Intelligent ring main unit fault real-time diagnosis method and system

Abstract

The invention relates to the technical field of data processing and state monitoring, and discloses an intelligent ring main unit fault real-time diagnosis method and system, and the method comprises the steps: constructing a fault feature over-complete dictionary; collecting and preprocessing high-frequency operation data streams of the ring main unit in real time; carrying out online sparse decomposition and manifold mapping based on the dictionary, and converting a high-dimensional signal into a low-dimensional sparse coefficient and manifold features; and a lightweight classification model is used for rapid reasoning, and fault early warning and data uploading are triggered when abnormity is detected. The system comprises a data acquisition module, a preprocessing unit, an edge computing engine, an intelligent diagnosis module and a communication gateway. According to the method, depth dimension reduction and feature enhancement are realized through a manifold sparse coding technology, the bandwidth pressure and the calculation overhead are reduced on the premise of ensuring the fault feature integrity, and the real-time performance and the accuracy of diagnosis are improved.

Description

Technical Field

[0001] This invention belongs to the field of data processing and status monitoring, and specifically relates to a method and system for real-time fault diagnosis of intelligent ring main unit. Background Technology

[0002] With the accelerated development of smart distribution networks, smart ring main units, as core nodes and key components of the power distribution system, are directly related to the safety and reliability of the power grid as a whole through real-time monitoring of their operating status and intelligent operation and maintenance. The ring main unit integrates circuit breakers, load switches, transformers, and various environmental monitoring sensors. The stable operation of these components is an important foundation for ensuring continuous power supply and efficient dispatch of the power distribution network.

[0003] Fault diagnosis technology is a core means to improve the intelligence level of ring main unit operation and maintenance. By collecting and logically analyzing the electrical parameters, environmental parameters and high-frequency characteristic signals generated by the ring main unit during operation, it aims to achieve early detection and accurate location of potential defects in the equipment, effectively reduce the risk of unplanned power outages and extend the service life of power distribution equipment.

[0004] Traditional ring main unit (RMU) fault diagnosis technologies suffer from several drawbacks: First, with the continuous increase in sampling frequency, the data volume generated by high-frequency partial discharge signals and vibration waveforms grows exponentially, leading to immense pressure on data acquisition and storage. Second, there is a conflict between the real-time nature of data processing and transmission bandwidth. Uploading all raw data to the cloud results in severe communication delays and bandwidth consumption, while simple frequency reduction or mean compression at the edge loses key waveform details reflecting the essence of the fault, significantly reducing diagnostic accuracy. Third, edge computing resources are relatively limited, lacking efficient algorithms capable of simultaneously achieving dimensionality reduction of massive data and enhancement of complex features, making it difficult to complete deep feature mining under limited computing power. Finally, the system's ability to represent nonlinear and non-stationary fault signals is insufficient, easily generating false alarms or missed alarms in complex operating environments, failing to meet the requirements of smart grids for millisecond-level fault response and accurate judgment. These problems collectively limit the application effectiveness of smart RMU fault diagnosis systems in actual power production environments. Summary of the Invention

[0005] The purpose of this invention is to provide a real-time fault diagnosis method and system for intelligent ring main units, which can effectively solve the problems mentioned in the background. Addressing the core contradiction between the explosive data volume and limited transmission bandwidth generated by intelligent ring main units in high-frequency sampling environments, and the challenge of achieving both real-time performance and detailed feature preservation in edge-side diagnostics due to limited computing resources, this invention introduces manifold sparse coding technology to achieve deep dimensionality reduction and lightweight real-time diagnosis of massive high-frequency data while ensuring the integrity of fault features.

[0006] To achieve the above objectives, this invention proposes a real-time fault diagnosis method for intelligent ring main units, comprising the following steps: Step S1, constructing an overcomplete fault feature dictionary to store basis functions that can sparsely represent various typical fault features; Step S2, acquiring high-frequency operating data streams of the ring main unit in real time, and preprocessing and segmenting the original signals; Step S3, performing online sparse decomposition and manifold mapping on the real-time high-frequency data stream based on the overcomplete fault feature dictionary, converting the high-dimensional original signals into sparse coefficients and manifold features in a low-dimensional space; Step S4, using a lightweight classification model to perform fast inference on the sparse coefficients, and triggering fault warnings and data transmission when abnormal distributions are detected.

[0007] Preferably, step S1 specifically includes the following steps: Step S11: Obtain multi-source fault waveform samples of the smart ring network cabinet during its historical operation cycle, including arc fault waveforms, partial discharge feature fingerprints, and mechanical vibration feature sequences; Step S12: Normalize and align the various fault waveform samples to construct an original training sample set; Step S13: Iteratively optimize the original training sample set using an offline learning algorithm to extract atomic features that can characterize the essential attributes of different fault types, and combine them to form a complete fault feature dictionary.

[0008] Preferably, step S3 specifically includes the following steps: Step S31, using the matching pursuit algorithm to project the real-time high-frequency data stream onto an overcomplete dictionary of fault features, and finding the atomic basis function that has the greatest correlation with the current signal residual; Step S32, obtaining the linear combination representation of the original signal on the dictionary through the iterative projection process, and generating a sparse coding vector composed of non-zero coefficients and their indices; Step S33, using the manifold learning algorithm to map the sparse coding vector to a preset low-dimensional manifold space, and extracting nonlinear structural features that reflect the fault evolution trend.

[0009] Preferably, the basis functions in the overcomplete fault feature dictionary are configured to cover common electrical faults and mechanical damage modes of ring main units. During the construction process, the system extracts feature bases with translation invariance and scale scalability by performing time-frequency domain decomposition on historical partial discharge signals. These basis functions can not only characterize normal sinusoidal waveforms and their harmonic components, but more importantly, they can accurately fit the details of sudden and transient fault pulses.

[0010] Preferably, the preprocessing includes anti-aliasing filtering and analog-to-digital conversion with a specific step size on the analog electrical signal acquired by the sensor. The system sets a preset sampling rate according to the sampling theorem to ensure that the envelope characteristics of the high-frequency partial discharge signal are not lost. The segmented quantization process divides the continuous data stream into several data frames according to a preset time period. The size of each data frame is dynamically adjusted according to the available capacity of the edge-side memory buffer to ensure the continuity of the processing logic.

[0011] Preferably, during execution, the matching pursuit algorithm selects the atomic components that best represent the signal structure by calculating the inner product of the current residual signal and each atom in the dictionary. In each iteration, the system subtracts the projected components of the selected atoms from the original signal until the energy of the residual decreases below a preset energy threshold, or the number of iterations reaches a preset step size limit. This approach ensures that even with a very small number of coefficients, the original waveform containing key fault distortion features can be reconstructed, achieving an effective conversion from a high-dimensional signal space to a sparse coefficient space.

[0012] Preferably, the sparse coding vector has extremely high sparsity. In actual operation, the high-dimensional vector with a large number of original sampling points is compressed into sparse features containing only a small number of non-zero values. This compression method differs from traditional lossy compression; it preserves the physical meaning of the signal in the dictionary space, so that each non-zero coefficient corresponds to a specific physical excitation or fault feature, providing a solid foundation for subsequent logical judgments.

[0013] Preferably, the manifold learning algorithm employs a local linear embedding or isometric mapping mechanism to preserve the local topological structure of high-dimensional features during dimensionality reduction. Through manifold mapping, the system uncovers the low-dimensional intrinsic structure hidden behind complex fault signals. In the low-dimensional manifold space, fault features of the same type (such as partial discharges in the same phase) exhibit a clear clustering trend due to their similar generation mechanisms; while different types of faults or normal disturbances are pushed to different regions of the manifold space, achieving automatic decoupling and enhancement of fault features.

[0014] Preferably, the lightweight classification model employs discrimination logic based on support vector machines or a specific topology. Since the input consists of sparse coefficients and manifold features that have undergone dimensionality reduction, the computational complexity of the classification model is significantly reduced. The model delineates normal operation regions and various fault risk regions within a preset feature space. When the feature points calculated in real-time fall outside the risk region or meet specific gradient deviation conditions, the system immediately determines it as a suspected fault state.

[0015] Preferably, the fault warning and data upload mechanism includes multi-level linkage logic. When the classification result indicates a minor anomaly, the edge-side system only uploads the compressed feature coefficients and the corresponding fault type index, informing the cloud management platform with extremely low bandwidth usage. When a serious fault or instantaneous deterioration is determined, the system uses the recorded sparse coefficients to reconstruct key segments of the original fault waveform in real time, and prioritizes uploading this high-fidelity waveform segment, providing first-hand original evidence for in-depth analysis and fault location by the backend expert system.

[0016] Preferably, this solution also includes a dictionary adaptive update strategy based on a time sliding window. The system monitors the residual energy distribution during the sparse decomposition process in real time. If the residual energy remains higher than a preset threshold for a long period, it indicates that the current overly complete fault feature dictionary cannot effectively characterize newly emerging interference or equipment aging features. At this time, the system will start a background update task, collect abnormal signals in the background update task, and feed them back to the offline learning module for dictionary expansion, ensuring that the diagnostic system has long-term robustness.

[0017] A real-time fault diagnosis system for intelligent ring main unit (RNB) cabinets, used to implement the above-mentioned method, includes: a data acquisition module for acquiring electrical signals, vibration signals, and environmental parameters of key nodes inside the RNB cabinet; a preprocessing unit connected to the data acquisition module for filtering, sampling, and segmenting the raw signals; an edge computing engine, which has a pre-built fault feature dictionary and is configured to receive preprocessed data frames and perform matching pursuit operations and manifold feature mapping to output sparse coded vectors; an intelligent diagnosis module connected to the edge computing engine for classifying and judging the sparse coded vectors based on a lightweight classification model and outputting fault status labels; and a communication gateway for selectively sending compressed feature data or alarm information to a remote monitoring center based on the judgment results of the intelligent diagnosis module.

[0018] Preferably, the edge computing engine includes a storage unit for storing an overcomplete dictionary of fault characteristics. This dictionary is loaded into a high-speed random access memory during system initialization to support millisecond-level online sparse decomposition tasks. Simultaneously, the engine integrates a hardware acceleration unit specifically for performing matrix inner product operations and iterative search processes, ensuring no data backlog occurs in high-frequency sampling environments.

[0019] Preferably, the data acquisition module integrates a high-frequency partial discharge sensor, a current transformer, and an accelerometer. The multi-source heterogeneous signals generated by each sensor are synchronized in the preprocessing unit to ensure consistency of different physical quantities along the time axis. This multi-dimensional signal fusion processing enables manifold learning to extract cross-modal correlation features, further improving the accuracy of identifying complex fault modes.

[0020] Preferably, the communication gateway has a bandwidth adaptive adjustment function. In the case of poor network communication environment, the gateway automatically increases the data compression ratio and only sends the most critical fault discrimination factors; when the network bandwidth is sufficient and a major incident occurs, the gateway switches to full detail forwarding mode and assists in incident tracing by uploading reconstructed waveforms.

[0021] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0022] 1. This invention constructs a lightweight diagnostic system suitable for edge computing environments by deeply coupling manifold learning and sparse coding techniques. First, the introduction of an overcomplete fault feature dictionary fundamentally changes the traditional data compression approach. It utilizes a priori fault knowledge base to project redundant high-frequency sampled signals onto a very small set of basis functions. This method not only achieves a data compression ratio of hundreds of times, alleviating the transmission bandwidth pressure on ring main units under millisecond-level sampling, but also ensures that the key distortion details of the fault are completely preserved in the sparse coefficients due to the clear physical directionality of the basis functions, avoiding the feature smoothing and misdiagnosis problems caused by traditional compression methods. Second, the manifold mapping process effectively extracts the nonlinear intrinsic structure in the signal, allowing noise interference and weak fault features that were originally intertwined and difficult to distinguish in the original space to be clearly separated on the low-dimensional manifold surface. This feature enhancement effect improves the system's ability to detect early defects in ring main units. In summary, this invention achieves efficient processing and accurate diagnosis of high-dimensional massive data on the edge side with limited computing resources, truly meeting the business needs of smart distribution networks for real-time fault response and high-reliability monitoring.

[0023] 2. By employing the matching pursuit algorithm for online sparse decomposition, the system can adaptively select the number of atoms based on the complexity of the signal, achieving a dynamic balance between computational overhead and representation accuracy. This means that for stable, normal data, the system requires minimal computational resources to process it, while for complex transient fault processes, the system automatically allocates more computational resources for in-depth reconstruction, optimizing the allocation efficiency of edge computing power.

[0024] 3. This invention achieves a superior technical solution in the field of intelligent ring main unit fault diagnosis by synergistically combining manifold sparse coding technology, matching pursuit algorithm, low-dimensional manifold mapping, and lightweight classification logic, balancing processing accuracy, response speed, and resource consumption. It enhances the intelligent sensing depth of power grid end nodes, providing crucial technical support for building a highly reliable and resilient modern distribution network. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the overall technical solution architecture proposed in this invention; Figure 2 This is a schematic diagram of the core principle framework for fault feature extraction based on manifold sparse coding in this invention; Figure 3 This is a flowchart illustrating the logical flow of fault feature overcomplete dictionary construction and atomic feature extraction in this invention. Figure 4 This is a flowchart of the multi-level linkage logic for edge-side lightweight inference and fault early warning in this invention. Figure 5This is a schematic diagram of the multi-level interaction relationship and data flow of the intelligent ring main unit fault real-time diagnosis system proposed in this invention. Detailed Implementation

[0026] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0027] Example 1: Please refer to the appendix Figure 1 To be continued Figure 5 In this embodiment, for 10 kV smart ring main units operating under high load in urban core areas, the present invention provides a real-time fault diagnosis method for smart ring main units. Because 10 kV smart ring main units bear dense commercial and residential power loads, their internal electromagnetic environment is complex, and the transmission of high-frequency sampling data is limited in narrowband carrier or low-power wide area network environments.

[0028] Step S1: Construct an incomplete dictionary of fault features. This step is the knowledge base construction stage of the entire diagnostic logic. Its core lies in extracting physically representative feature basis functions from massive historical operational data offline.

[0029] In step S11, multi-source fault waveform samples of the intelligent ring main unit over a 5-year operating cycle are acquired. These samples cover partial discharge signals caused by aging cable terminal insulation, weak arc waveforms caused by poor contact, and abnormal mechanical vibration characteristic sequences caused by wear of the operating mechanism. Each type of waveform sample includes typical performance under different load rates and ambient temperatures. In step S12, the acquired raw waveform samples are standardized. First, physical range normalization is performed, converting the analog voltage or current acquired by the sensors into a floating-point sequence between -1 and +1, eliminating numerical offsets caused by different sensor ranges. Next, feature alignment is performed by finding the starting trigger point or zero-crossing point of the signal to ensure phase consistency of all similar fault samples on the time axis, laying the foundation for subsequent atomic extraction. In step S13, the sample set is deeply mined using an offline learning algorithm. The system initializes a set of random basis vectors and iteratively adjusts the direction of the basis vectors using the residual minimization principle. In each iteration, the algorithm searches for the set of parameters that can represent the original waveform with the fewest possible combinations of basis vectors. After thousands of iterations, atomic features that characterize the transient spikes of the arc, partial discharge pulse oscillations, and the energy distribution of mechanical impacts are extracted. These atomic features are arranged in an ordered manner to form a large-scale overcomplete dictionary, which is stored in the flash memory of the edge device.

[0030] In one specific embodiment, the offline learning algorithm used in step S13 is specifically the K-SVD (K-Singular Value Decomposition) algorithm. Specifically, the original training sample set constructed in step S12 is represented as a matrix. ,in The number of sampling points for a single waveform sample. The total number of training samples. The system initializes an overcomplete dictionary matrix. (in (This indicates that the number of atoms is much larger than the signal dimension) and the corresponding sparse coefficient matrix. The "residual minimization principle" is quantified using the following objective function:

[0031]

[0032] in, The Frobenius norm of the matrix is ​​used to measure the global residual energy between the reconstructed signal and the original signal; Represents the sparse coefficient matrix The Column vector; This represents the vector Norm, which is the number of non-zero elements; The preset upper limit for sparsity constraint is set to 32. In each iteration, the algorithm fixes the dictionary. The sparse coefficient matrix is ​​updated using the orthogonal matching pursuit algorithm. ; then fixed Update the dictionary column by column. The atomic vectors in the equation are used until the descent gradient of the objective function is less than... Or, after reaching the maximum number of iterations of 500, the resulting matrix will be solidified. The complete dictionary of fault characteristics is stored in the storage unit of the edge device.

[0033] Step S2: Real-time acquisition of high-frequency operating data streams from the ring main unit, followed by preprocessing and segmented quantization of the raw signals. High-precision charge sensors, through-hole current transformers, and triaxial accelerometers are deployed at each key node of the intelligent ring main unit. The analog signals acquired by the sensors first enter the preprocessing unit. The preprocessing process includes anti-aliasing filtering using a low-pass Butterworth filter to remove high-frequency noise interference higher than half the sampling frequency. The analog-to-digital converter discretizes the analog signal at a sampling frequency of 20 MHz to ensure the capture of partial discharge pulse envelopes with durations only in microseconds. The segmented quantization process divides the continuously generated sampling point sequence into data frames of 4096 points each. After each frame of data enters the memory buffer, the quantization accuracy is dynamically adjusted based on the current remaining memory space and processor load. Signals with stable operation are stored with lower bit depths; signals with drastic waveform changes retain higher accuracy.

[0034] In a preferred embodiment, the squared magnitude characteristic of the transfer function of the low-pass Butterworth filter satisfies the following formula:

[0035]

[0036] in, The system transfer function of the filter. For the angular frequency of the analog signal, This is the cutoff frequency of the filter. This refers to the order of the filter. Specifically, this is determined by the preset sampling rate of the analog-to-digital converter. Set to 20MHz, and according to the Nyquist sampling theorem, set the cutoff frequency. corresponding frequency To filter out high-frequency noise interference higher than half the sampling frequency; order It is set to order 6 to provide the flattest amplitude-frequency response within the passband, ensuring distortion-free passage of the microsecond-level partial discharge pulse envelope.

[0037] For the dynamic adjustment of quantization accuracy, the preprocessing unit monitors the remaining capacity percentage of the memory buffer in real time. .when At that time, a 16-bit quantization depth is used; when At that time, quantization depth According to the formula Dynamically calculate and round down (minimum 8 bits guaranteed); when At this time, 8-bit quantization is forced to ensure the continuity of data flow when the processor is under high real-time load.

[0038] Step S3: Perform online sparse decomposition and manifold mapping on the real-time high-frequency data stream based on the overcomplete dictionary of fault features. In step S31, the edge computing engine reads the preset overcomplete dictionary of fault features and processes the current data frame using matching pursuit logic. Specifically, it calculates the inner product between the current signal vector and each atom vector in the dictionary. The size of the inner product reflects the similarity between the signal and the atom. The system searches for the atom with the largest inner product value as the first principal component. In step S32, the projection component of the current principal component in the atom direction is subtracted from the original signal to obtain the residual signal. The atom with the strongest correlation to the residual signal continues to be searched in the remaining atom library. This process of projection, subtraction, and optimization is repeated until the energy of the residual is less than 5% of the total energy of the original signal, or the number of selected atoms reaches the preset upper limit of 32. Finally, the original high-dimensional data frame is transformed into a sparse coded vector composed of these 32 non-zero coefficients and their index positions in the dictionary. This transformation compresses the redundant information of the original thousands of sampling points into a very small space. In step S33, the sparse coding vector is further processed using a manifold learning algorithm. The system searches for a low-dimensional embedding representation that can maintain this proximity relationship by calculating the local geometric distance between feature points in the high-dimensional space. Through an isometric mapping mechanism, the fault features, which originally exhibit a nonlinear curved distribution, are unfolded in a three-dimensional or four-dimensional Euclidean space. In the low-dimensional space, the distribution of features follows topological consistency, allowing subtle trends in fault evolution to be magnified and observed.

[0039] Specifically, the online sparse decomposition logic of the matching pursuit algorithm is implemented through the following rigorous mathematical iterative process: for the preprocessed input current data frame signal vector Initialize residual signal Initialize the atomic index set Initialize the number of iterations In each iteration, the current residual signal and the overcomplete dictionary of fault characteristics are calculated. Each atomic vector The inner product similarity is used to search for the atom index with the largest absolute value of the inner product using the following formula. :

[0040]

[0041] in, This represents the dot product operation of vectors. For the first The residual signal after the next iteration. For dictionary The first in Atom vectors. After finding the optimal atom, update the atom index set. The projection coefficients (i.e., sparsity coefficients) of the current signal in the selected atomic direction are calculated using the following formula. :

[0042]

[0043] Subsequently, the projected component is subtracted from the current residual using the following formula to update the residual signal. :

[0044]

[0045] The above iterative process is executed repeatedly until the following preset convergence threshold condition is met:

[0046]

[0047] in, For the first The energy of the residual signal after the next iteration The total energy of the original signal. The preset convergence threshold is 0.05, or 5%. The maximum number of atoms is preset (value 32). After stopping the iteration, a generator with non-zero coefficients is generated. and its position index in the dictionary The resulting sparse coding vector.

[0048] In one specific embodiment, the manifold learning algorithm employs an isometric mapping mechanism. Specifically, it first maps the manifolds generated within a continuous time window... The sparse encoded vectors are arranged in a high-dimensional feature space of the dataset. By calculating the local Euclidean distance between high-dimensional feature points, a nearest neighbor graph reflecting the global intrinsic structure is constructed: for any two points in the dataset... and ,like belong of Neighbors ( If the value is 5, then an undirected edge is established between the two nodes, and its initial edge weight is the Euclidean distance. If the points are not neighbors, the edge weights are set to infinity. Then, the Floyd-Warshall algorithm is used to calculate the shortest path distance between any two points, resulting in the geodesic distance matrix. To unfold the fault characteristics of the nonlinear bending distribution in a low-dimensional space, an inner product matrix is ​​constructed. :

[0049]

[0050] in, Represents the geodesic distance matrix The matrix obtained by squaring each element in the matrix. For a centered matrix, It is the identity matrix. It is a column vector consisting entirely of 1s. The inner product matrix... Perform eigenvalue decomposition and select the largest value. eigenvalues The value is 3, which means it is mapped to a three-dimensional manifold space and its corresponding eigenvector. The low-dimensional manifold coordinate matrix that preserves the local topology can be calculated using the following formula. :

[0051]

[0052] Through the topological consistency of this manifold space, similar fault characteristics are... Clustering enhancement occurs, and different types of fault modes achieve nonlinear decoupling.

[0053] Step S4: Utilize a lightweight classification model for rapid inference of sparse coefficients. The support vector machine classifier pre-installed on the edge side, upon receiving low-dimensional manifold features, immediately determines the location on a preset decision plane. If the feature point falls within the central steady-state region, the equipment is considered to be operating well; if the feature point shifts towards a partial discharge risk region, and the velocity gradient of the shift exceeds a preset threshold, the system immediately triggers a fault warning. The fault warning and data transmission mechanism employs a hierarchical strategy. When a suspected arc fault is detected, the communication gateway immediately blocks unnecessary background heartbeat packets and prioritizes sending high-priority data packets containing the current fault index, sparse coefficient values, and the time of the fault occurrence to the scheduling center. If the fault is determined to be a severe fault caused by insulation breakdown, the system retrieves the original data frame from memory, reconstructs the waveform using sparse coefficients, and transmits the reconstructed high-fidelity fault waveform segment for post-incident tracing by maintenance personnel.

[0054] As a preferred implementation, the lightweight classification model employs a support vector machine based on radial basis function kernels. Specifically, the low-dimensional manifold eigenfeatures extracted in step S33 are... As input, the decision function of the support vector machine determines the position using the following formula:

[0055]

[0056] in, The discrimination result output by the model. For symbolic functions, The number of support vectors. For the first The Lagrange multipliers corresponding to each support vector (obtained during offline training by maximizing the objective function of the margin). The class label for the training samples (+1 for normal, -1 for abnormal). The hyperparameters of the RBF kernel function (taking values ​​of...) , (The kernel width is determined to be 0.5 based on cross-validation). This is a category bias term.

[0057] For the determination logic of "offset velocity gradient exceeding preset threshold", the system calculates continuously in real time. The displacement trajectories of feature points on the low-dimensional manifold within a time window are calculated relative to the normal operating centroid using the following formula. Deviation gradient :

[0058]

[0059] in, The manifold characteristics at the current moment, The value is 5. When the judgment result satisfies... And deviates from gradient (Preset threshold) When the value is 0.15, the system confirms that the feature point falls outside the normal operating area and shows an accelerating deviation trend. It immediately triggers a fault warning signal and transmits the judgment result to the communication gateway to perform multi-level linkage data transmission.

[0060] Example 2: In this example, for distributed energy access ring network cabinets installed in coastal high salt spray and high humidity environments, the fault diagnosis method is based on Example 1, with specific logical optimizations made for interference environment and multi-energy fluctuation characteristics.

[0061] In the process of constructing the dictionary in step S1, considering the more complex electromagnetic noise background in coastal environments, this embodiment of the second case introduces additional environmental noise suppression atoms when constructing the overcomplete dictionary of fault features. In step S11, when acquiring samples, interference waveforms under typhoon, heavy fog, and high salt spray conditions were specifically collected. Through iterative optimization in step S13, the dictionary not only includes equipment fault atoms but also noise atoms that can accurately fit the environmental background noise. This allows the system to preferentially remove environmental noise components in subsequent sparse decomposition, improving the purity of fault feature extraction.

[0062] In the preprocessing stage of step S2, since the distributed energy access causes frequent voltage fluctuations, the system introduces adaptive gain control. When a large fluctuation in the bus voltage is detected due to photovoltaic grid connection, the front-end signal conditioning circuit automatically adjusts the feedback resistor ratio of the operational amplifier to ensure that the signal is always within the optimal quantization range of the analog-to-digital converter, avoiding signal clipping or a decrease in signal-to-noise ratio due to excessively small resistors.

[0063] In the online sparse decomposition of step S3, this embodiment employs an improved basis pursuit denoising logic. Unlike the simple matching pursuit in Embodiment 1, this embodiment incorporates a regularization term reflecting signal smoothness during the projection process. When calculating the inner product, the system considers not only the similarity between atoms and signals but also the continuity of the selected atom combinations on the time axis. This is effective in identifying atypical harmonic interference generated during the backfeeding of distributed energy resources. In the manifold mapping stage, a locally linear embedding algorithm is used to focus on mining the nonlinear distortion characteristics of fault waveforms caused by the impedance changes in the distribution network due to the access of new energy sources.

[0064] In the diagnostic logic of step S4, the lightweight classification model employs a multi-weighted discrimination mechanism. For high-salt-fog environments, the system increases the sensitivity weight to partial discharge characteristics, while a dedicated shielding time window is set for transient pulses generated by photovoltaic inverter switching. When a fault warning occurs, the communication gateway detects the current wireless signal strength. In cases of poor coverage at coastal base stations, the gateway ensures the arrival rate of alarm messages through multiple retransmissions and caches a large amount of reconstructed waveform data on an industrial-grade SD card at the edge. Once the network link bandwidth recovers to above a preset threshold, the breakpoint resume transmission is initiated.

[0065] This differentiated implementation approach, tailored to specific application scenarios, ensures the diagnostic robustness of the smart ring main unit under complex geographical and environmental conditions.

[0066] Example 3: In this example, for smart ring main units in industrial parks containing a large number of high-power inductive load motors, this invention provides a real-time fault diagnosis system and method based on multi-source heterogeneous signal fusion. The characteristics of ring main units in industrial parks are that they are frequently subjected to heavy-load impact vibrations and electromagnetic transients generated by the starting and stopping of large motors.

[0067] In this embodiment, the data acquisition module integrates higher-specification physical layer devices. In addition to traditional electrical sensors, it also incorporates distributed fiber optic temperature sensors and high-frequency acoustic sensors. The preprocessing unit features multi-channel synchronous timing capabilities, using an internal clock management chip to assign uniform nanosecond-level timestamps to electrical, vibration, and acoustic signals. This physical synchronization ensures that, in the event of a fault, the system can perform joint analysis from three dimensions: abrupt changes in electrical signals, frequency shifts in shell vibration, and ultrasonic signals generated by internal discharge.

[0068] The overcomplete fault feature dictionary within the edge computing engine is configured as a cross-modal composite dictionary. This dictionary contains not only atoms representing single signals but also composite atoms characterizing electro-resonant coupling and electro-acoustic coupling. During sparse decomposition in step S3, the engine concatenates multi-channel data into ultra-high-dimensional observation vectors. Through the isometric mapping mechanism in manifold learning, the system searches for the intrinsic correlations between these heterogeneous signals in high-dimensional space. For example, when an internal arc fault occurs, the system can extract not only the peak features of the current waveform but also the corresponding sound pressure pulse from the acoustic channel. These multidimensional features aggregate into highly distinctive feature clusters in the manifold space.

[0069] The lightweight classification model within the intelligent diagnostic module employs a topology-based clustering logic. The model not only determines which region a feature point falls into, but also predicts the severity of the fault by calculating the radial deviation rate of the feature point relative to the normal operating centroid. If the deviation rate increases exponentially, the system classifies it as a transient, severe fault.

[0070] In this embodiment, the communication gateway features deep protocol parsing and traffic shaping capabilities. Considering the bandwidth competition from other production and management data streams within the campus, the gateway automatically seizes bandwidth control using the Quality of Service (QoS) protocol when it detects a ring main unit fault alarm. At this time, the gateway categorizes the diagnostic results into three levels of data streams: the first priority is the fault type and action command, delivered in milliseconds using extremely simple messages; the second priority is the critical sparse coefficient stream, used for real-time trend display in the cloud; and the third priority is the original waveform reconstruction packet, sent up when the network is idle.

[0071] This cross-modal diagnostic scheme, implemented through the collaboration of hardware and algorithms, improves the accuracy of identifying subtle early defects in environments with strong power frequency interference.

[0072] Example 4: This example details a smart ring main unit fault diagnosis system and its dictionary adaptive update strategy. This system aims to solve the false alarm problem caused by the drift of original characteristics due to component aging throughout the equipment's life cycle.

[0073] The system includes a data acquisition module, a preprocessing unit, an edge computing engine, an intelligent diagnostic module, and a communication gateway. The edge computing engine is equipped with a storage unit that uses non-volatile storage technology to store a complete dictionary of currently active fault characteristics.

[0074] The core difference lies in the addition of adaptive dictionary update closed-loop logic in this embodiment. During system operation, the edge computing engine records the residual energy distribution after each sparse decomposition in real time. If, within a continuous 24-hour monitoring period, more than 10% of the data frames are found to have residual energy consistently higher than 15% of the original signal energy after sparse decomposition, it indicates that the existing overcomplete dictionary can no longer cover the current device characteristics, and new environmental interference sources have emerged or the device has undergone irreversible parameter evolution.

[0075] At this point, the system automatically initiates a background update task. The communication gateway marks these high-residual data fragments as "unknown feature samples" and asynchronously uploads them to the cloud management platform. The cloud-based expert system utilizes its more powerful computing resources to perform cluster analysis on these new samples. If stable, repeating structures are found, it is determined that a new feature has emerged. The cloud then generates corresponding patch atoms using offline learning algorithms and distributes these new atoms to the edge-side communication gateway.

[0076] Upon receiving the patched atoms, the edge computing engine dynamically loads the new atoms into the high-speed random access memory using the storage unit's update interface, without interrupting the diagnostic tasks. This incremental update method avoids the system downtime risk associated with a full dictionary replacement. After the update, the system's ability to sparsely represent newly emerging interference or aging features is significantly improved, and the residual energy returns to a preset low-level range.

[0077] To ensure the algorithm's execution speed, the edge computing engine integrates a hardware acceleration unit. This unit features a multi-stage pipeline structure specifically designed for matrix inner product operations. When executing the matching pursuit algorithm, the hardware accelerator simultaneously performs parallel projection calculations on 128 atoms, keeping the sparse decomposition time of a single frame signal within 50 microseconds. This hardware-level parallel processing capability is the physical foundation for achieving high-frequency real-time diagnostics.

[0078] The communication gateway features adaptive bandwidth adjustment. It can monitor the congestion level of the dedicated power grid in real time. During peak grid hours, when communication bandwidth is limited, the gateway automatically compresses the quantization bits of the sparse coding from 16 bits to 8 bits and transmits only the first 10 most critical coefficients. During periods of ample bandwidth, such as late at night, the gateway automatically uploads the complete sparse feature vector for more accurate equipment health assessment in the background.

[0079] This self-evolving system architecture ensures that the intelligent ring main unit maintains optimal fault diagnosis performance throughout its 10 to 15-year service life.

[0080] Example 5: In this example, for a large-scale intelligent ring network cabinet cluster inside a substation, the present invention provides a distributed collaborative real-time fault diagnosis scheme, focusing on the data association and discrimination logic between multiple ring network cabinet nodes.

[0081] In step S1, when constructing the dictionary, this embodiment not only considers the local fault atoms of a single device, but also constructs a global dictionary reflecting common-mode interference within the station. By analyzing the signals such as external short-circuit faults and systemic harmonics that occur synchronously in the historical operation of multiple ring network cabinets within the station, common basis functions with station-domain characteristics are extracted.

[0082] In the real-time acquisition phase of step S2, multiple ring main unit diagnostic nodes within the station area are synchronously triggered via an internal high-speed communication bus. When one of the ring main units detects an abnormal fluctuation, it sends a collaborative sampling request to neighboring nodes. Upon receiving the request, the preprocessing unit automatically extends the acquisition window of the current data frame to capture the complete disturbance propagation process.

[0083] In the sparse decomposition logic of step S3, this embodiment introduces a spatial constraint term. When the edge computing engine performs manifold mapping on local data, it references the sparse coefficients transmitted from neighboring nodes. If the feature points of multiple nodes in the low-dimensional manifold space exhibit consistent trajectory offsets, the system determines it as an external system-level fault; if only the feature points of local nodes experience abrupt changes, it is determined as a local fault within the cabinet. This collaborative mechanism reduces the risk of false alarms from individual machines caused by electromagnetic coupling within the station.

[0084] In the diagnostic output section of step S4, the intelligent diagnostic module is equipped with multi-level linkage logic. When a serious fault occurs in a certain loop and is confirmed, the system not only sends the alarm information of this cabinet, but also generates a comprehensive diagnostic report that includes the operating status of neighboring nodes. The communication gateway uses a priority-based message scheduling algorithm to package this comprehensive report and pushes it directly to the station monitoring platform through a secure encrypted channel.

[0085] To achieve efficient data processing, this embodiment simplifies the matching pursuit algorithm. When searching for the optimal atom, the system employs a hierarchical search-based skip strategy. First, a coarse search is performed within a small subset of the dictionary to determine the approximate physical category of the fault (e.g., electrical or mechanical). Then, a fine search is performed within the corresponding sub-atom library. This hierarchical search logic reduces the search step size by more than 60%, alleviating processor pressure when cluster nodes are operating simultaneously.

[0086] Through this distributed and collaborative implementation method, the intelligent ring main unit fault diagnosis system has evolved from a single individual monitoring system to a monitoring network with collective intelligence, which greatly improves the perception depth and discrimination accuracy of complex power distribution network status.

[0087] Example 6: In this example, the real-time fault diagnosis method for intelligent ring main unit is described in detail in response to the coupled fault of temperature rise and insulation degradation caused by extreme overload, aiming to illustrate the application of the present invention in multi-physics coupling analysis.

[0088] During the construction of the complete dictionary in step S1, partial discharge waveforms under different temperature gradients were simulated in the laboratory to address the changes in the polarization characteristics of the insulating material caused by temperature rise. In step S13, a set of temperature-sensitive atomic features were extracted. These atoms not only captured the amplitude changes of the pulse but also the fine-tuning of the waveform's slope on the rising and falling edges.

[0089] In step S2, the data acquisition module synchronously acquires infrared temperature measurement data of the busbar. The preprocessing unit uses the current temperature parameter as an additional dimension and encapsulates it with the high-frequency sampling signal. During quantization, if the temperature exceeds the preset threshold of 65 degrees Celsius, the system automatically activates high-gain mode to amplify weak insulation degradation signals.

[0090] In step S3, when the edge computing engine performs manifold-based feature extraction, it uses temperature parameters as constraint coordinates in the manifold space. Through manifold mapping, the system constructs a reference manifold surface that dynamically changes with temperature. During sparse decomposition, the matching pursuit algorithm dynamically selects the best-matching subset of atoms from the dictionary based on the current temperature coordinates. This dynamic dictionary selection mechanism eliminates the interference of natural waveform drift caused by temperature on diagnostic accuracy.

[0091] In the classification logic of step S4, the lightweight classification model incorporates trend evolution analysis. The system continuously calculates the displacement vectors of feature points on the low-dimensional manifold. If the direction of the displacement vector points to the pole of insulation failure, and the magnitude of the displacement is proportional to the rate of temperature increase, then a severe thermal insulation defect is identified.

[0092] Upon triggering a fault warning, the communication gateway executes a specific emergency data transmission strategy. The system compresses and sends a composite message containing temperature curves, sparse coefficient evolution diagrams, and partial discharge frequency statistics. Through this highly compressed but information-dense characteristic data, the cloud management platform can immediately outline the trend of physical state changes within the equipment and issue a power outage maintenance suggestion before complete insulation breakdown.

[0093] Through the detailed descriptions of the above embodiments, this invention resolves the contradiction between data volume and bandwidth in high-frequency sampling environments for intelligent ring network cabinets by employing manifold sparse coding technology. By constructing an overcomplete dictionary of fault features and utilizing a matching pursuit algorithm for online sparse decomposition, extremely high data compression ratios and preservation of key features are achieved. Combined with manifold learning algorithms to extract nonlinear structural features, and real-time discrimination by a lightweight classification model, efficient and accurate diagnosis is ensured even with limited computing power at the edge. A multi-level linkage early warning mechanism and an adaptive dictionary update strategy further enhance the system's practicality and long-term operational stability.

[0094] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for real-time fault diagnosis of intelligent ring main unit, characterized in that, Includes the following steps: Step S1: Construct an overcomplete dictionary of fault characteristics; Step S2: Real-time acquisition of high-frequency operating data streams from the ring main unit, and preprocessing and segmenting the raw signals; Step S3: Perform online sparse decomposition and manifold mapping on real-time high-frequency data stream based on the overcomplete dictionary of fault features: By calculating the inner product similarity between the data frame vector to be processed and each atomic vector in the overcomplete dictionary of fault features, the projection search and signal residual update are performed iteratively until the ratio of residual energy to the total energy of the original signal drops below the preset convergence threshold, a sparse coding vector consisting of non-zero coefficients and their position index in the dictionary is generated. Then, by calculating the local geometric distance between high-dimensional feature points, the sparse coding vector is mapped to a low-dimensional manifold space that preserves the local topology using manifold transformation, so as to extract nonlinear intrinsic features that reflect the fault evolution trend. Step S4: Use a lightweight classification model to quickly infer sparse coefficients and trigger fault warning and data upload when abnormal distribution is detected: Input the extracted low-dimensional manifold features into the preset lightweight classification model for position determination, monitor the displacement trajectory of feature points on the low-dimensional manifold surface and the deviation gradient relative to the normal operating centroid in real time, and when the determination result meets the preset risk area conditions, execute multi-level linkage data upload including compressed feature coefficients or reconstructed original waveform segments according to the severity level of the fault.

2. The real-time fault diagnosis method for a smart ring main unit according to claim 1, characterized in that, The construction of an overcomplete fault feature dictionary includes: acquiring multi-source fault waveform samples of smart ring network cabinets during historical operation cycles and performing physical range standardization processing; using the residual minimization principle to iteratively adjust the direction of randomly generated basis vectors; extracting atomic features that characterize the essential physical attributes of different fault types; and arranging the extracted atomic features in an orderly manner according to feature category and physical orientation to form an overcomplete fault feature dictionary with the ability to sparsely represent typical fault features.

3. The real-time fault diagnosis method for a smart ring main unit according to claim 2, characterized in that, Step S2 includes: acquiring multi-source analog signals through sensors deployed at key nodes of the ring network cabinet, performing anti-aliasing filtering to filter out high-frequency noise interference, and performing high-frequency discretization conversion. The generated sampling sequence is divided into continuous data frames according to a fixed data scale. At the same time, the data frames are adjusted in precision and segmented and encapsulated according to the real-time available capacity of the edge-side memory buffer.

4. The real-time fault diagnosis method for a smart ring main unit according to claim 3, characterized in that, Step S1 specifically includes: acquiring the arc fault waveform, partial discharge characteristic fingerprint and mechanical vibration characteristic sequence of the smart ring network cabinet in the historical operation cycle; Physical range normalization processing is performed on various fault waveform samples to convert the analog quantities collected by the sensor into floating-point number sequences within a fixed value range. Phase alignment processing is performed on similar fault samples through the starting trigger point of the positioning signal to construct the original training sample set. The original training sample set is iteratively optimized by an offline learning algorithm. In each iteration, a set of parameters representing the original waveform with a very small number of basis vector combinations is found. Atomic features characterizing the transient spikes of the electric arc, the partial discharge pulse oscillations and the energy distribution of mechanical action impact are extracted and stored in the storage unit of the edge device.

5. The real-time fault diagnosis method for a smart ring main unit according to claim 4, characterized in that, Step S2 specifically includes: using a low-pass Butterworth filter to perform anti-aliasing filtering on the analog electrical signal acquired by the sensor, filtering out high-frequency components that are higher than half of the sampling frequency; The analog signal is converted to digital according to the preset sampling rate to ensure that the microsecond-level partial discharge pulse envelope is captured; The continuously generated sampling point sequence is cut and packaged according to a preset length, and after the data frame enters the memory buffer, the quantization bit depth is dynamically adjusted according to the real-time load of the processor, and high-precision storage is retained for signal segments with drastic waveform changes.

6. The real-time fault diagnosis method for a smart ring main unit according to claim 5, characterized in that, The specific logic of online sparse decomposition in step S3 is as follows: the matching pursuit algorithm is used to calculate the inner product of the current data frame signal vector and each atom vector in the fault feature overcomplete dictionary, and the atom with the largest inner product value is searched out as the current principal component. Subtract the projection component of the principal component in the atomic direction from the current signal to obtain the residual signal, and continue to search for the atom with the strongest correlation to the residual signal in the remaining atom library; Repeat the projection, subtraction, and optimization process until the residual energy ratio is lower than the preset ratio, or the number of atoms selected reaches the preset upper limit, transforming the high-dimensional data frame into sparse features containing a small number of non-zero values.

7. The real-time fault diagnosis method for a smart ring main unit according to claim 6, characterized in that, The specific logic of manifold mapping in step S3 is as follows: the sparse coding vector generated is processed by local linear embedding or isometric mapping mechanism, and a nearest neighbor graph reflecting the global intrinsic structure is constructed by calculating the local Euclidean distance between high-dimensional feature points. In a low-dimensional Euclidean space, a coordinate embedding representation that preserves the local topological relationships of the nearest neighbor graph is sought, and the fault features of the nonlinear curved distribution are unfolded in the low-dimensional space. By leveraging the topological consistency of the manifold space, similar fault features are clustered and enhanced, while nonlinear decoupling is performed on different types of fault modes.

8. The real-time fault diagnosis method for a smart ring main unit according to claim 7, characterized in that, The specific logic for triggering fault warning and data transmission in step S4 is as follows: when the lightweight classification model determines that the feature point falls outside the normal operation area and the offset speed gradient exceeds the preset threshold, a fault warning signal is triggered. When the judgment result indicates a suspected fault state, the communication gateway prioritizes constructing a data packet containing the current fault index, sparse coefficient value, and the time of the fault and sending it to the management platform; When the judgment result indicates a severe fault or insulation breakdown, the original sampled data frame in memory is retrieved, the waveform is reconstructed using sparse coding vectors, and the reconstructed high-fidelity fault waveform segment is uploaded first.

9. The real-time fault diagnosis method for a smart ring main unit according to claim 8, characterized in that, The aforementioned method for real-time fault diagnosis of a smart ring main unit further includes a dictionary adaptive update step: the edge computing engine monitors the residual energy distribution during the sparse decomposition process in real time and calculates the proportion of data in which the residual energy is continuously higher than the preset proportion of the original signal energy within a preset monitoring period. When the proportion of the data exceeds a preset frequency threshold, the system automatically starts a background update task, marks the data fragments with abnormal residual energy as unknown feature samples and sends them to the cloud; The cloud management platform uses offline learning algorithms to perform cluster analysis on unknown feature samples and generate patch atoms, which are then distributed to the edge computing engine through the communication gateway.

10. A real-time fault diagnosis system for intelligent ring main unit, used to implement the method according to any one of claims 1 to 9, characterized in that, include: The data acquisition module is configured to acquire electrical signals, vibration signals, and environmental parameters of each key node inside the ring main unit; The preprocessing unit, connected to the data acquisition module, is configured to perform anti-aliasing filtering, high-frequency sampling, and segmented encapsulation on the original signal, and dynamically adjust the quantization accuracy according to the resource status on the edge side. The edge computing engine has a built-in fault feature overcomplete dictionary. It is configured to receive pre-processed data frames and perform online sparse decomposition and manifold feature mapping based on the matching pursuit algorithm, and output a sparse coding vector composed of non-zero coefficients and low-dimensional manifold eigenfeatures. The intelligent diagnostic module, connected to the edge computing engine, is configured to classify and discriminate low-dimensional manifold features based on a lightweight classification model and output fault status labels by monitoring feature displacement trajectories; the communication gateway is configured to perform bandwidth adaptive adjustment based on the fault status labels output by the intelligent diagnostic module and selectively send compressed feature data or reconstructed high-fidelity waveform information to the remote monitoring center.

Images