A fault detection method for low-voltage intelligent integrated distribution box

Abstract

The present application relates to the field of power distribution box fault detection, especially to a kind of fault detection method for low-voltage intelligent integrated power distribution box, the method comprises: obtaining the temperature data collected by the main circuit temperature sensor of low-voltage intelligent integrated power distribution box, to generate multi-scale short-term temperature fluctuation characteristics by temperature fluctuation characteristics extraction model based on TCN architecture;Multi-scale short-term temperature fluctuation characteristics are through temperature trend feature extraction model based on LSTM and time attention mechanism architecture, to generate temperature change trend feature;Current data and temperature change trend feature are through auto-encoder, to determine the hidden overheat overload fault warning type of low-voltage intelligent integrated power distribution box.The present application realizes the precise early warning of the overheat overload fault of its main circuit by the multi-scale temperature feature extraction of low-voltage intelligent power distribution box and the coupling modeling of temperature and current.

Description

Technical Field

[0001] This invention relates to the field of distribution box fault detection, and more particularly to a fault detection method for low-voltage intelligent integrated distribution boxes. Background Technology

[0002] Low-voltage intelligent integrated distribution boxes are complete sets of power distribution equipment at the end of low-voltage distribution networks with a rated voltage of 0.4kV. They are widely used in urban and rural power distribution areas, industrial and civil buildings, and new energy supporting power distribution. They undertake the functions of power distribution, metering and control, relay protection, and reactive power compensation. Their operational reliability directly determines the power supply continuity and power safety of the low-voltage power supply system.

[0003] Among various faults in low-voltage intelligent integrated distribution boxes, overheating and overload faults are the most frequent and pose the greatest hidden danger. In particular, latent overheating and overload faults, caused by factors such as loose terminal connections, poor contact at busbar joints, worn circuit breaker contacts, long-term minor overloads, and harmonic-induced heating, are characterized by slow development, weak characteristics, and no obvious early alarm signals. Traditional protection devices cannot effectively identify them, and they are prone to gradual deterioration until insulation breakdown, equipment burnout, or even electrical fires. According to relevant statistics, fires caused by poor contact, overload, and overheating in low-voltage power distribution systems account for more than 60% of all electrical fires. Therefore, achieving early and high-precision detection and warning of latent overheating and overload faults in low-voltage intelligent integrated distribution boxes is a core requirement for the safety protection of low-voltage power distribution systems.

[0004] Currently, detection technologies for overheating and overload faults in low-voltage distribution boxes are mainly divided into two categories: traditional electrical protection methods and data-driven intelligent diagnostic methods.

[0005] Traditional electrical protection methods are the mainstream application in current low-voltage distribution boxes. They mainly rely on the thermal trip unit of circuit breakers for overload protection, residual current devices for leakage protection, and fixed temperature thresholds for overheat alarms, along with simple overcurrent, overvoltage, and undervoltage relay protection logic. However, the warning judgment of such methods depends on fixed threshold triggers, and can only trip or alarm when the fault develops to a severe stage or when the electrical quantity or temperature exceeds the threshold. They cannot capture the weak characteristics of the fault in its infancy, lack early warning capabilities, and cannot perform fine-grained identification and graded warning of fault type and degradation degree, thus failing to support predictive maintenance of the power distribution system.

[0006] With the improvement of the sensing capabilities of low-voltage intelligent distribution boxes, real-time acquisition of multi-dimensional operating data such as temperature and current has become standard, supporting the development of data-driven intelligent fault diagnosis methods. In existing technologies, for example, the invention patent with announcement number CN119720048B discloses a fault diagnosis method and system for low-voltage switchgear. It uses a convolutional neural network (CNN) to extract current and voltage waveform features to identify short-circuit and arc faults, and uses a long short-term memory network (LSTM) to fit temperature time series data to achieve overheat warning.

[0007] However, existing technologies still suffer from insufficient feature extraction in detecting latent overheating and overload faults in low-voltage distribution boxes, failing to consider both the multi-scale characteristics of temperature signals and the electrothermal coupling nature of overheating faults. The core characteristic of overheating and overload faults is temperature anomaly, while the temperature characteristics of latent faults simultaneously include short-term fluctuations at the minute or even second level caused by poor contact, and long-term degradation trends at the hour or day level. Existing methods cannot simultaneously achieve refined extraction of multi-scale short-term fluctuation features and long-term trend features.

[0008] Furthermore, existing model architectures have poor adaptability and insufficient attention to subtle abnormal features. Existing multi-scale feature extraction methods mostly use multi-window moving averages or fixed convolution kernels, which cannot achieve adaptive extraction of temperature fluctuation features at different time scales and are prone to losing key fault information.

[0009] In summary, how to achieve accurate early warning of overheating and overload faults in the main circuit of low-voltage intelligent distribution boxes by extracting multi-scale temperature features and modeling temperature-current coupling is a technical problem that needs to be solved. Summary of the Invention

[0010] Therefore, the present invention provides a fault detection method for low-voltage intelligent integrated distribution boxes, which overcomes the problem in the prior art that it is impossible to achieve accurate early warning of overheating and overload faults in the main circuit of low-voltage intelligent distribution boxes through multi-scale temperature feature extraction and temperature-current coupling modeling.

[0011] To achieve the above objectives, this invention proposes a fault detection method for low-voltage intelligent integrated distribution boxes, comprising: Temperature data collected by the main circuit temperature sensor of the low-voltage intelligent integrated distribution box is obtained, and the temperature data is processed by a temperature fluctuation feature extraction model based on TCN architecture to generate multi-scale short-time temperature fluctuation features. The multi-scale short-term temperature fluctuation features are used to generate temperature change trend features through a temperature trend feature extraction model based on LSTM and time attention mechanism architecture. The current data collected by the main circuit current sensor of the low-voltage intelligent integrated distribution box is obtained, and the current data and the temperature change trend characteristics are used through an encoder to determine the latent overheating and overload fault warning type of the low-voltage intelligent integrated distribution box.

[0012] Furthermore, the process of generating multi-scale short-time temperature fluctuation features through the temperature fluctuation feature extraction model includes: The temperature data is passed through a first dilated convolution branch to generate temperature pulse features; The temperature data is passed through a second dilated convolution branch to generate heat dissipation features; The temperature data is passed through an upsampling branch to generate temperature trend features; The concatenated vector of the temperature pulse feature, heat dissipation feature, and temperature trend feature is processed through an output convolution operation to generate the multi-scale short-term temperature fluctuation feature. The temperature fluctuation feature extraction model includes a first dilated convolution branch, a second dilated convolution branch, an upsampling branch, and an output convolution operation.

[0013] Furthermore, the process of generating temperature pulse features through the first dilated convolution branch includes: The temperature data is passed through a first causal convolutional layer to generate initial temperature transformation features; The initial temperature transformation features are passed through a first dilated causal convolution residual block to generate temperature spike features; The temperature spike feature is passed through a second dilated causal convolution residual block to generate the temperature pulse feature; The first dilated convolution branch includes a first causal convolution layer, a first dilated causal convolution residual block, and a second dilated causal convolution residual block. The dilation rate of the first dilated causal convolution residual block is less than the dilation rate of the second dilated causal convolution residual block.

[0014] Furthermore, the process of generating heat dissipation features through the second dilated convolution branch includes: The temperature data is passed through a second causal convolutional layer to generate temperature rise segment features; The heat dissipation feature is generated by passing the temperature rise segment feature through a third dilated causal convolution residual block; The second dilated convolution branch includes a second causal convolution layer and a third dilated causal convolution residual block. The kernel size of the second causal convolution layer is larger than the kernel size of the first causal convolution layer, and the dilation rate of the third dilated causal convolution residual block is greater than the dilation rate of the second dilated causal convolution residual block.

[0015] Furthermore, the process of generating temperature trend features through upsampling branches includes: The temperature data is passed through an average pooling layer to generate initial temperature trend features; The initial temperature trend features are passed through a convolutional layer to generate the temperature trend features; The upsampling branch includes an average pooling layer and a convolutional layer.

[0016] Furthermore, the process of generating temperature change trend features through the temperature trend feature extraction model includes: The multi-scale short-term temperature fluctuation features are passed through a bidirectional LSTM layer to generate temperature development trajectory features; The temperature development trajectory features are weighted at their time steps using a time attention mechanism to generate the temperature change trend features; The temperature trend feature extraction model includes a bidirectional LSTM layer and a time attention mechanism.

[0017] Furthermore, the process of determining the type of latent overheating and overload fault warning through a self-encoder includes: The current data is passed through a current sub-encoder to generate current timing features; The concatenated vectors of the current timing features and the temperature change trend features are fused through a fully connected layer to generate temperature-current fused features. The temperature-current fusion features are passed through a temperature decoder to generate reconstructed temperature features; The type of latent overheating and overload fault warning is determined based on the reconstructed temperature characteristics and the reconstruction error of the temperature data. The self-encoder includes a current sub-encoder, a fused fully connected layer, and a temperature decoder.

[0018] Furthermore, the process of generating current timing features through a current sub-encoder includes: The current data is passed through a global average pooling layer to generate load level characteristics; The load level characteristics are encoded through a fully connected layer to generate the current timing characteristics; The current sub-encoder includes a global average pooling layer and an encoding fully connected layer.

[0019] Furthermore, the process of determining the latent overheating and overload fault warning type based on the reconstructed temperature characteristics and the error of the temperature data includes: The reconstructed temperature features and temperature data are processed using a mean square error function to obtain the reconstruction error. The warning threshold is calculated based on the quantiles of the reconstruction error from the training set. The type of latent overheating and overload fault warning is determined based on whether the reconstruction error is greater than the warning threshold and whether the current data is in the danger range.

[0020] Furthermore, fault detection methods also include: The temperature-current fusion features of the training set are used to calculate the KL divergence loss term. The reconstruction error and KL divergence loss term of the training set are weighted and summed to calculate the loss function; The temperature fluctuation feature extraction model, temperature trend feature extraction model, and autoencoder are trained collaboratively based on the loss function.

[0021] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention adapts to short-term weak temperature rise caused by poor contact of busbars and joints in low-voltage intelligent integrated distribution boxes, and temperature peaks caused by overload impacts through the first dilated convolution branch; it focuses on capturing temperature rise segments and heat dissipation attenuation characteristics on a scale of several minutes through the second dilated convolution branch, accurately characterizing the slow temperature rise caused by hidden overloads and the temperature drift characteristics caused by abnormal heat dissipation circuits; and it extracts the global baseline trend features of the temperature sequence through the upsampling branch, effectively filtering invalid fluctuations caused by on-site electromagnetic interference and sensor random noise, providing a global benchmark for multi-scale features, thereby realizing the accurate early warning of overheating and overload faults in the main circuit of low-voltage intelligent distribution boxes through multi-scale temperature feature extraction.

[0022] In particular, this invention uses the global average pooling layer of the current sub-encoder to accurately extract the load level and load fluctuation baseline features directly related to heat generation, filtering out invalid noise caused by high-frequency harmonics of the current and electromagnetic interference on site. By fusing the fully connected layer to learn the strong coupling physical mapping law of load level and temperature response under normal operating conditions through deep learning of current timing features and temperature change trend features, the temperature decoder completes the directional reconstruction of temperature features based on the fused features. Under normal operating conditions, the coupling relationship between current and temperature is stable and the reconstruction error is minimal. When a latent fault occurs, the coupling relationship is broken, and the reconstruction error will be amplified sharply, which improves the specificity of the reconstruction error for overheating and overload faults. Thus, it realizes the coupled modeling of temperature and current in the low-voltage intelligent distribution box to achieve accurate early warning of overheating and overload faults in its main circuit. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating a fault detection method for a low-voltage intelligent integrated distribution box according to an embodiment of the present invention. Figure 2 This is a flowchart illustrating the temperature fluctuation feature extraction model of the fault detection method for low-voltage intelligent integrated distribution boxes according to an embodiment of the present invention. Figure 3This is a schematic diagram of the forward propagation process of the temperature trend feature extraction model for the fault detection method of a low-voltage intelligent integrated distribution box according to an embodiment of the present invention. Figure 4 This is a flowchart illustrating the self-encoder in the fault detection method for a low-voltage intelligent integrated distribution box according to an embodiment of the present invention. Detailed Implementation

[0024] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0025] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0026] It should be noted that in the description of this invention, the terms "upper", "lower", "left", "right", "inner", "outer", etc., which indicate directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings. This is only for the convenience of description and is not intended to indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.

[0027] Furthermore, it should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0028] like Figures 1 to 4 As shown, the present invention provides a fault detection method for low-voltage intelligent integrated distribution boxes, which overcomes the problem in the prior art that it is impossible to achieve accurate early warning of overheating and overload faults in the main circuit of low-voltage intelligent distribution boxes through multi-scale temperature feature extraction and temperature-current coupling modeling.

[0029] like Figure 1 As shown in the figure, this embodiment proposes a fault detection method for low-voltage intelligent integrated distribution boxes, including: Temperature data collected by the main circuit temperature sensor of the low-voltage intelligent integrated distribution box is obtained, and the temperature data is processed by a temperature fluctuation feature extraction model based on TCN architecture to generate multi-scale short-time temperature fluctuation features. The multi-scale short-term temperature fluctuation features are used to generate temperature change trend features through a temperature trend feature extraction model based on LSTM and time attention mechanism architecture. The current data collected by the main circuit current sensor of the low-voltage intelligent integrated distribution box is obtained, and the current data and the temperature change trend characteristics are used through an encoder to determine the latent overheating and overload fault warning type of the low-voltage intelligent integrated distribution box.

[0030] Specifically, the low-voltage intelligent integrated distribution box is a complete set of low-voltage switchgear for 50Hz low-voltage distribution networks with a rated voltage of AC 0.4kV (380V / 220V). It integrates power distribution, metering and control, relay protection, reactive power compensation, intelligent operation and maintenance, edge computing, and fault diagnosis. Temperature sensors are installed at the main busbar connection point, the connection between the main busbar and the incoming circuit breaker, the incoming main circuit breaker, the contact points of moving and stationary contacts, the incoming and outgoing line terminals, and the fuse base contacts. Three-phase current sensors are also installed on the main circuit incoming line and main feeder circuits of the low-voltage intelligent integrated distribution box at the factory.

[0031] Therefore, the temperature data collected by the temperature sensor may include abnormal temperature rise data caused by poor contact, oxidation, loosening, or overload in the low-voltage intelligent integrated distribution box. The current data collected by the current sensor includes real-time load current signals. By fusing the current data and the temperature change trend characteristics through an encoder, the current data can be used as conditional information to assist in reconstructing the normal temperature pattern, thereby accurately identifying hidden overheating and overload faults and reducing false alarms caused by load changes.

[0032] like Figure 2 As shown, the process of generating multi-scale short-time temperature fluctuation features through the temperature fluctuation feature extraction model further includes: The temperature data is passed through a first dilated convolution branch to generate temperature pulse features; The temperature data is passed through a second dilated convolution branch to generate heat dissipation features; The temperature data is passed through an upsampling branch to generate temperature trend features; The concatenated vector of the temperature pulse feature, heat dissipation feature, and temperature trend feature is processed through an output convolution operation to generate the multi-scale short-term temperature fluctuation feature. The temperature fluctuation feature extraction model includes a first dilated convolution branch, a second dilated convolution branch, an upsampling branch, and an output convolution operation.

[0033] Specifically, the temperature data is a time-series sequence vector with a sampling rate of 2 times / minute and a time window of 60 minutes.

[0034] Specifically, the output convolution operation uses a 1-size convolution kernel with a stride of 1. It does not span time steps and only performs a linear transformation on the channel vector at each time step. It also uses the GELU activation function to perform a fully connected linear transformation on the multi-scale feature vectors only once, without changing the time axis length or temporal dependencies. The LSTM can still model based on the complete time series.

[0035] Furthermore, the process of generating temperature pulse features through the first dilated convolution branch includes: The temperature data is passed through a first causal convolutional layer to generate initial temperature transformation features; The initial temperature transformation features are passed through a first dilated causal convolution residual block to generate temperature spike features; The temperature spike feature is passed through a second dilated causal convolution residual block to generate the temperature pulse feature; The first dilated convolution branch includes a first causal convolution layer, a first dilated causal convolution residual block, and a second dilated causal convolution residual block. The dilation rate of the first dilated causal convolution residual block is less than the dilation rate of the second dilated causal convolution residual block.

[0036] Specifically, the first causal convolutional layer uses a 3-size convolutional kernel with a stride of 1 and a ReLU activation function. Therefore, the first causal convolutional layer slides on the time axis to extract local waveform patterns. The 3-size convolutional kernel extracts local waveform patterns, that is, joint changes within 1.5 minutes. Its output initial temperature change features can initially extract the most basic time change features such as temperature difference and inflection point between adjacent time moments, providing detailed input for subsequent residual blocks.

[0037] Specifically, the structures of the first dilated causal convolution residual block and the second dilated causal convolution residual block are, in sequence, dilated causal convolution, weight normalization (WN), GELU activation function, dropout layer and residual connection layer.

[0038] The first dilated causal convolution residual block has a dilation rate of 1 and a receptive field of 3. Its function is to further refine local temperature fluctuation patterns, such as small upward bulges or downward edges. The second causal convolution layer has a dilation rate of 5 and a receptive field of 5. The total effective receptive field of 5 can reach about 3.5 minutes, which means that each output feature of this layer integrates temperature evolution information over several minutes, but still retains the fine structure on the time axis.

[0039] Weight normalization reparameterizes the weight vectors of the convolutional kernels to promote more consistent gradient flow across different time steps and feature dimensions, improving the model's generalization ability to temperature fluctuation patterns. The GELU activation function enables the network to learn complex temperature time-series patterns, attenuating small noise inputs while amplifying larger anomalous fluctuations, making the model more sensitive to weak early signals of latent overheating. The Dropout layer uses SpatialDropout with a dropout rate of 0.1, randomly zeroing out the entire feature channel across all time steps to prevent the model from over-relying on features from a specific time step or a single temperature measurement node, reducing overfitting. The residual connection layer applies a 1-dimensional convolution to the original input features of the first and second dilated causal convolutional residual blocks for dimension matching, then adds it to the features output by the Dropout layer to obtain the output features of the first and second dilated causal convolutional residual blocks, i.e., temperature spike features or temperature pulse features. The residual connection layer allows gradient backflow, protecting detailed information from being buried by deep networks.

[0040] Therefore, the first dilated causal convolutional residual block specifically amplifies extremely short-duration temperature spikes, namely the intermittent instantaneous rise of 0.5 to 1 degree Celsius caused by loose connections. Even under normal load fluctuations, this anomalous abrupt change can be reinforced by residual learning. The second dilated causal convolutional residual block, while preserving local details, captures the persistence and transition of short-term anomalous patterns. That is, the trend of the temperature rise rate changing from 0.1°C / min to 0.3°C / min over several minutes in the early stages of latent overload is perceived by the dilated convolution.

[0041] Furthermore, the process of generating heat dissipation features through the second dilated convolution branch includes: The temperature data is passed through a second causal convolutional layer to generate temperature rise segment features; The heat dissipation feature is generated by passing the temperature rise segment feature through a third dilated causal convolution residual block; The second dilated convolution branch includes a second causal convolution layer and a third dilated causal convolution residual block. The kernel size of the second causal convolution layer is larger than the kernel size of the first causal convolution layer, and the dilation rate of the third dilated causal convolution residual block is greater than the dilation rate of the second dilated causal convolution residual block.

[0042] Specifically, the second causal convolutional layer uses a 5-size convolutional kernel with a stride of 1 and the GELU activation function. Therefore, the 5-size convolutional kernel of the second causal convolutional layer has a wider receptive field than the 3-size convolutional kernel of the first causal convolutional layer, and can directly sense more continuous temperature change arcs, such as a complete small waveform from the starting point of heating to the local extreme value.

[0043] Specifically, the structure of the third dilated causal convolutional residual block is the same as that of the first and second dilated causal convolutional residual blocks, consisting of dilated causal convolution, weight normalization, GELU activation function, Dropout layer, and residual connection layer. The dilated causal convolution of the third dilated causal convolutional residual block has a dilation rate of 5 and a receptive field of 4. Therefore, the features at each time step in the temperature rise segment feature sequence are calculated based on the temperature information of the current time step and the previous 16 time steps, equivalent to reviewing approximately 8 minutes of history. Combined with the 5-size convolutional kernel of the preceding second causal convolutional layer, the effective receptive field of the entire second dilated convolutional branch can reach approximately 21 time steps, covering a medium-length temperature evolution process. The Dropout layer of the third dilated causal convolutional residual block uses Spatial Dropout with a dropout rate of 0.1. The residual connection layer of the third dilated causal convolutional residual block performs dimension matching by applying a 1-size convolution to the original input features of the temperature rise segment features, and then adds it to the features output by the Dropout layer to obtain the heat dissipation features.

[0044] Therefore, the heat dissipation features generated by the second dilated convolution branch effectively capture implicit overheating modes that are neither instantaneous spikes nor ultra-long-period drifts, such as slowly developing poor contact and stable temperature rise acceleration caused by overload, providing the model with feature representations that are highly complementary to the short-term and macro-term branches.

[0045] Furthermore, the process of generating temperature trend features through upsampling branches includes: The temperature data is passed through an average pooling layer to generate initial temperature trend features; The initial temperature trend features are passed through a convolutional layer to generate the temperature trend features; The upsampling branch includes an average pooling layer and a convolutional layer.

[0046] Specifically, the average pooling layer has a pooling window size of 2 and a step size of 2. Therefore, it averages the feature values ​​of two consecutive time steps and merges them into one value, which is equivalent to performing low-pass filtering and downsampling on the original sequence. This can eliminate high-frequency fluctuations and random noise of the temperature sensor, and only retain the more macroscopic temperature change trend.

[0047] Specifically, the convolutional layer uses a 3-size kernel with a stride of 1. Since the time step has been halved, the effective receptive field of the kernel is doubled on the original time axis, thus enabling it to characterize the local shape of macro trends, such as the slope, plateau, and inflection point of long-term temperature rise. Therefore, the multi-scale short-term temperature fluctuation features, which include a combination of short-term spikes and moderate trend increases, are better able to characterize the onset of latent overheating, and the fused features have stronger discriminative power.

[0048] like Figure 3 As shown, the process of generating temperature change trend features through the temperature trend feature extraction model further includes: The multi-scale short-term temperature fluctuation features are passed through a bidirectional LSTM layer to generate temperature development trajectory features; The temperature development trajectory features are weighted at their time steps using a time attention mechanism to generate the temperature change trend features; The temperature trend feature extraction model includes a bidirectional LSTM layer and a time attention mechanism.

[0049] Specifically, the Bidirectional LSTM layer (BiLSTM) includes a forward LSTM layer and a backward LSTM layer, both based on the standard BiLSTM architecture. The forward LSTM layer captures the evolution of temperature fluctuation features from the past to the future, while the backward LSTM layer captures the dependencies from the future to the past. The combination of the two can provide a complete understanding of the temperature change background before and after each time point. Therefore, the Bidirectional LSTM layer can integrate the local fluctuation features extracted by TCN into long-range time-series dependencies, identify the continuous trend of temperature rise or fall, the turning points of acceleration or deceleration, and the overall trajectory of abnormal temperature rise.

[0050] Specifically, the fully connected layer of the time attention mechanism scores the time steps of the temperature trajectory features output by the bidirectional LSTM layer. After softmax normalization, this score is used as the importance weight for each time step. The importance weight is then multiplied element-wise with the temperature trajectory features to obtain the temperature change trend features. Therefore, these temperature change trend features automatically focus on the critical time periods when temperature changes are most drastic or when they are most indicative of faults, such as the period when the temperature suddenly begins to rise rapidly, or the turning point from normal to abnormal.

[0051] like Figure 4 As shown, the process of determining the type of latent overheating and overload fault warning via a self-encoder further includes: The current data is passed through a current sub-encoder to generate current timing features; The concatenated vectors of the current timing features and the temperature change trend features are fused through a fully connected layer to generate temperature-current fused features. The temperature-current fusion features are passed through a temperature decoder to generate reconstructed temperature features; The type of latent overheating and overload fault warning is determined based on the reconstructed temperature characteristics and the reconstruction error of the temperature data. The self-encoder includes a current sub-encoder, a fused fully connected layer, and a temperature decoder.

[0052] Specifically, the current data is a time sequence vector with a sampling rate of 1 time / minute and a time window of 60 minutes.

[0053] Specifically, the fully connected layer is a fully connected layer architecture that uses the ReLU activation function to achieve deep integration of temperature trends and current conditions in the latent space, enabling the temperature decoder to know the current load conditions for conditional temperature reconstruction.

[0054] Specifically, the temperature decoder includes a sequence length recovery layer, an LSTM layer, and an output projection layer. The sequence length recovery layer is used to replicate the temperature-current fusion features along the time dimension, providing the initial input for each time step of the LSTM. The LSTM layer forms an autoregressive decoder. The output projection layer is a fully connected layer architecture that independently maps the features of each time step back to the temperature values ​​of multiple temperature nodes, forming a reconstructed temperature feature with the same number of temperature values ​​as the temperature data.

[0055] Therefore, this autoencoder detects deviations by reconstructing normal temperature patterns under current conditions, exhibiting strong anti-interference capabilities and fault type localization capabilities. After training, the autoencoder only demonstrates excellent reconstruction capabilities under normal operating conditions. Given a specific current condition, the decoder can highly reproduce the normal temperature sequence that each temperature measurement node should exhibit under that condition.

[0056] Furthermore, the process of generating current timing features through a current sub-encoder includes: The current data is passed through a global average pooling layer to generate load level characteristics; The load level characteristics are encoded through a fully connected layer to generate the current timing characteristics; The current sub-encoder includes a global average pooling layer and an encoding fully connected layer.

[0057] Specifically, the Global Average Pooling layer averages each current value along the time dimension of the current data, thereby capturing the average load level, average harmonics, and imbalance within the entire window, while discarding specific temporal variation details. The encoding fully connected layer is a fully connected layer architecture using the ReLU activation function, which plays a role in feature extraction and dimensionality reduction, preparing for subsequent fusion.

[0058] Furthermore, the process of determining the latent overheating and overload fault warning type based on the reconstructed temperature characteristics and the error of the temperature data includes: The reconstructed temperature features and temperature data are processed using a mean square error function to obtain the reconstruction error. The warning threshold is calculated based on the quantiles of the reconstruction error from the training set. The type of latent overheating and overload fault warning is determined based on whether the reconstruction error is greater than the warning threshold and whether the current data is in the danger range.

[0059] Specifically, the warning threshold is the 99.7th percentile of the reconstruction error in the training set, corresponding to the 3σ criterion. If the reconstruction error exceeds the warning threshold for four consecutive time windows, a latent overheating / overload fault warning is issued. Specifically, if the reconstruction error of a single node is significantly greater than that of other nodes and the current data is not overloaded, a fault warning for poor joint contact and oxidation loosening is issued. If the reconstruction errors of multiple nodes increase synchronously and the accompanying current data is within the rated danger range, a latent line overload fault warning is issued. If the neutral line node reconstruction error is large and the three-phase current imbalance is within the unbalanced danger range, a three-phase unbalanced overheating fault warning is issued. The rated danger range is the ratio of the maximum effective value of any phase current to the rated current of the distribution box circuit, and the unbalanced danger range is 15% of the three-phase current imbalance.

[0060] Furthermore, fault detection methods for low-voltage intelligent integrated distribution boxes also include: The temperature-current fusion features of the training set are used to calculate the KL divergence loss term. The reconstruction error and KL divergence loss term of the training set are weighted and summed to calculate the loss function; The temperature fluctuation feature extraction model, temperature trend feature extraction model, and autoencoder are trained collaboratively based on the loss function.

[0061] Specifically, the training set is a time window dataset of the normal operating conditions of the low-voltage intelligent integrated distribution box, confirming that there are no hidden faults. It consists of 30 consecutive days of temperature and current data to cover a complete load cycle.

[0062] Specifically, the mean and standard deviation of the temperature-current fusion characteristics are calculated using the standard KL divergence formula to obtain the KL divergence loss term. This KL divergence loss term is then multiplied by a weighting coefficient of 0.1 and added to the reconstruction error to obtain the loss function. Therefore, using the mean squared error function as the main loss term for calculating the difference between the predicted temperature and the actual future temperature forces the model not only to reconstruct historical windows but also to understand the dynamic laws of temperature change, making it more sensitive to overheating trends.

[0063] Specifically, a PT100 platinum resistance temperature sensor and a Hall current sensor are configured in the main circuit of a general-purpose low-voltage intelligent integrated distribution box to construct normal operating condition datasets and fault operating condition datasets. After wavelet soft thresholding denoising and Min-Max normalization processing, these datasets are input into the following models: Control group 1 uses a basic LSTM model, Control group 2 uses a TCN and LSTM model, Control group 3 uses a TCN and temporal attention model, and Control group 4 uses a TCN, LSTM, temporal attention, and fully connected classification model. The experimental group uses the temperature fluctuation feature extraction model, temperature trend feature extraction model, and autoencoder described in this embodiment. Control groups 1, 2, 3, 4, and the experimental group are all trained using the training set described above. Their accuracy rates on the fault operating condition dataset are 82.37%, 91.56%, 89.74%, 95.29%, and 99.12%, respectively, and their latent fault identification rates on the fault operating condition dataset are 68.34%, 85.62%, 82.17%, 92.58%, and 98.36%, respectively. It is evident that the temperature fluctuation feature extraction model, temperature trend feature extraction model, and autoencoder described in this embodiment have strong extraction capabilities for multi-scale temperature fluctuation features, as well as the ability to focus on key changes in temperature trends and to uncover the implicit features of current and temperature coupling.

[0064] In this embodiment, the first dilated convolution branch is used to adapt to short-term weak temperature rises caused by poor contact between the busbar and connectors of the low-voltage intelligent integrated distribution box, as well as temperature spikes caused by overload impacts. The second dilated convolution branch focuses on capturing temperature rise segments on the order of minutes and heat dissipation attenuation characteristics, accurately characterizing the slow temperature rise caused by hidden overloads and the temperature drift characteristics caused by abnormal heat dissipation circuits. The upsampling branch extracts the global baseline trend features of the temperature sequence, effectively filtering out invalid fluctuations caused by on-site electromagnetic interference and sensor random noise, providing a global benchmark for multi-scale features, and thus realizing the accurate early warning of overheating and overload faults in the main circuit of the low-voltage intelligent distribution box through multi-scale temperature feature extraction. By using the global average pooling layer of the current sub-encoder, the load level and load fluctuation baseline features directly related to heat generation are accurately extracted, filtering out invalid noise caused by high-frequency harmonics of the current and electromagnetic interference on site. By fusing the fully connected layer to learn the strong coupling physical mapping law of load level and temperature response under normal operating conditions through deep learning of current time sequence features and temperature change trend features, the temperature decoder completes the targeted reconstruction of temperature features based on the fused features. Under normal operating conditions, the coupling relationship between current and temperature is stable and the reconstruction error is minimal. When a latent fault occurs, the coupling relationship is broken, and the reconstruction error will be amplified sharply, which improves the specificity of the reconstruction error for overheating and overload faults. This enables the accurate early warning of overheating and overload faults in the main circuit of the low-voltage intelligent distribution box by modeling the coupling of temperature and current.

[0065] Those skilled in the art will recognize that the modules and algorithm steps of the examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0066] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0067] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A fault detection method for low-voltage intelligent integrated distribution boxes, characterized in that, include: Temperature data collected by the main circuit temperature sensor of the low-voltage intelligent integrated distribution box is obtained, and the temperature data is processed by a temperature fluctuation feature extraction model based on TCN architecture to generate multi-scale short-time temperature fluctuation features. The multi-scale short-term temperature fluctuation features are used to generate temperature change trend features through a temperature trend feature extraction model based on LSTM and time attention mechanism architecture. The current data collected by the main circuit current sensor of the low-voltage intelligent integrated distribution box is obtained, and the current data and the temperature change trend characteristics are used through an encoder to determine the latent overheating and overload fault warning type of the low-voltage intelligent integrated distribution box.

2. The fault detection method for low-voltage intelligent integrated distribution boxes according to claim 1, characterized in that, The process of generating multi-scale short-time temperature fluctuation features through a temperature fluctuation feature extraction model includes: The temperature data is passed through a first dilated convolution branch to generate temperature pulse features; The temperature data is passed through a second dilated convolution branch to generate heat dissipation features; The temperature data is passed through an upsampling branch to generate temperature trend features; The concatenated vector of the temperature pulse feature, heat dissipation feature, and temperature trend feature is processed through an output convolution operation to generate the multi-scale short-term temperature fluctuation feature. The temperature fluctuation feature extraction model includes a first dilated convolution branch, a second dilated convolution branch, an upsampling branch, and an output convolution operation.

3. The fault detection method for low-voltage intelligent integrated distribution boxes according to claim 2, characterized in that, The process of generating temperature pulse features through the first dilated convolution branch includes: The temperature data is passed through a first causal convolutional layer to generate initial temperature transformation features; The initial temperature transformation features are passed through a first dilated causal convolution residual block to generate temperature spike features; The temperature spike feature is passed through a second dilated causal convolution residual block to generate the temperature pulse feature; The first dilated convolution branch includes a first causal convolution layer, a first dilated causal convolution residual block, and a second dilated causal convolution residual block. The dilation rate of the first dilated causal convolution residual block is less than the dilation rate of the second dilated causal convolution residual block.

4. The fault detection method for low-voltage intelligent integrated distribution boxes according to claim 3, characterized in that, The process of generating heat dissipation features through the second dilated convolution branch includes: The temperature data is passed through a second causal convolutional layer to generate temperature rise segment features; The heat dissipation feature is generated by passing the temperature rise segment feature through a third dilated causal convolution residual block; The second dilated convolution branch includes a second causal convolution layer and a third dilated causal convolution residual block. The kernel size of the second causal convolution layer is larger than the kernel size of the first causal convolution layer, and the dilation rate of the third dilated causal convolution residual block is greater than the dilation rate of the second dilated causal convolution residual block.

5. The fault detection method for a low-voltage intelligent integrated distribution box according to claim 2, characterized in that, The process of generating temperature trend features through upsampling branches includes: The temperature data is passed through an average pooling layer to generate initial temperature trend features; The initial temperature trend features are passed through a convolutional layer to generate the temperature trend features; The upsampling branch includes an average pooling layer and a convolutional layer.

6. The fault detection method for a low-voltage intelligent integrated distribution box according to claim 1, characterized in that, The process of generating temperature change trend features through a temperature trend feature extraction model includes: The multi-scale short-term temperature fluctuation features are passed through a bidirectional LSTM layer to generate temperature development trajectory features; The temperature development trajectory features are weighted at their time steps using a time attention mechanism to generate the temperature change trend features; The temperature trend feature extraction model includes a bidirectional LSTM layer and a time attention mechanism.

7. The fault detection method for a low-voltage intelligent integrated distribution box according to claim 1, characterized in that, The process of determining the type of latent overheating and overload fault warning using a self-encoder includes: The current data is passed through a current sub-encoder to generate current timing features; The concatenated vectors of the current timing features and the temperature change trend features are fused through a fully connected layer to generate temperature-current fused features. The temperature-current fusion features are passed through a temperature decoder to generate reconstructed temperature features; The type of latent overheating and overload fault warning is determined based on the reconstructed temperature characteristics and the reconstruction error of the temperature data. The self-encoder includes a current sub-encoder, a fused fully connected layer, and a temperature decoder.

8. The fault detection method for a low-voltage intelligent integrated distribution box according to claim 7, characterized in that, The process of generating current timing features using a current sub-encoder includes: The current data is passed through a global average pooling layer to generate load level characteristics; The load level characteristics are encoded through a fully connected layer to generate the current timing characteristics; The current sub-encoder includes a global average pooling layer and an encoding fully connected layer.

9. The fault detection method for a low-voltage intelligent integrated distribution box according to claim 7, characterized in that, The process of determining the type of latent overheating and overload fault warning based on the reconstructed temperature characteristics and the error of the temperature data includes: The reconstructed temperature features and temperature data are processed using a mean square error function to obtain the reconstruction error. The warning threshold is calculated based on the quantiles of the reconstruction error from the training set. The type of latent overheating and overload fault warning is determined based on whether the reconstruction error is greater than the warning threshold and whether the current data is in the danger range.

10. The fault detection method for a low-voltage intelligent integrated distribution box according to claim 7, characterized in that, Also includes: The temperature-current fusion features of the training set are used to calculate the KL divergence loss term. The reconstruction error and KL divergence loss term of the training set are weighted and summed to calculate the loss function; The temperature fluctuation feature extraction model, temperature trend feature extraction model, and autoencoder are trained collaboratively based on the loss function.

Images