A low-voltage cabinet assembly fault root cause tracing and self-healing method based on digital twinning

Abstract

The application discloses a low-voltage cabinet assembly fault root cause tracing and self-healing method based on digital twinning, comprising the following steps: time stamp synchronization and interpolation of multi-modal time series data are performed by using a physical information constraint model to generate a high-fidelity digital twinning state matrix; multi-dimensional static atlases are initialized by fusing electrical, thermal radiation and electromagnetic rules, and the directed edge weights thereof are dynamically evolved based on vector autoregressive residual statistics; when an alarm is triggered, hidden physical root cause nodes are accurately extracted by using atlas reverse transposition and random walk algorithm with restart, and a subgraph is extracted accordingly; a multi-agent deep deterministic policy gradient model is used for multi-track deduction in a digital twinning sand table, and a flexible self-healing control instruction with the maximum cumulative reward value is issued. The application penetrates the concurrent alarm appearance, realizes flexible isolation and global dynamic optimal regulation and control under the premise of uninterrupted power supply.

Description

Technical Field

[0001] This invention relates to the field of digital twin technology, and in particular to a method for tracing the root cause of faults and self-healing of low-voltage switchgear components based on digital twins. Background Technology

[0002] With the rapid development of new power systems and the Industrial Internet of Things (IIoT), low-voltage switchgear, as the core hub for power distribution and control at the end of the power grid, directly affects the safety of electricity use in industrial production and residential communities. In recent years, digital twin technology and artificial intelligence algorithms have been widely introduced into the intelligent operation and maintenance management of electrical equipment. Existing intelligent operation and maintenance systems for low-voltage switchgear typically utilize various sensors deployed within the cabinet to collect current, voltage, and temperature data, and construct corresponding digital models in a virtual space. In terms of fault diagnosis, mainstream technologies largely rely on predefined threshold-based alarm rules or employ conventional feedforward deep learning models. These models perform forward pattern recognition on single-modal or simply pieced-together operational data to monitor, classify, and execute intervention actions for abnormal states of electrical equipment.

[0003] However, existing fault diagnosis and self-healing technologies exhibit significant limitations when faced with the extremely compact physical space and complex electromagnetic-thermal coupling environment inside low-voltage switchgear. First, the inherent sampling rate differences and communication delays of multi-source heterogeneous sensors result in a lack of microsecond-level high-fidelity alignment mechanisms for the input system's time series, directly impacting causal logic misjudgments in subsequent models. Second, traditional diagnostic models are mostly based on fixed static wiring topologies, failing to capture the dynamic cascading fault evolution process caused by surges in thermal radiation or electromagnetic interference under actual operating conditions. This means that when faced with concurrent alarms from multiple nodes, existing systems can only identify the superficial symptoms at the end, making it difficult to reverse-engineer the complex spatiotemporal coupling delays and accurately trace the underlying hidden physical causes. Furthermore, existing self-healing strategies often employ a single, brute-force "threshold trigger - relay trip" hard isolation logic, lacking the ability to perform multi-dimensional strategy deduction in virtual space and global flexible control under uninterrupted power supply conditions, easily leading to power grid oscillations and irreversible secondary power outages. Summary of the Invention

[0004] The purpose of this section is to outline some aspects of embodiments of the present invention and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents; however, such simplifications or omissions should not be construed as limiting the scope of the invention.

[0005] In view of the aforementioned existing problems, this invention is proposed. Therefore, this invention provides a method for tracing the root causes and self-healing of low-voltage switchgear component failures based on digital twins, to solve the problems mentioned in the background art.

[0006] To address the aforementioned technical problems, this invention provides the following technical solution: a method for tracing the root cause of faults and achieving self-healing of low-voltage switchgear components based on digital twins, comprising: Acquire multimodal time-series observation data of physical low-voltage switchgear components, input the multimodal time-series observation data into a pre-constructed physical information constraint model for timestamp synchronization and missing interpolation, and generate a digital twin state matrix; Extract the node feature vectors from the digital twin state matrix, and initialize the static adjacency weight matrix of the multidimensional heterogeneous causal knowledge graph according to the preset low-voltage cabinet component spatial topology and multi-physics coupling rules. Multivariate time series features within a sliding time window are extracted, and a residual statistic based on vector autoregression is constructed. Based on the test results of the residual statistic, the directed edge weights of the multidimensional heterogeneous causal knowledge graph are dynamically smoothed and evolved. When the target component node is detected to have triggered the preset alarm threshold, the dynamic directed edge weight matrix at the current moment is frozen and the matrix is ​​transposed. Based on the random walk probability iteration algorithm with restart, the node with the largest probability steady-state convergence value is extracted from the reverse transition matrix as the physical root cause node. In the digital twin space, an isolated subgraph containing the physical root cause node is extracted to construct a Markov decision environment. A multi-agent reinforcement learning algorithm is used to perform multi-trajectory replay simulation in the Markov decision environment to generate the control instruction sequence with the maximum cumulative reward value and issue it for execution.

[0007] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By introducing a physical information neural network (PINN) with an embedded physical energy conservation law and combining it with the DTW algorithm, instead of simply relying on data smoothing, it uses high-frequency electrical excitation features to drive the deduction, achieving absolute alignment and high-fidelity interpolation of high- and low-frequency heterogeneous data at the microsecond time cross section, thereby achieving zero time error for traceability.

[0008] 2. This invention abandons the traditional static topology that relies solely on electrical wiring. It integrates three-dimensional physical fields—electrical admittance, spatial thermal radiation, and mutual inductance electromagnetic interference—to generate a multidimensional heterogeneous causal spectrum, and combines online residual statistical tests for dynamic edge weight smoothing. Combined with a restart-enabled random walk probability iterative algorithm (RWR) and its inverse transpose matrix, it can penetrate massive concurrent alarm phenomena and accurately pinpoint the hidden physical root causes caused by airborne heat transfer or electromagnetic coupling.

[0009] 3. Furthermore, this invention avoids the power flow oscillations caused by traditional "one-size-fits-all" hard isolation tripping. It constructs an MDP environment by extracting the physical root causes from the isolation subgraph in the digital twin space and uses a deterministic policy gradient model for multi-track extrapolation. Combined with a multi-dimensional reward function that includes positive power protection, over-temperature prediction penalties, and mechanical depreciation, it intelligently outputs control commands ranging from fan cooling and flexible current limiting to precise tripping within the upper limit of ensuring physical safety, thereby improving power supply continuity and system self-healing capabilities. Attached Figure Description

[0010] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 This is an overall flowchart of a digital twin-based method for tracing the root cause of low-voltage switchgear component failures and for self-healing, as described in one embodiment of the present invention. Detailed Implementation

[0011] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0012] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0013] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0014] This invention is described in detail with reference to the schematic diagrams. When detailing the embodiments of this invention, for ease of explanation, the cross-sectional views illustrating the device structure may be partially enlarged, not adhering to the usual scale. Furthermore, the schematic diagrams are merely examples and should not be construed as limiting the scope of protection of this invention. In actual fabrication, the three-dimensional spatial dimensions of length, width, and depth should be included.

[0015] Furthermore, in the description of this invention, it should be noted that the terms "upper," "lower," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. These terms are used solely for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. In addition, the terms "first," "second," or "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0016] Unless otherwise explicitly specified and limited, the terms "installation," "connection," and "joining" in this invention should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; similarly, they can refer to mechanical connections, electrical connections, or direct connections, or indirect connections through an intermediate medium, or internal connections between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0017] Example 1 Reference Figure 1 This is the first embodiment of the present invention, which provides a method for tracing the root cause of faults and achieving self-healing of low-voltage switchgear components based on digital twins, including: S1. Obtain multimodal time-series observation data of physical low-voltage switchgear components, input the multimodal time-series observation data into the pre-constructed physical information constraint model for timestamp synchronization and missing interpolation, and generate a digital twin state matrix.

[0018] It should be noted that in the actual physical operating environment of low-voltage distribution cabinets, due to the compact internal space and strong electromagnetic-thermal coupling effects, single-dimensional data cannot comprehensively characterize their health status. Therefore, this embodiment first acquires multimodal time-series observation data in real time, including high-frequency current sequences, voltage sequences, low-frequency infrared temperature sequences, and acoustic emission (partial discharge) sequences, through a heterogeneous sensor network deployed within the physical low-voltage cabinet. Secondly, due to the fundamental differences in the hardware sampling mechanisms and communication protocols of various sensors, the acquired data exhibits severe inconsistencies in sampling rates (for example, the sampling rate of a power sensor can reach 10kHz, while infrared array temperature measurement is typically limited to only 1Hz to 10Hz due to the limitations of thermal imaging components). This data gap between high- and low-frequency modes, along with unavoidable network packet loss, leaves a large number of missing data points on the timeline. Directly inputting these into conventional diagnostic models would lead to serious causal logic confusion. To overcome the shortcomings of the prior art, this step introduces and constructs a Physical Information Neural Network (PINN) as a data interpolation and reconstruction engine for the digital twin. To determine the intrinsic relationships among multimodal input data, this invention sets the input layer of PINN to contain a two-dimensional feature tensor: the first feature tensor is the spatial dimension coordinates. and target timestamp The second feature tensor is at that timestamp High-frequency and complete real-time current observations in the vicinity The output layer of PINN is then set to the predicted temperature value at the corresponding spatiotemporal point. .

[0019] Specifically, the backbone network of this model consists of 5 fully connected layers, using Tanh as the nonlinear activation function to ensure smooth calculation of physical derivatives. Furthermore, the multiphysics evolution within the low-voltage cabinet is embedded in the model's loss function in the form of partial differential equations (PDEs). During PINN training, the Adam optimizer is used for initial backpropagation gradient descent with an initial learning rate of 0.001. After the loss value decreases gradually, the L-BFGS second-order optimization algorithm is switched for fine-tuning until the loss function reaches its minimum convergence.

[0020] Furthermore, the loss function of this physical information neural network Data fitting error term Penalty terms in partial differential equations The weighted average is calculated using the following formula: in, This indicates the number of samples with actual infrared temperature sampling points. This is the actual measured temperature value. This is an adaptive weighting coefficient with a value range of (0,1], used to balance the gradient contribution of data-driven and physical constraints. This represents the number of randomly sampled configuration points in the spatiotemporal domain, used to force the network to obey physical laws even at non-sampled points. This represents the first derivative of the network-predicted temperature with respect to time, reflecting the rate of change of temperature over time. It is the thermal diffusivity of low-voltage switchgear components (such as copper busbars and contactor contacts). This is the spatial heat diffusion term (Laplace operator), representing the conduction and diffusion of heat in three-dimensional space. These are high-frequency current observations acquired synchronously. This is the equivalent contact resistance of the component; For component material density; Specific heat capacity. This is the Joule heat source term derived based on Joule's law.

[0021] It should be noted that traditional pure data-driven interpolation methods (such as linear interpolation, spline interpolation, or conventional RNN networks) often only blindly smooth low-frequency temperature data based on historical curve trends, failing to detect the transient thermal effects caused by sudden current changes at the current moment. However, this invention, by configuring the above... The penalty term is equivalent to injecting prior physical laws into the neural network. This penalty term forces the network to predict that the rate of temperature change must be strictly equal to the difference between heat conduction dissipation and Joule heat source input (i.e., the law of conservation of physical energy). Since the current sequence is high-frequency, continuous, and without gaps, the model can utilize the high-density current Joule thermal excitation characteristics to deduce and complete the missing state points of the low-frequency temperature sequence at the microsecond time cross-section, thereby achieving high-fidelity up-dimensional interpolation across frequency modes.

[0022] Furthermore, although PINN completes the data points, due to the jitter of multi-node communication in the underlying industrial Ethernet, there may still be slight nonlinear timestamp misalignments between multimodal sequences. Therefore, after interpolation, we need to perform absolute alignment of the multiple sequences based on the Dynamic Time Warping (DTW) algorithm. The specific processing procedure is as follows: Calculate the cumulative distance matrix between each completed feature sequence and the preset digital twin reference clock sequence. , where any element in the matrix This represents the shortest warped distance from one point in the sequence to another. Subsequently, a dynamic programming algorithm is used to backtrack through the cumulative distance matrix to find the optimal backtracking path that minimizes the global matching cost. Based on this optimal backtracking path, the timestamps of all sequences are elastically stretched or compressed to forcibly smooth out the nonlinear time warping caused by network jitter.

[0023] It should be noted that, after the above-mentioned missing interpolation based on physical constraints and timestamp alignment based on DTW, this invention can reconstruct the originally uneven, asynchronous and incomplete original multi-source data into multi-dimensional tensors that are absolutely matched with each other at every microsecond time segment, and finally output a high-fidelity digital twin state matrix. (in, This represents the total number of nodes in the monitoring components within the low-voltage cabinet. (The aligned multimodal feature dimensions).

[0024] S2. Extract the node feature vectors from the digital twin state matrix, and initialize the static adjacency weight matrix of the multidimensional heterogeneous causal knowledge graph based on the preset spatial topology and multi-physics coupling rules of the low-voltage cabinet components.

[0025] It should be noted that after obtaining the high-fidelity digital twin state matrix through S1... Then, the matrix is ​​sliced ​​in the spatial dimension to extract the node feature vectors corresponding to each key component (such as circuit breaker, contactor, thermal relay, copper busbar, etc.) in the physical low-voltage cabinet. In order to accurately recreate the complex fault transmission mechanism between these components in virtual space, this invention abandons the conventional approach of relying solely on a single electrical wiring diagram to construct the topology, and introduces a three-dimensional multi-physics coupling rule that includes electrical connectivity, thermal radiation, and electromagnetic interference, thereby initializing the static adjacency weight matrix of the multi-dimensional heterogeneous causal knowledge graph.

[0026] Furthermore, firstly, the electrical admittance between nodes is calculated based on the circuit topology to generate an electrical connectivity matrix. .

[0027] Specifically, in low-voltage switchgear, overload current, short-circuit surges, or harmonic voltages are primarily physically conducted along the conductive loop. According to Kirchhoff's laws and Ohm's law, the electrical conduction strength between nodes is inversely proportional to their impedance. Therefore, this invention calculates the electrical conduction strength between any two component nodes based on a pre-imported digital twin circuit schematic. and Equivalent electrical admittance between The electrical connectivity matrix is ​​constructed by normalizing the matrix elements. The calculation formula is: It should be noted that this matrix represents the conduction path of hard-wired faults caused by current overload or voltage drop. A larger admittance value (i.e., a smaller resistance / impedance) indicates lower losses during the propagation of electrical anomalies between the two components, and a stronger causal relationship.

[0028] Furthermore, the spatial centroid distance between nodes in the digital twin 3D model is extracted, and a thermal radiation coupling matrix is ​​generated through Gaussian kernel function mapping. .

[0029] Specifically, the internal space of a low-voltage switchgear is extremely compact. When a component (such as a loose busbar connector) experiences an abnormal temperature rise, even without electrical connection to surrounding components, it can cause nearby components (such as the trip unit of a nearby miniature circuit breaker) to malfunction due to heat through intense thermal radiation and convection. To quantify this implicit spatial coupling, this invention extracts the three-dimensional spatial centroid coordinates of the component nodes from a three-dimensional digital twin model and calculates the Euclidean distance. And it uses a Gaussian kernel function to map it into thermal radiation coupling weights, whose matrix elements The calculation formula is: in, The dimensional parameter characterizing the range of heat diffusion inside the low-voltage switchgear is dynamically and adaptively calibrated based on the cooling fan speed and heat dissipation channel structure inside the switchgear.

[0030] It should be noted that because the Gaussian kernel function has the inherent mathematical property of exponential decay with distance, it perfectly matches the physical law in thermodynamics that the intensity of thermal radiation decreases inversely with spatial distance. Therefore, this matrix can effectively compensate for the technical blind spot of traditional electrical topologies that cannot detect cascading faults caused by heat conduction over distance.

[0031] Furthermore, an electromagnetic interference matrix is ​​then generated based on the mutual inductance coefficients between component nodes. .

[0032] Specifically, high-frequency transient surge currents are generated at the moment of inverter startup or the opening and closing of large contactors. According to Faraday's law of electromagnetic induction, this would couple extremely high interference voltages onto adjacent parallel low-voltage control lines in space, causing the secondary circuit controller to malfunction or issue false commands. Based on this, this invention uses the Biot-Savart law to calculate the mutual inductance coefficient of adjacent components (or their connecting wires) in space. Generate an electromagnetic interference matrix, whose matrix elements are... The calculation formula is: It should be noted that the mutual inductance coefficient can be seen in the above electromagnetic interference matrix formula. This is strongly correlated with the parallelism and proximity of two components / wires in three-dimensional space. Therefore, by introducing this matrix, the knowledge graph can accurately capture the root causes of cross-level (primary main circuit interference with secondary control circuit) logical faults caused by high-frequency electromagnetic radiation.

[0033] Furthermore, finally, a linear weighted sum is performed on the three individual physics field matrices to generate the initial static adjacency weight matrix. .

[0034] Specifically, the matrices representing different physical conduction mechanisms are fused using multi-dimensional features, and the calculation formula is as follows: in, These are multiphysics weighting coefficients, and satisfy... .

[0035] It should be noted that, through the above processing, the initial static adjacency weight matrix constructed by this invention is not only a simple network reflecting circuit connections, but also a multidimensional heterogeneous causal topological graph that deeply integrates electrical conduction, spatial thermodynamic diffusion, and high-frequency electromagnetic induction. This graph provides a graph structure foundation consistent with the actual operating logic of the underlying physical world for subsequent use of deep learning algorithms to accurately track and locate hidden fault roots that appear to have no electrical connections but are actually conducted through heat or magnetism.

[0036] S3. Extract the multivariate time series features within the sliding time window, construct a residual statistic based on vector autoregression, and dynamically smooth the weights of the directed edges of the multidimensional heterogeneous causal knowledge graph based on the test results of the residual statistic.

[0037] It should be noted that the initial static adjacency weight matrix constructed in S2 This paper characterizes the potential physical channels or maximum theoretical coupling capacity for cascading faults between components within a low-voltage switchgear solely from the perspective of three-dimensional physical space and inherent electrical connections. However, under actual operating conditions, the occurrence and propagation of faults is a highly dynamic evolutionary process (for example, the heating process caused by increased contact resistance due to aging of contactors is slow, while the electromagnetic mutation caused by a short circuit is instantaneous). If fault tracing is performed solely based on static topology, it is easily interfered with by inactive physical channels, leading to misjudgments. Therefore, this invention introduces Granger causality analysis logic based on time series, utilizing online sliding time-series observation data to perform real-time activation and dynamic pruning of the static physical topology.

[0038] Specifically, the length is set to [length value] in the time dimension. The sliding time window, from the high-fidelity digital twin state matrix output by S1 In the middle, extract the source node of the directed edge connection. With the target node State feature sequence within the sliding time window and .

[0039] Furthermore, a restricted autoregressive model is constructed. This model assumes that the target node... The current state is completely independent of the source node. Only using the target node Predicting the current state by its own historical state sequence The calculation formula is as follows: in, The order of the autoregressive model represents the depth of tracing back to historical time steps; The autoregressive coefficients are based solely on their own historical data.

[0040] Furthermore, based on the above formula, the constrained sum of squared residuals between the predicted value and the actual observed value is calculated. : It needs to be explained that this restricted residual sum of squares This means that if the source node is not considered... For the target node When the target node is affected The baseline error for predicting its own operating status.

[0041] Furthermore, an unrestricted autoregressive model is constructed. This model jointly utilizes the source nodes. With the target node The historical state sequences of both can be used to jointly predict the target node. Current state The calculation formula is as follows: in, The regression coefficients of the target node itself; Let be the cross-regression coefficients between the source node and the target node. Similarly, calculate the unrestricted sum of squared residuals. get: Furthermore, if the source node A fault has indeed occurred, and it is being transmitted to the target node via thermal radiation or electrical loops. Conduction (e.g.) The abnormal temperature rise led to (The temperature begins to rise), then the source node The historical sequence must contain prediction nodes. The key incremental information of the current state. At this point, the prediction accuracy of the unrestricted model, after incorporating the source node information, will be significantly improved, leading to... Significantly smaller than Based on this logic, the difference between the restricted residual sum of squares and the unrestricted residual sum of squares is calculated, and the ratio of this difference to the unrestricted residual sum of squares is used as a statistical test metric for evaluating the strength of causal transmission from the source node to the target node. ,get: Furthermore, the directed edge weights of the multidimensional heterogeneous causal knowledge graph are dynamically and smoothly evolved.

[0042] Specifically, the above statistical test measures Mapped to open intervals using a sigmoid activation function with threshold bias. Causal activation coefficient within : in, The causal significance threshold is determined based on the white noise variance under normal operating conditions of the low-voltage switchgear. The mapping slope parameter controls the steepness of activation.

[0043] It should be noted that this causal activation coefficient acts as a soft switch, effectively filtering out spurious causal relationships caused by minor fluctuations under normal operating conditions. Only when the statistical test value exceeds the threshold will the causal relationship be effectively filtered out. Only then is it considered that fault propagation has actually occurred, and the output approaches [a certain value]. The activation coefficient.

[0044] Furthermore, finally, we only need to use the Exponential Moving Average (EMA) mechanism to update the directed edge dynamic weight matrix at the current time step. That's it. The update rules are as follows: in, To smooth the update rate, the evolutionary inertia of the map weights is determined; The weight of the directed edge in the previous time step.

[0045] It should be noted that in the above formula, the static matrix acts as the physical constraint channel, while the real-time causality coefficient acts as the data-driven valve. The product of the two ensures that the high-weighted connections in the graph must simultaneously satisfy two stringent conditions: (1) they must be feasible for transmission in physical space or electrical wiring; (2) they must exhibit a strong Granger causality effect in the real-time multimodal sequence of the current sliding time window. At the same time, the introduction of the EMA mechanism effectively absorbs the sensor impulse noise that occasionally occurs in the industrial field, avoiding severe oscillations in the entire causal graph structure due to single-point data jumps.

[0046] S4. When the target component node is detected to have triggered the preset alarm threshold, the dynamic directed edge weight matrix at the current moment is frozen and the matrix is ​​transposed. Based on the random walk probability iteration algorithm with restart, the node with the largest probability steady-state convergence value is extracted from the reverse transition matrix as the physical root cause node.

[0047] Furthermore, in the actual operation of low-voltage distribution cabinets, the maintenance system often only receives alarm signals from the end-point manifestations (e.g., a miniature circuit breaker trips due to overheating). However, since we have already revealed the complex cascading effects within the cabinet in S3, this alarming circuit breaker is very likely just an action being executed, while the actual executor (e.g., a loose and overheated copper busbar connector in a distant location causes the circuit breaker to malfunction due to increased ambient temperature caused by spatial heat radiation) does not directly trigger the system alarm. Therefore, it is necessary to transform the dynamic graph representing the forward propagation of faults (from cause to effect) evolved in S3 into a navigation map for reverse fault tracing (from effect to cause) in mathematical space, and combine it with a graph walk algorithm for precise source tracing.

[0048] Specifically, when the cabinet monitoring system detects a target component node... When the temperature or current characteristics exceed the preset safety alarm threshold, an alarm timestamp is immediately generated on the timeline. At this point, extract and freeze the dynamic directed edge weight matrix generated by iteration S3 at that moment. .

[0049] Furthermore, in order to achieve cause-and-effect reasoning, we transposed the frozen matrix to obtain the inverse weight matrix. ,Right now .

[0050] It needs to be explained that, since in the original matrix, the elements The energy representing the fault originates from the node. Transmission to nodes The causal strength. After matrix transpose, the elements in the reverse matrix represent nodes. What is the probability that the abnormal state is caused by a node? It is transmitted from there. Therefore, in graph theory, the transpose operation is equivalent to reversing the direction of the arrows on all directed edges, thus providing the basic topology for the subsequent virtual wandering particles to backtrack along the physical path of the fault propagation.

[0051] Furthermore, the transposed inverse weight matrix is ​​normalized in the column direction to generate the Markov state transition probability matrix. Its matrix elements The calculation formula is as follows: in, This represents the total number of component nodes included in the digital twin model.

[0052] It should be noted that, through normalization, the sum of the elements in each column of the matrix is ​​made strictly equal to 1, thus transforming the original physical transmission weights into a probability distribution that conforms to the properties of a Markov chain. At this point... This can be defined as: during the tracing process, from node Tracing back to the node The single-step state transition probability.

[0053] Furthermore, the source tracing probability distribution column vector is then initialized. As a length of The column vector will be the target component node that triggers the alarm. The probability value at the corresponding position is set to The probability values ​​for all other node positions are set to 0. It needs to be explained that initializing the source tracing probability distribution column vector is equivalent to injecting a source tracing particle at the target node in the virtual topology graph. This is because, in this invention, we assume that this node is 100% the point where the fault manifests, so the initial source tracing probabilities are all distributed around it.

[0054] Furthermore, based on the probabilistic iterative algorithm of Random Walk with Restart (RWR), an iterative walk is performed in the reverse transition matrix.

[0055] Specifically, set the restart damping probability parameter. In the In each iteration, the source tracing probability distribution column vector of the current round is updated. Its core state transition equation is as follows: in, This is the column vector of the probability distribution from the previous iteration; The restart damping probability parameter is set to between 0.15 and 0.20 in this embodiment.

[0056] It should be noted that if traditional graph-based shortest path algorithms (such as Dijkstra's algorithm) are used, the algorithm will only backtrack along the single path with the maximum weight, which completely violates the real-world environment of multiple physics fields (thermal, magnetic, and electrical) being coupled concurrently within a low-voltage switchgear. In contrast, the RWR algorithm... This allows the tracing particles to simultaneously diverge and trace back along all possible physical conduction paths (electrical coupling lines, thermal radiation space, etc.) in a probabilistic proportion. This ensures that the tracing particle has [something] at each step of its movement. The probability of forcibly transmitting back to the alarm origin. It physically simulates the spatial decay effect, effectively preventing wandering particles from getting trapped in a dead loop between two highly coupled nodes (such as two busbar joints that are extremely close to each other heating each other), ensuring that the calculated probability distribution can be firmly anchored to the real root cause that has the highest physical correlation with the alarm node and the closest topological distance.

[0057] Furthermore, the convergence conditions and extraction methods for extracting physical root cause nodes are executed.

[0058] Specifically, after each iteration, the source probability distribution column vectors of the two adjacent iterations are calculated. norm difference : Specifically, in the above formula, when the norm difference value Less than the preset minimum value (For example When the process reaches a global steady state, it can be determined that the Markov process has reached a global steady state. At this point, the iteration is stopped, and the steady-state source probability distribution column vector is obtained. .

[0059] Furthermore, the values ​​in the steady-state traceability probability distribution column vector are sorted in descending order. This is because the target component node that triggered the alarm... Because the restart mechanism has an extremely high probability value, it must be removed from the sequence (since finding itself as the root cause is meaningless). After removal, the node with the highest probability value is extracted and identified as the physical root cause node. .

[0060] It should be noted that, through the above processing, this invention can penetrate the chaotic and massive concurrent alarms within the low-voltage cabinet. Whether the fault is a short-circuit current creeping along the wires or abnormal heat transmitted over a distance, this invention can provide precise targeting for the upcoming multi-agent self-healing isolation simulation in the digital twin space.

[0061] S5. Extract isolated subgraphs containing physical root cause nodes from the digital twin space to construct a Markov decision environment. Use a multi-agent reinforcement learning algorithm to perform multi-trajectory replay simulation in the Markov decision environment, generate the control instruction sequence with the largest cumulative reward value, and issue it for execution.

[0062] It should be noted that in traditional low-voltage distribution cabinet operation and maintenance solutions, once a fault is diagnosed through some means, the system often adopts a single and crude threshold trigger-relay trip hard isolation logic. Although this "one-size-fits-all" power outage method protects the equipment, it is very likely to cause severe power flow oscillations and irreversible secondary power outages (such as causing unplanned shutdowns of entire industrial production lines). Therefore, based on this problem, this invention uses a high-fidelity physical model and dynamic causal graph constructed from S1 to S3, and the physical root cause node precisely located by S4, to conduct trial and error in advance in a digital twin sandbox (virtual space), and uses artificial intelligence algorithms to seek the optimal dynamic balance between ensuring the upper limit of physical safety and maintaining the maximum continuous power supply, thereby achieving global flexible control and self-healing.

[0063] Furthermore, an isolated subgraph containing physical root cause nodes is extracted to construct a Markov Decision Process (MDP).

[0064] Specifically, due to the large number of components in the low-voltage switchgear, including all nodes in reinforcement learning would lead to an explosion in the action space and difficulty in convergence. Therefore, the physical root cause nodes extracted by S4 are used instead. Centered on the dynamic directed edge weight matrix frozen in S3 In the process, breadth-first search (BFS) is used to extract the set of relevant nodes whose hop count is within a preset range (e.g., 2 hops) and whose causal propagation weight is higher than a set threshold, thus forming an isolated subgraph. .

[0065] Furthermore, in the isolated subgraph above, the state space and action space of the Markov decision process are defined.

[0066] Specifically, for the state space : Isolate subgraph The real-time state features of all component nodes within the model serve as the perceptual input. At time 1, the state vector Defined as: in, For components Current real-time temperature; For components Current current load factor (the ratio of actual current to rated current).

[0067] It should be noted that temperature reflects the current thermodynamic urgency of the equipment, while current load rate reflects the capacity margin of the electrical circuit. The combination of the two constitutes the underlying physical constraint base for the system's self-healing decision.

[0068] Specifically, regarding the action space The definition of this space differs from the traditional discrete action of controlling only the opening and closing of switches. In this embodiment, the space is defined as a hybrid action space of continuous and discrete actions. Moment, Action Vector Defined as: Among them, continuous variables Indicates the speed regulation ratio of the inverter cooling fan inside the cabinet (0 for stop, 1 for full load); continuous variable This represents the power limitation ratio for underlying flexible loads (such as non-core lighting and adjustable-speed water pumps); discrete variable. Indicates the opening and closing status of key relays or circuit breakers (1 for closed and energized, 0 for tripped and open).

[0069] Furthermore, the reward value for a single step is calculated based on a preset comprehensive self-healing reward function. .

[0070] Specifically, in order to guide the agent to learn the optimal strategy for ensuring system safety and stability with minimal power outage costs, this invention integrates positive rewards and penalty terms with physical dimensions into the reward function. The calculation logic is as follows: Specifically, in the above formula, the positive score of the retained power supply is... Represented as: in, The rated power of the node; This is a positive incentive factor. This term encourages the agent to keep the switch on as much as possible. Close and maintain a high load operating ratio. To maximize power supply continuity.

[0071] Specifically, the score for the over-temperature penalty term in the above formula... This item primarily aims to endow intelligent agents with the ability to foresee the future, by observing the current state. and the actions performed The input is fed into the pre-built Physical Information Constraint Model (PINN) in S1, and the PINN calculates the future... Predicted temperature after time This item can be represented as: in, The maximum safe temperature limit allowed for the physical components; It is a very high penalty coefficient.

[0072] It should be noted that this penalty is a variant of ReLU. As long as PINN predicts that the action will lead to a future temperature exceeding the limit, it will impose an exponential negative reward, forcing the agent to avoid dangerous actions that may cause a fire.

[0073] Specifically, in the above formula, the mechanical depreciation penalty score... Represented as: in, This is the mechanical wear conversion factor.

[0074] It should be noted that, due to the limited lifespan of the mechanical contacts of physical circuit breakers (e.g., only capable of opening and closing tens of thousands of times), frequent tripping and reclosing will cause severe wear. Therefore, this item mainly calculates the absolute value of the difference between the switching states of adjacent time steps, thereby effectively penalizing the agent's ineffective oscillating operations of repeated switching.

[0075] Furthermore, a multi-agent deep deterministic policy gradient model, comprising an action output network (Actor) and a value evaluation network (Critic), is subsequently constructed, and action sampling and training iterations are performed in a digital twin sandbox.

[0076] Specifically, to suit the actual scenario of multi-dimensional control of low-voltage switchgear, this invention adopts a centralized training and distributed execution architecture, instantiating three heterogeneous intelligent agents in the Actor network: Thermodynamic regulation agent (Agent 1): The input is a subset of temperature states within an isolated subgraph, which is extracted by a multilayer perceptron and then output as a continuous variable through a sigmoid activation function. Independent control of cooling fan speed.

[0077] Flexible load agent (Agent 2): Input is a subset of current load rate states, output is a continuous variable. Independently adjust the limiting ratio of the underlying flexible load.

[0078] Topologically isolated agent (Agent 3): Integrating input temperature rise and current state, it introduces the Gumbel-Softmax reparameterization technique to output an approximately discrete pseudo-probability distribution. The decision-making process involves determining the opening and closing status of key switches.

[0079] It is important to emphasize that this reparameterization technique allows the computational graph of discrete actions to remain differentiable everywhere.

[0080] Specifically, in the construction of the Critic network (global evaluation network), the Critic network receives the complete global state during the training phase. Joint action output by three intelligent agents As a joint input, evaluate the future cumulative expected Q value of performing this joint action in the current state, i.e. .

[0081] Furthermore, within the digital twin sandbox, the intelligent agent continuously interacts with the PINN physics engine, recording the trajectory segments it experiences. The data is stored in the experience replay pool. Furthermore, during the parameter update (fine-tuning) phase, the Critic network updates its parameters to approximate the true Bellman equation by minimizing the temporal difference error (TD-Error); while the Actor network updates its parameters along the Q-value output of the Critic network with respect to the action. The parameters are updated by ascending along the gradient direction in order to find a strategy that maximizes the Q value.

[0082] Furthermore, the sequence of control instructions with the highest cumulative reward value is generated and issued for execution.

[0083] Specifically, after the model completes training and fine-tuning, when an actual low-voltage switchgear triggers an alarm and the root cause node is located, the system freezes the current physical twin state as the initial environment. Then, the trained Actor network is used to perform Monte Carlo multi-trajectory replay simulations in the virtual twin space (i.e., generating multiple decision paths for future directions in parallel). The cumulative reward value is calculated and extracted through the Critic network. ( To predict the horizon, The system identifies the golden trajectory with the largest discount factor. Subsequently, the system extracts the output tensor of the first action combination in the golden trajectory and maps and parses it into a Modbus / Profinet control instruction sequence that the underlying PLC (Programmable Logic Controller) can recognize (e.g., first increase the speed of fan No. 3 to 85%, then limit the current of non-critical branches to 50%, and if the PINN predicts that the temperature still does not drop, finally execute the main circuit of the contactor to cut off).

[0084] It should be noted that, through the above operations, this invention changes the passive situation of electrical equipment burning out / tripping first and then being dealt with. By using high-precision physical information digital twins as a simulation sandbox and combining it with long-term value assessment through reinforcement learning, it can target the hidden physical root causes and implement flexible degradation and dynamic cooling control. Under the premise of ensuring uninterrupted power supply to industrial loads to the greatest extent, it can achieve true fully automatic self-healing of low-voltage power distribution systems.

[0085] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of this application can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0086] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0087] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0088] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0089] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0090] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for tracing the root cause of faults and achieving self-healing of low-voltage switchgear components based on digital twins, characterized in that, include: Acquire multimodal time-series observation data of physical low-voltage switchgear components, input the multimodal time-series observation data into a pre-constructed physical information constraint model for timestamp synchronization and missing interpolation, and generate a digital twin state matrix; Extract the node feature vectors from the digital twin state matrix, and initialize the static adjacency weight matrix of the multidimensional heterogeneous causal knowledge graph according to the preset low-voltage cabinet component spatial topology and multi-physics coupling rules. Initializing the static adjacency weight matrix includes: Calculate the electrical admittance between nodes based on the circuit topology and generate an electrical connectivity matrix; Extract the spatial centroid distance between nodes in the digital twin 3D model and generate a thermal radiation coupling matrix through Gaussian kernel function mapping; An electromagnetic interference matrix is ​​generated based on the mutual inductance coefficients between component nodes. The electrical connectivity matrix, thermal radiation coupling matrix, and electromagnetic interference matrix are linearly weighted and summed to generate an initial static adjacency weight matrix. Multivariate time series features within a sliding time window are extracted, and a residual statistic based on vector autoregression is constructed. Based on the test results of the residual statistic, the directed edge weights of the multidimensional heterogeneous causal knowledge graph are dynamically smoothed and evolved. When the target component node is detected to have triggered the preset alarm threshold, the dynamic directed edge weight matrix at the current moment is frozen and the matrix is ​​transposed. Based on the random walk probability iteration algorithm with restart, the node with the largest probability steady-state convergence value is extracted from the reverse transition matrix as the physical root cause node. In the digital twin space, an isolated subgraph containing the physical root cause node is extracted to construct a Markov decision environment. A multi-agent reinforcement learning algorithm is used to perform multi-trajectory replay simulation in the Markov decision environment to generate the control instruction sequence with the maximum cumulative reward value and issue it for execution.

2. The method for tracing the root cause of low-voltage switchgear component failures and achieving self-healing based on digital twins as described in claim 1, characterized in that, The generation of the digital twin state matrix includes: Acquire multimodal time-series observation data including high-frequency current sequences, voltage sequences, low-frequency infrared temperature sequences, and acoustic emission sequences; A physical information neural network is constructed, and a data fitting error term and a partial differential equation penalty term are configured in the loss function of the physical information neural network. The missing state points of the low frequency sequence are completed by using high frequency sequence deduction. The cumulative distance matrix between each completed sequence and the digital twin reference clock is calculated based on the dynamic time warping algorithm. The timestamps of all sequences are then extracted and the optimal backtracking path is used to reconstruct a high-fidelity digital twin state matrix.

3. The method for tracing the root cause of low-voltage switchgear component failures and achieving self-healing based on digital twins as described in claim 2, characterized in that, The methods for handling the penalty terms in the partial differential equations include: Calculate the time first derivative of the network-predicted temperature, the difference between the sum of the spatial heat diffusion term and the Joule heat source term, and use the sum of the squares of the difference as a physical energy conservation constraint to iteratively update the network parameters.

4. The method for tracing the root cause of low-voltage switchgear component failures and achieving self-healing based on digital twins as described in claim 1, characterized in that, The construction of residual statistics based on vector autoregression includes: Extract the state feature sequences of the source and target nodes connected by directed edges within the sliding time window; Construct a restricted autoregressive model that uses only the historical state sequence of the target node to predict the current state of the target node and calculates the restricted residual sum of squares. Construct an unrestricted autoregressive model, jointly use the historical state sequences of the source node and the target node to predict the current state of the target node, and calculate the unrestricted sum of squared residuals. The difference between the restricted residual sum of squares and the unrestricted residual sum of squares is calculated, and the ratio of this difference to the unrestricted residual sum of squares is used as a statistical test metric for evaluating the causal transmission strength from the source node to the target node.

5. The method for tracing the root cause of low-voltage switchgear component failures and achieving self-healing based on digital twins as described in claim 4, characterized in that, The directed edge weights of the aforementioned multidimensional heterogeneous causal knowledge graph undergo dynamic smoothing evolution, including: The statistical test statistic is mapped to a causal activation coefficient within an open interval using a sigmoid activation function. Using the exponential moving average mechanism, the edge weights of the previous time step are weighted and fused with the product of the causal activation coefficient and the initial static weights to update the directed edge weight matrix of the current time step.

6. The method for tracing the root cause of low-voltage switchgear component failures and achieving self-healing based on digital twins as described in claim 1, characterized in that, Based on a random walk probability iterative algorithm with restart, the physical root cause nodes are extracted, including: The column-direction norm normalization is performed on the transposed dynamic directed edge weight matrix to generate the Markov state transition probability matrix. Initialize the source tracing probability distribution column vector, set the value of the corresponding position of the target component node to one, and set the value of the other node positions to zero; Set the restart damping probability parameter. In each iteration, multiply the source tracing probability distribution column vector of the previous round with the Markov state transition probability matrix, and superimpose the initial source tracing probability distribution column vector after being weighted by the restart damping probability parameter, and update the source tracing probability distribution column vector of the current round.

7. The method for tracing the root cause of low-voltage switchgear component failures and achieving self-healing based on digital twins as described in claim 6, characterized in that, The convergence conditions and extraction methods for physical root cause nodes include: Calculate the norm difference between the column vectors of the source probability distributions in two consecutive iterations; When the norm difference is less than the preset minimum value, the iteration stops and the steady-state source tracing probability distribution column vector is obtained. The values ​​in the steady-state tracing probability distribution column vector are sorted in descending order. After removing the target component nodes that trigger alarms, the node with the highest probability value is extracted and output as the physical root cause node.

8. The method for tracing the root cause of low-voltage switchgear component failures and achieving self-healing based on digital twins as described in claim 1, characterized in that, The control instruction sequence that generates the maximum cumulative reward value includes: The real-time temperature and current load factor of the components within the isolated subgraph are defined as the state space of a Markov decision process. The combination of continuous cooling fan speed regulation ratio, flexible load limiting ratio, and discrete switch opening and closing states is defined as the action space; Construct a multi-agent deep deterministic policy gradient model that includes an action output network and a value evaluation network, and perform action sampling in a digital twin sandbox; The action reward is calculated based on the preset comprehensive self-healing reward function, and the model parameters are updated by backpropagation until the action combination with the largest cumulative reward value in the Monte Carlo trajectory is extracted and mapped into a control command sequence.

9. The method for tracing the root cause of low-voltage switchgear component failures and achieving self-healing based on digital twins as described in claim 8, characterized in that, The calculation logic of the comprehensive self-healing reward function includes: The positive score of the power supply retained after the action is executed is calculated, the score of the over-temperature penalty item for the future temperature predicted by the physical information constraint model exceeding the safe limit is subtracted, and the mechanical depreciation penalty score generated by the discrete action of the relay is subtracted. The sum is the action reward value of a single step.

Images