Fault diagnosis method based on fusion of multi-signal flow graph and gradient boosting tree

Abstract

The invention discloses a fault diagnosis method based on fusion of a multi-signal flow graph and a gradient boosting tree, and belongs to the technical field of complex system intelligent operation and maintenance. The method comprises four stages: a first stage, performing anomaly detection on system operation parameters through a period template, and generating parameter anomaly vectors and anomaly time periods; in the second stage, a system multi-signal flow graph is constructed according to an FMECA table, and fault group and suspicious group diagnosis is carried out on abnormal parameters in combination with an improved PTEAMS-RT method to obtain a fuzzy group; in the third stage, feature extraction and screening are carried out on the parameters, a gradient boosting tree classifier is input, and the prediction probability of each fault type is output; and a fourth stage, fusing, normalizing and sorting the fuzzy group and the probability output by the gradient boosting tree to obtain a final fault judgment result. According to the method, by fusing the advantages of causal reasoning and data-driven classification, high-precision, high-robustness and interpretable fault diagnosis is realized, and the method is suitable for fault detection and positioning of spacecrafts, aircrafts, ships and other complex electromechanical systems.

Description

Technical Field

[0001] This invention relates to fault diagnosis methods, and more particularly to a fault diagnosis method based on the fusion of multi-signal flow graphs and gradient boosting trees. This method can be widely applied to fault detection and localization in spacecraft, aircraft, ships, and other complex electromechanical systems. Background Technology

[0002] As the scale and complexity of modern engineering systems (such as aerospace, energy and power, rail transportation, and industrial production equipment) continue to increase, their reliability and safety have become core issues that urgently need to be addressed. Traditional model-based diagnostic methods (such as expert systems and inference methods based on multi-signal flow graphs) have advantages in processing system structural information and causal reasoning, but they are prone to decreased diagnostic accuracy and misjudgments when facing complex nonlinear systems, sensor failures, or noise interference. On the other hand, data-driven machine learning methods (such as support vector machines, neural networks, and gradient boosting trees) can learn the implicit features of a system through a large amount of historical operating data to classify and predict faults, but their diagnostic results often lack interpretability, and their performance degrades significantly when training samples are insufficient or unevenly distributed.

[0003] Therefore, how to effectively integrate the causal reasoning capability based on multi-signal flow graphs with the feature learning capability of data-driven methods to improve the accuracy, robustness and interpretability of fault diagnosis is an important research problem and technical challenge in the field of intelligent operation and maintenance of complex systems. Summary of the Invention

[0004] The purpose of this invention is to overcome the problems of insufficient accuracy, poor robustness, and weak interpretability of existing single diagnostic methods, and to propose a fault diagnosis method based on the fusion of multi-signal flow graphs and gradient boosting trees.

[0005] Figure 1This is a schematic diagram of the overall process of the fault diagnosis method based on the fusion of multi-signal flow graphs and gradient boosting trees according to the present invention. The overall process of the method includes four main stages: S1 Parameter anomaly detection method based on periodic templates, including the training stage S101 and the detection stage S102. A health model is built based on the periodic template, anomalies are detected in real-time parameter data, and the parameter anomaly vector and anomaly time period are output; S2 Fault diagnosis method combining multi-signal flow graphs with "pessimistic" TEAMS-RT (PessimisticTEAMS-RT, PTEAMS-RT, GBDT), firstly, a multi-signal flow graph of the system is established based on the FMECA table, a D matrix is ​​generated, and the improved PTEAMS-RT method is used to diagnose fault groups and suspected groups, resulting in ambiguous groups. This process includes steps S201, S202, S203, and S204; S3 Fault diagnosis method based on feature extraction and gradient boosting trees, including the training stage S301 and the online diagnosis stage S302. Feature extraction is performed using tsfresh, feature filtering is performed using RFE, and the feature is input into the gradient boosting tree (Gradient Boosting Tree). The Tree classifier is used to classify fault types and predict probabilities. The S4 fault diagnosis results are fused by normalizing the fault probabilities of the fuzzy groups and gradient boosting tree outputs obtained by the PTEAMS-RT method, sorting and filtering the Top-N candidate faults, and finally outputting the main fault types and auxiliary diagnosis results.

[0006] Compared with the prior art, the present invention has the following beneficial effects:

[0007] (1) By using periodic templates to achieve anomaly detection, potential abnormal signals can be identified in advance, improving the early warning capability before diagnosis;

[0008] (2) The multi-signal flow graph diagnosis based on PTEAMS-RT has good causal interpretability and remains robust under complex conditions such as sensor failure and missing test points.

[0009] (3) Combining feature extraction and gradient boosting tree methods can capture complex nonlinear relationships and achieve refined classification;

[0010] (4) By using the result fusion mechanism, the advantages of the two methods are complemented, which significantly improves the accuracy, reliability and generalization ability of fault diagnosis. Attached Figure Description

[0011] Figure 1 This is a schematic diagram of the overall process of the fault diagnosis method based on the fusion of multi-signal flow graph and gradient boosting tree of the present invention;

[0012] Figure 2 This is a schematic diagram of the parameter anomaly detection process based on a periodic template.

[0013] Figure 3 This is a schematic diagram of the fault diagnosis process based on the multi-signal flow graph and PTEAMS-RT method;

[0014] Figure 4 This is a schematic diagram of the training and detection process based on the feature extraction and gradient boosting tree method;

[0015] Figure 5 A flowchart illustrating the process of fusing fault diagnosis results. Detailed Implementation Plan

[0016] The following detailed description, with reference to the accompanying drawings, illustrates a fault diagnosis method based on the fusion of multi-signal flow graphs and gradient boosting trees provided by the present invention.

[0017] This invention provides a fault diagnosis method based on the fusion of multi-signal flow graphs and gradient boosting trees. The overall process includes four main stages: parameter anomaly detection, fault diagnosis based on multi-signal flow graphs and the "pessimistic" TEAMS-RT inference method (PTEAMS-RT), fault diagnosis based on feature extraction and gradient boosting trees, and fusion of fault diagnosis results. First, in the parameter anomaly detection stage, a periodic template representing the healthy state is generated using normal operating data. Then, real-time collected parameter data is compared with the template, and anomalies are determined by calculating similarity and anomaly scores, obtaining the parameter anomaly vector and the anomaly time period. Next, a system multi-signal flow graph is constructed based on the system's FMECA table, generating a D matrix. Using the PTEAMS-RT method, combined with the D matrix and parameter anomaly vectors, fault groups and suspected groups are diagnosed as fuzzy groups in the diagnostic results. Subsequently, feature processing is performed on all parameters: features are first extracted using tsfresh, and then feature filtering is performed using the recursive feature elimination (RFE) method. The processed features are input into the gradient boosting tree classifier in the scikit-learn tool to complete the fault type classification. Finally, the probability estimates of all fault types in the fuzzy group are normalized, and the fault type with the highest probability is taken as the final diagnosis result, thus obtaining the fused fault determination.

[0018] S1: A parameter anomaly detection method based on periodic templates, the flowchart of which is shown below. Figure 2 As shown, a system health cycle template is constructed during the training phase, and real-time parameters are compared with the template during the detection phase to quantify the degree of anomaly, generate parameter anomaly vectors and anomaly time periods, and provide input for subsequent fault diagnosis.

[0019] S101: The main purpose of the training phase is to construct a periodic template representing the healthy state using normal system operation data, providing a reference benchmark for subsequent anomaly detection. By performing differential, denoising, normalization, frequency domain analysis, and period division on the parameter sequence, key periodic features of the system are extracted, thereby forming an accurate and stable health template for rapid identification of abnormal states during the detection phase. The specific steps are as follows: (1) Acquire training data: Collect parameter data during normal system operation to construct the health template. (2) Differential processing: Perform differential operation on the training data to reduce the trend in the sequence. (3) Denoising processing: Denoise the differential sequence to reduce the impact of noise on subsequent analysis. (4) Normalization processing: Normalize the denoised sequence so that each parameter can be analyzed on the same scale. (5) Fourier transform: Perform Fourier transform on the normalized sequence to convert it to the frequency domain to obtain spectral information. (6) Period division: Divide the spectrum into periods and extract key periodic features. (7) Generate periodic template: Combine the period division results to generate a periodic template representing the healthy state of the system, providing a reference for subsequent detection.

[0020] S102: The main task of the detection phase is to compare the real-time collected parameter sequence with the healthy period template generated in the training phase. Through methods such as sequence window division, LB_Keogh distance calculation, template passband matching and waveform alignment, the abnormality of each data frame is quantified, and the abnormality of the data frame is judged according to the preset threshold, thereby forming a parameter abnormality vector and abnormal time period, providing input for subsequent fault diagnosis. The specific steps are as follows: (1) Obtain the detection sequence: Collect the real-time parameter data to be detected. (2) Form a sequence window: Divide the detection sequence into a fixed-length window for frame-by-frame analysis. (3) LB_Keogh distance calculation: Calculate the LB_Keogh distance between the sequence window and the period template to quantify the similarity. The calculation formula is shown in formula (1). (4) Template passband matching: Match the window data with the pre-generated period template to determine whether it falls into the healthy passband. (5) Waveform alignment: Align the detection data with the period template through the waveform alignment protocol to improve the accuracy of similarity comparison. (6) Calculate the abnormal score of the data frame: Calculate the abnormal score of each data frame based on similarity and deviation. (7) Anomaly determination: Based on the preset threshold, the data frame is classified as exceeding the threshold (abnormal) or not exceeding the threshold (normal), forming a parameter anomaly vector and anomaly time period.

[0021]

[0022] The query sequence in formula (1) is Q = [q1, q2, ..., q n The candidate sequence is C = [c1, c2, ..., c]. n The preorder sequence is C = [c1, c2, ..., c].n The lower bound sequence is L = [L1, L2, ..., L]. n ].

[0023] S2: A fault diagnosis method combining multi-signal flow graphs and PTEAMS-RT, the flowchart of which is shown below. Figure 3 As shown. In this step, a multi-signal flow graph is first established based on the system's functional structure and Fault Mode and Effects Analysis (FMECA), and a corresponding D matrix is ​​generated based on this flow graph to characterize the dependency relationship between faults and test points in the system. Subsequently, the parameter anomaly vector output from the previous anomaly detection stage is used as input to introduce an improved PTEAMS-RT inference algorithm for fault diagnosis. This method can improve the diagnostic accuracy and robustness under complex conditions such as sensor failure, based on the traditional PTEAMS-RT method. The specific process is as follows:

[0024] S201: Initialize and create four empty sets: unknown group, normal group, fault group, and suspected group, and put all fault modes into the unknown group set;

[0025] S202: Traverse the abnormal parameters, move the only fault source that causes the parameter abnormality from the unknown group to the fault group set, move all other faults that are related to the failed test but are not unique from the unknown group to the suspected group set, and move the remaining fault modes in the unknown group to the normal group set.

[0026] S203: For each fault mode in the suspected group set, iterate through them. If only one of their associated parameters is not included in the current abnormal parameter set, then the fault mode is considered highly suspicious. At this point, combine this fault mode with the fault modes in the fault group set to expand the existing abnormal parameter set FT to FT∪FT. ext This involves treating all parameters associated with these failure modes as potentially abnormal parameters, thereby constructing an expanded set of abnormal parameters.

[0027] S204: Based on the expanded set of abnormal parameters, continue traversing the suspected group set. If it is found that the parameter set associated with a certain fault mode cannot be completely covered by the expanded set of abnormal parameters, then it is removed from the suspected group and transferred to the normal group set. Finally, the fault group and the suspected group are merged and defined as a fuzzy group, which serves as the fault diagnosis output for this stage.

[0028] This method can effectively reduce the probability of misjudgment while maintaining the comprehensiveness of diagnosis, and compared with the traditional TEAMS-RT method, it can still maintain high diagnostic accuracy and robustness in complex situations such as sensor failure and missing measurement points.

[0029] S3: A fault diagnosis method based on feature extraction and gradient boosting trees, the flowchart of which is shown below. Figure 4As shown, the original parameter data is first stratified and divided into training and test sets in a 7:3 ratio, and the order is then shuffled. All preprocessing and feature engineering steps are fitted only on the training set, and the resulting parameters are fixed for use in the testing and online stages to avoid data leakage.

[0030] S301: The specific steps in the training phase are as follows: (1) Data cleaning and outlier removal: missing values ​​are processed in the training data (deletion or imputation), and outliers are removed or truncated using the 3σ principle to reduce the impact of extreme values ​​on the model. (2) Feature extraction: rich statistical and frequency domain features (such as mean, variance, skewness, kurtosis, autocorrelation / partial autocorrelation, Fourier coefficient, energy, entropy, etc.) are automatically extracted from time series data using tsfresh to form a high-dimensional candidate feature set. (3) Feature selection: recursive feature elimination (RFE) combined with cross-validation is used to iteratively select features on the training set based on the feature importance of the learner to obtain a feature subset with strong discriminative power and low redundancy (if necessary, correlation constraints are first applied to remove highly collinear features). (4) Model Training (GBDT): The gradient boosting tree classifier in scikit-learn is used to learn based on the training set and selected features. Parameters can be tuned through cross-validation (number of trees, learning rate, maximum depth, minimum number of sample splits, etc.), and sample weights can be set according to the class distribution to alleviate class imbalance. After training, the model and the preprocessing / feature selector of the entire process are stored. (5) Performance Validation: The performance of the trained gradient boosting tree classification model is validated using the previously divided test data. By inputting test data, multi-dimensional performance indicators are calculated and output, including precision, accuracy, recall, F1 score, and confusion matrix, to comprehensively evaluate the model's fault diagnosis capability and generalization performance. The formulas for the relevant evaluation indicators are shown in formulas (2)-(5).

[0031]

[0032]

[0033] Where TP is the number of true positive cases, FP is the number of false positive cases, TN is the number of true negative cases, and FN is the number of false negative cases.

[0034] S302: In the online diagnosis phase, the trained model and fixed feature engineering process are used for real-time data processing and fault diagnosis. The specific steps are as follows: (1) Data preprocessing and feature generation: Perform the same data cleaning and outlier removal operations on the real-time collected parameter data as in the training phase, and extract and generate the corresponding online feature vectors based on the fixed feature selector in the training phase. (2) Fault probability prediction: Input the generated online feature vectors into the trained gradient boosting tree classifier, and output the predicted probability distribution corresponding to each fault type, providing input basis for subsequent fuzzy group fusion and final fault determination.

[0035] S4: Fault Diagnosis Result Fusion The main task of this step is to fuse the diagnostic results based on multi-signal flow graphs and PTEAMS-RT with those based on feature extraction and gradient boosting trees. This overcomes the uncertainties and limitations of single methods, improving overall diagnostic accuracy and reliability. A flowchart is shown below. Figure 5 As shown. The specific steps are as follows: (1) Fuzzy group reception: First, receive the diagnostic results output from S2 (PTEAMS-RT inference method), that is, the fuzzy group set containing the fault group and the suspected group. As the candidate fault set for result fusion. (2) Fault probability estimation: Combine the output of S3 (gradient boosting tree classifier) ​​to estimate the probability of all candidate faults in the fuzzy group. That is, calculate the predicted probability distribution corresponding to each fault type through the trained classification model. (3) Normalization processing: Normalize the fuzzy group set. The fault prediction probabilities within are normalized so that their sum is 1, as shown in formula (6), to ensure the comparability of different fault probabilities. (4) Sorting and filtering: The normalized probability values ​​are sorted and filtered. Sort the results to obtain the Top-N most likely faults; the Top-1 faults are used as the final main diagnostic results, and the remaining faults are used as auxiliary diagnostic results. (5) Output the results: Output the fused fault diagnosis results, including the most likely fault type and its probability value, while retaining the second most likely fault candidates for subsequent verification and correction by maintenance personnel.

[0036]

[0037] In summary, this invention proposes a fault diagnosis method based on the fusion of multi-signal flow graphs and gradient boosting trees. This method effectively identifies abnormal signals during system operation through periodic template construction and parameter anomaly detection. Combining the inference capabilities of multi-signal flow graphs and the PTEAMS-RT inference method ensures diagnostic robustness under complex conditions such as sensor failure and missing measurement points. Furthermore, the introduction of an intelligent learning method based on feature extraction and gradient boosting trees enables modeling and classification prediction of complex nonlinear relationships. Finally, through a result fusion mechanism, the advantages of the two diagnostic methods are complemented, significantly improving the accuracy, reliability, and generalization ability of fault diagnosis. This invention can be widely applied to fault detection and intelligent operation and maintenance of complex engineering systems (such as aerospace, energy and power, and transportation), possessing high engineering practical value and promising prospects for widespread application.

Claims

1. A fault diagnosis method based on the fusion of multi-signal flow graphs and gradient boosting trees, characterized in that, Includes the following steps: Parameter anomaly detection: Generate a health cycle template using normal system operation data, compare real-time collected parameters with the template, calculate anomaly scores, and generate parameter anomaly vectors and anomaly time periods; Multi-signal flow graph and PTEAMS-RT fault diagnosis: A multi-signal flow graph is established based on the system FMECA table to generate a D matrix; the parameter anomaly vector is input into the improved PTEAMS-RT algorithm to diagnose fault groups and suspected groups, forming fuzzy groups; Feature extraction and gradient boosting tree diagnosis: Features are extracted and filtered from the parameter data and input into the gradient boosting tree classifier to predict the probability of each fault type; Fault diagnosis result fusion: The probabilities of candidate faults in the fuzzy group are normalized, sorted and filtered to output the Top-N faults, and the Top-1 is used as the final fault diagnosis result.

2. The method according to claim 1, characterized in that, The periodic template includes key periodic features extracted through differential analysis, denoising, normalization, frequency domain analysis, and period division.

3. The method according to claim 1, characterized in that, The PTEAMS-RT algorithm includes: initializing an unknown group, a normal group, a fault group, and a suspected group; moving the unique fault source to the fault group, moving related non-unique faults to the suspected group, and moving the remaining faults to the normal group; expanding the parameter matching of the suspected group to improve the suspicion determination; and removing suspected group faults that cannot cover abnormal parameters.

4. The method according to claim 1, characterized in that, The feature extraction is performed by obtaining time-series statistics and frequency domain features through tsfresh, and the feature selection is completed by combining recursive feature elimination (RFE) with cross-validation.

5. The method according to claim 1, characterized in that, The gradient boosting tree classifier learns from the training set and outputs the predicted probability of each fault type, which is then fused with the fuzzy group.

6. The method according to claim 1, characterized in that, The result fusion includes fuzzy group probability normalization, ranking, and Top-N candidate fault screening, where Top-1 is the final primary diagnostic result and the rest are output as auxiliary results.

7. The method according to claim 1, characterized in that, It is applicable to real-time fault detection and location in spacecraft, aircraft, ships and other complex electromechanical systems.

Images