A distribution box with a fault early warning function

Abstract

The application provides a distribution box with a fault early warning function, and belongs to the technical field of distribution boxes.The distribution box comprises a distribution box body, a current sensor, a voltage sensor, a temperature sensor, a frequency detector, a communication interface and a control chip.Eight-dimensional time sequence data vectors are collected, preprocessed, and the local feature patterns of different time scales are extracted in parallel by using three different receptive field convolution kernels of a multi-scale feature extraction layer.A comprehensive feature vector is generated by a feature fusion integration layer, a specialized fault recognition and probability distribution calculation are performed by running an adaptive gating expert network model, a sampling frequency is dynamically adjusted by executing an adaptive processing frequency adjustment mechanism, a false judgment is avoided by implementing an abnormal state detection confirmation mechanism through continuity verification, an expert selection strategy is optimized by implementing a gating weight dynamic adjustment mechanism, and finally a hierarchical early warning output layer is generated to realize accurate fault types, thereby solving the technical problem of insufficient accuracy of fault early warning of a power distribution system.

Description

Technical Field

[0001] This invention belongs to the field of distribution box technology, and more specifically, relates to a distribution box with fault early warning function. Background Technology

[0002] As core equipment in power systems, distribution boxes are widely used in various power-consuming locations such as industrial parks, commercial complexes, residential communities, data centers, and hospitals. Their fault early warning technology primarily relies on traditional single-parameter threshold monitoring methods and simple alarm systems. Traditional distribution box monitoring systems typically use fixed threshold settings to independently monitor basic electrical parameters such as current, voltage, and temperature. An alarm signal is triggered when a single parameter exceeds a preset range. This technology has been widely used in power operation and maintenance management, intelligent building control, and factory automation monitoring. However, its technical architecture is relatively simple, mainly based on empirical thresholds and static analysis methods. Traditional technologies have significant technical shortcomings: single-parameter monitoring methods cannot fully reflect the complex operating states of the power distribution system; fixed threshold settings lack adaptability to dynamically changing environments; and simple alarm mechanisms struggle to accurately distinguish between normal fluctuations and actual faults, leading to numerous false alarms and missed alarms. Furthermore, traditional methods lack the ability to analyze the correlations between multi-dimensional parameters, failing to identify complex fault modes and early, subtle fault signs. In existing technologies, due to the lack of multi-dimensional parameter fusion analysis technology and intelligent feature recognition capabilities, traditional distribution box monitoring systems often suffer from low identification accuracy, long response time, and high false alarm rates when facing complex situations such as compound faults, intermittent faults, and early faults. Consequently, they cannot provide reliable technical support for the safe and stable operation of power systems. In other words, existing technologies suffer from insufficient accuracy in early warning of power distribution system faults. Summary of the Invention

[0003] In view of this, the present invention provides a power distribution box with fault early warning function, which can solve the technical problem of insufficient accuracy of fault early warning in the existing power distribution system.

[0004] This invention is implemented as follows: It provides a distribution box with fault early warning function, including a distribution box body, a current sensor, a voltage sensor, a temperature sensor, a frequency detector, a communication interface, and a control chip. The distribution box body internally houses multiple circuit breaker units and terminals. The current sensor is installed at the output terminal of each circuit breaker unit. The voltage sensor is connected between the input terminal of the distribution box body and the output terminals of each branch circuit. Temperature sensors are installed at key heat-generating locations inside the distribution box body. The frequency detector is connected to the output terminal of the voltage sensor via a signal line. The control chip includes a fault early warning intelligent analysis module. This module synchronously acquires an eight-dimensional time-series data vector by constructing a multi-dimensional electrical parameter sampling base layer, performs signal preprocessing enhancement on the eight-dimensional time-series data vector, and establishes multi-scale feature extraction. The system employs one-dimensional convolutional kernels with different receptive field sizes to process standardized data sequences in parallel, extracting local feature patterns at different time scales. A feature fusion and integration layer is then executed to fuse multi-scale feature vectors into a comprehensive feature vector. A lightweight fault identification model is run, inputting the comprehensive feature vector into an adaptive gating expert network model to output an initial fault probability distribution. An adaptive processing frequency adjustment mechanism is implemented to dynamically adjust the signal sampling and processing frequency based on the system load intensity parameter. An abnormal state detection and confirmation mechanism is implemented to adjust the capacity factor parameter of the adaptive gating expert network model based on the deviation metric function value. A dynamic gating weight adjustment mechanism is implemented to adjust the weight allocation function of the gating network based on the system stability balance value. Finally, a hierarchical early warning output layer is generated, classifying the fault probability values ​​into four levels: normal state, attention state, early warning state, and emergency state.

[0005] The step of constructing a multi-dimensional electrical parameter sampling base layer involves synchronously acquiring three-phase current data from a current sensor, three-phase voltage data from a voltage sensor, temperature data from a temperature sensor, and frequency data from a frequency detector at a frequency of 200Hz. A circular buffer with a length of 1024 sampling points is established, and a time-series data matrix is ​​generated for subsequent processing.

[0006] The signal preprocessing enhancement step specifically involves removing DC components, bandpass filtering, and normalizing the eight-dimensional time-series data vector. A bandpass filter of 0.1Hz to 100Hz is used to filter out high-frequency noise, and the Z-score normalization method is used to eliminate the dimensional differences between different sensors, outputting a preprocessed normalized data sequence.

[0007] Specifically, the multi-scale feature extraction layer uses three one-dimensional convolutional kernels with different receptive field sizes of 16, 32, and 64 sampling points respectively to process the standardized data sequence in parallel. Each convolutional kernel contains 128 filters to extract local feature patterns at different time scales and generate a multi-scale feature tensor set.

[0008] Specifically, the feature fusion integration layer fuses three feature vectors at different scales using a weighted average method. The fusion weights are dynamically determined based on the distinguishing ability of each scale feature in historical fault samples, generating a 384-dimensional comprehensive feature vector and outputting the fused feature representation.

[0009] The adaptive gated expert network model consists of 5 expert subnetworks and 1 gated network. Each expert subnetwork is composed of a hidden layer with 128 neurons. The gated network dynamically selects the number of experts to be activated based on the input features.

[0010] The adaptive frequency adjustment mechanism specifically involves calculating the current system load strength parameter value. When the system load strength parameter value a∈[0, 0.3], the signal sampling processing frequency is reduced to 50% of the original frequency. When a∈(0.3, 0.7], the current signal sampling processing frequency is maintained. When a∈(0.7, 1.0], the signal sampling processing frequency is increased to 150% of the original frequency.

[0011] The abnormal state detection and confirmation mechanism specifically involves calculating the current electrical parameter deviation measurement function value, dynamically adjusting the capacity factor parameter of the adaptive gating expert network model based on the range of the deviation measurement function value, setting the capacity factor to 1.2 when the deviation measurement function value b∈[0, 0.3), setting the capacity factor to 1.5 when b∈[0.3, 0.7), and setting the capacity factor to 2.0 when b∈[0.7, 1.0].

[0012] The abnormal state detection and confirmation mechanism includes a step where, if five consecutive detection results exceed a preset threshold of 0.8, the state is determined to be a continuous abnormal state and an emergency warning process is initiated. If only a single detection result exceeds the threshold, the detection frequency is increased for secondary confirmation.

[0013] The steps of the dynamic adjustment mechanism for gating weights are as follows: calculate the system stability equilibrium value, adjust the weight allocation function of the gating network according to the equilibrium value range, use a uniform weight allocation function when the equilibrium value c∈[0, 0.4), use an exponential weight allocation function when c∈[0.4, 0.8), and use a logarithmic weight allocation function when c∈[0.8, 1.0].

[0014] The communication interfaces include RS485 communication ports and Ethernet communication ports, and the control chip uses a 32-bit microcontroller.

[0015] The eight-dimensional time-series data vector is formed by the three-phase current data from the current sensor, the three-phase voltage data from the voltage sensor, the temperature data from the temperature sensor, and the frequency data from the frequency detector. It includes the numerical combination of eight electrical parameters in the time series: A-phase current, B-phase current, C-phase current, A-phase voltage, B-phase voltage, C-phase voltage, temperature, and system frequency. The resulting multi-dimensional vector is used to characterize the operating status of the power distribution system.

[0016] Among them, the circular buffer refers to a ring-shaped data storage structure in which new data overwrites the oldest data when the buffer is full, and a fixed-length data window is maintained for real-time analysis and processing.

[0017] The Z-score standardization method involves calculating the mean and standard deviation of the data, transforming the original data into standard normal distribution data with a mean of 0 and a standard deviation of 1, thus eliminating numerical differences between parameters of different dimensions.

[0018] The receptive field size is specifically the range of data points covered by the convolutional kernel in the time series. Different receptive field sizes are used to capture fault feature patterns over different time spans.

[0019] The weighted average method refers to a feature fusion method that assigns different weights to each feature vector based on its importance, and then calculates the weighted sum to achieve effective integration of multi-scale information.

[0020] The distribution box body shell is also equipped with a status indicator array. The control chip is electrically connected to the status indicator array, which is used to display the operating status and fault warning information of each branch of the distribution box.

[0021] This invention achieves synchronous monitoring of key parameters such as current, voltage, temperature, and frequency. Compared with traditional single-parameter monitoring methods, it can comprehensively capture the operating status characteristics of the power distribution system, providing a sufficient data foundation for accurate fault identification and significantly improving the data integrity and status characterization capability of fault early warning. Through the collaborative work of a multi-scale feature extraction layer and an adaptive gated expert network model, this invention utilizes convolutional kernels with different receptive field sizes to extract fault feature patterns across multiple time scales in parallel. Through specialized analysis of multiple expert sub-networks and intelligent selection of the gated network, it achieves accurate identification of complex fault patterns and effective capture of early, weak fault signs, overcoming the technical limitations of traditional methods in identifying complex faults and significantly reducing false alarm and false negative rates. By introducing an adaptive processing mechanism and an abnormal state detection and confirmation mechanism, this invention dynamically adjusts the monitoring strategy and network parameters based on the system load intensity and deviation metric function value, achieving adaptive response to different operating states. Through continuous verification and multi-level confirmation mechanisms, it avoids occasional misjudgments, ensuring the reliability and accuracy of fault early warning results. In summary, this invention solves the technical problem of insufficient accuracy in power distribution system fault early warning mentioned in the background art. Attached Figure Description

[0022] Figure 1 This is a flowchart of the method of the present invention.

[0023] Figure 2 This is a schematic diagram of the overall structure of the distribution box with fault warning function in Example 2. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0025] A power distribution box with fault early warning function includes: a power distribution box body, a current sensor, a voltage sensor, a temperature sensor, a frequency detector, a communication interface, and a control chip. The power distribution box body internally houses multiple circuit breaker units and terminals. The current sensor is installed at the output terminal of each circuit breaker unit, collecting current by surrounding the conductor with a clip-on mounting method. The voltage sensor is connected between the input terminal of the power distribution box body and the output terminals of each branch circuit, converting the high-voltage signal into a standard 4-20mA or 0-10V DC signal output. The temperature sensors are installed at key heat-generating components inside the power distribution box body, monitoring temperature through digital signal output. The frequency detector is connected to the output terminal of the voltage sensor via a signal line to achieve system... The system accurately measures the frequency; the communication interface includes an RS485 communication port and an Ethernet communication port, connecting to various sensor devices via an RS485 bus and communicating with the host computer system via the Ethernet port; the control chip uses a 32-bit microcontroller, acquiring and interacting with the current sensor, voltage sensor, temperature sensor, and frequency detector via the RS485 communication bus, transmitting data via an SPI interface to the communication interface module, and controlling the display status of the status indicator array via GPIO ports; the control chip includes a fault early warning intelligent analysis module, used to identify weak signals of early fault characteristics in the power distribution system, extracting deep fault characteristics through a multi-scale parallel processing architecture, and achieving early fault warning;

[0026] like Figure 1 As shown, the fault early warning intelligent analysis module is used to perform the following steps:

[0027] S01. Construct a multi-dimensional electrical parameter sampling base layer, synchronously collect three-phase current data from the current sensor, three-phase voltage data from the voltage sensor, temperature data from the temperature sensor, and frequency data from the frequency detector at a frequency of 200Hz, form an eight-dimensional time-series data vector, establish a circular buffer with a length of 1024 sampling points, and generate a time-series data matrix for subsequent processing.

[0028] S02. Implement a signal preprocessing enhancement layer to remove DC components, perform bandpass filtering and normalization on the eight-dimensional time-series data vector, use a bandpass filter of 0.1Hz to 100Hz to filter out high-frequency noise, eliminate the difference in dimensions between different sensors through the Z-score normalization method, and output the preprocessed normalized data sequence.

[0029] S03. Establish a multi-scale feature extraction layer, using three one-dimensional convolutional kernels with different receptive field sizes of 16, 32, and 64 sampling points respectively, to process the standardized data sequence in parallel. Each convolutional kernel contains 128 filters to extract local feature patterns at different time scales and generate a multi-scale feature tensor group.

[0030] S04. The feature fusion and integration layer is executed, which fuses the feature vectors of three different scales by weighted averaging. The fusion weights are dynamically determined according to the distinguishing ability of each scale feature in historical fault samples, generating a 384-dimensional comprehensive feature vector and outputting the fused feature representation.

[0031] S05. Run the lightweight fault identification model and input the comprehensive feature vector into the adaptive gated expert network model. The adaptive gated expert network model contains 5 expert subnetworks and 1 gated network. Each expert subnetwork consists of a hidden layer with 128 neurons. The gated network dynamically selects the number of experts to be activated according to the input features and outputs the initial fault probability distribution.

[0032] S06. Execute the adaptive processing frequency adjustment mechanism, calculate the current system load strength parameter value, when the system load strength parameter value is in the range of [0, 0.3], reduce the signal sampling processing frequency to 50% of the original frequency, when the system load strength parameter value is in the range of (0.3, 0.7], maintain the current signal sampling processing frequency, when the system load strength parameter value is in the range of (0.7, 1.0], increase the signal sampling processing frequency to 150% of the original frequency, and re-acquire the eight-dimensional time series data vector;

[0033] S07. Implement an abnormal state detection and confirmation mechanism, calculate the current electrical parameter deviation measurement function value, and dynamically adjust the capacity factor parameter of the adaptive gating expert network model according to the range of the deviation measurement function value. When the deviation measurement function value belongs to [0, 0.3), the capacity factor is set to 1.2; when the deviation measurement function value belongs to [0.3, 0.7), the capacity factor is set to 1.5; when the deviation measurement function value belongs to [0.7, 1.0], the capacity factor is set to 2.0. When the detection results exceed the preset threshold of 0.8 for 5 consecutive times, it is determined to be a continuous abnormal state and the emergency warning process is initiated. When only a single detection result exceeds the threshold, the detection frequency is increased for secondary confirmation.

[0034] S08. Implement a dynamic adjustment mechanism for gating weights, calculate the system stability equilibrium value, and adjust the weight allocation function of the gating network according to the equilibrium value range. When the equilibrium value is in [0, 0.4), a uniform weight allocation function is used; when the equilibrium value is in [0.4, 0.8), an exponential weight allocation function is used; and when the equilibrium value is in [0.8, 1.0], a logarithmic weight allocation function is used to optimize the expert network selection strategy.

[0035] S09. Generate a hierarchical early warning output layer, which is divided into four levels: normal state, attention state, early warning state, and emergency state according to the adjusted fault probability value. At the same time, it outputs the fault type identification result and the expected fault occurrence time, and sends the early warning information to the host computer through the communication interface to complete the fault early warning process.

[0036] The eight-dimensional time-series data vector specifically includes the numerical combination of eight electrical parameters in the time series: A-phase current, B-phase current, C-phase current, A-phase voltage, B-phase voltage, C-phase voltage, temperature, and system frequency. The resulting multi-dimensional vector is used to characterize the operating state features of the power distribution system.

[0037] The circular buffer refers to a ring-shaped data storage structure where new data overwrites the oldest data when the buffer is full, maintaining a fixed-length data window for real-time analysis and processing.

[0038] The Z-score standardization method includes calculating the mean and standard deviation of the data, converting the original data into standard normal distribution data with a mean of 0 and a standard deviation of 1, and eliminating numerical differences between parameters of different dimensions.

[0039] The receptive field size is specifically the range of data points covered by the convolutional kernel in the time series. Different receptive field sizes are used to capture fault feature patterns over different time spans.

[0040] The weighted average method refers to a feature fusion method that assigns different weights to each feature vector based on its importance, and then calculates the weighted sum to achieve effective integration of multi-scale information.

[0041] The specific structure of the adaptive gated expert network model is a four-layer architecture consisting of an input layer, a gated network layer, an expert network layer, and an output layer. The input layer receives a 384-dimensional comprehensive feature vector. The gated network layer calculates the expert selection weights through a fully connected layer with 64 neurons. The expert network layer contains 5 parallel expert subnetworks, each consisting of 128 neurons. The output layer generates a fault probability distribution through weighted fusion. The gated network uses the softmax activation function to ensure that the sum of the weights is 1, and the expert network uses the ReLU activation function to improve its nonlinear expression capability.

[0042] The steps for establishing the training dataset for the adaptive gating expert network model specifically include collecting eight-dimensional time-series data vectors of 1000 distribution boxes under normal and fault conditions. 30 days of continuous data are collected for each device. The fault samples include five types: overload fault, short-circuit fault, poor contact fault, frequency offset fault, and temperature rise fault. 200 samples are collected for each fault type. 1000 samples are collected for normal operation. All samples are labeled and preprocessed. The training set, validation set, and test set are divided in a 7:2:1 ratio to establish a complete training dataset containing 3000 samples.

[0043] The specific steps for training the adaptive gated expert network model include: first, initializing the network parameters using random weights; then, performing gradient descent training using the Adam optimization algorithm, with a learning rate of 0.001, a batch size of 32, and 200 training epochs; using the cross-entropy loss function to calculate the difference between the predicted result and the true label; employing an early stopping mechanism to prevent overfitting during training; stopping training when the validation set loss does not decrease for 10 consecutive epochs; and finally, saving the model parameters with the highest validation set accuracy as the optimal model.

[0044] The system load strength parameter is used to adjust the signal sampling and processing frequency. The system load strength parameter is calculated based on the average value of the three-phase current RMS value, the average value of the three-phase voltage RMS value, and the difference between the temperature and the standard temperature to obtain a system load strength parameter value. When the system load strength parameter value belongs to different ranges, it is used to adjust the signal sampling and processing frequency parameter by different frequency multiples.

[0045] The deviation metric function is used to adjust the capacity factor parameter of the adaptive gated expert network model. The deviation metric function is calculated based on the Euclidean distance between the current eight-dimensional time-series data vector and the historical baseline data vector, the difference between the current system frequency and the standard frequency of 50Hz, and the difference between the current temperature and the normal temperature. When the deviation metric function value falls within different ranges, it is used to adjust the number of expert activations in the adaptive gated expert network model using different capacity factor values.

[0046] The gating weight function is used to adjust the gating weight allocation of the adaptive gating expert network model. The gating weight function is calculated based on the system voltage fluctuation variance, current harmonic content, and temperature change rate to obtain a balance value. When the balance value is in the range [0, 0.4), it is used to adjust the gating weight parameters using a uniform weight allocation function. When the balance value is in the range [0.4, 0.8), it is used to adjust the gating weight parameters using an exponential weight allocation function. When the balance value is in the range [0.8, 1.0], it is used to adjust the gating weight parameters using a logarithmic weight allocation function.

[0047] The emergency warning process refers to a rapid response mechanism triggered when the system detects a continuous abnormal state. It includes a complete process of immediately sending an emergency alarm signal to the communication interface, activating the red warning lights of the status indicator array, recording the time of the fault and the parameter status, and starting the backup monitoring mode.

[0048] Optionally, the outer casing of the distribution box is also provided with a status indicator array, and the control chip is electrically connected to the status indicator array. The status indicator array is used to display the operating status and fault warning information of each branch of the distribution box.

[0049] Key heat-generating components in the distribution box include busbar connection points, busbar-circuit breaker connection terminals, and cable inlet terminals. These locations are prone to heat generation due to contact resistance and high current flow. Circuit breaker contacts, circuit breaker outgoing terminals, and disconnector contacts generate heat due to arcing and contact resistance during current flow and switching operations. The cores of current transformers and voltage transformers, and the electronic components of relay protection devices generate heat due to core losses and power consumption. Transformer winding connection terminals, including the high-voltage and low-voltage sides, as well as the transformer core and windings themselves, are also significant heat sources. Furthermore, other components such as the dielectric losses of reactive power compensation capacitors, the coil resistance losses of current-limiting reactors, and the connection points between the grounding busbar and the equipment casing also generate varying degrees of heat during normal operation. These locations require real-time monitoring using temperature sensors to prevent overheating faults and ensure the safe and reliable operation of the power distribution system. By default, temperature sensors are deployed at busbar connection points, busbar-circuit breaker connection terminals, and cable inlet terminals.

[0050] The temperature collected is the average temperature from multiple temperature sensors.

[0051] The specific implementation methods of the above steps are described in detail below.

[0052] The specific implementation of step S01 involves constructing a multi-dimensional electrical parameter sampling base layer. This step first configures a 32-bit microcontroller's analog-to-digital converter module, setting the sampling clock frequency to 200Hz to ensure synchronous acquisition of electrical parameters from the power distribution system. The current sensor utilizes the Hall effect principle to convert the AC current signal passing through the conductor into a corresponding analog voltage signal, with an output range of 0–5V, corresponding to a current measurement range of 0–1000A. The voltage sensor, based on the resistive voltage divider principle, proportionally reduces the high-voltage signal to a safe low-voltage signal, achieving signal conversion through a precision resistor network, and outputting a standard 4–20mA DC signal. The temperature sensor uses a digital temperature chip, achieving temperature measurement based on the PN junction temperature characteristics, with a measurement accuracy of ±0.5℃ and a measurement range of -40℃ to 125℃. The frequency detector uses a zero-crossing detection circuit to measure the system frequency, with a detection accuracy of ±0.01Hz. The control chip synchronously reads the values ​​of eight sensors at preset time intervals, forming an eight-dimensional numerical vector containing three-phase current, three-phase voltage, temperature and frequency. Each vector element has a precision of 16 bits. The sampled data is stored in a circular buffer of 1024 sampling points in chronological order. When the buffer is full, the newly collected data automatically overwrites the oldest stored data, keeping the real-time data window continuously updated.

[0053] The specific implementation of step S02 involves implementing a signal preprocessing enhancement layer. This step uses digital signal processing technology to filter and standardize the original sampled data. DC component removal is based on the principle of a high-pass filter, with a cutoff frequency set to 0.1Hz to eliminate the effects of sensor zero-point drift and DC bias. The bandpass filter uses a Butterworth filter design, achieving a passband characteristic of 0.1Hz to 100Hz through cascading low-pass and high-pass filters, filtering out power frequency harmonic interference and high-frequency noise components. The filter order is set to 4th order to ensure sufficient cutoff frequency steepness. Normalization uses the Z-score normalization algorithm, calculating the mean and standard deviation of each dimension of data to transform the original data into a standard normal distribution with a mean of 0 and a standard deviation of 1, eliminating dimensional and numerical range differences between different physical quantities and improving the processing effect of subsequent feature extraction algorithms. The preprocessed data is stored in a standardized data sequence buffer, providing standardized input data for subsequent multi-scale feature extraction.

[0054] The specific implementation of step S03 involves establishing a multi-scale feature extraction layer. This step utilizes the principle of a one-dimensional convolutional neural network to extract features at different time scales. The convolution operation moves across the time series using a sliding window, with receptive field sizes set to 16, 32, and 64 sampling points, corresponding to time window lengths of 0.08 seconds, 0.16 seconds, and 0.32 seconds, respectively, to capture fault feature patterns of short-term transients, medium-term changes, and long-term trends. Each scale is configured with 128 convolutional kernels, whose weights are obtained through random initialization and training. The convolution stride is set to 1 to ensure the completeness of feature extraction. After the convolution operation, the ReLU activation function is applied to introduce nonlinear transformation capabilities, enhancing the network's ability to express complex fault patterns. Max pooling is performed with a pooling window size of 2, reducing feature dimensionality while retaining key feature information. The three convolutional layers at different scales process the standardized data sequence in parallel, outputting feature tensors of corresponding scales to form a multi-scale feature tensor set, providing a multi-level fault feature representation for subsequent feature fusion.

[0055] The specific implementation of step S04 involves executing a feature fusion integration layer. This step utilizes the attention mechanism to effectively integrate multi-scale features. The weighted average fusion algorithm calculates the classification accuracy of feature vectors at each scale in historical fault samples to determine the fusion weight coefficients. The weight calculation is based on the information gain principle, evaluating the contribution of each scale feature to fault identification. Short-scale feature weights are typically set to 0.3, medium-scale feature weights to 0.4, and long-scale feature weights to 0.3. The weight coefficients are dynamically adjusted according to the feature distribution characteristics of different fault types. Feature fusion employs an element-wise weighted summation method, fusing three 128-dimensional scale feature vectors into a single 384-dimensional comprehensive feature vector. Feature normalization is applied during the fusion process to ensure that the numerical range of the fused feature vector remains within a reasonable range, preventing a single scale feature from dominating the entire feature representation. The fused comprehensive feature vector contains operational status information of the power distribution system at different time scales, providing rich input features for subsequent fault identification models.

[0056] The specific implementation of step S05 involves running a lightweight fault identification model. This step achieves intelligent fault mode identification based on the principle of a hybrid expert system. The adaptive gated expert network model employs a divide-and-conquer strategy, dynamically selecting the most suitable combination of expert sub-networks for the current input features through the gated network. The gated network receives a 384-dimensional comprehensive feature vector as input, calculates the activation weights of each expert through a fully connected layer with 64 neurons, and applies the softmax activation function to ensure weight normalization, with the weight sum always equal to 1. Five expert sub-networks correspond to a fault identification task, each expert sub-network adopts a single hidden layer structure containing 128 neurons, and uses the ReLU activation function to enhance nonlinear mapping capabilities. The outputs of the expert networks are weighted and fused through the gated weights to generate the final fault probability distribution vector, with a vector dimension of 5, corresponding to the probability values ​​of the four fault types and the normal state. The model inference process uses the forward propagation algorithm, with a computational complexity of O(n), meeting the requirements for real-time processing.

[0057] The specific implementation of step S06 involves executing an adaptive frequency adjustment mechanism. This step dynamically adjusts the sampling frequency based on the system load adaptive principle. The system load strength parameter is calculated based on power theory, using a comprehensive evaluation of the average three-phase effective current, the average three-phase effective voltage, and the temperature deviation. The effective current value is calculated using the root mean square algorithm, summing the squares of the instantaneous current values ​​within one cycle and then taking the square root. The temperature deviation is calculated based on the difference between the current temperature and the rated operating temperature of 25°C. The load strength parameter value is normalized and mapped to the 0-1 range. When the parameter value is below 0.3, it indicates a light load state, and the sampling frequency is adjusted to 100Hz to reduce the system's computational load. When the parameter value is between 0.3 and 0.7, the standard sampling frequency of 200Hz is maintained. When the parameter value exceeds 0.7, it indicates a heavy load state, and the sampling frequency is increased to 300Hz to enhance the sensitivity to monitoring changes in system state. Frequency adjustment is achieved by modifying the division coefficient of the microcontroller timer, ensuring smooth switching of the sampling frequency.

[0058] The specific implementation of step S07 is to implement an abnormal state detection and confirmation mechanism. This step is based on the statistical process control principle to achieve reliable detection of abnormal states. The deviation measurement function is calculated using the Euclidean distance algorithm, which measures the degree of deviation of the system state by the distance between the current eight-dimensional time-series data vector and the historical baseline data vector. The baseline data vector is obtained by statistical averaging of 1000 samples under normal operating conditions and serves as a reference standard for normal system operation. Frequency deviation is calculated by the absolute difference between the current frequency and the standard power frequency of 50Hz, and temperature deviation is evaluated based on the difference between the current temperature and the normal operating temperature. The deviation measurement function value is normalized to the range of 0 to 1. When the deviation value is below 0.3, the capacity factor is set to 1.2, activating part of the expert network. When the deviation value is in the range of 0.3 to 0.7, the capacity factor is set to 1.5, increasing the participation of the expert network. When the deviation value exceeds 0.7, the capacity factor is set to 2.0, activating the entire expert network to improve detection accuracy. Continuous anomaly judgment adopts a sliding window statistical method. When five consecutive detection results exceed the threshold of 0.8, a continuous abnormal state judgment is triggered, and the emergency warning process is initiated. A single anomaly detection employs a two-stage verification mechanism, increasing the detection frequency to twice the original frequency for verification.

[0059] The specific implementation of step S08 involves implementing a dynamic adjustment mechanism for gating weights. This step optimizes the weight allocation of the gating network based on the principle of system stability assessment. The system stability equilibrium value calculation comprehensively considers three indicators: voltage fluctuation variance, current harmonic content, and temperature change rate. Voltage fluctuation variance assesses voltage stability by calculating the variance of the three-phase voltage within a time window; a larger variance indicates greater system instability. Current harmonic content is calculated based on the Fast Fourier Transform algorithm to analyze the frequency domain characteristics of the current signal, calculating the total harmonic distortion rate of the 2nd to 50th harmonic components. The temperature change rate is calculated using the first derivative of temperature with respect to time, reflecting the changing trend of the system's thermal state. When the equilibrium value is below 0.4, the system is in a stable state, and a uniform weight allocation function is used, with equal weights for each expert network to ensure even participation in decision-making by all experts. When the equilibrium value is in the range of 0.4 to 0.8, the system stability is moderate, and an exponential weight allocation function is used, allocating weights exponentially based on the experts' historical performance, with high-performing experts receiving higher weights. When the balance value exceeds 0.8, the system becomes unstable. A logarithmic weighting function is adopted to enhance the leading role of the best expert and improve the recognition accuracy under complex working conditions.

[0060] The specific implementation of step S09 involves generating a tiered early warning output layer. This step utilizes fuzzy logic principles to classify fault states. Fault probability values, after being adjusted by gating weights, are input to the tiered judgment module, where state classification is achieved by setting threshold ranges. Normal states correspond to fault probability values ​​of 0–0.2, watchful states 0.2–0.5, warning states 0.5–0.8, and emergency states 0.8–1.0. Fault type identification is based on the maximum value index of the probability distribution vector; when the probability value of a certain fault type is significantly higher than other types, it is determined to be the corresponding fault type. The expected fault occurrence time is calculated using a trend extrapolation algorithm, predicting the possible time window of the fault based on the current growth trend of the fault probability value and historical fault development patterns. The early warning information is encapsulated using a standard communication protocol format, including fields such as timestamp, fault level, fault type, probability value, and expected occurrence time. It is sent to the host computer system via RS485 and Ethernet interfaces, completing the entire fault early warning processing flow.

[0061] The detailed structure of the adaptive gated expert network model comprises a four-layer architecture. The input layer receives a 384-dimensional comprehensive feature vector, which, after feature standardization, is input to both the gated network layer and the expert network layer. The gated network layer employs a single-hidden-layer fully connected neural network structure with 64 neurons. The tanh activation function enhances the smoothness of the gated decision. The output layer uses the softmax activation function to generate selection weights for five experts, with weight values ​​ranging from 0 to 1 and summing to 1. The expert network layer contains five parallel expert subnetworks, each using the same network structure but with different weight parameters. Each expert subnetwork consists of an input layer, a hidden layer, and an output layer. The hidden layer contains 128 neurons and uses the ReLU activation function to improve the network's non-linear expressive power. The output layer contains five neurons corresponding to five state classification results. The output layer weights and fuses the prediction results of each expert using the gate weights to generate the final fault probability distribution vector, where each element represents the probability of occurrence of the corresponding fault type.

[0062] The detailed steps for establishing the training dataset began with a data collection phase. 1000 distribution boxes of different capacities and load types were selected as data sources. Each device was continuously monitored for 30 days at a sampling frequency of 200Hz, resulting in a total of 2.592 billion data sampling points. Fault sample collection was achieved through a combination of artificial fault injection and natural fault recording. Overload faults were induced by gradually increasing the load current to 120%–150% of the rated value; short-circuit faults were induced by creating phase-to-phase or phase-to-ground short circuits under safe conditions; poor contact faults were induced by loosening connection terminals; frequency deviation faults were induced by adjusting the power supply frequency to the 48–52Hz range; and temperature rise faults were induced by increasing the ambient temperature or blocking heat dissipation channels. In the data labeling phase, experienced electrical engineers labeled the data samples according to the fault occurrence time and type, with labeling accuracy down to the second-level timestamp. In the data preprocessing phase, all samples underwent the same filtering and standardization operations to ensure consistency between the training data and the actual application data. The dataset is divided into a training set of 2100 samples, a validation set of 600 samples, and a test set of 300 samples in chronological order to ensure the temporal independence of the test data and avoid data leakage issues.

[0063] The adaptive gating expert network model is suitable for solving the technical problem of fault early warning in distribution boxes because it can effectively handle the complexity and diversity of fault modes in power distribution systems. Compared with traditional single neural network models, gating expert networks achieve accurate identification of different fault types through specialized division of labor. Each expert network focuses on learning the feature representation of different fault modes, avoiding the feature confusion problem of a single model when dealing with multiple types of faults. Compared with existing support vector machine classification methods, the adaptive gating mechanism can dynamically select the most suitable expert combination according to the characteristics of the input features, improving the model's generalization ability and recognition accuracy under complex operating conditions. The adaptability of the gating network allows the model to dynamically adjust the participation level of experts according to the feature distribution of the current electrical parameters, reducing computational resource consumption when the system is stable and enhancing detection capabilities in abnormal states, achieving a balance between computational efficiency and detection accuracy. The parallel processing architecture of the expert network fully utilizes the multi-core computing power of modern processors, significantly improving real-time performance compared to serial processing methods, meeting the stringent requirements of power distribution systems for fault early warning response time.

[0064] The model training employs an end-to-end learning approach, simultaneously optimizing the gating network and expert networks. The optimal gating strategy and expert parameters are automatically learned through backpropagation. The Adam optimization algorithm is used during training, with a learning rate of 0.001, a batch size of 32, and a training cycle of 200 epochs. The loss function is a weighted combination of cross-entropy loss and gating sparsity loss. Cross-entropy loss ensures classification accuracy, while sparsity loss encourages the gating network to select a small number of high-quality experts for decision-making, improving model interpretability and computational efficiency. An early stopping mechanism is implemented by monitoring the trend of the validation set loss. Training automatically stops when the validation set loss does not decrease for 10 consecutive epochs to prevent overfitting, and the model parameters with the highest validation set accuracy are saved as the final deployment version.

[0065] It should be explained that the key technical ideas of this invention mainly include four aspects: multi-scale parallel feature extraction, adaptive gated expert network, dynamic parameter adjustment, and hierarchical early warning mechanism. Multi-scale parallel feature extraction technology extracts short-term, medium-term, and long-term fault feature patterns simultaneously using convolutional kernels with different receptive field sizes. Compared with traditional single-scale feature extraction methods, it can more comprehensively capture multi-timescale features in the fault evolution process of the power distribution system, especially showing stronger identification ability for early weak signals of progressive faults, overcoming the limitation of traditional methods that can only identify obvious fault features. Adaptive gated expert network technology achieves accurate identification of different types of faults through specialized division of labor and dynamic selection mechanisms. Compared with traditional single classifier methods, it avoids the problem of mutual interference of multiple types of fault features in the same model, improving the identification accuracy of various faults, especially significantly enhancing the ability to distinguish similar fault patterns. Dynamic parameter adjustment technology adjusts model parameters in real time according to the system operating status and load conditions. Compared with traditional methods with fixed parameters, it can adapt to changes in the operating conditions of the power distribution system, reducing computational resource consumption under light load and improving monitoring accuracy under heavy load, achieving a dynamic balance between computational efficiency and detection performance. The tiered early warning mechanism achieves progressive status judgment from normal to emergency through multi-level threshold settings and trend prediction. Compared with the traditional binary judgment method, it provides more granular status information, enabling operation and maintenance personnel to take corresponding countermeasures according to different early warning levels, effectively avoiding false alarms and missed alarms.

[0066] The synergistic effect of the four key technical approaches described above generates significant technical advantages. Multi-scale feature extraction provides rich input information for the gating expert network, enabling each expert to learn more accurate fault mode representations. Dynamic parameter adjustment ensures the optimal performance of the gating mechanism under different operating conditions. The hierarchical early warning mechanism transforms complex fault identification results into intuitive state levels, forming a complete technical chain from data acquisition to early warning output. Compared with existing single-point monitoring or simple threshold alarm systems, the collaborative technical architecture of this invention achieves comprehensive, multi-level, and intelligent monitoring of the operating status of the power distribution system, especially with significant advantages in early fault identification. It can provide accurate early warning information before fault characteristics become apparent, providing strong technical support for the preventive maintenance and safe operation of the power distribution system.

[0067] Optionally, the current sensor uses an Acrel AKH-0.66 series switchable current transformer, the voltage sensor uses a Chint NVT-10 series voltage transmitter, the temperature sensor uses a Delixi DT100 series intelligent temperature sensor, the frequency detector uses a Huali HLF series frequency meter, and the communication interface uses an Acrel...

[0068] The Anet-2E4SM series communication management unit uses a Renesas R5F562N8 series 32-bit microcontroller with a main frequency of 100MHz, and has 1MB of flash memory and 256KB of random access memory.

[0069] Optionally, the distribution box also includes power distribution equipment such as circuit breakers, contactors, relays, disconnect switches, fuses, and current transformers. Circuit breakers, as protective switches, can connect, carry, and disconnect current under normal circuit conditions. They have three protection functions: overload protection, short-circuit short-delay protection, and short-circuit long-delay protection. When a serious overload, short circuit, or undervoltage fault occurs, they can automatically disconnect the circuit to protect downstream equipment and personnel safety. Contactors, as control switches, can achieve remote control and frequent load switching. They are mainly used for motor start / stop control, heater control, and dual power supply switching, achieving remote operation by controlling the coil voltage. Relays are mainly used in control circuits for signal detection, transmission, conversion, or processing. They have multiple pairs of normally open and normally closed contacts and can be used in different control circuits to achieve logic control and signal transmission functions. Disconnect switches are used to disconnect circuits without load current, providing a clearly visible disconnect point during equipment maintenance to ensure the personal safety of maintenance personnel. They are mainly used to disconnect and isolate a portion of the network before maintenance. As an overcurrent protection device, a fuse melts and breaks the circuit when the current in the circuit exceeds its rated value, protecting equipment from overload and short-circuit damage. Instrument transformers include current transformers and voltage transformers. Current transformers proportionally transform large currents into smaller currents for use by measuring instruments and relay protection devices, while voltage transformers proportionally transform high voltages into low voltages to provide voltage signals for measuring instruments, relay protection devices, and automatic devices. These devices, through proper configuration and coordinated operation, together constitute a complete power distribution protection and control system.

[0070] Optionally, the distribution box may also include fault diagnosis equipment such as insulation monitoring devices, residual current monitors, partial discharge detection devices, arc fault circuit breakers, intelligent power monitoring instruments, fault indicators, and live-line detection devices. Insulation monitoring devices, such as the Acrel AIM-T series industrial insulation monitoring devices, are mainly used to monitor the insulation status of the IT system to ground, and promptly alarm when an insulation fault occurs in the system to remind electrical maintenance personnel to troubleshoot the fault. Residual current monitors, such as the Acrel ARCM series intelligent online power monitoring devices, can detect leakage current changes in TT and TN system circuits below 0.4kV in real time, and issue an alarm signal when the leakage current exceeds a set threshold to prevent electric shock accidents and electrical fires. Partial discharge detection devices, such as the Acrel APD series partial discharge monitoring devices and Guangzhou Zhihao...

[0071] The MCSG-PD series products utilize UHF, ultrasonic AE, and TEV sensor technologies to monitor partial discharge in switchgear and ring main units. Timely and effective partial discharge monitoring can prevent accidents. Arc fault circuit breakers, such as the Chint NXB-63 series and Delixi CDM1-63 series, can detect arc faults and automatically trip, preventing electrical fires caused by arc faults. Intelligent power monitoring instruments, such as the Acrel PZ series multi-functional power meters and the Chint PMC series network power meters, integrate multiple functions including power metering, power quality analysis, and harmonic analysis, enabling real-time monitoring of power distribution system operating parameters and data recording and analysis. Fault indicators, such as the NARI WXH series and XJ WFZ series, are installed at critical points on the line. When a short circuit or ground fault occurs, they can quickly locate the faulty section, shortening fault location time. Live-line detection devices include Huaguang Technology HG series infrared thermometers, Zhongshi Holdings ZS series ultraviolet detectors, and Baoding Sidel STL series ultrasonic detectors, which are used to detect the operating status of equipment without power interruption and to discover potential fault hazards. These domestically produced fault diagnosis devices provide comprehensive technical support for the safe and reliable operation of power distribution systems through multi-dimensional monitoring and intelligent analysis.

[0072] It should be noted that the fault early warning intelligent analysis module of this invention can identify five main types of faults: overload faults, short circuit faults, poor contact faults, frequency deviation faults, and temperature rise faults. Overload faults are identified by continuous monitoring of the effective value of the three-phase current exceeding the rated value; short circuit faults are judged by the characteristic combination of current sudden change and voltage drop; poor contact faults are detected by the local temperature rise and voltage fluctuation pattern caused by increased resistance; frequency deviation faults are identified by the degree of deviation of the system frequency from the standard 50Hz frequency; and temperature rise faults are determined by the abnormal temperature rise trend and hot spot distribution monitored by temperature sensors. At the same time, the module can also identify complex fault modes formed by the coupling of multiple basic fault types, such as the combined faults of overload and temperature rise, and the associated faults of poor contact and voltage fluctuation. Through the multi-expert collaborative analysis mechanism of the adaptive gating expert network model, accurate identification, classification, and early warning of single and complex faults are achieved.

[0073] It should be noted that this invention also solves the following technical problem: the low efficiency of computing resource utilization in traditional distribution box monitoring systems. In existing technologies, distribution box monitoring systems typically employ fixed-frequency data acquisition and static computing resource allocation strategies. Regardless of changes in system operating status, the same monitoring intensity and computational overhead are maintained, leading to a significant waste of computing resources during normal operation with light system loads, while insufficient computing power is provided to ensure monitoring accuracy during periods of heavy system load or abnormal fluctuations. This invention utilizes an adaptive processing frequency adjustment mechanism to dynamically adjust the signal sampling and processing frequency based on different ranges of system load intensity parameters. When the load intensity is low, the processing frequency is reduced to 50% of the original frequency to save computing resources; when the load intensity is high, the processing frequency is increased to 150% of the original frequency to improve monitoring accuracy, thus achieving intelligent on-demand allocation of computing resources. Meanwhile, through the capacity factor dynamic adjustment function in the abnormal state detection and confirmation mechanism, the number of activated expert networks and the computational scale are adaptively adjusted according to the range of the deviation measurement function value. Under normal operating conditions, the participation of expert networks is reduced to reduce computational overhead, while under abnormal conditions, the number of expert networks is increased to improve analytical capabilities. This effectively solves the problem of low utilization efficiency of computing resources and achieves the optimal balance between monitoring accuracy and computational efficiency.

[0074] This invention also solves the technical problem of difficult identification of complex fault modes in power distribution systems. In practical applications, power distribution systems often experience complex fault situations where multiple fault types occur simultaneously or influence each other, such as the coupling of overload faults and temperature rise faults, the correlation between poor contact faults and voltage fluctuations, and the superposition of frequency offset faults and harmonic interference. Traditional single-fault-type diagnostic methods are difficult to handle these complex combinations of fault modes, easily leading to misjudgments or omissions. This invention uses a multi-expert parallel processing architecture based on an adaptive gated expert network model, setting up four expert sub-networks responsible for the identification and analysis of different types of faults. Each expert acquires professional identification capabilities for the corresponding fault modes through deep learning training. The gated network intelligently calculates the activation weights and participation levels of each expert based on the comprehensive characteristic state of the current system. The dynamic adjustment mechanism of the gate weights optimizes the expert selection strategy by calculating the system stability balance value in real time and using different weight allocation functions such as uniform weight allocation, exponential weight allocation, and logarithmic weight allocation. This ensures that multiple relevant experts can be activated simultaneously for collaborative analysis in complex complex fault scenarios, accurately identifying the correlation between primary and secondary fault types, and significantly improving the accuracy and reliability of complex fault mode identification and analysis.

[0075] Specifically, the principle of this invention is as follows: The core principle that enables this invention to solve the technical problem of insufficient accuracy in power distribution system fault early warning lies in the adoption of a technical architecture of multi-dimensional fusion analysis and intelligent feature recognition. The fundamental reason for the inaccuracy of traditional methods lies in incomplete information acquisition and overly simplistic analysis methods. This invention, through the construction of an eight-dimensional time-series data vector, simultaneously collects multiple key electrical parameters such as three-phase current, three-phase voltage, temperature, and frequency, forming a comprehensive monitoring of the operating status of the power distribution system. Compared with single-parameter monitoring, it can obtain richer and more complete system status information, laying a solid data foundation for accurate fault identification. The multi-scale feature extraction technology uses three one-dimensional convolutional kernels with different receptive field sizes to process in parallel, capturing fault feature patterns of different time spans such as short-term fluctuations, medium-term trends, and long-term changes, overcoming the limitations of traditional single-time-window analysis, and can identify early fault signs hidden in complex signals. The feature fusion integration layer effectively integrates multi-scale information through a weighted average method, forming a comprehensive representation of fault features, significantly improving the richness and discriminative power of feature expression. Adaptive gating expert network models are a key technology for improving early warning accuracy. Each expert subnetwork is specifically responsible for identifying a type of fault. The gating network dynamically selects the most suitable expert combination based on input features, achieving specialized analysis and collaborative decision-making. Compared to traditional single-model methods, it has a stronger ability to distinguish fault types. The adaptive adjustment mechanism dynamically monitors multiple dimensions such as system load intensity, deviation metric function value, and system stability balance value, adjusting network parameters and monitoring strategies in real time to ensure optimal identification performance under different operating environments. The abnormal state detection and confirmation mechanism effectively avoids misjudgments through continuous verification and multi-level confirmation, ensuring the reliability of early warning results.

[0076] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0077] The specific implementation of step S01 is to construct a multi-dimensional electrical parameter sampling base layer, and the specific representation of the eight-dimensional time-series data vector is as follows:

[0078] D(t)=[I A (t), I B (t), I C (t), U A (t), U B (t), U C (t), T(t), F(t)] T ;

[0079] In the formula, D(t) is the eight-dimensional time-series data vector at time t; I A (t), I B (t), I C(t) represents the instantaneous current values ​​of phases A, B, and C at time t, respectively, in A; U A (t), U B (t), U C T(t) represents the instantaneous voltage values ​​of phases A, B, and C at time t, in V; T(t) represents the temperature value at time t, in °C; and F(t) represents the system frequency value at time t, in Hz.

[0080] The formula for constructing a time series data matrix is:

[0081] X=[D(t-1023Δt), D(t-1022Δt),…, D(t-Δt), D(t)];

[0082] In the formula, X is a 1024×8 dimensional time series data matrix; Δt is the sampling time interval, with a value of 0.005s, corresponding to a sampling frequency of 200Hz; matrix X adopts a circular buffer storage structure, which overwrites the oldest data row when new data arrives.

[0083] The specific implementation of step S02 is to implement a signal preprocessing enhancement layer. The specific calculation formula for the Z-score normalization method is as follows:

[0084]

[0085] In the formula, X norm (i, j) represents the element in the i-th row and j-th column of the standardized data matrix; X(i, j) represents the element in the i-th row and j-th column of the original data matrix; i is the row index of the data matrix, ranging from 1 to 1024; j is the column index of the data matrix, ranging from 1 to 8, corresponding to the eight electrical parameter dimensions; μ j σ is the mean of the j-th dimension of data; j Let be the standard deviation of the j-th dimension of data.

[0086] Where, μ j and σ j The calculation formulas are as follows:

[0087]

[0088] In the formula, N is the number of sampling points, with a value of 1024; the Z-score standardization process eliminates the differences in dimensions between different sensors, so that the mean of the standardized data is 0 and the standard deviation is 1.

[0089] The specific implementation of step S03 is to establish a multi-scale feature extraction layer. For eight-dimensional input data, the convolution operation is performed independently in each dimension. The mathematical expression of a one-dimensional convolution operation is:

[0090]

[0091] In the formula, The output feature value of the k-th convolutional kernel at the l-th scale at position n for the j-th dimension data; Let be the m-th weight parameter of the k-th convolutional kernel at the l-th scale; n be the position index of the convolutional output sequence; and m be the weight index within the convolutional kernel, ranging from 0 to K. l -1; K l X represents the length of the convolutional kernel at the l-th scale, taking values ​​of 16, 32, and 64 respectively; norm (n+m, j) is the value in the (n+m)th row and jth column of the standardized data matrix; The bias parameter is the k-th convolutional kernel at the l-th scale; l∈{1,2,3} represents three different scales, and k∈{1,2,…,128} represents 128 convolutional kernels at each scale.

[0092] The features after processing with the ReLU activation function are represented as follows:

[0093]

[0094] In the formula, These are the feature values ​​after processing by the activation function; the ReLU activation function introduces nonlinear transformation capabilities, enhancing the network's ability to represent complex fault modes.

[0095] The specific implementation of step S04 is to execute the feature fusion integration layer. The mathematical expression of the weighted average fusion is:

[0096] F fused =w1·F 1 +w2·F 2 +w3·F 3 ;

[0097] In the formula, F fused The fused 384-dimensional comprehensive feature vector; F 1 F 2 F 3 These are three 128-dimensional feature vectors at different scales; w1, w2, and w3 are fusion weight coefficients, satisfying w1 + w2 + w3 = 1 and w i >0, the weight coefficient is determined by statistically analyzing the classification accuracy of features at each scale in historical fault samples.

[0098] The calculation of the fusion weight is based on the information gain principle, and the specific formula is as follows:

[0099]

[0100] In the formula, IG lThe information gain value of the l-th scale feature is calculated by the contribution of this scale feature to the fault identification accuracy in historical fault samples, and the value ranges from 0.2 to 0.5.

[0101] The specific implementation method of step S05 is the same as described above, and will not be repeated in detail here.

[0102] The specific implementation of step S06 is to execute an adaptive frequency adjustment mechanism, first calculating the effective values ​​of the current and voltage of each phase:

[0103]

[0104] In the formula, I RMS，p U represents the effective value of the p-phase current, in A. RMS，p N represents the effective value of the p-phase voltage, in V. c The number of sampling points in one period is 40; p∈{A, B, C} represents three phases; k is the index of the sampling point in the period.

[0105] The formula for calculating the system load strength parameter is:

[0106]

[0107] In the formula, L intensity This is the system load strength parameter value, ranging from 0 to 1; I rated This is the rated current value, in amperes (A), determined based on the rated power of the distribution box; U rated This is the rated voltage value, typically 380V; T current This is the current temperature value, in °C; T ref This is a reference temperature value, with a value of 25℃; T max The maximum allowable temperature is 80℃; α, β, and γ are weighting coefficients, with values ​​of 0.4, 0.4, and 0.2 respectively, satisfying α+β+γ=1. The weighting coefficients are determined by experimentally analyzing the correlation between the influence of each parameter on the load strength.

[0108] The formula for adjusting the sampling frequency is:

[0109] f new =f original ·R freq ;

[0110] In the formula, f new The adjusted sampling frequency is expressed in Hz; f original The original sampling frequency is 200Hz; R freq For frequency adjustment factor, based on L intensity The value is determined when L intensity When R ∈ [0, 0.3] freq=0.5, when L intensity When R ∈ (0.3, 0.7] freq =1.0, when L intensity When R ∈ (0.7, 1.0] freq =1.5.

[0111] The specific implementation of step S07 is to implement an abnormal state detection and confirmation mechanism, and the calculation formula of the deviation measurement function is as follows:

[0112]

[0113] In the formula, D deviation The deviation measurement function value ranges from 0 to 1; D(t) is the current eight-dimensional time series data vector; D baseline The historical baseline data vector is obtained by statistically averaging 1000 normally operating samples, calculated using the following formula: Where D normal (q) is the eight-dimensional data vector of the q-th normal sample; ||D(t)-D baseline || represents the Euclidean distance between two vectors; F standard The standard power frequency is 50Hz; T normal The normal operating temperature is 25℃; ω1, ω2, and ω3 are weighting coefficients, with values ​​of 0.6, 0.3, and 0.1 respectively. The weighting is determined based on the statistical analysis of the contribution of each indicator to fault identification.

[0114] The formula for adjusting the capacity factor parameter is:

[0115] C factor =1.2·θ1+1.5·θ2+2.0·θ3;

[0116] In the formula, C factor θ1, θ2, θ3 are capacity factor parameters; θ1, θ2, θ3 are Boolean indicator variables, when D deviation When ∈ [0, 0.3), θ1 = 1, and the rest are 0. deviation When ∈ [0.3, 0.7), θ2 = 1; otherwise, 0. deviation When ∈[0.7, 1.0], θ3 = 1, and the rest are 0.

[0117] The specific implementation of step S08 is to implement a dynamic adjustment mechanism for the gating weight, first calculating the temperature change rate:

[0118]

[0119] In the formula, T current (t) represents the temperature value at the current moment, T current (t-Δt tempThe temperature value at the previous moment is Δt, which is derived from real-time monitoring data from temperature sensors. temp The time interval for calculating temperature changes is 1 second.

[0120] The formula for calculating the system stability equilibrium value is:

[0121]

[0122] In the formula, S balance The system stability equilibrium value is 0 to 1; Var(U) is the variance of the three-phase voltage; THD(I) is the total harmonic distortion of the current; φ1, φ2, and φ3 are weighting coefficients, with values ​​of 0.5, 0.3, and 0.2 respectively. The weighting is determined based on the importance of each indicator in the power system stability analysis theory.

[0123] The formula for calculating voltage variance is as follows:

[0124]

[0125] In the formula, M represents the number of voltage sampling points, with a value of 200; U RMS，p (s) represents the effective voltage value at the s-th sampling point of the p-th phase; is the mean of the effective values ​​of the voltage of phase p; s is the index of the sampling point within the voltage variance calculation window.

[0126] The formula for calculating the total harmonic distortion (THD) of current is:

[0127]

[0128] In the formula, I h Ih is the effective value of the h-th harmonic current, in A; I1 is the effective value of the fundamental current, in A; h is the harmonic order, ranging from 2 to 50.

[0129] The gating weight allocation function is based on S balance Values ​​are assigned using different strategies:

[0130] Uniform weighting function:

[0131] Exponential weighting function:

[0132] Log-weighting function:

[0133] In the formula, W gate，i Let η be the gating weight of the i-th expert; i∈{1,2,3,4,5} represents a network of 5 experts; η i The exponential weighting parameter, ranging from 1 to 3, is determined through cross-validation experiments to find the optimal value; ζi The logarithmic weight parameter has a value range of 2 to 5 and is obtained through optimization using a grid search algorithm.

[0134] The specific implementation method of step S09 is the same as described above, and will not be repeated in detail here.

[0135] It needs to be explained that the eight-dimensional time series data vector is D(t) = [I A (t), I B (t), I C (t), U A (t), U B (t), U C (t), T(t), F(t)] T This vector representation, constructed by simultaneously monitoring key electrical parameters such as current, voltage, temperature, and frequency, constitutes a complete state characteristic space of the power distribution system. Compared to traditional single-parameter or dual-parameter monitoring methods, this vector representation can comprehensively reflect changes in the operating state of the power distribution system, especially demonstrating stronger identification capabilities for complex and progressive faults, providing a rich foundation of input information for subsequent intelligent analysis algorithms. The principle of the Z-score standardization formula is based on the standard normal transformation in probability and statistics theory. By eliminating dimensional and numerical range differences between various physical quantities, multidimensional heterogeneous data is transformed into a unified standard normal distribution. Compared to traditional maximum-minimum normalization methods, Z-score standardization is more robust to outliers, maintains the statistical characteristics of the data distribution, and improves the training effect and generalization ability of subsequent machine learning algorithms. The principle of the one-dimensional convolution formula is based on the convolution theorem in signal processing. By extracting local feature patterns from time series using a sliding window approach, convolutional kernels of different lengths can capture fault features at different time scales. Compared to traditional fully connected networks, convolutional operations feature parameter sharing and local connectivity, significantly reducing the number of model parameters while maintaining sensitivity to temporal patterns. This makes it particularly suitable for multi-scale representation of fault features in power distribution systems. The weighted fusion formula is based on ensemble learning theory, F... fused =w1·F 1 +w2·F 2 +w3·F 3 By allocating information gain weights, multi-scale features are effectively integrated. Compared to simple feature stitching methods, weighted fusion dynamically allocates weights based on the contribution of each scale feature to fault identification, highlighting the role of important features, suppressing interference from redundant information, and improving the discriminative ability of the fused features. The principle of the system load intensity parameter formula is based on power system load theory. By comprehensively evaluating the operating load level of the power distribution system using three key indicators—current, voltage, and temperature—compared to a single current monitoring method, this comprehensive indicator can more accurately reflect the actual load state of the system, providing a scientific basis for adaptive adjustment of the sampling frequency and achieving optimal allocation of computing resources. The principle of the deviation measurement function formula is based on multivariate statistical analysis theory. This function measures the deviation of the current system state from the normal baseline by combining Euclidean distance and relative deviation. Compared with the traditional single-threshold judgment method, this function can quantify the severity of anomalies, providing continuous control signals for the dynamic adjustment of the model capacity factor and improving the sensitivity and accuracy of anomaly detection. The principle of the system stability equilibrium value formula is based on the theory of power system stability analysis. The stability of the system is comprehensively evaluated by three indicators: voltage variance, harmonic distortion, and temperature change rate. Compared with the traditional single stability index, this composite index can comprehensively reflect the dynamic characteristics of the system, provide accurate control basis for the adaptive adjustment of gating weights, and optimize the selection strategy of expert network under different operating conditions.

[0136] To better understand and implement this invention, a specific application scenario is provided below as an example 2: During the system deployment phase, the technical team first installed eight Hall effect current sensors inside the distribution box to monitor the three-phase current of the three main branch circuits. The current sensors have a measurement range of 0–2000A and an accuracy class of 0.2, and are installed at the output terminals of each circuit breaker unit using a clip-on method. The voltage sensors employ the capacitive voltage divider principle, connected between the 10kV input terminal and the 0.4kV output terminals of each branch circuit, converting the high-voltage signal into a standard 4–20mA DC signal output with a measurement accuracy of 0.1. The temperature sensors are PT100 platinum resistance type, installed at six key heat-generating locations, including busbar connection points, transformer terminals, and main switching equipment, with a temperature monitoring range of -40℃ to 120℃ and a resolution of 0.1℃. The frequency detector uses high-precision phase-locked loop technology to accurately measure the system frequency, with a measurement range of 45Hz to 55Hz and an accuracy of ±0.01Hz.

[0137] The control chip uses a 32-bit ARM Cortex-M4 microcontroller with a main frequency of 168MHz, and has 512KB of built-in flash memory and 192KB of random access memory. This control chip establishes connections with various sensor devices via an RS485 communication bus, with a communication rate set at 115200bps and a data acquisition cycle of 5ms. The communication interface module includes RS485 and Ethernet communication ports. The RS485 port uses differential signal transmission, with a transmission distance of up to 1200m, while the Ethernet port supports a transmission rate of 100Mbps. Data communication with the host computer monitoring system is achieved via the TCP / IP protocol.

[0138] The fault early warning intelligent analysis module begins executing multi-dimensional parameter acquisition tasks. The system synchronously acquires three-phase current data, three-phase voltage data, temperature data, and frequency data at a frequency of 200Hz, forming an eight-dimensional time-series data vector containing A-phase current, B-phase current, C-phase current, A-phase voltage, B-phase voltage, C-phase voltage, temperature, and system frequency. During actual operation, the A-phase current stabilizes at approximately 480A, the B-phase current at 465A, and the C-phase current at 492A. The three-phase voltages are 398V, 402V, and 395V respectively, the system frequency remains at 49.98Hz, and the average temperature inside the distribution box is 42℃. The system establishes a circular buffer with 1024 sampling points to achieve data storage and updates within a 5.12s time window. The final distribution box with fault early warning function is shown below. Figure 2 As shown.

[0139] The signal preprocessing enhancement layer performs DC component removal, bandpass filtering, and normalization on the eight-dimensional time-series data vector. The bandpass filter's frequency range is set to 0.1Hz–100Hz to filter out power frequency interference and high-frequency noise signals.

[0140] The Z-score normalization method normalizes data from different sensors, eliminating numerical differences between parameters with different dimensions such as current, voltage, temperature, and frequency. After normalization, the parameter values ​​are converted into standard normal distribution data with a mean of 0 and a standard deviation of 1.

[0141] The multi-scale feature extraction layer employs three one-dimensional convolutional kernels with different receptive field sizes: 16, 32, and 64 sampling points, corresponding to time windows of 0.08s, 0.16s, and 0.32s, respectively. Each convolutional kernel contains 128 filters, processing the normalized data sequence in parallel. The 16-point kernel primarily captures rapidly changing transient features, the 32-point kernel extracts feature patterns at medium time scales, and the 64-point kernel identifies trends over long time spans. The feature extraction results at the three scales generate 128-dimensional feature vectors, which are then weighted and averaged by the feature fusion layer to form a 384-dimensional comprehensive feature vector. The fusion weights are dynamically determined based on the discriminative power of each scale feature in historical fault samples, with a weight of 0.35 for the 16-point scale, 0.40 for the 32-point scale, and 0.25 for the 64-point scale.

[0142] The lightweight fault identification model employs an adaptive gated expert network architecture, comprising five expert subnetworks and one gated network. The gated network calculates expert selection weights through a fully connected layer of 64 neurons, using a softmax activation function to ensure the weights sum to 1. The five expert subnetworks, each consisting of 128 neurons, focus on different types of fault mode recognition and utilize a ReLU activation function to enhance nonlinear expressiveness. The output layer generates fault probability distributions for five types of faults—overload faults, short-circuit faults, poor contact faults, frequency shift faults, and temperature rise faults—through weighted fusion.

[0143] As shown in Table 1, the system recorded detection statistics for different fault types during actual operation.

[0144] Table 1 Fault Type Detection Statistics

[0145] Fault type Number of tests Confirm number of faults Detection accuracy Overload fault 156 142 0.91 Short circuit fault 23 21 0.91 Poor contact fault 89 78 0.88 Frequency offset fault 34 31 0.91 Temperature rise fault 67 59 0.88

[0146] The adaptive frequency adjustment mechanism dynamically adjusts the signal sampling and processing frequency based on the system load intensity parameters. The system load intensity parameter is calculated to be 0.65 based on the average three-phase current RMS value of 479A, the average three-phase voltage RMS value of 398V, and the temperature difference from the standard temperature of 25℃ (17℃). Since this value falls within the range of 0.3 to 0.7, the system maintains the current signal sampling and processing frequency of 200Hz. During peak electricity consumption periods in the shopping mall, when the system load intensity parameter rises to 0.82, the system automatically increases the sampling frequency to 300Hz to enhance its ability to monitor abnormal conditions.

[0147] The abnormal state detection and confirmation mechanism dynamically adjusts by calculating the electrical parameter deviation metric function. The deviation metric function is calculated based on the Euclidean distance between the current eight-dimensional time-series data vector and the historical baseline data vector, the difference of 0.02 Hz between the current system frequency and the standard frequency of 50 Hz, and the difference of 17 ℃ between the current temperature and the normal temperature, resulting in a deviation metric function value of 0.45. Since this value falls within the range of 0.3 to 0.7, the system sets the capacity factor parameter to 1.5 and adjusts the number of activated expert networks. When the system detects that five consecutive results exceed the preset threshold of 0.8, it determines it as a persistent abnormal state and initiates the emergency warning process.

[0148] The gating weight dynamic adjustment mechanism is based on an expert network selection strategy that optimizes the system stability balance value. The system voltage fluctuation variance is 2.3V. 2 With a current harmonic content of 4.2% and a temperature change rate of 0.8℃ / min, the calculated equilibrium value is 0.52. Since this value falls within the range of 0.4 to 0.8, the system employs an exponential weight allocation function to adjust the weight distribution of the gating network, enabling the expert network to adaptively adjust its activation strategy based on the current system state.

[0149] Table 2 shows the gating weight allocation of the system under different operating states.

[0150] Table 2. Statistics on Gate Control Weight Allocation

[0151]

[0152]

[0153] The tiered early warning output layer classifies warning levels based on the adjusted fault probability value. When the fault probability value is less than 0.3, it is considered a normal state, and the green indicator light illuminates. When the fault probability value is between 0.3 and 0.6, it is considered a state of concern, and the yellow indicator light flashes. When the fault probability value is between 0.6 and 0.8, it is considered a warning state, the orange indicator light remains illuminated, and the system sends a warning message to the host computer. When the fault probability value exceeds 0.8, it is considered an emergency state, the red indicator light flashes rapidly, and the emergency warning process is initiated.

[0154] Compared to traditional distribution box monitoring methods, this invention achieves in-depth fault feature mining through the collaborative analysis of multi-dimensional electrical parameters. Traditional methods mainly rely on single parameter thresholds for judgment, which are easily affected by environmental interference and load fluctuations, resulting in a high false alarm rate. The multi-scale parallel processing architecture adopted in this invention can simultaneously capture the fault evolution process at different time scales, and achieve accurate identification of complex fault modes through an adaptive gating expert network. The adaptive processing frequency adjustment mechanism dynamically adjusts the monitoring accuracy according to the system load status, ensuring high sensitivity detection under abnormal conditions while avoiding resource waste under normal conditions. The abnormal state detection confirmation mechanism reduces the risk of false alarms through multiple verifications, and the dynamic adjustment mechanism of gating weights enables the system to automatically optimize the detection strategy according to the operating environment. These technological innovations fundamentally solve the key problems of lagging fault identification, high false alarm rate, and poor adaptability in traditional power distribution monitoring systems, providing an effective technical means for the intelligent operation and maintenance of power distribution systems.

[0155] It should be noted that the variables involved in this invention are explained in detail in Table 3.

[0156] Table 3. Variable Explanation Table

[0157]

[0158]

[0159] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A distribution box with fault early warning function, comprising a distribution box body, a current sensor, a voltage sensor, a temperature sensor, a frequency detector, a communication interface, and a control chip, wherein, The distribution box contains multiple circuit breaker units and terminals. Current sensors are installed at the output terminals of each circuit breaker unit, voltage sensors are connected between the input terminals of the distribution box and the output terminals of each branch circuit, temperature sensors are installed at key heat-generating parts inside the distribution box, and frequency detectors are connected to the output terminals of the voltage sensors via signal lines. The key feature is that the control chip incorporates a fault early warning intelligent analysis module. This module synchronously acquires an eight-dimensional time-series data vector by constructing a multi-dimensional electrical parameter sampling base layer. This vector contains the numerical combinations of eight electrical parameters—A-phase current, B-phase current, C-phase current, A-phase voltage, B-phase voltage, C-phase voltage, temperature, and system frequency—over time. The resulting multi-dimensional vector characterizes the operating state of the power distribution system. The eight-dimensional time-series data vector... The system performs signal preprocessing and enhancement, outputting a preprocessed standardized data sequence. A multi-scale feature extraction layer is established, using one-dimensional convolutional kernels with different receptive field sizes to process the standardized data sequence in parallel, extracting local feature patterns at different time scales and generating multi-scale feature vectors. A feature fusion and integration layer is executed to fuse the multi-scale feature vectors into a comprehensive feature vector. A lightweight fault identification model is run, inputting the comprehensive feature vector into an adaptive gated expert network model to output an initial fault probability distribution. The adaptive gated expert network model contains 5 expert subnetworks and 1 gated network. Each expert subnetwork consists of a hidden layer with 128 neurons. The gated network dynamically selects the number of activated experts based on the input features. An adaptive processing frequency adjustment mechanism is executed to dynamically adjust the signal sampling and processing frequency based on the system load intensity parameter value. The formula for calculating the system load strength parameter is as follows: ； In the formula, This is the system load strength parameter value, which ranges from 0 to 1. This is the rated current value, in amperes (A), determined based on the rated power of the distribution box. This is the rated voltage value, typically 380V; This is the current temperature value, in °C. This is a reference temperature value, with a value of 25℃. This is the maximum permissible temperature value, which is 80℃. These are weighting coefficients, with values ​​of 0.4, 0.4, and 0.2 respectively. for Effective value of phase current, in amperes (A); for The effective value of the phase voltage, in volts (V). Indicates three phases; An abnormal state detection and confirmation mechanism is implemented to adjust the capacity factor parameters of the adaptive gating expert network model based on the deviation metric function value; The formula for calculating the deviation measurement function is as follows: ； In the formula, This is the deviation measurement function value, which ranges from 0 to 1; This is the current eight-dimensional time series data vector; The historical baseline data vector is obtained by statistically averaging 1000 normally operating samples, calculated using the following formula: ,in For the first An eight-dimensional data vector of a normal sample; This represents the Euclidean distance between two vectors; This is the standard power frequency, with a value of 50Hz. This is the normal operating temperature, with a value of 25℃. These are weighting coefficients, with values ​​of 0.6, 0.3, and 0.1 respectively. for Temperature value at time, for The system frequency value at that moment; Implement a dynamic adjustment mechanism for gating weights to adjust the weight allocation function of the gating network based on the system stability balance value; The formula for calculating the system stability equilibrium value is: ； In the formula, This is the system stability equilibrium value, ranging from 0 to 1; This represents the variance of the three-phase voltage. The total harmonic distortion of the current; These are weighting coefficients, with values ​​of 0.5, 0.3, and 0.2 respectively. The generated hierarchical early warning output layer is divided into four levels based on the adjusted fault probability value: normal state, attention state, early warning state, and emergency state.

2. The distribution box according to claim 1, characterized in that, The steps for constructing a multi-dimensional electrical parameter sampling base layer are as follows: synchronously collect three-phase current data from a current sensor, three-phase voltage data from a voltage sensor, temperature data from a temperature sensor, and frequency data from a frequency detector at a frequency of 200Hz; establish a circular buffer with a length of 1024 sampling points; and generate a time-series data matrix for subsequent processing.

3. The distribution box according to claim 2, characterized in that, The signal preprocessing enhancement steps specifically involve removing DC components, bandpass filtering, and normalizing the eight-dimensional time-series data vector. A bandpass filter of 0.1Hz to 100Hz is used to filter out high-frequency noise, and the Z-score normalization method is used to eliminate the dimensional differences between different sensors, outputting a preprocessed normalized data sequence.

4. The distribution box according to claim 3, characterized in that, The multi-scale feature extraction layer specifically uses three one-dimensional convolutional kernels with different receptive field sizes of 16, 32, and 64 sampling points respectively to process the standardized data sequence in parallel. Each convolutional kernel contains 128 filters to extract local feature patterns at different time scales and generate multi-scale feature vectors.

5. The distribution box according to claim 4, characterized in that, The feature fusion integration layer specifically involves fusing three feature vectors at different scales using a weighted average method. The fusion weights are dynamically determined based on the distinguishing ability of each scale feature in historical fault samples, generating a 384-dimensional comprehensive feature vector and outputting the fused feature representation.

6. The distribution box according to claim 5, characterized in that, The adaptive frequency adjustment mechanism involves calculating the current system load strength parameter value. When the system load strength parameter value a∈[0, 0.3], the signal sampling and processing frequency is reduced to 50% of the original frequency. When a∈(0.3, 0.7], the current signal sampling and processing frequency is maintained. When a∈(0.7, 1.0], the signal sampling and processing frequency is increased to 150% of the original frequency.

7. The distribution box according to claim 6, characterized in that, The steps of the abnormal state detection and confirmation mechanism are as follows: calculate the current electrical parameter deviation measurement function value, and dynamically adjust the capacity factor parameter of the adaptive gating expert network model according to the range of the deviation measurement function value. When the deviation measurement function value b∈[0, 0.3), the capacity factor is set to 1.2; when b∈[0.3, 0.7), the capacity factor is set to 1.5; and when b∈[0.7, 1.0], the capacity factor is set to 2.

0.

8. The distribution box according to claim 7, characterized in that, The abnormal state detection and confirmation mechanism also includes determining a continuous abnormal state and initiating an emergency warning process when five consecutive detection results exceed a preset threshold of 0.

8. When only a single detection result exceeds the threshold, the detection frequency is increased for secondary confirmation.

9. The distribution box according to claim 8, characterized in that, The steps of the dynamic adjustment mechanism for gating weights are as follows: calculate the system stability equilibrium value, adjust the weight allocation function of the gating network according to the equilibrium value range, use a uniform weight allocation function when the equilibrium value c∈[0, 0.4), use an exponential weight allocation function when c∈[0.4, 0.8), and use a logarithmic weight allocation function when c∈[0.8, 1.0].

Images