit breaker completes conventional disconnection within 5ms to 10ms.

[0037] In this embodiment, the preferred embodiment is as follows: Figure 1 As shown, the device also includes a communication unit and an alarm unit. The communication unit uses a wireless communication module to upload fault type, fault location, disconnection action record, and fault waveform clips to the cloud platform and receive remote control commands; the alarm unit is used to issue audible and visual alarm signals in the event of an arc fault, and enters a fault-locked state after the disconnection action, awaiting manual reset.

[0038] Furthermore, from the perspective of the current waveform distortion mechanism, the resistivity of the arc plasma channel is highly nonlinear, and its voltage-current characteristics exhibit a negative resistance effect. When the current crosses zero, the arc plasma temperature decreases and the conductivity drops sharply, causing the arc current to have a dwell plateau near the zero point. The zero-point dwell time of the series arc is relatively long (generally 5-30 μs in actual operation), while the zero-point dwell time of the parallel arc is relatively short (generally 1-5 μs in actual operation). Moreover, the nonlinear volt-ampere characteristics of the arc will excite rich harmonic components in the current waveform, among which even harmonics (2nd and 4th) account for an increased proportion in the series arc. From the perspective of electromagnetic radiation excitation mechanism, electrons in the arc plasma collide with neutral molecules and ions under the acceleration of the electric field, producing bremsstrahlung and recombination radiation. In the 30MHz–300MHz frequency band, the radiation spectrum of the series arc exhibits a bimodal characteristic around 50MHz and 150MHz. The radiation energy of the parallel arc is concentrated in the low-frequency band of 30–80MHz with extremely high pulse amplitude and narrow width. In contrast, the radiation energy of normal operating arcs, such as motor commutation sparks, is mainly concentrated above 400MHz with extremely short duration. From the perspective of ultraviolet radiation excitation mechanism, the high temperature of the arc causes thermal excitation and electron collision excitation of nitrogen molecules in the surrounding air. When excited nitrogen molecules de-excite, they emit ultraviolet photons of characteristic wavelengths. Among them, the second positive band system of nitrogen molecules has strong emission lines at 253.7nm and 280.4nm. The peak intensity of ultraviolet radiation from the parallel arc is extremely high but short in duration, while the intensity of ultraviolet radiation from the series arc is lower but persists throughout the entire arc period. The energy of a normal operating arc is far from sufficient to excite nitrogen molecules, and its ultraviolet radiation can be ignored.

[0039] Example 2, refer to Figures 4-10This is another embodiment of the present invention, based on the previous embodiment. The difference between this embodiment and the previous embodiment is that it provides an arc fault identification method based on multi-physics field fusion, including: Step S1: Synchronously acquire current waveform signals, electromagnetic radiation signals, and ultraviolet light signals; In this preferred embodiment, after the device is powered on, each sensor is calibrated first, followed by loading the neural network model and configuring the operating parameters. After self-testing, it enters normal operating mode. Sensor calibration includes zero-point calibration of the current transformer, sensitivity calibration of the electromagnetic radiation sensor, and dark current calibration of the ultraviolet light sensor. Neural network model loading refers to loading the pre-trained and quantized CNN-LSTM hybrid neural network model into the inference engine. Then, current waveform signals, electromagnetic radiation signals, and ultraviolet light signals are synchronously acquired at a sampling rate of 2MHz to form a three-channel synchronous data stream. Each of the three channels of data is converted into digital quantities by a 24-bit Δ-Σ ADC to form a synchronous data stream.

[0040] Step S2: Extract multi-dimensional features, and adaptively weight and fuse them according to real-time signal quality to obtain a comprehensive feature vector; In this preferred embodiment, notch filtering, baseline correction, denoising, and normalization are performed on the three-channel synchronous data streams, and multi-dimensional features are extracted. Preprocessing includes 50Hz power frequency notch filtering, adaptive sliding window baseline correction, wavelet threshold denoising, and data normalization. Then, current features, electromagnetic radiation features, and ultraviolet light features are extracted from the preprocessed signal, totaling 64 dimensions. Time-varying confidence coefficients are calculated based on the real-time signal quality of each channel, and weighted summation is performed to obtain a comprehensive feature vector. The feature vectors of the current, electromagnetic radiation, and ultraviolet light channels are then weighted and summed according to their respective time-varying confidence coefficients to obtain a comprehensive feature vector. Step S3: Input the time sequence of the comprehensive feature vector into the CNN-LSTM hybrid neural network model to identify the type of arc fault; In this preferred embodiment, the CNN-LSTM hybrid neural network model outputs three probability distributions: series arc, parallel arc, and normal working arc. If the probability of the largest class is not lower than the decision threshold, the recognition result is directly output; otherwise, the enhanced sampling mechanism is triggered to perform secondary inference.

[0041] Step S4: According to the identification results, the circuit is divided into different levels: parallel arcs are broken by zero arc, series arcs are broken by conventional circuit breakers, and normal operating arcs are not broken. In this preferred embodiment, if a parallel arc is detected, zero-arc interruption is immediately triggered, the SiCMOSFET is turned off within 500ns, and then the mechanical contactor closes to provide continuous mechanical isolation and current carrying. If a series arc is identified, the conventional circuit breaker is triggered to disconnect; if a normal operating arc is identified, the disconnection action is not performed, and the process returns to step S102 to continue monitoring.

[0042] Step S5: After the fault is broken, the fault location is calculated by co-location of time-domain reflection and electromagnetic radiation attenuation.

[0043] In this preferred embodiment, after the fault is disconnected, a time-domain reflectometry and electromagnetic radiation attenuation collaborative positioning algorithm is activated to calculate the distance between the fault point and the device installation point. The first-level positioning uses coarse positioning with time-domain reflectometry, and the second-level positioning uses electromagnetic radiation attenuation correction positioning. The final fault location is a weighted fusion of the two-level positioning results, with a positioning accuracy within ±0.5m. Then, the event is reported, alarms are triggered, and faults are locked. Fault information is uploaded and manual reset is awaited. The fault type, fault location, disconnection action record, and fault waveform segment are uploaded to the cloud platform. The local audible and visual alarm module is activated simultaneously, and the device enters a fault-locked state, awaiting manual confirmation and reset before normal operation can resume.

[0044] The adaptive weighted fusion uses time-varying confidence coefficients based on real-time signal quality. The feature vectors of the current, electromagnetic radiation, and ultraviolet light channels are weighted and summed according to the time-varying confidence coefficients of their respective channels to obtain a comprehensive feature vector. The time-varying confidence coefficients of each channel are dynamically calculated based on the real-time signal-to-noise ratio and signal integrity of the corresponding channel, and the sum of the time-varying confidence coefficients of each channel is 1.

[0045] In this preferred embodiment, when any channel sensor fails or the signal saturates, the time-varying confidence coefficient of that channel automatically drops to 0, relying on the remaining channels to maintain the arc fault identification function, thus achieving soft redundancy and fault-tolerant operation of the sensors. Data from each channel is converted into digital quantities by a 24-bit Δ-Σ ADC, forming a three-channel synchronous data stream. In one implementation, data acquisition is achieved through a DMA double-buffering mechanism, with each channel's data length being a 1024-point circular buffer. The preprocessing pipeline processing latency is no higher than 0.5ms, providing a data foundation for subsequent real-time fusion and identification.

[0046] It also includes adaptive digital preprocessing of the synchronously acquired current waveform signal, electromagnetic radiation signal and ultraviolet light signal before extracting multi-dimensional features. The adaptive digital preprocessing includes 50Hz notch filtering, adaptive baseline correction, wavelet threshold denoising and data normalization. Multidimensional characteristics include current characteristics, electromagnetic radiation characteristics, and ultraviolet light characteristics.

[0047] In this preferred embodiment, the adaptive digital preprocessing includes 50Hz notch filtering, adaptive baseline correction, wavelet threshold denoising, and data normalization. The 50Hz power frequency notch filtering eliminates fundamental frequency interference from the power grid; the adaptive sliding window baseline correction eliminates baseline drift caused by sensor temperature drift; wavelet threshold denoising preserves high-frequency components of the arc characteristics while suppressing random noise; and data normalization maps the features of each channel to a unified numerical range, providing high-confidence input for subsequent fusion and recognition.

[0048] Furthermore, the multi-dimensional features total 64 dimensions, including current characteristics, electromagnetic radiation characteristics, and ultraviolet light characteristics. Current characteristics include current distortion rate, amplitude of each harmonic from the 3rd to the 31st, zero-point dwell time, maximum current rate of change, and rate of change of current rise slope; electromagnetic radiation characteristics include mean and peak radiation intensity, pulse count rate, spectral centroid, energy proportion in the 30–80 MHz range, energy proportion in the 150–250 MHz range, and pulse width distribution; ultraviolet light characteristics include mean and peak light intensity, pulse rise time, pulse half-width at half-maximum, and pulse interval time.

[0049] Furthermore, the adaptive weighted fusion employs time-varying confidence coefficients based on real-time signal quality. The feature vectors of the current, electromagnetic radiation, and ultraviolet light channels are weighted and summed according to their respective channel's time-varying confidence coefficients to obtain a comprehensive feature vector, with the sum of the time-varying confidence coefficients of each channel being 1. The real-time signal-to-noise ratio of each channel can be obtained by calculating the ratio of signal power to noise floor power in real time using a sliding window, for example, with a window length of 1024 sampling points. Signal integrity can be determined by checking whether the signal is within the sensor's linear dynamic range, and the weight of the corresponding channel can be automatically reduced by detecting interference source characteristics in the background environment (such as electromagnetic pulses from motor startup). The physical basis for this is that the diagnostic confidence of different physical field signals varies under different operating conditions: electromagnetic radiation is most sensitive to arc faults, ultraviolet light has extremely high specificity in unobstructed environments, and current waveforms may be strongly interfered with when the load equipment operates within a large dynamic range. Therefore, the fusion weights should be dynamically adjusted according to the real-time signal quality rather than using fixed values. In the scenario of parallel electric arc in a residential building, the signal-to-noise ratios of the electromagnetic radiation and ultraviolet light channels are 45dB and 40dB, respectively. The confidence coefficients of the current, electromagnetic radiation and ultraviolet light channels are automatically adjusted to 0.15, 0.50 and 0.35, respectively, so that the recognition results are not affected by the interference of the current channel.

[0050] Furthermore, when any channel sensor fails or its signal saturates, the time-varying confidence coefficient of that channel automatically drops to 0 (or close to 0), relying on the remaining channels to maintain the arc fault identification function, thus achieving soft redundancy and fault-tolerant operation of the sensors. In scenarios where the ultraviolet sensor fails due to long-term contamination, after detecting that the signal integrity index of the ultraviolet channel is below the threshold, the confidence coefficient of the ultraviolet channel automatically drops to 0.02, and the confidence coefficients of the current and electromagnetic radiation channels are renormalized to 0.42 and 0.56, respectively. Relying only on the current and electromagnetic radiation channels for feature extraction and fusion, the CNN-LSTM hybrid neural network model still outputs the correct series arc identification result (generally, the probability of a series arc is 0.989), ensuring the correct execution of the conventional series arc breaking action and guaranteeing the continuity of the arc protection function in the event of a single sensor failure.

[0051] The CNN-LSTM hybrid neural network model includes a one-dimensional convolutional neural network front-end, a two-layer LSTM temporal network, and a Softmax classification output layer. The one-dimensional convolutional neural network front-end is used to extract short-term local temporal features, the two-layer LSTM temporal network is used to model the long-range temporal dependence of arc features, and the Softmax classification output layer is used to output the three probability distributions of series arc, parallel arc, and normal working arc.

[0052] In this preferred embodiment, when the maximum class probability is lower than the decision threshold of 0.85, the observation window is expanded from 128 frames to 256 frames to perform secondary inference. The CNN-LSTM hybrid neural network model includes a one-dimensional convolutional neural network front-end, a two-layer LSTM temporal network, and a Softmax classification output layer. The front end of a one-dimensional convolutional neural network receives a sequence of 128 × 64-dimensional temporal feature frames. Short-term local temporal features are extracted through three layers of one-dimensional convolutional layers, batch normalization layers, ReLU activation layers, and max pooling layers. These features include the duration of current zero-point dwell, the slope of the rising edge of electromagnetic radiation pulses, and the full width at half maximum (FWHM) of ultraviolet light peaks. The kernel sizes are 32, 16, and 8, respectively, and the number of kernels per layer is 64 → 128 → 256. The output is a spatiotemporal feature map. The number of hidden units in each layer of the two-layer LSTM temporal network is 128. The long-range temporal dependencies of arc features are modeled through gating mechanisms such as forget gates, input gates, and output gates. This makes it suitable for capturing the dynamic evolution features of a series arc from its inception to stable arcing. Its output is reduced to 64 dimensions through a fully connected layer. The Softmax classification output layer outputs three probability distributions for series arcs, parallel arcs, and normal operating arcs through two fully connected layers.

[0053] Furthermore, the CNN-LSTM hybrid neural network model was trained on no fewer than 100,000 samples of different types of arc faults and normal operating arcs. The training data included measured data from standard arc generating devices conforming to UL1699 and GB / T31143 standards, simulation data generated based on the Cassie-Mayr arc physics model, and real waveforms covering 33 load scenarios across 10 categories, including resistive, inductive, capacitive, rectifier, and frequency converter drives. Training employed the Adam optimizer with an initial learning rate of 0.001, an L2 regularization coefficient of 0.0001, a Dropout rate of 0.3, and 200 training epochs, combined with an early stopping strategy. The CNN-LSTM hybrid neural network model, after TFLite compression and INT8 quantization, was optimized using the CMSIS-NN acceleration library, resulting in a single inference latency of no more than 1.5ms. This end-to-end feature learning method eliminates the need for manually designed feature threshold rules; the CNN-LSTM hybrid neural network model automatically learns the optimal discrimination boundary of arc features from a large amount of labeled data, which helps improve recognition accuracy.

[0054] Furthermore, refer to Figure 8 The multimodal fusion processing unit employs a main control MCU running hierarchical modular embedded software, including a hardware abstraction layer encapsulating sensor driving and communication peripheral operations; a data acquisition and preprocessing layer that achieves synchronous acquisition and preprocessing of three channels through DMA double buffering; a feature extraction and fusion layer that performs fusion by sliding extraction of 64-dimensional feature vectors and calculating confidence coefficients; a deep learning inference layer that runs a quantized CNN-LSTM model; and a decision and execution layer that completes arc fault type determination, fault location, disconnection control, and event reporting. This layered architecture decouples the functions of each layer and allows for independent upgrades. For example, sensor selection upgrades in the sensing unit do not require changes to the disconnection execution logic, and iterative updates of the deep learning model do not affect the hardware execution path, facilitating engineering deployment and large-scale production.

[0055] Co-location of temporal reflectance and electromagnetic radiation attenuation includes first-level localization and second-level localization; The first-level positioning is coarse positioning using the time-domain reflection method, which calculates the coarse positioning distance based on the arrival time difference between the forward current wave and the reflected wave, combined with the line wave velocity. The second level of positioning is electromagnetic radiation attenuation correction positioning. Based on the characteristic that the electromagnetic radiation field strength decreases exponentially with increasing distance, the correction distance is calculated from the field strength at the reference distance. The final fault location is a weighted fusion of the coarse positioning distance and the correction distance, and the fusion weight is dynamically adjusted according to the positioning environment.

[0056] In this preferred embodiment, the coarse positioning method using time-domain reflectometry calculates the coarse positioning distance based on the arrival time difference between the forward current wave and the reflected wave, combined with the line wave velocity. Specifically, half of the product of the arrival time difference and the line wave velocity is used as the coarse positioning distance from the fault point to the device installation point. The line wave velocity depends on the cable dielectric constant and can be approximated as 2 × 10⁸ m / s. By increasing the sampling rate to 2 MHz and combining it with a pulse peak detection algorithm, the coarse positioning accuracy of the time-domain reflectometry method can be improved to within ±75m. However, the accuracy of the time-domain reflectometry method is limited in environments with complex branch lines and difficult reflected wave identification; therefore, a second-stage positioning correction is required.

[0057] Furthermore, electromagnetic radiation attenuation correction positioning utilizes the characteristic that electromagnetic radiation field strength decreases exponentially with increasing distance. The correction distance is calculated by combining the field strength at a reference distance with the path loss index and the cable dielectric loss factor. The path loss index is taken as 2.5–3.5 in power distribution corridor environments, and the cable dielectric loss factor is taken as 0.02–0.05 m−1 for PVC insulated cables. The final fault location is a weighted fusion of the coarse positioning distance and the correction distance. The fusion weight is dynamically adjusted according to the positioning environment: in complex branch line environments, the weight of electromagnetic radiation attenuation correction positioning is automatically increased; in environments with severe electromagnetic obstruction, the weight of coarse positioning using time-domain reflectometry is automatically increased. This ensures positioning accuracy within ±0.5m in various real-world power distribution scenarios.

[0058] Zero arc interruption involves two SiC MOSFETs connected back-to-back with common source, driven by an isolation driver chip, to complete bidirectional current turn-off within 500ns. The turn-off transient energy is absorbed by a parallel transient voltage suppressor and an RC absorption network. Subsequently, the mechanical contacts of the electromagnetic contactor close to provide continuous mechanical isolation and current carrying.

[0059] In this preferred embodiment, if a parallel arc is detected, zero-arc interruption is immediately triggered: the two SiC MOSFETs connected back-to-back with common source are driven by the isolation driver chip to complete bidirectional current turn-off within 500ns. The turn-off transient energy is absorbed by the parallel transient voltage suppressor and RC absorption network. Subsequently, the electromagnetic contactor mechanical contacts close to provide continuous mechanical isolation and current carrying. In one implementation, the isolation driver chip can be a UCC21750, and the SiC MOSFETs can be 650V / 90mΩ devices. When the SiC MOSFETs are turned off, the parasitic inductance of the power circuit generates a reverse electromotive force, the magnitude of which is proportional to the rate of change of current at the moment of turn-off. This energy is absorbed by the transient voltage suppressor and RC absorption network to ensure that the drain-source voltage of the SiC MOSFETs does not exceed the rated withstand voltage. The mechanical contactor completes mechanical closing switching within approximately 8ms. In a residential parallel arc scenario, the insulation between the live wire and the ground wire was damaged due to renovation work. At the moment the arc occurred, the current jumped from the normal load current of 15A to about 1200A (limited by the line impedance). The multiphysics sensing unit captured the current jump (current change rate of 1.2×109A / s), the strong radiation pulse in the 30-80MHz frequency band, and the ultraviolet light spike with a peak intensity 1200 times that of the normal background value within 2μs after the arc occurred. The CNN-LSTM hybrid neural network model output the parallel arc probability as 0.996. After receiving the decision result, the SiCMOSFET was immediately turned off. The measured total time from the occurrence of the parallel arc to the current cutoff was 72.5μs. The significance of this design lies in the fact that the SiCMOSFET is normally closed and conducts under normal conditions, only performing a rapid turn-off action when a parallel arc is detected. Subsequently, a mechanical contactor takes over the isolation task. This avoids the conduction losses and heat dissipation problems caused by the SiCMOSFET's long-term current flow, and reserves the expensive and fast solid-state breaking resources only for parallel arc faults that truly require ultra-fast response. In practical applications, conventional circuit breakers complete the breaking process within 5ms to 10ms using electromagnetic circuit breakers.

[0060] Furthermore, if a series arc is identified, the conventional circuit breaker is triggered to trip, while the electromagnetic circuit breaker completes the trip within 5ms to 10ms. In the scenario of a series arc in an old building, the incoming terminals become loose due to long-term oxidation (contact resistance of approximately 5Ω), resulting in a continuous series arc under a load current of 8A. The current transformer detects that the zero-point dwell time increases from 2μs under normal operating conditions to 18μs, the second harmonic content increases from 0.5% to 12.3%, and the fourth harmonic increases from 0.2% to 5.8%. The electromagnetic radiation sensor detects a medium-intensity 50MHz and 150MHz double-peak radiation signal (pulse frequency of approximately 100Hz, synchronized with the power frequency's two zero-crossings per cycle). The ultraviolet light sensor detects continuous ultraviolet radiation with an intensity approximately 35 times the background value that persists throughout the entire arc's existence period. The model outputs a series arc probability of 0.994. The conventional tripping module sends a tripping signal to the electromagnetic circuit breaker's tripping coil, and the circuit breaker completes mechanical tripping within approximately 7ms. In this scenario, the amplitude of the series arc current is lower than the rated value, and traditional overload protection cannot respond. The embodiment of this application achieves accurate identification and disconnection through multi-physics field collaborative sensing.

[0061] If the arc is identified as a normal operating arc, the disconnection action is not performed, and the process returns to step S102 to continue monitoring. The energy and radiation characteristics of a normal operating arc (such as a motor commutation spark) are both below the fault discrimination boundary. The model outputs a dominant probability of a normal operating arc, and the device maintains normal monitoring status to avoid unnecessary malfunctions and device lifespan loss.

[0062] Synchronous data acquisition is performed using current transformers, electromagnetic radiation sensors, and ultraviolet light sensors.

[0063] In this preferred embodiment, current waveform signals, electromagnetic radiation signals, and ultraviolet light signals are simultaneously acquired using a current transformer, an electromagnetic radiation sensor, and an ultraviolet light sensor. The current transformer employs a Rogowski coil with a sampling frequency of at least 2MHz and a rated current of at least 100A. The non-magnetic core design of the Rogowski coil ensures it does not saturate under high current impact, allowing for accurate measurement of fault currents of several thousand amperes in parallel arcs. The electromagnetic radiation sensor uses a broadband microstrip antenna operating in the 30MHz–300MHz frequency band. This frequency band selection is based on the analysis of arc plasma radiation characteristics: the radiation spectrum of series arcs exhibits a bimodal distribution around 50MHz and 150MHz; parallel arcs radiate greater energy but are dominated by low-frequency components; while the radiation frequency band of normal operating arcs, such as motor commutation sparks, is mainly above 400MHz and has an extremely short duration. This frequency band selection maximizes the distinguishability of the electromagnetic radiation characteristics of the three types of arcs. The ultraviolet light sensor uses a solar-blind ultraviolet photodiode with a response wavelength of 240nm to 280nm and a response time of no more than 500ns. The 240nm to 280nm solar-blind band corresponds to the emission spectrum of nitrogen molecule transitions, and sunlight in this band is completely absorbed by the atmospheric ozone layer. Therefore, the ultraviolet light sensor will not be affected by natural light in indoor or outdoor environments.

[0064] Furthermore, refer to Figure 9 The three physical signals have a synergistic and complementary relationship in the time domain, frequency domain, and physical mechanism: In the time domain, the electromagnetic radiation signal is sensitive to the moment of arc ignition (nanosecond-level response) and can be used to quickly start the detection process; the ultraviolet light signal is sensitive to the duration of the arc and can be used to confirm whether the arc has been extinguished. The time correlation of the three enables multiple confirmations of the arc event; In the frequency domain, the three channels cover the complete spectrum distribution of arc energy from low-frequency electromagnetic to high-frequency electromagnetic to light radiation, with no blind spots in perception; In terms of physical mechanism, the current waveform reflects the macroscopic electrical characteristics of the arc circuit, the electromagnetic radiation reflects the microscopic electronic oscillation characteristics of the arc plasma, and the ultraviolet light reflects the temperature and ionization degree of the arc plasma. The three physical quantities describe the same arc event from different perspectives, and their intersection information is conducive to improving the confidence of fault diagnosis.

[0065] After calculating the fault location, the fault type, fault location, and disconnection action record are uploaded to the cloud platform, and an audible and visual alarm signal is issued locally. The device then enters a fault-locked state and awaits manual reset.

[0066] In this preferred embodiment, the fault type, fault location, disconnection action record, and fault waveform segment are uploaded to the cloud platform via the communication unit. Simultaneously, the local audible and visual alarm module is activated. The audible and visual alarm can employ a flashing red LED and a pulsating buzzer to support remote monitoring, data analysis, and preventative maintenance. After the disconnection action, the device enters a fault-locked state. This fault-locking mechanism ensures traceability of fault handling and safe operating procedures. The device only resumes normal operation after manual confirmation and reset, preventing the device from being put back into operation before the fault is resolved.

[0067] Reference Figure 10 In the normal operation of a variable frequency air conditioner, the PWM switching current generated by the variable frequency drive circuit contains rich high-frequency harmonic components (3rd to 9th harmonic content of 5.1%) and zero-point distortion (zero-point dwell time of 3.5μs), and the current change rate reaches the order of 107A / s, which can easily be misjudged as a series arc. The high-frequency radiation pulses detected by the electromagnetic radiation sensor are mainly distributed in the 400MHz to 600MHz range (harmonic radiation from the PWM switch). The radiation energy in the 30MHz to 300MHz frequency band is extremely low (signal-to-noise ratio of only 5dB), far lower than the radiation level of the series arc. The ultraviolet light sensor did not detect ultraviolet radiation signals in the 240nm to 280nm band. After automatic adjustment of the confidence coefficient, the confidence coefficient of the electromagnetic radiation channel dropped to 0.15, and the confidence coefficient of the ultraviolet light channel was 0.05. Although the current waveform has distortion characteristics, it lacks cross-validation in the electromagnetic and optical channels. The CNN-LSTM model outputs a normal working probability of 0.976, correctly judging it as a normal working state, and does not execute the disconnection action, making the misjudgment rate in the variable frequency load scenario less than 0.1%, thereby reducing the misjudgment rate.

[0068] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon. When executed by a processor, the computer program can perform any step of the arc fault identification method based on multiphysics fusion described above. The computer-readable storage medium includes, but is not limited to, disk storage, CD-ROM, optical storage, flash memory, solid-state drives, etc.

[0069] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products, and may take the form of entirely hardware embodiments, entirely software embodiments, or embodiments combining software and hardware aspects. They may also take the form of computer program products implemented 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. Each flowchart and / or block in the flowchart and / or block diagram, and combinations of flowcharts and / or blocks, may be implemented by computer program instructions; these instructions may be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus, such that instructions executed via the processor create means for implementing the functions specified in the flowchart and / or block diagram; these instructions may also be stored in a computer-readable storage medium, such that instructions stored in the storage medium create an article of manufacture including instruction means; these instructions may also be loaded onto a computer or other programmable apparatus, such that a series of operational steps performed on the computer or other programmable apparatus provide computer-implemented steps.

[0070] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An arc fault identification method based on multi-physics field fusion, characterized in that, include: Step S1: Synchronously acquire current waveform signals, electromagnetic radiation signals, and ultraviolet light signals; Step S2: Extract multi-dimensional features, and adaptively weight and fuse them according to real-time signal quality to obtain a comprehensive feature vector; Step S3: Input the time sequence of the comprehensive feature vector into the CNN-LSTM hybrid neural network model to identify the arc fault type; Step S4: According to the identification results, the circuit is divided into different levels: parallel arcs are broken by zero arc, series arcs are broken by conventional circuit breakers, and normal operating arcs are not broken. Step S5: After the fault is broken, the fault location is calculated by co-location of time-domain reflection and electromagnetic radiation attenuation.

2. The arc fault identification method based on multiphysics field fusion as described in claim 1, characterized in that: The adaptive weighted fusion uses time-varying confidence coefficients based on real-time signal quality. The feature vectors of the current, electromagnetic radiation, and ultraviolet light channels are weighted and summed according to the time-varying confidence coefficients of their respective channels to obtain the comprehensive feature vector. The time-varying confidence coefficients of each channel are dynamically calculated based on the real-time signal-to-noise ratio and signal integrity of the corresponding channel, and the sum of the time-varying confidence coefficients of each channel is 1.

3. The arc fault identification method based on multiphysics field fusion as described in claim 2, characterized in that: It also includes adaptive digital preprocessing of the synchronously acquired current waveform signal, electromagnetic radiation signal and ultraviolet light signal before extracting multi-dimensional features. The adaptive digital preprocessing includes 50Hz notch filtering, adaptive baseline correction, wavelet threshold denoising and data normalization. The multidimensional features include current features, electromagnetic radiation features, and ultraviolet light features.

4. The arc fault identification method based on multiphysics field fusion as described in claim 3, characterized in that: The CNN-LSTM hybrid neural network model includes a one-dimensional convolutional neural network front-end, a two-layer LSTM temporal network, and a Softmax classification output layer. The one-dimensional convolutional neural network front-end is used to extract short-term local temporal features, the two-layer LSTM temporal network is used to model the long-range temporal dependencies of arc features, and the Softmax classification output layer is used to output three probability distributions: series arc, parallel arc, and normal operating arc.

5. The arc fault identification method based on multiphysics field fusion as described in claim 4, characterized in that: The time-domain reflection and electromagnetic radiation attenuation co-location includes first-level positioning and second-level positioning; The first-level positioning is coarse positioning using the time-domain reflection method, which calculates the coarse positioning distance based on the arrival time difference between the forward current wave and the reflected wave, combined with the line wave velocity. The second level of positioning is electromagnetic radiation attenuation correction positioning. Based on the characteristic that the electromagnetic radiation field strength decreases exponentially with increasing distance, the correction distance is calculated from the field strength at the reference distance. The final fault location is a weighted fusion of the coarse positioning distance and the correction distance, and the fusion weight is dynamically adjusted according to the positioning environment.

6. The arc fault identification method based on multiphysics field fusion as described in claim 5, characterized in that: The zero-arc interruption involves two SiC MOSFETs connected back-to-back with common source, driven by an isolation driver chip, to complete bidirectional current turn-off within 500ns. The turn-off transient energy is absorbed by a parallel transient voltage suppressor and an RC absorption network. Subsequently, the mechanical contacts of the electromagnetic contactor close to provide continuous mechanical isolation and current carrying.

7. The arc fault identification method based on multiphysics field fusion as described in claim 6, characterized in that: The synchronous data acquisition is performed using a current transformer, an electromagnetic radiation sensor, and an ultraviolet light sensor.

8. The arc fault identification method based on multiphysics field fusion as described in claim 7, characterized in that: After calculating the fault location, the fault type, fault location, and disconnection action record are uploaded to the cloud platform, and an audible and visual alarm signal is issued locally. The device then enters a fault-locked state and awaits manual reset.

9. An arc fault identification device based on multiphysics field fusion, characterized in that, include: A multi-physics sensing unit is used to simultaneously acquire current waveform signals, electromagnetic radiation signals, and ultraviolet light signals. The multimodal fusion processing unit is used to adaptively weighted fuse multi-dimensional features, identify the type of arc fault through a CNN-LSTM hybrid neural network model, and calculate the fault location through time-domain reflection and electromagnetic radiation attenuation. The graded disconnection execution unit is used to perform graded disconnection according to the identification results, including disconnection of parallel arc zero-arc circuit breaker and disconnection of series arc conventional circuit breaker.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the arc fault identification method based on multiphysics field fusion as described in claim 1.

Images