Shield tunneling machine fault prediction method based on cloud side-end cooperation

Abstract

The invention provides a shield tunneling machine fault prediction method based on cloud edge-end cooperation, and belongs to the technical field of shield tunneling, and the method comprises the following steps: S1, a sensor as end equipment collects the data of shield equipment in real time, and uploads the collected data to edge equipment; s2, after the edge device carries out data fusion, the fused data are uploaded to the cloud; s3, using the uploaded fused data to train a fault recognition model at the cloud, and then issuing the generated fault recognition model to an edge device; according to the method, data services are reasonably divided through edge cloud collaborative operation, and a part of functions are distributed to edge equipment, so that real-time data services are provided for an opposite end, real-time identification is divided to the edge, the data transmission distance is shortened, faults can be identified more quickly, and the data transmission efficiency is improved through a self-adaptive data acquisition method. The situation that too much redundant information is collected is avoided, waste of storage resources is reduced, and meanwhile more useful information can be collected.

Description

Technical Field

[0001] The present invention relates to the technical field of shield machines, and in particular to a shield machine fault prediction method based on cloud-edge-end collaboration. Background Art

[0002] Tunnel boring machines are heavy equipment for a country, with a huge structure, complex mechanical structure, changeable engineering environment, and customized discrete manufacturing. In the operation of tunnel equipment in engineering construction, due to information asymmetry and inadequate prediction capabilities, when abnormal construction conditions such as tool wear and mud cakes occur, they are not perceived in time, and construction personnel are unable to take effective measures in advance to deal with them, which ultimately leads to major problems and seriously affects the construction progress. In terms of repair and maintenance, since the status of components cannot be predicted in advance according to the status of the equipment, spare parts are not prepared in advance, resulting in untimely repairs and frequent downtime. In design and manufacturing, equipment selection and design are mostly based on experience. Many equipment have poor operating results during the trial operation phase due to problems such as the incompatibility of equipment design with geology and the lack of coordination between equipment components. The equipment needs to be remodeled, resulting in a huge waste of resources.

[0003] Since the entire shield machine system is huge, including multiple subsystems such as the cutter head system, soil bin system, articulation system, and shield tail sealing system, the operating environment of the shield machine is relatively complex. Once a failure of the shield machine occurs, it may cause casualties and generate huge economic losses.

[0004] After analyzing the data of failures in a slurry shield tunnel from September 2019 to November 2021, it can be seen that: from September 2019 to November 2021, a total of 900 failures occurred; among them, the most frequent one was "HBW1 pump counting overspeed", which occurred 8852 times; the least frequent one was "main drive pump 1 oil replenishment pressure abnormality" and other 77 failures, which occurred once each, Figure 1 These are the top 20 categories with the most failures from September 2019 to November 2021.

[0005] It can be seen that the types of failures that occur in shield machines during actual operations are diverse and uneven.

[0006] The difficulties in identifying shield machine faults are summarized as follows:

[0007] (1) Various types of faults and unbalanced data. The shield machine system is huge and contains multiple subsystems, so various types of faults may occur.

[0008] Figure 1 Among the top 20 types of faults with the highest number of occurrences, the least frequent "P2.3 pump sealing water tank liquid level high limit" also occurred 1965 times; but during this period, 77 types of faults only occurred once, which shows that there is a serious imbalance in the fault data.

[0009] (2) The relationship between sensor data and faults is complex. It is difficult to use artificial experience and simple formula derivation to characterize the complex relationship between sensor data and faults. The causes of each fault are uncertain, and there are cases where multiple faults occur at the same time. It is impossible to directly train a deep learning model to classify various faults. In addition, data imbalance reduces the performance of the classifier, and the classification ability of small samples is insufficient, resulting in a decrease in the accuracy of the classifier and weak fault recognition ability. Summary of the invention

[0010] The present invention provides a shield machine fault prediction method based on cloud-edge-end collaboration to solve the technical problem of low model accuracy in the prior art.

[0011] In order to solve the above problems, the shield machine fault prediction method based on cloud-edge-end collaboration provided by the present invention adopts the following technical solution, including the following steps:

[0012] S1, as a sensor of the end device, collects data of the shield equipment in real time and uploads the collected data to the edge device;

[0013] S2, after the edge device performs data fusion, the fused data is uploaded to the cloud;

[0014] S3, training a fault identification model using the uploaded fused data in the cloud, and then sending the generated fault identification model to the edge device;

[0015] S4, the edge device uses the fault identification model decentralized from the cloud to perform real-time identification on the data collected in real time;

[0016] S5, when the accuracy of the fault identification model provided by the cloud to the edge device is lower than the set threshold, steps S1 to S3 are repeated, the cloud retrains the fault identification model, and the latest fault identification model is transferred to the edge device.

[0017] As a further improvement, in S1, the data collection frequency is positively correlated with the data change amount. When the data change amount becomes smaller, the data collection frequency decreases; when the data change amount becomes larger, the data collection frequency increases.

[0018] As a further improvement, in S1, the data collection method includes an adaptive data collection method based on univariate linear regression, an adaptive method based on a revolving door algorithm, and an adaptive method based on data change stability.

[0019] As a further improvement, in S2, the K-Means algorithm is used to fuse the fault data, and the fused results and the sensor data collected in real time to reflect the actual situation of the shield machine are combined into a data set, which is input into the fault identification model library to complete the fault pattern recognition.

[0020] As a further improvement, in S3, the fault data fused by the K-Means algorithm is used as a label; the sensor data is used as a feature; and the fault-free data corresponding to the fault data is created to form a fault-free data set, and then the fault recognition model is trained.

[0021] As a further improvement, the features are extracted, and the extracted features include but are not limited to sensor data including the cutter disc speed, cutter disc torque, propulsion group A pressure, propulsion group A thrust, VWT guided horizontal front, and VMT guided vertical front.

[0022] As a further improvement, the fault identification model includes but is not limited to a neural network classifier TabNet, a tree classifier XGBoost and a support vector machine.

[0023] As a further improvement, the edge device includes a data storage unit, a data processing unit and a model execution unit.

[0024] The beneficial effects of the above technical solution of the present invention are as follows:

[0025] 1. Through edge-cloud collaboration, data storage, data processing and data services are jointly realized. Edge devices only store recent data. Edge devices will upload data that requires complex calculations, has a large amount of data, and will not be used by edge devices in the near future to the cloud for storage, realizing edge-cloud collaborative storage; for data processing, edge devices perform low-level data fusion, and then upload the fused data to the cloud for further processing. Edge devices mainly perform some real-time calculations to reduce the burden on the cloud; data services are reasonably divided and part of the functions are delegated to the edge, so as to provide real-time data services to the other end, and real-time identification is divided to the edge, which shortens the data transmission distance and can identify faults more quickly.

[0026] 2. The present invention adopts an adaptive data collection method, that is, the data collection frequency is positively correlated with the data change amount. When the data change amount becomes smaller, the data collection frequency is reduced to avoid collecting too much redundant information and reduce the waste of storage resources; when the data change amount becomes larger, the data collection frequency is increased to collect more useful information. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] By reading the following detailed description with reference to the accompanying drawings, the above and other objects, features and advantages of the exemplary embodiments of the present invention will become readily understood. In the accompanying drawings, several embodiments of the present invention are shown in an exemplary and non-limiting manner, and the same or corresponding reference numerals represent the same or corresponding parts, wherein:

[0028] Figure 1It is a statistical chart of the top 20 types of faults with the most faults occurring within a certain period of time;

[0029] Figure 2 It is a trend chart of the change of the cutter head power and wear pressure over time of the shield machine fault prediction method based on cloud-edge-end collaboration of the present invention;

[0030] Figure 3 The variation trend of the motor torque and wear pressure over time of the shield machine fault prediction method based on cloud-edge-end collaboration of the present invention;

[0031] Figure 4 This is a graph showing the prediction results of the cutter disc torque and cutter disc wear pressure of the shield machine fault prediction method based on cloud-edge-end collaboration of the present invention. DETAILED DESCRIPTION

[0032] 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. Those skilled in the art should know that the embodiments described below are part of the embodiments of the present disclosure, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0033] The present invention adopts a three-layer edge computing architecture of end-edge-cloud. Each shield machine is arranged to work in different regions, which are geographically thousands of kilometers apart. Sensors are used to monitor the real-time status of each subsystem of the shield machine.

[0034] The slave acquisition terminal uses serial ports and digital-to-analog conversion to collect real-time data from various sensors, and uses LoRa wireless LAN for communication. The main acquisition terminal uses PLC communication to collect operating status parameters and working condition data collected from the slave acquisition terminal. The main acquisition terminal transmits data to the big data cloud platform through 5G base station communication.

[0035] The cloud computing architecture has only two layers: cloud servers and data producers. The data producers are responsible for the data source layer, and other data governance layers are completed by the cloud.

[0036] The end, edge, and cloud collaborate to complete data governance. As the data producer, the end device is responsible for the data source layer and delivers data to the edge device. The edge and cloud collaborate, and even the end, edge, and cloud collaborate to complete other data governance layers. The division of labor is different.

[0037] In order to solve the problem of multiple and unbalanced shield fault data, the present invention fuses the basic data and then uploads it to the cloud. On the one hand, it reduces irrelevant calculations, and on the other hand, it can reduce the uncertainty and noise of a single data source and improve the overall quality and reliability of the data.

[0038] In terms of data collection, the present invention adopts an adaptive data collection method to dynamically adjust the data collection frequency to avoid failing to collect more useful information when the data changes dramatically, or collecting too much redundant information when the data changes slowly, resulting in a waste of storage resources.

[0039] After introducing the basic principle of the present invention, various non-limiting embodiments of the present invention are described in detail below. The number of any element in the drawings is for illustration and not for limitation, and any naming is only for distinction and does not have any limiting meaning.

[0040] The principle and spirit of the present invention are explained in detail below with reference to several representative embodiments of the present invention.

[0041] Embodiment 1 of the shield machine fault prediction method based on cloud-edge-end collaboration provided by the present invention:

[0042] The following steps are involved:

[0043] S1, as the sensor of the end device, collects data of the shield equipment in real time and uploads the collected data to the edge device;

[0044] S2, after the edge device performs data fusion, the fused data is uploaded to the cloud;

[0045] S3, training a fault identification model using the uploaded fused data in the cloud, and then sending the generated fault identification model to the edge device;

[0046] S4, the edge device uses the fault identification model decentralized from the cloud to perform real-time identification on the data collected in real time;

[0047] S5, when the accuracy of the fault identification model provided by the cloud to the edge device is lower than the set threshold, steps S1 to S3 are repeated, the cloud retrains the fault identification model, and the latest fault identification model is transferred to the edge device.

[0048] In S1, the edge device includes a data storage unit, a data processing unit, and a model execution unit. The data processing unit is used for data fusion and preprocessing, and the model execution unit is used to identify faults using the model sent from the cloud. The storage capacity of edge devices is relatively weak, so they will store more recent data. The main function of edge devices is to perform real-time calculations. Data storage and data processing are not the primary tasks of edge devices, but only to reduce the burden on the cloud.

[0049] In S1, an adaptive data collection method is used during data collection to ensure that more useful information is collected while reducing the situation where too much redundant information is collected.

[0050] The collected data mainly include equipment operation status parameters and working condition data. Equipment operation status parameters mainly refer to the data taken from sensors and controllers that can reflect the equipment operation conditions and health status, that is, monitoring data. Its characteristics are high sampling frequency and the most complex variables collected;

[0051] The operating condition data of the equipment mainly refers to the setting information of the equipment's load, speed, operating mode and other working conditions. Such data can often be obtained from the controller. Only by comparing and analyzing the equipment status parameters under the same working conditions can the changes in the equipment's health status and performance be reflected.

[0052] There are many devices in the field of shield construction. The status of these devices requires sensors to collect data, and the collected data is stored in the database for researchers to conduct in-depth research. Usually, sensors collect data at equal time intervals, but the law of data change is not taken into account.

[0053] Simply reducing the collection frequency will result in not collecting more useful information when the data changes dramatically.

[0054] Simply increasing the collection frequency will result in collecting too much redundant information when the data changes slowly, resulting in a waste of storage resources.

[0055] In this embodiment, an adaptive data collection method is adopted. When the data change amount is not large, the collection time interval is increased, that is, the collection frequency is reduced. When the data change amount becomes large, the collection time interval is dynamically reduced, that is, the collection frequency is increased. In this way, the collected data can be of guaranteed quality and quantity.

[0056] The core idea of ​​the adaptive acquisition algorithm is to adjust the acquisition interval according to the degree of data change. When the data changes slowly, the acquisition interval will be increased, and when the data changes dramatically, the acquisition interval will be reduced.

[0057] The adaptive data collection methods include an adaptive data collection method based on univariate linear regression, an adaptive method based on a revolving door algorithm, and an adaptive method based on the stability of data changes.

[0058] The adaptive data collection method based on univariate linear regression is to obtain the univariate linear regression line of the collected data within the current period of time, make a judgment based on the fluctuation ratio between the collected data and the regression line, so as to realize the dynamic change of the collection frequency and select the appropriate collection time interval to collect the status data.

[0059] For example, data is collected once every 1 second, and the 10 most recent data points are used to perform univariate linear regression modeling to obtain the regression line. The vertical distance (residual) between each data point and the regression line is calculated, and the root mean square error (RMSE) of all residuals is calculated and used as an evaluation indicator for the data change. If the root mean square error is less than the threshold value of 0.1, it means that the data changes slowly, and the collection frequency can be reduced to once every 5 seconds. If the root mean square error is greater than the threshold value of 1.0, it means that the data changes dramatically, then the collection frequency should be increased and adjusted to once every 0.5 seconds.

[0060] The adaptive method based on the revolving door algorithm is a relatively fast-responding linear fitting algorithm with the advantages of small computational complexity and the ability to track changes in process data. The essence of this algorithm is to replace a series of continuous data points with a straight line determined by the starting point and the end point, thereby compressing massive data into data segments with a linear trend and ensuring that the data compression error is within a certain compression threshold.

[0061] For example, collect data every 1 second, set a fixed-size sliding window with 10 data points, calculate the difference between the maximum and minimum values ​​of the data in the window, and use the difference between the maximum and minimum values ​​as the evaluation indicator of data change. If the data change is less than the threshold value of 0.1, it means that the data changes slowly, and the collection frequency can be reduced to every 5 seconds; if the data change is greater than the threshold value of 1.0, it means that the data changes dramatically, and the collection frequency needs to be increased to every 0.5 seconds.

[0062] The adaptive method based on data change stability is to use the data change stability index to quantitatively evaluate the data fluctuation according to the change law of the collected parameter data, judge the data fluctuation degree of the collected parameters, and dynamically adjust the time interval of data collection. When data collection is just beginning, it is not possible to apply data change stability to accurately judge the data change situation. The initial data is collected with the minimum time interval. When a certain amount of data is collected and the minimum window for calculating the data change stability is reached, the data change stability is calculated, and then the collection time interval of the adaptive data collection is adjusted to obtain the time interval for the next collection. The mobile data window is used to evaluate the change of the next segment of data, and this cycle is repeated until the collection is completed.

[0063] For example, data is collected once every 1 second, and the standard deviation is calculated using the most recent 10 data points. The standard deviation is used as an evaluation indicator for the amount of data change. If the standard deviation is less than the threshold of 0.1, it means that the data changes slowly, and the collection frequency can be appropriately reduced to once every 5 seconds. If the standard deviation is greater than the threshold of 1.0, it means that the data changes dramatically, and the collection frequency needs to be increased to once every 0.5 seconds.

[0064] It should be noted that different sensors have different parameters such as acquisition frequency, and the above data are only for example.

[0065] In S2, before the fault data is fused, data preprocessing is performed first, and the influence of the data dimension on the result is eliminated through normalization. The normalization formula is as follows:

[0066]

[0067] Where X' is the normalized result, X is the original data, and X mean and X std represent the mean and variance of the original data respectively.

[0068] In S2, the K-Means algorithm is used to fuse the fault data, and the fused results and the sensor data collected in real time to reflect the actual situation of the shield machine are combined into a data set, which is input into the fault identification model library to complete the fault pattern identification. The role of the K-Means algorithm is to fuse the fault data and fuse the unbalanced and multi-type fault data into a new fault category.

[0069] The processing flow of the K-Means algorithm is as follows: first, randomly select k data objects, each of which represents a cluster center, that is, select k initial centers; for each remaining object, assign it to the cluster corresponding to the cluster center that is most similar to it according to its similarity (distance) to each cluster center; then recalculate the average value of all objects in each cluster as the new cluster center. Repeat the above process until the criterion function converges, that is, the cluster center does not change significantly.

[0070] The fault data fused by the K-Means algorithm is used as a label; the sensor data is used as a feature; the fault-free data corresponding to the fault data is created to form a fault-free data set, and then the fault recognition model is trained. Before training, the features are extracted first, and the extracted features include but are not limited to the sensor data of the cutter head speed, cutter head torque, propulsion group A pressure, propulsion group A thrust, VWT guide horizontal front, and VMT guide vertical front.

[0071] We regard failure mode prediction as a classification problem.

[0072] In this classification problem, the data detected by each sensor point constitutes features, and the type of fault that occurred (or no fault occurred) is used as a label. The fault data fed back from the shield construction site has the characteristics of a wide variety of faults, low frequency of occurrence, and a small number of partial faults, which cannot be directly used as a classifier label. Therefore, the K-Means clustering algorithm is used to cluster the fault data collected on site and divide them into k types of faults as labels for the fault mode classifier.

[0073] In this embodiment, three types of classifiers, namely, neural network classifier TabNet, tree classifier XGBoost and support vector machine, are used to construct a fault identification model. Among them, XGBoost has the best performance of up to 96.31%, while support vector machine and neural network classifier Tabnet can also reach 84.16% and 60.30% respectively.

[0074] In this embodiment, TDevices is used to construct a data processing architecture. Based on the tunneling parameters of the shield machine, we find that the tunneling parameters have the following characteristics:

[0075] 1: All equipment data such as machines, devices, sensors, etc. are time-series and have timestamps;

[0076] 2: The data is structured;

[0077] 3: Data is rarely updated or deleted;

[0078] 4: No need for traditional database transaction processing;

[0079] 6: Users are concerned about the trend over a period of time rather than the value at a specific point in time;

[0080] 7: Data has a retention period;

[0081] 8: Data query and analysis must be based on time periods and geographical areas;

[0082] 9: In addition to storage queries, various statistical and real-time computing operations are often required;

[0083] 10: The amount of data is huge, and the data collected in one day can exceed hundreds of thousands.

[0084] These seem simple, but due to the large number of data records, real-time data writing becomes a bottleneck, query analysis is extremely slow, and becomes a new technical challenge.

[0085] TDengi ne adopts a combination of virtual nodes, data nodes and management nodes to realize node operation status monitoring and maintenance, and uses asynchronous replication to synchronize data to virtual nodes, ensuring that data is copied on multiple physical nodes. TDengi ne supports horizontal expansion, so that any hardware failure or software error in any single or multiple servers will not affect the availability and reliability of the system.

[0086] The following is an implementation example of a shield machine fault prediction method based on cloud-edge-end collaboration.

[0087] The data set used is the 350th to 375th ring of a tunnel (data for about 14 days). The mean absolute error (Mean Abso l ute Error) is used as the loss function, and the L2 norm is used as the penalty term for loss value testing.

[0088] First, select the point to be predicted. When predicting, regularize the historical data of 25 rings and convert it to statistics every minute (or every 30 seconds, etc.). If there are multiple data records in one minute, the data of that minute will be recorded as the average of all data records in that minute. And normalize the data to facilitate subsequent learning and use.

[0089] After that, we learn the historical data, and learn 25 rings (or less or more) of historical data. In terms of prediction, fine-grained prediction is used, that is, the time to be predicted is set, which can be set to 15 minutes, 30 minutes, 60 minutes, etc. The predicted data is compared with the actual value, and all its trend graphs are drawn.

[0090] from Figure 2 and Figure 3 It can be seen that the predicted trends or sizes are roughly the same.

[0091] The loss value, Figure 3 It represents the loss value of cutter disc torque and cutter disc wear pressure.

[0092] As can be seen from the curve, their initial loss value quickly converged to 10 in the 100th round. -3 level.

[0093] Although this specification has shown and described a plurality of embodiments of the present invention, it is obvious to those skilled in the art that such embodiments are provided by way of example only. Those skilled in the art will think of many changes, modifications and alternatives without departing from the thought and spirit of the present invention. It should be understood that in the process of practicing the present invention, various alternatives to the embodiments of the present invention described herein may be adopted. The attached claims are intended to define the scope of protection of the present invention, and therefore cover the modular composition, equivalent or alternatives within the scope of protection of these claims.

Claims

1. A shield machine fault prediction method based on cloud-edge-end collaboration, characterized in that: The following steps are involved: S1, as a sensor of the end device, collects data of the shield equipment in real time and uploads the collected data to the edge device; S2, after the edge device performs data fusion, the fused data is uploaded to the cloud; S3, training a fault identification model using the uploaded fused data in the cloud, and then sending the generated fault identification model to the edge device; S4, the edge device uses the fault identification model deployed from the cloud to perform real-time identification on the data collected by the sensor in real time; S5, when the accuracy of the fault identification model provided by the cloud to the edge device is lower than the set threshold, steps S1 to S3 are repeated, the cloud retrains the fault identification model, and the latest fault identification model is transferred to the edge device.

2. The shield machine fault prediction method based on cloud-edge-end collaboration according to claim 1 is characterized in that: In S1, the data collection frequency is positively correlated with the data change amount. When the data change amount becomes smaller, the data collection frequency decreases; when the data change amount becomes larger, the data collection frequency increases.

3. The shield machine fault prediction method based on cloud-edge-end collaboration according to claim 2 is characterized in that: In S1, the data collection methods include an adaptive data collection method based on univariate linear regression, an adaptive method based on a revolving door algorithm, and an adaptive method based on data change stability.

4. The shield machine fault prediction method based on cloud-edge-end collaboration according to claim 1 is characterized in that: In S2, the K-Means algorithm is used to fuse the fault data. The fused results and the sensor data collected in real time to reflect the actual situation of the shield machine are combined into a data set, which is input into the fault identification model library to complete the fault pattern recognition.

5. The shield machine fault prediction method based on cloud-edge-end collaboration according to claim 4 is characterized in that: In S3, the fault data fused by the K-Means algorithm is used as the label; the sensor data is used as the feature; Create non-fault data corresponding to the fault data to form a non-fault data set, and then train the fault recognition model.

6. The shield machine fault prediction method based on cloud-edge-end collaboration according to claim 5 is characterized in that: The features are extracted, and the extracted features include but are not limited to sensor data including cutter head rotation speed, cutter head torque, propulsion group A pressure, propulsion group A thrust, VWT guided horizontal front, and VMT guided vertical front.

7. The shield machine fault prediction method based on cloud-edge-end collaboration according to any one of claims 4 to 6 is characterized in that: The fault identification model includes but is not limited to a neural network classifier TabNet, a tree classifier XGBoost and a support vector machine.

8. The shield machine fault prediction method based on cloud-edge-end collaboration according to claim 1 is characterized in that: The edge device includes a data storage unit, a data processing unit and a model execution unit.

Images