Shield tunneling machine fault prediction method based on operation big data

Abstract

The invention discloses a shield tunneling machine fault prediction method based on operation big data, and the method comprises the steps: collecting real-time operation parameters of a shield tunneling machine, carrying out the parameter correction, carrying out the working condition self-adaptive normalization processing of the corrected parameters, extracting monotonic degradation characteristics, calculating a comprehensive degradation index, and carrying out the fault prediction of the shield tunneling machine. And inputting a historical sequence corresponding to the comprehensive degradation index before the current moment into a fault probability prediction model to perform fault prediction to obtain a fault probability, and finally setting a graded early warning probability threshold to perform graded early warning reminding. According to the method, interference factors can be effectively eliminated, and the equipment degradation process is quantified, so that the fault probability is accurately predicted.

Description

TECHNICAL FIELD

[0001] The application relates to a shield machine data processing technical field, in particular to a shield machine fault prediction method based on operation big data. BACKGROUND

[0002] The shield machine is large and complex equipment in tunnel construction, and its operation system covers a cutter system, a propulsion system, a hydraulic system, an electrical system and a muck handling system and the like. The shield machine is long-term operated under complex strata and variable working conditions, and the equipment parameters are easily disturbed by strata conditions, construction conditions, operation behaviors and sensor noise and the like.

[0003] The existing shield machine fault diagnosis and prediction methods are mostly based on instantaneous parameter threshold judgment or black box machine learning models, and have the following disadvantages: (1) it is difficult to distinguish the parameter fluctuation caused by the working condition change and the real equipment degradation; (2) the correlation between the multiple source operation parameters is not considered enough, and the degradation degree is easily amplified repeatedly; (3) the continuity and accumulation of the fault evolution process are limited; (4) the explicit modeling of the uncertainty factors is lacked, and the stability of the prediction result is insufficient. SUMMARY

[0004] The technical problem to be solved by the application is to provide a shield machine fault prediction method based on operation big data, which can effectively exclude interference factors and quantify the equipment degradation process, so as to realize accurate prediction of the fault probability.

[0005] The technical scheme of the application is: A shield machine fault prediction method based on operation big data, specifically comprising the following steps: (1) shield machine real-time operation parameter acquisition: various operation parameters of the shield machine are acquired in real time by using a sensor network; (2) operation parameter correction: the reliability coefficient of the operation parameter is calculated based on the long-term health real benchmark value and the real-time acquired operation parameter, the real-time acquired operation parameter is corrected based on the reliability coefficient of each operation parameter, and the corrected parameter corresponding to each operation parameter is obtained; (3) self-adaptive normalization processing of the corrected parameter: first, a plurality of typical working condition clusters are divided based on the historical operation parameters, then it is confirmed that each corrected parameter belongs to which working condition category corresponding to the typical working condition cluster, and finally each corrected parameter is normalized based on the working condition category corresponding to each corrected parameter, so that each corrected parameter is normalized to a normalized parameter; (4) Extracting monotonically degenerate features and calculating the comprehensive degradation index: First, set a sliding time window for the multidimensional normalized parameter sequence, then calculate the covariance matrix of the multidimensional normalized parameters within the time window, extract the first m principal components in descending order of eigenvalues ​​in the covariance matrix to obtain the principal component sequence, accumulate the bias of the principal component sequence, extract multiple monotonically degenerate features based on the accumulated bias value, and finally perform weighted fusion of multiple monotonically degenerate features to obtain the comprehensive degradation index; (5) Construct a fault probability prediction model, input the historical sequence corresponding to the comprehensive degradation index before the current time into the fault probability prediction model to predict the fault and obtain the fault probability; (6) Set a graded early warning probability threshold, analyze and compare the fault probability predicted by the fault probability prediction model with the graded early warning probability threshold, confirm the early warning level, and issue an early warning reminder at the corresponding level.

[0006] The various operating parameters of the tunnel boring machine include its mechanical parameters, hydraulic parameters, electrical parameters, and working condition parameters.

[0007] The correction of the aforementioned operating parameters specifically includes the following steps: S21. Calculate the first according to the following formula (1). Error variance of each operating parameter : (1); In equation (1), This represents the total number of runtime parameters collected in real time. The data collected in real time at time t represents the first data point of the tunnel boring machine. One operating parameter; This represents the first long-term statistical result under fault-free conditions. The average value of each operating parameter, i.e., the true baseline value for long-term health, is a constant. S22. Calculate the first according to the following formula (2). Reliability coefficient of each operating parameter : (2); S23. Based on the reliability coefficient, the real-time collected operating parameters are corrected, as shown in the following formula (3): (3); In equation (3), The number of samples collected at time t The corrected parameters are obtained after parameter adjustment of each running parameter.

[0008] The correction parameters undergo adaptive normalization processing under operating conditions, specifically including the following steps: S31. Using the operating condition parameters from the historical operating big data of the tunnel boring machine as feature vectors, an unsupervised clustering algorithm is used to divide the historical operating big data, resulting in multiple typical operating condition clusters. , Represents the total number of typical operating condition clusters, and each typical operating condition cluster... This corresponds to a stable construction state; S32, in each typical operating condition cluster Within this range, the healthy status data segment is selected, which is the historical interval with no fault records and stable parameters. For each corrected parameter... Calculate the statistical distribution, i.e., the mean, within this historical interval. and standard deviation ; S33. For each time t, calculate the distance or similarity between the operating condition parameters at each time t and the centers of each typical operating condition cluster, and then adjust each correction parameter. Assign to the cluster of typical working conditions with the smallest distance or the most similarity In this typical operating condition cluster The corresponding mean is The corresponding standard deviation is ; S34. Normalize each correction parameter based on the working condition category corresponding to each correction parameter, as shown in the following formula (4): (4); In equation (4), Representing the Each operating parameter is collected at time t, and its corresponding normalized parameter is collected.

[0009] The calculation of the comprehensive degradation index specifically includes the following steps: S41. For multidimensional normalized parameter sequences Set a sliding time window, with a window length of [value missing]. Thus, a vector matrix is ​​formed within each time window. ; S42. Calculate the covariance matrix of the multidimensional normalized parameters within each time window, and extract the top m principal components in descending order of eigenvalues ​​from the covariance matrix to obtain the principal component sequence. , ; S43. Accumulate the biases of the principal component sequences, as shown in the following formula (5): (5); In equation (5), Represents the number of times up to time t. The cumulative deviation of each principal component; This represents the historical time index, which is the historical time from the initial time to the current time t; Represents the time interval from the initial time to the current time t, for the th... All historical values ​​of each principal component are summed one by one. If the value is positive, the sum is added; if the value is negative, it is treated as 0. The final sum of the deviations is obtained. ; S44. Extract multiple monotonically degenerate features based on the accumulated bias value, i.e., use isotonic regression to... By fitting the data, a strictly monotonically increasing monotonically degenerate characteristic is obtained. ; S45. For m monotonically degenerate features The weighted fusion is performed as shown in the following formula (6): (6); In equation (6), This represents the comprehensive degradation index corresponding to each time window; The weights corresponding to each monotonically degenerate feature are determined using the entropy weight method. Multiple time windows corresponding to The historical sequence corresponding to the comprehensive degradation index before the current moment is formed and input into the fault probability prediction model for fault prediction.

[0010] The aforementioned fault probability prediction model is a Long Short-Term Memory (LSTM) network, comprising an input layer, 2-3 layers of LSTM units, and a fully connected output layer. During training, multiple degradation trajectories are extracted from the historical operational big data of the tunnel boring machine (TBM). Each degradation trajectory corresponds to a single process from healthy to faulty operation. The input features are... , The time length of the degradation trajectory is defined, and each degradation trajectory is labeled with a real label. Dropout technique is used to train a fault probability prediction model, resulting in a trained fault probability prediction model. This model is then used to predict faults, outputting the fault probability. See the following formula (7) for details: (7); In equation (7), Represents the original value output by the fully connected output layer; Represents the sigmoid activation function; This represents the mean function, which calculates the average value and outputs the 95% confidence interval.

[0011] When analyzing and comparing the fault probability predicted by the fault probability prediction model with the graded early warning probability threshold, the risk assessment value is first calculated according to the following formula (8). : (8); In equation (8), This represents the safety factor, with a value ranging from 1.65 to 2. The standard deviation of multiple predictions made by the failure probability prediction model; Set an emergency-level warning probability threshold, a warning-level warning probability threshold, and a caution-level warning probability threshold, with the emergency-level warning probability threshold being greater than the warning-level warning probability threshold, and the warning-level warning probability threshold being greater than the caution-level warning probability threshold; when the risk assessment value... When the probability of a warning is greater than the warning level threshold but less than the warning level threshold, a warning at the attention level will be issued. When the risk assessment value... When the probability of a risk assessment value is greater than the warning level threshold but less than the emergency level threshold, a warning level alert will be issued. When the probability exceeds the emergency warning threshold, an emergency warning will be issued.

[0012] The aforementioned fault probability prediction model performs continuous predictions and outputs predicted fault probability curves for multiple consecutive future time points. It also outputs the current health status, confidence interval, and remaining time estimate in real time. The remaining time estimate is the time remaining between the predicted future time of the fault and the current time. The model also provides a visual interface and early warning signals.

[0013] Advantages of this invention: (1) This invention effectively suppresses observation errors and improves the accuracy of subsequent fault prediction by calculating the confidence coefficient and correcting the real-time collected operating parameters through the confidence coefficient.

[0014] (2) In order to further eliminate the fluctuation of normal parameters caused by different working conditions, the present invention performs working condition adaptive normalization processing. After normalization processing, the normalized parameter in the healthy state is approximately equal to zero, the fluctuation of normal working conditions is compressed to the range of [-1,1], while the actual degradation will lead to a continuous deviation from zero, thereby achieving effective suppression of working condition interference.

[0015] (3) Although the normalized parameters obtained after the adaptive normalization processing of the working conditions in this invention have removed the interference of the working conditions, the parameters are highly correlated. Direct fusion is likely to cause repeated amplification of degradation information and the features are not monotonic. Therefore, monotonic degradation features are extracted to calculate the comprehensive degradation index. The comprehensive degradation index is close to zero in the healthy state and increases monotonically with degradation. The comprehensive degradation index has anti-interference, monotonicity and low redundancy at the same time, and can accurately depict the fault evolution process.

[0016] (4) The present invention sets a graded early warning probability threshold to realize graded early warning, reminding staff to perform corresponding on-site inspection operations according to the early warning level, and simultaneously outputs the predicted fault probability curve, current health status, confidence interval and remaining time estimate, so that staff can keep abreast of the working status of the tunnel boring machine in real time and perform corresponding analysis and processing based on the output information. Attached Figure Description

[0017] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] See Figure 1 A method for predicting tunnel boring machine (TBM) failures based on big data analysis includes the following steps: (1) Real-time acquisition of shield machine operating parameters: Various operating parameters of the shield machine are collected in real time using a sensor network. These parameters include mechanical parameters (cutterhead speed and torque, etc.), hydraulic parameters (cylinder pressure and flow, etc.), electrical parameters (current, voltage and other electrical signals), and working condition parameters (propulsion mode, earth pressure and stratum type, etc.). (2) Correction of operating parameters, specifically including the following steps: S21. Calculate the first according to the following formula (1). Error variance of each operating parameter : (1); In equation (1), This represents the total number of runtime parameters collected in real time. The data collected in real time at time t represents the first data point of the tunnel boring machine. One operating parameter; This represents the first long-term statistical result under fault-free conditions. The average value of each operating parameter, i.e., the true baseline value for long-term health, is a constant. S22. Calculate the first according to the following formula (2). Reliability coefficient of each operating parameter : (2); S23. Based on the reliability coefficient, the real-time collected operating parameters are corrected, as shown in the following formula (3): (3); In equation (3), The number of samples collected at time t The corrected parameters are obtained after parameter correction of each operating parameter; (3) Correcting parameters and performing adaptive normalization processing for operating conditions, specifically including the following steps: S31. Using the operating condition parameters from the historical operating big data of the tunnel boring machine as feature vectors, an unsupervised clustering algorithm is used to divide the historical operating big data, resulting in multiple typical operating condition clusters. , Represents the total number of typical operating condition clusters, and each typical operating condition cluster... This corresponds to a stable construction state; S32, in each typical operating condition cluster Within this range, the healthy status data segment is selected, which is the historical interval with no fault records and stable parameters. For each corrected parameter... Calculate the statistical distribution, i.e., the mean, within this historical interval. and standard deviation ; S33. For each time t, calculate the distance or similarity between the operating condition parameters at each time t and the centers of each typical operating condition cluster, and then adjust each correction parameter. Assign to the cluster of typical working conditions with the smallest distance or the most similarity In this typical operating condition cluster The corresponding mean is The corresponding standard deviation is ; S34. Normalize each correction parameter based on the working condition category corresponding to each correction parameter, as shown in the following formula (4): (4); In equation (4), Representing the Each operating parameter is collected at time t, and its corresponding normalized parameter is collected. (4) Extract monotonic degradation features and calculate the comprehensive degradation index, which includes the following steps: S41. For multidimensional normalized parameter sequences Set a sliding time window, with a window length of [value missing]. Thus, a vector matrix is ​​formed within each time window. ; S42. Calculate the covariance matrix of the multidimensional normalized parameters within each time window, and extract the top m principal components in descending order of eigenvalues ​​from the covariance matrix to obtain the principal component sequence. , m is determined by a cumulative variance contribution rate > 95%; S43. Accumulate the biases of the principal component sequences, as shown in the following formula (5): (5); In equation (5), Represents the number of times up to time t. The cumulative deviation of each principal component; This represents the historical time index, which is the historical time from the initial time to the current time t; Represents the time interval from the initial time to the current time t, for the th... All historical values ​​of each principal component are summed one by one. If the value is positive, the sum is added; if the value is negative, it is treated as 0. The final sum of the deviations is obtained. ; S44. Extract multiple monotonically degenerate features based on the accumulated bias value, i.e., use isotonic regression to... By fitting the data, a strictly monotonically increasing monotonically degenerate characteristic is obtained. ; S45. For m monotonically degenerate features The weighted fusion is performed as shown in the following formula (6): (6); In equation (6), This represents the comprehensive degradation index corresponding to each time window; The weights corresponding to each monotonically degenerate feature are determined using the entropy weight method. Multiple time windows corresponding to The historical sequence corresponding to the comprehensive degradation index before the current moment is formed; (5) Construct a fault probability prediction model. The fault probability prediction model is a long short-term memory network, including an input layer, 2-3 layers of LSTM units, and a fully connected output layer. During the training of the fault probability prediction model, multiple degradation trajectories are extracted from the historical operating big data of the tunnel boring machine. Each degradation trajectory corresponds to a process from health to failure. The input features are... , The time length of the degradation trajectory is defined, and each degradation trajectory is labeled with a real label. Dropout technology is used to train the fault probability prediction model (loss function is binary cross-entropy or a combination of mean squared error). Early stopping and learning rate decay are employed to prevent overfitting. Model validation uses k-fold cross-validation. Evaluation metrics include accuracy, false positive rate, average early warning time, and Brier score. This yields the trained fault probability prediction model. The historical sequence corresponding to the comprehensive degradation index before the current time is input into the trained fault probability prediction model for fault prediction, and the output fault probability is... See the following formula (7) for details: (7); In equation (7), Represents the original value output by the fully connected output layer; Represents the sigmoid activation function; This represents the mean function, which calculates the average value and outputs the 95% confidence interval. (6) Set graded warning probability thresholds and issue graded warning reminders; the graded warning probability thresholds include emergency level warning probability threshold = 0.95, warning level warning probability threshold = 0.8 and attention level warning probability threshold = 0.6; The failure probability predicted by the failure probability prediction model is analyzed and compared with the graded early warning probability threshold. First, the risk assessment value is calculated according to the following formula (8). : (8); In equation (8), This represents the safety factor, with a value ranging from 1.65 to 2. The standard deviation of multiple predictions made by the failure probability prediction model; When the risk assessment value When the value is greater than 0.6 and less than 0.8, a warning alert at the attention level will be issued. When the risk assessment value... When the value is greater than 0.8 and less than 0.95, a warning level alert will be issued. When the risk assessment value... When the value is greater than 0.95, an emergency-level early warning will be issued. (7) The fault probability prediction model makes continuous predictions and outputs the predicted fault probability curves for multiple consecutive time points in the future. At the same time, it outputs the current health status, confidence interval and remaining time estimate in real time. The remaining time estimate is the time from the current time to the predicted future time of the fault. It also provides a visualization interface and early warning signal.

[0020] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for tunneling machine fault prediction based on operation big data, characterized in that: Specifically comprising the following steps: (1) Real-time acquisition of shield machine operation parameters: various operation parameters of the shield machine are acquired in real time by using a sensor network; (2) Correction of operation parameters: a reliability coefficient of the operation parameters is calculated based on long-term healthy true benchmark values and the real-time acquired operation parameters, and the real-time acquired operation parameters are corrected based on the reliability coefficients of the operation parameters to obtain correction parameters corresponding to the operation parameters; (3) Self-adaptive normalization processing of the correction parameters: first, a plurality of typical working condition clusters are divided based on historical operation parameters, then it is determined to which working condition category the correction parameters belong, and finally the correction parameters are normalized based on the working condition categories corresponding to the correction parameters, so that the correction parameters are normalized into normalized parameters; (4) Extraction of monotonic degradation features and calculation of a comprehensive degradation index: first, a sliding time window is set for the multi-dimensional normalized parameter sequence, then a covariance matrix of the multi-dimensional normalized parameters in the time window is calculated, the first m principal components in descending order of eigenvalues in the covariance matrix are extracted to obtain a principal component sequence, the principal component sequence is subjected to bias accumulation, a plurality of monotonic degradation features are extracted based on the bias accumulation values, and finally the plurality of monotonic degradation features are weighted and fused to obtain the comprehensive degradation index; (5) Construction of a fault probability prediction model, input of a historical sequence corresponding to the comprehensive degradation index at the current time into the fault probability prediction model for fault prediction, and obtaining of a fault probability; (6) Setting of a hierarchical early warning probability threshold, analysis and comparison of the fault probability predicted and output by the fault probability prediction model with the hierarchical early warning probability threshold, determination of an early warning level, and early warning reminding at the corresponding level.

2. The shield tunnel fault prediction method based on operation big data according to claim 1, characterized in that: The various operation parameters of the shield machine include mechanical parameters, hydraulic parameters, electrical parameters and working condition parameters of the shield machine.

3. The shield tunnel fault prediction method based on operation big data according to claim 2, characterized in that: The correction of the operation parameters specifically comprises the following steps: S21, calculate the error variance of the i-th operating parameter according to the following formula (1) : σ2i= 1 n ∑ (xi - x)2 : （1）； In equation (1), This represents the total number of runtime parameters collected in real time. The data collected in real time at time t represents the first data point of the tunnel boring machine. One operating parameter; This represents the first long-term statistical result under fault-free conditions. The average value of each operating parameter, i.e., the true baseline value for long-term health, is a constant. S22, calculate the reliability coefficient of the i-th operating parameter according to the following formula (2) :​ （2）； S23, parameter correction of the real-time acquired operation parameters based on the reliability coefficient, specifically as shown in the following formula (3): （3）； In formula (3), the modified parameter obtained after the parameter correction of the i-th operating parameter collected at the representative time t. the modified parameter obtained after the parameter correction of the i-th operating parameter collected at the representative time t.

4. The shield tunnel fault prediction method based on operation big data according to claim 3, characterized in that: The self-adaptive normalization processing of the correction parameters specifically comprises the following steps: S31, using the working condition parameters in the historical operation big data of the shield tunneling machine as feature vectors, using an unsupervised clustering algorithm to divide the historical operation big data to obtain multiple typical working condition clusters , represent the total number of typical working condition clusters, and each typical working condition cluster corresponds to a stable construction state; S32, in each typical operating condition cluster Within this range, the healthy status data segment is selected, which is the historical interval with no fault records and stable parameters. For each corrected parameter... Calculate the statistical distribution, i.e., the mean, within this historical interval. and standard deviation ; S33、For each time t, calculate the distance or similarity between the working condition parameter of each time t and each typical working condition cluster center, and distribute each modified parameter to the typical working condition cluster with the smallest distance or the most similarity , the typical working condition cluster corresponding to the mean value , the corresponding standard deviation , the corresponding standard deviation ; S34, normalization processing of the correction parameters based on the working condition categories corresponding to the correction parameters, specifically as shown in the following formula (4): （4）； In formula (4), represent the first operating parameter at time t corresponding to the normalized parameter.

5. The method according to claim 4, wherein: The calculation of the comprehensive degradation index specifically comprises the following steps: S41、to the multi-dimensional normalization parameter sequence A sliding time window is set, the length of the time window is , so as to form a vector matrix in each time window ; S42, calculate the covariance matrix of the multi-dimensional normalization parameters in each time window, extract the first m principal components in descending order of eigenvalues in the covariance matrix to obtain a principal component sequence , ; S43, bias accumulation of the principal component sequence, specifically as shown in the following formula (5): （5）； In formula (5), represents the deviation accumulated value of the i-th principal component at the current time t; represents the history time index, i.e. the history time between the initial time and the current time t; represents that all history values of the i-th principal component from the initial time to the current time t are processed by item-by-item accumulation, and if the value is positive, the value is accumulated, and if the value is negative, the value is processed as 0, and finally the deviation accumulated value is obtained ;​​ S44, based on the deviation cumulative value, extract a plurality of monotone degradation features, that is, use isometric regression to fit to obtain strictly monotonically increasing monotone degradation features ; S45, m monotone degradation features The weighted fusion is performed, and the specific formula (6) is as follows. （6）； In formula (6), representing the comprehensive degradation index corresponding to each time window; representing the weight corresponding to each monotonic degradation feature, which is determined by an entropy weight method. The plurality of time windows correspond to The history sequence corresponding to the current moment before the comprehensive degradation index is input into the failure probability prediction model for failure prediction.

6. The shield tunnel fault prediction method based on operation big data according to claim 5, characterized in that: The failure probability prediction model is a long short-term memory network, comprising an input layer, 2-3 LSTM units and a fully connected output layer; during training of the failure probability prediction model, a plurality of degradation trajectories are extracted from historical operation big data of the shield tunneling machine, each degradation trajectory corresponds to a process from health to failure, the input features are , is the time length of the degradation trajectory, and each degradation trajectory is labeled with a true label; the dropout technique is used to train the failure probability prediction model, to obtain a trained failure probability prediction model; the trained failure probability prediction model is used for failure prediction, and the output failure probability Specifically, as shown in the following formula (7): （7）； In formula (7), represent the raw values output by the fully connected output layer; represent the sigmoid activation function; represent the mean function, i.e., to average the values, output the 95% confidence interval.

7. The shield tunnel fault prediction method based on operation big data according to claim 6, characterized in that: When the failure probability predicted according to the failure probability prediction model is compared with the graded early warning probability threshold, a risk assessment value is first calculated according to the following formula (8) : （8）； In formula (8), represents a safety factor, and has a value of 1.65-2; represents a standard deviation of multiple predictions of the failure probability prediction model; The emergency level warning probability threshold, the warning level warning probability threshold and the attention level warning probability threshold are set, and the emergency level warning probability threshold is greater than the warning level warning probability threshold, and the warning level warning probability threshold is greater than the attention level warning probability threshold; when the risk assessment value is greater than the attention level warning probability threshold and less than the warning level warning probability threshold, an attention level warning reminder is performed, when the risk assessment value is greater than the warning level warning probability threshold and less than the emergency level warning probability threshold, a warning level warning reminder is performed, when the risk assessment value is greater than the emergency level warning probability threshold, an emergency level warning reminder is performed.

8. The shield tunnel fault prediction method based on operation big data according to claim 1, characterized in that: The fault probability prediction model performs continuous prediction, outputs a predicted fault probability curve at a plurality of continuous future time points, simultaneously outputs a current health state, a confidence interval and a remaining time estimate, the remaining time estimate is a time length from the current time to a future time when the predicted fault occurs, and a visual interface and an early warning reminding signal are provided.

Images