A method and system for monitoring the operating state of a distribution box

Abstract

The application belongs to the technical field of state monitoring, and particularly relates to a power distribution box operation state monitoring method and system, which comprises the following steps: acquiring multi-modal data of the power distribution box and extracting thermal imaging space features, and aligning with electrical parameter time series to obtain a multi-modal input sequence; after time series reorganization, the multi-modal input sequence is mapped into a query matrix, a key matrix and a value matrix; a dot product basic matrix is added to a frequency domain time series bias matrix to generate an initial attention score; the initial attention score is divided by a temperature scaling coefficient for scaling and normalization to generate an attention weight; the attention weight is weighted with the value matrix to obtain a preliminary context vector, and each mode local kurtosis is mapped into a mode weight; after weighting each subsequence, the preliminary vector is spliced, and after network aggregation and dimension reduction, a classifier is input to output an operation state. The scheme can weaken redundant noise interference and improve the accuracy of the operation state monitoring result.

Description

Technical Field

[0001] This invention relates to the field of condition monitoring technology. More specifically, this invention relates to a method and system for monitoring the operating status of a distribution box. Background Technology

[0002] As a crucial piece of equipment in the power system, the distribution box has an extremely dense distribution of internal components and operates in an exceptionally complex environment. During its actual service life, it is susceptible to early-stage faults with highly concealed characteristics, such as overload heating and poor contact at wiring terminals, due to the combined effects of frequent fluctuations in grid load and alternating external ambient temperatures. These faults typically manifest as weak and non-stationary multimodal physical quantity fluctuations during their incubation phase. Accurately extracting degradation characteristics reflecting the true operating state of the equipment from these dynamic operating conditions mixed with strong background noise has become a pressing technical problem to be solved in ensuring a stable power supply system.

[0003] Currently, monitoring data is acquired by deploying current and voltage transformers and basic point-based temperature probes within the distribution box. The acquired single-dimensional time-series signals are then compared and judged against pre-set fixed safety thresholds. A deep learning algorithm framework is used to initially combine thermal imaging spatial features with electrical parameter time-series sequences, and a basic attention mechanism is introduced to process this multi-dimensional, multi-modal data. Based on these techniques, alarm interception for serious operational faults in the distribution box and preliminary classification and identification of the overall equipment operating status can be achieved.

[0004] However, the aforementioned technical solutions can alleviate the safety risks caused by severe power distribution box failures to some extent and provide basic data perception support. However, existing technologies, when processing non-stationary multimodal sequences, largely rely on conventional linear mapping networks to directly generate query matrices, key matrices, and value matrices, lacking the ability to deeply reconstruct complex temporal features. Furthermore, they fail to integrate orthogonal information from the frequency domain within the attention mechanism, preventing existing models from simultaneously modeling the periodic evolution trend and transient anomalous changes of equipment operation within the same feature space. Simultaneously, traditional attention calculation mechanisms generally use static, fixed scaling factors for score adjustment, detaching from the global statistical distribution characteristics of the input signal itself on a macroscopic time scale. This results in a flattened and insufficiently sparse distribution of attention weights when weak anomalies occur in the early stages. It is difficult to effectively amplify and highlight the latent fault mutation characteristics in the context of strong noise. In addition, in the final multimodal feature aggregation stage, existing technologies usually adopt simple feature splicing or assign fixed static weights to each mode, which seriously ignores the differences in information value of different modal physical quantities due to the different abnormal response speeds in specific local time segments. It fails to establish a dynamic weighting mechanism based on the local data distribution characteristics. This rigid processing at the feature fusion level further restricts the realization of the advantages of multimodal data synergy, ultimately making it difficult to effectively reduce the false alarm rate and false alarm rate of the distribution box when facing complex changing operating conditions and early weak faults. Summary of the Invention

[0005] To address the technical problems of insufficient non-stationary time-series feature extraction capabilities and sensitivity deficiencies caused by fixed attention allocation and static feature fusion mechanisms, this invention provides a method and system for monitoring the operating status of distribution boxes.

[0006] In a first aspect, the present invention provides a method for monitoring the operating status of a distribution box, comprising: S1, acquiring multimodal data of the distribution box and extracting thermal imaging spatial features, aligning the thermal imaging spatial features, temperature time series data, and electrical parameter time series sequences by time steps to obtain a multimodal input feature sequence; S2, mapping the multimodal input feature sequence to a query matrix, a key matrix, and a value matrix after time-series reorganization; obtaining a dot product basis matrix based on the query matrix and the key matrix, and adding the dot product basis matrix to the frequency domain time-series bias matrix obtained based on the multimodal input feature sequence to generate an initial... Attention score; calculate the global kurtosis value of the multimodal input feature sequence and construct a temperature scaling factor. Use the temperature scaling factor to divide and scale the initial attention score, and generate attention weights after normalization. S3, weight the attention weights with the value matrix to obtain the initial context vector. Calculate the local kurtosis value of each modal subsequence and map it to the modal weight. Use the modal weight to weight each modal subsequence and concatenate it with the initial context vector in terms of feature dimensions. After network aggregation and dimensionality reduction, generate the context feature vector and input it into the classifier. Output the power distribution box operation status monitoring results.

[0007] This invention extracts multimodal spatial features and aligns them with electrical parameters in time steps. It uses a gated projection unit with update and reset gates to achieve time-series reassembly, combines orthogonal frequency components to generate a frequency domain time-series bias matrix, calculates the global kurtosis scaling attention score, and integrates local kurtosis to construct modal weights for weighted dimensionality reduction. This allows for accurate modeling of periodic trends and instantaneous changes in the operation of distribution boxes, improves the sparsity and visibility of weak anomaly features, reduces redundant noise interference, and thus improves the accuracy of multimodal operation status monitoring results for distribution boxes.

[0008] Preferably, the step of aligning the thermal imaging spatial features, temperature time-series data, and electrical parameter time-series sequences by time steps to obtain a multimodal input feature sequence includes: extracting the effective value envelope sequences of the thermal imaging spatial feature sequence, temperature time-series data, and high-frequency sampling signals of current and voltage; aligning and sampling the effective value envelope sequences, temperature time-series data, and thermal imaging spatial feature sequences by a linear interpolation algorithm at a fixed time step; standardizing the aligned sequences by subtracting the mean of each modal sequence sample and dividing by the corresponding standard deviation; and performing vector concatenation along the feature channel dimension of each standardized modal sequence to generate a multimodal input feature sequence.

[0009] This invention eliminates clock and dimension differences between different sensors through linear interpolation and standardized mapping, ensuring strict alignment of data feature dimensions.

[0010] Preferably, the step of temporally reorganizing the multimodal input feature sequence and mapping it to a query matrix, a key matrix, and a value matrix includes: generating a reset gate vector and an update gate vector based on the input feature sequence at the current time step and the hidden state at the previous time step; performing a Hadamard product operation on the reset gate vector and the hidden state at the previous time step, concatenating it with the input feature sequence, multiplying it by the corresponding weight matrix, adding a bias term, and obtaining a candidate hidden state through a hyperbolic tangent activation function; performing a Hadamard product operation on the difference between the constant 1 and the update gate vector and the hidden state at the previous time step, and adding the result of the Hadamard product operation on the update gate vector and the candidate hidden state to output a fused hidden vector; concatenating the fused hidden vectors from all steps to form a hidden state matrix, and multiplying it by the weight matrices of the three initialized fully connected network layers to output the query matrix, the key matrix, and the value matrix.

[0011] This invention introduces a gated hidden state operation mechanism with update and reset gates, enabling the network to adaptively determine the proportion of forgetting past historical states and the proportion of accepting current new information based on the dynamic changes of the input feature sequence. This overcomes the limitation of conventional linear mapping in being unable to perform temporal reorganization of non-stationary multimodal sequences and enhances the expressive power of model feature mapping.

[0012] Preferably, the step of generating the reset gate vector and update gate vector based on the input feature sequence of the current time step and the hidden state of the previous time step includes: concatenating the input feature sequence of the current time step and the hidden state of the previous time step in the column direction; multiplying the concatenated vector by the network weight matrix of the reset gate and the network weight matrix of the update gate respectively using matrix multiplication, and then adding the bias terms of the reset gate and the update gate respectively, and then mapping the output value through the Sigmoid function to generate the reset gate vector and the update gate vector respectively.

[0013] Preferably, obtaining the dot product basis matrix based on the query matrix and the key matrix includes: multiplying the query matrix and the transpose of the key matrix, dividing by the square root of the column dimension of the query matrix, and obtaining the dot product basis matrix.

[0014] Preferably, the frequency domain time-series bias matrix obtained based on the multimodal input feature sequence includes: performing a discrete Fourier transform on the multimodal input feature sequence to extract low-frequency and high-frequency components and reconstructing them through inverse Fourier transform to obtain low-frequency and high-frequency time-series features; multiplying the low-frequency and high-frequency time-series features by the corresponding learnable mapping weight matrix and adding them together, and then multiplying them by their own transpose matrix to generate the frequency domain time-series bias matrix.

[0015] This invention constructs a bias matrix by extracting low-frequency and high-frequency components through discrete Fourier transform of the input multimodal sequence. This allows the attention mechanism to take into account both the similarity of time-domain features and the abrupt changes caused by sudden anomalies in the frequency domain, thereby improving the sensitivity and detection capability of the monitoring model for transient weak fault signals in complex noise environments.

[0016] Preferably, the step of calculating the global kurtosis value of the multimodal input feature sequence and constructing the temperature scaling factor includes: statistically analyzing the sample mean and standard deviation of the input sequence; calculating the fourth power of the difference between each element and the sample mean and dividing it by the fourth power of the standard deviation; summing the results, taking the average value, and subtracting a constant 3 to obtain the global kurtosis value; taking the maximum value of the global kurtosis value and the constant 0, adding a constant 1, and calculating the reciprocal of the square root to obtain the temperature scaling factor.

[0017] This invention dynamically calculates global kurtosis based on the fourth standard moments of the overall data and constructs a negative correlation temperature scaling coefficient. When an abnormal polarity change occurs, this coefficient can significantly widen the score difference between the normal change step and the background noise step. Through subsequent normalization operations, the network weights are highly focused on a very small number of time steps with abnormal responses, which greatly improves the attention sparsity characteristics of abnormal segments.

[0018] Preferably, the step of weighting the attention weights and the value matrix to obtain the preliminary context vector, and calculating the local kurtosis value of each modal subsequence, includes: performing matrix multiplication on the attention weights and the value matrix to obtain the preliminary context vector; setting an overlapping sliding window with a fixed width and a sliding step size for the input subsequence of each data modality along the time axis, and keeping the sequence length unchanged by combining edge padding operation; using the data sample group enclosed by each overlapping sliding window, calculating the fourth central moment of the sample group and dividing it by the square of the second central moment to obtain the local kurtosis value of the corresponding modal time.

[0019] This invention establishes a local kurtosis assessment method that can finely quantify the degree of anomaly and thick-tail characteristics of different modal flows within a specific local time segment by implementing an overlapping sliding window of fixed width along the time axis and solving the central moment index. This provides a highly reliable basis for data distribution characteristics for subsequent capture of instantaneous non-stationary disturbances and accurate weight allocation.

[0020] Preferably, the step of weighting each modal subsequence with modal weights and concatenating it with the initial context vector along the feature dimension, and generating a context feature vector through network aggregation and dimensionality reduction, includes: performing point-to-point numerical multiplication on the time sequence values ​​contained in the corresponding input subsequences using modal weights to obtain weighted modal subsequences; concatenating each weighted modal subsequence with the initial context vector along the feature dimension to form a fused feature tensor; inputting the fused feature tensor into a multilayer feedforward neural network with a two-layer network structure to perform two mapping dimensionality reductions, and performing a global average pooling operation to output a context feature vector.

[0021] This invention utilizes a perceptron model to dynamically map the local kurtosis values ​​of each modal flow into normalized modal weights, which are then applied to the corresponding subsequences to perform point-to-point multiplication amplification. This enhances the modal features containing transient anomaly information and effectively reduces data redundancy and information dilution interference caused by anomaly-free modal sequences.

[0022] Secondly, the present invention provides a power distribution box operation status monitoring system, including a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-mentioned power distribution box operation status monitoring method is implemented.

[0023] By adopting the above technical solution, a computer program is generated from the above-mentioned method for monitoring the operating status of a distribution box, and stored in a memory for loading and execution by a processor. This allows for the creation of a terminal device based on the memory and processor, making it convenient to use.

[0024] The technical solution of the present invention has the following beneficial technical effects:

[0025] This invention acquires current, voltage, temperature, and thermal imaging data of a distribution box, extracts spatial features of the thermal imaging, and aligns them with the time-series sequence of electrical parameters to obtain a unified multimodal input feature sequence. This ensures that the spatial distribution information reflected by the thermal imaging corresponds to the temporal change information reflected by the electrical parameters at the same time, thus overcoming the shortcomings of monitoring only a single electrical parameter with a single dimension and providing a consistent input for representing the true operating status of the equipment.

[0026] This invention uses a gated projection unit with built-in update and reset gates to perform temporal recombination of the input feature sequence and map it into a query matrix, a key matrix, and a value matrix. Compared with conventional linear mapping, the recombinated sequence can express both the gradual load change trend and the instantaneous change in non-stationary multimodal data, and enhances the feature mapping expressive power of the query matrix, key matrix, and value matrix.

[0027] This invention superimposes a frequency domain temporal bias matrix onto the dot product of the query matrix and the key matrix. The frequency domain temporal bias matrix is ​​generated by extracting orthogonal frequency components and weighting them through a discrete Fourier transform of the input feature sequence. This allows the attention score to carry both periodic trend information and instantaneous mutation information, solving the problem that existing attention mechanisms cannot incorporate frequency domain information and are difficult to model both the periodic trend of equipment operation and instantaneous abnormal mutations at the same time. This enhances the ability to detect transient features caused by faults in noisy environments.

[0028] This invention calculates the global kurtosis value of the input feature sequence, constructs a temperature scaling factor that is negatively correlated with the global kurtosis value, and uses the temperature scaling factor to perform division scaling on the attention score after superimposed bias. This tightens the attention distribution when anomalies occur, improves the attention sparsity of the anomaly features, and overcomes the shortcomings of fixed scaling factors that do not take into account the global statistical distribution characteristics of the input signal and the insufficient sparsity of the attention distribution when anomalies occur. This amplifies and highlights weak anomaly features.

[0029] This invention maps the local kurtosis values ​​of each modal subsequence to modal weights. After weighting each subsequence with modal weights, it concatenates the feature dimensions with the initial context vector. Then, it generates a context feature vector through network aggregation and dimensionality reduction. The weights are dynamically allocated according to the differences in information value of different modal subsequences in local time segments, highlighting the modal features of abnormal signal segments. This makes up for the shortcomings of simple feature concatenation or static weight allocation that misses the differences in local information value, and reduces the interference of redundant noise.

[0030] By integrating time step alignment, gated timing reassembly, frequency domain timing bias superposition, global kurtosis temperature scaling, and local kurtosis mode weighting, this invention achieves joint modeling of equipment cycle trends and instantaneous changes, as well as highlighting the value of local information of abnormal modes, thereby improving the accuracy of the monitoring results of the operating status of the distribution box under complex working conditions and early minor faults. Attached Figure Description

[0031] Figure 1 A flowchart of a method for monitoring the operating status of a distribution box; Figure 2 This is a schematic diagram of the aligned multimodal input feature sequence; Figure 3 This is a schematic diagram of modal weight allocation based on local kurtosis. Detailed Implementation

[0032] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0033] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0034] This invention discloses a method for monitoring the operating status of a distribution box, referring to... Figure 1 This includes the following steps: S1. Extract data and generate input feature sequences. Specifically, acquire current data, voltage data, temperature data, and thermal imaging data from the distribution box, extract thermal imaging spatial features, align the thermal imaging spatial features, temperature time series data, and electrical parameter time series with time steps to obtain a unified multimodal input feature sequence, and input the multimodal input feature sequence into the monitoring model.

[0035] The system reads three-phase current data, three-phase voltage data, and time-series one-dimensional data from a smart meter in the distribution box via RS485 serial port, as well as from a contact-type multi-channel temperature sensor. Cross-channel averages are calculated for the three-phase current data, three-phase voltage data, and multi-channel temperature data to obtain one-dimensional current, voltage, and temperature sequences representing the overall state. Using the OpenCV library's VideoCapture interface, a two-dimensional thermal imaging video stream from an infrared thermal imager is acquired at a set frame rate. The ResNet50 residual network model in the PyTorch deep learning framework is called to extract two-dimensional spatial feature maps from each frame of the thermal imaging image. These feature maps are then flattened into one-dimensional thermal imaging feature vectors using an average pooling layer.

[0036] To address the issue of inconsistent sampling frequencies across multiple sensors, a fixed time step is used as the baseline time axis to resample and align the temperature data and thermal imaging feature vectors. The aligned current, voltage, and temperature data, along with the thermal imaging feature vectors, are then concatenated along the feature dimension to obtain a unified multimodal input feature sequence tensor. This multimodal input feature sequence tensor is then loaded in batches into a GPU-accelerated operational status monitoring model using PyTorch's DataLoader module.

[0037] In an optional embodiment, high-frequency AC current signals and high-frequency AC voltage signals are acquired using current transformers and voltage transformers configured in the distribution box, at a sampling rate preferably ranging from 5kHz to 10kHz, for example, set to a sampling rate of 8kHz. A non-overlapping sliding window with a width of 20ms is used to calculate the root mean square value, generating an effective value envelope sequence with an update rate of 50Hz. The mean of the three-phase envelopes of the effective value envelope sequence is calculated, and this mean is used as a one-dimensional representative sequence of electrical parameters.

[0038] Simultaneously, an infrared thermal imager fixed inside the distribution box acquires a 160x120 resolution thermal imaging matrix at a frequency of 1 to 5 frames per second. This thermal imaging matrix is ​​then input into a pre-trained lightweight convolutional neural network (CNN). The CNN's structure includes multiple cascaded convolutional layers and activation function layers for extracting local spatial features, as well as a global average pooling layer at the end. The output of the CNN is a spatial feature sequence. After removing the fully connected layers, the CNN performs global average pooling, outputting a 128-dimensional spatial feature sequence. For multi-channel temperature sensors, Celsius time-series data is output at a frequency of 1 Hz.

[0039] During the alignment and stitching stage, due to the differences in the sampling clocks of the multimodal sensor hardware, a fixed time step is used as the reference time axis. The fixed time step is set to 100ms or 200ms, for example, 100ms is set as the reference time axis. A first-order linear interpolation algorithm is used to resample and align the effective value envelope sequence, temperature time series data, and spatial feature sequence to ensure that each modal data has corresponding samples at the same time.

[0040] The aligned sequence enters the Z-score normalization module to eliminate dimensional and numerical range differences caused by voltage, current, temperature, and eigenvector data.

[0041] The standardized modal data are sequentially concatenated along the feature channel dimensions. For example, a one-dimensional voltage sequence, a one-dimensional current sequence, a one-dimensional temperature sequence are concatenated with a 128-dimensional thermal imaging spatial feature sequence to generate a unified input feature sequence tensor with 131 dimensions at each time step.

[0042] S2. Calculate attention weights. Specifically, the input feature sequence is temporally reorganized and mapped into a query matrix, key matrix, and value matrix using a gated projection unit with built-in update and reset gates. A frequency domain temporal bias matrix is ​​superimposed on the dot product of the query matrix and key matrix. The frequency domain temporal bias matrix extracts orthogonal frequency components by performing a discrete Fourier transform on the input feature sequence and weights them to generate modeling periodic trends and instantaneous abrupt changes. The global kurtosis value of the input feature sequence is calculated, and a temperature scaling coefficient negatively correlated with the global kurtosis value is constructed. The attention score after superimposed bias is divided and scaled using the temperature scaling coefficient to improve the attention sparsity of abnormal features. Attention weights are then generated after normalization.

[0043] In PyTorch, a gated projection unit is constructed, internally instantiating two fully connected linear layers. These layers, combined with the Sigmoid activation function, generate update and reset gate vectors, respectively. The reset gate vector is used to perform element-wise multiplication of the historical hidden states to calculate candidate hidden states. The update gate vector is then used to perform a weighted sum of the historical and candidate hidden states to complete temporal reassembly. Three linear mapping layers (Linear) are invoked to project the reassembled sequence into query, key, and value matrices, respectively. The dot product of the query and key matrices (transposes) is performed to obtain the dot product basis matrix. This basis matrix is ​​then added to the frequency domain temporal bias matrix obtained from the multimodal input feature sequence to obtain the initial attention score matrix.

[0044] Simultaneously, a one-dimensional discrete Fourier transform is performed on the multimodal input feature sequence to extract low-frequency and high-frequency components, which are then reconstructed through inverse Fourier transform. These components are multiplied by their corresponding mapping weight matrices and then fused. The fused frequency domain features are multiplied by their transposes to construct a frequency domain temporal bias matrix with the same dimension as the attention score matrix. The frequency domain temporal bias matrix is ​​then element-wise superimposed onto the dot product basis matrix using the addition operator to obtain the initial attention score matrix.

[0045] The global fourth-order standard moment (GAS) of the multimodal input feature sequence in the time dimension is calculated, i.e., the global kurtosis value. The maximum value of the global kurtosis value and the constant 0 is added to a constant 1, and the reciprocal of the square root is calculated to construct a temperature scaling factor scalar that is negatively correlated with the global kurtosis value. The attention score matrix after superimposed bias is divided by this temperature scaling factor using a division operator. When the global kurtosis value is large, indicating the presence of anomalous abrupt changes, the temperature scaling factor decreases, resulting in a larger variance in the score distribution after scaling, thereby improving the attention sparsity of anomalous features. Exponential normalization is performed in the last dimension to generate the attention weight matrix.

[0046] The implementation of the gated projection unit is a variant of a recurrent neural network. The input is a feature sequence containing multiple time steps. The structure includes built-in update gates and reset gates as well as associated weight matrices. The output is a fused hidden vector.

[0047] The input data is a feature sequence generated by the front-end module, containing multiple time steps, for example, a time window length of 100 steps, and each time step having a feature dimension of 131. The gated projection unit adopts a variant GRU structure with 256 hidden units.

[0048] Based on the well-known computational principles of gated recurrent units in deep learning, and considering the settings for the input and hidden state dimensions, a 131-dimensional input feature sequence is concatenated with the 256-dimensional hidden state from the previous time step, resulting in a 387-dimensional combined vector. The network weight matrix sizes for both the reset and update gates are set to 256×387, and the bias terms for both are set to 256 dimensions. Matrix multiplication is used to multiply the concatenated vector by the network weight matrices for both the reset and update gates, and then the bias terms for both are added. The output values ​​are then mapped to the 0-1 range using the Sigmoid function, generating reset and update gate vectors to determine the proportion of past states to be forgotten and the proportion of new information to be received, respectively. For example, when a sudden short circuit is detected in the distribution box, a specific dimension of the update gate will approach 1, prompting the hidden state to accept the new information.

[0049] The reset gate vector is applied to the hidden state of the previous time step to perform a Hadamard product operation. The result is concatenated with the input feature sequence, multiplied by the candidate hidden state network weight matrix of the same size (256×387), and a bias term of the candidate hidden state is added. Then, it is processed by the hyperbolic tangent activation function to generate a 256-dimensional candidate hidden state with a value range between -1 and 1.

[0050] The difference between the constant 1 and the update gate vector is multiplied by the Hadamard product of the hidden state at the previous time step. This result is then added to the Hadamard product of the update gate vector and the candidate hidden state at the current time step, outputting the fused hidden vector for each time step. The fused hidden vectors from 100 time steps are stacked row-wise to construct a hidden state matrix of dimension 100×256.

[0051] The hidden state matrix is ​​multiplied by three fully connected network layer weight matrices initialized using a uniform Xavier distribution. These three weight matrices include the query matrix weight matrix, the key matrix weight matrix, and the value matrix weight matrix, all with a size of 256×256. The input to each fully connected network layer is the hidden state matrix. Each fully connected network layer consists of a single linear weight matrix layer, and its outputs are the mapped query, key, and value matrices. Matrix multiplication generates query, key, and value matrices with the same dimension of 100×256.

[0052] In an optional embodiment, based on the well-known common sense of scaling dot product calculation in the attention mechanism, and combined with the scaling mechanism of feature dimensions, batch matrix multiplication is performed using the transpose of the key matrix with a time window length of 100 and a feature dimension of 256 and a dimension of 256×100 to obtain a basic relevance matrix of size 100×100. Simultaneously, the basic relevance matrix is ​​divided by the square root of the column dimension of the query matrix, 16, to alleviate the problem of inner product numerical inflation caused by excessively large feature dimensions, thus obtaining the dot product basis matrix.

[0053] Based on the theory of time-frequency feature fusion, a one-dimensional fast Fourier transform is applied to the input feature sequence with a time step length of 100 and a channel number of 131 during the frequency domain feature extraction stage. By setting a truncation frequency threshold, for example, setting 2Hz at the sampling frequency as the boundary line, which is determined based on the frequency band interval between the main frequency of the daily load fluctuation of the distribution box and the starting frequency of the fault transient component, the complex spectrum is divided into low-frequency and high-frequency parts. Inverse Fourier transforms are then performed on these parts to extract 100×131-dimensional low-frequency time-series features representing the fundamental wave operating trend and the stable load fluctuation law, as well as 100×131-dimensional high-frequency time-series features detecting the violent fluctuations and transient changes in the effective value envelope caused by arc discharge or sudden overload. The extracted low-frequency and high-frequency time-series features are then subjected to linear projection dimensionality reduction through learnable low-frequency mapping weight matrices and learnable high-frequency mapping weight matrices, respectively. The size of the learnable low-frequency mapping weight matrix and the size of the learnable high-frequency mapping weight matrix are set to 131×64, so that the dimensions of the low-frequency time series features after dimensionality reduction and the dimensions of the high-frequency time series features after dimensionality reduction are both unified to 100×64.

[0054] Matrix operations are used to add and fuse the dimensionality-reduced low-frequency and high-frequency time-series features. A 100×100 square matrix is ​​constructed as the frequency-domain time-series bias matrix by multiplying the fused result with its transpose. Element-wise tensor addition is then used to stack the frequency-domain time-series bias matrix positionally onto the dot product base matrix, generating a 100×100 initial attention score matrix that balances temporal feature similarity with frequency-domain abrupt changes. This enhances the monitoring model's ability to detect transient envelope anomalies caused by faults in noisy environments.

[0055] In an optional embodiment, based on the kurtosis calculation theory in mathematical statistics, a statistical analysis of the global kurtosis value is performed to flatten the current time window. For example, setting the time step to 100 steps and the feature dimension to 131, all input feature sequence elements within this window are obtained, resulting in a sample set of 13,100 data points. The microprocessor iterates through the sample set, calculates the fourth power of the difference between each element in the sequence and the sample mean, divides it by the fourth power of the standard deviation, sums the calculation results for all elements, takes the average, and then subtracts a constant of 3 to obtain the global kurtosis value. For example, when the distribution box is operating normally and stably, the sample distribution is close to a Gaussian distribution, and the absolute value of the global kurtosis tends to be 0 or negative; when polarity anomalies such as partial discharge occur, the long-tail characteristics of the data become apparent, and the global kurtosis can increase to 4.5 or even higher.

[0056] Based on the temperature scaling principle in the attention mechanism, combined with the anomaly long-tail feature detection mechanism, the maximum value of the global kurtosis and the constant 0 is added to the constant 1, and then the reciprocal of its square root is calculated to obtain the temperature scaling factor that decreases as the positive anomaly kurtosis value increases. Continuing with the anomaly example, if the global kurtosis is equal to 4.5, and the maximum value after taking the value with 0 is still 4.5, then the temperature scaling factor is equal to 0.426; if it is in a stable and normal state, such as a global kurtosis of -1.5, and the maximum value after taking the value with 0 is truncated to 0, then the temperature scaling factor is equal to 1.

[0057] The model performs a scalar division operation on all elements within the 100×100-dimensional initial attention score matrix generated in the previous steps, dividing them uniformly by the generated temperature scaling factor. In abnormal scenarios, dividing by 0.426 is equivalent to magnifying each element of the score matrix by nearly 2.35 times, thus widening the absolute difference between the abnormal mutation step and the background noise step.

[0058] Applying the Softmax function along the rows of the 100×100 matrix widens the score gap after exponentialization, causing most weights to focus on at least a few time steps with anomalous responses, generating an attention weight matrix with sparse properties and reducing redundant interference data.

[0059] S3. Generate context feature vectors and output monitoring results. Specifically, the attention weights are weighted with the value matrix to obtain a preliminary context vector, and the local kurtosis value of each modal subsequence is calculated. The local kurtosis value is mapped to the modal weight. The modal weights are used to weight each subsequence and then concatenated with the preliminary context vector in terms of feature dimension. The network aggregates and reduces the dimensionality to generate a context feature vector. The context feature vector is input into the classifier to output the monitoring results of the power distribution box operation status.

[0060] The attention weight matrix and the value matrix are multiplied to obtain a preliminary context vector that integrates global temporal correlations. The input sequence is split into four subsequences based on modality: current, voltage, temperature, and thermal imaging. Local kurtosis values ​​are calculated for each of the four subsequences at each time step. The four local kurtosis values ​​at the current time step are concatenated into a one-dimensional vector. This vector is then input into a weight generation network containing a single linear layer and a Softmax activation function, mapping out four dynamic modal weight values ​​whose sum equals 1. Tensor multiplication is used to perform point-to-point numerical multiplication between the dynamic modal weight values ​​at each time step and the corresponding modal subsequence, completing the weighting.

[0061] The four weighted subsequences are concatenated and fused with the initial context vector along the feature dimension. The concatenated tensor is then input into a multilayer perceptron network consisting of two fully connected linear layers and a GELU activation function. Spatial mapping and feature dimension compression and aggregation are performed layer by layer, and global average pooling is performed along the time dimension to generate a context feature vector with fixed dimensions.

[0062] The context feature vector is input into a classifier consisting of a single-layer linear mapping layer and a LogSoftmax activation function layer. The log probability distribution of each operating status category is calculated, and the index corresponding to the maximum probability value is extracted to obtain the operating status monitoring results of the distribution box.

[0063] In an optional embodiment, the attention weighting operation is computed via matrix multiplication. An attention weight matrix with both row and column dimensions of 100×100 is multiplied by a value matrix of dimension 100×256 to generate a preliminary context vector of dimension 100×256. This preliminary context vector is then fused with information based on the global correlations of the current feature sequence at each time step.

[0064] For single-modal subsequences extracted from the input sequence, such as one-dimensional current and voltage sequences, local sliding window analysis is performed along the time axis. A zero-value padding strategy is employed, for example, adding 4 to 5 zero values ​​at both ends of the sequence. This ensures that when the sliding window width is set to 10 and the sliding step size is set to 1, the length of the processed output local sequence remains unchanged at the original 100 time steps. For subsequences with a feature dimension greater than 1, such as 128-dimensional thermal imaging spatial feature sequences, the average value is calculated along the feature channel dimension before calculating the local kurtosis. This average value is then converted into a one-dimensional time series before local sliding window analysis is performed. The sliding window width is set to 10, approximately twice the duration of a typical fault transient, ensuring that the abnormal event is completely encompassed. The sliding step size is set to 1, allowing independent calculation of the local kurtosis for each time step.

[0065] Within a specific sample group enclosed by an overlapping sliding window, the microcontroller obtains the local mean through an accumulation and summation algorithm, calculating the second-order central moment representing the variance and the fourth-order central moment representing the thick-tailed characteristic of extreme values. The fourth-order central moment of the sample group is calculated and divided by the square of its second-order central moment, and the result is assigned as the local kurtosis value for each data mode at that moment. For a current data mode window at a certain moment, assuming that a transient arcing phenomenon is caused by contact loosening, the high-frequency glitches contained in the current data mode window will lead to an increase in the fourth-order central moment. The calculated local kurtosis value for the current moment is, for example, equal to 6.2. Meanwhile, the adjacent pure temperature data mode, due to its slow change in thermal inertia, has a calculated local kurtosis value maintained at around 1.8. By solving for the local kurtosis values, the model generates a kurtosis index that represents the degree of transient anomaly of each data mode at each specific moment.

[0066] In an optional embodiment, the mapping and transformation process for multimodal local kurtosis values ​​is defined, assuming the input includes four modal data streams: current mode data stream, voltage mode data stream, temperature mode data stream, and image mode data stream. The four-dimensional local kurtosis value vector concatenated at the current time step is fed into a single-layer perceptron model for normalization. The input to the perceptron model is the concatenated local kurtosis value vector. The perceptron model structure includes a single-layer linear mapping layer and a Softmax activation function layer. The output of the perceptron model is the modal weights corresponding to each data mode.

[0067] The perceptron model uses a pre-defined 4×4-dimensional mapping parameter matrix and a four-dimensional bias vector to linearly project the concatenated vector of local kurtosis values. This is followed by a Softmax layer to perform normalization calculations, yielding modal weights for the four data modes. The sum of the output weights for the four data modes equals 1. For example, if insulation failure occurs at a certain moment, causing leakage, the current mode anomaly results in a calculated weight allocation of 0.6, while the voltage, temperature, and image modes are assigned modal weights of 0.15, 0.1, and 0.15, respectively.

[0068] Reference Figure 3 When anomalies occur in the current mode within the 40-60 time step interval, the model increases the weight of the current mode to 0.6 based on the local kurtosis calculation results, while correspondingly reducing the weights of the voltage, temperature, and image modes. This weight allocation phenomenon indicates that when a mode experiences a local kurtosis increase within a specific time interval, the mode weighting mechanism can automatically allocate higher weights to the anomalous mode, making the features of the anomalous mode stand out in the fusion stage. This verifies that dynamic mode weighting based on local kurtosis can distinguish the information value of each mode in local time segments.

[0069] The obtained modal weights are multiplied point-to-point with the time sequence values ​​contained in the corresponding input subsequence, thereby enhancing the numerical representation of the abnormal current modes on the feature map.

[0070] After weighting, the model concatenates the four weighted modal feature subsequences to maintain a total dimension of 131. The model then concatenates this 131-dimensional result with the pre-calculated 256-dimensional initial context vector along the feature channel axis, constructing a fused feature tensor with 100 time steps and a feature channel thickness of 387 dimensions. This fused feature tensor is then fed into a multilayer feedforward neural network with a two-layer structure for aggregation and dimensionality reduction. The input to the multilayer feedforward neural network is the fused feature tensor. The multilayer feedforward neural network consists of two fully connected mapping layers, each equipped with a GELU activation function. The output of the multilayer feedforward neural network is the dimensionality-reduced context feature vector.

[0071] In the specific computation, the fused feature tensor is multiplied by the first-layer network node mapping parameter matrix with a dimension of 387×512 and a bias term of the first-layer network node mapping parameter matrix is ​​added. After processing by the GELU activation function, it is multiplied again by the second-layer network node mapping parameter matrix with a dimension of 512×256 and a bias term of the second-layer network node mapping parameter matrix is ​​added. After processing by the GELU activation function again, a sequence feature matrix with a size maintained at 100×256 is obtained. The model calculates the average value of the sequence feature matrix along the dimension of 100 time steps, that is, performs global average pooling, compresses and aggregates the two-dimensional sequence data, eliminates the dimension of time steps, and outputs a single 256-dimensional context feature vector. The context feature vector is fed into the downstream classifier module to output the monitoring and diagnostic results of the normal operation status, overload status, short circuit status, or poor contact status of the distribution box.

[0072] The experiment was conducted based on a historical operational dataset of a distribution box, containing 5000 normal samples and 1500 abnormal samples, with anomaly types including overload, short circuit, and poor contact. The model training was set with a learning rate of 0.1%, using the Adam algorithm as the optimizer, a batch size of 64, and a total of 100 iterations. The learning rate was set according to the default recommended value of the Adam optimizer. The evaluation metrics were accuracy, F1 score, and anomaly detection recall. The comparative experiment included four network groups: a basic network without any additional processing modules, a network with only a frequency domain temporal bias module, a network with a global kurtosis temperature scaling module, and a complete model containing a frequency domain temporal bias module, a global kurtosis temperature scaling module, and a local kurtosis mode weighting module.

[0073] Regarding the specific experimental data and results, the accuracy of the basic network was 89.4%, the F1 score was 87.2%, and the anomaly detection recall was 84.5%. After adding the frequency domain temporal bias module, the model's accuracy increased to 92.1%, the F1 score reached 90.8%, and the anomaly detection recall was 89.3%. After adding the global kurtosis temperature scaling module to the basic network, the accuracy improved to 94.6%, the F1 score reached 93.5%, and the anomaly detection recall was 94.1%. The complete model applying all three additional processing modules performed best, achieving an accuracy of 97.8%, an F1 score of 97.2%, and an anomaly detection recall of 98.5%.

[0074] This invention also discloses a power distribution box operation status monitoring system, including a processor and a memory. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, a power distribution box operation status monitoring method according to the present invention is implemented.

[0075] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for monitoring the operating status of a distribution box, characterized in that, include: S1. Acquire multimodal data of the distribution box and extract thermal imaging spatial features. Align the thermal imaging spatial features, temperature time series data and electrical parameter time series with time steps to obtain multimodal input feature sequence. S2. After temporal reorganization of the multimodal input feature sequence, it is mapped into a query matrix, a key matrix, and a value matrix; based on the query matrix and the key matrix, the dot product basis matrix is ​​obtained, and the dot product basis matrix is ​​added to the frequency domain temporal bias matrix obtained based on the multimodal input feature sequence to generate the initial attention score; The global kurtosis value of the multimodal input feature sequence is calculated and a temperature scaling factor is constructed. The initial attention score is then divided and scaled using the temperature scaling factor, and attention weights are generated after normalization. S3. The attention weights are weighted with the value matrix to obtain the initial context vector. The local kurtosis values ​​of each modal subsequence are calculated and mapped to modal weights. The modal weights are used to weight each modal subsequence and then concatenated with the initial context vector in terms of feature dimensions. After network aggregation and dimensionality reduction, the context feature vector is generated and input into the classifier to output the power distribution box operation status monitoring results.

2. The method for monitoring the operating status of a distribution box according to claim 1, characterized in that, The step of aligning thermal imaging spatial features, temperature time-series data, and electrical parameter time-series sequences by time steps to obtain a multimodal input feature sequence includes: extracting the effective value envelope sequences of the thermal imaging spatial feature sequence, temperature time-series data, and high-frequency sampling signals of current and voltage; aligning and sampling the effective value envelope sequences, temperature time-series data, and thermal imaging spatial feature sequences at a fixed time step using a linear interpolation algorithm; standardizing the aligned sequences by subtracting the mean of each modal sequence sample and dividing by the corresponding standard deviation; and performing vector concatenation along the feature channel dimension of each standardized modal sequence to generate a multimodal input feature sequence.

3. The method for monitoring the operating status of a distribution box according to claim 2, characterized in that, The process of temporally reorganizing the multimodal input feature sequence and mapping it to a query matrix, a key matrix, and a value matrix includes: generating a reset gate vector and an update gate vector based on the input feature sequence at the current time step and the hidden state at the previous time step; performing a Hadamard product operation on the reset gate vector and the hidden state at the previous time step, concatenating it with the input feature sequence, multiplying it by the corresponding weight matrix, adding a bias term, and then applying the hyperbolic tangent activation function to obtain candidate hidden states; performing a Hadamard product operation on the difference between the constant 1 and the update gate vector and the hidden state at the previous time step, and adding the result of the Hadamard product operation on the update gate vector and the candidate hidden state to output a fused hidden vector; concatenating the fused hidden vectors from all steps to form a hidden state matrix, and multiplying it by the weight matrices of the three initialized fully connected network layers to output the query matrix, the key matrix, and the value matrix.

4. The method for monitoring the operating status of a distribution box according to claim 3, characterized in that, The step of generating reset gate vectors and update gate vectors based on the input feature sequence of the current time step and the hidden state of the previous time step includes: concatenating the input feature sequence of the current time step and the hidden state of the previous time step in the column direction; multiplying the concatenated vectors by the network weight matrix of the reset gate and the network weight matrix of the update gate respectively using matrix multiplication, and then adding the bias terms of the reset gate and the update gate respectively; and finally mapping the output values ​​through the Sigmoid function to generate reset gate vectors and update gate vectors respectively.

5. The method for monitoring the operating status of a distribution box according to claim 1, characterized in that, The step of obtaining the dot product basis matrix based on the query matrix and the key matrix includes: multiplying the query matrix and the transpose of the key matrix, and dividing by the square root of the column dimension of the query matrix to obtain the dot product basis matrix.

6. The method for monitoring the operating status of a distribution box according to claim 1, characterized in that, The frequency domain time-series bias matrix obtained based on the multimodal input feature sequence includes: performing a discrete Fourier transform on the multimodal input feature sequence to extract low-frequency and high-frequency components and reconstructing them through inverse Fourier transform to obtain low-frequency and high-frequency time-series features; multiplying the low-frequency and high-frequency time-series features by the corresponding learnable mapping weight matrix and adding them together, and then multiplying them by their own transpose matrix to generate the frequency domain time-series bias matrix.

7. The method for monitoring the operating status of a distribution box according to claim 1, characterized in that, The calculation of the global kurtosis value of the multimodal input feature sequence and the construction of the temperature scaling factor include: statistically analyzing the sample mean and standard deviation of the input sequence; calculating the fourth power of the difference between each element and the sample mean and dividing it by the fourth power of the standard deviation; summing the results, taking the average value, and subtracting a constant 3 to obtain the global kurtosis value; taking the maximum value of the global kurtosis value and the constant 0, adding a constant 1, and calculating the reciprocal of the square root to obtain the temperature scaling factor.

8. The method for monitoring the operating status of a distribution box according to claim 1, characterized in that, The step of weighting the attention weights and the value matrix to obtain a preliminary context vector, and calculating the local kurtosis value of each modal subsequence, includes: performing matrix multiplication on the attention weights and the value matrix to obtain a preliminary context vector; setting an overlapping sliding window with a fixed width and a sliding step size for the input subsequence of each data modality along the time axis, and keeping the sequence length unchanged by combining edge padding operation; using the data sample group enclosed by each overlapping sliding window, calculating the fourth central moment of the sample group and dividing it by the square of the second central moment to obtain the local kurtosis value of the corresponding modal time.

9. A method for monitoring the operating status of a distribution box according to claim 8, characterized in that, The process of weighting each modal subsequence with modal weights and concatenating it with the initial context vector along the feature dimension, followed by network aggregation and dimensionality reduction to generate a context feature vector, includes: performing point-to-point numerical multiplication on the time sequence values ​​contained in the corresponding input subsequences using modal weights to obtain weighted modal subsequences; concatenating each weighted modal subsequence with the initial context vector along the feature dimension to form a fused feature tensor; inputting the fused feature tensor into a multilayer feedforward neural network with a two-layer network structure to perform two mapping dimensionality reduction operations, and performing a global average pooling operation to output a context feature vector.

10. A power distribution box operation status monitoring system, characterized in that, include: A processor and a memory, wherein the memory stores computer program instructions that, when executed by the processor, implement a method for monitoring the operating status of a distribution box according to any one of claims 1-9.

Images