Intelligent fault early warning maintenance management system for distribution box

Abstract

The invention belongs to the technical field of distribution boxes, and discloses an intelligent fault early warning maintenance management system for a distribution box. Comprising the steps of obtaining multi-dimensional monitoring data of a full life cycle of a distribution box, and comprehensively describing a health state of the distribution box; dividing the whole life cycle into a plurality of relevance operation stages, identifying a hidden fault mode from an equipment operation event sequence by using a time sequence relevance mining algorithm, and calculating an abnormal evolution coefficient of each stage; based on the abnormal evolution coefficient and the hidden fault mode, constructing a dynamic risk entropy value of the current stage, and dynamically evaluating the fault risk; obtaining a gradient change trend of a risk entropy value, positioning a key degradation component set, generating a priority diagnosis map containing a fault root path, adaptively adjusting a strategy decision tree, and outputting an optimization maintenance scheme set; the operation reliability and the maintenance efficiency of the distribution box are obviously improved, and stable operation of a power system is guaranteed.

Description

Technical Field

[0001] The present invention relates to the technical field of distribution boxes, and more particularly to an intelligent fault early warning maintenance and management system for distribution boxes. Background Art

[0002] At present, the distribution box is a key equipment responsible for the distribution and control of electric energy in the power system. Its operating status is directly related to the stability and reliability of the entire power grid. However, there are many deficiencies in the existing operation and maintenance management of the distribution box, which brings hidden dangers to the safe operation of the power system. First, the existing technology lacks comprehensive monitoring and analysis of the entire life cycle of the distribution box. During the long-term operation of the distribution box, the components will show different degrees of aging and wear, but the existing technology cannot effectively obtain and integrate the multi-dimensional data of the distribution box, making it difficult to accurately evaluate the health status of the distribution box and unable to timely discover potential fault hazards. Secondly, the existing technology cannot deeply explore the hidden fault modes during the operation of the distribution box. The failure of the distribution box is often due to the gradual deterioration of the performance of the components, which manifests as some subtle abnormal modes, but the existing technology cannot effectively obtain and integrate the multi-dimensional data of the distribution box. It is difficult to accurately evaluate the health status of the distribution box and cannot timely discover potential fault hazards. The technology lacks the ability to identify and analyze these hidden failure modes, resulting in the neglect of failure risks and the missed opportunity for prevention and correction. Furthermore, existing technologies have difficulty accurately locating the root cause of distribution box failures and lack analysis of the fault propagation path. Once a distribution box fails, existing technologies are unable to quickly identify the key components that cause the failure and their mutual influence. Maintenance personnel can only passively wait for the failure to further expand, resulting in low maintenance efficiency and increased failure losses. Finally, existing technologies lack the ability to adaptively optimize distribution box maintenance strategies. Due to the inability to accurately assess the status of the distribution box, existing technologies often use empirical fixed maintenance cycles, which not only wastes a lot of maintenance resources, but also cannot dynamically adjust maintenance timing and plans according to the actual condition of the distribution box, resulting in low maintenance efficiency.

[0003] In view of this, the present invention proposes an intelligent fault warning maintenance and management system for distribution boxes to solve the above problems. Summary of the Invention

[0004] In order to overcome the above-mentioned shortcomings of the prior art and achieve the above-mentioned objectives, the present invention provides the following technical solutions: an intelligent fault warning maintenance and management system for a distribution box, comprising: a data integration module for acquiring multi-dimensional monitoring data of the entire life cycle of the distribution box, wherein the entire life cycle is divided into N related operation stages, each related operation stage comprising several groups of equipment operation event sequences, and the multi-dimensional monitoring data comprising an electrical parameter matrix, a mechanical state waveform, an environmental factor map, and a maintenance intervention log;

[0005] The pattern fitting module is used to analyze the hidden fault modes in several sets of equipment operation event sequences for each correlation operation stage through the time series correlation mining algorithm, and calculate the abnormal evolution coefficient of each correlation operation stage;

[0006] The risk assessment module is used to construct the dynamic risk entropy value of the current correlation operation stage based on the abnormal evolution coefficients of the previous K correlation operation stages and the hidden failure modes in the corresponding equipment operation event sequence;

[0007] A graph construction module is used to obtain the gradient change trend of the dynamic risk entropy value, locate the key degraded component set based on the gradient change trend, use the fault tree reverse deduction algorithm to match the responsibility weight of each component in the key degraded component set in historical failure cases, and generate a priority diagnosis graph containing the fault root cause path;

[0008] The maintenance module adaptively adjusts and maintains a policy decision tree based on the priority diagnostic map, and uses the policy decision tree to output an optimized maintenance plan set including predictive maintenance windows, key component replacement sequences, and environmental control parameters to the maintenance terminal; each module is connected via wired and / or wireless means to achieve data transmission between modules.

[0009] Furthermore, the electrical parameter matrix includes the voltage fluctuation harmonic spectrum, current imbalance vector and insulation resistance attenuation curve; the environmental factor map integrates the temperature and humidity time series matrix, dust particle concentration gradient and electromagnetic interference intensity distribution; the maintenance intervention log uses a blockchain structure to record the operator's identity fingerprint, spare parts replacement batch code and maintenance action timestamp.

[0010] Furthermore, the method for identifying the hidden fault mode includes:

[0011] Construct a sliding time window, the length of the sliding time window m satisfies ΔT is the preset time window length;

[0012] The device operation event sequence is segmented into multiple granularities using a sliding time window to obtain several events. An event encoder is then used to convert the events in each time window into high-dimensional feature vectors. The cosine similarity matrix of the high-dimensional feature vectors of adjacent time windows is calculated.

[0013] Based on the cosine similarity matrix of high-dimensional feature vectors of adjacent time windows, the latent fault mode across time windows is identified; the fault mode recognition threshold is preset and the similarity decay rate of adjacent time windows is calculated; the calculation formula is:

[0014] Among them, δ_k is the similarity decay rate between the kth time window and the k+1th time window, W_k is the cosine similarity matrix of the kth time window, and W_(k+1) is the cosine similarity matrix of the k+1th time window. The empty event reference template is used to calculate the distance between W_k and W_(k+1) and the empty template. DTW() is used to solve the dynamic time warping distance. If the similarity decay rate of three consecutive time windows exceeds the fault mode recognition threshold, the latent fault mode is identified, and the union of the high-dimensional feature vectors of these three consecutive time windows is used as the representation of the latent fault mode.

[0015] Furthermore, the calculation method of the abnormal evolution coefficient includes:

[0016] Define the device state space S = {s_1, s_2, s_3}, where s_1 is the stable state, s_2 is the warning state, and s_3 is the fault state. Based on the similarity decay rate of adjacent time windows, the state transition probability is calculated. The calculation formula for the state transition probability is:

[0017] Wherein, μ_ij is the mean similarity decay rate when transferring from state s_i to s_j in the historical data; i and j are both the indexes of the state in the device state space, and they do not index the same state at the same time; σ_ij is the standard deviation of the similarity decay rate when transferring from state s_i to s_j in the historical data; N(δ_k|μ_ij,σ_ij) is the probability density value of δ_k under the Gaussian distribution of μ_ij and σ_ij; P(s_i→s_j) is the state transition probability from state s_i to s_j; μ_ik is the mean similarity decay rate when transferring from state s_i to the state corresponding to the kth time window in the historical data; σ_ik is the standard deviation of the similarity decay rate when transferring from state s_i to the state corresponding to the kth time window in the historical data;

[0018] Based on the state transition probability, a calculation model for the abnormal evolution coefficient E is constructed. The calculation model is:

[0019]

[0020] Where K is the index of the current time window, P(s_2→s_3)_k is the state transition probability from the warning state s_2 to the fault state s_3 in the k-th time window; is the sum of the similarity decay rates accumulated from the 1st time window to the kth time window, and δ_t is the similarity decay rate between the tth time window and the t+1th time window.

[0021] Furthermore, the dynamic risk entropy value is constructed in the following ways:

[0022] The abnormal evolution coefficients of the first K correlated operation stages are sequenced, denoted as the coefficient sequence EL. The time domain characteristics of the coefficient sequence are extracted, including trend components, fluctuation intensity, and mutation characteristics. The mode correlation is calculated based on the characterization of latent fault modes. The cumulative effect of equipment aging, C_A, is defined based on the coefficient sequence EL and the trend component Et. The environmental disturbance coupling degree, C_E, is calculated using an improved principal component analysis method based on the environmental factor map and the coefficient sequence EL.

[0023] The dynamic risk entropy value H_risk is obtained by synthesizing the mode correlation, equipment aging cumulative effect, mode maintenance effect and environmental disturbance coupling degree;

[0024] in represents the nonlinear fusion based on Kumaraswamy bipolar aggregation operator; ω_A is the weight of the cumulative effect of equipment aging, ω_M is the weight of the mode maintenance effect, and ω_E is the weight of the environmental disturbance coupling degree; is the sigmoid function of H(a,s);

[0025] Define the distribution function of the above weights. The formula of the distribution function is:

[0026] Where x is the index of the cumulative effect of equipment aging, mode maintenance effect, or environmental disturbance coupling, that is, x∈{A,M,E}; is the partial derivative of the sample entropy value He with respect to C_x; C_x is the cumulative effect of equipment aging, mode maintenance effect or environmental disturbance coupling; softmax() is the soft maximum function.

[0027] Furthermore, the long-term trend term is obtained by decomposing the coefficient sequence through Hodrick-Prescott filtering, which is the trend component Et;

[0028] Construct a histogram of the coefficient sequence, and divide the range of the coefficient sequence into n equal intervals in the histogram; calculate the probability p_I of each interval I, then the sample entropy He = -∑(p_I×log(p_I)); the sample entropy is the fluctuation intensity; calculate the standard deviation σ_EL of the coefficient sequence, and smooth the coefficient sequence to obtain the smoothed sequence EL′;

[0029] Initialize two scan line pointers p and q, and calculate the difference ΔEq between adjacent time points in the smooth sequence. If |ΔEq| is greater than L×σ_EL, where L is a preset threshold multiple, then add the later time points in the adjacent time points to the abnormal mutation point set. Perform Mean Shift local density clustering on the preliminary abnormal mutation point set, retaining only the clustered core points. Repeat the above process until reaching the last time point in the smooth sequence. The number of core points in the final abnormal mutation point set is used as the mutation feature.

[0030] Furthermore, the calculation formula of pattern correlation is:

[0031] Among them, H(a,s) is the mode correlation between the latent fault mode a and the latent fault mode s, is the similarity function; F_a is the representation of the latent fault mode a, F_s is the representation of the latent fault mode s, FN is the service life of the equipment, SE is the design life of the equipment, k_a is the starting time point of the latent fault mode a, and k_s is the starting time point of the latent fault mode s;

[0032] Cumulative effects of equipment aging Among them, α1 is the aging weight factor, β1 is the decay rate, γ is the trend correction coefficient; T is the current time, Et is the trend component; t′ is the past time;

[0033] The maintenance effectiveness index is calculated based on the maintenance intervention log. The maintenance effectiveness index is 1-(actual maintenance interval-theoretical optimal interval) / theoretical optimal interval. The maintenance effect of the model C_M=tanh(∑ J μJ×log(NF_J+1)); where μJ is the maintenance effectiveness index of the J-th correlation operation stage, and NF_J is the number of latent failure modes in the J-th correlation operation stage;

[0034] Based on the environmental factor map and coefficient sequence EL, the improved principal component analysis method is used to calculate the environmental disturbance coupling degree C_E=||UΣV T′ ||_F×ρ(Γ,EL); where UΣV T′ is the singular value decomposition of the environmental factor map, U is the left singular matrix, V is the right singular matrix, Σ is the singular value diagonal matrix, V T′ is the transpose of the right singular matrix; ||||_F is the Frobenius norm, ρ(Γ,EL) is the Spearman correlation coefficient between the environmental parameter and the coefficient sequence EL; Γ is the environmental factor map.

[0035] Furthermore, the method for locating the key degraded component set includes:

[0036] The distribution box system consists of M components. For each component v, the partial derivative of the dynamic risk entropy value H_risk with respect to the component v is calculated. Among them, x_v represents the health status feature vector of the component v;

[0037] Forming the gradient vector based on the calculated partial derivatives Set a sliding time window with a size of W; calculate the gradient change at time R through the sliding time window The gradient change per unit time is the gradient change trend; is the gradient vector at time R, is the gradient vector at time RW; set the gradient change threshold θ_grad. For each component v, if the gradient change per unit time is greater than θ_grad, the corresponding component will be regarded as a preliminary critical degraded component; all preliminary critical degraded components constitute a preliminary critical degraded component set; calculate the correlation coefficient ρ_vu between components v and u in the preliminary critical degraded component set; construct a component correlation network graph, the nodes of the component correlation network graph are the components in the critical degraded component set, the weight of the edge is the correlation coefficient ρ_vu, and only retain the edges with |ρ_vu|>θ_corr; θ_corr is the preset correlation threshold; use the Dinic maximum flow algorithm in the correlation network graph to find densely connected component clusters, and the components in the component clusters constitute the critical degraded component set.

[0038] Furthermore, an M×B matrix G is initialized, where M is the number of components and B is the number of failure cases. For each failure case f, the set of failed components involved in the failure case, C_f, is marked. The failed component set contains several failed components. For each component c in the failed component set, the element G[c, f] at the corresponding position in the matrix G is set to 1. Then, all elements in the matrix G are obtained.

[0039] Calculate the probability of occurrence p_f for each failure case f, and for each component c, calculate its responsibility weight For each component in the critical degraded component set, if its responsibility weight is greater than the preset empirical threshold, the corresponding component is marked as a priority node and a priority diagnosis map D is constructed;

[0040] In the priority diagnosis graph D, the priority nodes are used as nodes of the priority diagnosis graph D. The edges between the priority nodes represent the fault propagation path. For each edge in the priority diagnosis graph D, a case set S is found in the historical cases where the failure of one priority node on both sides of the edge causes the failure of another priority node. If a priority node q1 causes the failure of another priority node q2, and the proportion of the case set S is greater than the preset threshold, the corresponding edge is given a direction, from the priority node q1 to the priority node q2. At this time, the edge is recorded as a directed edge. The weight of the directed edge is defined as follows: Where Δt_q1q2 is the failure time difference between priority nodes q1 and q2, and then the final priority diagnosis map D is obtained.

[0041] Furthermore, the components in the priority diagnosis map are sorted in descending order according to their responsibility weights, and the first component after sorting is used as the root node of the decision tree. For the root node, possible maintenance strategies are enumerated as child nodes of the root node. For each child node, possible maintenance strategies are continued to be enumerated as child nodes, and the above process is repeated until all components in the priority diagnosis map are covered to form an initial multi-layer decision tree; the cost function of the multi-layer decision tree is defined as the weighted sum of the quantitative values of maintenance cost and failure loss; for the current multi-layer decision tree, the value of its cost function is calculated by simulating the execution of different maintenance strategies for each node, and the heuristic search algorithm is used to optimize the multi-layer decision tree with the goal of minimizing the value of the cost function. In each iteration, the multi-layer decision tree is randomly perturbed to obtain a new decision tree for each iteration, the value of the cost function of the new decision tree is calculated, and the new decision tree is accepted or rejected according to the preset probability; the above process is repeated until the preset maximum number of iterations is reached, and an adaptively adjusted policy decision tree is obtained.

[0042] Traverse all paths of the adjusted decision tree. Each path corresponds to a maintenance plan, which includes a predictive maintenance window, a key component replacement sequence, and environmental control parameters. The predictive maintenance window is the time when the maintenance plan needs to be executed based on the current system status; the key component replacement sequence is the node on the path and its corresponding maintenance strategy; the environmental control parameter is the recommended value of the environmental parameter that needs to be adjusted based on the coupling degree of environmental disturbance; the maintenance plans corresponding to all paths are collected as output.

[0043] The technical effects and advantages of the intelligent fault early warning maintenance management system for distribution boxes of the present invention are as follows:

[0044] By integrating multi-source heterogeneous data, the present invention establishes a multi-dimensional model that comprehensively describes the health status of the distribution box, which can timely discover potential fault hazards, thereby reducing the risk of distribution box failure. At the same time, the system can also accurately locate the root cause of the fault and analyze the fault propagation path, providing an important basis for formulating targeted maintenance strategies. More importantly, it adopts an adaptive optimization algorithm, which can dynamically adjust the maintenance plan according to the real-time status of the distribution box, including predictive maintenance windows, key component replacement sequences, and environmental control parameters, thereby maximizing maintenance efficiency and extending the service life of the distribution box. It can not only significantly reduce the economic losses caused by distribution box failures, but also reduce power supply interruptions caused by failures, and improve the stability and reliability of the power grid. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 This is a schematic diagram of an intelligent fault warning maintenance and management system for distribution boxes according to the present invention;

[0046] Figure 2 This is a schematic diagram of an intelligent fault warning maintenance and management method for a distribution box according to the present invention. DETAILED DESCRIPTION

[0047] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0048] Example 1

[0049] See also Figure 1 As shown, the intelligent fault warning maintenance and management system for distribution boxes described in this embodiment includes:

[0050] A data integration module is used to obtain multi-dimensional monitoring data for the entire life cycle of the distribution box. The entire life cycle is divided into N related operation stages, each of which contains several sets of equipment operation event sequences. The multi-dimensional monitoring data includes electrical parameter matrices, mechanical status waveforms, environmental factor maps, and maintenance intervention logs.

[0051] The pattern fitting module is used to analyze the hidden fault modes in several sets of equipment operation event sequences for each correlation operation stage through the time series correlation mining algorithm, and calculate the abnormal evolution coefficient of each correlation operation stage;

[0052] The risk assessment module is used to construct the dynamic risk entropy value of the current correlation operation stage based on the abnormal evolution coefficients of the previous K correlation operation stages and the hidden failure modes in the corresponding equipment operation event sequence;

[0053] A graph construction module is used to obtain the gradient change trend of the dynamic risk entropy value, locate the key degraded component set based on the gradient change trend, use the fault tree reverse deduction algorithm to match the responsibility weight of each component in the key degraded component set in historical failure cases, and generate a priority diagnosis graph containing the fault root cause path;

[0054] The maintenance module adaptively adjusts and maintains a policy decision tree based on the priority diagnostic map, and uses the policy decision tree to output an optimized maintenance plan set including predictive maintenance windows, key component replacement sequences, and environmental control parameters to the maintenance terminal; each module is connected via wired and / or wireless means to achieve data transmission between modules.

[0055] Dividing the entire life cycle of a distribution box into N related operating stages is intended to better capture and analyze the state changes and failure modes of the distribution box at different operating periods. Specifically, from commissioning to scrapping, the distribution box will experience different working environments, load conditions, and maintenance interventions. These factors will cause the performance of each component of the distribution box to gradually degrade and may trigger a series of related failure events. Dividing the entire life cycle into multiple related operating stages, each stage containing a sequence of equipment operation events over a period of time, can more finely characterize the operating status of the distribution box in that stage.

[0056] For example, the initial stage of operation of the distribution box may belong to a correlation operation stage, during which the equipment performance is good and the operation event sequence mainly reflects normal switch operations; then it enters the high-load operation period, which may be divided into another correlation operation stage, during which some minor faults and maintenance records may appear in the operation event sequence; then it enters the aging period, which may be divided into a new correlation operation stage, during which fault events and maintenance interventions in the operation event sequence will be more frequent.

[0057] This division allows for better correlation and exploration of failure modes within the same stage, as well as the discovery of the evolution patterns of failure modes between different stages. For example, a latent failure mode discovered during the aging stage may be due to the neglect of minor failures in previous stages. Furthermore, the abnormal evolution coefficients of different stages can be analyzed to assess whether the risk of failure is accelerating.

[0058] In addition, dividing the entire life cycle into related operation stages is also conducive to combining the actual operating environment and maintenance records of the distribution box to establish a more accurate health status assessment model, providing a basis for formulating targeted predictive maintenance strategies.

[0059] Traditional distribution box monitoring systems mostly use single-dimensional electrical parameter collection and lack cross-modal data fusion of mechanical vibration, environmental factors and maintenance records, resulting in one-sided health status assessment.

[0060] The electrical parameter matrix includes the voltage fluctuation harmonic spectrum, current imbalance vector and insulation resistance attenuation curve; the mechanical state waveform collects the contactor action impact response characteristics through a three-axis vibration sensor; the environmental factor map integrates the temperature and humidity time series matrix, dust particle concentration gradient and electromagnetic interference intensity distribution; the maintenance intervention log uses a blockchain structure to record the operator's identity fingerprint, spare parts replacement batch code and maintenance action timestamp.

[0061] Specifically, a power quality analyzer is connected to the distribution box busbar to perform high-precision sampling of the voltage waveform and perform a fast Fourier transform (FFT) on the sampled data to obtain the voltage fluctuation harmonic spectrum. A Hall current sensor is installed in each phase current loop of the distribution box to sample the three-phase current in real time and calculate the imbalance of the three-phase current. The imbalance can be expressed as a negative sequence component, a zero sequence component, or the ratio of the maximum value to the average value. The three-phase imbalance components are combined to form an imbalance vector, which is the current imbalance vector.

[0062] Use a high-voltage megohmmeter to measure the insulation resistance of the distribution box, perform time series measurement of the insulation resistance, and obtain the insulation resistance attenuation curve over time, which is the insulation resistance attenuation curve.

[0063] A triaxial acceleration sensor is installed on the contactor component of the distribution box. When the contactor operates, the acceleration sensor collects the vibration waveform of the action impact response, which is the mechanical state waveform;

[0064] Temperature and humidity sensors are arranged inside the distribution box to collect temperature and humidity data in real time, and a two-dimensional temperature and humidity time series matrix is constructed. A dust particle counting sensor is installed at the air inlet of the distribution box to monitor the changes in dust particle concentration at the air inlet in real time, and calculate the gradient distribution of dust concentration in space and time; this is the dust particle concentration gradient; an electromagnetic field strength probe is arranged inside the distribution box to collect the electromagnetic field intensity distribution inside the distribution box in real time, and the electromagnetic field intensity is filtered in the time and space domain to remove high-frequency noise, which is the electromagnetic interference intensity distribution.

[0065] The multi-source heterogeneous data obtained above are integrated into a unified time and space coordinate system through data fusion to form multi-dimensional monitoring data that comprehensively describes the health status of the distribution box; through the multi-source heterogeneous data fusion of electrical parameter matrices, mechanical status waveforms, environmental factor maps and blockchain maintenance logs, full life cycle monitoring is formed in a unified time and space coordinate system, breaking through the limitations of traditional single-dimensional diagnosis and providing three-dimensional data support for hidden fault analysis.

[0066] Conventional fault pattern recognition methods rely on event statistics in fixed time windows, which makes it difficult to capture implicit correlation features across time windows, and lack quantitative indicators for the abnormal evolution process.

[0067] Construct a sliding time window. The length m of the sliding time window is dynamically adjusted according to the device type and satisfies: ΔT is the preset time window length;

[0068] The device operation event sequence is segmented into multiple granularities using a sliding time window to obtain several events. An event encoder is then used to convert the events in each time window into high-dimensional feature vectors. The cosine similarity matrix of the high-dimensional feature vectors of adjacent time windows is calculated.

[0069] Based on the cosine similarity matrix of high-dimensional feature vectors of adjacent time windows, the latent fault mode across time windows is identified. The identification methods of latent fault mode include:

[0070] The fault mode recognition threshold is preset and the similarity decay rate of adjacent time windows is calculated; the calculation formula is:

[0071] Among them, δ_k is the similarity decay rate between the kth time window and the k+1th time window, W_k is the cosine similarity matrix of the kth time window, and W_(k+1) is the cosine similarity matrix of the k+1th time window. The empty event reference template is used to calculate the distance between W_k and W_(k+1) and the empty template. DTW() is used to solve the dynamic time warping distance. If the similarity decay rate of three consecutive time windows exceeds the fault mode recognition threshold, the latent fault mode is identified, and the union of the high-dimensional feature vectors of these three consecutive time windows is used as the representation of the latent fault mode.

[0072] The calculation methods of the anomaly evolution coefficient include:

[0073] Define the device state space S = {s_1, s_2, s_3}, where s_1 is the stable state, s_2 is the warning state, and s_3 is the fault state. Based on the similarity decay rate of adjacent time windows, the state transition probability is calculated. The calculation formula for the state transition probability is:

[0074] Among them, μ_ij is the mean similarity decay rate when transferring from state s_i to s_j in historical data (distribution box monitoring data and fault records accumulated during past operation); i and j are both indexes of states in the device state space, and they do not index the same state at the same time; σ_ij is the standard deviation of the similarity decay rate when transferring from state s_i to s_j in historical data; N(δ_k|μ_ij,σ_ij) is the probability density value of δ_k under the Gaussian distribution of μ_ij and σ_ij; P(s_i→s_j) is the state transition probability from state s_i to s_j; μ_ik is the mean similarity decay rate when transferring from state s_i to the state corresponding to the kth time window in historical data; σ_ik is the standard deviation of the similarity decay rate when transferring from state s_i to the state corresponding to the kth time window in historical data.

[0075] Based on the state transition probability, a calculation model for the abnormal evolution coefficient E is constructed. The calculation model is:

[0076]

[0077] Where K is the index of the current time window, P(s_2→s_3)_k is the state transition probability from the warning state s_2 to the fault state s_3 in the k-th time window; is the sum of the similarity decay rates accumulated from the 1st time window to the kth time window, and δ_t is the similarity decay rate between the tth time window and the t+1th time window.

[0078] Dynamic time warping algorithm and multi-granularity event coding are adopted to detect continuous abnormal features through the cosine similarity decay rate of the sliding time window. The abnormal evolution coefficient is quantified by combining the state transition probability model to achieve explainable modeling of hidden fault modes.

[0079] Typically, existing risk assessment models use static weight allocation and are unable to adapt to the dynamic coupling effects of environmental disturbances, equipment aging, and maintenance interventions.

[0080] The abnormal evolution coefficients of the first K correlation operation stages are sequenced and recorded as coefficient sequence EL. The time domain characteristics of the coefficient sequence are extracted. The time domain characteristics include trend component, fluctuation intensity and mutation characteristics.

[0081] The long-term trend term, namely the trend component Et, is obtained by decomposing the coefficient sequence through Hodrick-Prescott filtering; the sample entropy He of the coefficient sequence is calculated. Specifically, a histogram of the coefficient sequence is constructed, and the value range of the coefficient sequence is equally divided into n intervals in the histogram; the probability p_I of each interval I is calculated (time point in the interval / total time points), then the sample entropy He = -∑(p_I×log(p_I)); the sample entropy is the fluctuation intensity, reflecting the complexity of the abnormal pattern.

[0082] Calculate the standard deviation σ_EL of the coefficient sequence, smooth the coefficient sequence to obtain the smoothed sequence EL′; (the smoothing method can be moving average, exponential smoothing or wavelet transform, etc.); initialize two scan line pointers p and q; (p points to the current point, q points to the next point), calculate the difference ΔEq between adjacent time points of the smoothed sequence; if |ΔEq| is greater than L×σ_EL, where L is a preset threshold multiple, usually 3; then add the later time points in the adjacent time points to the abnormal mutation point set; perform Mean Shift local density clustering on the preliminary abnormal mutation point set, retaining only the core points after clustering, and repeat the above process until reaching the last time point of the smoothed sequence; the number of core points in the final abnormal mutation point set is used as the mutation feature.

[0083] The mode correlation is calculated based on the characterization of the latent fault mode. The calculation formula of the mode correlation is:

[0084] Among them, H(a,s) is the mode correlation between the latent fault mode a and the latent fault mode s, is a similarity function, such as cosine similarity, Euclidean distance, etc.; F_a is the representation of latent fault mode a, F_s is the representation of latent fault mode s, FN is the service life of the equipment, SE is the design life of the equipment, k_a is the starting time point of latent fault mode a, and k_s is the starting time point of latent fault mode s.

[0085] Define the equipment aging cumulative effect C_A based on the coefficient sequence EL and trend component Et;

[0086] Among them, α1 is the aging weight factor, the experimental optimal value is 0.7, β1 is the decay rate, the experimental optimal value is 0.05; γ is the trend correction coefficient, the experimental optimal value is 0.3; T is the current time, Et is the trend component; t′ is the past time; Tt′ is the time difference from the current time T, that is, how long the past time t′ is from now;

[0087] The maintenance effectiveness index is calculated based on the maintenance intervention log. The maintenance effectiveness index is 1-(actual maintenance interval-theoretical optimal interval) / theoretical optimal interval. The maintenance effect of the model C_M=tanh(∑ J μJ×log(NF_J+1)); where μJ is the maintenance effectiveness index of the J-th correlation operation stage, and NF_J is the number of latent failure modes in the J-th correlation operation stage;

[0088] Based on the environmental factor map and coefficient sequence EL, the improved principal component analysis method is used to calculate the environmental disturbance coupling degree C_E=||UΣVT′ ||_F×ρ(Γ,EL); where UΣV T′ is the singular value decomposition of the environmental factor map, U is the left singular matrix, V is the right singular matrix, Σ is the singular value diagonal matrix, V T′ is the transpose of the right singular matrix; ||||_F is the Frobenius norm, ρ(Γ,EL) is the Spearman correlation coefficient between the environmental parameter and the coefficient sequence EL; Γ is the environmental factor map.

[0089] The dynamic risk entropy value H_risk is obtained by synthesizing the mode correlation, equipment aging cumulative effect, mode maintenance effect and environmental disturbance coupling degree;

[0090] in It represents the nonlinear fusion based on Kumaraswamy bipolar aggregation operator, which operates as follows: ω_A is the weight of the cumulative effect of equipment aging, ω_M is the weight of the mode maintenance effect, and ω_E is the weight of the environmental disturbance coupling degree; is the sigmoid function of H(a,s);

[0091] Define the distribution function of the above weights. The formula of the distribution function is:

[0092] Where x is the index of the cumulative effect of equipment aging, mode maintenance effect, or environmental disturbance coupling, that is, x∈{A,M,E}; is the partial derivative of the sample entropy value He with respect to C_x; C_x is the cumulative effect of equipment aging, mode maintenance effect or environmental disturbance coupling; softmax() is the soft maximum function, which maps a set of input values to the interval (0, 1) and satisfies that the sum of all values is 1. It is often used as a probability output in multi-classification problems.

[0093] Traditional fault location relies on expert experience rule bases, which makes it difficult to handle multi-component coupling degradation scenarios, and the correlation analysis between historical cases and real-time data is insufficient.

[0094] The distribution box system consists of M components. For each component v (1≤v≤M), the partial derivative of the dynamic risk entropy value H_risk with respect to the component v is calculated. Among them, x_v represents the health state characteristic vector of the component v; the health state characteristic vector is obtained through the electrical parameter matrix; specifically, the statistical characteristics of the voltage fluctuation harmonic spectrum are extracted, and the statistical characteristics include the fundamental amplitude, total harmonic distortion (THD), main harmonic order and main harmonic amplitude; these statistical characteristics are used as part of the health state characteristic vector; the current imbalance vector is used as part of the health state characteristic vector.

[0095] Insulation characteristics, including the initial insulation resistance R0 and the time constant, are extracted from the insulation resistance decay curve. These insulation characteristics are used as part of x_v. These parts are sequentially concatenated to form a health status feature vector. This fully utilizes multiple electrical parameters, such as voltage, current, and insulation, to comprehensively describe the electrical health status of the component and extract physically meaningful statistics and parameters that are easy to understand and analyze.

[0096] Forming the gradient vector based on the calculated partial derivatives Set a sliding time window with a size of W; calculate the gradient change at time R through the sliding time window The amount of gradient change per unit time is the gradient change trend. By monitoring the change trend, it can be determined whether the system risk is accelerating. is the gradient vector at time R, is the gradient vector at time RW.

[0097] Set the gradient change threshold θ_grad. For each component v, if the gradient change per unit time is greater than θ_grad, the corresponding component will be regarded as a preliminary key degraded component. All preliminary key degraded components constitute the preliminary key degraded component set.

[0098] Calculate the correlation coefficient ρ_vu between components v and u in the preliminary critical degraded component set; construct a component correlation network graph, where the nodes of the component correlation network graph are the components in the critical degraded component set, the edge weights are the correlation coefficient ρ_vu, and only retain edges with |ρ_vu|>θ_corr; θ_corr is the preset correlation threshold; use the Dinic maximum flow algorithm in the correlation network graph to find densely connected component clusters, and the components in the component clusters constitute the critical degraded component set; the health status of the components in these critical degraded component sets is the key factor leading to the rapid growth of system risks.

[0099] Existing maintenance strategies mostly adopt fixed-cycle plans, which do not consider the coordinated optimization of component degradation priority and environmental regulation, resulting in low resource utilization.

[0100] Initialize an M×B matrix G, where M is the number of components and B is the number of fault cases. For each fault case f, mark the set of failed components C_f involved in the fault case. The failed component set contains several failed components.

[0101] For each component c in the set of failed components, set the element G[c, f] at the corresponding position in the matrix G to 1; then obtain all the elements in the matrix G, indicating whether the component failed in the fault case;

[0102] Calculate the probability of occurrence p_f for each failure case f (based on historical data statistics), and for each component c, calculate its responsibility weight That is, for all fault cases involving the component, the inverse of the log-likelihood probability is weighted summed.

[0103] For each component in the critical degraded component set, if its responsibility weight is greater than the preset empirical threshold, the corresponding component is marked as a priority node and a priority diagnosis map D is constructed;

[0104] In the priority diagnosis graph D, the priority nodes are used as nodes of the priority diagnosis graph D (representing components), and the edges between the priority nodes represent the fault propagation path. For each edge in the priority diagnosis graph D, a case set S is found in the historical cases where the failure of one priority node on both sides of the edge causes the failure of another priority node; if a priority node q1 causes the failure of another priority node q2, when the proportion in the case set S is greater than the preset threshold, the corresponding edge is given a direction, from the priority node q1 to the priority node q2, and the edge at this time is recorded as a directed edge; the weight of the directed edge is defined Where Δt_q1q2 is the failure time difference between priority nodes q1 and q2.

[0105] A correlation matrix between components and historical failure cases was established; then, based on the correlation matrix, the comprehensive responsibility weight of each component in all related failure cases was calculated, reflecting the importance of the component to the system failure.

[0106] The resulting priority diagnosis map D is a directed weighted graph containing the root cause path of the fault, which intuitively reflects the propagation logic and importance ranking of component failures in the system; it provides an important reference for formulating targeted maintenance strategies.

[0107] Traditional maintenance plans lack the ability to predict risk evolution trends, and the maintenance window settings and component replacement sequences are subjective and arbitrary.

[0108] The components in the priority diagnosis map are sorted in descending order according to their responsibility weights. The first component after sorting is used as the root node of the decision tree. For the root node, possible maintenance strategies (such as replacement, repair, monitoring, etc.) are enumerated as child nodes of the root node. For each child node, possible maintenance strategies are continued to be enumerated as child nodes. The above process is repeated until all components in the priority diagnosis map are covered, forming an initial multi-layer decision tree.

[0109] The cost function of the multi-layer decision tree is defined as the weighted sum of the quantified values of maintenance cost and failure loss. For the current multi-layer decision tree, the value of its cost function is calculated by simulating the execution of different maintenance strategies for each node. A heuristic search algorithm (such as simulated annealing) is used to optimize the multi-layer decision tree with the goal of minimizing the value of the cost function. In each iteration, the multi-layer decision tree is randomly perturbed (such as exchanging subtrees, modifying maintenance strategies, etc.) to obtain a new decision tree for each iteration. The value of the cost function of the new decision tree is calculated, and the new decision tree is accepted or rejected according to a preset probability. The above process is repeated until the preset maximum number of iterations is reached, and an adaptively adjusted strategy decision tree is obtained.

[0110] Traverse all paths of the adjusted decision tree (root node to leaf node). Each path corresponds to a maintenance plan, which includes a predictive maintenance window, a key component replacement sequence, and environmental control parameters. The predictive maintenance window is the prediction of when the maintenance plan needs to be executed based on the current system status; the key component replacement sequence is the node on the path and its corresponding maintenance strategy (replacement, repair, etc.); the environmental control parameter is the recommended value of the environmental parameter that needs to be adjusted based on the degree of environmental disturbance coupling; the maintenance plans corresponding to all paths are collected as the output.

[0111] Specifically, based on the current dynamic risk entropy value and its gradient change trend, a risk entropy value threshold and a gradient change threshold are set. Starting from the current time point, the predicted risk entropy value and gradient change amount at each future time point are recursively calculated; a time series model (such as ARIMA) is used to model the risk entropy value sequence and predict the risk entropy value; based on the predicted risk entropy value sequence, the gradient change amount at adjacent time points is calculated; the first time point that satisfies the risk entropy value greater than the risk entropy value threshold and the gradient change amount greater than the gradient change threshold is determined. This time point is the predictive maintenance window period, which is the inflection point where the system risk begins to accelerate.

[0112] Find a path P from the root node to the leaf node in the priority diagnosis graph D. Each node in the path P corresponds to a component. The node strategy is the maintenance strategy of the component. According to the order of the nodes in the path P, a key component replacement sequence is formed.

[0113] Calculate the matrix EP of the environmental disturbance coupling degree at the current time point, perform singular value decomposition on the matrix EP to obtain the left and right singular matrices and singular values, analyze the singular values, and find the left and right singular vectors u1 and v1 corresponding to the largest k1 singular values. u1 reflects the principal component of the environmental factor, and v1 reflects the projection of the coefficient sequence on the principal component. According to the weight of each environmental factor in u1, determine the environmental parameters that need to be regulated first. According to the sign of v1 (positive or negative), determine the regulation direction (increase or decrease). The environmental parameters that need to be regulated first and the regulation direction are used as the recommended output of the environmental regulation parameters, that is, the recommended values of the environmental parameters that need to be adjusted.

[0114] This process converts the component importance information in the priority diagnostic map into a decision tree. By adaptively adjusting the minimization cost function, it obtains a set of multiple candidate optimized maintenance plans and gives specific maintenance windows, component repair sequences, and environmental control recommendations for each plan, providing a comprehensive reference for actual maintenance decisions.

[0115] In this embodiment, by integrating multi-source heterogeneous data, a multi-dimensional model that comprehensively describes the health status of the distribution box is established, which can timely discover potential fault hazards, thereby reducing the risk of distribution box failure. At the same time, the system can also accurately locate the root cause of the fault and analyze the fault propagation path, providing an important basis for formulating targeted maintenance strategies. More importantly, an adaptive optimization algorithm is adopted, which can dynamically adjust the maintenance plan according to the real-time status of the distribution box, including predictive maintenance windows, key component replacement sequences, and environmental control parameters, thereby maximizing maintenance efficiency and extending the service life of the distribution box. It can not only significantly reduce the economic losses caused by distribution box failures, but also reduce power supply interruptions caused by failures, and improve the stability and reliability of the power grid.

[0116] Example 2

[0117] See also Figure 2 As shown, for the parts not described in detail in this embodiment, please refer to the description of Example 1. A method for intelligent fault warning maintenance and management of a distribution box is provided, comprising:

[0118] Step 1: Acquire multi-dimensional monitoring data for the entire life cycle of the distribution box. The entire life cycle is divided into N related operation stages, each of which contains several sets of equipment operation event sequences. The multi-dimensional monitoring data includes an electrical parameter matrix, mechanical state waveforms, environmental factor maps, and maintenance intervention logs.

[0119] Step 2: For each correlation operation stage, the hidden fault modes in several sets of equipment operation event sequences are analyzed using a time series correlation mining algorithm, and the abnormal evolution coefficient of each correlation operation stage is calculated;

[0120] Step 3: Based on the abnormal evolution coefficients of the previous K correlation operation stages and the hidden failure modes in the corresponding equipment operation event sequence, the dynamic risk entropy value of the current correlation operation stage is constructed;

[0121] Step 4: Obtain the gradient change trend of the dynamic risk entropy value, locate the key degraded component set based on the gradient change trend, use the fault tree reverse derivation algorithm to match the responsibility weight of each component in the key degraded component set in historical failure cases, and generate a priority diagnosis map containing the fault root cause path;

[0122] Step 5: Based on the priority diagnosis map, adaptively adjust and maintain a policy decision tree, and use the policy decision tree to output an optimized maintenance plan set including predictive maintenance windows, key component replacement sequences, and environmental control parameters to the maintenance terminal.

[0123] Example 3

[0124] This embodiment discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the operation mode of the intelligent fault warning maintenance management method for a distribution box provided above is implemented.

[0125] Since the electronic device introduced in this embodiment is an electronic device used to implement a distribution box intelligent fault warning maintenance management method in the embodiment of this application, based on the distribution box intelligent fault warning maintenance management method introduced in the embodiment of this application, technical personnel in this field can understand the specific implementation of the electronic device of this embodiment and its various variations, so how the electronic device implements the method in the embodiment of this application will not be introduced in detail here. As long as technical personnel in this field implement the electronic device used in the distribution box intelligent fault warning maintenance management method in the embodiment of this application, it falls within the scope of protection of this application.

[0126] The above formulas are all dimensionless and numerical calculations. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters and thresholds in the formulas are set by technicians in this field according to actual conditions.

[0127] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiment. All technical solutions based on the principles of the present invention are within the scope of protection of the present invention. It should be noted that for users of ordinary skill in the art, various improvements and modifications that do not depart from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. An intelligent fault warning maintenance and management system for distribution boxes, characterized in that: include: A data integration module is used to obtain multi-dimensional monitoring data for the entire life cycle of the distribution box. The entire life cycle is divided into N related operation stages, each of which contains several sets of equipment operation event sequences. The multi-dimensional monitoring data includes electrical parameter matrices, mechanical status waveforms, environmental factor maps, and maintenance intervention logs. The pattern fitting module is used to analyze the hidden fault modes in several sets of equipment operation event sequences for each correlation operation stage through the time series correlation mining algorithm, and calculate the abnormal evolution coefficient of each correlation operation stage; The risk assessment module is used to construct the dynamic risk entropy value of the current correlation operation stage based on the abnormal evolution coefficients of the previous K correlation operation stages and the hidden failure modes in the corresponding equipment operation event sequence; A graph construction module is used to obtain the gradient change trend of the dynamic risk entropy value, locate the key degraded component set based on the gradient change trend, use the fault tree reverse deduction algorithm to match the responsibility weight of each component in the key degraded component set in historical failure cases, and generate a priority diagnosis graph containing the fault root cause path; The maintenance module adaptively adjusts and maintains a policy decision tree based on the priority diagnostic map, and uses the policy decision tree to output an optimized maintenance plan set including predictive maintenance windows, key component replacement sequences, and environmental control parameters to the maintenance terminal; each module is connected via wired and / or wireless means to achieve data transmission between modules.

2. The intelligent fault warning maintenance and management system for distribution boxes according to claim 1 is characterized in that: The electrical parameter matrix includes voltage fluctuation harmonic spectrum, current imbalance vector and insulation resistance attenuation curve; The environmental factor map integrates the temperature and humidity time series matrix, dust particle concentration gradient and electromagnetic interference intensity distribution; the maintenance intervention log uses a blockchain structure to record the operator's identity fingerprint, spare parts replacement batch code and maintenance action timestamp.

3. The intelligent fault warning maintenance and management system for distribution boxes according to claim 2 is characterized in that: The identification method of the hidden failure mode includes: Construct a sliding time window, the length of the sliding time window m satisfies ΔT is the preset time window length; The device operation event sequence is segmented into multiple granularities using a sliding time window to obtain several events. An event encoder is then used to convert the events in each time window into high-dimensional feature vectors. The cosine similarity matrix of the high-dimensional feature vectors of adjacent time windows is calculated. Based on the cosine similarity matrix of high-dimensional feature vectors of adjacent time windows, the latent fault mode across time windows is identified; the fault mode recognition threshold is preset and the similarity decay rate of adjacent time windows is calculated; the calculation formula is: and the similarity decay rate between the k+1th time window, W_k is the cosine similarity matrix of the kth time window, W_(k+1) is the cosine similarity matrix of the k+1th time window, The empty event reference template is used to calculate the distance between W_k and W_(k+1) and the empty template. DTW() is used to solve the dynamic time warping distance. If the similarity decay rate of three consecutive time windows exceeds the fault mode recognition threshold, the latent fault mode is identified, and the union of the high-dimensional feature vectors of these three consecutive time windows is used as the representation of the latent fault mode.

4. The intelligent fault warning maintenance and management system for distribution boxes according to claim 3 is characterized in that: The calculation method of the abnormal evolution coefficient includes: Define the device state space S = {s_1, s_2, s_3}, where s_1 is the stable state, s_2 is the warning state, and s_3 is the fault state. Based on the similarity decay rate of adjacent time windows, the state transition probability is calculated. The calculation formula for the state transition probability is: Wherein, μ_ij is the mean similarity decay rate when transferring from state s_i to s_j in the historical data; i and j are both the indexes of the state in the device state space, and they do not index the same state at the same time; σ_ij is the standard deviation of the similarity decay rate when transferring from state s_i to s_j in the historical data; N(δ_k|μ_ij,σ_ij) is the probability density value of δ_k under the Gaussian distribution of μ_ij and σ_ij; P(s_i→s_j) is the state transition probability from state s_i to s_j; μ_ik is the mean similarity decay rate when transferring from state s_i to the state corresponding to the kth time window in the historical data; σ_ik is the standard deviation of the similarity decay rate when transferring from state s_i to the state corresponding to the kth time window in the historical data; Based on the state transition probability, a calculation model for the abnormal evolution coefficient E is constructed. The calculation model is: Where K is the index of the current time window, P(s_2→s_3)_k is the state transition probability from the warning state s_2 to the fault state s_3 in the k-th time window; is the sum of the similarity decay rates accumulated from the 1st time window to the kth time window, and δ_t is the similarity decay rate between the tth time window and the t+1th time window.

5. The intelligent fault warning maintenance and management system for distribution boxes according to claim 4 is characterized in that: The dynamic risk entropy value is constructed in the following ways: The abnormal evolution coefficients of the first K correlated operation stages are sequenced, denoted as the coefficient sequence EL. The time domain characteristics of the coefficient sequence are extracted, including trend components, fluctuation intensity, and mutation characteristics. The mode correlation is calculated based on the characterization of latent fault modes. The cumulative effect of equipment aging, C_A, is defined based on the coefficient sequence EL and the trend component Et. The environmental disturbance coupling degree, C_E, is calculated using an improved principal component analysis method based on the environmental factor map and the coefficient sequence EL. The dynamic risk entropy value H_risk is obtained by synthesizing the mode correlation, equipment aging cumulative effect, mode maintenance effect and environmental disturbance coupling degree; in represents the nonlinear fusion based on Kumaraswamy bipolar aggregation operator; ω_A is the weight of the cumulative effect of equipment aging, ω_M is the weight of the mode maintenance effect, and ω_E is the weight of the environmental disturbance coupling degree; is the sigmoid function of H(a,s); Define the distribution function of the above weights. The formula of the distribution function is: Where x is the index of the cumulative effect of equipment aging, mode maintenance effect, or environmental disturbance coupling, that is, x∈{A,M,E}; is the partial derivative of the sample entropy value He with respect to C_x; C_x is the cumulative effect of equipment aging, mode maintenance effect or environmental disturbance coupling; softmax() is the soft maximum function.

6. The intelligent fault warning maintenance and management system for distribution boxes according to claim 5, characterized in that: The long-term trend term is obtained by decomposing the coefficient sequence through Hodrick-Prescott filtering, which is the trend component Et; Construct a histogram of the coefficient sequence, and divide the range of the coefficient sequence into n intervals in the histogram; Calculate the probability p_I of each interval I, then the sample entropy He = -∑(p_I×log(p_I)); the sample entropy is the fluctuation intensity; calculate the standard deviation σ_EL of the coefficient sequence, smooth the coefficient sequence, and obtain the smoothed sequence EL′; Initialize two scan line pointers p and q, and calculate the difference ΔEq between adjacent time points in the smooth sequence. If |ΔEq| is greater than L×σ_EL, where L is a preset threshold multiple, then add the later time points in the adjacent time points to the abnormal mutation point set. Perform Mean Shift local density clustering on the preliminary abnormal mutation point set, retaining only the clustered core points. Repeat the above process until reaching the last time point in the smooth sequence. The number of core points in the final abnormal mutation point set is used as the mutation feature.

7. The intelligent fault warning maintenance and management system for distribution boxes according to claim 6, characterized in that: The calculation formula of pattern correlation is: Among them, H(a,s) is the mode correlation between the latent fault mode a and the latent fault mode s, is the similarity function; F_a is the representation of the latent fault mode a, F_s is the representation of the latent fault mode s, FN is the service life of the equipment, SE is the design life of the equipment, k_a is the starting time point of the latent fault mode a, and k_s is the starting time point of the latent fault mode s; Cumulative effects of equipment aging Among them, α1 is the aging weight factor, β1 is the decay rate, γ is the trend correction coefficient; T is the current time, Et is the trend component; t′ is the past time; The maintenance effectiveness index is calculated based on the maintenance intervention log. The maintenance effectiveness index is 1-(actual maintenance interval-theoretical optimal interval) / theoretical optimal interval. The maintenance effect of the model C_M=tanh(∑ J μJ×log(NF_J+1)); where μJ is the maintenance effectiveness index of the J-th correlation operation stage, and NF_J is the number of latent failure modes in the J-th correlation operation stage; Based on the environmental factor map and coefficient sequence EL, the improved principal component analysis method is used to calculate the environmental disturbance coupling degree C_E=||UΣV T′ ||_F×ρ(Γ,EL); where UΣV T′ is the singular value decomposition of the environmental factor map, U is the left singular matrix, V is the right singular matrix, Σ is the singular value diagonal matrix, V T′ is the transpose of the right singular matrix; ||||_F is the Frobenius norm, ρ(Γ,EL) is the Spearman correlation coefficient between the environmental parameter and the coefficient sequence EL; Γ is the environmental factor map.

8. The intelligent fault warning maintenance and management system for distribution boxes according to claim 7, characterized in that: Methods for locating the key degraded component set include: The distribution box system consists of M components. For each component v, the partial derivative of the dynamic risk entropy value H_risk with respect to the component v is calculated. Among them, x_v represents the health status feature vector of the component v; Forming the gradient vector based on the calculated partial derivatives Set the sliding time window to W; calculate the gradient change at time R through the sliding time window. The gradient change per unit time is the gradient change trend; is the gradient vector at time R, is the gradient vector at time RW; set the gradient change threshold θ_grad. For each component v, if the gradient change per unit time is greater than θ_grad, the corresponding component will be regarded as a preliminary critical degraded component; all preliminary critical degraded components constitute a preliminary critical degraded component set; calculate the correlation coefficient ρ_vu between components v and u in the preliminary critical degraded component set; construct a component correlation network graph, the nodes of the component correlation network graph are the components in the critical degraded component set, the weight of the edge is the correlation coefficient ρ_vu, and only retain the edges with |ρ_vu|>θ_corr; θ_corr is the preset correlation threshold; use the Dinic maximum flow algorithm in the correlation network graph to find densely connected component clusters, and the components in the component clusters constitute the critical degraded component set.

9. The intelligent fault warning maintenance and management system for distribution boxes according to claim 8, characterized in that: Initialize an M×B matrix G, where M is the number of components and B is the number of failure cases. For each failure case f, mark the set of failed components C_f involved in the failure case. The failed component set contains several failed components. For each component c in the set of failed components, set the element G[c, f] at the corresponding position in the matrix G to 1; then obtain all the elements in the matrix G; Calculate the probability of occurrence p_f for each failure case f, and for each component c, calculate its responsibility weight For each component in the critical degraded component set, if its responsibility weight is greater than the preset empirical threshold, the corresponding component is marked as a priority node and a priority diagnosis map D is constructed; In the priority diagnosis graph D, the priority nodes are used as nodes of the priority diagnosis graph D. The edges between the priority nodes represent the fault propagation path. For each edge in the priority diagnosis graph D, a case set S is found in the historical cases where the failure of one priority node on both sides of the edge causes the failure of another priority node. If a priority node q1 causes the failure of another priority node q2, and the proportion of the case set S is greater than the preset threshold, the corresponding edge is given a direction, from the priority node q1 to the priority node q2. At this time, the edge is recorded as a directed edge. The weight of the directed edge is defined as follows: Where Δt_q1q2 is the failure time difference between priority nodes q1 and q2, and then the final priority diagnosis map D is obtained.

10. The intelligent fault warning maintenance and management system for distribution boxes according to claim 9, characterized in that: The components in the priority diagnosis map are sorted in descending order according to their responsibility weights, and the first component after sorting is used as the root node of the decision tree. For the root node, possible maintenance strategies are enumerated as child nodes of the root node. For each child node, possible maintenance strategies are continued to be enumerated as child nodes. The above process is repeated until all components in the priority diagnosis map are covered to form an initial multi-layer decision tree; the cost function of the multi-layer decision tree is defined as the weighted sum of the quantitative values of maintenance cost and failure loss; for the current multi-layer decision tree, different maintenance strategies are executed on each node by simulation, and the value of its cost function is calculated. The heuristic search algorithm is used to optimize the multi-layer decision tree, with the goal of minimizing the value of the cost function. In each iteration, the multi-layer decision tree is randomly perturbed to obtain a new decision tree for each iteration, and the value of the cost function of the new decision tree is calculated. The new decision tree is accepted or rejected according to the preset probability; the above process is repeated until the preset maximum number of iterations is reached, and an adaptively adjusted policy decision tree is obtained; Traverse all paths in the adjusted decision tree. Each path corresponds to a maintenance plan, which includes a predictive maintenance window, a key component replacement sequence, and environmental control parameters. The predictive maintenance window is the predicted time when the maintenance plan needs to be executed based on the current system status. The key component replacement sequence is the node on the path and its corresponding maintenance strategy. The environmental control parameters are given as recommended values of the environmental parameters that need to be adjusted according to the degree of coupling of environmental disturbances; the maintenance plans corresponding to all paths are collected as output.

Images