A method for predicting the remaining lifespan of intelligent power distribution boxes on construction sites

Abstract

This invention discloses a method for predicting the remaining lifespan of intelligent distribution boxes on construction sites, relating to the field of industrial intelligence. The method includes: filtering historical distribution box lifecycle data related to faults based on fault diagnosis results; constructing a degradation trajectory library; selecting state features from a structured dataset using feature engineering and constructing state feature vectors; selecting a sandbox environment as a virtual testing platform; configuring a reinforcement learning environment based on the degradation trajectory library and state feature vectors; training a reinforcement learning model in the sandbox environment; obtaining the degradation trajectory model; calculating the expected time between the current state and failure of the intelligent distribution box; and outputting the remaining lifespan prediction result. This invention achieves dynamic quantitative expression of parameter causal relationships, solving the problem of low accuracy in predicting the remaining lifespan of intelligent distribution boxes.

Description

Technical Field

[0001] This invention relates to the field of industrial intelligence, and in particular to a method for predicting the remaining life of intelligent power distribution boxes on construction sites. Background Technology

[0002] In recent years, the lifespan monitoring of distribution boxes has shifted from routine manual inspections to automated data analysis. The application of industrial-grade sensors and data processing technologies has enabled real-time monitoring of multi-dimensional parameters. Some studies have introduced time series analysis to capture the dynamic changes in the operating status of distribution boxes, while reinforcement learning technology has also provided new ideas for predicting remaining lifespan. By simulating the equipment degradation process in a virtual environment, maintenance strategies can be optimized to extend the lifespan of distribution boxes.

[0003] Existing methods for predicting the remaining life of smart distribution boxes have several areas for improvement. Many conventional life prediction methods rely on static datasets or threshold analysis of single parameters, lacking modeling of the dynamic degradation relationships between multidimensional parameters. This makes them difficult to adapt to load fluctuations and climate changes in construction site environments. Furthermore, existing reinforcement learning models often neglect the guiding role of fault diagnosis in remaining life prediction, resulting in insufficient sensitivity to key degradation patterns. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a method for predicting the remaining life of intelligent power distribution boxes on construction sites, which solves the problems of insufficient dynamic degradation modeling and low accuracy of remaining life prediction for intelligent power distribution boxes on construction sites.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a method for predicting the remaining life of intelligent distribution boxes on construction sites, comprising: collecting data from intelligent distribution boxes and preprocessing it to obtain a structured dataset;

[0008] By using causal discovery algorithms to analyze structured datasets, causal relationships between the states of internal components of intelligent distribution boxes are identified. A directed acyclic graph is constructed, and reverse tracing and intervention effect analysis are performed to generate fault diagnosis results.

[0009] Based on the fault diagnosis results, historical distribution box lifecycle data related to the fault are screened, a degradation trajectory library is constructed, and feature engineering is used to select state features from the structured dataset and construct a state feature vector.

[0010] A sandbox environment is selected as the virtual testing platform. A reinforcement learning environment is configured based on the degradation trajectory library and state feature vectors. The reinforcement learning model is trained in the sandbox environment to obtain the degradation trajectory model and calculate the expected time between the current state and failure of the smart distribution box. The remaining lifetime prediction result is then output.

[0011] As a preferred embodiment of the remaining life prediction method for intelligent distribution boxes on construction sites according to the present invention, the intelligent distribution box data includes leakage current, circuit breaker coil current, contactor contact temperature, bus voltage, insulation resistance, ambient humidity, and contactor resistance.

[0012] As a preferred embodiment of the remaining life prediction method for intelligent power distribution boxes on construction sites according to the present invention, the method for obtaining the structured dataset specifically includes the following steps:

[0013] The data of the intelligent distribution box is denoised by using discrete wavelet transform to obtain the denoised intelligent distribution box data.

[0014] Cubic spline interpolation was used to fill in the denoised smart distribution box data, and the filled smart distribution box data was then standardized to obtain a structured dataset.

[0015] As a preferred embodiment of the remaining life prediction method for intelligent power distribution boxes on construction sites according to the present invention, the construction of the directed acyclic graph specifically includes the following steps:

[0016] Collect historical fault data, and select the parameters with the highest frequency of occurrence of the internal component status of the smart distribution box from the structured dataset and historical fault data as selected variables;

[0017] Set a standardization threshold; when the calculated Pearson correlation coefficient between the selected variables is less than the standardization threshold, a pair of selected variables is obtained.

[0018] Set an independence threshold, perform a chi-square test on the selected variable pairs, and determine the causal edge of the selected variable pairs when the calculated chi-square statistic of the selected variable is less than the set independence threshold.

[0019] By treating the selected variable as a node and connecting it to the causal edges of the filtered variable pairs, a directed acyclic graph is obtained.

[0020] As a preferred embodiment of the remaining life prediction method for intelligent power distribution boxes on construction sites according to the present invention, the generation of the fault diagnosis report specifically includes the following steps:

[0021] The directed acyclic graph is optimized by using time series data from historical fault data, and weights are assigned to the causal edges of nodes and filtering variable pairs to obtain a dynamic causal graph.

[0022] Abnormal signals from smart distribution boxes are detected using structured datasets and mapped to nodes in a dynamic cause-effect graph to identify abnormal nodes.

[0023] The breadth-first search algorithm is used to traverse the abnormal nodes in reverse, obtain the reverse path, and collect the starting nodes of all reverse paths as potential sources of failure. The confidence value of the potential sources of failure is then calculated.

[0024] Set an intervention effect threshold. When the calculated intervention effect value from the abnormal node to the potential source of failure is greater than the intervention effect threshold, the core node is identified.

[0025] The core nodes, potential sources of failure, and their confidence values ​​are compiled into JSON format to generate fault diagnosis results.

[0026] As a preferred embodiment of the remaining life prediction method for intelligent power distribution boxes on construction sites according to the present invention, the step of constructing a degradation trajectory library and selecting state features from a structured dataset using feature engineering to construct a state feature vector specifically includes the following steps:

[0027] Based on the core nodes and potential fault sources in the fault diagnosis results, historical distribution box lifecycle data related to the fault are screened and missing values ​​are filled to generate degradation curves;

[0028] Based on the annotation of core events in the historical power distribution box lifecycle data related to faults, the health of the smart power distribution box is calculated and the smart power distribution box is divided into health regions. The degradation curve and the health regions are combined to form a degradation trajectory library.

[0029] Calculate the feature importance score for each parameter in the structured dataset, and select the top M parameters with the highest feature importance scores for each parameter as state features;

[0030] Calculate the feature values ​​of the state features and convert them into z-score values. Arrange the z-score values ​​in a fixed order to form the state feature vector.

[0031] As a preferred embodiment of the remaining life prediction method for intelligent power distribution boxes on construction sites according to the present invention, the step of obtaining the degradation trajectory model specifically includes the following steps:

[0032] Configure the reinforcement learning environment based on the degenerate trajectory library and state feature vectors;

[0033] A deep Q-network was selected as the agent. The agent was trained in a sandbox environment according to the reinforcement learning environment. At the same time, the feature values ​​of the state features were mapped using the degradation curves in the degradation trajectory library to generate virtual time series.

[0034] By updating the feature values ​​of the state characteristics and the health level of the health region through virtual time series, the state transition of the intelligent distribution box between different health regions is repeatedly simulated until the agent reaches convergence, and then the degradation trajectory model is obtained.

[0035] As a preferred embodiment of the remaining life prediction method for intelligent power distribution boxes on construction sites according to the present invention, the output of the remaining life prediction result specifically includes the following steps:

[0036] Calculate the state transition probability of the intelligent distribution box;

[0037] The state feature vector and the health level in the health region are input into the degradation trajectory model. The input is matched with the virtual time series. Based on the state transition probability of the smart distribution box, the expected single-step transition time of the smart distribution box is calculated and accumulated to obtain the total expected transition time of the smart distribution box. The remaining lifetime prediction result is then output.

[0038] In a second aspect, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the remaining life prediction method for intelligent power distribution boxes at construction sites as described in the first aspect of the present invention.

[0039] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the remaining life prediction method for intelligent power distribution boxes at construction sites as described in the first aspect of the present invention.

[0040] The beneficial effects of this invention are as follows: This invention constructs a directed acyclic graph by analyzing structured datasets, optimizes edge directions and assigns weights through time lag analysis to generate a dynamic causal graph, detects abnormal signals in the distribution box and traces back to generate fault diagnosis results, realizing the dynamic quantitative expression of causal relationships between multidimensional parameters and the accurate location of fault sources. Through causal discovery and dynamic causal graphs, it accurately captures the causal mechanism in the degradation process of the distribution box, greatly improving the depth and pertinence of dynamic degradation modeling. Furthermore, by using a deep Q-network to train the degradation trajectory model, calculating the remaining lifetime through state transition probabilities and outputting prediction results, it achieves dynamic simulation of the entire life cycle of the distribution box and accurate prediction of the remaining lifetime, significantly improving the accuracy and environmental adaptability of the remaining lifetime prediction of the intelligent distribution box, effectively reducing prediction errors in harsh construction site environments, and extending the life of the intelligent distribution box. Attached Figure Description

[0041] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0042] Figure 1 This is a flowchart of the method for predicting the remaining lifespan of intelligent power distribution boxes on construction sites according to the present invention.

[0043] Figure 2 A flowchart for causal discovery and dynamic causal graph construction.

[0044] Figure 3 A flowchart generated from the fault diagnosis results.

[0045] Figure 4 This is a flowchart for training the degradation trajectory model and predicting remaining lifetime. Detailed Implementation

[0046] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0047] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0048] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0049] Reference Figure 1 As one embodiment of the present invention, this embodiment provides a method for predicting the remaining life of a smart distribution box on a construction site, comprising the following steps:

[0050] S1 collects data from the smart distribution box and preprocesses it to obtain a structured dataset.

[0051] Includes the following steps:

[0052] S1.1, A sensor array is installed inside the intelligent distribution box on the construction site. The sensor array uses industrial-grade equipment, including current sensors, temperature sensors, voltage sensors, humidity sensors, and vibration sensors. All sensors are set to the same sampling frequency to capture transient signals. All sensors are installed in fixed positions within the intelligent distribution box; that is, the current and voltage sensors are fixed to the input terminals of the circuit breaker in the distribution box, the temperature sensor is attached to the contact surface of the contactor in the distribution box, and the humidity and vibration sensors are placed in the internal environmental monitoring points of the distribution box. The collected data from the intelligent distribution box includes leakage current, circuit breaker coil current, contactor contact temperature, bus voltage, insulation resistance, ambient humidity, and contactor resistance.

[0053] S1.2, preprocess the data from the smart distribution box.

[0054] Specifically, a discrete wavelet transform algorithm is used to denoise the smart distribution box data. This involves using the Daubechies wavelet (D4) as the basis function and setting the decomposition level to 5 to filter out high-frequency noise, such as noise from random signals caused by electrical interference or mechanical vibration. The smart distribution box data is segmented, and wavelet decomposition is performed on each segment, including retaining low-frequency signals and removing high-frequency noise. The signal is then reconstructed through inverse transform to obtain the denoised smart distribution box data.

[0055] Data imputation is performed on the denoised smart distribution box data. Since the denoised smart distribution box data contains missing data, such as missing values ​​due to sensor disconnection or communication interruption, cubic spline interpolation (algorithm) is used for imputation. Specifically, based on the ten data points before and after each missing value in the denoised smart distribution box data, a cubic polynomial is used to fit these ten data points into a smooth curve, thereby filling in the missing value and obtaining the imputed smart distribution box data.

[0056] Step S1.3 involves standardizing the augmented smart distribution box data. Specifically, the z-score method is used. For each parameter in the augmented smart distribution box data, such as the circuit breaker coil current, the mean and standard deviation over a past period are calculated. Several data points are selected from the augmented smart distribution box data; the mean of each parameter is subtracted, and then divided by the standard deviation, converting the data into dimensionless values ​​with a mean of 0 and a standard deviation of 1. The z-score method is applied to each parameter, and the standardization process forms a multi-dimensional feature vector, such as [current z-score: 1.2, temperature z-score: -0.45, voltage z-score: 0.78, ...]. The multi-dimensional feature vector is indexed by timestamps and saved in JSON format, outputting a structured dataset.

[0057] S2 uses a causal discovery algorithm to analyze structured datasets, identify causal relationships between the states of internal components of the smart distribution box, construct a directed acyclic graph, perform reverse tracing and intervention effect analysis, and generate fault diagnosis results.

[0058] See Figure 2 and Figure 3 This includes the following steps:

[0059] S2.1 Collect historical fault data. The historical fault data comes from the operation records of multiple smart distribution boxes of the same model, including the labeling of fault events, such as circuit breaker aging, abnormal contact resistance, insulation and deterioration, etc.

[0060] Parameters of the internal component status of the smart distribution box are selected as selected variables from structured datasets and historical fault data, such as contactor contact temperature, circuit breaker coil current and insulation resistance. The selection of variables depends largely on the highest frequency of the fault-related parameters of the distribution box in the historical fault data and structured datasets. For example, circuit breaker aging and abnormal contact resistance have occurred hundreds of times, and have the highest frequency among all parameters, so they are included in the priority selection.

[0061] S2.2, The Pearson correlation coefficient between the selected variables is calculated using the PC algorithm, and the expression is:

[0062] ;

[0063] in, This represents the Pearson correlation coefficient between the selected variables, with a numerical range of -1 to 1. This represents the total number of sampling points for the selected variable. Indicates the index of the sampling point, ranging from 1 to... , This indicates that the variable representing the selected internal component of the first smart distribution box is in the [missing information]. The observed values ​​at each sampling point, such as the circuit breaker coil current at the th sampling point. At each sampling point, the observed value was 10A (where A represents the unit of current). Indicates the selected variable for the internal components of the second intelligent distribution box in the _____th ... Observations at each sampling point This represents the average observation of selected variables within the first smart distribution box over the past hour. This represents the average observed value of the selected variables of the internal components of the second smart distribution box over the past hour. The selection of one hour as the time window for calculating the average value is based on the typical periodicity of the operation of the distribution box on the construction site, such as the load change of the distribution box, which is usually measured in hours.

[0064] A standardized threshold is set according to statistical convention: a Pearson correlation coefficient less than 0.05 is considered highly correlated at a 90%–95% confidence level. When the Pearson correlation coefficient between selected variables is less than the standardized threshold, the selected variables are retained as strongly correlated pairs; when the Pearson correlation coefficient is greater than or equal to the standardized threshold, they are discarded, and the filtered variable pairs are output.

[0065] S2.3, Perform a chi-square test on the selection variable pairs and calculate the chi-square statistic of the selection variables. The expression is:

[0066] ;

[0067] ;

[0068] in, The chi-square statistic represents the correlation between selection variables. The larger the value, the smaller the p-value, indicating a higher correlation between the selected variables. This represents the observational efficiency between the screening variables, i.e., the actual joint efficiency of the screening variables in historical fault data. For example, if two screening variables are selected, such as abnormal current and increased resistance, the two screening variables can be divided into a 2×2 contingency table. It refers to the total number of actual observations for each cell in the contingency table, that is, the number of times when the current is abnormal and the resistance also increases. This indicates the expected efficiency among the selection variables. This represents the total number of sampling points for the selected variable. The total number of outliers in the first filter variable. This represents the total number of sampling points for the selected variable. The total number of outliers in the second filter variable.

[0069] An independence threshold is set based on conventions in statistics and causal discovery, with a value of 0.01 (which can be adjusted according to actual circumstances). This means that in the PC algorithm, conditional independence testing requires high confidence, as incorrect causal edges can lead to biases in fault analysis. A value of 0.01 is more stringent than the standard level threshold, suitable for scenarios involving intelligent distribution boxes on construction sites, where distribution box faults, such as increased contact resistance due to circuit breaker aging, involve high risks and require ensuring the reliability of causal relationships.

[0070] When the chi-square statistic of a screening variable is less than the independence threshold, it indicates that there is no conditional independence between the screening variables, meaning there is a correlation. In this case, the correlation between the screening variables is retained as an undirected edge between the screening variable pairs. For example, after temperature control, the p-value for current and contact resistance is 0.005 < 0.01, and the undirected edge is X—Y, where X represents current and Y represents contact resistance. A direction is assigned to the undirected edge. For instance, if 80%–90% of abnormal events in historical fault data show that current abnormalities precede increases in contact resistance, then the causal edge between the screening variable pairs is determined to be X→Y. When the chi-square statistic of a screening variable is greater than or equal to the independence threshold, it indicates that there is conditional independence between the screening variables, meaning there is no correlation. In this case, the correlation between the screening variables is eliminated.

[0071] S2.4. Select the variables as nodes, and connect the causal edges between the nodes and the selected variable pairs to form a directed acyclic graph.

[0072] Specifically, nodes are viewed as vertices in a graph structure, each representing the state of a component within the intelligent distribution box. Causal edges, representing selected variable pairs, are directed connections, indicating the causal relationships between the states of the internal components. Nodes are connected by these causal edges, forming graph paths, such as circuit breaker coil current → increased contact resistance → contactor contact temperature. A depth-first search (DFS) algorithm is used to check the acyclicity of the graph paths. Starting from each node, all causal edges are traversed sequentially, recording the visited paths. If a node is visited again in the current path, a cycle exists, and the causal edge with the highest p-value is removed. DFS re-checks after removal until no cycles are found.

[0073] S2.5 Optimize the directed acyclic graph using time series data from historical fault data and assign weights to the causal edges of nodes and filtering variable pairs to obtain a dynamic causal graph.

[0074] Specifically, the time series in historical fault data consists of timestamps and corresponding component state variable values, such as circuit breaker coil current, contactor contact temperature, and insulation resistance, reflecting the changes in the operating status of the intelligent distribution box over time. The time series specifically includes abnormal events that occur in the intelligent distribution box, the frequency of fault occurrence, and the proportion of causal events.

[0075] Optimization of directed acyclic graphs (DAGs) includes optimizing the edge directions of the DAG through time lag analysis. This involves extracting abnormal events from the time series, such as current > 10.5A and contact resistance > 0.1Ω, and statistically analyzing the time difference of these abnormal events. When 80% to 90% of the abnormal events satisfy the condition that a certain variable occurs first, for example, the time of current abnormality is 3 seconds later than the time of contact resistance, the edge direction is updated, i.e., current abnormality → contact resistance increase.

[0076] Assign weights to nodes in a directed acyclic graph. This involves counting the number of times a selected variable appears in a fault event based on the frequency of distribution box faults in the time series. For example, if the distribution box experiences current anomalies 400 times out of 1000 fault events, then the node is assigned a weight of 0.4.

[0077] Assign weights to the causal edges of the selected variable pairs in the directed acyclic graph. Based on the occurrence ratio of causal events in the time series, count the abnormal events that occur in the smart distribution box. For example, if 700 out of 1000 events of increased contact resistance are caused by abnormal circuit breaker coil current, then the weight of this causal edge is 0.8.

[0078] The optimized directed acyclic graph forms a dynamic causal graph, which includes nodes, causal edges of the filter variable pairs, node weights, and causal edge weights of the filter variable pairs. The dynamic causal graph is stored in GraphML format.

[0079] S2.6 Detect abnormal signals from the distribution box using a structured dataset, and map these abnormal signals to nodes in the dynamic cause-effect graph to obtain the abnormal nodes.

[0080] Specifically, an abnormal signal in a distribution box refers to a parameter value among the selected variables that exceeds the normal range, indicating a potential fault or operational abnormality in the distribution box. For example, the insulation resistance under normal conditions is greater than 5MΩ, and the bus voltage fluctuation is within the normal range of ±2V. When the parameter value exceeds the normal range, it is determined to be an abnormal signal in the distribution box and is matched with a node in the dynamic cause-effect diagram. For example, if an abnormal signal is excessive leakage current, this abnormal signal is matched with a node in the dynamic cause-effect diagram to form a leakage current node.

[0081] Breadth-first search (BFS) is used to traverse the abnormal nodes in reverse to obtain the reverse path. The starting nodes of all reverse paths are collected as potential sources of failure, and the confidence value of the potential sources of failure is calculated.

[0082] Specifically, the process identifies input edges pointing to anomalous nodes, such as insulation aging → leakage current, moisture intrusion → leakage current, and poor grounding → leakage current. Breadth-first search (BFS) is used to automatically backtrack from the current anomalous node layer by layer to the source node, i.e., the node without input edges. For example, the first layer finds insulation aging and moisture intrusion, and the second layer finds increased contact resistance → insulation aging. After traversal, each reverse path is recorded, such as leakage current ← insulation aging ← increased contact resistance, and the path length of the reverse path is set to avoid excessively long reverse paths reducing confidence. All reverse paths are collected together as potential fault sources for each reverse path, and the confidence value of each potential fault source is calculated by multiplying the weights of the causal edges assigned in the dynamic cause-effect graph.

[0083] The intervention effect value from the abnormal node to the potential source of failure (i.e., the causal path) is calculated using the following expression:

[0084] ;

[0085] in, This represents the intervention effect value from the abnormal node to the potential source of the fault, and indicates the magnitude of the causal effect of the intervention variable on the outcome variable. The resulting variables, i.e., the values ​​of the abnormal nodes, such as leakage current, are obtained from the structured dataset. This represents the causal variable, that is, the value of the potential source of failure, such as insulation aging. 1 indicates that aging has occurred, and 0 indicates that it has not occurred. Indicates to Intervention, coercion A value of 1 indicates isolation. right causal effect Indicates to Intervention, coercion A value of 0 is used to compare the situation without intervention. value, Indicates under intervention The expected value, that is, under forced The value is 1, for example, when insulation aging occurs. The average value, for example, in historical fault data, the average leakage current when insulation aging occurs is 0.1mA, then It is 0.1mA. Indicates under intervention The expected value, that is, the forced The value is 0, assuming there is no insulation aging. The average value, for example, if the average leakage current is 0.02mA without insulation aging, then It is 0.02mA.

[0086] An intervention effect threshold is set dynamically based on the actual needs of the construction site's power distribution box faults. Typically, the threshold ranges from 0.8, suitable for high-risk scenarios. When the intervention effect value from the abnormal node to the potential fault source exceeds the threshold, i.e., based on the causal path with an intervention effect value greater than 0.8, the core node is identified. The potential fault source, its confidence value, and the core node are then compiled into a JSON format to obtain the fault diagnosis result.

[0087] S3. Based on the fault diagnosis results, filter historical distribution box lifecycle data related to the fault, construct a degradation trajectory library, and use feature engineering to select state features from the structured dataset and construct a state feature vector.

[0088] See Figure 4This includes the following steps:

[0089] S3.1, based on the core nodes and potential fault sources in the fault diagnosis results, filter historical lifecycle data of distribution boxes related to the fault from the cloud database, including timestamps and core event annotations, such as operating records related to insulation aging or increased contact resistance of smart distribution boxes, to ensure that the main fault modes are reflected. Extract core events, including the first occurrence of anomalies, critical states, and failure points of the distribution box. The first occurrence of anomalies refers to, for example, the first situation where the current exceeds the normal value range; critical states refer to, for example, the critical state where the insulation resistance is about to reach the normal value range; and failure points refer to the situation where the distribution box shuts down.

[0090] The missing values ​​in the lifecycle data of historical distribution boxes are filled using cubic spline interpolation to generate degradation curves.

[0091] Specifically, missing segments in the historical distribution box lifecycle data are identified. Sampling points with timestamps approximately ten seconds before and after each missing segment are taken. Based on these timestamps, missing values ​​are automatically filled using a cubic polynomial, and a degradation curve is generated. The horizontal axis of the degradation curve represents the operating time span of the distribution box in seconds, and the vertical axis represents the health of the distribution box, ranging from 0 to 1, where 1 represents brand new and 0 represents failed. The health of the distribution box is calculated based on the core event annotations in the historical distribution box lifecycle data, expressed as:

[0092] ;

[0093] in, Indicates the time of the intelligent distribution box Health status of the area This indicates that the smart distribution box was in the previous time step health This represents the core event set in the lifecycle data of historical distribution boxes. Indicates the core event The severity coefficient is set as follows: for example, 0.1 is set for the first abnormality of the distribution box, 0.3 is set for the critical state, and 1 is set directly when the failure point is reached. Indicates the core event The weights are set based on the frequency of occurrence of core events in the historical lifecycle data of the distribution box. For example, the first anomaly is set to 10%, the critical state to 30%, and the failure point to 60%. Indicates the time when the core event occurred. This represents the time decay constant, used to reflect the duration of the degradation effects on the distribution box. The specific value is set to 10. 6 Seconds, the reason is when Too small, for example, 104 If the time frame is too short, the impact of core events will decay too quickly, leading to an overestimation of health. It will also cause the predicted lifespan of the distribution box to be too long, resulting in unnecessary errors. The same applies to excessively large numbers, such as 10. 8 The time frame of seconds makes the impact of core events too long to reflect the rate of degradation of the distribution box, resulting in low health and an excessively short predicted lifespan. This represents an exponentially decaying function, with a value ranging from 0 to 1, and represents the core event. Regarding time The extent to which the impact of reduced health decays over time.

[0094] Based on the health status of the distribution box, the distribution box is divided into health status regions. The degradation curve and the health status regions are combined to form a degradation trajectory library.

[0095] Specifically, a health threshold is set according to the IEC 61439 standard for distribution box reliability. The value range of the health threshold can be dynamically adjusted according to the actual situation of the distribution box. Assuming the health threshold range is 0.5~0.8, when the health of the distribution box is greater than or equal to 0.8, it means that most distribution boxes are at a low failure rate level, and the remaining lifespan is generally greater than 6 months. These distribution boxes are classified as green zone. When the health of the distribution box is less than 0.8 but greater than 0.5, it means that the distribution box has a failure problem, and the remaining lifespan is mostly around 2~6 months. These distribution boxes are classified as yellow zone. When the health of the distribution box is less than or equal to 0.5, the working capability of the distribution box is close to failure, and the remaining lifespan is very low, usually 2 months or even less. These distribution boxes are classified as red zone.

[0096] S3.2, use feature engineering to extract state features from the structured dataset and construct them into a state feature vector.

[0097] Specifically, using the feature importance ranking method, the feature importance score for each parameter in the structured dataset is calculated, as expressed by:

[0098] ;

[0099] in, The feature importance score represents the parameter, ranging from 0 to 1. This indicates the number of times a parameter occurs in a fault event. This can be achieved by iterating through the fault event annotations in historical fault data and counting the number of times the parameter occurs outside its normal value range. This indicates the total number of fault events. This indicates the frequency of a parameter's occurrence within the health zone. It is obtained by dividing the number of anomalies of the parameter in the historical fault data time series by the total number of time points in the health zone. For example, if the red zone has 100 million points and insulation resistance anomalies occurred 50 million times, the frequency would be 0.5. The weight of the health zone is represented by a value ranging from 0 to 1. It can be set according to the actual severity of degradation of the smart distribution box. For example, the weight is 0.2 for the green zone, 0.3 for the yellow zone, and 0.5 for the red zone.

[0100] The top M(7) parameters in terms of feature importance score for each parameter are selected as state features. State features include the rate of increase of contactor contact temperature, the increase of contactor resistance, the fluctuation range of bus voltage, the abnormal rate of circuit breaker coil current, the rate of decrease of insulation resistance, the mean of leakage current, and the peak value of ambient humidity.

[0101] Calculate the eigenvalues ​​of the state characteristics.

[0102] Specifically, the rate of increase of the contactor contact temperature is calculated, which is the rate of change of the contactor contact temperature during the operation time of the distribution box. The specific values ​​of the contactor contact temperature are extracted from the structured dataset, the difference between the maximum and minimum values ​​of the contactor contact temperature is calculated, and the difference is divided by the actual operation time of the distribution box to obtain the rate of increase of the contactor contact temperature.

[0103] To calculate the fluctuation range of the bus voltage, we need to calculate the average value of the bus voltage in the structured dataset, calculate the difference between the bus voltage value and the average value, square the difference, sum the results, and then divide by the total number of bus voltage values.

[0104] Calculate the abnormal frequency of the circuit breaker coil current. This involves iterating through all the circuit breaker coil current values ​​and counting the number of times the circuit breaker coil current exceeds the normal range. For example, if 10 abnormalities are found and the time range of the circuit breaker coil current values ​​is 1 hour, then the abnormal frequency of the circuit breaker coil current is 10 times per hour.

[0105] Calculate the rate of decrease in insulation resistance. Specifically, take the insulation resistance values ​​from the structured dataset. Assuming the time range of the insulation resistance values ​​is 1 hour, take the start and end values ​​of the insulation resistance values ​​within the 1-hour range. Subtract the start value from the end value and divide by the number of data points within the time range of the insulation resistance values ​​to obtain the rate of decrease in insulation resistance.

[0106] Calculate the mean leakage current. Based on the leakage current values ​​in the structured dataset, sum all leakage current values ​​and divide by the number of data points within the time range of the leakage current values ​​to obtain the average leakage current.

[0107] Calculate the peak value of ambient humidity. That is, iterate through the ambient humidity values ​​in the structured dataset and find the largest ambient humidity value as the peak value of ambient humidity.

[0108] Calculate the contactor resistance increment. Specifically, take all contactor resistance values ​​from the structured dataset, and subtract the initial value from the end value of each contactor resistance value to obtain the contact resistance increment. For example, if the end value of the contactor resistance value is 55 milliohms and the initial value is 50 milliohms, then the contactor resistance increment is 5 milliohms.

[0109] S3.3, the z-score method is used to convert the feature values ​​of the state features into z-score values, and the z-score values ​​are arranged in a fixed order to form the state feature vector.

[0110] Specifically, the eigenvalues ​​of all state features are summarized, the mean and standard deviation of each eigenvalue are calculated, and the z-score method is used to convert a eigenvalue into a z-score value. The expression is as follows:

[0111] ;

[0112] in, The z-score value represents the transformed feature value of a certain state feature. The feature value represents a characteristic of a certain state. The mean of the feature values ​​representing a certain state characteristic. The standard deviation of the eigenvalues ​​representing a certain state characteristic.

[0113] Set the sorting order as follows: contactor contact temperature rise rate, bus voltage fluctuation range, circuit breaker coil current abnormal frequency, insulation resistance decrease rate, average leakage current, peak ambient humidity, and contactor resistance increment. Arrange the seven feature values ​​into an array format according to their sorting order, such as [0.5, -0.3, 0.8, -0.1, 0.2, 0.9, 0.4], corresponding to the z-score values ​​of the seven state features, and store them in JSON format. Output the state feature vector.

[0114] S4. Select the sandbox environment as the virtual test platform. Based on the degradation trajectory library and state feature vector, train the reinforcement learning model in the sandbox environment to obtain the degradation trajectory model and predict the remaining life of the distribution box, and output the remaining life prediction result.

[0115] Includes the following steps:

[0116] S4.1, Construct a reinforcement learning environment based on the degenerate trajectory library and state feature vectors.

[0117] Specifically, the reinforcement learning environment is defined by its state space, action space, and reward function. The state space includes state feature vectors and health zones. The action space includes detection frequency and maintenance cycle. Detection frequency refers to the frequency at which the distribution box sensors collect data, ranging from 1Hz to 100Hz. This needs to be comprehensively defined based on the actual operating requirements and hardware configuration of the distribution box. The lower limit is the minimum detection frequency, such as once per second, suitable for high-health zones (e.g., green zones). The upper limit is the maximum detection frequency, such as 100 times per second, suitable for low-health zones (e.g., red zones), capable of capturing rapidly changing abnormal signals such as sudden increases in leakage current. The maintenance cycle refers to the preventative maintenance interval of the distribution box, which needs to be set according to the operating characteristics of the distribution box, typically ranging from 7 days to 90 days, with 7 days as the lower limit and 90 days as the upper limit. The maintenance cycle defines the actions the agent schedules for maintenance, such as checking circuit breakers, contactors, and other components of the distribution box every 30 days to extend their lifespan or prevent failures.

[0118] The reward function includes the reward objective, namely the minimum maintenance cost and failure loss of the distribution box. A positive reward is defined as 0.5 units of reward points when the agent passes through high-frequency detection or short maintenance cycles to avoid distribution box failure. When the agent is in the green zone of the health status and selects low-frequency detection or long maintenance cycles, 0.2 units of reward points are set to ensure the health of the distribution box and low resource consumption. A negative reward is defined as deducting a certain amount of reward points, such as 0.3, based on the minimum maintenance cost of the distribution box whenever maintenance is required. If the distribution box completely fails, 1 unit of reward points is directly deducted. If the distribution box completely fails during the rainy season or extreme weather, more reward points are deducted, such as 1.5 (the reinforcement learning environment is configured in a sandbox environment).

[0119] S4.2, a deep Q-network is selected as the agent. The agent is trained in a sandbox environment according to the reinforcement learning environment. At the same time, the degradation curves in the degradation trajectory library are used to simulate the entire life cycle of the distribution box to obtain the degradation trajectory model.

[0120] Specifically, an intelligent agent refers to a decision-making entity in reinforcement learning, specifically a computational model of the Deep Q-Network (DQN) algorithm. By observing the state of the distribution box, it selects actions in the action space and learns optimization strategies based on the feedback of the reward function to extend the life of the distribution box and reduce maintenance costs. The essence of an intelligent agent is a software program used to simulate a decision-maker.

[0121] Configure the agent by setting a randomized update strategy, such as a detection frequency of 5 minutes and a maintenance cycle of 90 days. After configuration, by observing the status of the distribution box in the sandbox environment, including the state feature vector and the partition labels in the health zone, the agent adjusts the parameters of the deep Q network based on the reward feedback for the performed action, such as +0.5 or -1, prioritizing high-reward actions. For example, the detection frequency in the yellow zone is adjusted to 10Hz and the maintenance time to 7 days, while the detection frequency is increased to 15Hz during the rainy season.

[0122] The simulated lifecycle of a distribution box is as follows: The core events in the degradation curve are used to map the feature values ​​of seven state characteristics. The mapping rules are statistically derived from a degradation trajectory library. For example, 95% of the yellow zone distribution boxes experience a temperature rise rate of 0.08~0.12 degrees Celsius per minute, generating a time series with a sampling frequency of approximately 10kHz. For health status, cubic spline interpolation is used to fill the degradation curve, generating continuous values, such as 0.9 in year 2 to 0.6 in year 4. The feature values ​​of the state characteristics are randomly sampled according to the health status partition. For example, the temperature rise rate in the yellow zone is sampled from a statistical distribution, such as a mean of 0.1 and a standard deviation of 0.02, generating one point every 0.0001 seconds to obtain a virtual time series.

[0123] The virtual time series model illustrates the state transitions of the distribution box, which refers to the process by which the distribution box changes from one operating state to another in a sandbox environment. Based on the core events of the degradation curve, the virtual time series model automatically updates the health level and feature values ​​of the state characteristics. The agent performs action interactions, i.e., observes the current state of the distribution box and selects the corresponding action. The sandbox environment automatically updates the state of the distribution box according to the degradation curve and provides corresponding rewards. The entire life cycle of the distribution box is simulated repeatedly, with each cycle covering different degradation paths from a brand-new distribution box to a failed distribution box. When all simulations end, if the rate of change of the Q-value (expected long-term reward of the state-action pair) is less than 0.01, the cumulative reward is stable. If 70% of the actions receive positive rewards and the update strategy no longer changes, it indicates that the agent has converged, and the degradation trajectory model is output.

[0124] S4.3 predicts the remaining lifespan of the intelligent distribution box and outputs the remaining lifespan prediction results.

[0125] Specifically, the state transition probability of the intelligent distribution box is calculated by analyzing and statistically analyzing the degradation curves. The curves corresponding to the actions are then selected from the degradation curves. Taking the number of times the intelligent distribution box's health status appears in the yellow zone and the number of transitions from the yellow zone to the red zone as an example, the state transition probability of the intelligent distribution box is:

[0126] ;

[0127] in, This represents the state transition probability of the intelligent distribution box. This indicates the number of times the health status of the smart distribution box transitions from the yellow zone to the red zone. This indicates the number of times the health status of the smart distribution box appears in the yellow zone. The state transition probability is expressed in the form of conditional probability, for example, P(red zone|yellow zone, action = 10Hz detection, 7-day maintenance, probability value 0.8).

[0128] The state feature vector and the health level in the health region are input into the degradation trajectory model. Based on the state transition probability of the smart distribution box, the expected time between the current state and failure of the smart distribution box is calculated, and the remaining lifetime prediction result is obtained.

[0129] Specifically, after the state feature vector and the level of the health zone are input into the degradation trajectory model, the degradation trajectory model matches the input results with the virtual time series to locate similar states, such as a health level of 0.6, a yellow zone, and a temperature rise rate of 0.1 degrees Celsius per minute. After matching, the previously calculated state transition probability of the distribution box is queried. For example, the state transition probability from yellow zone to red zone without maintenance measures is 0.9, with an average transition time of 3 months. The state transition probability from red zone to failure without maintenance measures is 0.95, with an average transition time of 15 days. The expected single-step transition time of the smart distribution box is calculated using the following expression:

[0130] ;

[0131] in, This indicates the expected single-step transfer time of the intelligent distribution box. This represents the state transition probability of the intelligent distribution box. This indicates the average transfer time.

[0132] The expected transfer times of each step in the smart distribution box are summed up to obtain the total expected transfer time. The remaining lifespan prediction result is output, including the maintenance strategy recommended by the degradation trajectory model based on the remaining lifespan of the distribution box. For example, in the yellow zone, maintenance is recommended within 48 hours, and the detection frequency is 10Hz.

[0133] This embodiment also provides a computer device, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the remaining life prediction method for intelligent power distribution boxes on construction sites as proposed in the above embodiment.

[0134] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0135] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the remaining life prediction method for intelligent power distribution boxes at construction sites as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0136] In summary, this invention constructs a directed acyclic graph by analyzing structured datasets, optimizes edge directions and assigns weights through time lag analysis to generate a dynamic causal graph, detects abnormal signals in the distribution box, and traces back to generate fault diagnosis results. This achieves dynamic quantitative expression of causal relationships between multidimensional parameters and precise location of fault sources. Through causal discovery and dynamic causal graphs, it accurately captures the causal mechanisms in the degradation process of the distribution box, greatly improving the depth and relevance of dynamic degradation modeling. Furthermore, by using a deep Q-network to train the degradation trajectory model and calculating the remaining lifetime through state transition probabilities and outputting prediction results, it achieves dynamic simulation of the entire life cycle of the distribution box and accurate prediction of its remaining lifetime. This significantly improves the accuracy and environmental adaptability of the remaining lifetime prediction of intelligent distribution boxes, effectively reduces prediction errors in harsh construction site environments, and extends the lifespan of intelligent distribution boxes.

[0137] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for predicting the remaining life of intelligent distribution boxes on construction sites, characterized in that: include: Data from intelligent distribution boxes is collected and preprocessed to obtain a structured dataset; The causal discovery algorithm is used to analyze structured datasets to identify causal relationships between the states of internal components of intelligent distribution boxes. A directed acyclic graph is constructed, and backtracking and intervention effect analysis are performed to generate fault diagnosis results. The specific steps include: The directed acyclic graph is optimized by using time series data from historical fault data, and weights are assigned to the causal edges of nodes and filtering variable pairs to obtain a dynamic causal graph. Abnormal signals from smart distribution boxes are detected using structured datasets and mapped to nodes in a dynamic cause-effect graph to identify abnormal nodes. The breadth-first search algorithm is used to traverse the abnormal nodes in reverse, obtain the reverse path, and collect the starting nodes of all reverse paths as potential sources of failure. The confidence value of the potential sources of failure is then calculated. Set an intervention effect threshold. When the calculated intervention effect value from the abnormal node to the potential source of failure is greater than the intervention effect threshold, the core node is identified. The core nodes, potential sources of failure, and their confidence values ​​are compiled into JSON format to generate fault diagnosis results. Based on the fault diagnosis results, historical distribution box lifecycle data related to the fault are screened to construct a degradation trajectory database. Feature engineering is used to select state features from the structured dataset and construct a state feature vector. The specific steps include: Based on the core nodes and potential fault sources in the fault diagnosis results, historical distribution box lifecycle data related to the fault are screened and missing values ​​are filled to generate degradation curves; Based on the annotation of core events in the historical power distribution box lifecycle data related to faults, the health of the smart power distribution box is calculated and the smart power distribution box is divided into health regions. The degradation curve and the health regions are combined to form a degradation trajectory library. Calculate the feature importance score for each parameter in the structured dataset, and select the top M parameters with the highest feature importance scores for each parameter as state features; Calculate the feature values ​​of the state features and convert them into z-score values. Arrange the z-score values ​​in a fixed order to form the state feature vector. A sandbox environment is selected as the virtual testing platform. A reinforcement learning environment is configured based on the degenerate trajectory library and state feature vectors. A reinforcement learning model is trained within the sandbox environment to obtain the degenerate trajectory model. The specific steps include: Configure the reinforcement learning environment based on the degenerate trajectory library and state feature vectors; A deep Q-network was selected as the agent. The agent was trained in a sandbox environment according to the reinforcement learning environment. At the same time, the feature values ​​of the state features were mapped using the degradation curves in the degradation trajectory library to generate virtual time series. The eigenvalues ​​of the state features and the health level of the health region are updated by virtual time series. The state transition of the intelligent distribution box between different health regions is repeatedly simulated until the agent reaches convergence, and then the degradation trajectory model is obtained. It calculates the expected time between the current state and failure of the intelligent distribution box, and outputs the remaining life prediction result, specifically including the following steps: Calculate the state transition probability of the intelligent distribution box; The state feature vector and the health level in the health region are input into the degradation trajectory model. The input is matched with the virtual time series. Based on the state transition probability of the smart distribution box, the expected single-step transition time of the smart distribution box is calculated and accumulated to obtain the total expected transition time of the smart distribution box. The remaining lifetime prediction result is then output.

2. The method for predicting the remaining lifespan of intelligent distribution boxes on construction sites as described in claim 1, characterized in that: The data from the intelligent distribution box includes leakage current, circuit breaker coil current, contactor contact temperature, bus voltage, insulation resistance, ambient humidity, and contactor resistance.

3. The method for predicting the remaining lifespan of intelligent distribution boxes on construction sites as described in claim 2, characterized in that: Obtaining the structured dataset specifically includes the following steps: The data of the intelligent distribution box is denoised by using discrete wavelet transform to obtain the denoised intelligent distribution box data. Cubic spline interpolation was used to fill in the denoised smart distribution box data, and the filled smart distribution box data was then standardized to obtain a structured dataset.

4. The method for predicting the remaining lifespan of intelligent distribution boxes on construction sites as described in claim 3, characterized in that: The construction of the directed acyclic graph specifically includes the following steps: Collect historical fault data, and select the parameters with the highest frequency of occurrence of the internal component status of the smart distribution box from the structured dataset and historical fault data as selected variables; Set a standardization threshold; when the calculated Pearson correlation coefficient between the selected variables is less than the standardization threshold, a pair of selected variables is obtained. Set an independence threshold, perform a chi-square test on the selected variable pairs, and determine the causal edge of the selected variable pairs when the calculated chi-square statistic of the selected variable is less than the set independence threshold. By treating the selected variable as a node and connecting it to the causal edges of the filtered variable pairs, a directed acyclic graph is obtained.

5. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the remaining life prediction method for intelligent power distribution boxes on construction sites as described in any one of claims 1 to 4.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the remaining life prediction method for intelligent power distribution boxes on construction sites as described in any one of claims 1 to 4.

Images