Fault monitoring system and method for low-voltage power distribution cabinet

Abstract

The invention relates to the technical field of power equipment monitoring, and discloses a low-voltage power distribution cabinet fault monitoring system and method. According to the system, multiple paths of original operation signals are collected and preprocessed in real time to form a standard time sequence data stream, frequency domain feature extraction, transient feature capture and load fluctuation analysis are carried out, and a dynamic multi-dimensional feature reference library is constructed; iteratively calculating a feature-fault association weight based on the feature library and a pre-stored fault case graph, and establishing a dynamic mapping relation network which can be corrected in real time; according to the fault probability distribution output by the mapping network, the sampling frequency, the feature priority and the diagnosis threshold are adaptively adjusted, and an optimal monitoring strategy is formed; and finally, generating a standardized diagnosis report containing fault types, positions and severity levels through online analysis. According to the invention, comprehensive perception of fault features, autonomous optimization of a diagnosis strategy and accurate output of a fault report are realized, and the intelligent level and reliability of fault monitoring of the low-voltage power distribution cabinet are significantly improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power equipment monitoring, in particular to a low-voltage power distribution cabinet fault monitoring system and method thereof. BACKGROUND

[0002] The current low-voltage power distribution cabinet fault monitoring mainly adopts the threshold alarm and regular maintenance mode. The existing technology relies on single parameter threshold judgment for operation state monitoring, and multi-source signals such as voltage, current and temperature cannot be deeply fused and analyzed. The feature extraction method is simple, and the frequency domain features, transient features and load characteristic features are not systematically combined and mined. The fault diagnosis model is fixed, based on fixed rules or simple classifiers, which cannot adapt to the change of equipment state. The monitoring strategy is static, and the sampling frequency, feature priority and diagnosis threshold cannot be dynamically adjusted according to the actual operation risk. The historical fault data is not effectively transformed into diagnostic knowledge. The existing method needs to solve the key technical problems of multi-source data fusion, intelligent feature extraction, adaptive monitoring and accurate diagnosis.

[0003] The traditional power distribution cabinet monitoring system has obvious deficiencies in accuracy and adaptability. The signal acquisition accuracy is low, and the key transient process features are easy to lose. The data preprocessing method is rough, and the noise interference affects the signal quality. The frequency domain analysis resolution is insufficient, and the harmonic feature extraction is incomplete. The transient process capture algorithm is complex, and the real-time processing efficiency is low. The load characteristic analysis dimension is single, and cannot reflect the complex power consumption mode. The feature fusion method is simple, and the multi-dimensional feature synergistic effect cannot be fully played. The benchmark library updating mechanism lags behind, and cannot adapt to the change of equipment state. The mapping relationship network training data is insufficient, and the fault mode correlation accuracy is limited. The monitoring strategy adjustment is experienced, and cannot accurately match the actual risk level. The diagnosis report generation template is fixed, and the information integrity is insufficient. SUMMARY

[0004] The purpose of the present application is to provide a low-voltage power distribution cabinet fault monitoring system and method thereof to solve the problems raised in the background art.

[0005] To achieve the above purpose, the present application provides a low-voltage power distribution cabinet fault monitoring system, which comprises: A signal acquisition and preprocessing module is used for real-time acquisition of multi-channel original operation signals of a low-voltage power distribution cabinet, and denoising and normalization processing of the multi-channel original operation signals to generate a standard time sequence signal data stream; A multi-dimensional feature generation module is used for receiving the standard time sequence signal data stream and synchronously executing signal frequency domain feature extraction, transient process feature capture and load characteristic fluctuation analysis, and constructing a dynamically updated multi-dimensional feature benchmark library after fusing the extracted features; a fault mode dynamic mapping module, configured to establish and correct a dynamic mapping relationship network between the features and the fault modes by iteratively calculating the correlation weight between the features and the fault modes based on the multi-dimensional feature benchmark library and in combination with a pre-stored fault case atlas; a monitoring strategy adaptive generation module, configured to automatically adjust a sampling frequency of signal monitoring, a priority of feature extraction, and a logic judgment threshold of fault diagnosis according to a current fault mode probability distribution output by the dynamic mapping relationship network, and to generate an optimal monitoring strategy; a fault decision and output module, configured to execute the optimal monitoring strategy, perform online analysis on real-time signals, generate a diagnostic report containing a fault type, a fault location, and a fault severity level, and output the diagnostic report after standardization and encapsulation.

[0006] Preferably, the implementation of the signal acquisition and preprocessing module comprises: a plurality of types of sensors deployed inside a low-voltage power distribution cabinet are used to collect original operation signals including current, voltage, temperature, and vibration signals; the collected original operation signals are sequentially subjected to moving average filtering and wavelet threshold denoising processing to suppress environmental noise and electromagnetic interference; each of the denoised signals is divided by its corresponding rated range value to complete amplitude normalization, and a sampling time stamp is unified to generate the standard time series signal data stream.

[0007] Preferably, the implementation of the multi-dimensional feature generation module comprises: the input standard time series signal data stream is subjected to fast Fourier transform to calculate its frequency spectrum energy distribution, and the energy proportion of a specific frequency band is extracted as a frequency domain feature; a window sliding method is used to capture waveform mutation points in the standard time series signal data stream, and the amplitude change rate and duration of the mutation points are calculated as transient process features; statistical characteristics of the standard time series signal data stream within a set time window are analyzed, including mean, variance, and skewness, as load characteristic fluctuation features; the frequency domain features, the transient process features, and the load characteristic fluctuation features are aligned and spliced in time sequence to form a high-dimensional feature vector, which is stored in the multi-dimensional feature benchmark library.

[0008] Preferably, the implementation of the fault mode dynamic mapping module comprises: the latest high-dimensional feature vector is read from the multi-dimensional feature benchmark library; the high-dimensional feature vector is subjected to similarity matching with historical feature vectors in a pre-stored fault case atlas, and the pre-stored fault case atlas contains a feature vector set corresponding to a known fault type; Based on the matching result, the gradient descent method is used to iteratively update the contribution weight of each feature dimension to different fault modes; According to the updated contribution weight, a dynamic mapping relationship network reflecting the nonlinear relationship between features and fault modes is constructed.

[0009] Preferably, the implementation mode of the monitoring strategy adaptive generation module comprises: The dynamic mapping relationship network is calculated for the current input high-dimensional feature vector, and a set of probability values corresponding to different fault modes is output; According to the size of the probability value, the set of fault modes most likely to occur is determined; Query the strategy knowledge base to obtain the recommended monitoring parameters matched with the set of fault modes most likely to occur, the recommended monitoring parameters including sampling frequency, feature extraction item list and diagnostic threshold; Based on the recommended monitoring parameters, the optimal monitoring strategy for guiding the next monitoring period is generated.

[0010] Preferably, the implementation mode of the fault decision and output module comprises: According to the sampling frequency and feature extraction items set by the optimal monitoring strategy, real-time feature calculation is performed on the subsequent incoming standard time series signal data stream; The calculated real-time features are input into the dynamic mapping relationship network to obtain real-time fault mode probability distribution; Select the fault mode whose probability exceeds the diagnostic threshold in the optimal monitoring strategy as the candidate fault; Combine the topological structure information of the power distribution cabinet to perform positioning analysis on the candidate fault, and determine the severity level according to its probability value and historical data; Integrate the finally determined fault type, fault location and fault severity level into a structured diagnostic report.

[0011] Preferably, the fast Fourier transform of the input standard time series signal data stream comprises the following steps: Set the window length and overlap ratio of fast Fourier transform, and divide the standard time series signal data stream into multiple time windows; Apply the Hanning window function to each time window to perform windowing processing to reduce spectral leakage; Perform fast Fourier transform algorithm to convert time domain signal to frequency domain signal and obtain complex form of frequency spectrum data; Calculate the modulus of the frequency spectrum data to obtain the power spectral density distribution; Extract the power spectral density value in the pre-defined frequency band range as the frequency domain feature vector.

[0012] Preferably, the iterative updating of the contribution weight of each feature dimension to different fault modes by using the gradient descent method comprises the following steps: Initialize the weight matrix corresponding to each feature dimension to different fault modes, set the learning rate and the maximum number of iterations; Construct a loss function based on the difference between the predicted fault mode and the real fault mode under the current weight matrix; Calculate the gradient of the loss function with respect to the weight matrix to obtain the gradient direction; Update the weight matrix in the opposite direction of the gradient according to the learning rate to reduce the loss function value; Repeat the gradient calculation and weight update steps until the maximum number of iterations is reached or the loss function converges within a threshold.

[0013] Preferably, the calculation of the dynamic mapping relationship network on the current input high-dimensional feature vector comprises the following steps: Input the current input high-dimensional feature vector into the input layer of the dynamic mapping relationship network for linear transformation and bias addition; Perform non-linear activation function processing through multiple hidden layers to pass the feature representation layer by layer; Apply the softmax function in the output layer to convert the hidden layer output into a probability distribution form; Output a set of probability values corresponding to different fault modes, each probability value representing the possibility of the current input belonging to the corresponding fault mode.

[0014] Preferably, the present application further comprises a low-voltage power distribution cabinet fault monitoring method, which includes all the modules and method processes of the above-mentioned low-voltage power distribution cabinet fault monitoring system.

[0015] Compared with the prior art, the present application has the following advantages: By synchronously performing frequency domain feature extraction, transient process feature capture and load characteristic fluctuation analysis on the standard time sequence signal data stream, multi-dimensional feature fusion is realized. The frequency domain feature extraction uses Fourier transform to obtain the frequency spectrum characteristics, the transient process feature is identified by sudden change detection, and the load characteristic analysis uses time series modeling to extract the operation rules. The multi-dimensional features are processed by weighted fusion and dimension reduction to construct a dynamically updated feature reference library, improving the accuracy of state representation.

[0016] Based on the feature reference library and the pre-stored fault case atlas, a dynamic mapping relationship network between features and fault modes is established through iterative calculation. The mapping network uses a graph neural network structure, and the correlation weight is optimized by a gradient descent algorithm to realize accurate matching between features and fault modes. The network supports online learning and adapts to new fault modes.

[0017] According to the fault mode probability distribution output by the dynamic mapping network, the monitoring strategy parameters are automatically adjusted. The sampling frequency is dynamically adjusted according to the fault risk level, the feature extraction priority is rearranged according to the correlation degree, and the diagnosis threshold is adaptively optimized according to the confidence. Through the adaptive adjustment of the strategy parameters, the optimal configuration of the monitoring resources is realized.

[0018] The optimized monitoring strategy is executed to perform online analysis on real-time signals, and generate standardized diagnostic reports. The dynamic mapping relationship network is used to identify fault types, the fault points are located through topology analysis, and the severity level is evaluated based on the influence range. The report generation adopts a standardized template to ensure information integrity and system compatibility.

[0019] Through the synergistic effect of multi-dimensional feature fusion, dynamic mapping network and adaptive strategy, the fault monitoring precision and efficiency are improved. Feature fusion enhances the state representation ability, mapping network improves the diagnosis accuracy, adaptive strategy optimizes resource utilization, and finally realizes precise and efficient fault monitoring. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 The working principle diagram of the low-voltage power distribution cabinet fault monitoring system described in the application; Figure 2 The flowchart implemented by the signal acquisition and preprocessing module; Figure 3 The flowchart implemented by the multi-dimensional feature generation module; Figure 4 The radar chart of the low-voltage power distribution cabinet fault mode probability distribution in the strategy generation stage; Figure 5 The correlation diagram of the sampling frequency and the monitoring cycle dynamic adjustment. DETAILED DESCRIPTION

[0021] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0022] Please refer to Figure 1The application provides a low-voltage power distribution cabinet fault monitoring system and method, the system comprising: a signal acquisition and preprocessing module, a multi-dimensional feature generation module, a fault mode dynamic mapping module, a monitoring strategy adaptive generation module, and a fault decision and output module. The signal acquisition and preprocessing module is responsible for real-time acquisition of multiple original operating signals of the low-voltage power distribution cabinet, and performs denoising and normalization processing on these signals to generate a standard time sequence signal data stream. The multi-dimensional feature generation module receives the standard time sequence signal data stream, synchronously performs signal frequency domain feature extraction, transient process feature capture, and load characteristic fluctuation analysis, and constructs a dynamically updated multi-dimensional feature benchmark library after fusing the extracted features. The fault mode dynamic mapping module, based on the multi-dimensional feature benchmark library and in combination with pre-stored fault case atlases, establishes and real-time corrects a dynamic mapping relationship network between features and fault modes by iteratively calculating the correlation weight between features and fault modes. The monitoring strategy adaptive generation module automatically adjusts the sampling frequency of signal monitoring, the priority of feature extraction, and the logic judgment threshold of fault diagnosis according to the current fault mode probability distribution output by the dynamic mapping relationship network, and generates an optimal monitoring strategy. The fault decision and output module executes the optimal monitoring strategy to perform online analysis on real-time signals, generates a diagnostic report containing fault type, fault location, and fault severity level, and outputs the diagnostic report after standardization packaging.

[0023] Embodiment 1: refer to Figure 2 In specific implementation, through the deployment of multiple types of sensors inside the low-voltage power distribution cabinet, original operating signals including current, voltage, temperature, and vibration signals are collected. The multiple types of sensors are configured to continuously monitor the operating parameters of the power distribution cabinet to ensure the comprehensiveness and real-time nature of the data. The sensor output is preliminarily amplified and filtered through a signal conditioning circuit to adapt to the input range of the subsequent processing module. In some embodiments, the multiple types of sensors include current transformers, voltage dividers, thermocouples, and accelerometers, which are installed on busbars, switch contacts, and cabinet structures to capture changes in different physical quantities. The sensor data is converted into digital signals by a high-precision analog-to-digital converter for processing by a microprocessor. Optionally, the sensor sampling frequency is dynamically adjusted according to signal characteristics, for example, high sampling rates are used for current and voltage signals, while lower sampling rates are used for temperature and vibration signals to balance processing load and data accuracy.

[0024] In specific implementation, the collected original operating signals are sequentially subjected to sliding average filtering and wavelet threshold denoising processing to suppress environmental noise and electromagnetic interference. Sliding average filtering smooths high-frequency noise by calculating the average value of the signal within a moving window, while wavelet threshold denoising eliminates impulse interference by decomposing the signal through wavelet transform and threshold processing of detail coefficients. It can be understood that the mathematical expression of sliding average filtering is:

[0025] wherein: represents the value of the original operating signal at the sampling point , represents the value of the filtered signal at the sampling point , represents the length of the sliding window (100 for power frequency signals and 32 for transient signals), represents the weighting coefficient specific to the power distribution cabinet signal , for power frequency, for sampling frequency), used to enhance the retention of power frequency signals and suppress harmonic interference. Wavelet threshold denoising involves multi-scale wavelet decomposition of the signal, processing of high-frequency coefficients through a soft threshold function, and reconstruction of the signal to retain useful components. In specific implementations, the specific implementation of the soft threshold function includes threshold processing of high-frequency coefficients after wavelet decomposition, by setting a suitable threshold parameter, setting high-frequency coefficients with absolute values less than the threshold to zero, and performing shrinkage processing on high-frequency coefficients with absolute values greater than or equal to the threshold, i.e. subtracting the threshold value but retaining the sign, thereby effectively suppressing noise components while retaining the main characteristics of the signal. The mathematical expression of this process is:

[0026] where: x represents the high-frequency coefficient after wavelet decomposition, T is the threshold parameter, is the sign function, is the coefficient after soft threshold processing.

[0027] The threshold parameter can be adjusted adaptively according to the signal noise level, for example based on the statistical distribution of wavelet coefficients or by setting an empirical value, to ensure the balance between denoising effect and signal fidelity. The high-frequency coefficients after soft threshold processing are combined with the unprocessed low-frequency coefficients, and the signal is reconstructed through wavelet inverse transform to complete the denoising process and generate a cleaner standard time sequence signal data stream. In some embodiments, the window length of the moving average filter is pre-set according to the frequency characteristics of the signal, for example, a fixed window is used for power frequency signals and an adaptive window is used for transient signals, and the wavelet threshold denoising selects a suitable wavelet basis and threshold rule to optimize the denoising effect.

[0028] In specific implementations, the de-noised signals are divided by their corresponding rated range values respectively to complete amplitude normalization, and the sampling time stamps are unified to generate a standard time sequence signal data stream. Amplitude normalization scales the signals to a unified range, facilitating feature extraction and comparison, while time stamp unification aligns the data of each channel through a synchronous clock source to ensure timing consistency. In some embodiments, the rated range values are pre-stored based on sensor nominal values, such as dividing the current signal by the rated current value, dividing the voltage signal by the rated voltage value, dividing the temperature signal by the maximum range value, and dividing the vibration signal by the maximum acceleration value. Optionally, the sampling time stamps are marked by a high-precision timer, and an interpolation method is used to compensate for the slight delay between channels to form a synchronous standard time sequence signal data stream.

[0029] Embodiment 2: Refer to Figure 3 In specific implementations, the input standard time sequence signal data stream is subjected to a fast Fourier transform to calculate its spectral energy distribution, and the energy proportion of a specific frequency band is extracted as a frequency domain feature. Fast Fourier transform converts time domain signals into frequency domain representation, thereby revealing the frequency composition characteristics of the signals. In some embodiments, the window length of the fast Fourier transform is set to 1024 sampling points, and the overlap ratio is 50%. The standard time sequence signal data stream is divided into multiple consecutive time windows, and the signal data in each time window is windowed using a Hanning window function to reduce spectral leakage. The fast Fourier transform algorithm is executed to convert the time domain signal into complex frequency spectrum data, calculate the modulus square of the frequency spectrum data to obtain the power spectral density distribution, extract the power spectral density values within a predefined frequency band range, and calculate their proportion of total energy to form a frequency domain feature vector.

[0030] In specific implementations, a window sliding method is used to capture waveform mutation points in the standard time sequence signal data stream, and the amplitude change rate and duration of the mutation points are calculated as transient process features. The window sliding method detects sharp changes in signal waveform by continuously moving the analysis window. It can be understood that the amplitude change rate formula for transient characteristics of power distribution cabinet short circuit, overload, etc. is as follows:

[0031] Wherein: represents the signal amplitude at time point , represents the signal amplitude at time point , , is the interval between the two adjacent sampling time points (not more than 1ms), represents the variance of the current signal, represents the reference variance of the signal when the power distribution cabinet is running normally (obtained by statistical data), and the variance correction term enhances the discrimination between fault transients and normal fluctuations.

[0032] In some embodiments, the window sliding uses a fixed-length analysis window, for example containing 64 sampling points, and identifies abrupt changes by calculating the first-order difference of the signal within the window. The rate of change of amplitude and duration are quantified as numerical features to describe the intensity and persistence of the transient process.

[0033] In practical implementation, the statistical characteristics of the standard time-series signal data stream within a set time window are analyzed, including the mean, variance, and skewness, as load characteristic fluctuation features. These statistical characteristics reflect the distribution pattern and variation pattern of the signal over time. Optionally, the time window length is set to 1 second. The arithmetic mean of the signal sample values ​​within the window is calculated as the mean; the average of the squared deviations of the sample values ​​from the mean is calculated as the variance; and the third central moment of the asymmetry in the sample value distribution is calculated as the skewness. It can be understood that the mean describes the baseline level of the signal, the variance characterizes the fluctuation range of the signal, and the skewness indicates the shape characteristics of the signal distribution. These statistical quantities together constitute the numerical representation of the load characteristic fluctuation features.

[0034] In practical implementation, frequency domain features, transient process features, and load characteristic fluctuation features are aligned and concatenated according to time series to form a high-dimensional feature vector, which is then stored in a multi-dimensional feature benchmark library. Feature alignment ensures that features from different sources maintain consistency in timestamps. Optionally, feature alignment uses interpolation methods to compensate for the time delay in each feature extraction process. Feature concatenation connects the frequency domain feature vector, transient process feature vector, and load characteristic fluctuation feature vector in the dimensional direction to form a unified high-dimensional feature vector. The high-dimensional feature vector is assigned a timestamp identifier and stored in the multi-dimensional feature benchmark library in a circular buffer structure to achieve dynamic data updates and historical backtracking.

[0035] Example 3: In specific implementation, the latest high-dimensional feature vector is read from a multi-dimensional feature benchmark library. This high-dimensional feature vector includes frequency domain features, transient process features, and load characteristic fluctuation features extracted from standard time-series signal data streams. The multi-dimensional feature benchmark library stores feature data in a circular buffer structure to ensure real-time access to the latest vectors. The high-dimensional feature vector is then matched with historical feature vectors in a pre-stored fault case map. The pre-stored fault case map contains a set of feature vectors corresponding to known fault types. The similarity matching uses a cosine similarity calculation method to calculate the directional consistency between the current high-dimensional feature vector and historical feature vectors in the feature space. In some embodiments, the pre-stored fault case map is constructed through an offline training phase, covering feature samples of various typical fault modes. The similarity matching results are used to filter out the subset of historical cases most relevant to the current input, serving as a reference for weight updates. It can be understood that cosine similarity calculation is based on vector dot product and modulus operations, with an output value range between -1 and 1. The closer the value is to 1, the more similar the feature vectors are.

[0036] In practice, based on the matching results, gradient descent is used to iteratively update the contribution weights of each feature dimension to different failure modes. Gradient descent optimizes the weight matrix by minimizing the prediction error, thereby enhancing the accuracy of the association between features and failure modes. The weight matrix for each feature dimension corresponding to different failure modes is initialized, and the learning rate and maximum number of iterations are set. The dimensions of the weight matrix correspond to the length of the feature vector and the number of failure modes. The learning rate controls the step size of the weight update, and the maximum number of iterations prevents infinite loops. A loss function is constructed, based on the difference between the predicted failure mode and the actual failure mode under the current weight matrix. The loss function uses cross-entropy to measure the degree of difference between the predicted probability distribution and the true label. The formula for the cross-entropy loss function is as follows:

[0037] in: Indicates the loss value. This represents the number of categories of failure modes. Indicates the first The true label (0 or 1) for each failure mode. Indicates the prediction is the first The probability value of each failure mode.

[0038] The gradient of the loss function with respect to the weight matrix is ​​calculated to obtain the gradient direction. This gradient calculation is performed using the backpropagation algorithm, involving the chain rule for calculating partial derivatives. The weight matrix is ​​then updated in the opposite direction of the gradient according to the learning rate, reducing the loss function value. The weight update formula is:

[0039] in: This represents the current weight matrix. This represents the updated weight matrix. Indicates the learning rate. Indicates the loss function at The gradient is calculated at a given point. The gradient calculation and weight update steps are repeated until the maximum number of iterations is reached or the loss function converges within a threshold, which is preset to a value where the change in the loss function is less than a specific value. In some embodiments, the learning rate employs an adaptive adjustment strategy, with an initial value of 0.01, dynamically scaled according to the rate of decrease of the loss function during iteration, and a maximum number of iterations set to 1000 to ensure computational efficiency. Optionally, the weight matrix is ​​initialized using a random normal distribution with a mean of 0 and a standard deviation of 0.1 to avoid getting trapped in local optima.

[0040] In practical implementation, fault diagnosis during the daily operation of low-voltage distribution cabinets is particularly important in locations requiring rapid identification of various typical fault modes such as overload, short circuit, and grounding faults, including industrial plants, commercial buildings, and residential communities that rely on low-voltage distribution cabinets for stable power supply. Based on the updated contribution weights, a dynamic mapping network reflecting the nonlinear relationship between features and fault modes is constructed. This dynamic mapping network employs a multilayer perceptron structure: the input layer receives high-dimensional feature vectors, the hidden layer performs nonlinear transformations, and the output layer generates fault mode probabilities. The data processed by the dynamic mapping network is a high-dimensional feature vector with clear technical meaning. This vector is formed by concatenating the frequency domain features, transient process features, and load characteristic fluctuation features of the low-voltage distribution cabinet according to a time series alignment. The frequency domain features originate from the Fast Fourier Transform results of a standard time-series signal data stream, while the transient process features are derived from the amplitude change rate and duration of signal waveform abrupt change points. The amplitude change rate is calculated based on the amplitude difference between adjacent sampling points, as shown in the following formula:

[0041] in: Indicates a point in time The signal amplitude, Indicates a point in time The signal amplitude, , These are two adjacent sampling times.

[0042] Load characteristic fluctuation features are the statistical characteristics of the signal within a set time window, such as mean, variance, and skewness, all of which are directly related to the core theme of low-voltage switchgear fault monitoring. High-dimensional feature vectors comprehensively characterize the real-time operating status of the low-voltage switchgear. The dynamic mapping network, through the nonlinear transformation of a multilayer perceptron, associates and matches these state features with historical features in a pre-stored fault case map, ultimately outputting the probability of occurrence of each fault mode. That is, the more significant the abnormal state features reflected in the feature vector, the higher the output probability of the corresponding fault mode, achieving a precise mapping from operating state features to fault probability. In essence, the hidden layer of the dynamic mapping network uses the ReLU activation function to introduce nonlinear processing capabilities, and the output layer uses the softmax function to convert the hidden layer output into a probability distribution form. Optionally, the number of layers and neurons in the dynamic mapping network can be configured according to the feature dimension and fault mode complexity. For example, the number of neurons in the input layer can be equal to the feature vector length, with two hidden layers, each containing 128 neurons, and the number of neurons in the output layer equal to the number of fault mode categories. The dynamic mapping relationship network continuously adjusts the connection weights through iterative training, enabling the network output to accurately map input features to corresponding fault modes. The weight update formula is as follows:

[0043] in: This represents the current weight matrix. This represents the updated weight matrix. Indicates the learning rate. Indicates the loss function at The gradient at that point.

[0044] Example 4: In specific implementation, the dynamic mapping network calculates the high-dimensional feature vector of the current input and outputs a set of probability values ​​corresponding to different fault modes. The dynamic mapping network receives the high-dimensional feature vector read from the multi-dimensional feature benchmark library as input and generates a fault mode probability distribution through a forward propagation process. The high-dimensional feature vector of the current input is input to the input layer of the dynamic mapping network, where a linear transformation and bias addition are performed. The linear transformation multiplies the high-dimensional feature vector with the weight matrix, and the bias addition introduces an offset to adjust the output. The output of the input layer is used as the input of the hidden layer. Nonlinear activation functions are applied through multiple hidden layers to pass the feature representation layer by layer. The hidden layers use the ReLU activation function to introduce a nonlinear transformation, enhancing the network's ability to fit complex patterns. The softmax function is applied to the output layer to convert the hidden layer output into a probability distribution form, outputting a set of probability values ​​corresponding to different fault modes. Each probability value represents the probability that the current input belongs to the corresponding fault mode. It can be understood that the output layer of the dynamic mapping network uses a softmax function adapted to the fault probability distribution characteristics of the distribution cabinet, as shown in the following formula:

[0045] in: Indicates the first The probability value of each failure mode. Indicates the output layer number The input value of each neuron, This indicates the total number of failure modes. Represents an exponential function. The fault mode calibration coefficient (set based on the historical fault occurrence rate of the distribution cabinet, with 1.2 for high-incidence faults and 0.9 for low-incidence faults) is used to improve the sensitivity of probability identification of high-risk faults.

[0046] In some embodiments, the number of neurons in the input layer of the dynamic mapping relationship network is consistent with the dimension of the high-dimensional feature vector, two hidden layers are set, each containing 64 neurons, and the number of neurons in the output layer is the same as the number of predefined fault mode categories, such as overload, short circuit and ground fault.

[0047] In implementation, the most likely fault modes are determined by sorting them according to their probability values. The probability values ​​are sorted in descending order, and the fault modes with the highest probabilities are included in the candidate set. The strategy knowledge base is queried to obtain recommended monitoring parameters that match the most likely fault mode set. These recommended parameters include sampling frequency, a list of feature extraction items, and diagnostic thresholds. The strategy knowledge base stores the mapping relationship between fault modes and monitoring parameters in tabular form. Based on the recommended monitoring parameters, an optimal monitoring strategy is generated to guide the next monitoring cycle. The optimal monitoring strategy is output as a configuration file, specifying the specific parameter settings for signal acquisition and feature extraction. Optionally, referring to Table 1, the strategy knowledge base is constructed using historical operating data and learning algorithms and is updated periodically to reflect system changes.

[0048] Table 1: Strategy Knowledge Base Table Fault mode Recommended sampling frequency (Hz) Feature extraction item list Diagnosis threshold Overload fault 1000 Current mean, temperature variance, spectral energy 0.85 Short circuit fault 5000 Current transient rate of change, voltage sag, high frequency oscillation 0.90 Ground fault 2000 Zero sequence current, insulation resistance, ground current harmonics 0.80 In some embodiments, the probability values ​​are sorted to select the top N fault modes as the set of most likely fault modes. N is set to a fixed value or dynamically adjusted according to the system configuration; for example, N=3 indicates that the three fault modes with the highest probability are retained. It is understood that recommended monitoring parameters are designed based on the characteristics of the fault modes. For example, high-frequency fault modes correspond to high sampling frequencies, transient fault modes correspond to feature extraction items, and diagnostic thresholds are used to filter low-probability events in subsequent decisions. Optionally, the optimal monitoring strategy generation process includes parameter verification steps to ensure that the sampling frequency is within the hardware's allowable range, the feature extraction items are compatible with the multi-dimensional feature generation module, and the diagnostic threshold is set within a reasonable range to avoid false alarms. The generated optimal monitoring strategy is encapsulated into a data structure and transmitted to the signal acquisition and preprocessing module and the fault decision and output module to achieve adaptive adjustment of the monitoring process.

[0049] See Figure 4 The graph presents the probability distribution of different fault modes. Specifically, the dynamic mapping relationship network receives high-dimensional feature vectors from a multi-dimensional feature benchmark library. After forward propagation (linear transformation of the input layer, ReLU nonlinear activation of the hidden layer, and softmax normalization of the output layer), it outputs the probability values ​​of six fault modes, including overload faults and short-circuit faults. The probabilities corresponding to each fault mode in the graph are as follows: abnormal heat dissipation 0.15, insulation aging 0.28, poor contact 0.35, grounding fault 0.62, short-circuit fault 0.76, and overload fault 0.89. A diagnostic threshold (0.80) is also marked as the probability boundary for fault determination. Among them, the probability of overload fault (0.89) has exceeded the threshold, and the probability of short-circuit fault (0.76) is close to the threshold. These two are high-risk fault modes that need to be focused on. The distribution results will be directly used to query the strategy knowledge base and generate optimal monitoring strategies (such as adjusting the sampling frequency and feature extraction priority).

[0050] Example 5: In specific implementation, based on the sampling frequency and feature extraction items set by the optimal monitoring strategy, real-time feature calculation is performed on the subsequent standard time-series signal data stream. The optimal monitoring strategy specifies the sampling frequency value and feature extraction item list in the form of a configuration file. The real-time feature calculation process calls the algorithm in the multi-dimensional feature generation module to extract frequency domain features, transient process features, and load characteristic fluctuation features for current, voltage, temperature, and vibration signals. The sampling frequency is dynamically adjusted according to the fault mode; for example, it is set to 5000 Hz for high-frequency transient faults and 1000 Hz for steady-state faults. The calculated real-time features are input into the dynamic mapping relationship network to obtain the real-time fault mode probability distribution. The dynamic mapping relationship network processes the input feature vector through forward propagation and outputs a set of probability values ​​representing the probability of occurrence of each fault mode. The probability distribution is stored in vector form and used for subsequent decision logic. In some embodiments, the feature extraction item list is parsed from the optimal monitoring strategy, and only the specified feature types are calculated to optimize processing efficiency. The output of the dynamic mapping relationship network is normalized by the softmax function to ensure that the sum of all probability values ​​is 1.

[0051] In practical implementation, fault modes with probabilities exceeding the diagnostic thresholds in the optimal monitoring strategy are selected as candidate faults. The diagnostic thresholds are set independently for different fault modes; for example, the overload fault threshold is 0.85, and the short-circuit fault threshold is 0.90. Probability comparison is achieved through element-by-element judgment, and fault modes exceeding the thresholds are identified as candidate sets. Combining the topology information of the distribution cabinet, candidate faults are located and analyzed, and their severity levels are determined based on their probability values ​​and historical data. The topology information is stored in graph form, describing the connection relationships between components within the distribution cabinet. The location analysis uses a graph traversal algorithm to search for abnormal areas starting from sensor detection points. Historical data is retrieved from the fault case map to find records of similar faults. The severity level calculation formula for the importance of the distribution cabinet components can be understood as follows:

[0052] in: Indicates the severity level (0-1). This represents the probability value of a candidate fault. Represents the normalized historical fault frequency (0-1). This represents the importance coefficient of the fault-related components. , , These represent the weighting coefficients for probability values, historical data, and component importance, respectively, comprehensively considering the likelihood and scope of a failure. The weighting coefficients are set by the system based on the equipment type, service life, historical failure data, and component importance in the low-voltage switchgear's topology, combined with statistical analysis results from pre-stored failure case maps.

[0053] In practical implementation, the final determined fault type, fault location, and fault severity level are integrated into a structured diagnostic report. The fault type is identified by extracting a name identifier from candidate faults, the fault location is described as a string through topology analysis, such as "bus connection point A," and the fault severity level is represented as a floating-point number with two decimal places. Optionally, the diagnostic report is encapsulated in JSON format, containing fixed fields to record the fault description, location coordinates, and severity level. After generation, the diagnostic report is transmitted to the monitoring system via a communication interface. In some embodiments, the structured diagnostic report includes a timestamp, device identifier, and fault details for easy parsing and storage by external systems. Fault location analysis is combined with cross-validation of multi-sensor data to improve location accuracy. Optionally, the diagnostic report undergoes format validation before output to ensure field integrity and reasonable data range; for example, severity level values ​​are between 0 and 1, and the fault location string conforms to predefined naming conventions.

[0054] See Figure 5 The graph simultaneously presents the correlation between sampling frequency (red bars) and monitoring period (cyan line) under different fault modes / operating states. Specifically, the raw data corresponds to different operating scenarios of the low-voltage distribution cabinet (high-frequency transient faults, steady-state faults, etc.). The sampling frequency quantifies the signal acquisition density in Hertz, and the monitoring period reflects the data acquisition interval in seconds. The dynamic correlation between the two is as follows: when the equipment is in a high-frequency transient fault state, the sampling frequency reaches 5000Hz and the monitoring period is shortened to about 2 seconds to achieve high-timeliness data capture; while under normal operating conditions, the sampling frequency drops to about 500Hz and the monitoring period is extended to 40 seconds, balancing monitoring cost and efficiency. During the parameter mapping process, the sampling frequency and monitoring period exhibit a negative correlation and dynamic adaptation. Their numerical configuration is determined based on the characteristic response requirements of the fault mode. For example, high-frequency transient faults correspond to the highest sampling density and the shortest monitoring interval.

[0055] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0056] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A low voltage switchgear failure monitoring system, characterized in that, The system comprises: a signal acquisition and preprocessing module for acquiring multiple original operation signals of a low-voltage power distribution cabinet in real time, and performing denoising and normalization processing on the multiple original operation signals to generate a standard time sequence signal data stream; a multi-dimensional feature generation module for receiving the standard time sequence signal data stream and synchronously performing signal frequency domain feature extraction, transient process feature capture and load characteristic fluctuation analysis, fusing the extracted features to construct a dynamically updated multi-dimensional feature benchmark library; a fault mode dynamic mapping module for establishing and real-time correcting a dynamic mapping relationship network between features and fault modes based on the multi-dimensional feature benchmark library and in combination with pre-stored fault case graphs by iteratively calculating the correlation weight between features and fault modes; a monitoring strategy self-adaptive generation module for automatically adjusting the sampling frequency of signal monitoring, the priority of feature extraction and the logic judgment threshold of fault diagnosis according to the current fault mode probability distribution output by the dynamic mapping relationship network to generate an optimal monitoring strategy; a fault decision and output module for executing the optimal monitoring strategy, performing online analysis on real-time signals, generating a diagnostic report containing fault type, fault location and fault severity level, and outputting the diagnostic report after standardized packaging.

2. The low voltage switchgear failure monitoring system of claim 1, wherein, The implementation of the signal acquisition and preprocessing module comprises: acquiring original operation signals including current, voltage, temperature and vibration signals through multiple types of sensors deployed inside the low-voltage power distribution cabinet; performing sliding average filtering and wavelet threshold denoising processing on the acquired original operation signals in sequence to suppress environmental noise and electromagnetic interference; dividing each denoised signal by its corresponding rated range value to complete amplitude normalization, and unifying the sampling time stamp to generate the standard time sequence signal data stream.

3. The low voltage switchgear failure monitoring system of claim 2, wherein, The implementation of the multi-dimensional feature generation module comprises: performing fast Fourier transform on the input standard time sequence signal data stream to calculate its frequency spectrum energy distribution and extract the energy proportion of a specific frequency band as a frequency domain feature; capturing waveform mutation points in the standard time sequence signal data stream in a window sliding manner, calculating the amplitude change rate and duration of the mutation points as transient process features; analyzing the statistical characteristics of the standard time sequence signal data stream within a set time window, including mean, variance and skewness, as load characteristic fluctuation features; aligning and splicing the frequency domain features, transient process features and load characteristic fluctuation features in time sequence to form a high-dimensional feature vector, and storing it in the multi-dimensional feature benchmark library.

4. The low voltage switchgear failure monitoring system of claim 3, wherein, The implementation of the fault mode dynamic mapping module comprises: reading the latest high-dimensional feature vector from the multi-dimensional feature benchmark library; performing similarity matching between the high-dimensional feature vector and historical feature vectors in a pre-stored fault case graph, which contains a feature vector set corresponding to a known fault type; based on the matching result, iteratively updating the contribution weight of each feature dimension to different fault modes using the gradient descent method; according to the updated contribution weight, constructing a dynamic mapping relationship network reflecting the nonlinear relationship between features and fault modes.

5. The low voltage switchgear failure monitoring system of claim 4, wherein, The implementation mode of the monitoring strategy self-adaptive generation module comprises: The dynamic mapping relationship network is used to calculate a high-dimensional feature vector input at present, and a group of probability values corresponding to different fault modes is output; According to the size of the probability values, a set of fault modes most likely to occur at present is determined; A strategy knowledge base is queried to obtain recommended monitoring parameters matched with the set of fault modes most likely to occur, the recommended monitoring parameters comprising a sampling frequency, a feature extraction item list and a diagnosis threshold; Based on the recommended monitoring parameters, an optimal monitoring strategy for guiding a next monitoring period is generated.

6. The low voltage switchgear failure monitoring system of claim 5, wherein, The implementation mode of the fault decision and output module comprises: According to the sampling frequency and the feature extraction item set in the optimal monitoring strategy, real-time feature calculation is performed on a subsequent standard time sequence signal data stream; The real-time features calculated are input into the dynamic mapping relationship network to obtain a real-time fault mode probability distribution; A fault mode with a probability exceeding the diagnosis threshold in the optimal monitoring strategy is selected as a candidate fault; In combination with topological structure information of the power distribution cabinet, the candidate fault is positioned and analyzed, and a severity level is determined according to the probability value and historical data of the candidate fault; The final determined fault type, fault position and fault severity level are integrated into a structured diagnosis report.

7. The low voltage switchgear failure monitoring system of claim 3, wherein, The fast Fourier transform of the input standard time sequence signal data stream comprises the following steps: The window length and overlap ratio of the fast Fourier transform are set, and the standard time sequence signal data stream is divided into a plurality of time windows; A Hanning window function is applied to each time window to perform windowing processing on the signal data in the time window, so as to reduce spectral leakage; A fast Fourier transform algorithm is executed to convert the time domain signal into a frequency domain signal, and complex frequency spectrum data are obtained; The modulus of the frequency spectrum data is calculated to obtain a power spectral density distribution; The power spectral density values in a predefined frequency band range are extracted as a frequency domain feature vector.

8. The low voltage switchgear failure monitoring system of claim 4, wherein, The gradient descent method is used to iteratively update the contribution weight of each feature dimension to different fault modes, comprising the following steps: A weight matrix of each feature dimension corresponding to different fault modes is initialized, and a learning rate and a maximum number of iterations are set; A loss function is constructed, the loss function being based on the difference between the predicted fault mode and the real fault mode under the current weight matrix; The gradient of the loss function with respect to the weight matrix is calculated to obtain a gradient direction; The weight matrix is updated in the direction opposite to the gradient according to the learning rate, and the loss function value is reduced; The gradient calculation and weight update steps are repeated until the maximum number of iterations is reached or the loss function converges within a threshold.

9. The low voltage switchgear failure monitoring system of claim 5, wherein, The calculation of the dynamic mapping relationship network on the high-dimensional feature vector input at present comprises the following steps: The high-dimensional feature vector input at present is input into an input layer of the dynamic mapping relationship network for linear transformation and bias addition; Nonlinear activation function processing is performed through a plurality of hidden layers, and feature representation is transmitted layer by layer; A softmax function is applied to an output layer to convert the hidden layer output into a probability distribution form; A group of probability values corresponding to different fault modes is output, each probability value representing the possibility of the input belonging to the corresponding fault mode.

10. A method of low voltage switchgear failure monitoring, characterized in that, The low-voltage power distribution cabinet fault monitoring system comprises all the modules and method processes in any one of claims 1 to 9.

Images