Power distribution network fault online diagnosis method and system based on edge computing

Abstract

The application discloses a power distribution network fault online diagnosis method and system based on edge computing, which effectively removes random noise and extracts energy spectrum entropy reflecting the essential characteristics of fault arc by adaptive modal separation of original signals on the edge side, and then performs high-dimensional nonlinear space mapping by using a pruned and quantized lightweight deep separable convolutional neural network to compensate for the insufficient adaptability of traditional artificial discrimination threshold in a dynamic environment. In this way, not only can the weak high-resistance ground fault fingerprint be accurately captured on the resource-limited edge terminal, greatly reducing the false positive and false negative rates, but also the low-delay response of the diagnosis process is ensured through the lightweight model, finally realizing the online accurate identification and high-confidence reasoning of the power distribution network fault, and providing reliable technical support for the safe and stable operation of the intelligent power distribution network.

Description

Technical Field

[0001] This application relates to the field of intelligent diagnostics, and more specifically, to a method and system for online fault diagnosis of power distribution networks based on edge computing. Background Technology

[0002] As a crucial link in the power system serving end users, the safety and reliability of the distribution network are of great significance to ensuring social production and daily life. With the in-depth advancement of the construction of new power systems, the large-scale integration of distributed power sources has led to increasingly complex distribution network topologies and frequent changes in load conditions. In order to capture system anomalies in real time, shorten fault repair time, and improve power supply continuity, the construction of an efficient and robust online fault diagnosis solution for the distribution network has become an urgent need for intelligent operation and maintenance of the power grid.

[0003] However, existing online fault diagnosis solutions for distribution networks still face significant challenges in practical applications. Current mainstream diagnostic strategies largely rely on centralized processing at the master station or local discrimination logic based on fixed thresholds. Due to the presence of numerous nonlinear loads and random electromagnetic interference in distribution networks, especially during high-impedance ground faults (HIF), the fault current is typically much smaller than the system load current, and the waveform exhibits significant random nonlinear characteristics. In this non-stationary random noise context, traditional linear analysis methods struggle to effectively isolate fault characteristics, easily leading to numerous false alarms or missed alarms when distinguishing between genuine weak faults and normal load fluctuations. Furthermore, current centralized cloud computing architectures face limitations such as high communication bandwidth consumption and data transmission latency, making it difficult to meet the real-time requirements of fault diagnosis. Even though some solutions attempt to introduce deep learning to improve recognition rates, standard models often suffer from excessive computational loads, making them difficult to deploy on resource-constrained edge hardware, resulting in a lack of adaptive analysis strategies for complex fault signals at the edge.

[0004] Therefore, an optimized online fault diagnosis method for distribution networks based on edge computing is needed. Summary of the Invention

[0005] To address the aforementioned technical issues, this application provides a method and system for online fault diagnosis in power distribution networks based on edge computing.

[0006] According to one aspect of this application, an online fault diagnosis method for distribution networks based on edge computing is provided, comprising: S1: By deploying edge acquisition terminals at each monitoring node of the power distribution network, the acquired raw sensor data stream is truncated by sliding window, zero-drift calibration and normalization to obtain preprocessed data frames; S2: Perform adaptive mode separation on the preprocessed data frame for non-stationary signals to obtain several intrinsic mode components; S3: Extract time-frequency energy spectrum entropy features from several intrinsic mode components to obtain energy spectrum entropy feature vectors; S4: Using a lightweight, deeply separable convolutional neural network that has been pre-deployed in an edge computing unit and pruned and quantized, the energy spectrum entropy feature vector is mapped to a high-dimensional nonlinear space to obtain a deep feature representation tensor. S5: Perform fault probability calculation and diagnostic reasoning on the deep feature representation tensor to obtain the final diagnostic result, which includes fault type, fault phase and confidence level.

[0007] According to another aspect of this application, an online fault diagnosis system for distribution networks based on edge computing is provided, comprising: The data preprocessing module is used to perform sliding window truncation, zero drift calibration and normalization on the acquired raw sensor data stream through edge acquisition terminals deployed at each monitoring node of the power distribution network to obtain preprocessed data frames. The adaptive mode separation module is used to perform adaptive mode separation on the preprocessed data frame for non-stationary signals to obtain several intrinsic mode components; The time-frequency energy spectrum entropy feature extraction module is used to extract time-frequency energy spectrum entropy features from several intrinsic mode components to obtain energy spectrum entropy feature vectors; The nonlinear spatial mapping module is used to perform high-dimensional nonlinear spatial mapping on the energy spectrum entropy feature vector to obtain the deep feature representation tensor by utilizing a lightweight, deeply separable convolutional neural network that has been pre-deployed in the edge computing unit after model pruning and quantization. The diagnostic result generation module is used to calculate the fault probability and perform diagnostic inference on the deep feature representation tensor to obtain the final diagnostic result, which includes fault type, fault phase and confidence level.

[0008] Compared with existing technologies, this application provides an online fault diagnosis method and system for distribution networks based on edge computing. It effectively removes random noise and extracts the energy spectrum entropy reflecting the essential characteristics of the fault arc by performing adaptive mode separation on the original signal at the edge. Subsequently, it utilizes a lightweight, deeply separable convolutional neural network with pruned quantization to perform high-dimensional nonlinear spatial mapping, using deep feature representation to compensate for the insufficient adaptability of traditional manual judgment thresholds in dynamic environments. This approach not only enables accurate capture of weak, high-resistance grounding fault fingerprints at resource-constrained edge terminals, significantly reducing false alarms and missed alarms, but also ensures low-latency response during the diagnostic process through a lightweight model. Ultimately, it achieves accurate online identification and high-confidence inference of distribution network faults, providing reliable technical support for the safe and stable operation of smart distribution networks. Attached Figure Description

[0009] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0010] Figure 1 This is a flowchart of an online fault diagnosis method for distribution networks based on edge computing, according to an embodiment of this application. Figure 2 This is a schematic diagram of the data flow of the online fault diagnosis method for distribution networks based on edge computing according to an embodiment of this application; Figure 3 This is a flowchart of step S3 in the online fault diagnosis method for distribution networks based on edge computing according to an embodiment of this application; Figure 4 This is a block diagram of an online fault diagnosis system for distribution networks based on edge computing, according to an embodiment of this application. Detailed Implementation

[0011] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0012] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not specifically singular and may include plural forms. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0013] While this application makes various references to certain modules of the systems according to embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The modules described are merely illustrative, and different aspects of the systems and methods may use different modules.

[0014] Flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0015] The technical solution of this application proposes an online fault diagnosis method for power distribution networks based on edge computing. Figure 1 This is a flowchart of an online fault diagnosis method for distribution networks based on edge computing, according to an embodiment of this application. Figure 2 This is a system architecture diagram of an online fault diagnosis method for distribution networks based on edge computing, according to an embodiment of this application. Figure 1 and Figure 2 As shown, the online fault diagnosis method for distribution networks based on edge computing according to an embodiment of this application includes the following steps: S1, using edge acquisition terminals deployed at each monitoring node of the distribution network, the acquired raw sensor data stream is subjected to sliding window truncation, zero drift calibration, and normalization to obtain a preprocessed data frame; S2, the preprocessed data frame is subjected to adaptive mode separation of non-stationary signals to obtain several intrinsic mode components; S3, the time-frequency energy spectrum entropy features of the several intrinsic mode components are extracted to obtain an energy spectrum entropy feature vector; S4, using a lightweight deep separable convolutional neural network pre-deployed in the edge computing unit after model pruning and quantization, the energy spectrum entropy feature vector is mapped to a high-dimensional nonlinear space to obtain a deep feature representation tensor; S5, the deep feature representation tensor is subjected to fault probability calculation and diagnostic inference to obtain a final diagnostic result, the diagnostic result including fault type, fault phase, and confidence level.

[0016] Specifically, in step S1, the acquired raw sensor data stream is truncated using a sliding window, calibrated for zero drift, and normalized by edge acquisition terminals deployed at each monitoring node of the distribution network to obtain a preprocessed data frame. The raw sensor data stream includes three-phase current, zero-sequence current, and their corresponding sampling timestamps. Due to the significant electromagnetic interference in the distribution network environment, and the susceptibility of front-end hardware sensors to zero-point drift during long-term operation or under fluctuating ambient temperatures, non-characteristic DC offsets are superimposed on the acquired raw current data. Furthermore, the current amplitude varies greatly in magnitude under different voltage levels or load intensities; directly inputting this into a deep learning model would lead to difficulties in model convergence. Therefore, in the technical solution of this application, the acquired raw sensor data stream is truncated using a sliding window, calibrated for zero drift, and normalized to eliminate measurement errors introduced by the hardware, unify data dimensions, and impart spatial structural features to the time-series data through a sliding window mechanism, thereby ensuring the stability and robustness of the diagnostic algorithm.

[0017] Among them, edge acquisition terminals refer to intelligent hardware devices deployed at monitoring nodes such as power distribution lines, transformers or switch cabinets, which have the ability to capture data in real time and perform preliminary local computing power.

[0018] In specific implementation, firstly, in step 1.1, the original sensor data stream is framed to obtain a truncated data matrix. During this process, the edge acquisition terminal continuously acquires data streams containing three-phase current and zero-sequence current at a fixed sampling rate. To perform online, real-time diagnostics, a sliding window technique is used to extract data segments of fixed time lengths from the infinitely long continuous data stream, according to a preset window size and sliding step size, for analysis. During execution, a two-dimensional matrix is ​​extracted from the continuous time-series stream containing four-channel current data, with a set step size and window length. The number of rows in this matrix equals the number of sampling points contained within the window length, and the number of columns corresponds to the four current channels, thus obtaining the truncated data matrix, where the rows represent the time series and the columns represent the four data channels: three-phase current and zero-sequence current.

[0019] Next, in step 1.2, the truncated data matrix is ​​zero-drift calibrated to obtain a calibration data matrix. Since the sensor and signal conditioning circuitry may introduce DC offset during long-term operation, this step aims to eliminate the DC offset of each channel, bringing the signal centerline to zero. In this process, for the truncated data matrix obtained in step 1.1, the arithmetic mean of each column (i.e., each current channel) is calculated, and the corresponding average is subtracted from all elements in that column. During execution, the average of all elements in each column of the data matrix is ​​calculated, representing the estimated DC component of that channel within the current time window. Then, this average is subtracted from each element in that column, thereby removing the DC offset and obtaining a new data column with a mean of zero. This process is repeated for all four channels to obtain the zero-drift calibrated matrix, i.e., the calibration data matrix. This process ensures that the signal fluctuates around zero, eliminating the interference of the DC component on subsequent signal analysis (especially frequency domain analysis).

[0020] Furthermore, in step 1.3, the calibration data matrix is ​​scaled by channel-by-channel amplitude normalization to obtain a preprocessed data frame. It should be understood that in a distribution network, current amplitude can vary drastically due to factors such as load size and the distance to the fault point. To eliminate amplitude influence and allow the model to focus on the relative changes and patterns of the signal, amplitude normalization is required for each channel. During this process, for the calibration data matrix obtained in the previous step, the maximum absolute value of each column is calculated, and then each element of that column is divided by this maximum value. During execution, for each column of the calibrated matrix, the maximum absolute value of all elements in that column is found; this value represents the peak amplitude of the channel signal within the current time window. It is worth noting that to prevent calculation errors caused by a maximum value of zero, a very small positive number is usually added to the denominator as a protection factor. Then, each element of that column is divided by (peak amplitude plus the protection factor), thereby scaling all values ​​in that column to the range of [0, 1]. This operation is repeated for all four channels, eventually generating a matrix in which all element values ​​are constrained to the range of [0, 1]. This is the preprocessed data frame after complete preprocessing, which will be provided as input data to the subsequent S2 step (adaptive modality separation module) for further analysis.

[0021] Specifically, S2 involves performing adaptive mode separation on the preprocessed data frame to obtain several intrinsic mode components (IMCs). It should be understood that complex operating conditions such as high-resistance grounding faults in power distribution networks exhibit strong non-stationarity, with amplitude and frequency fluctuating drastically over time. Traditional Fourier transforms lack time localization capabilities, making it difficult to capture the transient characteristics of the signal. To achieve accurate fault feature extraction, the technical solution of this application decomposes the original signal into a series of narrowband signals using an adaptive mode separation algorithm. Each narrowband component reflects the local oscillation characteristics of the signal within a specific frequency band, thereby eliminating the interleaving interference of background noise and irrelevant load harmonics, providing a pure base component for accurate feature extraction downstream.

[0022] In specific implementation, firstly, in step 2.1, variational constraint modeling is performed on the preprocessed data frame based on the preset number of mode decompositions to obtain the augmented Lagrangian function model. Since distribution network fault signals often contain multiple physical components such as the fundamental frequency, higher harmonics, and attenuated DC components, these components are intertwined in the time and frequency domains. Traditional decomposition methods, such as Empirical Mode Decomposition (EMD), are easily affected by mode aliasing and are sensitive to noise. Through variational constraint modeling, each component to be decomposed can be set as a narrowband signal with a specific bandwidth. Utilizing the global search capability of variational optimization, the center frequency and envelope of each component can be extracted more fundamentally, thus laying the foundation for subsequently obtaining pure intrinsic mode components.

[0023] In this process, the Hilbert transform is first used to perform analytical operations on each component to obtain its spectrum. Then, complex exponential modulation is used to shift the spectrum of each component to its corresponding estimated center frequency. Subsequently, the system quantitatively describes the demodulated bandwidth by calculating the square of the second norm of the signal gradient operator. To ensure the fidelity of the decomposition, a quadratic penalty term is introduced into the logic to constrain the deviation between the sum of the components and the original signal, and a Lagrange multiplier term is introduced simultaneously to ensure the rigor of the constraint. The final combinational logic is to linearly weight and accumulate the sum of the demodulated bandwidth estimates of all components, the square of the reconstruction error, and the inner product of the Lagrange multiplier and the reconstruction error, thereby encapsulating them into a complete augmented Lagrange function target object.

[0024] Next, in step 2.2, the augmented Lagrange function model is mapped to the frequency domain using a Fourier transform. The frequency domain estimates of each modal component, along with the center frequency and Lagrange multipliers, are iteratively updated using the alternating direction multiplier method until the convergence error is less than a preset tolerance limit, at which point the iteration stops, thus obtaining the optimal mode matrix in the frequency domain. It should be understood that the augmented Lagrange model constructed in the previous step contains complex convolution terms and variational operations in the time domain, and direct solution has extremely high computational complexity and the risk of the curse of dimensionality. Mapping it to the frequency domain using a Fourier transform allows the energy functional in the time domain to be transformed into an arithmetic-algebraic form in the frequency domain using Passevar's theorem, thereby significantly reducing the computational difficulty. The Alternating Direction Multiplier Method (ADMM) is introduced to decompose the high-dimensional co-optimization problem into multiple low-dimensional single-variable subproblems. By repeatedly updating the modal components, center frequency, and multipliers, the algorithm can exhibit strong robustness and adaptability in noisy environments, and finally lock the optimal frequency domain component distribution, thereby solving the bottleneck of the difficulty in removing non-stationary signal characteristics from the distribution network.

[0025] In this process, firstly, all signal components and residual terms in the previously constructed augmented Lagrange function model are mapped to the frequency domain. Next, the frequency domain estimates of each modal component are updated based on the alternating direction multiplier method. Specifically, based on the Wiener filtering concept, the spectrum of the original signal is subtracted from the spectra of all other known components, and the residual bias provided by the Lagrange multipliers is added. Then, this difference is divided by a weighting factor related to the frequency offset. This logic ensures that each modal component has the highest energy gain near its center frequency, while components far from the center frequency are suppressed, thus achieving narrowband filtering. Furthermore, the power value of the modal component at each frequency point is used as a weight, multiplied by its corresponding angular frequency, and these weighted energies are integrated across the entire frequency band. Finally, this is divided by the component, which is proportionally adjusted according to the signal reconstruction error at the current stage. This global energy integration of the spectrum allows the center frequency to adaptively converge at the frequency points where signal components are most concentrated. Subsequently, specifically, the difference between the original signal spectrum and the sum of all extracted components is multiplied by a preset update step size and accumulated onto the existing multiplier vector, thereby dynamically strengthening the constraints of the reconstruction process. Finally, the changes of all variables in this round relative to the previous round are calculated, and the iteration is considered to have converged when the sum of the relative L2 norms of the changes is lower than a preset tolerance limit.

[0026] Furthermore, in step 2.3, the inverse Fourier transform is used to perform a time-domain inverse transform on each frequency-domain mode signal in the frequency-domain optimal mode matrix, and the real part is extracted to obtain several intrinsic mode components. Because in the preliminary processing of variational mode separation, in order to utilize the alternating direction multiplier method for efficient optimization and simplify the computational overhead of Wiener filtering, all mode updates and center frequency evolution are completed in the complex frequency domain. Although the frequency-domain optimal mode matrix obtained at this time locks the fault component in terms of energy distribution, it is a set of complex vectors with angular frequency as the independent variable, and cannot be directly used for subsequent time-frequency energy spectrum entropy feature extraction or deep neural network inference. Since the three-phase current and zero-sequence current in the actual operation of the distribution network are real-valued physical quantities that change with time, these spectral components are further restored to time-domain waveforms through a time-domain inverse transform, while simultaneously eliminating the analytic signal virtual terms introduced in the variational modeling process, thereby extracting intrinsic mode components consistent with the physical characteristics of the original signal.

[0027] In this process, firstly, after detecting that the variational iteration satisfies the convergence criterion, the system sequentially extracts each order of frequency-domain modal signal from the optimal frequency-domain modal matrix. For a specific frequency-domain modal component, the system first initiates the inverse fast Fourier transform operator to transform it from the frequency index space back to the discrete-time point space. Since this component is constructed as an analytic signal form after Hilbert transform during frequency-domain calculation, it contains complex imaginary part information. Therefore, the real part extraction operation is performed on the complex time-domain sequence output by the inverse Fourier transform, that is, the imaginary coordinate component is directly stripped, and only the real component is retained as the final time-domain result of this order. The system iterates through all component columns of the matrix, performing the above operation repeatedly, until all the preset number of time-domain eigenmode components are obtained. These components satisfy the energy conservation property on the time axis and maintain good analytic orthogonality between components.

[0028] Specifically, in step S3, time-frequency energy spectrum entropy features are extracted from several intrinsic mode components to obtain energy spectrum entropy feature vectors. It should be understood that when complex faults such as high-resistance grounding faults occur in a distribution network, the energy distribution of its intrinsic mode components undergoes significant and regular shifts, which are difficult to intuitively reflect in traditional time-domain amplitude analysis. By introducing information entropy theory, the disorder and uniformity of spectral energy distribution of the signal across different frequency bands can be quantitatively characterized, thereby effectively distinguishing energy topology changes caused by normal load fluctuations, transient impacts, and real faults. The resulting energy spectrum entropy features not only possess translation invariance but also amplify the spectral structure differences caused by weak faults, providing highly discriminative feature inputs for the nonlinear mapping of downstream deep learning models.

[0029] Figure 3 This is a flowchart of step S3 in the edge computing-based online fault diagnosis method for distribution networks according to an embodiment of this application. Figure 3 As shown, in the first embodiment of this application, step S3 includes: S3.1, performing discrete-time series energy distribution quantization on several intrinsic mode components to obtain a mode energy sequence; S3.2, performing energy normalization and probability mapping on the mode energy sequence to obtain a set of energy probability distributions; and S3.3, performing information entropy estimation on the set of energy probability distributions to obtain a global information entropy value, and vertically splicing and assembling the global information entropy value and the probability values ​​of each frequency band in the set of energy probability distributions to obtain an energy spectrum entropy feature vector.

[0030] Specifically, in step S3.1, discrete-time energy distribution quantization is performed on several intrinsic mode components to obtain a mode energy sequence. It should be understood that the intrinsic mode components after adaptive separation are still discrete-time sequences containing a large number of sampling points. The amplitudes of these components fluctuate drastically over time and cannot be directly used as quantitative classification feature inputs. To characterize the physical intensity of each frequency band component in the distribution network, the technical solution of this application extracts the overall activity intensity of each mode within a sliding window through energy distribution quantization. Specifically, by calculating the cumulative energy in the form of a mean square value, the amplitude abrupt changes caused by fault transient signals can be significantly amplified, while suppressing a small amount of random clutter, thus providing the most basic physical quantity measure for subsequently constructing the energy spectrum entropy reflecting the energy topology distribution.

[0031] In this process, firstly, all discrete sampling points of the intrinsic mode components of a specific order within the current preset sliding window length are traversed, and the amplitude of each sampling point is squared sequentially to ensure that the obtained energy index has positive definite physical characteristics and can sensitively reflect the instantaneous power of the signal. Secondly, after completing the squaring operation, the squared values ​​corresponding to all sampling points within the current sliding window are summed to obtain a single discrete energy value reflecting the overall activity level of that mode. The system repeats the above point-by-point squaring and summing calculation process for all preset number of intrinsic mode components, and finally outputs a mode energy sequence vector containing energy scalars of each order, as the intermediate data basis for subsequent probability mapping calculations.

[0032] Specifically, in step S3.2, the modal energy sequence is normalized and probabilistically mapped to obtain a set of energy probability distributions. It should be understood that the raw current amplitude collected from the distribution network is highly susceptible to the real-time load intensity of the lines, resulting in a difference of several times or even tens of times in the absolute energy values ​​between the fault occurrence time and the normal operation time. If the absolute energy value is directly used as the diagnostic input, the algorithm will struggle to maintain stable identification under dynamically changing operating conditions. By performing normalization and probabilistic mapping, energy sequences of different magnitudes can be uniformly mapped to a unit probability interval, allowing the diagnostic system to focus on analyzing the energy transfer trends and dispersion characteristics across frequency bands, thereby improving the sensitivity to capturing weak, high-resistance grounding fault signals.

[0033] This process begins by acquiring the modal energy sequence generated in the previous stage. This sequence contains the absolute energy values ​​of a predetermined number of intrinsic modal components. Next, iterates through all component scalars in the modal energy sequence, performing a full summation to obtain the total accumulated energy value of all modal components within the current sliding window. Then, while maintaining the correspondence between modal orders, each independent intrinsic modal energy value is used as the dividend, and the previously calculated global total accumulated energy value is used as the divisor, performing a digit-wise proportional division. Through this logic, each order of energy is transformed into a dimensionless probability percentage value between zero and one. The system iteratively executes the above logic until all orders of energy have been processed, ultimately encapsulating these probability scalars in their original order into a set of energy probability distributions, which serves as the standard data input for subsequent information entropy estimation.

[0034] Specifically, in step S3.3, information entropy estimation is performed on the energy probability distribution set to obtain a global information entropy value. This global information entropy value is then vertically concatenated with the probability values ​​of each frequency band in the energy probability distribution set to obtain an energy spectrum entropy feature vector. It should be understood that although the energy probability distribution obtained in the preceding steps has initially revealed the relative intensity of the signal in different frequency bands, this discrete probability description still lacks a quantitative assessment of the overall topological structure of the signal spectrum. Under normal operating conditions of the distribution network, energy is highly concentrated in the fundamental frequency band, exhibiting extremely high orderliness and extremely low information uncertainty. However, nonlinear processes such as high-resistance grounding faults cause energy to diffuse to multiple high-frequency components, significantly increasing the randomness and complexity of the spectrum distribution. By introducing information entropy estimation, a scalar index that highly summarizes the signal spectrum characteristics can be obtained. Concatenating this index with the original probability values ​​of each order not only preserves the local energy distribution details but also introduces global topological complexity information. This multi-dimensional feature combination can significantly enhance the discriminative power of fault signals in the nonlinear mapping space of the neural network, thereby improving the accuracy of diagnosis.

[0035] In this process, the system first obtains the data from the previous stage, which includes... The set of energy probability distributions for each element. Then, based on the Shannon entropy principle, logarithmic mapping and weighting are performed on each non-zero probability component in the set. By summing all modal probabilities and their corresponding negative logarithmic gains, the degree of disorder in the spectrum is quantified. This process can be expressed by the formula: in, The calculated global information entropy value, For the first The energy probability value of each modal component after topological coupling weighting This represents a base-2 logarithmic operation. Furthermore, using the data reconstruction operator, the calculated single scalar information entropy is... Place it at the beginning of the feature sequence, and then place the rest... The probability values ​​are arranged sequentially after the entropy value according to the center frequency of the corresponding mode, from low to high. This vertical concatenation method generates a dimension... The spectral entropy eigenvector. This process can be expressed by the formula: in, The output energy spectrum entropy eigenvector is... This represents the transpose operation. After the system completes the structured encapsulation of the vector, it outputs it to the subsequent neural network model for high-dimensional mapping.

[0036] It is worth noting that the mapping strategy adopted in the first embodiment implies a basic assumption: the contributions of all intrinsic mode components to fault detection are independent and equally weighted, and each frequency band is regarded as an information island, whose position in the probability space is determined only by the relative magnitude of its absolute energy value.

[0037] However, in the actual physical process of HIF in the distribution network, this independent and equal weight assumption is fundamentally contradictory to the real spectral dynamics mechanism of arc discharge. The arc in a high-resistivity ground fault has a nonlinear characteristic of intermittent extinguishing-reignition. When the arc extinguishes near the zero-crossing point of the AC current and breaks down again after the voltage recovers, this violent nonlinear jump process does not only trigger isolated energy pulses in a single frequency band, but also simultaneously excites strongly correlated energy co-transitions in multiple adjacent frequency bands. For example, the energy surge of the third harmonic component is often accompanied by the time-synchronous energy response of its adjacent fifth and seventh harmonic components. This adjacent frequency band energy co-transition mode constitutes the unique topological fingerprint of HIF that distinguishes it from normal load fluctuations, capacitor switching transients, and other short-circuit faults.

[0038] The simple proportional normalization in the first embodiment treats the energy of each frequency band as an isolated scalar unrelated to each other, completely discarding the core topological relationship information of the energy coupling gradient between adjacent frequency bands. This leads to two serious consequences: First, when an isolated energy anomaly occurs in a single frequency band due to random electromagnetic interference or nonlinear load harmonic injection, the probability value of that frequency band is indiscriminately amplified under equal-weight normalization due to the lack of a cross-verification mechanism for neighborhood coupling relationships, causing a shift in the downstream information entropy characteristics and thus triggering false alarms. Second, when the HIF arc does indeed trigger a coordinated energy mutation in multiple adjacent frequency bands, the equal-weight normalization mechanism evenly spreads these mutation energies across the entire probability distribution, diluting the topological mutation characteristics of cross-frequency band linkage, causing the downstream information entropy value to fail to fully reflect the structural differences of the fault signal and thus triggering missed alarms.

[0039] To address the aforementioned shortcomings, this application proposes a preferred example, which introduces a topology-aware adaptive weighting strategy based on the energy coupling gradient of adjacent frequency bands in step S3.2, thereby improving the mapping from equal-weight probability mapping to topology-coupled probability mapping through the following three steps.

[0040] Specifically, firstly, the energy coupling gradient between adjacent frequency bands is calculated to obtain the coupling gradient sequence. In the HIF (High-Intensity Interval) diagnosis scenario of a distribution network, the intermittent extinguishing-reignition process of the fault arc does not manifest as an isolated abrupt change in a single frequency band, but rather as a spatially continuous energy disturbance propagation between adjacent frequency bands. To capture the intensity of this energy difference between adjacent frequency bands (i.e., the topological coupling gradient) at a quantitative level, it is necessary to first calculate the energy drop between each modal component and its nearest neighbor modal component.

[0041] Specifically, for the modal energy sequence (Modal_Energy_Sequence) arranged from low to high center frequency The energy value of the modal, for the ... Each modal component is calculated to be related to its left neighbor (the first). (one) and right adjacent The absolute value of the energy difference between modal components, and the average of the differences in two directions, are used to quantify the energy gradient activity of the frequency band in the local spectral topology; for boundary modal components ( Only the gradient of its unique one-sided neighborhood is calculated; ultimately all The coupled gradient values ​​of each modal component are assembled in the original frequency order to generate a coupled gradient sequence. This process can be expressed by the formula: in, This represents the energy coupling gradient value of the adjacent frequency bands of the k-th modal component, which forms a coupling gradient sequence as elements; The absolute energy value of the k-th intrinsic mode component extracted from the modal energy sequence; It represents the absolute energy value of the left adjacent (low-frequency side) modal component of the k-th modal component; is the absolute energy value of the right adjacent (high-frequency side) modal component of the k-th modal component; K is the total number of intrinsic modal components.

[0042] This gradient characterizes the energy difference between each frequency band and its nearest neighboring bands. When a multi-band coordinated mutation is induced by an HIF arc, multiple adjacent frequency bands involved will exhibit high gradient activity, while a single-band anomaly caused by isolated noise will only generate a gradient within that frequency band itself, with extremely low gradients in its neighborhood. This difference is the key topological clue for distinguishing between real faults and random interference.

[0043] It can be understood that by performing gradient quantization, the local topological relationships that were originally hidden between the energy values ​​of adjacent frequency bands are explicitly extracted as independent data objects, providing a quantitative topological coupling metric basis for subsequent adaptive weight synthesis, thus enabling the system to perceive the spatial correlation structure between frequency bands.

[0044] Next, the coupled gradient sequence and modal energy sequence are synthesized using topological coupling adaptive weights to obtain a weighted energy sequence. It should be understood that simply obtaining the original values ​​of the coupled gradients is insufficient to directly improve the quality of the probability mapping. In the field environment of the distribution network, the absolute energy levels of each frequency band differ significantly under different operating conditions. A nonlinear modulation mechanism is needed to transform gradient information into adaptive weights with selective amplification capabilities, enabling the high-coupling frequency band (HIF arc co-transition region) to obtain significant gains in energy weighting, while the low-coupling frequency band (isolated noise region) is effectively suppressed.

[0045] Specifically, the coupling gradient sequence is first normalized by dividing each gradient value by the sum of all gradient values ​​to obtain the normalized gradient weights, which reflect the relative activity of each frequency band in the local topology. This process can be expressed by the formula: in, For the first Normalized gradient weights for each modal component; The first one extracted from the coupled gradient sequence The original coupling gradient values ​​of each modal component; This represents the original gradient value of the i-th component in the coupled gradient sequence; The total number of intrinsic modal components; Use the local accumulation index variable for global summation.

[0046] Subsequently, the topological coupling modulation index is introduced. ( This involves constructing nonlinear topologically coupled adaptive weights by applying an exponential amplification operation to the normalized gradient weights. This process can be expressed by the following formula: in, For the first The final topologically coupled adaptive weights obtained for each modal component satisfy the normalization reduction ; Let be the normalized gradient weight of the i-th modal component; The topological coupling modulation index is a preset positive real hyperparameter used to control the intensity of nonlinear amplification. It degenerates into linear weights when Superlinear selective amplification is achieved in real time; The total number of intrinsic modal components; Use the local accumulation index variable for global summation.

[0047] Finally, the topology-coupled adaptive weights are multiplied element-wise with the original energy values ​​of the corresponding components in the modal energy sequence to generate a weighted energy sequence that incorporates the inter-band topological relationships. This process can be expressed by the following formula: in, For the first The energy values ​​of each modal component after topological coupling weighting are used as elements to form a weighted energy sequence; For the first Topological coupling adaptive weights for each modal component; The first one extracted from the modal energy sequence The original energy value of each component.

[0048] This step utilizes an exponential nonlinear modulation function to amplify the weight distribution with a winner-takes-all effect. In the HIF scenario, the multi-band cooperative abrupt change region caused by the intermittent arcing gains significant weighting, while isolated noise bands are adaptively suppressed due to a lack of neighborhood coupling support. Specifically, when When set to a value greater than 1, the weights of frequency bands with high topological coupling are superlinearly amplified, while the weights of frequency bands with low topological coupling are accelerated attenuation. This nonlinear differentiation characteristic is highly consistent with the multi-band synergistic mutation physical mechanism of HIF arc.

[0049] In this way, by combining exponential nonlinear modulation with element-wise weighted multiplication, the topological coupling information between frequency bands is injected into the energy value itself, so that the subsequent probability normalization is no longer an indiscriminate equal weighting operation, but a structured transformation with fault topology sensitivity, which fundamentally solves the defect that the HIF cooperative mutation characteristics are flattened and diluted under the original equal weighting mechanism.

[0050] Then, the weighted energy sequence is coupled and weighted energy probability normalized to obtain a set of energy probability distributions. It should be understood that after the first two steps, the elements in the weighted energy sequence have incorporated the topological coupling information between frequency bands, but their numerical scale is still in the absolute energy dimension, which does not yet meet the probability distribution form required by the subsequent information entropy calculation (i.e., the sum of each element is strictly equal to 1). Therefore, the final probability normalization needs to be performed.

[0051] Specifically, traversing all of the weighted energy sequences The weighted global total energy after topological coupling modulation is obtained by summing the weighted energy values. Then, each weighted energy value is divided by the total weighted global energy to perform proportional normalization, outputting a set of energy probability distributions that satisfy the probability axiom constraints. This process can be expressed by the formula: in, For the first The energy probability values ​​of each modal component after topological coupling weighting are used as elements to form the set of output energy probability distributions; The first one extracted from the weighted energy sequence Each weighted energy value; For the weighted energy sequence, the first Each weighted energy value; The total number of intrinsic modal components; Use the local accumulation index variable for global summation.

[0052] Thus, the probability distribution after coupling and weighting is no longer the original equal-weighted energy proportion, but a structured probability distribution embedded with the topological coupling fingerprint between frequency bands. In the frequency band where HIF arcs induce cooperative mutations, the probability value is significantly increased; in the isolated noise frequency band, the probability value is effectively compressed. This allows the information entropy calculation in the downstream sub-step S3.3 to more sensitively capture the topological structural features of the fault signal, rather than merely reflecting the coarse-grained dispersion of spectral energy.

[0053] This step completes the final mapping from the weighted energy space to the probability space, ensuring that the output energy probability distribution set is completely consistent with the output of the first embodiment in terms of naming and interface format, guaranteeing seamless connection with the information entropy calculation module of the downstream sub-step S3.3, and enabling the information entropy operator to fully perceive and utilize the topological coupling structure information encoded in the probability distribution.

[0054] As described above, by introducing a topology-aware adaptive weighting strategy based on the energy coupling gradient of adjacent frequency bands to optimize the energy normalization and probability mapping mechanism of the original sub-step 3.2, the following technical effects and objectives are achieved: In terms of reducing false alarm rate, since the improved mechanism applies differentiated weights to each frequency band through topological coupling gradient, when a single frequency band has an isolated energy anomaly due to random electromagnetic interference or nonlinear load harmonic injection, the frequency band lacks neighborhood gradient coupling support, and its adaptive weight is significantly compressed under exponential modulation, thereby effectively suppressing the phenomenon of falsely high probability of isolated noise frequency bands and reducing the false fault alarms caused by it.

[0055] In terms of reducing the false alarm rate, when the arc of a high-resistivity ground fault does indeed cause a coordinated energy mutation in multiple adjacent frequency bands, these frequency bands all exhibit high activity in gradient calculation and obtain superlinear weight gain under exponential modulation. The topological fingerprint of the coordinated mutation is selectively amplified in the probability distribution rather than uniformly diluted, which significantly improves the sensitivity of the downstream information entropy feature to the real fault signal and effectively reduces the risk of false alarm.

[0056] In terms of improving the efficiency of edge computing resource utilization, the three sub-steps of the improved mechanism (gradient calculation, weight synthesis, and probability normalization) are all lightweight vector operations, with the introduced additional computational overhead being only O(n log n). )level( The modal component count (usually between 4 and 8) is negligible compared to the computational cost of subsequent deep neural network inference, and does not impose an additional burden on the real-time performance of edge devices. It achieves a fundamental improvement in fault identification accuracy with almost no increase in edge computing power consumption.

[0057] Specifically, in S4, a lightweight, deeply separable convolutional neural network, pre-deployed in the edge computing unit and pruned and quantized, is used to perform high-dimensional nonlinear spatial mapping on the energy spectral entropy feature vector to obtain a deep feature representation tensor. Distribution network fault signals (especially high-resistance grounding faults) and normal fluctuations or load disturbances exhibit significant spatial overlap in the original numerical domain, making it difficult to achieve high-confidence classification using simple linear thresholds or predefined rule bases. Deep convolutional neural networks can automatically uncover complex correlations between data through multi-layer nonlinear transformations. However, traditional convolutional neural network architectures are large, parameter-redundant, and computationally intensive, making efficient operation difficult on distribution network edge terminals with limited computing resources and storage constraints. By introducing model pruning to remove redundant connections and using quantization techniques to compress floating-point numbers into fixed-point integers, combined with the reconstruction of the computational flow using deeply separable convolution, the inference power consumption and memory footprint of the neural network can be significantly reduced, ensuring real-time diagnostic response at the edge without sacrificing diagnostic accuracy. It is worth mentioning that the lightweight deep separable convolutional neural network is a neural network architecture that decomposes standard convolution into single-channel spatial feature learning (deep convolution) and cross-channel feature combination (pointwise convolution), which can greatly reduce the amount of computation.

[0058] In specific implementation, firstly, in step 4.1, the energy spectrum entropy feature vector is reconstructed in dimension and subjected to initial convolution mapping to obtain an initial feature channel expansion tensor. Since the energy spectrum entropy feature vector generated in the preceding steps is mathematically a one-dimensional numerical sequence, and the hierarchical feature extraction logic of a convolutional neural network (CNN) is designed based on the topological structure of a multi-channel tensor, the original feature vector lacks depth (i.e., channel dimension) and cannot directly participate in the spatial filtering operation of the convolution operator. Dimension reconstruction assigns spatial coordinate attributes to the original features; while initial convolution mapping projects low-dimensional energy spectrum information into a higher-dimensional feature stream, playing a role in feature augmentation. This allows subsequent layers to analyze the nonlinear correlation patterns of fault signals from more observation dimensions (channels), thus providing a more expressive feature foundation for edge-side classification decisions.

[0059] In this process, firstly, the energy spectral entropy feature vector is mapped to a three-dimensional coordinate space adapted to the input specification of the convolutional network, thus adding a depth dimension index of one value to the data to form the input tensor. Next, feature projection is performed on the input tensor using convolutional layers pre-deployed at the edges. For each preset output channel, the system performs element-wise multiplication with the input tensor using the channel-specific convolutional kernel weight matrix, and then performs a global summation of all multiplication results. After completing the element-wise multiplication and summation, the system adds a preset bias adjustment value to achieve linear translation. Then, a specific nonlinear function (such as a rectified linear unit) is used to perform a nonlinear transformation on the value, thereby preserving key features and filtering redundant components. By traversing and performing the above multiplication, summation, and activation calculations for all output channels, the system finally assembles and generates a tensor with significantly expanded channel dimensions, i.e., the initial feature channel expanded tensor.

[0060] Next, in step 4.2, spatial filtering is performed independently on each channel of the initial feature channel expansion tensor using a depthwise convolution kernel to obtain the depthwise convolution result. Then, a pointwise convolution kernel of size one is used to perform a cross-channel linear combination mapping on the depthwise convolution result to obtain a multi-scale hidden layer feature tensor. In this process, firstly, an initial feature expansion tensor containing multiple channels is obtained from the previous submodule. Secondly, each channel is treated as an independent information dimension for spatial filtering. Specifically, a dedicated set of convolution weights is assigned to each input feature channel. During the sliding filtering process, the system only performs element-wise multiplication operations on the feature value within the current specific channel and the corresponding convolution kernel weight, and performs arithmetic accumulation on the predicted product values ​​within the local window. After completing the sliding multiplication accumulation at all spatial positions of the channel and adding a specific offset adjustment value, the system obtains the intermediate output of that channel. The system assembles and generates the depthwise convolution result by synchronously performing the above non-interfering filtering operations on all input channels. Then, a cross-channel linear combination mapping is performed using a pointwise convolution kernel. Specifically, feature channel recombination and information fusion are achieved through linear weighting. During execution, for the intermediate tensor generated in the previous step, the system extracts the feature values ​​of all input channels at the same location at each discrete spatial coordinate point and performs multiplication operations with a set of preset output mapping weights. Subsequently, the system performs global summation of the weighted product results of all channels at that location and adds a preset output bias value. This process is repeated for all preset output channels, thereby aggregating local features scattered across different frequency bands into a new semantic space and outputting the final multi-scale hidden layer feature tensor.

[0061] Furthermore, in step 4.3, the multi-scale hidden layer feature tensor is compressed using nonlinear activation and global average pooling to obtain the depth feature representation tensor. It should be understood that the hidden layer feature tensor output by the preceding depthwise separable convolutional layer is essentially a set of locally linear combinations extracted by multiple convolutional kernels, and its data dimension still carries significant spatial redundancy. To enable the diagnostic model to accurately fit the complex nonlinear transient processes of the power system, a nonlinear transformation capability is introduced through a nonlinear activation function, thereby endowing the network with the ability to identify weak arc waveform distortions. Simultaneously, to achieve centralized aggregation of feature information within the entire sliding window and effectively reduce the computational overhead of subsequent classification layers, global average pooling logic is used to compress features distributed at different locations. This processing method significantly enhances the model's robustness to signal translation or local distortion caused by sampling bias, transforming the multidimensional hidden layer state into a fixed-length and highly abstract feature vector, driving the diagnostic system to evolve from local, fragmented signal perception to global fault attribute representation.

[0062] In this process, firstly, for each discrete data point in the multi-scale hidden layer feature tensor, it is input as an independent variable into the activation algorithm. The value is transformed in real time according to a preset nonlinear mapping rule (such as rectification transformation or hardware-friendly switching mapping). Through this point-to-point nonlinear transformation, nonlinear descriptive capability is injected while preserving the feature distribution topology. Next, each feature channel in the tensor is traversed sequentially. For a specific channel, the system first captures all neuron values ​​contained in that channel and performs a global summation operation. Subsequently, the system calculates the total number of elements in that channel and divides the aforementioned summation by this total number of elements. Through this division logic, the feature layer, originally having spatial width or length, is compressed into a single scalar mean, reflecting the feature activity intensity of that channel globally. After processing all channels in a loop, the system reassembles these mean scalars to output the deep feature representation tensor.

[0063] Specifically, in step S5, fault probability calculation and diagnostic reasoning are performed on the deep feature representation tensor to obtain the final diagnostic result, which includes fault type, fault phase, and confidence level. Due to the extremely complex environment of the distribution network, even with deep features acquired, features within a single time window may still be affected by instantaneous strong interference (such as random electromagnetic pulses generated by high-power equipment switching), leading to false alarms. Furthermore, simple nonlinear mapping can only provide the classification tendency at the current moment and cannot provide evidence reflecting the persistence and stability of physical faults. Therefore, in the technical solution of this application, by performing quantitative calculation of fault probability and introducing a dynamic adaptive threshold and time-series decision queue based on a sliding window, a multi-window comparison method is used to confirm weak faults, thereby filtering out the true fault fingerprint in a noisy background and outputting a comprehensive diagnostic report including fault type, phase, and reliability score.

[0064] In specific implementation, firstly, in step 5.1, the deep feature representation tensor is linearly combined and mapped using a fully connected weight matrix to obtain logistic regression values ​​for each category. Then, the logistic regression values ​​are probabilistically normalized using a normalized exponential function to obtain the fault probability distribution vector. Although the deep feature representation tensor generated in the preceding steps contains extremely rich fault waveform semantics, its representation is a high-dimensional, highly abstract numerical sequence, which cannot be directly understood by power dispatchers or automated operation and maintenance terminals. Therefore, a fully connected layer is used to globally weight and aggregate all candidate feature dimensions to deeply explore the coupling relationships between different feature dimensions. Simultaneously, since the original scores obtained after linear mapping have arbitrary real-valued distributions, they cannot intuitively reflect the degree of certainty in the diagnostic results. Therefore, a normalized exponential function is introduced to map the output to a unit probability space, quantitatively giving the probability of occurrence for each preset fault type, thus providing rigorous mathematical support for subsequent confidence-based fault confirmation.

[0065] In this process, firstly, the system uses a pre-trained weight matrix deployed at the edge to perform matrix multiplication with the input deep feature vector to achieve weighted aggregation of global features. After completing the multiplication, the system then performs element-wise biased addition on the resulting vector, using a pre-defined bias vector to achieve a linear translation from the feature space to the decision space, ultimately generating logistic regression value vectors for each fault category. Next, for each category element in the logistic regression value vector, an exponentiation operation with the natural constant as the base is performed to obtain an exponential score reflecting the relative strength of that category. Subsequently, the system calculates the global arithmetic sum of the exponential scores for all categories. Finally, the exponential score of each independent category is used as the dividend, and the previously calculated global sum is used as the divisor for division. Through this logic, the system maps the original raw scores to discrete probability values ​​between zero and one, with a sum of one, thus obtaining a complete fault probability distribution vector.

[0066] Next, in step 5.2, the fault probability distribution vector is dynamically and adaptively updated with a threshold and compared with a single window to obtain the single-window discrimination state. It should be understood that the background noise in the actual operating environment of the distribution network is not constant. Random load switching and the dynamic evolution of the harmonic environment can cause certain statistical fluctuations in the fault probability value under normal conditions. If a fixed hard threshold is used as the judgment criterion, it is easy to cause frequent false alarms under strong noise conditions, while it may cause missed alarms under ultra-light load conditions. By performing dynamic adaptive threshold updates, the judgment criterion can be adaptively scaled in real time according to the intensity of the background noise, thereby significantly reducing non-fault false judgments caused by random environmental disturbances while maintaining recognition sensitivity.

[0067] In this process, firstly, the system continuously maintains a buffer containing the highest fault probability values ​​under a preset number of fault-free conditions. By calculating the arithmetic mean of these historical statistical probability values ​​and adding the product of a preset sensitivity coefficient and the standard deviation of that sample group, the current judgment threshold value is dynamically synthesized. Next, the logic of comparing the highest fault probability in the fault probability distribution vector with the dynamically adaptive threshold to generate an out-of-limit Boolean flag is applied. Specifically, the maximum value in the current probability distribution vector is extracted and compared with the dynamically generated threshold. If the maximum value exceeds the threshold, the output flag is set to a true Boolean value representing an out-of-limit condition; otherwise, it is set to a false value. Then, using a data encapsulation operator, the Boolean flag generated at that moment, the corresponding highest probability fault label name, and the specific probability value are ordered and compressed to generate a single-window discrimination object.

[0068] Furthermore, in step 5.3, by caching multiple consecutive single-window discrimination states in the first-in-first-out (FIFO) timing decision queue, a logical AND operation is performed on all out-of-limit Boolean flags in the queue to confirm whether the fault continues to occur. When a preset number of consecutive sliding windows are all in out-of-limit states, the fault is confirmed, and the arithmetic mean of the probabilities of candidate faults in the queue is calculated as the confidence level. The final diagnostic result, including the fault type, fault phase, and confidence level, is output. It should be understood that the actual operating environment of the distribution network is extremely harsh. Instantaneous electromagnetic interference frequently occurs on the lines due to high-power load switching, transformer inrush current, or capacitor bank switching. These random interferences may exhibit fingerprint characteristics highly similar to real weak faults within a very short time window, thus inducing isolated false alarms in the neural network. Since physical power system faults usually have significant persistent characteristics, by introducing a timing decision queue and performing continuity verification, occasional disturbances without temporal regularity can be effectively filtered out, ensuring the physical credibility of diagnostic commands from an algorithmic perspective.

[0069] In this process, after obtaining the single-window discrimination state at the current sampling time, it is first pushed to the end of the first-in-first-out timing decision queue, and the oldest state record is popped. Then, the out-of-limit Boolean flags of all windows cached in the decision queue are extracted, and a logical AND operation is performed on these flags. Only when all flags of every sampling window in the queue are true values ​​representing out-of-limit events is the system determined that a fault has officially occurred; otherwise, it is determined to be a transient disturbance. Next, the system traverses each storage unit in the decision queue, retrieves the highest candidate fault probability value recorded for each window, and performs a global summation operation on these probability values. Then, the system divides this summation by the total length of the decision queue. Through this averaging logic, the system obtains a reliability index after time smoothing. Finally, the previously identified specific fault labels, phase information, and the aforementioned calculated average confidence scores are assembled in an orderly manner to generate the final diagnostic result.

[0070] In summary, the edge computing-based online fault diagnosis method for distribution networks according to the embodiments of this application is explained. It effectively removes random noise and extracts the energy spectrum entropy reflecting the essential characteristics of the fault arc by performing adaptive mode separation on the original signal at the edge. Subsequently, a lightweight, deeply separable convolutional neural network with pruned quantization is used to perform high-dimensional nonlinear spatial mapping, using deep feature representation to compensate for the insufficient adaptability of traditional manual judgment thresholds in dynamic environments. In this way, not only can the fingerprint of weak, high-resistance grounding faults be accurately captured at resource-constrained edge terminals, significantly reducing false alarm and false negative rates, but the lightweight model also ensures low-latency response in the diagnostic process. Ultimately, it achieves accurate online identification and high-confidence inference of distribution network faults, providing reliable technical support for the safe and stable operation of smart distribution networks.

[0071] Furthermore, an online fault diagnosis system for power distribution networks based on edge computing is also provided.

[0072] Figure 4 This is a block diagram of an online fault diagnosis system for distribution networks based on edge computing, according to an embodiment of this application. Figure 4As shown, the edge computing-based online fault diagnosis system 300 for power distribution networks according to an embodiment of this application includes: a data preprocessing module 310, used to perform sliding window truncation, zero-drift calibration, and normalization on the acquired raw sensor data stream through edge acquisition terminals deployed at each monitoring node of the power distribution network to obtain a preprocessed data frame; an adaptive mode separation module 320, used to perform adaptive mode separation on the preprocessed data frame for non-stationary signals to obtain several intrinsic mode components; a time-frequency energy spectrum entropy feature extraction module 330, used to extract time-frequency energy spectrum entropy features from several intrinsic mode components to obtain an energy spectrum entropy feature vector; a nonlinear space mapping module 340, used to perform high-dimensional nonlinear space mapping on the energy spectrum entropy feature vector using a lightweight, deeply separable convolutional neural network that has been pre-deployed in the edge computing unit after model pruning and quantization to obtain a deep feature representation tensor; and a diagnosis result generation module 350, used to perform fault probability calculation and diagnostic inference on the deep feature representation tensor to obtain a final diagnosis result, the diagnosis result including fault type, fault phase, and confidence level.

[0073] As described above, the edge computing-based online fault diagnosis system 300 for distribution networks according to embodiments of this application can be implemented in various wireless terminals, such as servers with edge computing-based online fault diagnosis algorithms for distribution networks. In one possible implementation, the edge computing-based online fault diagnosis system 300 for distribution networks according to embodiments of this application can be integrated into the wireless terminal as a software module and / or a hardware module. For example, the edge computing-based online fault diagnosis system 300 for distribution networks can be a software module in the operating system of the wireless terminal, or it can be an application developed for the wireless terminal; of course, the edge computing-based online fault diagnosis system 300 for distribution networks can also be one of many hardware modules of the wireless terminal.

[0074] Alternatively, in another example, the edge computing-based online fault diagnosis system 300 for power distribution networks and the wireless terminal can also be separate devices, and the edge computing-based online fault diagnosis system 300 for power distribution networks can be connected to the wireless terminal via wired and / or wireless networks, and transmit interactive information in accordance with an agreed data format.

[0075] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. An edge-computing-based power distribution network fault online diagnosis method, characterized in that, include: S1: By deploying edge acquisition terminals at each monitoring node of the power distribution network, the acquired raw sensor data stream is truncated by sliding window, zero-drift calibration and normalization to obtain preprocessed data frames; S2: Perform adaptive mode separation on the preprocessed data frame for non-stationary signals to obtain several intrinsic mode components; S3: Extract time-frequency energy spectrum entropy features from several intrinsic mode components to obtain energy spectrum entropy feature vectors; S4: Using a lightweight, deeply separable convolutional neural network that has been pre-deployed in an edge computing unit and pruned and quantized, the energy spectrum entropy feature vector is mapped to a high-dimensional nonlinear space to obtain a deep feature representation tensor. S5: Perform fault probability calculation and diagnostic reasoning on the deep feature representation tensor to obtain the final diagnostic result, which includes fault type, fault phase and confidence level.

2. The edge-computing-based power distribution network fault online diagnosis method according to claim 1, characterized in that, The raw sensor data stream contains three-phase current, zero-sequence current and their corresponding sampling timestamps.

3. The online fault diagnosis method for distribution networks based on edge computing according to claim 2, characterized in that, Step S1 includes: Step 1.1: Frame segmentation and truncation of the original sensor data stream to obtain a truncated data matrix; Step 1.2: Perform zero-drift calibration on the truncated data matrix to obtain the calibration data matrix; Step 1.3: Perform channel-by-channel amplitude normalization scaling on the calibration data matrix to obtain the preprocessed data frame.

4. The online fault diagnosis method for distribution networks based on edge computing according to claim 1, characterized in that, Step S2 includes: Step 2.1: Based on the preset number of mode decompositions, perform variational constraint modeling on the preprocessed data frame to obtain the augmented Lagrangian function model; Step 2.2: Map the augmented Lagrange function model to the frequency domain through Fourier transform. Based on the alternating direction multiplier method, iteratively update the frequency domain estimates of each modal component, the center frequency, and the Lagrange multipliers until the convergence error is less than the preset tolerance limit, and then stop the iteration to obtain the optimal mode matrix in the frequency domain. Step 2.3: Perform inverse time-domain transformation on each frequency-domain mode signal in the optimal frequency-domain mode matrix using inverse Fourier transform and extract the real part to obtain several eigenmode components.

5. The online fault diagnosis method for distribution networks based on edge computing according to claim 1, characterized in that, Step S3 includes: Step 3.1: Perform discrete-time series energy distribution quantization on several intrinsic mode components to obtain the mode energy sequence; Step 3.2: Perform energy normalization and probability mapping on the modal energy sequence to obtain a set of energy probability distributions; Step 3.3: Estimate the information entropy of the energy probability distribution set to obtain the global information entropy value. Then, vertically splice and assemble the global information entropy value with the probability values ​​of each frequency band in the energy probability distribution set to obtain the energy spectrum entropy feature vector.

6. The online fault diagnosis method for distribution networks based on edge computing according to claim 1, characterized in that, Step S4 includes: Step 4.1: Perform dimension reconstruction and initial convolution mapping on the energy spectral entropy feature vector to obtain the initial feature channel expansion tensor; Step 4.2: Spatially filter each channel of the initial feature channel expansion tensor independently using a depth convolution kernel to obtain the depth convolution result. Then, use a pointwise convolution kernel of size one to perform a cross-channel linear combination mapping on the depth convolution result to obtain the multi-scale hidden layer feature tensor. Step 4.3: Perform nonlinear activation and global average pooling compression on the multi-scale hidden layer feature tensor to obtain the depth feature representation tensor.

7. The online fault diagnosis method for distribution networks based on edge computing according to claim 1, characterized in that, Step S5 includes: Step 5.1: Linearly combine and map the deep feature representation tensor through the fully connected weight matrix to obtain the logistic regression values ​​for each category. Perform probability normalization operation on the logistic regression values ​​based on the normalized exponential function to obtain the fault probability distribution vector. Step 5.2: Perform dynamic adaptive threshold update and single-window comparison on the fault probability distribution vector to obtain the single-window discrimination state; Step 5.3: Cache multiple consecutive single-window discrimination states in the first-in-first-out timing decision queue, perform a logical AND operation on all out-of-limit Boolean flags in the queue to confirm whether the fault continues to occur. When a preset number of consecutive sliding windows are all in out-of-limit states, the fault is confirmed and the arithmetic mean of the probabilities of candidate faults in the queue is calculated as the confidence level. The final diagnosis result containing the fault type, fault phase and confidence level is output.

8. The online fault diagnosis method for distribution networks based on edge computing according to claim 7, characterized in that, Step 5.2 includes: The dynamic adaptive threshold is calculated based on the probability fluctuation variance of background noise within the past fault-free history window. The highest failure probability in the failure probability distribution vector is numerically compared with the dynamic adaptive threshold to generate an out-of-limit Boolean flag. The out-of-limit Boolean flag, along with the candidate fault type and candidate fault probability, are encapsulated in a structured manner to obtain the single-window discrimination state.

9. The online fault diagnosis method for distribution networks based on edge computing according to claim 5, characterized in that, Step 3.2 includes: The coupled gradient sequence is obtained by calculating the energy coupling gradient between adjacent frequency bands of the modal energy sequence. A topologically coupled adaptive weighted synthesis of coupled gradient sequences and modal energy sequences is performed to obtain a weighted energy sequence. The weighted energy sequence is coupled with weighted energy probability normalization to obtain a set of energy probability distributions.

10. An online fault diagnosis system for power distribution networks based on edge computing, characterized in that, include: The data preprocessing module is used to perform sliding window truncation, zero drift calibration and normalization on the acquired raw sensor data stream through edge acquisition terminals deployed at each monitoring node of the power distribution network to obtain preprocessed data frames. The adaptive mode separation module is used to perform adaptive mode separation on the preprocessed data frame for non-stationary signals to obtain several intrinsic mode components; The time-frequency energy spectrum entropy feature extraction module is used to extract time-frequency energy spectrum entropy features from several intrinsic mode components to obtain energy spectrum entropy feature vectors; The nonlinear spatial mapping module is used to perform high-dimensional nonlinear spatial mapping on the energy spectrum entropy feature vector to obtain the deep feature representation tensor by utilizing a lightweight, deeply separable convolutional neural network that has been pre-deployed in the edge computing unit after model pruning and quantization. The diagnostic result generation module is used to calculate the fault probability and perform diagnostic inference on the deep feature representation tensor to obtain the final diagnostic result, which includes fault type, fault phase and confidence level.

Images