Low-voltage series fault arc AI detection method and device based on intelligent electric meter

Abstract

The invention provides a low-voltage series fault arc AI detection method and device based on an intelligent electric meter. The method comprises the steps of collecting current data of the intelligent electric meter and transmitting the current data to a data processing module of an embedded development board; performing feature extraction and standardization processing on the acquired current data by adopting a data processing module to obtain standardized feature data; inputting the standardized feature data into the trained neural network model to obtain a fault probability value; judging whether a fault arc state exists or not according to the fault probability value; and if the fault arc state exists, triggering an alarm device to output an alarm signal. According to the invention, current data are collected, time domain and frequency domain features are extracted, a double-branch neural network model is used for fault judgment, and an alarm is given when a fault arc is detected so as to guarantee circuit safety; the problem that a traditional method is insufficient in detection precision in a complex load environment is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of power system safety signal detection, and particularly to a low-voltage series arc fault AI detection method based on a smart meter. BACKGROUND

[0002] With the wide application of power systems and electronic devices, electrical safety has become an important issue to protect public safety and personal property. In low-voltage distribution systems, series arc fault is a common and extremely dangerous form of electrical fault. Because its current is usually small, it is often below the action threshold of traditional circuit breakers, which makes it difficult for existing protection devices to detect and cut off the fault circuit in time, thus becoming a significant vulnerability in the existing electrical protection system. Fault arc is caused by air breakdown or loose electrical connection due to aging, damage, and moisture of electrical insulation such as wires. It not only has huge energy and extremely high temperature, but also can easily cause fire and even explosion, which seriously threatens the safety of equipment and personnel.

[0003] In view of the above problems, domestic and foreign scholars have conducted extensive and in-depth research on fault arc detection. Early methods mainly rely on physical phenomena such as arc light, arc sound, temperature, and electromagnetic radiation to identify fault arcs. However, this method is limited by the installation location of the sensor and environmental factors, making it difficult to be widely applied in actual power lines. In recent years, with the development of digital signal processing technology and machine learning algorithms, methods based on current waveform feature analysis have gradually become the mainstream research direction. Specifically, researchers use Fourier transform, wavelet transform, empirical mode decomposition, and other tools to extract harmonic components and other characteristic quantities in the current signal, which are used as the basis for determining whether there is a fault arc. In addition, in order to improve the accuracy and adaptability of fault arc identification, many works have begun to explore the use of deep learning methods to automatically learn the deep features of fault arcs. Although significant progress has been made, most current methods still have certain limitations, such as potential interference with metering and billing during sampling, poor adaptability to nonlinear load conditions, false positives, and difficulty in applying to low-power embedded devices. SUMMARY

[0004] The purpose of the present application is to address the problem of increased difficulty in detecting low-voltage alternating current series arc faults due to the large differences in current characteristics under different load types. A low-voltage series arc fault AI detection method and device based on a smart meter are proposed.

[0005] The technical solution of the present application is:

[0006] The present application provides a low-voltage series arc fault AI detection method based on a smart meter, which comprises the following steps:

[0007] Data acquisition steps: Collect the current data of the smart meter and transmit it to the data processing module of the embedded development board;

[0008] Feature extraction and standardization steps: The data processing module is used to extract and standardize the collected current data to obtain standardized feature data;

[0009] AI model prediction steps: Input standardized feature data into the trained neural network model to obtain the fault probability value;

[0010] Status determination and alarm steps: Determine whether there is a fault arc state based on the fault probability value; if it is determined that there is a fault arc state, trigger the alarm device to output an alarm signal.

[0011] Furthermore, the data acquisition steps specifically include:

[0012] The smart meter's metering module is activated by sending a trigger signal to the smart meter management module via the GPIO interface of the embedded development board.

[0013] The smart meter is used as the SPI communication master device, and communicates with the embedded development board through the serial peripheral interface bus to transmit single-cycle current data.

[0014] Obtain voltage and current sampling points from single-cycle current data.

[0015] Furthermore, the feature extraction and standardization steps specifically include:

[0016] Dimensionless time-domain and frequency-domain feature indicators are extracted from the collected current data; wherein, the time-domain feature indicators include waveform indicators, peak indicators, pulse indicators, kurtosis indicators, margin indicators and energy indicators, and the frequency-domain feature indicators are obtained by obtaining the frequency centroid and each harmonic factor through Fast Fourier Transform (FFT).

[0017] The Z-score standardization method is used to process the extracted time-domain and frequency-domain feature indices to eliminate dimensional differences and obtain standardized feature data.

[0018] Furthermore, the AI ​​model prediction step specifically includes:

[0019] Standardized feature data is input into a pre-trained dual-branch neural network model, which includes an attention mechanism branch and a fully connected layer branch.

[0020] The non-linear relationships between features are captured through attention mechanism branches, and local features are extracted through fully connected layer branches.

[0021] The output features of the attention mechanism branch and the fully connected layer branch are fused. The output layer uses an activation function to process the fused feature data and outputs the fault probability value.

[0022] Furthermore, the training process of the dual-branch neural network model includes:

[0023] Construct a dataset containing various electrical load types, including normal current waveforms and fault arc waveforms;

[0024] Clean and filter the data in the constructed dataset to remove outlier data;

[0025] The cleaned data is divided into a training set and a test set;

[0026] Dynamic learning rate adjustment, early stopping mechanism and L2 regularization strategy are used to train the two-branch neural network model;

[0027] The performance of the dual-branch neural network model was evaluated using K-fold cross-validation to obtain the optimized model.

[0028] Furthermore, the electrical load types include resistive loads, inductive loads, electronic loads, switching power supplies, magnetrons, and eddy currents.

[0029] Furthermore, the deployment process of the neural network model includes:

[0030] The trained neural network model is quantized and optimized to generate a model version suitable for low-power embedded devices; the optimized neural network model is then integrated into the embedded development board to support local inference functionality.

[0031] The embedded development board receives the collected current data and executes inference tasks to output fault probability values, ensuring that the processing time of the inference tasks meets the requirements of real-time monitoring.

[0032] Furthermore, the status determination and alarm steps specifically include:

[0033] Receive the fault probability value output by the neural network model and compare the fault probability value with a preset threshold;

[0034] If the fault probability value is greater than or equal to the preset threshold, it is determined that there is a fault arc state and the alarm device is triggered; otherwise, it is determined to be a normal state and monitoring continues.

[0035] The device used in a low-voltage series fault arc detection method based on smart meters includes:

[0036] A smart meter is a data processing module used to collect current data and transmit it to an embedded development board via an SPI interface.

[0037] An embedded development board is used to control a smart meter via GPIO ports to trigger current data sampling and to receive current data from the smart meter.

[0038] The data processing module, running on the embedded development board, is used to perform feature extraction and standardization, AI model prediction, and state determination.

[0039] It also includes an alarm module, which is used to trigger an audible and visual alarm or remote notification when a faulty arc state is detected.

[0040] Furthermore, the neural network model is integrated into the embedded development board in the form of a static library or executable code, supporting online updates and model replacement, and adapting to changes in current characteristics under various electrical load environments.

[0041] The beneficial effects of this invention are:

[0042] This invention discloses an AI detection method and device for low-voltage series fault arc based on smart meters. The method collects current data, extracts time-domain and frequency-domain features, and uses a dual-branch neural network model to judge the fault. When a fault arc is detected, an alarm is issued to ensure circuit safety. This solves the problem of insufficient detection accuracy of traditional methods under complex load environments.

[0043] This invention employs quantization optimization technology to integrate the model into an embedded device, achieving low-power real-time detection. Simultaneously, it supports online model updates, adapting to changes in different load characteristics and enhancing the system's adaptability and reliability. It also enables accurate detection, timely alarm, and traceable recording of arc faults, providing effective protection for electrical safety.

[0044] This invention optimizes the algorithm and model structure, effectively capturing nonlinear relationships between features through a network structure combining attention mechanisms and fully connected layers, thereby improving detection accuracy. This makes it suitable for low-power embedded devices, ensuring efficient and accurate execution of fault arc detection tasks even in resource-constrained environments. Through this design, this invention can achieve reliable real-time monitoring and early warning in various electrical environments while maintaining low energy consumption and hardware resource requirements.

[0045] Other features and advantages of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0046] The above and other objects, features and advantages of the present invention will become more apparent from the more detailed description of exemplary embodiments of the invention in conjunction with the accompanying drawings, wherein the same reference numerals generally represent the same components in the exemplary embodiments of the invention.

[0047] Figure 1This is a schematic diagram of the structure of the AI ​​detection device for low-voltage series fault arc based on a smart meter according to the present invention.

[0048] Figure 2 This is a flowchart of the AI ​​detection method for low-voltage series fault arc based on smart meters according to the present invention.

[0049] Figure 3 Pearson correlation coefficient and Spearman coefficient for different eigenvalues.

[0050] Figure 4 Nonlinear intensity analysis diagrams for different eigenvalues.

[0051] Figure 5 This is a schematic diagram of a low-complexity bi-branch neural network model. Detailed Implementation

[0052] Preferred embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While preferred embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein.

[0053] This invention provides an AI-based method for detecting low-voltage series fault arcs based on smart meters, comprising the following steps:

[0054] Data Acquisition Steps: To acquire current data, the smart meter and the embedded development board communicate via an SPI interface in this invention. First, a high-level signal is sent to the COM_RQ pin of the smart meter management module through a GPIO port on the development board to trigger the metering module to start sending single-cycle current data. Specifically, when the smart meter detects that its COM_RQ pin is set to a high level, it will act as the master device in the SPI communication and initiate the data transmission process. During this process, the embedded development board acts as the slave device to receive data. Both parties communicate using SPI mode 1 (CPOL=0, CPHA=1) to ensure accurate data transmission. The communication rate is 819kHz, ensuring that 256 sampling points (128 for voltage and 128 for current) can be acquired per cycle, thus providing sufficient data resolution to support subsequent feature extraction and fault arc detection.

[0055] Feature extraction and standardization steps: Extract dimensionless time-domain indices (waveform indices, peak indices, impulse indices, kurtosis indices, margin indices, energy indices) based on probability density functions and frequency-domain feature indices (frequency centroid, harmonic factors) obtained based on Fast Fourier Transform (FFT). Apply the Z-score standardization method to standardize the extracted feature values, calculate the mean μ and standard deviation σ of each feature value, and perform X... normal = (X-μ) / σ is used to standardize the eigenvalues ​​and eliminate dimensional differences.

[0056] AI model prediction steps: Due to limited resources in embedded devices, the trained neural network model needs to be optimized by using quantization techniques to convert it into a form more suitable for hardware execution. Then, the optimized model is integrated into the C language environment to implement local inference functionality. Whenever new current cycle data is received, this inference module is immediately invoked to quickly calculate the probability value of the current state and determine whether a fault arc is present.

[0057] Status determination and alarm steps: After completing model training, newly collected current data is input into the model, and the existence of a fault arc is determined by the calculated probability value. If the probability value is greater than or equal to 0.5, it is determined to be a fault state; otherwise, it is determined to be a normal state.

[0058] The neural network model employs a dual-branch neural network model. The construction process includes: to reduce model complexity and adapt to the hardware requirements of embedded devices, this invention designs a hybrid architecture containing an attention mechanism and fully connected layers. The multi-head attention mechanism layer is used to capture non-linear relationships between features, while the fully connected layers are used to extract local features. The output layer uses the Sigmoid activation function to output the fault probability within the [0,1] interval, and a threshold of 0.5 is set to determine the state.

[0059] Dataset Creation and Model Training: Data samples from various typical electrical loads were collected according to the GB / 14287.4 standard, and then cleaned, filtered, and integrated to form a dataset containing 22,000 data points (11,000 fault waveforms and 11,000 normal waveforms). After the dataset was created, a two-branch neural network model incorporating attention mechanisms and fully connected layers was used for training. This process first used Z-score normalization to eliminate dimensional differences in the data features. Then, 70% of the data (15,400 data points) was used as the training set, and the remaining 30% (6,600 data points) was used as the test set to verify the model's performance. During training, optimization strategies such as dynamic learning rate adjustment, early stopping mechanisms, and L2 regularization were employed to ensure that the model did not overfit and maintained good generalization ability. Finally, the model achieved high accuracy and AUC on the test set, confirming its effectiveness and reliability in low-voltage series fault arc detection.

[0060] Example 1

[0061] 1. Data Sampling

[0062] The AI ​​arc detection module connects to the COM_RQ pin of the meter via GPIO, triggering the meter management module to control the metering module to start sending data. It receives the data sent by the meter via SPI, with an SPI communication rate of 819kHz. The metering module sends raw cycle data in real time, with 256 sampling points per cycle, 128 sampling points for voltage and 128 sampling points for current. The sampling frequency is 6.4kHz in a 220V / 50Hz environment.

[0063] The current signals of different electrical loads in the circuit exhibit complex linear and nonlinear characteristics. Therefore, it is necessary to enrich the types of loads during the data set production process. Referring to the commonly used typical electrical loads adopted in GB / 14287.4 standard to simulate actual production and living conditions, data collection and classification of the electrical appliance types in the table below should be carried out.

[0064] For each load, 1000 normal samples and 1000 arc samples were sampled, for a total of 22,000 data points, as shown in Table 1.

[0065] Table 1 Types of Load Devices

[0066]

[0067] 2. Dataset Creation

[0068] (1) Data cleaning

[0069] Because the arc generator may be disconnected during data sampling, and current value distortion may occur at the moment of conduction, it is necessary to remove abnormal data from the original sampled data before creating the dataset. Screening is performed based on electrical parameter thresholds to remove current values ​​of 0 and cycles that significantly exceed the load operating conditions.

[0070] (2) Screening criteria for arc frequency of different electrical equipment

[0071] Arc frequency screening standards for different electrical devices are set based on their respective load characteristics to distinguish between normal and fault states. For example, electric fans use the sum of absolute waveform values ​​combined with peak values ​​for screening; ceramic cooktops use the sum of absolute waveform values ​​and peak values ​​for judgment. Hair dryers use smoothness to screen fault frequencies based on their power. LED air purifiers use a combination of smoothness and peak values ​​for screening. Induction cookers use smoothness and peak values ​​as screening criteria. Microwave ovens use the sum of absolute values ​​and peak values. For combined loads such as ceramic cooktops and electric fans, or induction cookers and LED air purifiers, multiple characteristic indicators are considered comprehensively to accurately identify fault arcs, ensuring accurate detection under various load conditions. These standards work together to improve the accuracy and reliability of fault arc detection.

[0072] (3) Data integration

[0073] Among the fault waveforms selected from the above electrical equipment, 1000 fault waveforms were randomly selected for each type of electrical equipment, for a total of 11000 fault waveforms. 1000 normal waveforms for each type of electrical equipment were sampled separately, for a total of 11000 normal waveforms. The 11000 fault waveforms and 11000 normal waveforms were then combined to form a dataset of 22000 waveforms.

[0074] 3. Feature Data Filtering

[0075] To comprehensively quantify the correlation between 22 characteristic quantities and the single-cycle state of the current (arc = 1, normal = 0), the following two complementary correlation coefficients are used, and the calculation results are as follows: Figure 3 and Figure 4 As shown.

[0076] (1) Pearson correlation coefficient

[0077] This method assesses the linear correlation strength between features and state labels by calculating the normalized covariance between each feature x and the binary label y (arc = 1, normal = 0). The Pearson correlation coefficient is calculated as follows:

[0078]

[0079] Where: X i Y is the feature observation value of the i-th sample. i It is the state label of the i-th sample. and These are the sample means of each of the two variables, and n is the number of observations.

[0080] (2) Spearman rank correlation coefficient

[0081] This method converts feature values ​​and labels into ranking ranks, calculates the Pearson correlation coefficient of the ranks, and evaluates the monotonous nonlinear association between features and state labels. The calculation formula is as follows:

[0082]

[0083] Where: d i It is the difference between the feature rank and the label rank of the i-th sample.

[0084] 4. Arc Fault Detection Algorithm Based on Dual-Branch Neural Network

[0085] (1) Design of dual-branch neural network architecture

[0086] The architecture diagram of a two-branch neural network is as follows: Figure 5As shown, the structure includes the following: The attention branch uses the MultiHeadAttention mechanism to capture the non-linear relationships between features, and LayerNormalization stabilizes the training process. The fully connected branch uses a traditional fully connected layer structure, combined with BatchNormalization and Dropout to improve generalization ability. The fusion layer concatenates the output features of the two branches, combining the advantages of the two feature extraction methods. The output layer uses the Sigmoid activation function to output probability values ​​in the [0,1] interval.

[0087] The optimization strategy includes dynamic learning rate adjustment, automatic learning rate adjustment using ReduceLROnPlateau callback, early stopping mechanism to prevent overfitting through EarlyStopping, and L2 regularization to constrain weight parameters.

[0088] (2) Current single-cycle state determination mechanism

[0089] Judgment rule: The probability threshold is set to 0.5 (based on a balanced dataset). → Determined as a fault state (1), <0.5 → is considered normal (0).

[0090] 5. AI-based detection performance of low-voltage series fault arcs using smart meters

[0091] (1) Detection performance

[0092] Of the 22,000 data points in the dataset, 70% (15,400 data points) were used as the training set and 30% (6,600 data points) were used as the test set. Tables 2 and 3 show the prediction performance of the model on the 6,600 data points of the test set. A total of 8 data points were mispredicted, including 1 false positive and 7 false negatives.

[0093] Table 2 Performance of Fault Arc Detection Model

[0094] Accuracy AUC Accuracy Recall Training set performance 0.9994 0.9999 0.9999 0.9988 Test set performance 0.9988 0.9997 0.9997 0.9979

[0095] Table 3 Error Detection Information for Test Set

[0096]

[0097] (2) Detection speed

[0098] The current single-cycle data transmission takes 20ms, and the complete prediction process (data reception -> feature value calculation -> dual-branch neural network prediction -> result output) takes less than 1ms, which can meet the needs of real-time detection. The specific time details are shown in Table 5.

[0099] Table 5 Reasoning Speed

[0100]

[0101]

[0102] Table 6 Comparison of Existing Methods

[0103]

[0104] Table 6 shows the comparison results of the proposed method with other studies in terms of sampling frequency, recognition accuracy, and running time. Among the five methods in Table 6, neither the MFF-GNN nor the LSTM-CNN models have been tested and verified in embedded software, so their real-time performance cannot be proven. The HTFSCNN model requires a high-frequency coupled current sensor for data sampling at a frequency of 100kHz, which makes its engineering implementation quite difficult. The LiDtNet model is implemented on a Jetson Nano embedded device, resulting in relatively high hardware costs. The proposed solution has advantages in several dimensions. Without affecting the meter readings, it uses the original waveform provided by the meter, with a sampling frequency of 6.4kHz. The accuracy of arc detection is the highest among current solutions, while also having the lowest hardware overhead.

[0105] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments.

Claims

1. A method for detecting low-voltage series fault arcs based on smart meters, characterized in that, The method includes the following steps; Data acquisition steps: Collect the current data of the smart meter and transmit it to the data processing module of the embedded development board; Feature extraction and standardization steps: The data processing module is used to extract and standardize the collected current data to obtain standardized feature data; AI model prediction steps: Input standardized feature data into the trained neural network model to obtain the fault probability value; Status determination and alarm steps: Determine whether there is a fault arc state based on the fault probability value; if it is determined that there is a fault arc state, trigger the alarm device to output an alarm signal.

2. The AI ​​detection method for low-voltage series fault arc based on smart meters as described in claim 1, characterized in that, The data acquisition steps specifically include: The smart meter's metering module is activated by sending a trigger signal to the smart meter management module via the GPIO interface of the embedded development board. The smart meter is used as the SPI communication master device, and communicates with the embedded development board through the serial peripheral interface bus to transmit single-cycle current data. Obtain voltage and current sampling points from single-cycle current data.

3. The AI ​​detection method for low-voltage series fault arc based on smart meters as described in claim 1, characterized in that, The feature extraction and standardization steps specifically include: Dimensionless time-domain and frequency-domain feature indicators are extracted from the collected current data; wherein, the time-domain feature indicators include waveform indicators, peak indicators, pulse indicators, kurtosis indicators, margin indicators and energy indicators, and the frequency-domain feature indicators are obtained by obtaining the frequency centroid and each harmonic factor through Fast Fourier Transform (FFT). The Z-score standardization method is used to process the extracted time-domain and frequency-domain feature indices to eliminate dimensional differences and obtain standardized feature data.

4. The AI ​​detection method for low-voltage series fault arc based on smart meters as described in claim 1, characterized in that, The AI ​​model prediction steps specifically include: Standardized feature data is input into a pre-trained dual-branch neural network model, which includes an attention mechanism branch and a fully connected layer branch. The non-linear relationships between features are captured through attention mechanism branches, and local features are extracted through fully connected layer branches. The output features of the attention mechanism branch and the fully connected layer branch are fused. The output layer uses an activation function to process the fused feature data and outputs the fault probability value.

5. The AI ​​detection method for low-voltage series fault arc based on smart meters as described in claim 4, characterized in that, The training process of the dual-branch neural network model includes: Construct a dataset containing various electrical load types, including normal current waveforms and fault arc waveforms; Clean and filter the data in the constructed dataset to remove outlier data; The cleaned data is divided into a training set and a test set; Dynamic learning rate adjustment, early stopping mechanism and L2 regularization strategy are used to train the two-branch neural network model; The performance of the dual-branch neural network model was evaluated using K-fold cross-validation to obtain the optimized model.

6. The AI ​​detection method for low-voltage series fault arc based on smart meters as described in claim 5, characterized in that, The electrical load types include resistive loads, inductive loads, electronic loads, switching power supplies, magnetrons, and eddy currents.

7. The AI ​​detection method for low-voltage series fault arc based on smart meters as described in claim 1, characterized in that, The deployment process of the neural network model includes: The trained neural network model is quantized to generate a model version suitable for low-power embedded devices; the quantized neural network model is then integrated into the embedded development board to support local inference functionality. The embedded development board receives the collected current data and executes inference tasks to output fault probability values, ensuring that the processing time of the inference tasks meets the requirements of real-time monitoring.

8. The AI ​​detection method for low-voltage series fault arc based on smart meters as described in claim 1, characterized in that, The specific steps for status determination and alarm include: Receive the fault probability value output by the neural network model and compare the fault probability value with a preset threshold; If the fault probability value is greater than or equal to the preset threshold, it is determined that there is a fault arc state and the alarm device is triggered; otherwise, it is determined to be a normal state and monitoring continues.

9. An apparatus used in the AI ​​detection method for low-voltage series fault arc based on a smart meter as described in any one of claims 1-8, characterized in that... include; A smart meter is a data processing module used to collect current data and transmit it to an embedded development board via an SPI interface. An embedded development board is used to control a smart meter via GPIO ports to trigger current data sampling and to receive current data from the smart meter. The data processing module, running on the embedded development board, is used to perform feature extraction and standardization, AI model prediction, and state determination. It also includes an alarm module, which is used to trigger an audible and visual alarm or remote notification when a faulty arc state is detected.

10. The AI ​​detection device for low-voltage series fault arc based on a smart meter as described in claim 9, characterized in that, The neural network model is integrated into the embedded development board in the form of a static library or executable code, supporting online updates and model replacement, and adapting to changes in current characteristics under various electrical load environments.

Images