Bayesian network-based load action fault early warning method and device, terminal equipment and computer readable storage medium

Abstract

The invention discloses a load action fault early warning method and device based on a Bayesian network, terminal equipment and a computer readable storage medium. The method comprises the following steps: acquiring real-time telemetry data; performing data alignment, missing value filling and balanced sampling on real-time telemetry data to obtain telemetry data to be processed; inputting telemetering data to be processed into the trained Bayesian network model, and calculating the posterior probability of each telemetering point location in the current working state; the posterior probability is compared with a preset threshold value, if the posterior probability is lower than the preset threshold value, abnormal work early warning is triggered, the telemetering point location with the posterior probability lower than the preset threshold value is marked as a fault telemetering point location, and the fault telemetering point location is output. By implementing the method, the telemetering fault can be positioned according to the parameter dependency relationship of the Bayesian network model.

Description

Technical Field

[0001] The present invention relates to the field of satellite operation and maintenance, and in particular to a load action fault early warning method based on a Bayesian network. Background Art

[0002] The task of mining telemetry parameters associated with payload actions is of great significance in the industry. In actual scenarios, for a certain satellite, when it performs a specific task, some telemetry parameters will inevitably show certain regular changes. Specifically, when some tasks occur, parameters such as voltage and temperature will increase significantly, but for different tasks, the telemetry parameters that change are different. If they are mined and labeled manually, only a rough range can be given. At the same time, the actual parameters within the range may not be relevant to the task. Therefore, how to select key telemetry and how to monitor the mission status through the selected telemetry parameters are the current problems.

[0003] When solving such problems, decision trees and association rules are often used to model historical data. Using random forests for fault diagnosis can ensure that reasonable and normal predictions can be given in most scenarios, but there are deficiencies in the interpretability of the results. It is often impossible to know the interdependence between parameters and to locate the fault. In actual production, it is far from enough to know that a fault has occurred but cannot be located. Generally speaking, it is necessary to reversely infer which telemetry has an abnormality based on the occurrence of the fault condition. Therefore, there is an urgent need for a method that can provide correlation between parameters and fault location information. Summary of the invention

[0004] The embodiments of the present invention provide a load action fault early warning method, apparatus, terminal equipment and computer-readable storage medium based on a Bayesian network, which can realize telemetry fault location by utilizing the correlation between parameters of the Bayesian network.

[0005] An embodiment of the present invention provides a load action fault early warning method based on a Bayesian network, comprising:

[0006] Get real-time telemetry data;

[0007] The real-time telemetry data is obtained through data alignment, missing value filling and balanced sampling to obtain the telemetry data to be processed;

[0008] Input the telemetry data to be processed into the trained Bayesian network model to calculate the posterior probability of each telemetry point in the current working state;

[0009] The posterior probability is compared with the preset threshold value. If the posterior probability is lower than the preset threshold value, an abnormal working warning is triggered, and the telemetry point with the posterior probability lower than the preset threshold value is marked as a faulty telemetry point, and the faulty telemetry point is output.

[0010] Further, the training of the Bayesian network model includes:

[0011] Get some historical telemetry data;

[0012] For each historical telemetry data, the historical telemetry data is processed through data alignment, missing value filling and balanced sampling to obtain preprocessed telemetry data;

[0013] The pre-processed telemetry data are binned by a chi-square binning algorithm to obtain a number of telemetry data to be trained;

[0014] The feature importance of each telemetry data to be trained is calculated by random forest; the feature importance is compared with the preset feature screening threshold, and the telemetry data to be trained that meets the preset feature screening threshold is used as the key telemetry data;

[0015] The key telemetry parameters are input into the Bayesian network model to be trained for iterative structural training until the BIC score of the candidate network reaches the preset score threshold, thereby obtaining the model structure of the Bayesian network model;

[0016] The probability distribution function of each node in the model structure of the Bayesian network model is calculated by maximum likelihood estimation;

[0017] A Bayesian network model is constructed based on the model structure and the probability distribution function of each node in the model structure.

[0018] Further, telemetry data includes: telemetry parameters and load actions;

[0019] Perform data alignment by following the steps below:

[0020] Determine the start time and end time of the telemetry parameters according to the timestamp in the telemetry parameters;

[0021] Generate a time series of corresponding telemetry parameters according to the start time and end time of the telemetry parameters;

[0022] Supplement missing timestamps in time series through interpolation algorithms;

[0023] Select the timestamp of telemetry parameters or the timestamp of payload actions as the reference time axis;

[0024] Through the interpolation algorithm, the timestamps of telemetry parameters and load actions are aligned to the reference time axis.

[0025] Furthermore, missing values ​​are filled by following the steps below:

[0026] According to the timestamps of telemetry parameters, the telemetry parameters are divided into several continuous time series data;

[0027] For each first time series data, scan the first time series data from front to back, and when a first missing value is scanned, fill it with the first time series value closest to the first missing value to obtain the second time series data;

[0028] For each second time series data, the second time series data is scanned from back to front, and when a second missing value is scanned, the second time series value closest to the second missing value is used for filling.

[0029] Furthermore, the pre-processed telemetry data are binned by the chi-square binning algorithm to obtain a number of telemetry data to be trained, including:

[0030] For each preprocessed telemetry data, a telemetry parameter is selected as the initial binning variable and divided into several binning intervals;

[0031] Calculate the chi-square value for each binning interval;

[0032] According to the chi-square value of each bin interval, the adjacent intervals with the smallest chi-square value are merged;

[0033] Determine whether the preset minimum number of samples is greater than the number of bin intervals after merging. If so, output the telemetry parameters after merging the bin intervals. If not, continue to calculate the chi-square value of the bin intervals for merging.

[0034] The telemetry parameters and historical load actions after merging all bin intervals are used as the telemetry data to be trained.

[0035] Furthermore, the feature importance of each telemetry data to be trained is calculated by random forest, including:

[0036] Randomly select a number of telemetry data from the telemetry data to be trained to generate a number of decision trees;

[0037] For each telemetry data in each decision tree, calculate the reduction of the Gini coefficient of each telemetry data before and after the node splitting in the decision tree, and obtain the impurity reduction value of each telemetry data in all nodes;

[0038] The importance of each telemetry data in all decision trees is calculated by the impurity reduction value of each telemetry data in all nodes.

[0039] Furthermore, the key telemetry parameters are input into the Bayesian network model to be trained for iterative structural training until the BIC score of the candidate network reaches a preset score threshold, and the model structure of the Bayesian network model is obtained, including:

[0040] Inputting key telemetry parameters into the Bayesian network model to be trained;

[0041] The Bayesian network model randomly generates an initial Bayesian network as the current network based on the key telemetry data, and calculates the BIC score of the current network;

[0042] Generate candidate networks based on the current network;

[0043] Calculate the BIC score of each candidate network and select the candidate network with the highest BIC score as the best network;

[0044] The BIC score of the best network is compared with the preset score threshold. If the BIC score of the best network does not reach the preset score threshold, the current best network is replaced by the current network, and the current network is iteratively trained again until the BIC score of the best network reaches the preset score threshold; if the BIC score of the best network reaches the preset score threshold, the current best network is used as the model structure of the Bayesian network model.

[0045] Based on the above method embodiment, the present invention provides a corresponding device embodiment, including: a telemetry data acquisition module, a telemetry data preprocessing module, a posterior probability calculation module, and a threshold comparison module;

[0046] Telemetry data acquisition module, used to obtain real-time telemetry data;

[0047] The telemetry data preprocessing module is used to obtain the telemetry data to be processed through data alignment, missing value filling and balanced sampling of real-time telemetry data;

[0048] The posterior probability calculation module is used to input the telemetry data to be processed into the trained Bayesian network model to calculate the posterior probability of each telemetry point under the current working state;

[0049] The threshold comparison module is used to compare the posterior probability with the preset threshold. If the posterior probability is lower than the preset threshold, an abnormal operation warning is triggered, and the telemetry point with the posterior probability lower than the preset threshold is marked as a faulty telemetry point, and the faulty telemetry point is output.

[0050] Based on the above-mentioned method item embodiment, the present invention provides a corresponding terminal device item embodiment, including: a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, the steps of the load action fault warning method based on the Bayesian network as described in the present invention are implemented.

[0051] Based on the above-mentioned method embodiment, the present invention provides a corresponding computer-readable storage medium embodiment, including: a stored computer program, which controls the device where the computer-readable storage medium is located to execute the steps of the load action fault warning method based on the Bayesian network as described in the present invention when the computer program is running.

[0052] Compared with the prior art, the beneficial effects of the embodiment of this solution are:

[0053] The present invention acquires real-time telemetry data for real-time monitoring, processes the real-time telemetry data and inputs it into a trained Bayesian network model, calculates the posterior probability of each telemetry point under the current working state, compares the posterior probability with a preset threshold value, and triggers an abnormal working warning when the posterior probability is lower than the preset threshold value, and marks the telemetry point with the posterior probability lower than the preset threshold value as a faulty telemetry point, and outputs the faulty telemetry point. Because the Bayesian network model itself has a dependency relationship between parameters, it can use historical data and prior knowledge to calculate the posterior probability, thereby providing an accurate prediction of the current working state, and when the abnormal working warning is triggered, the parameter dependency relationship of the Bayesian network model can be used to locate the fault, and the potential fault can be discovered and resolved in time.

[0054] To sum up, the present invention calculates the posterior probability of each telemetry point in the current working state through the Bayesian network model, and compares it with the preset threshold value to determine the working state of the telemetry point and locate the fault point, so as to reversely infer which telemetry has an abnormality based on the occurrence of the fault condition. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is a flow chart of a load action fault early warning method based on a Bayesian network provided by an embodiment of the present invention;

[0056] Figure 2 It is a flowchart of the training process of the Bayesian network model provided by one embodiment of the present invention;

[0057] Figure 3 It is a flowchart of a telemetry data preprocessing process provided by an embodiment of the present invention;

[0058] Figure 4It is a flowchart of a chi-square algorithm binning process provided by an embodiment of the present invention;

[0059] Figure 5 is a schematic diagram of random forest feature screening provided by an embodiment of the present invention;

[0060] Figure 6 is a schematic diagram of the structure of a trained Bayesian network model provided by an embodiment of the present invention;

[0061] Figure 7 It is a structural schematic diagram of a load action fault early warning device based on a Bayesian network provided by an embodiment of the present invention;

[0062] Figure 8 It is a structural schematic diagram of another load action fault warning device based on a Bayesian network provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0063] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. 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 creative work are within the scope of protection of the present invention.

[0064] In the description of the present invention, it should be understood that the terms "first" and "second" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features.

[0065] like Figure 1 As shown, an embodiment of the present invention provides a load action fault early warning method based on a Bayesian network, the method comprising at least the following steps:

[0066] Step S1: Obtain real-time telemetry data;

[0067] For step S1, the real-time telemetry data obtained includes two parts: real-time telemetry parameters and real-time load actions; wherein the telemetry parameters may include: temperature, pressure, current, vibration and other parameters, which are used to describe the state of the system or environment, and the load action can be understood as a transaction, which includes the action type and the interval value of each telemetry parameter. In this embodiment, these real-time telemetry parameters and corresponding load actions are collected through temperature sensors, pressure sensors, vibration sensors and other corresponding sensors.

[0068] Step S2: The real-time telemetry data is subjected to data alignment, missing value filling and balanced sampling to obtain the telemetry data to be processed;

[0069] In a preferred embodiment, data alignment is performed by the following steps:

[0070] Determine the start time and end time of the telemetry parameters according to the timestamp in the telemetry parameters;

[0071] Generate a time series of corresponding telemetry parameters according to the start time and end time of the telemetry parameters;

[0072] Supplementing missing timestamps in the time series by an interpolation algorithm;

[0073] Select the timestamp of telemetry parameters or the timestamp of payload actions as the reference time axis;

[0074] Through an interpolation algorithm, the timestamp of the telemetry parameter and the timestamp of the load action are aligned to a reference time axis.

[0075] In a preferred embodiment, missing value filling is performed by the following steps:

[0076] According to the timestamps of telemetry parameters, the telemetry parameters are divided into several continuous time series data;

[0077] For each first time series data, scan the first time series data from front to back, and when a first missing value is scanned, fill it with the first time series value closest to the first missing value to obtain second time series data;

[0078] For each second time series data, the second time series data is scanned from back to front, and when a second missing value is scanned, the second time series value closest to the second missing value is used for filling.

[0079] For step S2, the real-time telemetry data collected by step S1 has a lot of missing data due to various limitations of data transmission and data storage. When the data is spliced ​​by seconds, we will find that there are missing parameters in the data, that is, the data of some parameters is missing in a specific time period. This situation will lead to parameter alignment problems. In order to solve these problems, the collected real-time telemetry data needs to be preprocessed, such as Figure 3 As shown, through data alignment, missing value filling and balanced sampling, the processed telemetry data is complete and consistent.

[0080] First, the data alignment process is explained: Data alignment is to ensure that telemetry parameters and load actions have consistent intervals and alignment in the time dimension, including two steps: timestamp completion and timestamp alignment. The following is a description of the two steps. Timestamp completion first finds the earliest and latest timestamps based on the timestamps in the telemetry parameters to determine the start time and end time of the telemetry parameters, and then creates a continuous time series between the start time and the end time based on the start time and the end time of the telemetry parameters. Finally, the interpolation algorithm is used to supplement the missing timestamps in the time series based on the existing timestamps; timestamp alignment is to ensure that the telemetry parameters and load actions are aligned on the same time axis. It is necessary to select the timestamp of the telemetry parameters or the timestamp of the load action as the reference time axis, and then use the interpolation algorithm to find the nearest existing timestamp for interpolation, and align the timestamps of the telemetry parameters and the timestamps of the load action to the reference time axis.

[0081] Secondly, the missing value filling process is explained: Missing value filling is to ensure the continuity of data, avoid breaks or intervals caused by missing values, and reduce errors in data processing and analysis. It includes three steps: time series grouping, forward filling, and backward filling. The following is an explanation of these three steps. First, the timestamps of telemetry parameters are grouped in time series, that is, the telemetry parameters are divided into several continuous time series data; for each first time series data, the first time series data is scanned from front to back. When the first missing value is scanned, that is, the time point at which data is missing, the nearest first time series value before the first missing value is used to fill it, that is, the missing timestamp is supplemented with the nearest existing timestamp. This process is called forward filling; the time series that has been forward filled is scanned from back to front. When a missing value is scanned, that is, the time point at which data is missing, the nearest time series value after the missing value is used to fill it, that is, the missing timestamp is supplemented with the nearest existing timestamp. This process is called backward filling.

[0082] It should be noted that after forward filling, there may still be some missing values, that is, missing values ​​that are not filled in the forward filling process. These missing values ​​may be located at the end of the time series. Therefore, in order to ensure the continuity and integrity of the data in the time dimension, backward filling is required. Through the combination of forward filling and backward filling, the missing values ​​in the data can be supplemented to the maximum extent, so that the data has consistent intervals and alignment in the time series.

[0083] Finally, the balanced sampling process is explained: in actual production tasks, since the data volume of the task segment is much smaller than the data volume of the non-task segment, the positive and negative samples are unbalanced. In order to solve this problem, the present invention performs balanced sampling and inputs the positive and negative samples into the Bayesian network model at a ratio of 1:1 to solve the problem of category imbalance in the data set and ensure that the number of samples in each category is relatively balanced when training the model.

[0084] Step S3: input the telemetry data to be processed into the trained Bayesian network model to calculate the posterior probability of each telemetry point under the current working state;

[0085] For step S3:

[0086] First, the training process of the Bayesian network model is described in detail:

[0087] like Figure 2 As shown, an embodiment of the present invention provides a training process of a Bayesian network model, and the training process includes at least the following steps:

[0088] Step S301: Obtain some historical telemetry data;

[0089] For step S301, several groups of historical telemetry data are obtained from the database, and the Bayesian network model is trained using the historical data, wherein the historical telemetry data includes telemetry parameters, load actions, and corresponding working states.

[0090] Step S302: for each historical telemetry data, the historical telemetry data is subjected to data alignment, missing value filling and balanced sampling to obtain pre-processed telemetry data;

[0091] For step S302, the historical telemetry data obtained in the database may have inconsistent timestamps, missing values, and unbalanced numbers of samples in different working states. Therefore, it is necessary to perform data preprocessing on each set of historical telemetry data. First, the timestamps of each telemetry parameter are aligned to ensure that each telemetry parameter has the same timestamp. In the aligned historical telemetry data, there may still be missing values. For each telemetry parameter, the missing values ​​are filled with existing data through forward filling and backward filling methods to maintain the continuity of the time series. At the same time, in order to maintain the balance of the samples, a balanced sampling method is adopted to randomly select the same number of samples from each working state to avoid the model's preference for certain working states during the training process.

[0092] Preprocessing telemetry data after data alignment, missing value filling and balanced sampling improves data consistency, completeness and balance, providing a more reliable data foundation for subsequent modeling and analysis.

[0093] Step S303: binning each pre-processed telemetry data using a chi-square binning algorithm to obtain a number of telemetry data to be trained;

[0094] In a preferred embodiment, each pre-processed telemetry data is binned by a chi-square binning algorithm to obtain a number of telemetry data to be trained, including:

[0095] For each preprocessed telemetry data, a telemetry parameter is selected as the initial binning variable and divided into several binning intervals;

[0096] Calculate the chi-square value for each binning interval;

[0097] According to the chi-square value of each bin interval, the adjacent intervals with the smallest chi-square value are merged;

[0098] Determine whether the preset minimum number of samples is greater than the number of bin intervals after merging. If so, output the telemetry parameters after merging the bin intervals. If not, continue to calculate the chi-square value of the bin intervals for merging.

[0099] The telemetry parameters and historical load actions after merging all bin intervals are used as the telemetry data to be trained.

[0100] For step S301, the pre-processed telemetry data is binned by the chi-square binning algorithm, and the telemetry parameters are divided into several binning intervals, thereby converting continuous data into discrete data. The following is a detailed description of the binning process:

[0101] like Figure 4 As shown, first, for each pre-processed telemetry data obtained in step S302, a telemetry parameter is selected as the initial binning variable, and it is divided into several binning intervals; for each binning interval, the chi-square value between the binning interval and the historical load action is calculated. The chi-square value can measure the correlation and significance between the binning interval and the load action, so the best binning interval can be selected by calculating the chi-square value; then, according to the chi-square value of each binning interval, the adjacent intervals with the smallest chi-square value are merged, and the purpose of the merger is to reduce the number of binning intervals; then, a minimum number of samples is preset to determine the binning interval after the merger. Whether the number of binning intervals after merging is less than or equal to the preset minimum number of samples. If the number of binning intervals after merging is less than or equal to the minimum number of samples, it means that the number of binning intervals after merging has been reduced to the requirement of the minimum number of samples, and the telemetry parameters after merging the binning intervals are output as the telemetry data to be trained. If the number of binning intervals after merging is greater than the minimum number of samples, it means that the number of binning intervals after merging has not yet reached the requirement of the minimum number of samples, and the chi-square value of the binning intervals is continued to be calculated for merging until the requirement of the minimum number of samples is met; finally, the telemetry parameters after merging all binning intervals, historical load actions and corresponding working states are taken together as the telemetry data to be trained.

[0102] Step S304: Calculate the feature importance of each telemetry data to be trained by random forest; compare the feature importance with the preset feature screening threshold, and use the telemetry data to be trained that meets the preset feature screening threshold as key telemetry data;

[0103] In a preferred embodiment, the feature importance of each telemetry data to be trained is calculated by random forest, including:

[0104] Randomly select a number of telemetry data from the telemetry data to be trained to generate a number of decision trees;

[0105] For each telemetry data in each decision tree, calculate the reduction of the Gini coefficient of each telemetry data before and after the node splitting in the decision tree, and obtain the impurity reduction value of each telemetry data in all nodes;

[0106] The importance of each telemetry data in all decision trees is calculated by the impurity reduction value of each telemetry data in all nodes.

[0107] For step S304, by calculating the feature importance through random forest, it is possible to screen out which telemetry parameters are more important for the prediction results of the model. The specific steps include: generating a decision tree, calculating the impurity reduction value, calculating the feature importance and feature screening.

[0108] like Figure 5 As shown, first, from the telemetry data to be trained obtained in step S303, several telemetry parameters are randomly selected to generate several decision trees, each decision tree uses different training samples and is generated by randomly selecting telemetry parameters; then, in the random forest, the Gini coefficient of each node is calculated by the following formula: Where D represents the set of nodes, K represents the number of categories, and p k represents the proportion of samples belonging to the kth class in the node; then, for each telemetry parameter in each decision tree, the reduction of the Gini coefficient before and after the node split is calculated by the following formula: ΔGini=Gini(D parent )-Gini(D split ), the impurity of the samples in the node can be measured by calculating the Gini coefficient. The greater the reduction of the Gini coefficient, the greater the contribution of the telemetry parameter to the splitting of the node. Next, for each telemetry parameter of each decision tree, the reduction of its impurity in all nodes is accumulated to obtain the importance of the telemetry parameter in the decision tree: in, represents the importance of telemetry parameter j in decision tree t, N trepresents the node set of telemetry parameter j in decision tree t, 1(feature(n)=j) represents the indicator function, which takes the value of 1 when node n is split using telemetry parameter j, otherwise it takes the value of 0; finally, the importance of telemetry parameter j in the entire random forest FI j is the result of averaging the importance of telemetry parameter j in all decision trees: Where T represents the number of decision trees generated in the random forest.

[0109] After calculating the feature importance of each telemetry parameter, an appropriate feature screening threshold is set, and then the feature importance of each telemetry parameter is compared with the preset feature screening threshold. If the feature importance of a telemetry parameter is greater than or equal to the feature screening threshold, the telemetry parameter is regarded as an important telemetry data. Through this process, telemetry data that meets the threshold requirements can be selected to generate a key telemetry data set. In this embodiment, six key telemetry data are screened from more than 100 telemetry parameters through the feature importance of random forest. These six key telemetry data are of high importance to the prediction results of the model, which can reduce the feature dimension and calculation complexity, while improving the efficiency and accuracy of the model.

[0110] Step S305: inputting the key telemetry parameters into the Bayesian network model to be trained to perform iterative training of the structure until the BIC score of the candidate network reaches a preset score threshold, thereby obtaining the model structure of the Bayesian network model;

[0111] In a preferred embodiment, the key telemetry parameters are input into the Bayesian network model to be trained for iterative structural training until the BIC score of the candidate network reaches a preset score threshold, and the model structure of the Bayesian network model is obtained, including:

[0112] Inputting key telemetry parameters into the Bayesian network model to be trained;

[0113] The Bayesian network model randomly generates an initial Bayesian network as the current network based on the key telemetry data, and calculates the BIC score of the current network;

[0114] Generate candidate networks based on the current network;

[0115] Calculate the BIC score of each candidate network and select the candidate network with the highest BIC score as the best network;

[0116] The BIC score of the best network is compared with the preset score threshold. If the BIC score of the best network does not reach the preset score threshold, the current best network is replaced by the current network, and the current network is iteratively trained again until the BIC score of the best network reaches the preset score threshold; if the BIC score of the best network reaches the preset score threshold, the current best network is used as the model structure of the Bayesian network model.

[0117] It should be noted that the Bayesian network is a probabilistic graphical model used to describe the dependencies between variables. It can infer the probability distribution of other variables through the observed variables, and thus be used for prediction and decision-making. A complete Bayesian network needs to be completed through structural training and parameter training.

[0118] As for step S305, this step is the structural training of the Bayesian network model, and its goal is to train multiple nodes into a directed acyclic graph with directionality. In this embodiment, the hill climbing search algorithm in the heuristic search algorithm is used to find the local optimal solution. The hill climbing search algorithm is a search algorithm based on a greedy strategy, which is used to find the local optimal solution in the search space. Specifically, in the structural training, it starts from an initial structure, changes the structure of the network by constantly trying to add, delete or modify edges, and then evaluates the Bayesian Information Criterion (BIC) score of each modified structure. If the BIC score of the modified structure is better than the current structure, it will become the new current structure and continue to the next step of modification. This process continues until no better structure can be found. The detailed steps are described as follows:

[0119] First, the six key telemetry data obtained in step S304 are input into the Bayesian network model to be trained, and the telemetry parameters in the key telemetry data are used as input nodes, the load actions in the key telemetry data are used as conditional nodes, and the working states corresponding to the telemetry parameters in the key telemetry data are used as output nodes to construct an initial Bayesian network model. The initial Bayesian network model will randomly generate an initial Bayesian network as the current network according to the key telemetry data; then, the Bayesian Information Criterion (BIC) score of the current network is calculated as an evaluation index of the current network. The BIC score is a criterion for model selection, which is used to measure the fitting ability and complexity of the model; then, Add or delete edge connections or change the dependencies between nodes to generate candidate networks. Then, calculate the BIC score of each candidate network and select the candidate network with the highest BIC score as the current best network. Compare the BIC score of the best network with the preset score threshold. If the BIC score of the best network does not reach the preset score threshold, it means that the complexity of the best network is high or the size of the data set is small. Then replace the current best network with the current network and continue iterative training until the BIC score of the best network reaches the preset score threshold. At this time, the current best network is used as the final model structure of the Bayesian network model.

[0120] Step S306: Calculate the probability distribution function of each node in the model structure of the Bayesian network model by maximum likelihood estimation;

[0121] For step S306, this step is the parameter training of the Bayesian network model. In the Bayesian network, each node has a conditional probability distribution function, which describes the probability distribution of the node given its parent node value. The present invention calculates the probability distribution function of each node of the Bayesian network model by maximum likelihood estimation (Maximum Likelihood Estimation, MLE). Specifically, for each node, its sample frequency under the condition of a given parent node value is calculated, and then these sample frequencies are used to estimate the conditional probability distribution function of the node. For discrete variables, the probability is calculated using frequency. For example, for an unobserved state of a node, the frequency of each possible state is calculated under the condition of a given parent node value, and it is used as the probability of the state; for continuous variables, statistical methods are used to approximate its probability distribution function, for example, Gaussian distribution is used to fit the data, and the sample mean and variance are used as estimates of the probability distribution.

[0122] Step S307: construct a Bayesian network model according to the model structure and the probability distribution function of each node in the model structure.

[0123] For step S307, after step S305 and step S306, the model structure of the Bayesian network and the probability distribution function of each node in the model structure are obtained. Combining them can obtain a probability distribution function such as Figure 6 The complete Bayesian network model shown in Figure 1 is a graph of a telemetry parameter. In this model, each node represents a telemetry parameter, and the connections between them represent their dependencies. Using this Bayesian network model, fault detection and fault location can be performed.

[0124] In one embodiment of the present invention, the telemetry data after data preprocessing in step S2 is input into the trained Bayesian network model. For each telemetry point, the Bayesian network model will make inferences based on the prior probability and conditional probability of the current working state, and calculate the posterior probability of each telemetry point under the current working state.

[0125] Step S4: Compare the posterior probability with the preset threshold value. If the posterior probability is lower than the preset threshold value, trigger an abnormal operation warning, mark the telemetry point with the posterior probability lower than the preset threshold value as a faulty telemetry point, and output the faulty telemetry point.

[0126] For step S4, the posterior probability obtained in step S3 is compared with the preset threshold value. Assume that the threshold value is set to 0.3, that is, the node with a posterior probability lower than 0.3 will be considered as a sign of failure. For each node, check whether its posterior probability is lower than 0.3. If the posterior probability of a node is lower than 0.3, it means that the telemetry parameter may be faulty, triggering an abnormal working warning, and marking these telemetry points as faulty telemetry points. Through the marked faulty telemetry points, we can further locate the fault. According to the structure and dependency of the Bayesian network model, the posterior probabilities of other nodes can be analyzed to determine the possible causes and locations of the fault. For example, if the posterior probability of the TMA0071 node is lower than 0.3, it means that the abnormal working may be related to the telemetry parameters of the TMA0071 node.

[0127] It should be noted that fault location based solely on the posterior probability of a single node may be uncertain because the nodes in the Bayesian network model are interdependent. Therefore, when locating the fault, the posterior probabilities of other nodes directly or indirectly connected to the TMA0071 node can be further analyzed to obtain a more comprehensive fault location result. For example, if the TMA0071 node is connected to the TMA0007 node, and the posterior probability of the TMA0007 node is also low, it may indicate that there is a correlation between the TMA0071 node and the TMA0007 node, and the combination of the two may cause the fault. By analyzing the posterior probabilities of multiple related nodes, we can more accurately locate the cause and location of the fault.

[0128] like Figure 7 As shown, based on the above method embodiment, a corresponding device embodiment is provided;

[0129] An embodiment of the present invention provides a load action fault early warning device based on a Bayesian network, comprising: a telemetry data acquisition module, a telemetry data preprocessing module, a posterior probability calculation module, and a threshold comparison module;

[0130] Telemetry data acquisition module, used to obtain real-time telemetry data;

[0131] The telemetry data preprocessing module is used to obtain the telemetry data to be processed through data alignment, missing value filling and balanced sampling of real-time telemetry data;

[0132] The posterior probability calculation module is used to input the telemetry data to be processed into the trained Bayesian network model to calculate the posterior probability of each telemetry point under the current working state;

[0133] The threshold comparison module is used to compare the posterior probability with the preset threshold. If the posterior probability is lower than the preset threshold, an abnormal operation warning is triggered, and the telemetry point with the posterior probability lower than the preset threshold is marked as a faulty telemetry point, and the faulty telemetry point is output.

[0134] like Figure 8 As shown, an embodiment of the present invention provides another load action fault early warning device based on Bayesian network;

[0135] In a preferred embodiment, the load action fault early warning device based on the Bayesian network further includes: a Bayesian network model training module;

[0136] The Bayesian network model training module includes: a historical data acquisition submodule, a historical data preprocessing submodule, a chi-square binning module, a feature screening submodule, a structure training submodule, a parameter training submodule, and a model building submodule;

[0137] The historical data acquisition submodule is used to obtain some historical telemetry data;

[0138] The historical data preprocessing submodule is used to obtain preprocessed telemetry data by aligning the historical telemetry data, filling missing values ​​and balancing sampling for each historical telemetry data;

[0139] The chi-square binning module is used to bin each pre-processed telemetry data using the chi-square binning algorithm to obtain a number of telemetry data to be trained;

[0140] The feature screening submodule is used to calculate the feature importance of each telemetry data to be trained through random forest; compare the feature importance with the preset feature screening threshold, and use the telemetry data to be trained that meets the preset feature screening threshold as the key telemetry data;

[0141] The structure training submodule is used to input the key telemetry parameters into the Bayesian network model to be trained to perform iterative training of the structure until the BIC score of the candidate network reaches a preset score threshold, thereby obtaining the model structure of the Bayesian network model;

[0142] A parameter training submodule, used for calculating the probability distribution function of each node in the model structure of the Bayesian network model by maximum likelihood estimation;

[0143] The model building submodule is used to build a Bayesian network model based on the model structure and the probability distribution function of each node in the model structure.

[0144] In a preferred embodiment, the telemetry data preprocessing module includes: a data alignment submodule and a missing value filling submodule;

[0145] The data alignment submodule is used to align data through the following steps:

[0146] Determine the start time and end time of the telemetry parameters according to the timestamp in the telemetry parameters;

[0147] Generate a time series of corresponding telemetry parameters according to the start time and end time of the telemetry parameters;

[0148] Supplementing missing timestamps in the time series by an interpolation algorithm;

[0149] Select the timestamp of telemetry parameters or the timestamp of payload actions as the reference time axis;

[0150] Through an interpolation algorithm, the timestamp of the telemetry parameter and the timestamp of the load action are aligned to a reference time axis.

[0151] The missing value filling submodule is used to fill missing values ​​through the following steps:

[0152] According to the timestamps of telemetry parameters, the telemetry parameters are divided into several continuous time series data;

[0153] For each first time series data, scan the first time series data from front to back, and when a first missing value is scanned, fill it with the first time series value closest to the first missing value to obtain second time series data;

[0154] For each second time series data, the second time series data is scanned from back to front, and when a second missing value is scanned, the second time series value closest to the second missing value is used for filling.

[0155] In a preferred embodiment, the chi-square binning module includes: a binning unit, a chi-square value calculation unit, an interval merging unit, a chi-square value judgment unit, and a data merging unit;

[0156] A binning unit, for selecting a telemetry parameter as an initial binning variable for each preprocessed telemetry data, and dividing it into a number of binning intervals;

[0157] A chi-square value calculation unit, used to calculate the chi-square value of each binning interval;

[0158] An interval merging unit, used for merging adjacent intervals with the smallest chi-square value according to the chi-square value of each binning interval;

[0159] The chi-square value judgment unit is used to judge whether the preset minimum number of samples is greater than the number of bin intervals after merging. If so, the telemetry parameters after merging the bin intervals are output; if not, the chi-square value of the bin intervals is continued to be calculated for merging;

[0160] The data merging unit is used to merge the telemetry parameters and historical load actions of all bin intervals as the telemetry data to be trained.

[0161] In a preferred embodiment, the feature screening submodule includes: a decision tree generation unit, a Gini coefficient calculation unit, and a feature importance calculation unit;

[0162] A decision tree generation unit, used for randomly selecting a number of telemetry data from the telemetry data to be trained to generate a number of decision trees;

[0163] A Gini coefficient calculation unit, for calculating the reduction of the Gini coefficient of each telemetry data before and after the node splitting in the decision tree for each telemetry data in each decision tree, and obtaining the impurity reduction value of each telemetry data in all nodes;

[0164] The feature importance calculation unit is used to calculate the importance of each telemetry data in all decision trees by the impurity reduction value of each telemetry data in all nodes.

[0165] In a preferred embodiment, the structure training submodule includes: a data input unit, an initial network construction unit, a candidate network generation unit, an optimal network selection unit, and an iterative training unit;

[0166] A data input unit, used for inputting key telemetry parameters into the Bayesian network model to be trained;

[0167] The initial network construction unit is used for the Bayesian network model to randomly generate an initial Bayesian network as the current network according to the key telemetry data, and calculate the BIC score of the current network;

[0168] A candidate network generation unit, used to generate various candidate networks according to the current network;

[0169] The best network selection unit is used to calculate the BIC score of each candidate network and select the candidate network with the highest BIC score as the best network;

[0170] The iterative training unit is used to compare the BIC score of the best network with a preset score threshold. If the BIC score of the best network does not reach the preset score threshold, the current best network is replaced by the current network, and the current network is iteratively trained again until the BIC score of the best network reaches the preset score threshold; if the BIC score of the best network reaches the preset score threshold, the current best network is used as the model structure of the Bayesian network model.

[0171] It can be understood that the above-mentioned device item embodiments correspond to the method item embodiments of the present invention, and can implement the Bayesian network-based load action fault warning method provided by any of the above-mentioned method item embodiments of the present invention.

[0172] It should be noted that the device embodiments described above are merely illustrative, and some or all of the modules may be selected according to actual needs to achieve the purpose of the present embodiment. In addition, in the accompanying drawings of the device embodiments provided by the present invention, the connection relationship between the modules indicates that there is a communication connection between them, which may be specifically implemented as one or more communication buses or signal lines. Those of ordinary skill in the art may understand and implement the present invention without creative work.

[0173] Based on the above-mentioned embodiment of the load action fault warning method based on Bayesian network, another embodiment of the present invention provides a terminal device, which includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, the load action fault warning method based on Bayesian network of any embodiment of the present invention is implemented.

[0174] Exemplarily, in this embodiment, the computer program may be divided into one or more modules, which are stored in the memory and executed by the processor to implement the present invention. The one or more module elements may be a series of computer program instruction segments capable of implementing specific functions, which are used to describe the execution process of the computer program in the terminal device.

[0175] The terminal device may be a computing device such as a desktop computer, a notebook, a PDA, a cloud server, etc. The terminal device may include, but is not limited to, a processor and a memory.

[0176] The processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc. The processor is the control center of the terminal device, and uses various interfaces and lines to connect various parts of the entire terminal device.

[0177] Based on the above method embodiment, another embodiment is provided: another embodiment of the present invention provides a computer-readable storage medium, including a stored computer program, wherein when the computer program is running, the device where the computer-readable storage medium is located is controlled to execute the load action fault warning method based on the Bayesian network described in any one of the above method embodiments of the present invention.

[0178] Wherein, the module / unit integrated with the load action fault warning device / terminal device based on the Bayesian network, if implemented in the form of a software functional unit and sold or used as an independent product, can be stored in a computer-readable storage medium. Based on such an understanding, the present invention implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium, and the computer program can implement the steps of the above-mentioned various method embodiments when executed by the processor. Wherein, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form, etc. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electric carrier signal, telecommunication signal and software distribution medium, etc.

[0179] The above is a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several improvements and modifications without departing from the principle of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A load action fault early warning method based on Bayesian network, characterized in that: include: Get real-time telemetry data; The real-time telemetry data is obtained through data alignment, missing value filling and balanced sampling to obtain the telemetry data to be processed; Input the telemetry data to be processed into the trained Bayesian network model to calculate the posterior probability of each telemetry point in the current working state; The posterior probability is compared with the preset threshold value. If the posterior probability is lower than the preset threshold value, an abnormal operation warning is triggered, and the telemetry point with the posterior probability lower than the preset threshold value is marked as a faulty telemetry point, and the faulty telemetry point is output.

2. The load action failure early warning method based on Bayesian network according to claim 1 is characterized in that: The training of the Bayesian network model includes: Get some historical telemetry data; For each historical telemetry data, the historical telemetry data is processed through data alignment, missing value filling and balanced sampling to obtain preprocessed telemetry data; The pre-processed telemetry data are binned by a chi-square binning algorithm to obtain a number of telemetry data to be trained; Calculate the feature importance of each telemetry data to be trained by random forest; compare the feature importance with a preset feature screening threshold, and use the telemetry data to be trained that meets the preset feature screening threshold as key telemetry data; Inputting the key telemetry parameters into the Bayesian network model to be trained to perform iterative training of the structure until the BIC score of the candidate network reaches a preset score threshold, thereby obtaining a model structure of the Bayesian network model; Calculating the probability distribution function of each node in the model structure of the Bayesian network model by maximum likelihood estimation; The Bayesian network model is constructed according to the model structure and the probability distribution function of each node in the model structure.

3. The load action failure early warning method based on Bayesian network according to claim 2 is characterized in that: Telemetry data, including: telemetry parameters and load actions; Perform data alignment by following the steps below: Determine the start time and end time of the telemetry parameters according to the timestamp in the telemetry parameters; Generate a time series of corresponding telemetry parameters according to the start time and end time of the telemetry parameters; Supplementing missing timestamps in the time series by an interpolation algorithm; Select the timestamp of telemetry parameters or the timestamp of payload actions as the reference time axis; Through an interpolation algorithm, the timestamp of the telemetry parameter and the timestamp of the load action are aligned to a reference time axis.

4. The load action failure early warning method based on Bayesian network according to claim 3 is characterized in that: The missing values ​​are filled by following the steps below: According to the timestamps of telemetry parameters, the telemetry parameters are divided into several continuous time series data; For each first time series data, scan the first time series data from front to back, and when a first missing value is scanned, fill it with the first time series value closest to the first missing value to obtain second time series data; For each second time series data, the second time series data is scanned from back to front, and when a second missing value is scanned, the second time series value closest to the second missing value is used for filling.

5. The load action failure early warning method based on Bayesian network according to claim 2 is characterized in that: The pre-processed telemetry data are binned by the chi-square binning algorithm to obtain a number of telemetry data to be trained, including: For each preprocessed telemetry data, a telemetry parameter is selected as the initial binning variable and divided into several binning intervals; Calculate the chi-square value for each binning interval; According to the chi-square value of each binning interval, adjacent intervals with the smallest chi-square value are merged; Determine whether the preset minimum number of samples is greater than the number of bin intervals after merging. If so, output the telemetry parameters after merging the bin intervals. If not, continue to calculate the chi-square value of the bin intervals for merging. The telemetry parameters and historical load actions after merging all bin intervals are used as the telemetry data to be trained.

6. The load action failure early warning method based on Bayesian network according to claim 2 is characterized in that: The importance of each telemetry feature to be trained is calculated by random forest, including: Randomly select a number of telemetry data from the telemetry data to be trained to generate a number of decision trees; For each telemetry data in each decision tree, calculate the reduction of the Gini coefficient of each telemetry data before and after the node splitting in the decision tree, and obtain the impurity reduction value of each telemetry data in all nodes; The importance of each telemetry data in all decision trees is calculated by the impurity reduction value of each telemetry data in all nodes.

7. The load action failure early warning method based on Bayesian network according to claim 2 is characterized in that: The key telemetry parameters are input into the Bayesian network model to be trained for iterative structural training until the BIC score of the candidate network reaches a preset score threshold, and the model structure of the Bayesian network model is obtained, including: Inputting the key telemetry parameters into the Bayesian network model to be trained; The Bayesian network model randomly generates an initial Bayesian network as the current network according to the key telemetry data, and calculates the BIC score of the current network; Generate candidate networks based on the current network; Calculate the BIC score of each candidate network and select the candidate network with the highest BIC score as the best network; The BIC score of the best network is compared with the preset score threshold. If the BIC score of the best network does not reach the preset score threshold, the current best network is replaced by the current network, and the current network is iteratively trained again until the BIC score of the best network reaches the preset score threshold; if the BIC score of the best network reaches the preset score threshold, the current best network is used as the model structure of the Bayesian network model.

8. A load action fault early warning device based on Bayesian network, characterized in that: include: Telemetry data acquisition module, telemetry data preprocessing module, posterior probability calculation module, and threshold comparison module; The telemetry data acquisition module is used to acquire real-time telemetry data; The telemetry data preprocessing module is used to obtain the telemetry data to be processed by aligning the real-time telemetry data, filling missing values ​​and balancing sampling; The posterior probability calculation module is used to input the telemetry data to be processed into the trained Bayesian network model to calculate the posterior probability of each telemetry point under the current working state; The threshold comparison module is used to compare the posterior probability with the preset threshold. If the posterior probability is lower than the preset threshold, an abnormal operation warning is triggered, and the telemetry point with the posterior probability lower than the preset threshold is marked as a faulty telemetry point, and the faulty telemetry point is output.

9. A terminal device, characterized in that: include: A processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein when the processor executes the computer program, the load action fault warning method based on the Bayesian network as described in any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium, characterized in that: include: A stored computer program, wherein when the computer program is running, the device where the computer-readable storage medium is located is controlled to execute the load action fault early warning method based on the Bayesian network as described in any one of claims 1 to 7.

Images