Fault diagnosis method of shield hydraulic propulsion system

Abstract

The invention discloses a fault diagnosis method for a shield hydraulic propulsion system, and the method comprises the steps: 1, regularly collecting sensor data of the hydraulic propulsion system in the operation process of a shield tunneling machine, storing the sensor data in an unknown sample database, manually marking a confirmable sample, and carrying out the feature extraction based on LightGBM; 2, performing initialization training on a diagnosis algorithm based on DBSCAN and LightGBM; 3, in the actual operation process of the shield, distinguishing whether the data collected at each moment is a known sample or an unknown sample through a DBSCAN clustering algorithm; 4, carrying out the recognition and classification of the known sample data through a LightGBM classification model; and 5, extracting data in the database to be updated, automatically training and updating the LightGBM classification model, training and updating the DBSCAN model by combining the sample of the database to be updated and the unknown sample data set, and emptying the database to be updated. The invention relates to the technical field of shield construction, and can solve the problems in the prior art.

Description

Technical Field

[0001] The invention relates to the technical field of shield construction, and in particular to a fault diagnosis method for a shield hydraulic propulsion system. Background Art

[0002] Shield tunneling construction has the characteristics of harsh environment, limited operating space and high safety requirements. The hydraulic propulsion system is one of the core systems of the shield machine. Once a failure occurs, the shield machine will be unable to operate, resulting in losses in manpower, funds and construction period. At present, there are still deficiencies in the fault diagnosis method for the hydraulic propulsion system of the shield machine in China. The main technical problems faced are as follows:

[0003] ① Most fault diagnoses rely on experience summary and are not universal. They are mainly diagnostic methods for one or more components. For example, after manually collecting the digital input or analog input signal value on the PLC, simple logical analysis and judgment are performed to give user prompt information. It is difficult to give effective processing methods for complex phenomena, and no systematic diagnostic system has been formed.

[0004] ② Fault diagnosis for complex systems such as shield propulsion devices requires the ability to diagnose known and unknown faults. Known faults can be summarized manually or historically occurred, but new types of faults or faults that are not considered manually may also occur during the operation of the actual system. Therefore, the fault diagnosis model needs to have the ability to diagnose unknown faults.

[0005] ③ Most of the data sets actually collected by the shield machine are normal sample data, and the fault data are extremely small. Therefore, the sample data of each type are extremely unbalanced, which makes the classification model very biased and causes misjudgment.

[0006] ④ Most of the data generated during the actual operation of the shield hydraulic propulsion system are not labeled, and it is also very difficult to locate the actual faults and data, making it difficult to apply methods that require a large amount of labeled data for training.

[0007] ⑤ Today, the hydraulic system of the shield machine is equipped with a large number of intelligent sensors, and the amount of feature data related to the propulsion system is huge, which requires feature selection, otherwise it will affect the program execution performance and interfere with the classification accuracy of the classification model.

[0008] At present, there have been some studies on fault diagnosis of shield hydraulic propulsion systems, such as the fault diagnosis method using variable weighted Fisher discriminant analysis, and the introduction of fuzzy logic theory and the design of fault prediction inference machine architecture using fuzzy neural networks. However, these data-driven fault diagnosis algorithms will encounter the above problems when applied to actual shield propulsion systems. In the research on diagnostic theory for identifying new fault categories, the distance rejection (DR) mechanism is introduced through the Bayesian network based on probabilistic reasoning, which can realize new fault detection and dynamically update the fault library; there are also studies that use neural networks to distinguish known category samples from new fault samples, and use dynamic structure BP networks to realize the detection of dynamically updated categories of chemical faults, but none of them can guarantee the elimination of misdiagnosis. To deal with the problem of sample imbalance, the weighted SVM algorithm can be used to add different weights to samples of different categories in the objective function to compensate for the impact of category imbalance. For the problem of data and fault location difficulties caused by unlabeled data, unsupervised models can be used to learn the features of unlabeled data for fault detection, or only a small amount of labeled data can be used for feature extraction and then classifier training.

[0009] Although the above method can improve the fault diagnosis accuracy of the shield hydraulic propulsion system to a certain extent, the following problems still exist: 1. It is too dependent on experience and manual operation and is not universal; 2. There is a possibility of misdiagnosis for the detection of new faults; 3. The imbalance of sample data of various types makes the classification model biased and it is easy to misjudge fault samples as normal; 4. It is difficult to locate faults in unlabeled data. Therefore, it is necessary to provide a fault diagnosis method for the shield hydraulic propulsion system that can solve the above problems of the prior art. Summary of the invention

[0010] The purpose of the present invention is to provide a fault diagnosis method for a shield hydraulic propulsion system, which can solve the above-mentioned problems in the prior art.

[0011] The present invention is achieved in that:

[0012] A fault diagnosis method for a shield hydraulic propulsion system comprises the following steps:

[0013] Step 1: The sensor data related to the hydraulic propulsion system of the shield machine during operation is collected regularly through the collection system and stored in the unknown sample database; the data of some identifiable samples are manually categorized and stored in the known sample database, and feature extraction is performed on the unknown sample database and the known sample database based on the feature importance ranking obtained by the LightGBM histogram algorithm;

[0014] Step 2: Select samples from the unknown sample database and the known sample database to initialize the training of the diagnosis algorithm based on DBSCAN and LightGBM;

[0015] Step 3: During the actual operation of the shield machine, the data collected at each moment is distinguished as known samples or unknown samples through the DBSCAN clustering algorithm of the anomaly monitoring model. If it is an unknown sample, the unknown sample is added to the unknown sample database or manually added to the database to be updated, and then go to step 5. If it is a known sample, execute step 4;

[0016] Step 4: The known sample data is identified and classified through the fault classification model, namely the LightGBM classification model, and the diagnosis results are output to complete the diagnosis process. The identification results are stored in the known sample database as the initial model parameters;

[0017] Step 5: When the sample size of the database to be updated exceeds the preset value N train When the data in the database to be updated is extracted, the LightGBM classification model is automatically trained and updated, the unknown sample data set is extracted from the unknown sample database, the DBSCAN model is trained and updated by combining the samples in the database to be updated with the unknown sample data set, and the database to be updated is cleared after the update, waiting for the next automatic training update.

[0018] The step 1 comprises the following sub-steps:

[0019] Step 1.1: Regularly collect sensor data related to the hydraulic propulsion system during the operation of the shield machine through the acquisition system to regularly update the unknown sample database;

[0020] Step 1.2: The technician can view the samples in the unknown sample database through the visual graphical interface software and mark the samples of identifiable categories. After confirmation and submission, the corresponding marked samples are automatically stored in the known sample database; the samples of unknown categories are set as U, and the marked samples of known categories are set as W;

[0021] Step 1.3: The number of sensors is N S , take the time series N K continuous points as a sample, that is, the data set of unknown samples is N T is the number of unknown samples, where the data set corresponding to the sensor in the i-th time window is:

[0022]

[0023] s k，j Represents the kth sensor data of the jth unknown category sample in the time series;

[0024] The data set of known samples is where y i is the tag type, N WFor a known sample (x i ,y i ), the union of the two is M = U ∪ W;

[0025] Step 1.4: The process of LightGBM feature selection is to sort the known category samples W, sort all samples from large to small according to the absolute value of the sample gradient; retain the top-a sample set to obtain the data subset A;

[0026] Step 1.5: Randomly sample a subset C of size b×|B| from the remaining dataset B in the known category samples W, and calculate the estimated variance gain of the subset A∪C To determine the j-th split node on the LightGBM decision tree

[0027]

[0028] g i is the negative gradient of the loss function relative to the output of the LightGBM model, g i is the negative gradient of the loss function relative to the model output, and A l ∪C l and A r ∪C r The number of sets; thereby counting the number of times each feature is used as a partition attribute in the LightGBM decision tree, and sorting the feature importance by the number of times.

[0029] The step 2 comprises the following sub-steps:

[0030] Step 2.1: The ε, minPts parameters and core point list of the anomaly monitoring model, i.e., the DBSCAN model, are initialized during the initialization training. The candidate ε and minPts parameter lists are generated based on the distribution characteristics of the dataset M itself. The final ε and minPts are determined when the clustering results tend to be stable. The core point list consists of the following two types of data:

[0031] i) In a known sample data set W;

[0032] ii) There are at least minPts points within its Euclidean distance ε;

[0033] Step 2.2: Calculate the Euclidean distance from each point in the data set M to other points. The calculation formula is:

[0034] Dist n×n ={dist(i, j)|1≤i≤n, 1≤j≤n};

[0035] Step 2.3: Calculate the mean of each column of Dist in ascending order Subtract its self-attenuation term to get the parameter list Eps of ε list :

[0036]

[0037] Step 2.4: Calculate the list of minPts:

[0038]

[0039] Among them, P j is the number of objects within the radius ε neighborhood of the jth object;

[0040] Step 2.5: Traverse Eps list and MinPt list List, when the number of clusters in the clustering result tends to be stable, select the corresponding ε and minPts;

[0041] The criterion for stabilization is: select the number of clusters with the largest number of consecutive clusters of the same number of clusters, and determine ε and minPts according to the data noise situation. They are divided into three categories: less, normal and more, and select the list subscript K corresponding to ε and minPts in turn. p ;

[0042]

[0043] eps=EPS list [K p ],minsp=minPt list [K p ];

[0044] Since the ratio of known samples to unknown samples is extremely unbalanced, the average accuracy is used to measure the accuracy of the DBSCAN model based on the following rules:

[0045]

[0046] Step 2.6: The dataset W of known category samples is sampled by BorderLine Smote algorithm to obtain a new dataset W′, and LightGBM is trained for classification tasks. LightGBM is a parallel ensemble decision tree algorithm. During the initial training process, r decision tree models are initially established. The dataset W is divided into k parts by k-fold cross-validation method, one of which is used as a validation set and the remaining k-1 parts are used as training sets. A total of k LightGBM classification models are obtained through comprehensive training. The arithmetic mean of the outputs of these k LightGBM classification models is calculated to obtain the final LightGBM model. The accuracy of the LightGBM model is defined as

[0047] Where n is the number of categories, T i is the number of samples of the i-th category correctly judged as the i-th category, N i is the total number of samples of the i-th category.

[0048] The step 3 comprises the following sub-steps:

[0049] Step 3.1: Assume that the data collected in a certain time window is x i , traverse the core point list, if there is a core point p j With x i If it is reachable in the DBSCAN definition, the corresponding sample is determined as a known sample and go to step 4; if there is no core point that meets the conditions, the corresponding sample is determined as an unknown sample and go to step 3.2;

[0050] Step 3.2: Store the unknown category samples in the unknown sample database. The technicians can view the information of the unknown samples in the visual graphical interface software and manually label the corresponding samples with the fault category. The labeled samples are automatically stored in the database to be updated, and the newly created category faults are also automatically added to the fault category library. Go to step 5.

[0051] Compared with the prior art, the present invention has the following beneficial effects:

[0052] 1. The present invention classifies and identifies all samples based on the LightGBM model, and sorts the features based on the LightGBM histogram algorithm. It can distinguish between known category samples and unknown fault samples according to the unclear characteristics of shield hydraulic system fault types, and realize data with a distribution similar to historical data, with higher classification accuracy, effectively reducing misdiagnosis.

[0053] 2. The clustering algorithm of the present invention is based on the DBSCAN model, combined with the borderline-smote and k-fold cross-validation LightGBM classification algorithm to identify, classify and diagnose known sample data, and distinguish and identify unknown samples and confirmable samples. It can make full use of a large amount of unlabeled and a small amount of labeled data in the actual engineering environment, overcome the bias problem of the classification model caused by the imbalance between fault samples and normal category samples, and the fault location of unlabeled data is more accurate.

[0054] 3. The present invention combines the LightGBM model with the DBSCAN model, can automatically extract samples from the database to be updated for automatic training, can analyze and select input features, has faster training speed, lower memory usage, and does not rely on manual operation and experience, and has good universality. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is a flow chart of the fault diagnosis method of the shield hydraulic propulsion system of the present invention;

[0056] Figure 2 is a comparison chart of the average accuracy of fault classification of the fault diagnosis method of the shield hydraulic propulsion system of the present invention and the traditional algorithm in Example 1;

[0057] Figure 3 It is a visualization result diagram after principal component analysis of the DBSCAN model training results in the fault diagnosis method of the shield hydraulic propulsion system of the present invention in Example 1. DETAILED DESCRIPTION

[0058] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

[0059] Please see attached Figure 1 , a fault diagnosis method for a shield hydraulic propulsion system, comprising the following steps:

[0060] Step 1: The sensor data related to the hydraulic propulsion system of the shield machine during operation is regularly collected through the acquisition system and stored in the unknown sample database; the data of some identifiable samples are manually labeled and stored in the known sample database, and feature extraction is performed on the unknown sample database and the known sample database based on the feature importance ranking obtained based on the histogram algorithm of LightGBM (Light Gradient Boosting Machine, a gradient boosting decision tree model).

[0061] The step 1 comprises the following sub-steps:

[0062] Step 1.1: The data collection system is used to regularly collect sensor data related to the hydraulic propulsion system during the operation of the shield machine, and is used to regularly update the unknown sample database.

[0063] Step 1.2: Technicians can view samples in the unknown sample database through the visual graphical interface software and mark samples of identifiable categories. After confirmation and submission, the corresponding marked samples are automatically stored in the known sample database; samples of unknown categories are set as U, and marked samples of known categories are set as W.

[0064] Step 1.3: The number of sensors is N S , take the time series N K continuous points as a sample, that is, the data set of unknown samples is N T is the number of unknown samples, where the data set corresponding to the sensor in the i-th time window is:

[0065]

[0066] s k，j Represents the kth sensor data of the jth unknown category sample in the time series.

[0067] The data set of known samples is where y i is the tag type, N W For a known sample (x i ,y i ), the union of the two is M = U ∪ W.

[0068] Step 1.4: The process of LightGBM feature selection is to sort the known category samples W, sort all samples from large to small according to the absolute value of the sample gradient; retain the top-a sample set to obtain the data subset A.

[0069] Step 1.5: Randomly sample a subset C of size b×|B| from the remaining dataset B in the known category samples W, and calculate the estimated variance gain of the subset A∪C To determine the j-th split node on the LightGBM decision tree

[0070]

[0071] g i is the negative gradient of the loss function relative to the output of the LightGBM model, g i is the negative gradient of the loss function relative to the model output, and A l ∪C l and A r ∪C r The number of sets of . This counts the number of times each feature is used as a partition attribute in the LightGBM decision tree, and sorts the feature importance by the number of times.

[0072] The input of the LightGBM model is sensor data, that is, the collected samples, and the output is the category of the samples, that is, known samples and unknown samples, which can be used to classify samples.

[0073] Step 2: Select samples from the unknown sample database and the known sample database to initialize the training of the diagnosis algorithm based on DBSCAN and LightGBM.

[0074] Before running the diagnostic algorithm for the first time, the diagnostic system needs to be initialized and trained to obtain the initial diagnostic model. The diagnostic model is based on DBSCAN and LightGBM. That is, after the LightGBM model completes the classification of the samples, the classified model is trained with the DBSCAN model. A manual training function for the diagnostic model is provided in the visual graphical interface software, and all samples from the known sample database and the unknown sample database are extracted for initialization training.

[0075] The step 2 comprises the following sub-steps:

[0076] Step 2.1: The ε, minPts parameters and core point list of the anomaly monitoring model, i.e., DBSCAN (Density-Based Spatial Clustering of Applications with Noise) model, are initialized during the initialization training. The candidate ε and minPts parameter list are generated by the distribution characteristics of the data set M itself. The final ε and minPts are determined when the clustering results tend to be stable. The core point list consists of the following two types of data:

[0077] i) In a known sample data set W;

[0078] ii) There are at least minPts points within its Euclidean distance ε.

[0079] Step 2.2: Calculate the Euclidean distance from each point in the data set M to other points. The calculation formula is:

[0080] Dist n×n ={dist(i,j)|1≤i≤n, 1≤j≤n}.

[0081] Step 2.3: Calculate the mean of each column of Dist in ascending order Subtract its self-attenuation term to get the parameter list Eps of ε list :

[0082]

[0083] Step 2.4: Calculate the list of minPts:

[0084]

[0085] Among them, P j is the number of objects within the neighborhood of radius ε of the jth object.

[0086] Step 2.5: Traverse Eps list and MinPt listList, when the number of clusters in the clustering result tends to be stable, select the corresponding ε and minPts.

[0087] The criterion for stabilization is: select the number of clusters with the largest number of consecutive clusters of the same number of clusters, and determine ε and minPts according to the data noise situation. They are divided into three categories: less, normal and more, and select the list subscript K corresponding to ε and minPts in turn. p .

[0088]

[0089] eps=EPS list [K p ],minsp=minPt list [K p ].

[0090] Since the ratio of known samples to unknown samples is extremely unbalanced, the average accuracy is used to measure the accuracy of the DBSCAN model based on the following rules:

[0091]

[0092] Step 2.6: The dataset W of known category samples is sampled by BorderLine Smote algorithm to obtain a new dataset W′, and LightGBM is trained for classification tasks. LightGBM is a parallel ensemble decision tree algorithm. During the initial training process, r decision tree models are initially established. The dataset W is divided into k parts by k-fold cross-validation method, one of which is used as a validation set and the remaining k-1 parts are used as training sets. A total of k LightGBM classification models are obtained through comprehensive training. The arithmetic mean of the outputs of these k LightGBM classification models is calculated to obtain the final LightGBM model. The accuracy of the LightGBM model is defined as

[0093] Where n is the number of categories, T i is the number of samples of the i-th category correctly judged as the i-th category, N i is the total number of samples of the i-th category.

[0094] Step 3: During the actual operation of the shield machine, the data collected at each moment is distinguished as a known sample or an unknown sample through the DBSCAN clustering algorithm of the anomaly monitoring model. If it is an unknown sample, the unknown sample is added to the unknown sample database or manually added to the database to be updated, and then go to step 5. If it is a known sample, execute step 4.

[0095] The step 3 comprises the following sub-steps:

[0096] Step 3.1: Assume that the data collected in a certain time window is x i , traverse the core point list, if there is a core point p j With x i If it is reachable in the DBSCAN definition, the corresponding sample is determined as a known sample and go to step 4; if there is no core point that meets the conditions, the corresponding sample is determined as an unknown sample and execute step 3.2.

[0097] Step 3.2: Store the unknown category samples in the unknown sample database. The technicians can view the information of the unknown samples in the visual graphical interface software and manually label the corresponding samples with the fault category. The labeled samples are automatically stored in the database to be updated, and the newly created category faults are also automatically added to the fault category library. Go to step 5.

[0098] Step 4: The known sample data is identified and classified through the fault classification model, namely the LightGBM classification model, and the diagnosis results are output to complete the diagnosis process. The identification results are stored in the known sample database as the initial model parameters.

[0099] Step 5: When the sample size of the database to be updated exceeds the preset value N train When the data in the database to be updated is extracted, the LightGBM classification model is automatically trained and updated, the unknown sample data set is extracted from the unknown sample database, the DBSCAN model is trained and updated by combining the samples in the database to be updated with the unknown sample data set, and the database to be updated is cleared after the update, waiting for the next automatic training update.

[0100] Embodiment 1:

[0101] Take the effect of the fault diagnosis system in the actual construction process in a certain working area of ​​the shield as an example. The acquisition system collects data every time the shield advances 20 mm. K =6, that is, every 6 sensor data in the time series are regarded as a group of samples, and there are 1891 original sensor signals.

[0102] The fault forms targeted by the experiment are too slow propulsion and inconsistent propulsion between cylinders. First, 810 sets of data are extracted from the ring report as the training set, which consists of 400 sets of normal propulsion data, 400 sets of slow propulsion data, and 10 sets of too slow propulsion data. Another 820 sets of data are extracted from the test set, including 400 sets of normal propulsion data, 400 sets of slow propulsion data, 10 sets of too slow propulsion data and 10 sets of inconsistent propulsion data, that is, the inconsistent propulsion fault is an unknown category relative to data set 1.

[0103] First, LightGBM is used to extract features and sort them in descending order according to feature importance. In the actual experiment, the first 64 variables are used as input variables.

[0104] The DBSCAN model is trained using dataset 1 as the training set and dataset 2 as the test set. The data noise level is level=more and the hyperparameter λ is a =0.8,λ b = 0.8, the program implementation relies on the sklearn machine learning library of the python language. The test results are visualized after principal component analysis (PCA) as shown in the attached figure. Figure 3 As shown, it is calculated according to the following definition

[0105]

[0106] Next, based on the ring report data mentioned above, we also use data set 1 as the training set, data set 2 as the test set, and the parameter K of the k-fold cross-validation method is 7. We compare the difference between using the oversampling algorithm BorderLineSmote to balance the number of samples and not using the oversampling algorithm. The oversampling balance strategy is to make the number of samples in each category the same. The average accuracy under multiple classifications is defined as follows:

[0107]

[0108] Where n is the number of categories, T i is the number of samples of the i-th category correctly judged as the i-th category, N i is the total number of samples of the i-th category.

[0109] The main setting parameters related to LightGBM are max_depth=10, num_leaves=31, learning_rate=0.1. The program depends on the lightgbm library in Python. Figure 2 As shown, compared with traditional BP network, random forest, XgBoost, and GBDT algorithms, the present invention combines the oversampling algorithm BorderLineSmote, the K-fold cross validation algorithm, and LightGBM to achieve the highest average accuracy in fault classification of the propulsion system test data set.

[0110] The above are only preferred embodiments of the present invention and are not intended to limit the scope of protection of the invention. Therefore, any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A fault diagnosis method for a shield hydraulic propulsion system, characterized by: The following steps are involved: Step 1: The sensor data related to the hydraulic propulsion system of the shield machine during operation is collected regularly through the collection system and stored in the unknown sample database; the data of some identifiable samples are manually categorized and stored in the known sample database, and feature extraction is performed on the unknown sample database and the known sample database based on the feature importance ranking obtained by the LightGBM histogram algorithm; Step 2: Select samples from the unknown sample database and the known sample database to initialize the training of the diagnosis algorithm based on DBSCAN and LightGBM; Step 3: During the actual operation of the shield machine, the data collected at each moment is distinguished as known samples or unknown samples through the DBSCAN clustering algorithm of the anomaly monitoring model. If it is an unknown sample, the unknown sample is added to the unknown sample database or manually added to the database to be updated, and then go to step 5. If it is a known sample, execute step 4; Step 4: The known sample data is identified and classified through the fault classification model, namely the LightGBM classification model, and the diagnosis results are output to complete the diagnosis process. The identification results are stored in the known sample database as the initial model parameters; Step 5: When the sample size of the database to be updated exceeds the preset value N train When the data in the database to be updated is extracted, the LightGBM classification model is automatically trained and updated, the unknown sample data set is extracted from the unknown sample database, the DBSCAN model is trained and updated by combining the samples in the database to be updated with the unknown sample data set, and the database to be updated is cleared after the update, waiting for the next automatic training update.

2. The fault diagnosis method for a shield hydraulic propulsion system according to claim 1 is characterized in that: The step 1 comprises the following sub-steps: Step 1.1: Regularly collect sensor data related to the hydraulic propulsion system during the operation of the shield machine through the acquisition system to regularly update the unknown sample database; Step 1.2: The technician can view the samples in the unknown sample database through the visual graphical interface software and mark the samples of identifiable categories. After confirmation and submission, the corresponding marked samples are automatically stored in the known sample database; the samples of unknown categories are set as U, and the marked samples of known categories are set as W; Step 1.3: The number of sensors is N S , take the time series N K continuous points as a sample, that is, the data set of unknown samples is N T is the number of unknown samples, where the data set corresponding to the sensor in the i-th time window is: s k，j Represents the kth sensor data of the jth unknown category sample in the time series; The data set of known samples is where y i is the tag type, N w For a known sample (x i ,y i ), the union of the two is M = U ∪ W; Step 1.4: The process of LightGBM feature selection is to sort the known category samples W, sort all samples from large to small according to the absolute value of the sample gradient; retain the top-a sample set to obtain the data subset A; Step 1.5: Randomly sample a subset C of size b×|B| from the remaining dataset B in the known category samples W, and calculate the estimated variance gain of the subset A∪C To determine the j-th split node on the LightGBM decision tree g i is the negative gradient of the loss function relative to the output of the LightGBM model, g i is the negative gradient of the loss function relative to the model output, and A l ∪C l and A r ∪C r The number of sets; thereby counting the number of times each feature is used as a partition attribute in the LightGBM decision tree, and sorting the feature importance by the number of times.

3. The fault diagnosis method for a shield hydraulic propulsion system according to claim 1 is characterized in that: The step 2 comprises the following sub-steps: Step 2.1: The ε, minPts parameters and core point list of the anomaly monitoring model, i.e., the DBSCAN model, are initialized in the initialization training. The candidate ε and minPts parameter lists are generated by the distribution characteristics of the dataset M itself. The final ε and minPts are determined when the clustering results tend to be stable. The core point list consists of the following two types of data: i) In a known sample data set W; ii) There are at least minPts points within its Euclidean distance ε; Step 2.2: Calculate the Euclidean distance from each point in the data set M to other points. The calculation formula is: Dist n×n ={dist(i,j)|1≤i≤n,1≤j≤n}; Step 2.3: Calculate the mean of each column of Dist in ascending order Subtract its self-attenuation term to get the parameter list Eps of ε list : Step 2.4: Calculate the list of minPts: Among them, P j is the number of objects within the radius ε neighborhood of the jth object; Step 2.5: Traverse Eps list and MinPt list List, when the number of clusters in the clustering result tends to be stable, select the corresponding ε and minPts; The criterion for stabilization is: select the cluster with the largest number of consecutive clusters with the same number of clusters, and determine ε and minPts according to the data noise, which are divided into three categories: less, normal and more, and select the list subscript K corresponding to ε and minPts in turn p ; eps=Eps list [K p ],minsp=minPt list [K p ]; Since the ratio of known samples to unknown samples is extremely unbalanced, the average accuracy is used to measure the accuracy of the DBSCAN model based on the following rules: Step 2.6: The dataset W of known category samples is sampled by BorderLine Smote (i.e. B-SMOTE) algorithm to obtain a new dataset W′, and LightGBM is trained for classification tasks. LightGBM is a parallel ensemble decision tree algorithm. During the initial training process, r decision tree models are initially established. The dataset W is divided into k parts by k-fold cross-validation method, one of which is used as a validation set and the remaining k-1 parts are used as training sets. A total of k LightGBM classification models are obtained through comprehensive training. The arithmetic mean of the outputs of these k LightGBM classification models is calculated to obtain the final LightGBM model. The accuracy of the LightGBM model is defined as Where n is the number of categories, T i is the number of samples of the i-th category correctly judged as the i-th category, N i is the total number of samples of the i-th category.

4. The fault diagnosis method for a shield hydraulic propulsion system according to claim 1 is characterized in that: The step 3 comprises the following sub-steps: Step 3.1: Assume that the data collected in a certain time window is x i , traverse the core point list, if there is a core point p j With x i If it is reachable in the DBSCAN definition, the corresponding sample is determined as a known sample and go to step 4; if there is no core point that meets the conditions, the corresponding sample is determined as an unknown sample and go to step 3.2; Step 3.2: Store the unknown category samples in the unknown sample database. The technicians can view the information of the unknown samples in the visual graphical interface software and manually label the corresponding samples with the fault category. The labeled samples are automatically stored in the database to be updated, and the newly created category faults are also automatically added to the fault category library. Go to step 5.

Images