Shield tail seal system pressure abnormal condition detection method

Abstract

The application provides a shield tail sealing system pressure abnormal state detection method, and belongs to the technical field of shield machines. Specifically, the following steps are included: S1, collecting pressure data recorded in a shield tail sealing system during shield machine construction; S2, preprocessing data recorded by a sensor; S3, using a time sequence segmentation method to segment and extract the pressure data and establishing a data set; S4, performing offline anomaly detection by using a density clustering DBSCAN method in unsupervised learning; S5, expanding abnormal sample points by using a Smote algorithm; and S6, training an online anomaly detection model by using the expanded data set by using a RUSBoost algorithm. The application provides a shield tail sealing pressure data processing method, so that shield tail sealing rear cavity pressure data recorded during construction can be fully utilized, so that whether the shield tail sealing rear cavity pressure state is abnormal can be quickly found during shield machine tunneling, and a certain guarantee is provided for the safe construction of the shield machine.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of shield machines, and particularly relates to a shield tail sealing system pressure abnormal state detection method. TECHNICAL BACKGROUND

[0002] A shield machine is a tunneling equipment and plays a huge role in tunnel traffic construction. However, many problems often occur in the working process of the shield machine due to complex construction conditions, and even accidents are caused. The shield tail sealing system is located at the tail of the shield head of the shield machine and directly contacts with the segment and the mud outside, and is generally a multilayer annular sealing pair composed of sealing grease and a shield tail brush. It is a system for ensuring the safe construction of the shield machine and can effectively cope with construction risks such as groundwater intrusion. Failure of the shield tail sealing system will cause damage to the inside of the shield machine and harm to the safety of the workers.

[0003] Shield tail sealing grease is a key material in the sealing system, which is generally a non-Newtonian fluid and has good lubricity and sealing property. At the same time, the water tightness of the grease also determines the upper limit of the sealing performance of the sealing cavity. However, the non-Newtonian characteristics of the pressure conduction of the sealing grease also bring some difficulties to the monitoring of the shield tail sealing state. The shield tail sealing system is in a small space, and the space volume will change with the adjustment of the posture of the shield machine. At the same time, the poor working environment and the pressure transmission characteristics of the non-Newtonian fluid make the fault phenomenon and fault position of the sealing system cannot be accurately observed. Therefore, the conventional supervised learning method cannot be used to detect whether the shield tail fails.

[0004] In the process of shield machine construction, the pressure of each cavity in the shield tail sealing system can be monitored and recorded in real time through a pressure sensor. Since the shield tail rear cavity is directly in contact with the mud, accidents generally occur at the rear cavity. Therefore, the pressure trend of the shield tail rear cavity can be judged by using an unsupervised learning method to determine whether the pressure change of the rear cavity is in an abnormal state, thereby providing a certain guarantee for the safe construction of the shield machine. SUMMARY

[0005] In view of the problems in the prior art, the application provides a shield tail sealing system pressure abnormal state detection method, which can ensure that the recorded shield tail sealing rear cavity pressure data in the construction process can be fully utilized, so that the shield machine can quickly find out whether the shield tail sealing rear cavity pressure state is abnormal in the tunneling process, thereby providing reliable protection for the safe construction of the shield machine.

[0006] In order to achieve the above purpose, the technical scheme adopted by the application is as follows:

[0007] The application discloses a shield tail sealing system pressure abnormal state detection method, which comprises the following steps: S1, collecting recorded rear cavity pressure data of a shield tail sealing system in a shield construction process; specifically,

[0008] According to the tunneling period and the advancing speed of the shield, effective data are selected, and the data record format is unified.

[0009] S2, pre-processing the rear cavity pressure data recorded by the sensor and removing noise data; specifically: a, statistical data difference: the pressure sensor used by the shield has a certain probability to generate obvious abnormal values, and the abnormal value duration is 1 second, so the current data is generally selected to be subtracted from the data of the previous second, and the data difference is statistically analyzed; the statistical result shows that the difference value when the abnormal condition occurs is outside the 6sigma range, so 6sigma is selected as the threshold, wherein sigma refers to the standard deviation calculated in the statistical data difference;

[0010] b, data cleaning: judging whether the difference value is less than the threshold value determined in step a, if the difference value exceeds the threshold value, the data of the previous second is used to replace the current data.

[0011] S3, using a time series segmentation method to segment the rear cavity pressure data and establishing a data set; specifically:

[0012] The sliding window method is used to segment the rear cavity pressure data pre-processed in step S2, the mean (u), the slope (k), the length (L) and the pressure difference value (dp) of each segment are extracted, wherein the pressure difference value is equal to the product of the slope and the length, and the pressure difference value is introduced so that the data set can better reflect the trend of pressure change. The extracted data of each segment is summarized to obtain a data set (the data set is composed of samples, and the samples include normal samples and abnormal samples), that is, a single sample is extracted from the segmented data, and the mean, the slope, the length and the pressure difference of each segment are extracted. After a plurality of construction data are processed in this way, a data set composed of a plurality of samples is obtained.

[0013] S4, performing offline anomaly detection, and performing offline anomaly detection on the data set obtained in step S3 by using a density clustering (DBSCAN) method in unsupervised learning, discovering abnormal sample points in the data set, and dividing the samples in the data set into normal samples and abnormal samples; specifically:

[0014] a, first, the K value is selected, the data set obtained after step S3 has four dimensions k, u, L and dp, according to the empirical formula of density clustering, the K value selected in the application is 2 times the number of dimensions minus 1;

[0015] b. Set neighborhood radius: After the value of K is determined, the K-distance plot is calculated and drawn, the elbow point can be found by the elbow method, and the value of the point is selected as the neighborhood radius;

[0016] c. Select the minimum number of samples of the adjustable parameter cluster: According to the empirical formula of density clustering, the minimum number of samples of the adjustable parameter cluster is selected as 2 times the data dimension;

[0017] d. Find abnormal samples: After the basic adjustable parameters of density clustering, neighborhood radius and adjustable parameter cluster minimum sample number are determined, the data set obtained in step S3 is processed using the density clustering algorithm, and abnormal samples are found and marked.

[0018] S5, the abnormal sample points are divided into two groups, and the Smote algorithm is used to expand the abnormal sample points, thereby reducing the imbalance of the samples; Specifically:

[0019] Generally, the number of abnormal samples is much smaller than the number of normal samples, and direct use for training can easily lead to low model accuracy. Therefore, in order to reduce the error caused by sample imbalance, the Smote algorithm is selected to expand the number of abnormal samples; The abnormal sample points found in step S4 are generally distributed on both sides of the normal sample points, and the distribution domains of the normal sample points and the abnormal sample points are symmetric about k=0. If the Smote algorithm is directly used to expand the abnormal sample points, there is a high probability that the expanded abnormal sample points will fall into the normal sample domain, which will adversely affect the classification result. In order to avoid interference between the expanded abnormal samples and the normal samples, the data set is divided into two parts according to the k value of the sample to distribute and expand, wherein the k value of the first group of abnormal samples is greater than 0, and the k value of the second group of abnormal samples is less than 0. Then the Smote algorithm is used to expand the two groups of abnormal samples respectively, and finally the expanded two groups of abnormal samples and the normal samples are mixed together to obtain a new data set that can be used for online detection model training.

[0020] S6, using the RUSBoost algorithm to train the online anomaly detection model using the expanded data set. Specifically:

[0021] a. The expanded new data set obtained in step S5 is divided into modeling samples (80%) and detection samples (20%);

[0022] b. Train the online anomaly detection model using the RUSBoost algorithm through the modeling samples; the online anomaly detection model is a RUSBoost classification model.

[0023] c. Verify the trained online anomaly detection model through the detection samples.

[0024] Finally, the problem that density clustering cannot perform real-time anomaly detection is effectively avoided by using the trained online anomaly detection model, and the rapid detection of the pressure state of the rear cavity after the shield tail sealing is realized.

[0025] Further, the time series segmentation in the step S3 refers to fitting the time series T with a length of n with K straight lines. Because K is usually much smaller than n, this representation makes the data storage, transmission and calculation more efficient.

[0026] The present application has the beneficial effects that:

[0027] (1) The present application uses a time series segmentation method to quickly extract the change characteristics of the pressure of the rear cavity after the shield tail sealing, and performs real-time detection on the change trend of the pressure through an online anomaly detection model, so that whether the pressure change trend is abnormal can be found in time during the construction process.

[0028] (2) The present application sets the mean (u), the slope (k), the length (L) and the pressure difference value (dp) as the four dimensions in the construction of the data set, wherein the introduction of the pressure difference value can better reflect the change trend of the pressure of the rear cavity.

[0029] (3) Meanwhile, the present application first finds abnormal samples through offline anomaly detection, and then trains an online anomaly detection model with labeled samples, which can effectively avoid the problem that density clustering cannot perform real-time anomaly detection. BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 is the pressure data before cleaning shown in the specific example of the present application.

[0031] Figure 2 is the pressure data after cleaning shown in the specific example of the present application.

[0032] Figure 3 is the pressure data after time series segmentation shown in the specific example of the present application.

[0033] Figure 4 is the K-distance graph shown in the specific example of the present application.

[0034] Figure 5 is the abnormal sample and normal sample found after density clustering in the specific example of the present application.

[0035] Figure 6 is the sample scatter plot after expansion by the Smote algorithm shown in the specific example of the present application.

[0036] Figure 7 is the confusion matrix of the online anomaly detection algorithm shown in the specific example of the present application. DETAILED DESCRIPTION

[0037] The calculation method of the shield tail sealing system is further described below with reference to an embodiment.

[0038] S1, collect the recorded back cavity pressure data in the shield tail sealing system during the shield construction process

[0039] In this example, a certain tunnel project is taken as the experimental object. First, the PLC in the shield machine is connected to read the back cavity pressure data stored in a specific address, the reading frequency is 1 s per piece, and the read data is written into the database and grouped according to the construction process. Finally, only the pressure data in the pushing process is retained.

[0040] S2, pre-process the data recorded by the sensor to remove noise data

[0041] Figure 1 The pressure data recorded by the back cavity pressure sensor before data cleaning is shown. It can be seen from Figure 1 that there are mutation values in the data. In order to remove the influence of mutation values, the data of the next second is subtracted from the data of the previous second, and the difference is counted. It is found that the difference value of the data generally conforms to the normal distribution, and the difference value at the abnormal data point is outside the 6σ range of the normal distribution. The 6σ here is about 12, so 12 can be selected as the threshold value. When the pressure difference value is greater than the threshold value, the data of the point is removed and the data of the previous second is filled.

[0042] Figure 2 The pressure data after cleaning is shown.

[0043] S3, use time series segmentation method to segment the pressure data, extract the mean (u), slope (k), length (L) and pressure difference (dp) of each piece of data, and establish a data set

[0044] Figure 3 The pressure data after time series segmentation is shown. The sliding window method is used to segment the pressure data according to its change trend, and the mean (u), slope (k) and length (L) of each segment are removed, and the pressure difference (dp) of each segment is calculated.

[0045] S4, offline anomaly detection, find abnormal sample points in the data set by density clustering (DBSCAN) method in unsupervised learning, and divide the samples into normal samples and abnormal samples

[0046] Figure 4 The K-distance diagram calculated in this example is shown. Figure 5 The scatter plot composed of abnormal samples and normal samples found in this example is shown.

[0047] Since the dimension of data is 4, the value of K is selected as 7 in this example, and the K-distance graph of all samples is calculated based on the selected value of K. Figure 4 The neighborhood radius is selected as 0.048 according to the calculation result in the formula (2) and the elbow method. Finally, the minimum sample number of the adjustable parameter cluster is selected as 8, and the density clustering is performed. The density clustering can automatically find the abnormal samples and mark them, and the final result is shown in FIG. 4. Figure 5

[0048] S5, divide the abnormal sample points into two groups and respectively expand the abnormal sample points through the Smote algorithm, so as to reduce the imbalance of the samples (according to the abnormal detection result, the abnormal sample points are distributed outside the normal sample points, as shown in FIG. 3, and it can be seen that the abnormal samples basically surround the normal samples. If the Smote algorithm is directly used for expansion, the expanded points may appear in the normal sample domain. This is related to the interpolation principle used by the Smote algorithm.); Figure 5

[0049] Figure 6 The sample scatter plot after expansion by the Smote algorithm in this example is shown in FIG. 5. In the original data set, the ratio of normal samples to abnormal samples is about 45:1, and the sample imbalance is too large, so it is necessary to expand the abnormal samples. At the same time, the abnormal samples are distributed outside the data set, and as shown in FIG. 3, the k value is symmetrically distributed about k = 0. In order to prevent the expanded data from interfering with the normal samples, the abnormal samples need to be divided into two groups and expanded respectively. The k value of the first group of abnormal samples is greater than 0, and the k value of the second group of abnormal samples is less than 0. The expansion data multiple is selected as 9 here, and the ratio of the normal samples to the abnormal samples after expansion is 5:1, as shown in FIG. 6. It can be seen that there is still a clear boundary between the abnormal samples and the normal samples. Then, the two groups of expanded abnormal samples and the normal samples are mixed together to obtain a new data set that can be used for online detection model training. Figure 5 Figure 6

[0050] S6, train the online anomaly detection model, divide the expanded and labeled data into modeling samples (80%) and detection samples (20%), and train the RUSBoost classification model.

[0051] Figure 7 ​​​​The confusion matrix of the online classification model in the present example is shown. RUSBoost is a commonly used unbalanced sample classification algorithm, which is a combination of RUS (random undersampling) and Adaboost. Although the data set has been balanced using the Smote algorithm, the imbalance in the data set still exists, so this algorithm is selected. The algorithm extracts a certain amount of majority samples and a balanced amount of minority samples to form a training sample set, and trains multiple weak classifiers through an iterative method. In the training process, the algorithm automatically adjusts the weights of the training samples to improve the classification accuracy. Finally, all weak classifiers are combined to obtain a high-precision classifier. In the present example, 80% of the data is randomly extracted as the training sample, and the remaining data is used as the detection sample, and the RUSBoost classification model is trained. The confusion matrix of the model is shown in Figure 7 The accuracy of the model on the detection data set is 99.36%, the precision is 98.2%, and the recall is 98.2%. The overall accuracy of the model is high and can meet the needs of online detection.

[0052] The method effectively solves the problem that density clustering cannot be used for real-time anomaly detection and the problem of unbalanced sample quantity. The online anomaly detection model trained through the above steps can realize real-time and rapid detection of the pressure state of the rear cavity of the shield tail seal.

[0053] The above-described embodiments only express the implementation of the present application, but should not be interpreted as limiting the scope of the patent of the present application. It should be noted that for those skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are all within the scope of protection of the present application.

Claims

1. A method for detecting abnormal pressure conditions in a shield tail sealing system, characterized in that, Includes the following steps: Step S1: Collect the pressure data of the rear chamber in the shield tail sealing system recorded during the tunnel boring machine construction process; Step S2: Preprocess the rear chamber pressure data recorded by the sensor and remove noise data; Step S3: Use time series segmentation method to segment the back chamber pressure data and build a dataset; Step S4: Perform offline anomaly detection. The dataset obtained in Step S3 is processed using the density-based clustering (DBSCAN) method in unsupervised learning to detect anomalous sample points and categorize the samples into normal and anomalous samples. Specifically: Step S4.1: First, select the K value. After processing in step S3, the resulting dataset has four dimensions: k, u, L, and dp. The K value is twice the number of dimensions minus 1. Step S4.2: Set the neighborhood radius: After the K value is determined, calculate and draw the K-distance graph, find the inflection point by using the elbow method, and select the value of that point as the neighborhood radius; Step S4.3: Select the minimum number of samples for the adjustable parameter cluster: Based on the empirical formula for density clustering, select the minimum number of samples for the adjustable parameter cluster as twice the data dimension; Step S4.4: Discovering anomalous samples: After determining the basic adjustable parameter neighborhood radius and the minimum number of samples per cluster for density clustering, the density clustering algorithm is used to process the dataset obtained in step S3 to discover and mark anomalous samples. Step S5: Divide the outlier sample points into two groups, and expand the outlier sample points using the Smote algorithm to reduce sample imbalance; specifically: To avoid interference between the expanded anomalous samples and normal samples, the dataset was divided into two parts based on the k-value of the samples for expansion. The k-values ​​of the first group of anomalous samples were all greater than 0, while the k-values ​​of the second group of anomalous samples were all less than 0. Then, the Smote algorithm was used to expand the two groups of anomalous samples separately. Finally, the two expanded groups of anomalous samples were mixed with the normal samples to obtain a new dataset that can be used for training online detection models. Step S6: Use the RUSBoost algorithm to train an online anomaly detection model using the expanded dataset.

2. The method for detecting abnormal pressure in a shield tail sealing system according to claim 1, characterized in that, The specific steps of S1 are as follows: Based on the tunneling cycle and advance speed of the tunnel boring machine, valid data is selected. Here, only the rear cavity pressure data recorded during continuous advance is selected, and the data recording format is standardized.

3. The method for detecting abnormal pressure in a shield tail sealing system according to claim 1, characterized in that, The specific steps of S2 are as follows: Step S2.1, Statistical data difference: Select the difference between the current data and the data of the previous second, and perform statistical analysis on the data difference. Select 6σ as the threshold, where σ refers to the standard deviation calculated in the statistical analysis of data difference. Step S2.2, Data Cleaning: Determine whether the difference is less than the threshold determined in step a. If the difference exceeds the threshold, use the data from the previous second to replace the current data.

4. The method for detecting abnormal pressure in a shield tail sealing system according to claim 1, characterized in that, The specific steps of step S3 are as follows: The sliding window method is used to segment the back chamber pressure data after preprocessing in step S2, and the mean u, slope k, length L and pressure difference value dp of each segment are extracted. The pressure difference value is equal to the product of the slope and the length. Its introduction can make the dataset better reflect the trend of pressure change. The extracted data of each segment are then summarized to obtain the dataset.

5. The method for detecting abnormal pressure in a shield tail sealing system according to claim 1, characterized in that, The specific steps of step S6 are as follows: Step S6.1: Divide the expanded new dataset obtained in step S5 into two parts: modeling samples and detection samples; Step S6.2: Train the online anomaly detection model using the RUSBoost algorithm based on the modeling samples; the online anomaly detection model is a RUSBoost classification model. Step S6.3: Validate the trained online anomaly detection model using test samples; Finally, by using a well-trained online anomaly detection model, the problem that density clustering cannot perform real-time anomaly detection was effectively avoided, and rapid detection of the pressure state of the rear cavity after the shield tail seal was achieved.

Images