A multi-parameter fusion fault early warning method of an intelligent distribution box

Abstract

The application particularly relates to a multi-parameter fusion fault early warning method of an intelligent power distribution box, and belongs to the technical field of electric digital data processing.The method comprises the following steps: collecting temperature and current data; establishing an initial topology according to sensor coordinates, and generating a dynamic heat conduction topology graph by updating edge weights on the basis of temperature gradient and load current change rate; extracting node embedding vectors from a de-meaned space-time temperature matrix by using a graph convolution network; establishing a node normal distribution center through historical embedding vectors, calculating an Euclidean distance on line, and obtaining an abnormality degree after environmental temperature compensation; locating a fault source by iteratively propagating a residual error along a reverse heat flow direction by using a topological definition of an attribution allocation coefficient and a node heat absorption coefficient; and setting a threshold value dynamically based on historical residual error statistics to perform early warning.The method can dynamically adapt to contact thermal resistance changes, suppress heat diffusion and environmental temperature drift, and realize early fault sensitive early warning and accurate tracing.

Description

Technical Field

[0001] This invention relates to a multi-parameter fusion fault early warning method for intelligent distribution boxes, belonging to the field of electrical digital data processing technology. More specifically, this invention relates to a multi-parameter fusion fault early warning method for intelligent distribution boxes. Background Technology

[0002] In industrial environments such as metal processing, the contact thermal resistance of electrical connection points in low-voltage distribution boxes is constantly and slowly changing due to surface oxidation and dust accumulation. Furthermore, production schedule changes cause a wide range of load current fluctuations. To monitor the potential for localized overheating caused by this, existing technologies typically construct a fixed topology network based on temperature sensors. Static thermal conductivity weights are preset according to the nominal thermal resistance and physical spacing of components. This allows for the extraction of temperature amplitude or short-term statistical characteristics to measure anomalies, and the location with the highest temperature or anomaly measurement among the monitoring points is directly identified as the fault source.

[0003] However, the aforementioned modeling method using fixed topology and constant thermal conductivity weights struggles to reflect in real time the actual changes in heat transfer efficiency caused by slow variations in contact thermal resistance and sudden load changes, leading to a systematic mismatch between the network model and the actual heat transfer path. Furthermore, fixed statistical detection thresholds based on short-term data cannot cover the wide-range fluctuations in ambient temperature across different seasons and normal operating conditions under production schedules, easily causing false alarms or missed alarms. More critically, due to the conductive effect of highly thermally conductive materials such as copper and aluminum busbars, heat generated by poor contact will rapidly diffuse outwards, causing downstream sensor nodes to also exhibit significant temperature rises. If faults are directly located based on temperature or local abnormal amplitude peaks, it is highly likely that secondary high-temperature points formed by heat transfer will be mistakenly identified as the source of heat, ignoring the true fault connection point, thus resulting in fault tracing errors.

[0004] Therefore, under complex working conditions in industrial sites, how to overcome the sensor network model mismatch, abnormal characterization baseline drift and heat conduction diffusion effects caused by the slow change of contact thermal resistance, load current fluctuation and ambient temperature change, and achieve sensitive early warning and accurate source tracing of early contact failures in low-voltage distribution boxes has become an urgent technical problem to be solved. Summary of the Invention

[0005] In existing technologies, under complex industrial conditions, overcoming the sensor network model mismatch, anomaly baseline drift, and heat conduction diffusion effects caused by the coupling of slow changes in contact thermal resistance, load current fluctuations, and ambient temperature changes, and achieving sensitive early warning and accurate tracing of early contact failures in low-voltage distribution boxes, has become an urgent technical problem to be solved. To address this problem, this invention provides solutions in the following aspects.

[0006] In a first aspect, the present invention provides a multi-parameter fusion fault early warning method for intelligent distribution boxes, employing the following technical solution: A multi-parameter fusion fault early warning method for intelligent distribution boxes includes: collecting temperature and current data; establishing an initial directed topology graph based on sensor coordinates; updating edge weights online based on temperature gradient and load current change rate to generate a dynamic heat conduction topology graph; using graph convolutional network encoding to obtain a mean-free spatiotemporal temperature matrix to obtain node embedding vectors; establishing a normal distribution center for nodes based on historical embedding vectors; calculating the Euclidean distance between the current embedding vector and the center online as the initial anomaly degree; and obtaining a compensated anomaly degree using ambient temperature compensation; defining an attribution allocation coefficient based on the dynamic heat conduction topology graph and combining it with a preset node heat absorption coefficient; iteratively propagating the residual of the compensated anomaly degree in the reverse heat conduction direction; and using the node with the largest residual as the fault source; and setting a dynamic early warning threshold based on historical residual statistics.

[0007] Preferably, establishing an initial directed topology graph based on sensor coordinates includes: mapping each temperature sensor to a node in the graph; determining the direction of the directed edges based on the physical heat transfer path between the electrical components monitored by each sensor; calculating the initial thermal conductivity weight of each edge based on the thermal conductivity, effective thermal cross-sectional area, and conduction path length of the heat transfer path material, according to Fourier's law of thermal conductivity; and generating an initial directed topology graph using the initial thermal conductivity weight as the edge weight.

[0008] Preferably, the online generation of a dynamic heat conduction topology based on updating edge weights using temperature gradient and load current change rate includes: during online operation, at each sampling time, calculating the temperature gradient normalization exponent between adjacent node pairs, where the temperature gradient normalization exponent is determined by the ratio of the absolute temperature difference between nodes to a preset maximum temperature difference; calculating the load current change rate factor for each node, where the load current change rate factor is determined by the ratio of the change in current at the current sampling time relative to the previous sampling time to the rated current of the circuit; multiplying the temperature gradient normalization exponent by the load current change rate factor, multiplying by the initial thermal conductivity weight, and then exponentially smoothing with the dynamic edge weights of the previous time step to obtain the dynamic edge weights at the current time step; and generating a dynamic heat conduction topology based on the updated edge weights.

[0009] Preferably, the node embedding vector is obtained by encoding the mean-reduced spatiotemporal temperature matrix using a graph convolutional network, including: taking the temperature values ​​of each node at the most recent preset number of sampling times as a reference to form a spatiotemporal temperature matrix; subtracting the mean of the temperature time series of each node to obtain the mean-reduced spatiotemporal temperature matrix; inputting the mean-reduced spatiotemporal temperature matrix and the adjacency weight matrix of the dynamic heat conduction topology graph into a graph feature propagation network, which includes a first edge aggregation convolutional layer and a second higher-order feature mapping convolutional layer; in the first edge aggregation convolutional layer, for each node, the temperature sequence vectors of its incoming edge neighbor nodes are weighted and summed using the current dynamic edge weights, and then processed by a linear transformation of the trainable parameter matrix and a nonlinear activation function to obtain a primary feature vector; the primary feature vector is input into the second higher-order feature mapping convolutional layer for nonlinear mapping and cross-node information integration, and the intrinsic embedding vector of each node is output.

[0010] Preferably, establishing a normal distribution center for nodes based on historical embedding vectors includes: collecting all intrinsic embedding vectors of each node within at least one complete load change cycle during a fault-free historical period; and for each node, calculating the arithmetic mean of the collected intrinsic embedding vectors element by element to obtain the normal operating condition embedding distribution center of that node.

[0011] Preferably, the Euclidean distance between the current embedded vector and the center is calculated online as the initial anomaly degree, and the compensated anomaly degree is obtained using environmental temperature compensation. This includes: during online operation, whenever a new intrinsic embedded vector is obtained, the Euclidean distance between the vector and the center of the normal operating condition embedded distribution of the corresponding node is calculated as the initial node anomaly degree; the current ambient temperature inside the chamber is obtained, and the ratio of the thermodynamic temperature of the reference ambient temperature to the thermodynamic temperature of the current ambient temperature inside the chamber is calculated as the compensation factor, based on a preset reference ambient temperature; the initial node anomaly degree is multiplied by the compensation factor to obtain the compensated anomaly degree after environmental temperature compensation.

[0012] Preferably, the attribution allocation coefficient is defined based on the dynamic heat conduction topology graph, combined with the preset node heat absorption coefficient, including: for any directed edge from the first node to the second node in the dynamic heat conduction topology graph, an attribution allocation coefficient is defined from the second node to the first node, the attribution allocation coefficient is equal to the ratio of the dynamic edge weight of the directed edge to the sum of the dynamic edge weights of all outgoing edges of the first node; a preset heat absorption coefficient is set for each node, the heat absorption coefficient is determined according to the type of electrical component represented by the node, and is used to characterize the heat dissipation consumption ratio of the node in the residual propagation process.

[0013] Preferably, the residual is iteratively propagated in the reverse heat conduction direction for the compensation anomaly degree, with the node with the largest residual as the fault source. This includes: using the compensation anomaly degree of each node as the initial residual; in each iteration, for each node, the sum of the residuals returned by all its downstream nodes according to the attribution allocation coefficient is added to the residual of the node in the previous round, and then the product of the residual of the node in the previous round and its own heat absorption coefficient is subtracted to obtain the residual updated in this round; the iteration continues until the maximum value of the change in the residual of all nodes in this round and the previous round is less than the preset convergence threshold, or the maximum number of iterations is reached; the node with the largest final residual value is selected as the fault source identifier.

[0014] Preferably, the dynamic early warning threshold is set based on historical residual statistics, including: during historical fault-free operation, recording the final residual of each node after each iteration process to form a residual background set; calculating the mean and standard deviation of the residual background set; adding three times the standard deviation to the mean as the dynamic early warning threshold; during online monitoring, if the final residual value corresponding to the fault source identifier exceeds the dynamic early warning threshold, a fault early warning signal is output, indicating the electrical connection position corresponding to the fault source identifier.

[0015] Secondly, this invention provides a multi-parameter fusion fault early warning system for intelligent distribution boxes, employing the following technical solution: A multi-parameter fusion fault early warning system for intelligent distribution boxes includes: a data acquisition module for acquiring temperature and current data; a dynamic topology construction module for establishing an initial directed topology graph based on sensor coordinates, and updating edge weights online based on temperature gradient and load current change rate to generate a dynamic heat conduction topology graph; a graph convolutional encoding module for obtaining node embedding vectors by encoding a mean-free spatiotemporal temperature matrix using a graph convolutional network; an anomaly generation module for establishing a normal distribution center of nodes based on historical embedding vectors, calculating the Euclidean distance between the current embedding vector and the center online as the initial anomaly, and obtaining a compensated anomaly using environmental temperature compensation; a fault tracing module for iteratively propagating the residuals of the compensated anomaly in the reverse heat conduction direction according to the attribution allocation coefficient defined in the dynamic heat conduction topology graph and a preset node heat absorption coefficient, with the node with the largest residual as the fault source; and an early warning module for setting a dynamic early warning threshold based on historical residual statistics.

[0016] The embodiments of the present invention have at least the following beneficial effects:

[0017] 1. The thermal conductivity weights of directed edges are updated in real time based on temperature gradient and load current change rate, enabling the topology graph to dynamically match changes in contact thermal resistance and fluctuations in heat conduction efficiency, thus solving the mismatch problem of traditional fixed network models. The graph convolutional network, combined with dynamic topology, encodes the mean-free spatiotemporal temperature matrix. The extracted node embedding vectors not only retain the thermal conduction structure constraints but also amplify early weak abnormal temperature rise features and suppress instantaneous noise, thereby significantly improving the detection sensitivity of early contact failure faults under complex working conditions.

[0018] 2. Establish normal embedded distribution centers for each node on a large timescale of fault-free history, covering the complete change cycle from low load to high load and the accompanying ambient temperature fluctuations, making the anomaly judgment criteria more comprehensive and effectively avoiding false alarms caused by short-term operating condition changes; use the absolute temperature scale ratio of ambient temperature to reference temperature to multiplicatively compensate for the initial anomaly degree, suppressing the anomaly degree baseline drift caused by seasonal or diurnal temperature differences, making the anomaly measurement more realistically reflect the thermal state deviation of the equipment body, and improving the robustness of online monitoring.

[0019] 3. Based on the attribution allocation coefficient and node heat absorption coefficient defined by dynamic topology, the residual iterative propagation in the reverse heat flow direction is performed to gradually transmit the abnormal signals of secondary conduction nodes back and converge them to the real fault source, eliminating the positioning deviation caused by heat conduction and diffusion; by pre-setting the heat absorption coefficient to distinguish the heat dissipation capacity of different components, the residual allocation is made more in line with the physical reality, further improving the positioning accuracy; the early warning threshold is adaptively determined by the statistical distribution of historical normal residuals, overcoming the empirical blindness of fixed thresholds, and taking into account the sensitivity of early fault warning and the effectiveness of false alarm suppression. Attached Figure Description

[0020] Figure 1 The schematic diagram illustrates the steps of a multi-parameter fusion fault early warning method for an intelligent distribution box according to the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0023] S1: Collect temperature and current data.

[0024] In this embodiment, a multi-parameter fusion fault early warning method for intelligent distribution boxes first performs the acquisition of temperature and current data.

[0025] Considering the coupling effect between high-current circuits and high-dust environments in industrial environments such as metal processing, the contact thermal resistance of electrical connection points due to oxidation and dust accumulation is in a state of continuous slow change. At the same time, the wide fluctuation range of load current caused by production schedule switching makes it necessary to obtain original measurement values ​​that can reflect local thermal state and circuit electrical stress. This provides high-time-resolution basic data for subsequent quantification of conduction efficiency and thermal anomaly offset. Therefore, this step synchronously obtains multiple temperature values ​​and circuit current values ​​through pre-deployed temperature sensors and current transformers, and forms a real-time data stream with a fixed sampling interval.

[0026] Specifically, N digital temperature sensors are installed at the busbar contact surfaces, main circuit breaker contacts, branch circuit terminals, and reference locations within the distribution box. Each sensor node is denoted as […]. , Simultaneously, current transformers are installed on the main incoming line and each major branch circuit to monitor the current of the corresponding circuit and record the circuit. The current is If there is no independent circuit current monitoring, then the total incoming current shall be used. As a common current value.

[0027] The temperature sensor uses a contact-type PT100 resistance temperature detector (RTD) probe, which is directly attached to the surface of the conductor being measured. Thermal grease is applied between the probe and the conductor to reduce contact thermal resistance. Its measurement accuracy is [insert accuracy here]. Response time constant Less than This ensures sufficient responsiveness to rapid temperature rises caused by sudden load changes.

[0028] The output signals of the temperature sensor and the current transformer are respectively connected to a data acquisition unit located in the distribution box. This acquisition unit has a built-in multi-channel synchronous sampling module, and all channels use a uniform configurable sampling interval. Perform synchronous analog-to-digital conversion, sampling interval Get experience points The unit is seconds (s), which can also be adjusted by the implementer according to the specific implementation scenario to achieve a balance between capturing dynamic changes and data storage costs.

[0029] At each sampling time The data acquisition unit performs a single sampling of the signals from each temperature sensor and converts them into Celsius temperature values ​​to obtain the node temperature. The unit is Celsius. Simultaneously, the outputs of the current transformers in each circuit are synchronously sampled and converted into ampere values ​​to obtain... The unit is amperes (A); in addition, the current ambient temperature measured by the ambient temperature sensor inside the chamber is also read. The unit is Celsius. ).

[0030] The data set obtained within a complete sampling window is a set of multidimensional vectors, which contains the current time step. All node temperature values, circuit current values, and ambient temperature values.

[0031] During the data acquisition process, if a sensor in a certain channel becomes open-circuited or exceeds the measurement range, the data acquisition unit marks the data of that channel as invalid and fills it with the measurement value of the last valid moment of that channel. At the same time, a channel status flag is generated to avoid subsequent calculations being interrupted due to missing values.

[0032] The above sampling process is performed at fixed intervals. Periodic triggering continuously generates a time-series data stream with equal intervals. This data stream is written to a circular buffer in real time for subsequent steps to read.

[0033] S2: Establish an initial directed topology graph based on the sensor coordinates.

[0034] In industrial scenarios such as metal processing, the contact thermal resistance of electrical connection points such as busbar lap surfaces and circuit breaker contacts in low-voltage distribution boxes is in a slowly changing state due to oxidation and dust accumulation on the contact surfaces. At the same time, the load current fluctuates within a large range due to production schedule changes, which causes the effective heat conduction capacity of the heat conduction path to change accordingly. The topology model with fixed edge weights cannot accurately reflect this real-time physical heat transfer state. Therefore, this step first establishes an initial directed topology graph based on the installation coordinates of the temperature sensor. Then, the edge weights are dynamically updated online based on the temperature gradient normalization exponent and the load current change rate factor to generate a dynamic heat conduction topology graph. Subsequently, a graph convolutional network is used to perform feature encoding on the mean-free spatiotemporal temperature matrix to obtain the thermal state embedding vector of each node.

[0035] Each temperature sensor pre-deployed in the distribution box is represented as a node in the diagram, and the set of nodes is denoted as . Based on the sensor's installation coordinates within the distribution box, the spatial relationship between nodes is determined. Combining this with the electrical components and physical heat transfer paths of the monitoring points, the direction of directed edges is identified. This direction points from the expected heat source node to the downstream node affected by its heat conduction, such as from the main circuit breaker contact to the branch terminal of the same circuit, thus forming a set of directed edges. For each directed edge The initial thermal conductivity base value of this path is calculated using circuit analogy based on Fourier's law of thermal conductivity. Used as the initial weight of the edge.

[0036]

[0037] In the formula, The thermal conductivity of the busbar material is determined by referring to a table based on the actual material, such as copper busbar, aluminum busbar, or air gap, and the unit is watts per meter Kelvin (W / m). ); The equivalent thermal conductivity cross-sectional area of ​​the path in the direction perpendicular to the heat flow is calculated from the conductor geometry and the connection area, with units of square meters. ); The length of the effective heat conduction path between sensors is measured based on the installation coordinates and component layout, and the unit is meters (m). ); the calculated The dimension is watt per Kelvin ( This reflects the theoretical thermal conductivity of the path in a clean, unoxidized state, thereby generating the initial directed topology graph. .

[0038] During online operation, the slow increase in contact thermal resistance or the drastic fluctuations in Joule heating caused by sudden load changes will alter the actual conduction efficiency, requiring adjustments at each sampling interval. To update edge weights, first calculate the weights of adjacent nodes. normalized exponent of temperature gradient between This index is calculated by comparing the current absolute temperature difference with a preset reference temperature difference. The ratio quantifies the transient heat conduction driving force.

[0039]

[0040] In the formula, and They are nodes With nodes At any moment The temperature values ​​are from the temperature time series collected by S1, and the unit is degrees Celsius. ); As the temperature gradient normalization benchmark, an empirical value of 50K is taken. This is a hyperparameter and can also be adjusted by the implementer based on the maximum allowable temperature difference of the distribution box. Taking the larger of the two values ​​in the denominator ensures... In Interval, dimensionless.

[0041] Next, calculate each node. Load current change rate factor This factor reflects the degree of fluctuation in the circuit or total incoming current of the node.

[0042]

[0043] In the formula, For nodes The associated loop at time The measured current value is used; if no independent loop current sensor is configured, the total incoming current is used. The unit is ampere (A); This is the current value at the previous sampling time; The rated current value for the corresponding circuit is obtained from the equipment nameplate, and the unit is amperes (A). It is dimensionless, and the larger its value, the more drastic the change in current.

[0044] get and Then, combined with the initial thermal conductivity baseline value The dynamic edge thermal conductivity weights are updated using an exponential smoothing method. .

[0045]

[0046] In the formula, The smoothing coefficient is taken as an empirical value of 0.25. It is dimensionless and is a hyperparameter. It can also be adjusted by the implementer according to the sampling noise level. The edge weights at the previous sampling time, and the initial time. Take as ; The choice of value strikes a trade-off between dynamic response speed and suppression of short-time measurement noise; after the update, it is... Using the non-zero elements of the adjacency matrix, generate the dynamic heat conduction topology graph for the current moment. .

[0047] To extract node thermal state features containing graph structure information from time-series temperature data, the current sampling time is used as the starting point. Based on the baseline, truncate the nearest [item / feature]. The temperature values ​​of each node at each sampling point are used to construct a... 3D spatiotemporal temperature matrix ,in The time window length is taken as an empirical value of 15; this is a hyperparameter, which can also be adjusted by the implementer according to the desired thermal dynamic timescale to be captured; matrix The Behavior Nodes Temperature time series vector within the window .

[0048] For each node Calculate the arithmetic mean of the given row sequence. and subtract each element in the sequence This yields the mean-free temperature time series vector. Its length is ; to all nodes Arranged row by row to form a mean-free spatiotemporal temperature matrix Mean removal can eliminate overall baseline drift caused by seasonal or diurnal temperature variations, allowing features to focus on relative temperature change patterns between nodes.

[0049] Remove the mean spatiotemporal temperature matrix With dynamic heat conduction topology diagram Adjacency weight matrix The common input is a graph feature propagation network containing two convolutional layers, which encodes the temperature sequence sequentially through a first edge aggregation convolutional layer and a second higher-order feature mapping convolutional layer.

[0050] In the first side of the aggregated convolutional layer, for each node... First, the aggregation vector is calculated using the temperature sequence vectors of its incoming neighbors and the current dynamic edge weights. .

[0051]

[0052] In the formula, express All nodes in the array are pointed to by directed edges. The set of neighboring nodes, i.e., incoming edge neighbors; For nodes The mean-free temperature time series vector; For the node Pointing to node Dynamic edge weights; The dimension is .

[0053] Then After trainable weight matrix Perform a linear transformation and activate the function using a linear rectified circuit. Processing to obtain primary feature vectors .

[0054]

[0055] In the formula, Let be the learnable parameter matrix of the edge-aggregated convolutional layer, with size . ,in The initial feature dimension is taken as an empirical value of 64, which is a hyperparameter and can also be adjusted according to computing resources. This is the ReLU activation function.

[0056] The primary feature vectors of all nodes are input into the second higher-order feature mapping convolutional layer. This layer also uses a neighbor aggregation method based on dynamic topology and applies independent trainable weight matrices. Linear mapping and nonlinear activation are performed to further map the primary features to a higher-order feature space, ultimately outputting each node. eigenembryo vectors Embedded vector dimension The empirical value of 32 is taken as the hyperparameter, but it can also be adjusted by the implementer according to the model capacity requirements. It serves as a compact representation of the node's thermal state at the current moment, integrating local topological constraints and temporal temperature dynamics.

[0057] S3: Establish the normal distribution center of nodes based on historical embedding vectors, calculate the Euclidean distance between the current embedding vector and the center online as the initial anomaly degree, and obtain the compensated anomaly degree using environmental temperature compensation.

[0058] In industrial environments such as metal processing, low-voltage distribution boxes are installed in non-temperature-controlled workshops, where the internal ambient temperature varies. The rate of Joule heating in the conductor fluctuates with the season, day and night, and heat dissipation from nearby equipment. At the same time, changes in load current cause the rate of Joule heating in the conductor to be inconsistent.

[0059] The node intrinsic embedding vector output by step S2 The spatiotemporal thermal state of each node under dynamic topology constraints at the current moment is encoded, but the absolute value of the instantaneous embedding vector alone cannot directly determine whether it has deviated from the safe operating range.

[0060] Under normal operating conditions, the embedding vectors exhibit a clustered distribution in the feature space around a certain center, which is determined by the long-term thermal behavior of the device under fault-free conditions.

[0061] Therefore, this step requires first establishing a normal distribution center for each node based on historical fault-free data, which can cover the entire load change cycle, as a comparison benchmark.

[0062] The process of establishing a normal distribution center for nodes is completed offline before the system is put into online early warning.

[0063] Collect the node embedding vectors of all sampling moments during the historical period when the distribution box is confirmed to be operating without faults. The length of the historical period is not less than a complete load change cycle, such as at least 24 hours, to ensure that the sampling set includes various operating conditions from low load to full load.

[0064] For each node Extract common data from this time period A series of embedded vector samples The arithmetic mean of all samples is calculated, and the resulting vector is the normal operating condition embedding center of that node. .

[0065] The calculation formula is as follows:

[0066]

[0067] In the formula, This represents the total number of valid samples within the normal operating window, and its value depends on the length of the historical period and the sampling interval. The ratio, expressed as a dimensionless count value; For nodes During a time of no failure in history The intrinsic embedding vectors are output by the graph feature propagation network in step S2, and their dimension is the same as that set in step S2. Consistent, take the empirical value of 32, dimensionless; the summation operation is performed on each component of the vector separately.

[0068] dimensionality and Similarly, it encodes the average thermal state pattern of the node under the combined effects of normal load cycles and environmental changes, stores it in system memory, and remains unchanged during online operation without being updated over time.

[0069] During online operation, whenever step S2 occurs at the current sampling time... Output new embedding vector The system immediately calculates the vector and the corresponding node's normal distribution center. The degree of deviation between them.

[0070] Using Euclidean distance as a measure of deviation, the initial node anomaly degree at the current time is obtained. The calculation formula is:

[0071]

[0072] In the formula, Representing vectors Norm, i.e., the calculation of The arithmetic square root of the sum of the squares of the components of the difference between two vectors in 3D space.

[0073] The dimension of the value is the same as that of the embedding vector; both are dimensionless scalars. The larger the value, the stronger the node. The further the current thermal state deviates from its normal distribution center in the feature space.

[0074] Since the embedded vector has undergone mean removal and graph convolutional encoding in step S2, this distance value mainly reflects the changes in the relative temperature patterns between nodes, and has a certain resistance to seasonal overall environmental temperature drift.

[0075] In power distribution network operation, conductor resistance increases approximately linearly with temperature. With a constant load current, the same Joule heat power will produce a higher rate of temperature rise. This means that even if the equipment itself is fault-free, the temperature rise at each node may be higher in summer than in winter, thus increasing the Euclidean distance between the embedded vector and the normal center. This generates baseline drift that is independent of the device's state.

[0076] To mitigate this environmental impact, this step adjusts the initial anomaly level. Introduce multiplicative compensation based on ambient temperature.

[0077] Obtain the current ambient temperature inside the chamber, which is synchronously collected in step S1. and preset reference ambient temperature The ratio of the two values ​​after converting them to thermodynamic temperatures is used as a compensation factor. This factor is then multiplied by the initial anomaly degree to obtain the compensation anomaly degree. .

[0078] The compensation calculation formula is:

[0079]

[0080] In the formula, For reference to ambient temperature, an empirical value of 25 is used, in degrees Celsius. This is an overparameter, which can also be adjusted by the implementer according to the typical operating environment of the distribution box; For step S1 at time... The ambient temperature inside the chamber was collected, in degrees Celsius. The constant 273 is used to convert Celsius temperature to an approximate thermodynamic temperature, with the unit being Kelvin (K); ratio factor. Dimensionless.

[0081] when Higher than When the factor is less than 1, the anomaly degree is reduced accordingly after compensation; otherwise, it is appropriately amplified, thereby suppressing the systematic shift in anomaly degree caused by changes in ambient temperature itself.

[0082] This is the final output of this step, with dimensions and... They are identical, dimensionless, and stored in memory for subsequent residual propagation steps.

[0083] S4: Based on the dynamic heat conduction topology, define the attribution allocation coefficient and combine it with the preset node heat absorption coefficient. Iteratively propagate the residual in the reverse heat conduction direction to compensate for the anomaly degree, and take the node with the largest residual as the fault source.

[0084] Because both copper and aluminum busbars in the low-voltage distribution box have high thermal conductivity, heat generated at poor contact points will rapidly diffuse along the conductors and their connecting structures to the surrounding area, causing a significant temperature rise in the sensor nodes located downstream of the heat flow direction. If the compensation anomaly output from step S3 is directly used... The highest point is identified as the fault location, which can easily lead to the misidentification of secondary high-temperature points formed by heat conduction as the source of heat. Therefore, this step is based on the dynamic heat conduction topology generated in step S2. The heat flow direction information contained therein is used to perform iterative residual propagation of the compensation anomaly degree along the reverse heat conduction direction by defining the attribution allocation coefficient and the introduced node heat absorption coefficient, so that the abnormal signal gradually converges to the real upstream fault source.

[0085] For dynamic heat conduction topology diagram Any directed edge Represents heat flow from nodes Transmission to nodes , i.e., node For upstream heat sources and nodes For the heated downstream node, to reverse-attribute the abnormal signals of the downstream node to their upstream source, define a slave node. Pointing to node Attribution coefficient The method for calculating this coefficient is as follows: Current dynamic edge weights Divide by node The sum of the weights of all outgoing dynamic edges is calculated using the following formula:

[0086]

[0087] In the formula, express China-Israel node The set of all neighboring nodes pointed to by directed edges at the tail, i.e., nodes. Out-of-border neighbor set; The dynamic edge weights updated in step S2, in watts per Kelvin (W). ); For nodes Weights of total thermal conductivity for each downstream path; attribution allocation coefficients Dimensionless, its numerical value quantifies the nodes What percentage of the current anomalies can be attributed to upstream nodes? The thermal conduction effect.

[0088] Meanwhile, to simulate the heat dissipation of each node during the back propagation of abnormal signals, a diagram is provided for each node. Preset heat absorption coefficient This coefficient is based on the node The types of electrical components being monitored are determined. For example, the main circuit breaker contacts are taken as empirically as 0.3 due to their large size and sufficient heat dissipation area, the small branch terminals are taken as empirically as 0.05 due to their small geometric size, and the reference point for the environment inside the box is taken as empirically as 0.01 because it is basically unaffected by conduction. All of these are dimensionless and are hyperparameters, which can also be adjusted by the implementer according to the actual heat dissipation conditions of the components. The larger the heat absorption coefficient, the higher the proportion of residual dissipated by the node in each iteration due to its own heat dissipation.

[0089] Compensation anomalies for each node generated online in step S3. Directly used as the initial residual of the corresponding node. , that is to say The residual is dimensionless, and at this point it still contains the abnormal distribution deviation caused by thermal diffusion. It is necessary to go through multiple iterations to gradually transfer the abnormal signal to the upstream heat source node.

[0090] In each iteration ( In ), for any node All its downstream nodes (i.e., there are directed edges in the original graph) The node satisfies According to the attribution coefficient The sum of the returned residuals and the node The residual from the previous iteration Add them together, then subtract the nodes. The residual portion consumed by heat dissipation The residuals after this round of updates are obtained. The update expression is:

[0091]

[0092] In the formula, For downstream nodes In the The residuals of each iteration are dimensionless. For self-node Pointing to node The attribution assignment coefficient, whose calculation method has been given in the aforementioned definition, is responsible for attributing downstream nodes. Abnormal signals are transmitted backward to upstream nodes according to the attribution ratio. ;product Simulated nodes Abnormal signals caused by heat dissipation naturally attenuate.

[0093] After each iteration, calculate all nodes. residual change maximum value If the maximum value is less than the preset convergence threshold Take experience points , dimensionless, is a hyperparameter, which can also be adjusted by the implementer according to the convergence accuracy requirements. If the number of iterations is... Reaching the preset maximum number of iterations The hyperparameter is set to an empirical value of 50, which can be adjusted according to the size of the graph. If this value is not set, the graph will be forcibly terminated to avoid infinite looping in abnormal situations.

[0094] After the iteration terminates, the final residuals of each node are calculated. The node with the largest median value is identified as the fault source node. ,Right now , The corresponding residual value is the quantification of the abnormality of the potential fault point. The node identifier and its final residual will be used as the output of this step for subsequent early warning decisions.

[0095] S5: Set dynamic early warning thresholds based on historical residual statistics.

[0096] After outputting the fault source node identifier and its final residual in step S4, it is not yet directly determined whether to trigger an alarm. This is because in industrial scenarios such as metal processing, load current fluctuations and seasonal changes can cause the residual amplitude under normal operating conditions to shift as a whole. Fixed warning thresholds are prone to misjudging the normal fluctuations of the residual as faults or missing early slow deterioration. Therefore, this step uses historical fault-free operation data accumulated before the system is put into operation to statistically analyze the background distribution characteristics of the residual and adaptively determine the dynamic warning threshold.

[0097] The historical fault-free running period is selected as a continuous period when the system confirms that there are no electrical abnormalities. The length of this period is not less than a complete load change cycle, and it typically covers at least 24 hours to ensure that the collected samples include natural changes in ambient temperature from low load to full load and day and night.

[0098] During this period, steps S1 to S4 are executed completely for each sampling moment. The final residual values ​​of all nodes are recorded after the iteration of step S4 is terminated. These values ​​are then merged into the same residual background set, regardless of which node is identified as the fault source at that moment.

[0099] After accumulating over a complete historical period, the total number of samples in the residual background set... Depends on the length of the historical period and the sampling interval The ratio, expressed as a dimensionless count value.

[0100] Calculate the arithmetic mean of all samples from the residual background set. and sample standard deviation The calculation process follows a well-known statistical definition.

[0101] Reflects the central level of the final residual under fault-free conditions. It reflects the degree of its fluctuation.

[0102] Since the residuals under normal operating conditions are mainly caused by measurement noise and minute transient changes in operating conditions, and approximately follow a one-sided skewed normal distribution, As a dynamic early warning threshold It can ensure the capture of real anomalies while keeping the false alarm probability of normal fluctuations at a low level.

[0103] Dynamic early warning threshold The formula for calculation is:

[0104]

[0105] In the formula, The mean of the residual background set is dimensionless. The standard deviation of the residual background set is dimensionless; coefficient 3 is an empirical value and a hyperparameter, which can also be adjusted by the implementer according to their tolerance for false positives and false negatives; the resulting threshold... Dimensionless.

[0106] threshold Once the calculations are completed during the historical learning phase, they are stored in the system memory and can be directly retrieved during online monitoring, eliminating the need for recalculation every cycle and ensuring consistency in the judgment criteria.

[0107] During online operation, each time step S4 identifies the fault source node at the current sampling time... and its final residual The system will immediately With threshold Compare them.

[0108] like If the abnormality of the potential fault point exceeds the normal background fluctuation range, indicating a real contact failure or local overheating hazard, the early warning module will output a fault warning signal, which will carry the fault source node information. The logo.

[0109] The node identifier is converted into a readable physical location description, such as "C-phase terminal of the 3rd branch circuit", through a pre-stored sensor number and electrical connection location mapping table. The creation and querying of this mapping table is a well-known technology and will not be described in detail here.

[0110] like If the current anomaly is within the acceptable range of normal operating conditions, no warning will be triggered, and the system will continue data collection and analysis for the next cycle.

[0111] Through the above adaptive threshold decision, the sensitivity of the warning is dynamically matched with the actual aging process of the equipment and the statistical characteristics of historical operating conditions, effectively reducing the subjective bias introduced by the fixed threshold.

[0112] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make several improvements and substitutions without departing from the technical principles of the present invention, and these improvements and substitutions should also be considered within the scope of protection of the present invention.

Claims

1. A multi-parameter fusion fault early warning method for intelligent distribution boxes, characterized in that, include: Collect temperature and current data; An initial directed topology graph is established based on sensor coordinates. The edge weights are updated online based on temperature gradient and load current change rate to generate a dynamic heat conduction topology graph. The node embedding vector is obtained by encoding the mean spatiotemporal temperature matrix using a graph convolutional network. The normal distribution center of nodes is established based on historical embedding vectors. The Euclidean distance between the current embedding vector and the center is calculated online as the initial anomaly degree, and the compensated anomaly degree is obtained by using environmental temperature compensation. Based on the dynamic heat conduction topology, the attribution allocation coefficient is defined and combined with the preset node heat absorption coefficient. The residual is iteratively propagated in the reverse heat conduction direction to compensate for the anomaly degree, and the node with the largest residual is taken as the fault source. Warnings are issued based on dynamic early warning thresholds set according to historical residual statistics.

2. The multi-parameter fusion fault early warning method for intelligent distribution boxes according to claim 1, characterized in that, The process of establishing an initial directed topology graph based on sensor coordinates includes: mapping each temperature sensor to a node in the graph, and determining the direction of the directed edges based on the physical heat transfer paths between the electrical components monitored by each sensor. Based on the thermal conductivity, effective thermal cross-sectional area, and conduction path length of the heat transfer path material, the initial thermal conductivity weight of each edge is calculated according to Fourier's law of thermal conductivity. The initial thermal conductivity weights are used as edge weights to generate an initial directed topology graph.

3. The multi-parameter fusion fault early warning method for intelligent distribution boxes according to claim 2, characterized in that, The online generation of dynamic heat conduction topology graph based on temperature gradient and load current change rate of edge weight update includes: during online operation, at each sampling time, calculating the temperature gradient normalization index between adjacent node pairs, wherein the temperature gradient normalization index is determined by the ratio of the absolute temperature difference between nodes to the preset maximum temperature difference. Calculate the load current change rate factor for each node, which is determined by the ratio of the change in current at the current sampling time relative to the previous sampling time to the rated current of the circuit. Multiply the temperature gradient normalization exponent by the load current change rate factor, multiply by the initial thermal conductivity weight, and then exponentially smooth the dynamic edge weights from the previous time step to obtain the dynamic edge weights for the current time step. A dynamic heat conduction topology graph is generated based on the updated edge weights.

4. The multi-parameter fusion fault early warning method for intelligent distribution boxes according to claim 3, characterized in that, The step of using graph convolutional network encoding to obtain the node embedding vector from the mean-reduced spatiotemporal temperature matrix includes: taking the temperature values ​​of each node from the most recent preset number of sampling times as a reference to form a spatiotemporal temperature matrix. Subtracting the mean of the temperature time series from the mean of the temperature time series at each node yields the mean-free spatiotemporal temperature matrix. The mean-removed spatiotemporal temperature matrix and the adjacency weight matrix of the dynamic heat conduction topology graph are input into the graph feature propagation network, which includes a first edge aggregation convolutional layer and a second higher-order feature mapping convolutional layer. In the first edge aggregation convolutional layer, for each node, the temperature sequence vectors of its incoming edge neighbor nodes are weighted and summed using the current dynamic edge weights, and then processed by a linear transformation of the trainable parameter matrix and a nonlinear activation function to obtain the primary feature vector. The primary feature vector is input into the second higher-order feature mapping convolutional layer for nonlinear mapping and cross-node information integration, and the intrinsic embedding vector of each node is output.

5. The multi-parameter fusion fault early warning method for intelligent distribution boxes according to claim 4, characterized in that, The establishment of a normal distribution center for nodes based on historical embedding vectors includes: collecting all intrinsic embedding vectors of each node within at least one complete load change cycle during a fault-free historical period. For each node, the arithmetic mean of the collected intrinsic embedding vectors is calculated element by element to obtain the normal operating condition embedding distribution center of that node.

6. The multi-parameter fusion fault early warning method for intelligent distribution boxes according to claim 5, characterized in that, The online calculation of the Euclidean distance between the current embedding vector and the center is used as the initial anomaly, and the compensation anomaly is obtained by using environmental temperature compensation. This includes: during online operation, whenever a new intrinsic embedding vector is obtained, the Euclidean distance between the vector and the center of the normal working condition embedding distribution of the corresponding node is calculated as the initial node anomaly. The current ambient temperature inside the chamber is obtained. Using a preset reference ambient temperature as a benchmark, the ratio of the thermodynamic temperature of the reference ambient temperature to the thermodynamic temperature of the current ambient temperature inside the chamber is calculated as a compensation factor. Multiply the initial node anomaly by the compensation factor to obtain the compensated anomaly after environmental temperature compensation.

7. The multi-parameter fusion fault early warning method for intelligent distribution boxes according to claim 1, characterized in that, The step of defining the attribution allocation coefficient based on the dynamic heat conduction topology graph and combining it with the preset node heat absorption coefficient includes: for any directed edge from the first node to the second node in the dynamic heat conduction topology graph, defining the attribution allocation coefficient from the second node to the first node, wherein the attribution allocation coefficient is equal to the ratio of the dynamic edge weight of the directed edge to the sum of the dynamic edge weights of all outgoing edges of the first node. A preset heat absorption coefficient is set for each node. The heat absorption coefficient is determined according to the type of electrical component represented by the node and is used to characterize the heat dissipation ratio of the node during residual propagation.

8. The multi-parameter fusion fault early warning method for intelligent distribution boxes according to claim 7, characterized in that, The iterative propagation of residuals in the reverse heat conduction direction for the compensation anomaly degree, with the node with the largest residual as the fault source, includes: using the compensation anomaly degree of each node as the initial residual; In each iteration, for each node, the sum of the residuals returned by all its downstream nodes according to the attribution allocation coefficient is added to the residual of the node in the previous round, and then the product of the residual of the node in the previous round and its own heat absorption coefficient is subtracted to obtain the residual after the current round. Continue iterating until the maximum value of the change in residuals of all nodes between the current round and the previous round is less than the preset convergence threshold, or until the maximum number of iterations is reached; The node with the largest final residual value is selected as the fault source identifier.

9. The multi-parameter fusion fault early warning method for intelligent distribution boxes according to claim 8, characterized in that, The method of setting a dynamic early warning threshold based on historical residual statistics includes: during historical fault-free operation, recording the final residual of each node after the end of each iteration process to form a residual background set; Calculate the mean and standard deviation of the residual background set; The mean plus three times the standard deviation is used as the dynamic early warning threshold; During online monitoring, if the final residual value corresponding to the fault source identifier exceeds the dynamic early warning threshold, a fault early warning signal is output, indicating the electrical connection location corresponding to the fault source identifier.

10. A multi-parameter fusion fault early warning system for an intelligent distribution box, characterized in that, include: The data acquisition module is used to collect temperature and current data; The dynamic topology construction module is used to establish an initial directed topology graph based on sensor coordinates and to generate a dynamic heat conduction topology graph by updating edge weights online based on temperature gradient and load current change rate. The graph convolutional coding module is used to encode the mean spatiotemporal temperature matrix using a graph convolutional network to obtain the node embedding vector. The anomaly generation module is used to establish the normal distribution center of nodes based on historical embedding vectors, calculate the Euclidean distance between the current embedding vector and the center online as the initial anomaly, and obtain the compensated anomaly using ambient temperature compensation. The fault tracing module is used to define the attribution allocation coefficient based on the dynamic heat conduction topology diagram and combine it with the preset node heat absorption coefficient to iteratively propagate the residual in the reverse heat conduction direction of the compensation anomaly degree, and take the node with the largest residual as the fault source. The early warning module is used to set dynamic early warning thresholds based on historical residual statistics.

Images