A method for intelligent alarm push in data center power distribution monitoring systems

By constructing a weighted directed capacity network graph in the data center power distribution monitoring system and using a gated clock recursive adaptive topology-aware model, combined with the network flow minimum cut pruning algorithm, the problem of being unable to quickly locate the source of faults under multi-node cascading faults is solved, and fast and accurate fault source location and alarm pruning are achieved.

CN123074115APending Publication Date: 2026-11-10QINGDAO HENGHUA COMPUTER-ROOM EQUIP & PROJECT CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611212499.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-11
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

In existing technologies, data center power distribution monitoring systems cannot effectively eliminate derivative alarms under multi-node cascaded failure conditions, making it difficult for maintenance personnel to quickly locate the fault source node.

Method used

A weighted directed capacity network graph is used in combination with a gated clock recursive adaptive topology sensing model and a network flow minimum cut pruning algorithm to collect electrical state variables of power distribution system nodes in real time, construct a dynamic topology graph, eliminate transient noise, locate fault source nodes through time series analysis and dynamic threshold comparison, and prune downstream derived alarms.

Benefits of technology

It enables rapid and accurate location of the fault source node in multi-node cascaded fault conditions, reducing the number of alarms faced by maintenance personnel and improving fault handling efficiency.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123074115A_ABST
    Figure CN123074115A_ABST
Patent Text Reader

Abstract

This invention provides an intelligent alarm push method for data center power distribution monitoring systems, belonging to the field of power distribution technology. This invention utilizes a double-buffered circular queue and a lock-free dynamic memory pool to achieve zero-copy rapid distribution of electrical messages. After removing transient noise using a composite filtering algorithm, a gated clock recursive adaptive topology sensing model is used to output the probability distribution of abnormal states and transient change rates of each node, triggering dynamic threshold comparison. Then, a network flow minimum cut pruning algorithm is used to solve for the minimum cut edge set on a weighted directed capacity network graph, accurately locating the fault source node and pruning all downstream derivative alarms. Finally, an AI-assisted alarm containing the fault time, source node identifier, fault type, and handling solution is pushed to maintenance personnel. This solves the technical problem that power distribution monitoring systems cannot effectively remove derivative alarms under multi-node cascaded fault conditions, making it difficult for maintenance personnel to quickly locate the fault source node.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of power distribution technology, and more specifically, relates to an intelligent alarm push method applied to a data center power distribution monitoring system. Background Technology

[0002] Data center power distribution monitoring systems typically employ alarm push methods based on fixed threshold rules or statistical characteristics. These methods involve real-time sampling of electrical parameters at each node and comparing them to thresholds. When a parameter exceeds a preset range, an alarm is triggered and pushed to the operations and maintenance platform. In scenarios with small-scale power distribution networks and simple fault conditions, this approach can meet basic monitoring needs. However, in large-scale data center power distribution systems, the power distribution topology is a multi-level tree structure with strong electrical coupling between nodes. When an upstream node fails, the electrical parameters of all its downstream subordinate nodes will simultaneously show abnormal responses, triggering a large number of derivative alarms and creating an alarm storm. Under these conditions, existing fixed-threshold rule-based alarm methods cannot distinguish between primary fault alarms and derivative alarms, making it difficult for operations and maintenance personnel to quickly identify the fault source node when faced with a massive number of alarms, thus delaying fault handling. In existing technologies, due to the lack of modeling and utilization of the causal relationship of power distribution topology, the fixed threshold method only judges the parameters of a single node independently and cannot perceive the electrical causal propagation path across nodes. In other words, existing technologies have the technical problem that power distribution monitoring systems cannot effectively eliminate derivative alarms under multi-node cascade fault conditions, making it impossible for maintenance personnel to quickly locate the fault source node. Summary of the Invention

[0003] In view of this, the present invention provides an intelligent alarm push method for data center power distribution monitoring systems, which can solve the technical problem in the prior art that the power distribution monitoring system cannot effectively eliminate derivative alarms under multi-node cascaded fault conditions, making it impossible for operation and maintenance personnel to quickly locate the fault source node.

[0004] This invention is implemented as follows: This invention provides an intelligent alarm push method applied to a data center power distribution monitoring system, comprising the following steps:

[0005] Based on the graph attention network dynamic topology adaptive reconstruction model, the electrical state quantities of each node in the power distribution system are collected in real time, a weighted directed capacity network graph is constructed, and the rated current carrying capacity is used as the edge capacity.

[0006] A double-buffered circular queue receives electrical messages from each node, performs zero-copy fast distribution via a lock-free dynamic memory pool, and synchronously starts a composite filtering algorithm to remove transient noise.

[0007] The gated clock recursive adaptive topology sensing model performs feature projection and timing analysis on multi-dimensional electrical parameters, outputs the probability distribution of abnormal states and transient change rate of each node, and triggers dynamic threshold comparison.

[0008] For the set of abnormal nodes that exceeds the dynamic threshold, the network flow minimum cut pruning localization algorithm is executed to solve for the minimum cut edge set and locate the source node of the power distribution fault.

[0009] Using the source node corresponding to the minimum cut edge set as the root, remove all downstream derived abnormal information and retain only the fault alarm record of the source node.

[0010] Alarms are pushed to operations and maintenance personnel, including the time of failure, the identifier of the source node, the type of failure, the possible causes of failure and the handling solution, and can be expanded to view the details of abnormal parameters of downstream sub-nodes.

[0011] Specifically, the construction of the weighted directed capacity network graph involves using the bus or source-end circuit breaker as the source point, the terminal load circuit breakers at each level as the sink point, and the circuit breakers of each distribution branch as intermediate nodes. The weight of the directed edges between each node is taken as the rated current carrying capacity of the corresponding circuit breaker. The graph attention network dynamic topology adaptive reconstruction model reconstructs the virtual topology connection matrix in reverse based on the instantaneous changes in the coupling of electrical state variables, so that the topology graph is consistent with the physical power distribution system.

[0012] Wherein, the virtual topology connection matrix is 3D sparse matrix, The matrix element represents the total number of monitoring nodes in the current power distribution system. The matrix element takes the real-time normalized value of the directed current between the corresponding nodes. When the normalized value of the current between the nodes is lower than the connectivity determination threshold, the corresponding matrix element is set to zero, triggering the topology reconstruction process. The connectivity determination threshold is determined by simulating various operating conditions in a controlled experimental environment, collecting the distribution data of the normalized current values ​​between each node, and taking the lower bound of the 95% confidence interval through iterative analysis.

[0013] The dual-buffered circular queue consists of two independent fixed-length memory buffers, with write and read pointers operating in different buffers and switching between them using atomic operations. The lock-free dynamic memory pool is implemented based on a linked list structure of pre-allocated fixed-size memory blocks. When a message arrives, it directly retrieves a free block from the memory pool and returns it after processing. Zero-copy fast distribution means that only memory block pointers, not the data itself, are transmitted during the entire transmission of message data from the acquisition gateway to the analysis thread.

[0014] The composite filtering algorithm consists of a sliding window time-varying median filter and a wavelet threshold denoising system connected in series; the sliding window time-varying median filter has a length of... A sliding window takes the median of the electrical timing signal point by point; wavelet thresholding denoising uses a soft thresholding method, with the threshold determined by the signal standard deviation estimate. The criteria are established. It is obtained by estimating the minimum absolute median deviation.

[0015] The input to the gated clock recursive adaptive topology sensing model is the time window of each node. The multidimensional electrical parameter acquisition sequence within, with dimensions of It includes five features: effective voltage value, effective current value, active power, power factor, and phase deviation. The input is linearly transformed by the pre-spatial projection layer and then enters the clock-gated loop layer. Each node is assigned an independent adaptive time step update frequency. The hidden layers of each time scale are connected by a long-span residual dense jump matrix. The output layer is a multi-node abnormal state probability classification head.

[0016] In the clock-gated loop layer, the node update frequency The state is dynamically determined by the exponentially weighted average of its historical transient change rate. When the transient change rate of a node exceeds the dynamic threshold, the hidden layer state dynamic backtracking mechanism is triggered, which resets the hidden state of the node to the most recent steady state before the failure, and sends an activation signal to the adjacent nodes in the topology graph to synchronously update the hidden state of the adjacent nodes.

[0017] The gated clock recursive adaptive topology sensing model internally nests the Dijkstra shortest path algorithm, using the anomaly probability value of each node in the topology graph as the edge weight. After each inference, it automatically calculates the minimum cost path from the source point to each abnormal node to assist in locating the direction of fault propagation.

[0018] Wherein, the dynamic threshold is the activation gain adjustment function. The output, Based on the average transient change rate of nodes in the current batch Current system load rate and historical false alarm rate The gain adjustment value is obtained by calculating three data points. ,in , , These are weighting coefficients, and the sum of the three is 1; when When the threshold exceeds the tightening threshold, the dynamic threshold tightens to 0.8 times the baseline threshold; when... When the threshold is lower than the relaxed threshold, the dynamic threshold is relaxed to 1.2 times the baseline threshold; the baseline threshold is calculated by collecting fault-free operation data and applying it to various parameters. The interval boundary value is obtained.

[0019] The network flow minimum cut pruning and localization algorithm uses the Ford-Fulkerson method to solve for the maximum flow from the source to the sink in a weighted directed capacity network graph, and obtains the minimum cut edge set by the maximum flow minimum cut theorem. The minimum cut edge set physically corresponds to the minimum set of circuit breakers that interrupt the current transmission of the current due to the current fault. After the algorithm completes the minimum cut solution, all nodes downstream of the minimum cut are marked as derived abnormal nodes and removed from the alarm queue, and only the alarms of the nodes associated with the minimum cut edge set are retained.

[0020] The establishment of the training dataset for the gated clock recursive adaptive topology sensing model includes: simulating typical fault conditions such as single-phase undervoltage, three-phase short circuit, circuit breaker malfunction, and poor line contact on a laboratory power distribution test bench; collecting time-series data of multi-dimensional electrical parameters of each node and labeling the fault type and fault source node number; simultaneously collecting normal operating condition data as negative samples; performing composite filtering preprocessing on the collected data; dividing the data into training set, validation set, and test set in an 8:1:1 ratio; and performing zero-mean unit variance normalization on the parameters of each dimension.

[0021] The training of the gated clock recursive adaptive topology sensing model uses cross-entropy loss as the classification loss, with an additional physical constraint loss term based on the transient evolution equation of the transient circuit. The physical constraint loss term requires that the hidden state entropy of the model be minimized when there is no fault and that the hidden state entropy increase be maximized when a fault occurs. The AdamW optimizer and cosine annealing strategy are used. The early stopping criterion is to minimize the weighted sum of the false positive rate and the false negative rate on the validation set. After training, the model weights are quantized into half-precision floating-point numbers before deployment.

[0022] The memory allocation strategy of the gated clock recursive adaptive topology-aware model adopts a layer-by-layer gradient checkpointing approach, retaining only the memory allocated by each layer during forward propagation. The activation values ​​of the first layer are calculated, and the activation values ​​of the remaining layers are recalculated during backpropagation. The connection weights between neurons are stored in a grouped sparse compressed storage format, which divides the weights within the same node into groups and stores only the non-zero block index and weight value. The hidden state tensors between layers are stored as half-precision floating-point numbers.

[0023] Wherein, the weighting coefficient , , To obtain the data, a validation dataset containing no fewer than 500 labeled fault samples was constructed on a simulated power distribution fault test bench. The objective function was the weighted sum of the false alarm rate and the false negative rate. To constrain the parameters, a grid search method is used to traverse all three combinations, and the parameter combination that minimizes the objective function is taken as the initial value. Then, gradient descent is used for local fine-tuning.

[0024] Wherein, the dimension of the long-span residual dense jump matrix is , The loop depth is 4 to 8; the window length of the sliding window time-varying median filter is 4 to 8. The value range is 5 to 21 sampling points; check the gradient points by layer. The value range is 2 to 4; the block size range of the sparse compressed storage format for the inter-neuron connection weights is 16 to 64; the typical range of the connectivity determination threshold is 0.05 to 0.15; the tightening threshold is 0.75, and the relaxation threshold is 0.4.

[0025] This invention models the power distribution system topology as a weighted directed capacity network graph, combines it with a gated clock recursive adaptive topology awareness model to output the probability distribution of abnormal states of each node, and then uses a network flow minimum cut pruning algorithm to solve for the minimum cut edge set. This accurately locates the fault source node and prunes all downstream derivative alarms, solving the technical problem that power distribution monitoring systems cannot effectively eliminate derivative alarms under multi-node cascaded fault conditions. The weighted directed capacity network graph explicitly encodes the electrical causal relationships of the power distribution topology as directed edges and edge capacities, making the minimum cut solution directly correspond to the physical circuit breaker location, ensuring the determinism and traceability of fault tracing. The gated clock recursive adaptive topology awareness model uses an asynchronous update mechanism to provide more dense state updates for high-frequency transient nodes, and combines it with a hidden layer state dynamic backtracking mechanism to eliminate the contamination of pre-fault detached states, improving the accuracy of the abnormal probability output, thus providing reliable node weight input for the minimum cut algorithm. In summary, this invention solves the technical problem mentioned in the background art: the power distribution monitoring system cannot effectively eliminate derivative alarms under multi-node cascaded fault conditions, making it impossible for maintenance personnel to quickly locate the fault source node. Attached Figure Description

[0026] Figure 1 This is a flowchart of the method of the present invention.

[0027] Figure 2 This is a schematic diagram of the fault propagation path based on Dijkstra's shortest path algorithm. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0029] like Figure 1 The diagram shows a flowchart of an intelligent alarm push method for a data center power distribution monitoring system provided by the present invention. This method includes the following steps:

[0030] S01. Based on the graph attention network dynamic topology adaptive reconstruction model, the electrical state quantities of each node in the power distribution system are collected in real time, a weighted directed capacity network graph is constructed, and the rated current carrying capacity is used as the edge capacity.

[0031] S02. The dual-buffered circular queue receives electrical messages from each node, performs zero-copy fast distribution through a lock-free dynamic memory pool, and synchronously starts a composite filtering algorithm to remove transient noise.

[0032] S03, the gated clock recursive adaptive topology sensing model performs feature projection and timing analysis on multi-dimensional electrical parameters, outputs the probability distribution of abnormal states and transient change rate of each node, and triggers dynamic threshold comparison.

[0033] S04. For the set of abnormal nodes that exceed the dynamic threshold, execute the network flow minimum cut pruning algorithm to solve for the minimum cut edge set and locate the source node of the power distribution fault.

[0034] S05. Using the source node corresponding to the minimum cut edge set as the root, remove all downstream derived abnormal information and retain only the fault alarm record of the source node.

[0035] S06. Push alarms to maintenance personnel, including the fault time, source fault node identifier, fault type, possible fault causes and handling solutions, and support expanding to view the details of abnormal parameters of downstream sub-nodes.

[0036] The weighted directed capacity network graph is constructed as follows: the bus or source-end circuit breaker is used as the source point, the terminal load circuit breakers at each level are used as the sink points, and the circuit breakers of each distribution branch are used as intermediate nodes. The weight of the directed edge between each node is taken as the rated current carrying capacity of the corresponding circuit breaker. The graph structure is maintained in real time by the graph attention network dynamic topology adaptive reconstruction model. When the operation mode of the power distribution equipment is adjusted due to maintenance or expansion, the model reconstructs the virtual topology connection matrix in reverse according to the instantaneous change coupling of the electrical state variables, so that the topology graph is consistent with the physical power distribution system.

[0037] Wherein, the virtual topology connection matrix is ​​a 3D sparse matrix, The matrix element represents the total number of monitoring nodes in the current power distribution system. The values ​​of the matrix elements are the real-time normalized values ​​of the directed edges between the corresponding nodes. When the normalized current value detected between nodes is lower than the connectivity threshold, the corresponding matrix element is set to zero, triggering the topology reconstruction process. The connectivity threshold is determined by simulating various operating conditions such as normal circuit breaker disconnection, fault undervoltage, and light line load in a controlled experimental environment, collecting the distribution data of the normalized current values ​​between nodes under different operating conditions, and taking the lower bound of the 95% confidence interval after multiple iterative analyses. The typical range is 0.05 to 0.15.

[0038] The dual-buffered circular queue is implemented by setting up two independent fixed-length memory buffers, with write and read pointers running in different buffers, and using atomic operations to switch pointers to avoid contention between writing and reading. The lock-free dynamic memory pool is implemented based on a linked list structure of pre-allocated fixed-size memory blocks. When a message arrives, it directly retrieves a free block from the memory pool, processes it, and returns it. No system-level memory allocation is required throughout the process, thus eliminating latency jitter in dynamic memory allocation. The zero-copy fast distribution means that during the entire transmission of message data from the acquisition gateway to the analysis thread, only memory block pointers are transmitted, not the data itself, to avoid redundant data copying.

[0039] The composite filtering algorithm consists of a sliding window time-varying median filter and a wavelet threshold denoising system connected in series; the sliding window time-varying median filter has a length of... The sliding window takes the median value of the electrical timing signal point by point. The value range is 5 to 21 sampling points, and the specific value is determined experimentally based on the on-site acquisition frequency and the expected noise suppression frequency band. Wavelet threshold denoising adopts a soft thresholding method, and the threshold is determined by estimating the signal standard deviation. The criteria are established. It is obtained by minimum absolute median deviation estimation to resist contamination of variance estimation by isolated impulses.

[0040] The gated clock recursive adaptive topology sensing model is based on deep learning, and its specific structure is as follows: the model input is the time window of each node. The multidimensional electrical parameter acquisition sequence within, with dimensions of ,in For batch size, For the number of nodes, For time steps, Each node has five feature dimensions: RMS voltage, RMS current, active power, power factor, and phase deviation. The input is linearly transformed by a pre-projection layer. Dimensional mapping to hidden dimensions Alignment Then it enters the clock-gated loop layer, which assigns an independent adaptive time step update frequency to each node. update frequency The hidden layer state is dynamically determined by the exponentially weighted average of its historical transient change rates. When a node detects that the transient change rate exceeds the dynamic threshold, a dynamic backtracking mechanism for the hidden layer state is triggered. This resets the node's hidden state to the most recent steady-state moment before the failure and sends activation signals to its neighboring nodes in the topology graph to synchronously update the hidden states of neighboring nodes. The hidden layers at each time scale are connected by a long-span residual dense jump matrix, with the jump matrix dimension being [missing information]. , The loop depth is set to 4-8, enabling lossless transfer of historical memory. The output layer is a multi-node abnormal state probability classification head, outputting four probability values ​​for each node: normal, single-phase abnormal, multi-phase abnormal, and undervoltage. The model internally embeds Dijkstra's shortest path algorithm, using the abnormal probability values ​​of each node in the topology graph as edge weights. After each inference, it automatically calculates the minimum cost path from the source node to each abnormal node, assisting in locating the fault propagation direction. The memory allocation strategy adopts a layer-by-layer gradient checkpointing approach, retaining only the minimum cost path for each node during forward propagation. The activation value of the layer, The value range is 2 to 4. Activations in other layers are recalculated during backpropagation to reduce peak memory usage. Memory allocation for inter-neuron connection weights uses a grouped sparse compressed storage format, grouping weights within the same node into blocks of 16 to 64 bytes each, storing only non-zero block indices and weight values. Batch scheduling in the data loop uses a priority queue, prioritizing nodes based on their transient rate of change; nodes with higher rates of change are queued first. Inter-layer hidden state tensors are stored as half-precision floating-point numbers to compress memory usage, restoring full precision only in the loss calculation layer. The dynamic threshold is an activation gain adjustment function. The output, Based on the average transient change rate of nodes in the current batch Current system load rate and historical false alarm rate The gain adjustment value is obtained by calculating three data points. ,in , , The weighting coefficients, summed to 1, are determined by collecting noise-free standard fault datasets on a simulated power distribution fault test bench, using gradient descent iterations for 50–200 rounds with the goal of minimizing the false alarm rate. When the dynamic threshold is tightened to 0.8 times the baseline threshold, the model becomes more sensitive to anomalies; when When the dynamic threshold remains unchanged, the baseline threshold is maintained; when In this case, the dynamic threshold is relaxed to 1.2 times the baseline threshold to suppress false alarms in low-quality data environments; the baseline threshold is calculated based on electrical parameters of various dimensions by collecting more than 30 consecutive days of fault-free operation data on-site. The interval boundary value is obtained.

[0041] The steps for establishing the training dataset of the gated clock recursive adaptive topology sensing model specifically include: simulating typical fault conditions such as single-phase undervoltage, three-phase short circuit, circuit breaker malfunction, and poor line contact on a laboratory power distribution test bench; collecting time-series data of multi-dimensional electrical parameters of each node and labeling the fault type and fault source node number; simultaneously collecting normal operating condition data as negative samples; performing composite filtering preprocessing on the collected data and dividing it into training set, validation set, and test set in an 8:1:1 ratio; and performing zero-mean unit variance normalization on each dimension parameter, with the normalized parameters obtained statistically from the training set and solidified for use in the inference stage.

[0042] The specific steps for training the gated clock recursive adaptive topology sensing model include: using the cross-entropy loss function as the classification loss, and adding a physical constraint loss term based on the transient evolution equation of the transient circuit. The physical constraint loss term requires that the hidden state entropy of the model be minimized when there is no fault, and that the hidden state entropy increase be maximized when a fault occurs. The two loss components are combined using a weighted sum. The AdamW optimizer is used, with an initial learning rate range of [range missing]. ~ The model weights are decayed using a cosine annealing strategy; the number of training rounds ranges from 100 to 500, with the minimum weighted sum of false positive and false negative rates on the validation set serving as the early stopping criterion; after training, the model is evaluated using a test set, and the model weights are quantized into half-precision floating-point numbers before being deployed to the power distribution monitoring platform.

[0043] The technical advantages of the gated clock recursive adaptive topology sensing model are as follows: the clock gating structure enables asynchronous reasoning by nodes with different update frequencies, avoiding the occupation of computing resources by low-frequency steady-state nodes and allowing high-frequency transient nodes to obtain more intensive state updates, thereby improving the instantaneous response sensitivity to faults without increasing the overall computational load; the hidden layer state dynamic backtracking mechanism eliminates the pollution of the hidden layer by abnormal free states before the fault, making the gradient flow at the moment of the fault more pure and accelerating the convergence of fault source localization; the long-span residual dense jump matrix retains historical memory at multiple time scales, enabling the model to distinguish between transient disturbances and persistent faults, reducing false alarms triggered by single sampling noise; the nested Dijkstra algorithm uses topological prior knowledge to constrain the reasoning path, reducing the search cost of the model in topology-independent directions, and compared with the pure data-driven path, the localization result is more consistent with the physical causal relationship of the power distribution system.

[0044] The principle and technical effect of the network flow minimum cut pruning localization algorithm are as follows: The algorithm models the power distribution system topology as a weighted directed capacity network graph, solves the maximum flow from the source to the sink using the Ford-Fulkerson method, and obtains the minimum cut edge set using the maximum flow minimum cut theorem; the minimum cut edge set physically corresponds to the set of the fewest circuit breakers that interrupt the current transmission of the current under the current fault, i.e., the location of the fault source; after completing the minimum cut solution, the algorithm marks all nodes downstream of the minimum cut as derived abnormal nodes and removes them from the alarm queue, retaining only the alarms of the nodes associated with the minimum cut edge set; the technical effect is that: by utilizing the strict mathematical guarantee of graph theory, under the condition that the topological relationship of the power distribution system is known, a large-scale cascading derived abnormality is converged to the minimum fault source set in a deterministic manner, overcoming the limitation of the fuzzy location boundary of the pure data-driven model in the scenario of strong topological constraints, making the alarm pruning process fully traceable and verifiable, while significantly reducing the alarm queue length and alleviating the push calculation pressure of the monitoring platform server.

[0045] The benchmark value for the dynamic threshold is obtained by continuously collecting electrical parameter data for no less than 30 days of normal, fault-free operation at the data center site or equivalent test bench, and calculating the mean and standard deviation for each parameter dimension. The interval boundary values ​​serve as the upper and lower limits of the baseline alarm thresholds for each parameter. For newly built systems where it is difficult to obtain long-term fault-free data, simulation operation data of no less than 7 days can be collected on the simulation test bench as a substitute, and on-site data can be continuously accumulated for correction and iteration after the system is put into operation.

[0046] Among them, the weight coefficients in the activation gain adjustment function , , The acquisition method is as follows: A validation dataset containing no fewer than 500 labeled fault samples is constructed on a simulated power distribution fault test bench. The objective function is the weighted sum of the false alarm rate and the false negative rate. To constrain the process, a grid search method is used in... Within the range, the three combinations are traversed with a step size of 0.05. The parameter combination corresponding to the minimum objective function is taken as the initial value, and then gradient descent is used for local fine-tuning.

[0047] In addition, as an optional implementation, the present invention also provides a computer-based intelligent alarm push system for data center power distribution monitoring. The system runs on a computer with an embedded readable storage medium, which stores program instructions. When the program instructions run on the computer, they drive the computer to execute all the above-mentioned method steps, including real-time construction and maintenance of a weighted directed capacity network graph, message reception and distribution of a double-buffered circular queue, reasoning of a gated clock recursive adaptive topology-aware model, execution of a network flow minimum cut pruning localization algorithm, and finally, generation and push of alarms.

[0048] The specific implementation of step S01 is as follows: First, a topology survey is conducted on all monitored nodes in the power distribution system. Busbars or source-end circuit breakers are identified as source points in the directed graph, and terminal load circuit breakers at each level are identified as sink points. Each distribution branch circuit breaker is used as an intermediate node. Directed edges are established between adjacent nodes according to the current flow direction, and the rated current capacity of the corresponding circuit breaker is used as the capacity weight of the edge, thus completing the initial construction of the weighted directed capacity network graph. During operation, the graph attention network dynamic topology adaptive reconstruction model continuously collects the electrical state variables of all nodes, calculates the normalized current values ​​between nodes, and writes them into the graph. A sparse virtual topology connection matrix is ​​used. When the normalized value of the current between a node pair falls below the connectivity threshold, the corresponding matrix element is set to zero, triggering the topology reconstruction process. The model inversely calculates the topology change based on the coupling of the instantaneous changes in the electrical state variables of each node, updating the edge set and weights of the directed graph to ensure that the topology graph remains dynamically consistent with the physical power distribution system. The typical reference range for the connectivity threshold is 0.05 to 0.15, determined by using data collected under multiple operating conditions in a controlled experimental environment and applying the lower bound of the 95% confidence interval.

[0049] The specific implementation of step S02 is as follows: The system sets up two independent fixed-length memory buffers to form a double-buffered circular queue. The write pointer and read pointer run in different buffers, and atomic operations are used to switch the pointers to ensure that writing and reading do not interfere with each other. The lock-free dynamic memory pool is implemented based on a linked list structure of pre-allocated fixed-size memory blocks. When the electrical message returned by each node arrives, it directly retrieves a free memory block from the memory pool. After the message is processed, the memory block is returned to the memory pool. No system-level memory allocation is required throughout the process, eliminating the latency jitter caused by dynamic memory allocation. Zero-copy fast distribution means that during the entire transmission of electrical message data from the acquisition gateway to the analysis thread, only the memory block pointer is transmitted, not the data itself, avoiding redundant data copying. After the message is distributed, a composite filtering algorithm is started synchronously to preprocess the original electrical timing signal. The composite filtering algorithm consists of two parts connected in series: a sliding window time-varying median filter and a wavelet threshold denoising filter. The sliding window time-varying median filter has a length of... The sliding window takes the median value of the signal point by point, suppressing random impulse noise. The window length is... The reference range is 5–21 sampling points; wavelet thresholding denoising uses a soft thresholding method, with the threshold obtained by estimating the minimum absolute median deviation. according to The criteria are established to resist contamination of variance estimation by isolated pulses and further suppress high-frequency transient noise.

[0050] The specific implementation of step S03 is as follows: The gated clock recursive adaptive topology sensing model receives the filtered multi-dimensional electrical parameter acquisition sequence of each node, with the input dimension being... The five-dimensional features include the RMS voltage, RMS current, active power, power factor, and phase deviation. The input sequence is first mapped to the hidden dimensions through a linear transformation by a pre-spatial projection layer. Alignment Then it enters the clock-gated loop layer. The clock-gated loop layer allocates an independent adaptive time step update frequency to each node. update frequency The hidden state is dynamically determined by the exponentially weighted mean of its historical transient change rate, allowing transient nodes to receive more frequent hidden state updates, while steady-state nodes reduce their update frequency to conserve computational resources. When the transient change rate of a node exceeds a dynamic threshold, a dynamic backtracking mechanism for the hidden layer state is triggered, resetting the node's hidden state to the most recent steady-state moment before the failure, eliminating the contamination of the hidden layer by the pre-failure detached state, and sending activation signals to adjacent nodes in the topology graph to synchronously update the hidden states of adjacent nodes. The hidden layers at each time scale are connected by a long-span residual dense jump matrix, with the jump matrix dimension being [missing information]. , The reference range is 4–8, enabling lossless transfer of historical memory. The output layer outputs four probability values ​​for each node: normal, single-phase anomaly, multi-phase anomaly, and undervoltage, forming an anomalous state probability distribution. The model internally embeds Dijkstra's shortest path algorithm, using the anomaly probability value of each node as the edge weight. After each inference, it automatically calculates the minimum cost path from the source node to each anomalous node, assisting in constraining the inference of the fault propagation direction. The dynamic threshold is adjusted by the activation gain function. Based on the current batch of nodes' average transient change rate System load rate and historical false alarm rate The gain adjustment value was calculated. ,when The threshold is tightened to 0.8 times the baseline threshold. When the baseline threshold remains unchanged, The threshold is relaxed to 1.2 times the baseline threshold, which is based on data from more than 30 consecutive days of fault-free operation in the field. The interval boundary values ​​are determined.

[0051] The specific implementation of step S04 is as follows: using the abnormal state probability values ​​of each node output in step S03 as the abnormal weights of each node in the weighted directed capacity network graph, the network flow minimum cut pruning algorithm is executed on the set of abnormal nodes that exceed the dynamic threshold. The algorithm uses the Ford-Fulkerson method to solve for the maximum flow value from the source to the sink in the weighted directed capacity network graph, and then obtains the minimum cut edge set by the maximum flow minimum cut theorem. The minimum cut edge set physically corresponds to the minimum set of circuit breakers involved in the current fault interruption current transmission, i.e., the location of the fault source.

[0052] The specific implementation of step S05 is as follows: taking the source node associated with the minimum cut edge set obtained in step S04 as the root node, traversing all its downstream nodes along the directed edge direction in the weighted directed capacity network graph, marking these downstream nodes as derived abnormal nodes and removing them all from the current alarm queue, retaining only the fault alarm record of the source node, thereby converging the large-scale cascading derived abnormalities to the minimum fault source set, and greatly shortening the alarm queue length.

[0053] The specific implementation of step S06 is as follows: The system organizes the source node fault alarm records retained in step S05 into a structured alarm message. The message includes a fault occurrence timestamp, source fault node identifier, fault type classification label, possible fault causes inferred from the gated clock recursive adaptive topology awareness model, and corresponding handling solutions. The abnormal parameter details of downstream child nodes are appended to the alarm message in an expandable form. The alarm message is sent to the maintenance personnel's terminal via a push channel. Maintenance personnel can directly obtain the fault source location results and expand the message as needed to view the electrical parameter details of each derived abnormal node, assisting in on-site fault investigation and handling decisions.

[0054] It should be noted that the key technologies of this invention include: weighted directed capacity network graph modeling technology, asynchronous inference mechanism of gated clock recursive adaptive topology-aware model, and network flow minimum cut pruning localization algorithm. Weighted directed capacity network graph modeling explicitly encodes the physical causal relationships of the power distribution system into a directed graph structure, ensuring that the subsequent minimum cut solution directly corresponds to the actual circuit breaker location, guaranteeing the physical interpretability of fault tracing. Compared to traditional alarm methods that rely solely on statistical features, it introduces topological prior constraints, fundamentally avoiding the obscuring of the fault source by downstream derived alarms. The asynchronous inference mechanism of the gated clock recursive adaptive topology-aware model allows nodes with different update frequencies to independently advance their hidden states, avoiding low-frequency steady-state nodes occupying the computational resources of high-frequency transient nodes. Simultaneously, the dynamic backtracking mechanism of the hidden layer state eliminates the contamination of pre-fault detached states, improving the accuracy of anomaly probability output. Compared to traditional synchronously updated recurrent neural networks, it exhibits higher state sensitivity at the moment of a fault. The network flow minimum cut pruning localization algorithm leverages the mathematical rigor of graph theory to deterministically converge cascading alarms to the minimum set of fault sources. Compared to rule-based or purely data-driven alarm suppression methods, it offers complete traceability. The three technologies work synergistically: topology modeling provides the correct graph structure for the minimum cut algorithm; asynchronous inference provides accurate node weights; and the minimum cut algorithm performs deterministic pruning based on the outputs of the first two. All three are indispensable, forming a complete technical loop from perception to localization to pruning.

[0055] It should be noted that when the operation mode of the power distribution system is frequently adjusted, inconsistencies may arise between the actual connection relationships of the power distribution topology and the pre-stored static topology file. This can cause the minimum cut algorithm to solve on an incorrect topology graph, outputting a minimum cut edge set that does not match the physical fault location, ultimately leading to missed or false alarms in the alarm pruning results. The reason for this technical problem is that large-scale data center power distribution systems frequently undergo circuit breaker switching, bus segment switching, and load transfer operations during daily maintenance. These operations all alter the physical power distribution topology connections. Traditional power distribution monitoring systems typically rely on manual input or periodic synchronization of topology files, failing to perceive topology changes in real time. This results in the topology graph maintained internally by the monitoring system lagging behind the actual state of the physical system. Under these conditions, all causal reasoning and algorithmic solutions based on the topology graph are built on an incorrect topology foundation, making fault location results unreliable. The usual solution to this technical problem is to periodically poll the topology file or rely on maintenance personnel to manually update the topology database to maintain topology consistency. However, periodic polling suffers from the problem of not being able to capture topology changes in a timely manner within the time window. This is especially problematic when operational adjustments and faults occur almost simultaneously, leading to topology lag and causing the minimum cut algorithm to solve on an incorrect graph structure, resulting in pruning results that do not match the actual fault location. Manual updates, on the other hand, are limited by the timeliness of operations personnel, and in emergency fault handling scenarios, it is even more difficult to guarantee the real-time nature of topology data, failing to fundamentally eliminate the topology lag problem. This invention effectively solves this technical problem by using a graph attention network dynamic topology adaptive reconstruction model to continuously monitor the normalized current values ​​between all nodes. When the normalized current values ​​between node pairs fall below the connectivity threshold, the model automatically reverse-calculates the topology connection changes based on the instantaneous changes in the electrical state variables of each node, reconstructing the virtual topology connection matrix in real time. This ensures that the weighted directed capacity network graph always remains consistent with the actual connection state of the physical power distribution system, eliminating the impact of topology lag on the correctness of the minimum cut algorithm. This dynamic reconstruction mechanism does not rely on manual input or periodic synchronization. It can drive topology updates using only the real-time electrical measurement data of each node. This mechanism ensures that the minimum cut algorithm always runs on the correct topology graph, so that the alarm pruning results still have physical reliability in the dynamic power distribution environment where the operating mode is frequently adjusted.

[0056] Specifically, the principle of this invention is as follows: This invention can solve the above-mentioned technical problems because its solution forms a complete causal chain closed loop from three levels: topology modeling, dynamic perception, and graph theory pruning. First, the weighted directed capacity network graph uses the bus or source-end circuit breaker as the source point and the terminal load circuit breakers at each level as the sink point, taking the rated current carrying capacity as the directed edge capacity, thus transforming the physical electrical causal relationship of the power distribution system into a strict mathematical directed graph structure. When the power distribution system adjusts its operating mode due to maintenance or expansion, the graph attention network dynamic topology adaptive reconstruction model reconstructs the virtual topology connection matrix in reverse based on the instantaneous changes in the coupling of electrical state variables, ensuring that the topology graph is consistent with the physical system in real time. This is the basis for the correctness of subsequent causal pruning. Second, the gated clock recursive adaptive topology perception model allocates an independent adaptive time step update frequency to each node, allowing nodes with drastic transient changes to receive more intensive hidden state updates, while steady-state nodes have a reduced update frequency to save computational resources. The hidden layer state dynamic backtracking mechanism resets the hidden state of a node to the most recent steady-state moment before the fault when a transient rate of change exceeds a dynamic threshold. This eliminates the contamination of the hidden layer by abnormal free states before the fault, making the abnormal state probability output by the model more accurately reflect the true fault initiation location. The long-span residual dense jump matrix retains historical memory across multiple time scales, enabling the model to distinguish between transient disturbances and persistent faults, reducing false alarms caused by single-sample noise. The nested Dijkstra's shortest path algorithm uses the abnormal probability of each node as edge weights to constrain the propagation of the inference path along the causal direction of the distribution topology, further improving the physical rationality of fault propagation direction localization. Finally, the network flow minimum cut pruning localization algorithm solves for the maximum flow from the source to the sink based on the Ford-Fulkerson method. The minimum cut edge set is obtained using the maximum flow minimum cut theorem. This edge set physically corresponds to the set of the fewest circuit breakers that interrupt current transmission under the current fault, i.e., the true fault source location. After solving for the minimum cut, the algorithm marks all downstream nodes of the minimum cut as derived anomaly nodes and removes them from the alarm queue, retaining only the alarms of nodes associated with the minimum cut edge set. This process has strict graph theory mathematical guarantees, deterministically converging large-scale cascading derived anomalies to the minimum fault source set. It overcomes the limitation of pure data-driven models in ambiguity of boundary location in scenarios with strong topological constraints, making the alarm pruning process fully traceable and verifiable. The synergy of these three levels enables this invention to form a mutually supportive technical closed loop between topology awareness, anomaly probability estimation, and causal pruning, fundamentally distinguishing between native fault alarms and derived alarms, and solving the core technical problems faced by existing technologies.

[0057] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0058] The specific implementation method of step S01 is as follows.

[0059] Using the busbar or source-end circuit breaker as the source point Using the terminal load circuit breakers at each level as the collection point A weighted directed capacity network diagram is constructed using the circuit breakers of each distribution branch as intermediate nodes. ,in For a set of nodes, Let be a set of directed edges. The capacity is a function of the edge capacity, where the capacity of each directed edge is taken as the rated current capacity of the corresponding circuit breaker. Virtual topology connection matrix. for A 2D sparse matrix is ​​defined as follows:

[0060] ;

[0061] In the formula, For nodes With nodes The matrix element values ​​of the directed edges between them are dimensionless normalized values. for Time Node Flow to Node The measured current value, in units of , This represents the rated current value of the corresponding circuit breaker, in units of... , This represents the total number of monitoring nodes in the current power distribution system. For node index, the value range is: to .when season And triggers topology reconstruction, in which The connectivity threshold is a dimensionless value, typically ranging from 0.05 to 0.15, determined by the distribution data of normalized current values ​​under different operating conditions collected in a controlled experiment and the lower bound of the 95% confidence interval. The attention network is based on... The real-time changes are reversed to reconstruct the topology, ensuring that the graph structure is consistent with the physical power distribution system.

[0062] The specific implementation method of step S02 is as follows.

[0063] A double-buffered circular queue uses two independent fixed-length memory buffers. Write and read pointers operate in different buffers, and atomic operations are used to switch pointers, avoiding read-write contention. A lock-free dynamic memory pool is based on a linked list structure with pre-allocated fixed-size memory blocks. When a packet arrives, it directly retrieves a free block from the memory pool, processes it, and returns it, with no system-level memory allocation throughout the process. Zero-copy fast distribution means that only memory block pointers are transmitted between the acquisition gateway and the analysis thread, not the data itself, eliminating redundant data copying. The composite filtering algorithm consists of a sliding window time-varying median filter and wavelet threshold denoising in series. The sliding window time-varying median filter is used for lengths of... The sliding window takes the midpoint point by point. The value range is 5–21 sampling points, specifically determined experimentally based on the on-site sampling frequency and the expected noise suppression frequency band. Wavelet thresholding denoising employs a soft thresholding method. The definition is as follows:

[0064] ;

[0065] In the formula, This is a wavelet domain soft threshold, with units consistent with the wavelet coefficients. This is an estimate of the signal standard deviation, with units consistent with the wavelet coefficients. It is obtained from the minimum absolute median deviation estimate, and the specific calculation is as follows:

[0066] ;

[0067] In the formula, This is the finest-scale wavelet coefficient vector. for The median of the absolute deviations of each element relative to its midpoint, 0.6745 is the proportional correction coefficient under the Gaussian distribution. This estimation method is highly resistant to isolated impulses.

[0068] The specific implementation method of step S03 is as follows.

[0069] The input dimension of the gated clock recursive adaptive topology sensing model is ,in For batch size, For the number of nodes, For time steps, It includes five dimensions: effective voltage value, effective current value, active power, power factor, and phase deviation. The preceding spatial projection layer uses linear transformation to... Dimensional mapping to hidden dimensions , record nodes In batch The input feature vector is The projection output is:

[0070] ;

[0071] In the formula, For nodes In batch The initial hidden state vector is as follows. The projection weight matrix is... Let be the projection bias vector, and the dimensions of both sides of the equation are 1. Dimensions are consistent. The clock-gated loop layer is a node. Assign adaptive update frequency It is dynamically determined by the exponentially weighted mean of its historical transient rate of change:

[0072] ;

[0073] In the formula, For nodes In the The update frequency of the time step is a dimensionless value. This is the exponentially weighted decay coefficient, a dimensionless value, with an empirical value of 0.9. For nodes In the The transient rate of change at each time step is a dimensionless value, calculated from the normalized eigendometric difference norm of adjacent time steps:

[0074] ;

[0075] In the formula, For nodes In the The 5-dimensional electrical feature vector after time step normalization. For nodes In the The 5-dimensional electrical feature vector after time step normalization. To represent the Euclidean norm, the normalization process uses the zero-mean, unit-variance method, so that... It is a dimensionless value. When Exceeding the dynamic threshold When this is triggered, the hidden layer state is dynamically backtracked, and the node is... The hidden state is reset to the most recent steady-state moment before the failure occurred. status It also sends activation signals to adjacent nodes in the topology graph to synchronously update their hidden states. The hidden layers at each time scale utilize long-span residual dense jump matrices. connect, The loop depth is the number of iterations, ranging from 4 to 8. The layer output is:

[0076] ;

[0077] In the formula, For nodes No. Hidden state vector of layer, jump matrix No. Line number Column elements are dimensionless scalars, representing the first element. Layer to the first The historical memory of each layer carries weight. For nodes No. The hidden state vector of the layer, the dimension of both sides of the equation is 1. Dimensionality is consistent. The output layer outputs four anomaly probability vectors for each node. These correspond to four states: normal, single-phase abnormal, multi-phase abnormal, and pressure loss, respectively. Dynamic threshold. Activation gain adjustment function The gain adjustment value is calculated as follows:

[0078] ;

[0079] In the formula, This is the gain adjustment value, which is a dimensionless value. The average transient rate of change of nodes in the current batch is a dimensionless value. The current system load rate is a dimensionless value, and its range is [value range missing]. , The historical false alarm rate is a dimensionless value, and its range is [value range missing]. , , , These are weighting coefficients, all of which are dimensionless values, satisfying... The range of values ​​for each coefficient is 100%. The initial values ​​were determined through 50-200 iterations of gradient descent, using a noise-free standard fault dataset collected on a simulated power distribution fault test bench, with the goal of minimizing the false alarm rate. Dynamic threshold. according to The value is adjusted in three levels, when hour, ;when hour, ;when hour, ,in For nodes The baseline threshold is a dimensionless value, calculated from normalized electrical parameters across all dimensions based on over 30 consecutive days of fault-free operation data from the field. The interval boundary value is obtained. For nodes The mean of the corresponding normalized parameter, For nodes The standard deviation of the corresponding normalized parameters. The model internally embeds the Dijkstra shortest path algorithm, using the anomaly probability value of each node as the edge weight, and automatically calculates the path from the source node after each inference step. The minimum cost path to each anomalous node is used to assist in locating the direction of fault propagation. During training, the loss function is a weighted sum of cross-entropy classification loss and physical constraint loss. The physical constraint loss requires that the hidden state entropy of the model be minimized when there is no fault and maximized when a fault occurs. Training uses the AdamW optimizer, with an initial learning rate ranging from [value missing]. ~ The model weights are decayed using a cosine annealing strategy, with a training round range of 100 to 500 rounds. The early stopping criterion is to minimize the weighted sum of the false positive rate and the false negative rate on the validation set. After training, the model weights are quantized into half-precision floating-point numbers and then deployed to the power distribution monitoring platform.

[0080] The specific implementation method of step S04 is as follows.

[0081] For the set of abnormal nodes exceeding a dynamic threshold, the network flow minimum cut pruning algorithm is executed. The algorithm uses a weighted directed capacity network graph. As input, the Ford-Folkson method is used to solve for the problem starting from the source point. to the remittance point Maximum Flow The minimum cut edge set is obtained by the maximum flow minimum cut theorem. The maximum flow and minimum cut equations are as follows:

[0082] ;

[0083] In the formula, Maximum flow value, in units of , A set of cuts, in which For including source point a subset of nodes For including remittance points a subset of nodes For the set of all cuts, , For the endpoint nodes of a directed edge, For directed edges The capacity, i.e., the rated current carrying capacity of the corresponding circuit breaker, is expressed in units of... The dimensions of both sides of the equation are Minimal cut edge set The minimum set of circuit breakers that physically corresponds to the current interruption of the fault transmission, i.e., the location of the fault source.

[0084] The specific implementation method of step S05 is as follows.

[0085] minimum cut edge set The associated source node is used as the root. All downstream derived abnormal nodes are marked by breadth-first traversal, and the derived abnormal records are removed from the alarm queue. Only the fault alarm records of the source node are retained, which greatly reduces the length of the alarm queue.

[0086] The specific implementation method of step S06 is as follows.

[0087] The system pushes alarms to operations and maintenance personnel, including the time of the failure, the identifier of the source failure node, the type of failure, the possible causes of the failure and the handling solution. It also supports expanding to view the details of abnormal parameters of downstream sub-nodes, realizing hierarchical presentation and source tracing of alarm information.

[0088] To better understand and implement this invention, the following is a specific application scenario of this invention, Example 2:

[0089] This embodiment uses a data center power distribution system as the test scenario. The system has 32 monitored nodes: one source-end circuit breaker node as the source, eight distribution branch circuit breaker nodes as intermediate nodes, and 23 terminal load circuit breaker nodes as sinks. Each node constructs a weighted directed capacity network graph with its rated current capacity as the edge capacity. A graph attention network dynamic topology adaptive reconstruction model maintains the virtual topology connection matrix in real time. The matrix dimension is 32×32, and the connectivity threshold is set to 0.10.

[0090] During the system initialization phase, 30 days of fault-free operation data were collected for each node. The mean and standard deviation of five electrical parameters—RMS voltage, RMS current, active power, power factor, and phase deviation—were calculated. The boundary values ​​of the interval determine the baseline alarm thresholds for each parameter. Taking the effective value of current as an example, the calculation results of the baseline alarm thresholds for each node are shown in Table 1.

[0091] Table 1. Reference alarm threshold table for RMS current values ​​at each node

[0092]

[0093] The weighting coefficients in the activation gain adjustment function are determined through a grid search, and the final value is... , , .

[0094] The test scenario is set as follows: During a certain operating cycle, the circuit breaker N03 corresponding to branch 2 experiences a single-phase contact failure. After the failure occurs, the effective current values ​​of the five downstream terminal nodes N11, N12, N13, N14, and N15 all fluctuate abnormally, exceeding their respective dynamic thresholds. At the same time, the phase deviation parameter of N03 itself also exceeds the reference threshold.

[0095] After receiving electrical messages from the aforementioned six nodes, the double-buffered circular queue performs zero-copy distribution via a lock-free dynamic memory pool. A composite filtering algorithm preprocesses the electrical timing signals of each node, and the window length of the sliding window time-varying median filter is determined. Set to 11 sampling points, wavelet threshold denoising Obtained by minimum absolute median deviation estimation and then... The criteria determine a soft threshold to filter out transient pulse interference.

[0096] The gated clock recursive adaptive topology sensing model infers the multidimensional electrical parameter sequences of 6 abnormal nodes and outputs the probability distribution of abnormal states of each node as shown in Table 2.

[0097] Table 2 Probability Distribution of Abnormal States for Each Abnormal Node

[0098]

[0099] like Figure 2 As shown, the Dijkstra shortest path algorithm nested within the model uses the anomaly probability value of each node as the edge weight to calculate the minimum cost path from the source node N01 to each anomaly node. The results show that the path cost from N01 to N11~N15 via N03 is higher than the path cost to directly reach N03, indicating that N03 is a critical node on the fault propagation path.

[0100] Average transient rate of change of nodes in the current batch System load rate Historical false alarm rate Calculate the gain adjustment value , The dynamic threshold remains unchanged compared to the baseline threshold.

[0101] For the set of 6 abnormal nodes that exceed the dynamic threshold, the network flow minimum cut pruning algorithm is executed. The maximum flow is solved on the weighted directed capacity network graph using the Ford-Fulkerson method. The minimum cut edge set is obtained by the maximum flow minimum cut theorem. The fault source node corresponding to the minimum cut edge set is N03, which is the location of the single-phase contact failure fault at the circuit breaker of branch 2.

[0102] With N03 as the root node, traverse its downstream nodes N11, N12, N13, N14, and N15, mark them all as derived abnormal nodes and remove them from the alarm queue. Finally, the alarm queue only retains the fault alarm record of N03. The length of the alarm queue before and after alarm pruning is compressed from 6 to 1.

[0103] The system pushes an alarm to maintenance personnel, including a fault timestamp, the source fault node identifier N03, the fault type as single-phase anomaly, the possible cause as poor line contact, and suggested solutions. Maintenance personnel can then view detailed information on the abnormal electrical parameters of downstream nodes N11 to N15 to assist in on-site troubleshooting.

[0104] From a technical principle perspective, the beneficial effects are as follows: The single-phase anomaly probability of N03 is 0.87, which is much higher than the single-phase anomaly probabilities of downstream nodes N11 to N15 (all below 0.56). This difference reflects that the dynamic backtracking mechanism of the hidden layer state in the gated clock recursive adaptive topology sensing model eliminates the contamination of pre-fault detached states, making the anomaly probability output of the fault source node more concentrated, while the anomaly probability distribution of downstream nodes is relatively dispersed due to the passive influence of electrical coupling. The minimum cut algorithm utilizes the topological constraints of the weighted directed capacity network graph to transform this probability difference into a deterministic graph theory pruning conclusion, enabling derived anomaly nodes to be systematically removed from the alarm queue without relying on manual experience or threshold adjustment, which is something that existing fixed threshold methods cannot achieve. As the scale of the power distribution system expands and the number of downstream nodes increases, the proportion of derived alarms in the alarm queue will continue to increase. The minimum pruning mechanism of this invention requires only a linear relationship with the number of edges in the graph, which makes the alarm compression capability more obvious as the system scale increases, ensuring that the push calculation pressure of the monitoring platform server in the scenario of large-scale alarm storm is within a controllable range.

[0105] It should be noted that the variables involved in this invention are explained in detail in Tables 3 and 4.

[0106] Table 3. Variable Explanation Table (Part 1)

[0107]

[0108] Table 4. Variable Explanation Table (Part Two)

[0109]

[0110] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent alarm push notification applied to a data center power distribution monitoring system, characterized in that, Includes the following steps: Based on the graph attention network dynamic topology adaptive reconstruction model, the electrical state quantities of each node in the power distribution system are collected in real time, a weighted directed capacity network graph is constructed, and the rated current carrying capacity is used as the edge capacity. A double-buffered circular queue receives electrical messages from each node, performs zero-copy fast distribution via a lock-free dynamic memory pool, and synchronously starts a composite filtering algorithm to remove transient noise. The gated clock recursive adaptive topology sensing model performs feature projection and timing analysis on multi-dimensional electrical parameters, outputs the probability distribution of abnormal states and transient change rate of each node, and triggers dynamic threshold comparison. For the set of abnormal nodes that exceeds the dynamic threshold, the network flow minimum cut pruning localization algorithm is executed to solve for the minimum cut edge set and locate the source node of the power distribution fault. Using the source node corresponding to the minimum cut edge set as the root, remove all downstream derived abnormal information and retain only the fault alarm record of the source node. Alarms are pushed to operations and maintenance personnel, including the time of failure, the identifier of the source node, the type of failure, the possible causes of failure and the handling solution, and can be expanded to view the details of abnormal parameters of downstream sub-nodes.

2. The intelligent alarm push method according to claim 1, characterized in that, The construction of the weighted directed capacity network graph is specifically based on the bus or source-end circuit breaker as the source point, the terminal load circuit breakers at all levels as the sink points, and the circuit breakers of each distribution branch as intermediate nodes. The weight of the directed edge between each node is taken as the rated current carrying capacity of the corresponding circuit breaker. The graph attention network dynamic topology adaptive reconstruction model reconstructs the virtual topology connection matrix in reverse based on the instantaneous changes and coupling of electrical state variables, so that the topology graph is consistent with the physical power distribution system.

3. The intelligent alarm push method according to claim 2, characterized in that, The virtual topology connection matrix is ​​an N×N sparse matrix, where N is the total number of monitoring nodes in the current power distribution system, and the matrix elements are the normalized values ​​of the real-time current of the directed edges between the corresponding nodes. When the normalized value of the current between nodes is lower than the connectivity determination threshold, the corresponding matrix element is set to zero, triggering the topology reconstruction process. The connectivity determination threshold is determined by simulating various working conditions in a controlled experimental environment, collecting the distribution data of normalized current values ​​between nodes, and then iteratively analyzing the lower bound of the 95% confidence interval.

4. The intelligent alarm push method according to claim 3, characterized in that, The dual-buffered circular queue sets up two independent fixed-length memory buffers, with write pointers and read pointers running in different buffers, and pointer switching is completed using atomic operations; the lock-free dynamic memory pool is implemented based on a linked list structure of pre-allocated fixed-size memory blocks, and when a message arrives, it directly retrieves a free block from the memory pool, returns it after processing; zero-copy fast distribution means that only memory block pointers are transmitted, rather than the data itself, during the entire transmission of message data from the acquisition gateway to the analysis thread.

5. The intelligent alarm push method according to claim 4, characterized in that, The composite filtering algorithm consists of a sliding window time-varying median filter and a wavelet threshold denoising component connected in series; the sliding window time-varying median filter has a length of... The sliding window takes the median of the electrical timing signal point by point; the wavelet thresholding denoising adopts the soft thresholding method, and the threshold is determined by the signal standard deviation estimate according to the 3σ criterion, and σ is obtained by the minimum absolute median deviation estimate.

6. The intelligent alarm push method according to claim 5, characterized in that, The input of the gated clock recursive adaptive topology sensing model is the multi-dimensional electrical parameter acquisition sequence of each node within the time window T, which includes five features: effective voltage value, effective current value, active power, power factor, and phase deviation. After linear transformation by the pre-spatial projection layer, the input enters the clock-gated loop layer, and each node is assigned an independent adaptive time step update frequency. The hidden layers of each time scale are connected by a long-span residual dense jump matrix. The output layer is a multi-node abnormal state probability classification head.

7. The intelligent alarm push method according to claim 6, characterized in that, In the clock-gated loop layer, the update frequency of a node is dynamically determined by the exponentially weighted average of its historical transient change rate. When the transient change rate of a node exceeds the dynamic threshold, the hidden layer state dynamic backtracking mechanism is triggered to reset the hidden state of the node to the most recent steady state before the fault occurred, and an activation signal is sent to the adjacent nodes in the topology graph to synchronously update the hidden state of the adjacent nodes.

8. The intelligent alarm push method according to claim 7, characterized in that, The gated clock recursive adaptive topology sensing model internally nests the Dijkstra shortest path algorithm, using the anomaly probability value of each node in the topology graph as the edge weight. After each inference, it automatically calculates the minimum cost path from the source point to each abnormal node to help locate the direction of fault propagation.

9. The intelligent alarm push method according to claim 8, characterized in that, The establishment of the training dataset for the gated clock recursive adaptive topology sensing model includes: simulating typical fault conditions such as single-phase undervoltage, three-phase short circuit, circuit breaker malfunction, and poor line contact on a laboratory power distribution test bench; collecting time-series data of multi-dimensional electrical parameters of each node and labeling the fault type and fault source node number; simultaneously collecting normal operating condition data as negative samples; performing composite filtering preprocessing on the collected data; dividing the data into training, validation, and test sets in an 8:1:1 ratio; and performing zero-mean unit variance normalization on the parameters of each dimension.

10. The intelligent alarm push method according to claim 9, characterized in that, The network flow minimum cut pruning localization algorithm uses the Ford-Fulkerson method to solve for the maximum flow from the source to the sink in a weighted directed capacity network graph, and obtains the minimum cut edge set by the maximum flow minimum cut theorem; The minimum cut edge set physically corresponds to the minimum set of circuit breakers that interrupt the transmission of current during the current fault. After the algorithm completes the minimum cut solution, it marks all downstream nodes of the minimum cut as derived abnormal nodes and removes them from the alarm queue, retaining only the alarms of the nodes associated with the minimum cut edge set.