Shield tunneling machine fault monitoring method and system based on slow feature analysis

Abstract

The invention discloses a shield tunneling machine fault monitoring method and system based on slow feature analysis, and the method comprises the steps: firstly obtaining the operation data of a shield tunneling machine under a normal working condition, extracting a key variable, carrying out the slow feature analysis of the key variable, and constructing a monitoring statistical magnitude; the monitoring statistics comprise steady-state monitoring statistics describing steady-state changes in the operation process of the shield tunneling machine and dynamic monitoring statistics describing dynamic change distribution of operation data of the shield tunneling machine, and the control limit of each monitoring statistics is calculated; secondly, real-time operation data of the shield tunneling machine are obtained, monitoring statistics are calculated based on slow feature analysis, and fault alarm is carried out according to the monitoring statistics and the control limit of the monitoring statistics. Fault monitoring can be performed in combination with internal and external conditions of the shield tunneling machine, and false alarms caused by external environment factor changes of the shield tunneling machine are reduced.

Description

Technical Field

[0001] The invention relates to the technical field of shield machine fault monitoring, in particular to a shield machine fault monitoring method and system based on slow feature analysis. Background Art

[0002] As an important tool in the development of underground space, shield machines are widely used due to their low environmental impact, fast excavation speed, safe and economical construction. With the increasing working intensity of shield machines and the need to adapt to complex and changing environments, shield machines are prone to different types of failures, such as propulsion system failure, cutter head failure, main bearing failure, hydraulic system failure, etc. In order to monitor the equipment, modern tunnel shield machines usually integrate a large number of sensors to record a series of operating parameters such as stress, current, flow and gas pressure, which is also the basis for establishing a fault monitoring model for shield machines.

[0003] With the development of computer technology, fault monitoring technology is gradually shifting from relying on expert knowledge to determine faults to quantitative analysis methods that use computers to analyze faults. However, the application of shield machine fault monitoring in the field is still in its infancy. Shield machines are large-scale underground operating equipment and are inevitably affected by the operating environment such as geology. Faced with different soil layers, the setting parameters of shield machines are different. At present, when monitoring the faults of shield machines, the changes in the operating data of shield machines caused by geological changes are not taken into account, and it is impossible to distinguish between abnormal conditions caused by external and internal factors of the shield machine, which causes false alarms due to changes in external environmental factors of the shield machine. Summary of the invention

[0004] Purpose of the invention: The purpose of the present invention is to provide a shield machine fault monitoring method and system based on slow feature analysis, which can perform fault monitoring in combination with the internal and external conditions of the shield machine, thereby reducing false alarms caused by changes in external environmental factors of the shield machine.

[0005] Technical solution: The shield machine fault monitoring method based on slow feature analysis of the present invention comprises the following steps: (1) Obtain the operating data of the shield machine under normal working conditions, preprocess the operating data, and extract key variables in the shield machine operating data; (2) constructing monitoring statistics based on the key variables and slow feature analysis, and calculating the control limit of each monitoring statistic; the monitoring statistics include steady-state monitoring statistics describing steady-state changes during the operation of the shield machine and dynamic monitoring statistics describing the dynamic change distribution of the shield machine operation data; (3) Obtain the real-time operation data of the shield machine and pre-process the real-time operation data; (4) Calculate monitoring statistics based on slow feature analysis according to the preprocessed real-time operation data; (5) Providing fault alarms based on the monitoring statistics and their control limits.

[0006] Furthermore, in step (2), constructing monitoring statistics based on the key variables and slow feature analysis includes: (21) performing slow feature analysis on the key variables to obtain slow features of the key variables; (22) Get the slowest ones in descending order J A slow feature; (23) For J slow features, calculate the slow features that are faster than all key variables as the fast-changing feature space, and the remaining slow features as the slow-changing feature space; (24) A steady-state monitoring statistic and a dynamic monitoring statistic are constructed in the fast-varying feature space and the slow-varying feature space respectively.

[0007] Furthermore, in step (24), the steady-state monitoring statistic , and dynamic monitoring statistics , The calculation method is: , , , , where the subscript d and e represent the slowly varying feature space and the rapidly varying feature space respectively, is the feature in the slowly varying feature space, for The first derivative of is the feature in the fast-changing feature space, for The first derivative of .

[0008] Furthermore, in step (2), the control limits of the monitoring statistics are calculated by kernel density estimation.

[0009] Furthermore, in step (5), performing a fault alarm according to the monitoring statistic and its control limit includes: If both the steady-state monitoring statistics and the dynamic monitoring statistics do not exceed the corresponding control limits, no alarm information will be issued; If the steady-state monitoring statistic exceeds the corresponding control limit, and the dynamic monitoring statistic does not exceed the corresponding control limit, an alarm message of structural abnormality is issued; If the steady-state monitoring statistic does not exceed the corresponding control limit, and the dynamic monitoring statistic exceeds the corresponding control limit, an alarm message to keep an eye on is issued; If both the steady-state monitoring statistics and the dynamic monitoring statistics exceed the corresponding control limits, an alarm message for shutdown adjustment will be issued.

[0010] Furthermore, in step (1), preprocessing the running data includes using The criterion method eliminates abnormal data, standard deviation Calculated according to Bessel's formula.

[0011] Furthermore, the key variables extracted from the shield machine operation data in step (1) include: (11) Select several operating data as expert knowledge variables based on expert knowledge, and construct an XGBoost regression tree model for each expert knowledge variable for regression prediction to obtain feature importance; (12) calculating the correlation between the preprocessed operating data and the expert knowledge variable according to the feature importance, sorting the correlation from high to low, and retaining the operating data with a correlation greater than a threshold; (13) De-dimensionalize and sum the operating data retained in step (12) to obtain the final key variables.

[0012] The shield machine fault monitoring system based on slow feature analysis of the present invention comprises: A key variable acquisition unit is used to acquire the operating data of the shield machine under normal working conditions, pre-process the operating data, and extract key variables from the shield machine operating data; A monitoring statistic construction unit is used to construct monitoring statistics based on the key variables and slow feature analysis, and calculate the control limit of each monitoring statistic; the monitoring statistics include steady-state monitoring statistics describing steady-state changes during the operation of the shield machine and dynamic monitoring statistics describing the dynamic change distribution of shield machine operation data; A real-time operation data acquisition unit is used to acquire the real-time operation data of the shield machine and pre-process the real-time operation data; A monitoring statistic calculation unit, used for calculating monitoring statistics based on slow feature analysis according to the pre-processed real-time operation data; A fault alarm unit is used to perform a fault alarm according to the monitoring statistic and its control limit.

[0013] The electronic device described in the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded into the processor, the shield machine fault monitoring method based on slow feature analysis is implemented.

[0014] The computer-readable storage medium of the present invention stores a computer program, and when the computer program is executed by a processor, the shield machine fault monitoring method based on slow feature analysis is implemented.

[0015] Beneficial effect: Compared with the prior art, the advantages of the present invention are: the present invention utilizes the existing shield machine operation data, applies the proposed fault monitoring method to the real data, can analyze the steady-state and dynamic conditions of the shield machine operation, and performs fault monitoring in combination with the internal and external conditions of the shield machine, thereby reducing false alarms caused by changes in the external environmental factors of the shield machine. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a flow chart of a shield machine fault monitoring method according to an embodiment of the present invention.

[0017] Figure 2 Schematic diagram of a fault alarm strategy according to an embodiment of the present invention. DETAILED DESCRIPTION

[0018] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0019] like Figure 1 As shown, the shield machine fault monitoring method based on slow feature analysis includes the following steps.

[0020] Step S1, obtaining shield machine operation data under normal working conditions, preprocessing the shield machine operation data, and extracting key variables in the shield machine operation data.

[0021] Step S11, as a major equipment, the shield machine uses many sensors to monitor its various excavation parameters in real time to ensure safe and stable construction. Various sensors deployed inside the shield machine are used to collect various excavation parameters during the shield construction process, and the shield machine operation data under normal working conditions are obtained from the shield machine control system. The shield machine operation data is preprocessed to remove abnormal data that is not conducive to offline modeling. This includes removing data that falsely reports shutdown when the shield machine is working and removing data that reaches steady state from the beginning of excavation. At present, the shield machine operation mechanism mainly relies on setting thresholds to simply determine whether the shield machine is faulty. This method may lead to false alarms. Once the system detects the alarm information, it will command the shield machine to pause. From the perspective of data application, it is necessary to eliminate such false shutdown data. The data of the shield machine reaching steady state at the beginning of each ring of excavation refers to the data generated when the cutter head speed is not stable. From the perspective of data application, this type of data is not conducive to fault monitoring, so it also needs to be eliminated.

[0022] use The criterion method removes abnormal data. Among them, the standard deviation According to Bessel's formula: ; in For the k Operation data, K is the number of samples; if a shield machine operation data The residual error Satisfy the absolute value , then the data is considered to be abnormal data and should be eliminated.

[0023] After the above preprocessing, the downtime data and noise in the shield machine operation data can be reduced as much as possible, and the original shield machine operation data After preprocessing, the preprocessed shield machine working data is obtained .

[0024] Step S12: As a large underground operation equipment, the shield machine itself is equipped with a large number of sensors, and the collected data is redundant, so it is necessary to further filter the variables of the pre-processed data. According to expert knowledge, the thrust, torque, penetration and propulsion speed of the shield machine are very important for the operation of the shield machine. The four variables of penetration, torque, propulsion speed and thrust are selected as expert knowledge variables, and an XGBoost regression tree model is constructed for each expert knowledge variable for regression prediction. , n is the total number of sensors, vector x Represents the measurement values ​​of different sensors at different times over a period of time.

[0025] The regression prediction of variables using XGBoost is based on the tree model. When constructing the tree model, each level greedily selects a feature segmentation point as a leaf node to maximize the overall gain of the entire tree. Here, the feature is the variable collected by different sensors. When each feature is divided, the greater its average gain, the more important the feature is. In the segmentation process, the weight of each leaf node can be expressed as ,in, , , training error Indicates the target value and predicted values The difference is Represents a leaf node a The sum of the first-order gradients of all samples on Represents a leaf node a The sum of the second-order gradients of all samples.

[0026] Traverse all features, for each feature, traverse all possible split points, and calculate the gain of splitting ( Gain ), select the feature and split point with the maximum gain from all splits of all features. The gain formula is: ; in, represents the total weight before splitting, and Respectively represent the total weights of the left child node and the right child node after the split.

[0027] The gain of each split point is expressed as the difference between the loss after the split and the loss before the split. For each feature, XGBoost evaluates its importance in the feature set according to the importance metric and sets the number of feature splits. FScore To measure feature importance: ; in, X is the set of leaf nodes classified by the desired features. FScore It indicates the number of times each feature is selected as a split feature when building all trees. It is a simple counting indicator to measure the importance of the feature.

[0028] Step S13, construct a regression tree model by the above method, take the four expert knowledge variables as regression prediction targets to train four XGBoost regression tree models in turn, perform feature importance analysis on each regression tree model, record the correlation degree between each variable in the preprocessed shield machine operation data and each expert knowledge variable, and use FScore Indicators are used to measure, thus obtaining 4 groups of variable correlation tables.

[0029] Step S14, de-dimensionalizing (normalizing) and summing the correlation degree of each variable in the four groups of expert knowledge variable correlation degree tables to obtain a total correlation degree score of each variable.

[0030] Step S15: sort the variables according to the total correlation degree, perform cumulative calculation on the sorted variables according to their importance proportion, and select the variables with the top 95% cumulative proportion. m variables, select this m variables as features used in subsequent work, that is, the currently selected m This variable is more concise and contains most of the information of shield machine operation.

[0031] Step S2, constructing offline monitoring statistics based on slow feature analysis according to the key variables extracted in step S1, and calculating the control limit of each offline monitoring statistic.

[0032] Step S21, performing slow feature analysis on the key variables obtained in step S1, mining the slowly changing information in the shield machine operation data, and dividing the feature space into a slowly changing feature space and a quickly changing feature space according to the slowness. Step S21 includes: Step S211, standardize the key variables in the shield machine operation data. The specific method is as follows: ; in, To standardize the data, is the standardized data, is the data average, is the standard deviation of the data.

[0033] The standardized operating data is , m Indicates the number of key variables.

[0034] Step S212, select L Shield machine timing working data after ring standardization , which is used to construct the subsequent offline monitoring statistics. m Represents the number of key variables. For a certain variable, its sensor observation data can be expressed as , Indicates l The sensor at the time t The measured value at time T means that the sensor collected data for T time moments in the L ring.

[0035] Step S213, setting the dimension of the slow feature J , calculate the target weight matrix according to the slow feature solution method W . Target weight matrix W The derivation process is as follows: Given m Dimensional input signal ,in Representing time, the Slow Feature Analysis Algorithm (SFA) seeks a transformation function , so that the output of the function Change as slowly as possible, , corresponding to a slow feature.

[0036] Slow feature SFs can be expressed as: ,in, is the weight matrix.

[0037] Will The covariance matrix is SVD Decomposition, that is ,in U is a unitary matrix, Λ is a diagonal matrix; let ,in To express m dimensional input signal matrix, then the expanded signal , SFs can be written , thus: ; ; The symbol Expressing the desire t The mean of the variable over the time range,I is the identity matrix, P is an orthogonal matrix. The final optimization problem of SFA is transformed into solving an orthogonal matrix P Make the following as small as possible: ,in, It is used to measure the speed of change of slow feature signals. Indicates i The first derivative of the slow feature, represents the first-order derivative of the extended signal, P i Represents an orthogonal matrix P The i By performing principal component analysis to solve the above optimization problem, The matrix is ​​used for principal component analysis, specifically: ; In the formula, P is the eigenvector matrix, , , which is the first J characteristic roots, and finally obtain the weight matrix: ,in P Indicates the first J The matrix consists of the eigenvectors corresponding to the eigenroots, where U is the unitary matrix in the above SVD decomposition process and Λ is a diagonal matrix.

[0038] Step S214, according to the weight matrix W , calculate the slow feature SF of key variables in shield machine operation data s , get the slowest to slowest J SFs. Among them, the slow feature , , is the weight matrix.

[0039] Step S215, SFs are divided into a slow-changing feature space and a fast-changing feature space. First, the SFs that are faster than all input variables are calculated by the following formula: ; J After being divided into two parts, It is also divided into two parts: ;in, J represents all slow features, Indicates the number of slower features among all slow features, Represents the number of faster features among all slow features, is the feature transformation matrix of the slower feature, is the feature transformation matrix of the faster feature.

[0040] Step S22: constructing offline monitoring statistics in the slowly varying feature space and the rapidly varying feature space respectively, and calculating the control limit of each offline monitoring statistic.

[0041] In this embodiment, the offline monitoring statistics are respectively the offline steady-state monitoring statistics describing the steady-state changes during the operation of the shield machine: , And offline dynamic monitoring statistics describing the dynamic distribution of shield machine operation data , : , , , ; , ; , ; represents the transpose of the L-ring data matrix in step S212, p and q They represent the feature dimensions in the slowly varying feature space and the fast varying feature space respectively. For the slower features derived, For the faster features derived, represents the first derivative of the slower feature in the feature space, Represents the first derivative of the faster features in feature space.

[0042] Statistics , It is a steady-state monitoring indicator that describes the steady-state changes during the operation of the shield machine. Its control limit is calculated by kernel density estimation. , Exceeding the limit indicates that the steady state of the operating state has changed, that is, the shield machine itself has an abnormal situation. , To dynamically monitor the indicators and describe the distribution of the dynamic changes in the shield machine operation data, the control limits are also calculated by kernel density estimation. , This indicates that potential abnormality in the dynamic process has been monitored, that is, the external environment of the shield machine has changed.

[0043] Step S3, obtaining online real-time operation data of the shield machine , and preprocess the real-time running data. The mean and variance used in the preprocessing process are the mean of the offline modeling data. μ and standard deviation σ The processed data is .

[0044] Step S4: based on the pre-processed real-time operation data , calculate online monitoring statistics based on slow feature analysis , , and The monitoring statistics are calculated by using the , .

[0045] Step S41, projecting the online data into the slower feature space and the faster feature space respectively, and obtaining the corresponding slower and faster features.

[0046] ; .

[0047] Step S42, calculating steady-state online monitoring statistics.

[0048] ; .

[0049] Step S43, calculating the dynamic online monitoring statistics through the first-order derivatives of the slower feature and the faster feature.

[0050] ; .

[0051] Step S5, providing a fault alarm strategy based on the online monitoring statistics and the control limits of each offline monitoring statistic calculated in step S2.

[0052] Fault alarm strategy such as Figure 2 As shown in the figure, different alarm messages are issued according to whether the four online monitoring statistics exceed the limit, which can be divided into four specific situations: 1) If the four online monitoring statistics , , and If the corresponding control limits are not exceeded, the shield machine is considered to be operating normally and no alarm information is issued; 2) When online dynamic monitoring statistics and None of them exceeded the corresponding control limits, and the online steady-state monitoring statistics or If the corresponding control limit is exceeded, it is considered that the shield machine itself has a structural failure and an alarm message of structural abnormality is issued; 3) When online dynamic monitoring statistics or If the corresponding control limit is exceeded, it means that the shield machine operating environment has changed suddenly. At this time, if the online steady-state monitoring statistics and If the corresponding control limits are not exceeded, it means that the current operating condition of the shield machine can adapt to the changes in the current environment. However, considering that dynamic changes may lead to abnormalities in the steady state of operation, it is necessary to keep a continuous eye on the subsequent operating status of the shield machine and issue an alarm message to keep paying attention. 4) When online dynamic monitoring statistics or exceeds the corresponding control limit, and the online steady-state monitoring statistics or It also exceeds the limit, indicating that the current operating conditions of the shield machine cannot adapt to the changes in the current environment. The shield machine should be shut down to adjust the working parameters and issue an alarm message for shutdown and adjustment.

[0053] The shield machine fault monitoring system based on slow feature analysis of the present invention comprises: A key variable acquisition unit is used to acquire the operating data of the shield machine under normal working conditions, pre-process the operating data, and extract key variables from the shield machine operating data; A monitoring statistic construction unit is used to construct monitoring statistics based on the key variables and slow feature analysis, and calculate the control limit of each monitoring statistic; the monitoring statistics include steady-state monitoring statistics describing steady-state changes during the operation of the shield machine and dynamic monitoring statistics describing the dynamic change distribution of shield machine operation data; A real-time operation data acquisition unit is used to acquire the real-time operation data of the shield machine and pre-process the real-time operation data; A monitoring statistic calculation unit, used for calculating monitoring statistics based on slow feature analysis according to the pre-processed real-time operation data; A fault alarm unit is used to perform a fault alarm according to the monitoring statistic and its control limit.

[0054] The electronic device described in the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded into the processor, the shield machine fault monitoring method based on slow feature analysis is implemented.

[0055] The computer-readable storage medium of the present invention stores a computer program, and when the computer program is executed by a processor, the shield machine fault monitoring method based on slow feature analysis is implemented.

Claims

1. A shield machine fault monitoring method based on slow feature analysis, characterized in that: The steps include: (1) Obtain the operating data of the shield machine under normal working conditions, preprocess the operating data, and extract key variables in the shield machine operating data; (2) constructing monitoring statistics based on the key variables and slow feature analysis, and calculating the control limit of each monitoring statistic; the monitoring statistics include steady-state monitoring statistics describing steady-state changes during the operation of the shield machine and dynamic monitoring statistics describing the dynamic change distribution of the shield machine operation data; (3) Obtain the real-time operation data of the shield machine and pre-process the real-time operation data; (4) Calculate monitoring statistics based on slow feature analysis according to the preprocessed real-time operation data; (5) Providing fault alarms based on the monitoring statistics and their control limits.

2. The shield machine fault monitoring method based on slow feature analysis according to claim 1 is characterized in that: In step (2), based on the key variables, constructing monitoring statistics based on slow feature analysis includes: (21) performing slow feature analysis on the key variables to obtain slow features of the key variables; (22) Get the slowest ones in descending order J A slow feature; (23) For J slow features, calculate the slow features that are faster than all key variables as the fast-changing feature space, and the remaining slow features as the slow-changing feature space; (24) A steady-state monitoring statistic and a dynamic monitoring statistic are constructed in the fast-varying feature space and the slow-varying feature space respectively.

3. The shield machine fault monitoring method based on slow feature analysis according to claim 2 is characterized in that: In step (24), the steady-state monitoring statistic , and dynamic monitoring statistics , The calculation method is: , , , , where the subscript d and e represent the slowly varying feature space and the rapidly varying feature space respectively, is the feature in the slowly varying feature space, for The first derivative of is the feature in the fast-changing feature space, for The first derivative of .

4. The shield machine fault monitoring method based on slow feature analysis according to claim 1 is characterized in that: In step (2), the control limits of the monitoring statistic are calculated using kernel density estimation.

5. The shield machine fault monitoring method based on slow feature analysis according to claim 1 is characterized in that: In step (5), the fault alarm is performed according to the monitoring statistic and its control limit, including: If both the steady-state monitoring statistics and the dynamic monitoring statistics do not exceed the corresponding control limits, no alarm information will be issued; If the steady-state monitoring statistic exceeds the corresponding control limit, and the dynamic monitoring statistic does not exceed the corresponding control limit, an alarm message of structural abnormality is issued; If the steady-state monitoring statistic does not exceed the corresponding control limit, and the dynamic monitoring statistic exceeds the corresponding control limit, an alarm message to keep an eye on is issued; If both the steady-state monitoring statistics and the dynamic monitoring statistics exceed the corresponding control limits, an alarm message for shutdown adjustment will be issued.

6. The shield machine fault monitoring method based on slow feature analysis according to claim 1 is characterized in that: In step (1), the preprocessing of the running data includes using The criterion method eliminates abnormal data, standard deviation Calculated according to Bessel's formula.

7. The shield machine fault monitoring method based on slow feature analysis according to claim 1 is characterized in that: The key variables extracted from the shield machine operation data in step (1) include: (11) Select several operating data as expert knowledge variables based on expert knowledge, and construct an XGBoost regression tree model for each expert knowledge variable for regression prediction to obtain feature importance; (12) calculating the correlation between the preprocessed operating data and the expert knowledge variable according to the feature importance, sorting the correlation from high to low, and retaining the operating data with a correlation greater than a threshold; (13) De-dimensionalize and sum the operating data retained in step (12) to obtain the final key variables.

8. A shield machine fault monitoring system based on slow feature analysis, characterized in that: include: A key variable acquisition unit is used to acquire the operating data of the shield machine under normal working conditions, pre-process the operating data, and extract key variables from the shield machine operating data; A monitoring statistic construction unit is used to construct monitoring statistics based on the key variables and slow feature analysis, and calculate the control limit of each monitoring statistic; the monitoring statistics include steady-state monitoring statistics describing steady-state changes during the operation of the shield machine and dynamic monitoring statistics describing the dynamic change distribution of shield machine operation data; A real-time operation data acquisition unit is used to acquire the real-time operation data of the shield machine and pre-process the real-time operation data; A monitoring statistic calculation unit, used for calculating monitoring statistics based on slow feature analysis according to the pre-processed real-time operation data; A fault alarm unit is used to perform a fault alarm according to the monitoring statistic and its control limit.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the computer program is loaded into the processor, the shield machine fault monitoring method based on slow feature analysis according to any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the shield machine fault monitoring method based on slow feature analysis according to any one of claims 1 to 7 is implemented.

Images