CN113982605B - Multi-layer shield method tunnel safety protection system and method - Google Patents

Abstract

The invention discloses a multi-level shield tunnel safety protection system and a method, wherein the tunnel safety protection system comprises a data acquisition module, an abnormality diagnosis module, a typical characteristic comparison module and an alarm result output module; the data acquisition module is used for acquiring data of the setting sensor, and the acquired data comprise real-time shield construction parameters; the abnormality diagnosis module is used for identifying abnormality from the selected real-time shield construction parameters; the typical characteristic comparison module is used for extracting characteristic indexes and threshold values from typical cases of the tail grease leakage event so as to match whether the data change characteristics in the tail grease pressure data accord with abnormal typical characteristics of the tail grease leakage event. The system and the method for protecting the safety of the multi-level shield tunnel are beneficial to timely finding out the problem of shield construction abnormality, can improve the monitoring level of data on a construction site, and help constructors to timely early warn construction abnormality and reasonably operate shield tunneling.

Description

Multi-layer shield method tunnel safety protection system and method

Technical Field

The invention belongs to the technical field of tunneling, relates to a tunneling protection system, and particularly relates to a multi-level shield tunneling safety protection system and method.

Background

The shield tunneling machine is a large-scale machine widely used in construction projects such as subways and tunnels, integrates light, mechanical, electric, liquid, sensing and information technologies, and can perform multiple functions such as soil body cutting and conveying, tunnel lining assembling, measurement, guiding and deviation correcting, and the like, so that the integrated construction of the tunnel is completed. The main advantages of the shield method include: the construction safety of the shield is high due to the excavation and lining work of the shield; the construction process can not affect the land and river channel of the upper surface; the construction progress is not affected by weather; the noise and vibration generated are less harmful to the surroundings. However, the environment of the shield construction method is very complex, and the shield machine has higher automation degree, and the internal equipment faults and the external environment changes can generate chain reactions, so that project progress can be influenced, economic losses are caused, safety of constructors and the periphery of a construction site can be endangered, and serious accident risks exist.

In actual engineering projects, a certain hysteresis exists in the knowledge of constructors on abnormal events, and when abnormality is found, the construction progress is often influenced. The construction site needs to analyze the data change of the construction parameters before and after the abnormal event is found by an inversion analysis method, and the construction experience is summarized so as to conveniently identify the abnormality from the specific construction parameters.

In order to ensure the efficiency and quality of shield construction, the sensors assembled on the shield machine body can capture information of corresponding equipment and systems, including operation parameters and feedback parameters of the equipment. These parameters are affected by complex factors and change in real time along with construction, so that the parameters are widely applied to aspects such as shield tunneling control and anomaly monitoring. Based on the sensing data recorded in real time, the construction risk of the shield machine hidden in the data can be found, so that measures can be further taken to process, the risk is avoided as much as possible, and the loss is reduced. Therefore, when monitoring the risk of the inside and the outside of the shield, monitoring and analyzing based on the construction parameter data acquired by the sensor becomes a focus of study.

In view of this, there is an urgent need to design a new tunneling protection method in order to overcome at least some of the above-mentioned drawbacks of the existing tunneling protection methods.

Disclosure of Invention

The invention provides a multi-level shield method tunnel safety protection system and a multi-level shield method tunnel safety protection method, which are beneficial to timely finding out the problem of shield construction abnormality, can improve the monitoring level of data on a construction site, and help constructors to timely early warn construction abnormality and reasonably operate shield tunneling.

In order to solve the technical problems, according to one aspect of the present invention, the following technical scheme is adopted:

a multi-level shield tunnel security protection system, the tunnel security protection system comprising:

The data acquisition module is used for acquiring data of the setting sensor, and the acquired data comprise real-time shield construction parameters;

The abnormality diagnosis module is used for identifying abnormality from the selected real-time shield construction parameters; constructing a subsequence data set by data cleaning and time delay method on shield tail grease pressure data in a section of normal construction mode, inputting the subsequence data set into a mode learning model, and learning a normal change mode of shield tail grease pressure; then updating real-time data, and identifying whether the change mode of the latest data section is changed beyond the set size according to the residual values of the actual grease pressure recursion diagram matrix and the predicted grease pressure recursion diagram matrix; if the residual value is higher than a threshold value set under engineering experience, the tail oil pressure is considered to have abnormal mode at the moment;

The typical characteristic comparison module is used for extracting characteristic indexes and threshold values from typical cases of the shield tail grease leakage event so as to match whether the data change characteristics in the shield tail grease pressure data accord with abnormal typical characteristics of the shield tail grease leakage event.

As an implementation mode of the invention, the data acquired by the data acquisition module comprises shield tail grease pressure, and the construction parameters related to the data acquisition module also comprise shield tail grease injection quantity, grease set value and shield tail gap;

The data acquisition module comprises: the real-time construction parameter extraction unit is used for selecting common construction parameters related to the abnormality of the shield machine from construction experience, extracting selected construction parameter data from shield real-time data, preprocessing time sequence data, and inputting the preprocessed time sequence data into the pattern learning model for preliminary abnormality identification.

As an embodiment of the present invention, the exemplary feature comparison module includes:

The typical characteristic extraction unit is used for extracting typical characteristics from the tail oil pressure and the oil injection quantity, including abrupt proportion indexes of the oil pressure and the oil injection quantity; after determining the threshold value under each construction parameter index, taking the indexes as visual basis for abnormality identification;

the typical characteristic identification unit is used for aiming at the shield tail grease leakage event from the preprocessed core parameter time sequence data and calculating characteristic index values related to the abnormal event; when the index value exceeds a set threshold value, the tail grease leakage event is considered to be possible;

the tunnel security protection system further comprises: and the alarm result output module is used for alarming the abnormality.

As an embodiment of the present invention, the abnormality diagnosis module includes at least one of a threshold value alarm unit, a heartbeat failure alarm unit, a control deviation diagnosis unit, and a data-driven abnormality diagnosis unit:

The threshold alarming unit is used for identifying individual equipment faults occurring in the shield tunneling machine; according to the method, distribution of recent sensing data D of target monitoring equipment is counted, the change range D _l,d_h of the normal working state of the equipment is determined after data denoising is carried out, and an upper-lower threshold D _l,d_h is set; when the value d _r recorded by the sensor in real time is lower than the lower limit d _l or higher than the upper limit d _h, the equipment is considered to be abnormal at the moment;

The heartbeat fault alarm unit is used for finding out abnormal communication between the shield machine and an external database; the basic form of the heartbeat protocol is: the shield PLC control system regularly transmits pulse signals at a certain frequency, and meanwhile, the intelligent tunneling control system receives and checks the continuity of heartbeat; the heartbeat detection rule is that if the current time and the last time the intelligent tunneling control system receives the heartbeat exceed the set timeout, the timeout is reached, and the heartbeat failure of the shield PLC control system is judged, namely the shield has communication faults;

And the control deviation diagnosis unit is used for identifying the control risk caused by the model defect in the shield posture. The control indexes of the shield posture comprise two horizontal deviation indexes (left and right) and two elevation deviation indexes (up and down), real-time monitoring data of the control indexes in the data layer record actual four control index values, and the four control index values comprise: left horizontal deviation b _left, right horizontal deviation b _right, upper elevation deviation b _high and lower elevation deviation b _low; the model real-time control output parameters are output values of the attitude control model to the current attitude control parameters of the shield, wherein the output values comprise a horizontal deviation left b '_left, a horizontal deviation right b' _right, an elevation deviation upper b '_high and an elevation deviation lower b' _low; if the deviation between the model output value and the actual control index is overlarge, the control parameter value output by the model is considered to be abnormal; therefore, the difference between the model output value of each index and the actual index is calculated, and the calculation method of the difference score _m is shown in formula (1):

score _m＝|b'_m-b_m×(1±k_m) I, m ε { left, right, low, high }, k ε [0%,100% ] formula (1)

Wherein m represents deviation indexes in different directions, b _m and b' _m represent control index values and model output recommended values of the current shield, and k _m represents a permissible floating range of a corresponding control value; meanwhile, a threshold s _m is set for the difference score _m; when score _m is larger than threshold s _m, the control parameters output by the model are considered to deviate from the current shield posture, and the model has defects and needs to be adjusted;

The data-driven abnormality diagnosis unit is used for acquiring construction parameters related to common abnormal events from real-time shield data, then learning a change mode of parameters in a normal working state from a time sequence formed by the construction parameters within a period of time, and identifying whether the change mode of the latest data section is greatly changed; if a large change occurs, it is considered that there is a pattern abnormality.

As one embodiment of the present invention, the data-driven abnormality diagnosis unit includes: the system comprises a real-time construction parameter extraction unit, a mode learning model construction unit and an abnormal mode identification unit; the real-time construction parameter extraction unit comprises a time sequence data cleaning unit and a time window sequence construction unit;

The time sequence data cleaning unit is used for supplementing the null value with the corresponding statistical value according to the actual distribution condition when the data of the selected construction parameter part time point is missing; the time sequence data cleaning unit is used for deleting repeated parts when the data of the selected construction parameter part time points are repeated; the time sequence data cleaning unit is used for considering the normal change interval and the construction state of the parameters when the data of the selected construction parameter part time points have zero values; selecting construction parameters which possibly have 0 values, not processing the construction parameters, and processing the construction parameters which are not zeroed according to missing values if the 0 values are sporadic, wherein the sensor is considered to be possibly abnormal in recording; if the 0 value appears repeatedly or continuously, considering that equipment abnormality or other shield abnormality time possibly exists, temporarily retaining the abnormal value;

The time window sequence constructing unit is used for intercepting a plurality of time windows with equal length from the original shield time sequence data, namely, a time delay method is adopted to divide an original data set X= { X ₁,x₂,...,x_i,...,x_n } into a plurality of time subsequences, wherein X _i＝{x_i,x_i+τ,...,x_i+(m-1)τ}, X_i represents an ith time sequence segment, X _i is an ith observed sample, m is an embedded dimension, and tau is a time delay factor; the obtained time window sequence data set X ₁,X₂,...,X_N represents the shield construction state at different moments, wherein N satisfies N=n- (m-1) tau;

the pattern learning model building unit is used for building a pattern learning module; the constructed model learning model is divided into two parts, namely construction parameter characteristic extraction and normal construction model self-learning;

the construction parameter characteristic extraction method is threshold-free recursion diagram conversion; the thresholdless recursive graph conversion method calculates the distance of the data between the time points in each time window as shown in formula (2):

R _i,j＝||X_i-X_j i, j=1, 2,..n' formula (2)

Wherein, I and I are distance norms, and N' is the number of time points contained in the time window; converting the shield construction data from a time sequence into a two-dimensional symmetric matrix, wherein each point in the matrix represents the distance between any two time points, so that the dynamic characteristics of the shield data on time sequence are highlighted;

the method adopted by the self-learning of the normal construction mode is an extreme learning machine self-encoder. After the non-threshold recursion diagram is converted into a recursion diagram matrix of a real-time construction parameter time subsequence, the subsequence recursion matrix is used as input and output of an extreme learning machine at the same time, and an output layer parameter set value is calculated through a random input layer parameter, so that a model reconstruction parameter is obtained; the method comprises the following specific steps:

(1) Setting the self-encoder of the extreme learning machine as a single-layer neural network of M input and output neurons and L hidden-layer neurons; constructing an input P and an output Q with M consecutive real-time construction parameter recursion matrices, wherein p=q= { R ₁,R₂,...,R_M }, the training samples are therefore M is less than N;

(2) Randomly distributing orthogonalization input weights a _i and deviation b _i based on a specific continuous probability distribution function, and solving an hidden layer H (P) = [ H ₁(p),...,h_L (P) ] corresponding to an input layer P through an activation function H (·), wherein the calculation mode of neurons of each hidden layer is shown in a formula (3);

(3) After calculating hidden layer neurons, under the condition of knowing the output Q, the optimal output weight which minimizes the residual error can be obtained through a formula (4) At this time a, b,The formed model parameter set is the result of model self-learning, namely the changing mode of the shield real-time parameter in the construction period;

An abnormal pattern recognition unit for recognizing an abnormal pattern after the normal construction pattern learning is completed; acquiring the latest construction parameters from the shield real-time data, and forming the latest time subsequence X _N+1 to be checked; after conversion to the thresholdless recursive matrix R _N+1, input a _i,b_i, In an extreme learning machine network formed by P '(P' =R _N+1), outputting a reconstruction result in a normal construction state by using a formula (5);

Where H (a, b, P ') represents the hidden layer neurons of the current matrix and Q' is the reconstructed thresholdless recursive matrix. In order to calculate the residual errors of the actual matrix P 'and the predicted matrix Q', the method adopts radial basis function calculation, and omega _N+1 represents the residual errors of the two matrices, namely the abnormal fraction of the time subsequence of the (n+1) th construction parameter, as shown in a formula (6);

And setting a threshold delta of the anomaly score by combining the change characteristics of different construction parameters and the actual construction requirements, and considering that the current shield construction has anomalies related to the construction parameters when omega is more than or equal to delta.

According to another aspect of the invention, the following technical scheme is adopted: a multi-layer shield tunnel safety protection method comprises the following steps:

A data acquisition step of acquiring data of a setting sensor, wherein the acquired data comprises real-time shield construction parameters;

An abnormality diagnosis step of identifying an abnormality from the selected real-time shield construction parameters; constructing a subsequence data set by data cleaning and time delay method on shield tail grease pressure data in a section of normal construction mode, inputting the subsequence data set into a mode learning model, and learning a normal change mode of shield tail grease pressure; then updating real-time data, and identifying whether the change mode of the latest data section is changed beyond the set size according to the residual values of the actual grease pressure recursion diagram matrix and the predicted grease pressure recursion diagram matrix; if the residual value is higher than a threshold value set under engineering experience, the tail oil pressure is considered to have abnormal mode at the moment;

And a typical characteristic comparison step, wherein characteristic indexes and thresholds are extracted from typical cases of the tail grease leakage event, so that whether the data change characteristics in the tail grease pressure data accord with abnormal typical characteristics of the tail grease leakage event is matched.

In the step of obtaining the data, the obtained data comprises shield tail grease pressure, and the construction parameters related to the data comprise shield tail grease injection quantity, grease set value and shield tail gap;

The data acquisition step includes: and a real-time construction parameter extraction step, namely selecting common construction parameters related to the abnormality of the shield machine from construction experience, extracting selected construction parameter data from shield real-time data, preprocessing time sequence data, and inputting the preprocessed time sequence data into a pattern learning model for preliminarily identifying the abnormality.

As an embodiment of the present invention, the exemplary feature comparing step includes:

Extracting typical characteristics from the tail oil pressure and the oil injection quantity, wherein the typical characteristics comprise abrupt proportion indexes of the oil pressure and the oil injection quantity; after determining the threshold value under each construction parameter index, taking the indexes as visual basis for abnormality identification;

A typical feature identification step, namely aiming at a shield tail grease leakage event from the preprocessed core parameter time sequence data, and calculating a feature index value related to an abnormal event; when the index value exceeds a set threshold value, the tail grease leakage event is considered to be possible;

The tunnel security protection method further comprises the following steps: and outputting an alarm result, namely alarming the abnormality.

As an embodiment of the present invention, the abnormality diagnosis step includes at least one of a threshold value alarm step, a heartbeat failure alarm step, a control deviation diagnosis step, and a data-driven abnormality diagnosis step:

The threshold alarming step comprises the following steps: identifying individual equipment faults occurring in the shield tunneling machine; according to the method, distribution of recent sensing data D of target monitoring equipment is counted, the change range D _l,d_h of the normal working state of the equipment is determined after data denoising is carried out, and an upper-lower threshold D _l,d_h is set; when the value d _r recorded by the sensor in real time is lower than the lower limit d _l or higher than the upper limit d _h, the equipment is considered to be abnormal at the moment;

The step of heartbeat fault alarming comprises the following steps: the communication abnormality between the shield machine and an external database is found; the basic form of the heartbeat protocol is: the shield PLC control system regularly transmits pulse signals at a certain frequency, and meanwhile, the intelligent tunneling control system receives and checks the continuity of heartbeat; the heartbeat detection rule is that if the current time and the last time the intelligent tunneling control system receives the heartbeat exceed the set timeout, the timeout is reached, and the heartbeat failure of the shield PLC control system is judged, namely the shield has communication faults;

The control deviation diagnosing step includes: and identifying the control risk caused by the model defect in the shield posture. The control indexes of the shield posture comprise two horizontal deviation indexes (left and right) and two elevation deviation indexes (up and down), real four control index values (the four control index values comprise horizontal deviation left b _left, horizontal deviation right b _right, elevation deviation upper b _high and elevation deviation lower b _low) are recorded in real-time monitoring data of the control indexes in the data layer, and the output values of the posture control model to the current posture control parameters of the shield are output in real-time control output parameters of the model (the output values comprise horizontal deviation left b '_left, horizontal deviation right b' _right, elevation deviation upper b '_high and elevation deviation lower b' _low); if the deviation between the model output value and the actual control index is overlarge, the control parameter value output by the model is considered to be abnormal; therefore, the difference between the model output value of each index and the actual index is calculated, and the calculation method of the difference score _m is shown in formula (1):

score _m＝|b'_m-b_m×(1±k_m) I, m ε { left, right, low, high }, k ε [0%,100% ] formula (1)

Wherein m represents deviation indexes in different directions, b _m and b' _m represent control index values and model output recommended values of the current shield, and k _m represents a permissible floating range of a corresponding control value; meanwhile, a threshold s _m is set for the difference score _m; when score _m is larger than threshold s _m, the control parameters output by the model are considered to deviate from the current shield posture, and the model has defects and needs to be adjusted;

The data-driven abnormality diagnosis step includes: acquiring construction parameters related to common abnormal events from real-time shield data, then learning a change mode of the parameters in a normal working state from a time sequence formed by the construction parameters within a period of time, and identifying whether the change mode of the latest data section is greatly changed; if a large change occurs, it is considered that there is a pattern abnormality.

As one embodiment of the present invention, the data-driven abnormality diagnosis step includes: a real-time construction parameter extraction step, a pattern learning model construction step and an abnormal pattern recognition step; wherein the real-time construction parameter extraction step comprises a time sequence data cleaning step and a time window sequence construction step

The time sequence data cleaning step comprises the following steps: when the data of the selected construction parameter part time point is missing, the null value is complemented by the corresponding statistical value according to the actual distribution condition; deleting the repeated part when the data of the selected construction parameter part time point is repeated; when the data of the selected construction parameter part time points have zero values, the normal change interval and the construction state of the parameters are considered; selecting construction parameters which possibly have 0 values, not processing the construction parameters, and processing the construction parameters which are not zeroed according to missing values if the 0 values are sporadic, wherein the sensor is considered to be possibly abnormal in recording; if the 0 value appears repeatedly or continuously, considering that equipment abnormality or other shield abnormality time possibly exists, temporarily retaining the abnormal value;

The time window sequence construction step includes: intercepting a plurality of equal-length time windows from original shield time sequence data, namely dividing an original data set X= { X ₁,x₂,...,x_i,...,x_n } into a plurality of time subsequences by adopting a time delay method, wherein X _i＝{x_i,x_i+τ,...,x_i+(m-1)τ}, X_i represents an ith time sequence segment, X _i is an ith observed sample, m is an embedded dimension, and tau is a time delay factor; the obtained time window sequence data set X ₁,X₂,...,X_N represents the shield construction state at different moments, wherein N satisfies N=n- (m-1) tau;

the pattern learning model construction step comprises constructing a pattern learning module; the constructed model learning model is divided into two parts, namely construction parameter characteristic extraction and normal construction model self-learning;

the construction parameter characteristic extraction method is threshold-free recursion diagram conversion; the thresholdless recursive graph conversion method calculates the distance of the data between the time points in each time window as shown in formula (2):

R _i,j＝||X_i-X_j i, j=1, 2,..n' formula (2)

Wherein, I and I are distance norms, and N' is the number of time points contained in the time window; converting the shield construction data from a time sequence into a two-dimensional symmetric matrix, wherein each point in the matrix represents the distance between any two time points, so that the dynamic characteristics of the shield data on time sequence are highlighted;

the method adopted by the self-learning of the normal construction mode is an extreme learning machine self-encoder. After the non-threshold recursion diagram is converted into a recursion diagram matrix of a real-time construction parameter time subsequence, the subsequence recursion matrix is used as input and output of an extreme learning machine at the same time, and an output layer parameter set value is calculated through a random input layer parameter, so that a model reconstruction parameter is obtained; the method comprises the following specific steps:

(1) Setting the self-encoder of the extreme learning machine as a single-layer neural network of M input and output neurons and L hidden-layer neurons; constructing an input P and an output Q with M consecutive real-time construction parameter recursion matrices, wherein p=q= { R ₁,R₂,...,R_M }, the training samples are therefore M is less than N;

(2) Randomly distributing orthogonalization input weights a _i and deviation b _i based on a specific continuous probability distribution function, and solving an hidden layer H (P) = [ H ₁(p),...,h_L (P) ] corresponding to an input layer P through an activation function H (·), wherein the calculation mode of neurons of each hidden layer is shown in a formula (3);

(3) After calculating hidden layer neurons, under the condition of knowing the output Q, the optimal output weight which minimizes the residual error can be obtained through a formula (4) At this time a, b,The formed model parameter set is the result of model self-learning, namely the changing mode of the shield real-time parameter in the construction period;

The abnormal pattern recognition step includes: identifying an abnormal mode after the normal construction mode learning is finished; acquiring the latest construction parameters from the shield real-time data, and forming the latest time subsequence X _N+1 to be checked; after conversion to the thresholdless recursive matrix R _N+1, input a _i,b_i, In an extreme learning machine network formed by P '(P' =R _N+1), outputting a reconstruction result in a normal construction state by using a formula (5);

Wherein H (a, b, P ') represents hidden layer neurons of the current matrix, Q' is a reconstructed thresholdless recursive matrix; in order to calculate the residual errors of the actual matrix P 'and the predicted matrix Q', the method adopts radial basis function calculation, as shown in a formula (6), and omega _N+1 represents the residual errors of the two matrices, namely the abnormal fraction of the (n+1) th construction parameter time subsequence.

The threshold delta of the anomaly score can be set by combining the change characteristics of different construction parameters and the actual construction requirements, and when omega is more than or equal to delta, the current shield construction is considered to have the anomaly related to the construction parameters.

The invention has the beneficial effects that: the system and the method for protecting the safety of the multi-level shield tunnel are beneficial to timely finding out the problem of shield construction abnormality, can improve the monitoring level of data on a construction site, and help constructors to timely early warn construction abnormality and reasonably operate shield tunneling.

Drawings

FIG. 1 is a schematic diagram of a multi-layer shield tunnel security protection system according to an embodiment of the present invention.

Fig. 2 is a schematic diagram of a center-jump pulse according to an embodiment of the invention.

FIG. 3 is a flowchart of a data driven anomaly diagnosis method according to an embodiment of the present invention.

Fig. 4 is a schematic diagram showing the composition of a data-driven abnormality diagnosis unit according to an embodiment of the present invention.

Detailed Description

Preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

For a further understanding of the present invention, preferred embodiments of the invention are described below in conjunction with the examples, but it should be understood that these descriptions are merely intended to illustrate further features and advantages of the invention, and are not limiting of the claims of the invention.

The description of this section is intended to be illustrative of only a few exemplary embodiments and the invention is not to be limited in scope by the description of the embodiments. It is also within the scope of the description and claims of the invention to interchange some of the technical features of the embodiments with other technical features of the same or similar prior art.

The description of the steps in the various embodiments in the specification is merely for convenience of description, and the implementation of the present application is not limited by the order in which the steps are implemented. "connected" in the specification includes both direct and indirect connections.

The invention discloses a multi-level shield tunnel safety protection system, and FIG. 1 is a schematic diagram of the composition of the multi-level shield tunnel safety protection system in an embodiment of the invention; referring to fig. 1, the tunnel security protection system includes: the system comprises a data acquisition module 1, an abnormality diagnosis module 2, a typical characteristic comparison module 3 and an alarm result output module 4.

The data acquisition module 1 is used for acquiring data of a set sensor, and the acquired data comprises real-time shield construction parameters.

The abnormality diagnosis module 2 is used for identifying abnormality from the selected real-time shield construction parameters; constructing a subsequence data set by data cleaning and time delay method on shield tail grease pressure data in a section of normal construction mode, inputting the subsequence data set into a mode learning model, and learning a normal change mode of shield tail grease pressure; then updating real-time data, and identifying whether the change mode of the latest data section is changed beyond the set size according to the residual values of the actual grease pressure recursion diagram matrix and the predicted grease pressure recursion diagram matrix; if the residual value is higher than a threshold value set under engineering experience, the tail oil pressure is considered to have abnormal mode.

The typical feature comparison module 3 is configured to extract a feature index and a threshold from a typical case of a tail grease leakage event, so as to match whether a data change feature in tail grease pressure data accords with an abnormal typical feature of the tail grease leakage event. The alarm result output module 4 is used for alarming the abnormality.

In an embodiment of the present invention, the data acquired by the data acquisition module 1 includes a tail grease pressure, and the construction parameters associated with the data include a tail grease injection amount, a grease setting value and a tail gap. The data acquisition module 1 comprises a real-time construction parameter extraction unit which is used for selecting common construction parameters related to the abnormality of the shield machine from construction experience, extracting selected construction parameter data from shield real-time data, preprocessing time sequence data, and inputting the time sequence data into a pattern learning model for preliminary abnormality identification.

In an embodiment of the present invention, the exemplary feature comparing module 3 includes: the typical feature extraction unit 301 and the typical feature recognition unit 302. The typical feature extraction unit 301 is configured to extract typical features from the tail oil pressure and the oil injection amount, including an abrupt proportion index of the oil pressure and the oil injection amount; after the thresholds under the construction parameter indexes are determined, the indexes are used as visual basis for abnormality identification. The typical feature recognition unit 302 is configured to calculate a feature index value related to an abnormal event for a tail grease leakage event from the preprocessed core parameter time sequence data; when the index value exceeds the set threshold, it is considered that there may be a tail grease leakage event.

In one embodiment of the invention, four safety protection methods aiming at different abnormal event types are provided for detecting the abnormal events which easily occur inside and outside the shield machine, and the safety protection methods are applied to a first diagnosis layer of a safety level to primarily diagnose the abnormality of real-time data.

With continued reference to fig. 1, the abnormality diagnosis module 2 includes at least one of a threshold alarm unit 201, a heartbeat fault alarm unit 202, a control deviation diagnosis unit 203, and a data-driven abnormality diagnosis unit 204.

The threshold value alarm unit 201 is used for counting the distribution of recent sensing data of the shield tail grease pressure, determining the change range of the normal working state of the equipment after data denoising, and setting upper and lower threshold values; when the value recorded in real time by the sensor is lower than the lower limit or higher than the upper limit, the device is considered to be abnormal at the moment.

In an embodiment, the threshold value alarm unit 201 is used for identifying individual equipment faults occurring inside the shield tunneling machine; according to the method, distribution of recent sensing data D of target monitoring equipment is counted, the change range D _l,d_h of the normal working state of the equipment is determined after data denoising is carried out, and an upper-lower threshold D _l,d_h is set; when the value d _r recorded by the sensor in real time is lower than the lower limit d _l or higher than the upper limit d _h, the equipment is considered to be abnormal at the moment; the method can simply and directly identify abnormal values of real-time data of the equipment and output equipment fault early warning.

The heartbeat fault alarm unit 202 is used for finding out abnormal communication between the shield machine and an external database; the basic form of the heartbeat protocol is: the shield PLC control system regularly transmits pulse signals at a certain frequency (fig. 2 is a schematic diagram of center jump pulse in an embodiment of the invention, refer to fig. 2), and meanwhile, the intelligent tunneling control system receives and checks the continuity of heartbeat; and the heartbeat detection rule is that if the current time and the last time the intelligent tunneling control system receives the heartbeat exceed the set timeout, the timeout is obtained, and the heartbeat failure of the shield PLC control system is judged, namely the communication failure of the shield is judged.

The control deviation diagnosis unit 203 is used for identifying the control risk caused by the model defect in the shield posture. The control indexes of the shield posture comprise two horizontal deviation indexes (left and right) and two elevation deviation indexes (up and down), real four control index values (the four control index values comprise horizontal deviation left b _left, horizontal deviation right b _right, elevation deviation upper b _high and elevation deviation lower b _low) are recorded in real-time monitoring data of the control indexes in the data layer, and the output values of the posture control model to the current posture control parameters of the shield are output in real-time control output parameters of the model (the output values comprise horizontal deviation left b '_left, horizontal deviation right b' _right, elevation deviation upper b '_high and elevation deviation lower b' _low); if the deviation between the model output value and the actual control index is overlarge, the control parameter value output by the model is considered to be abnormal; therefore, the difference between the model output value of each index and the actual index is calculated, and the calculation method of the difference score _m is shown in formula (1):

score _m＝|b'_m-b_m×(1±k_m) I, m ε { left, right, low, high }, k ε [0%,100% ] formula (1)

Wherein m represents deviation indexes in different directions, b _m and b' _m represent control index values and model output recommended values of the current shield, and k _m represents a permissible floating range of a corresponding control value; meanwhile, a threshold s _m is set for the difference score _m; when score _m is larger than threshold s _m, the control parameters output by the model are considered to deviate from the current shield posture, and the model has defects and needs to be adjusted;

The data-driven abnormality diagnosis unit 204 is configured to obtain construction parameters related to common abnormal events from real-time shield data, and then learn a change pattern of parameters in a normal working state from a time sequence formed by the construction parameters within a period of time, and identify whether the change pattern of the latest data segment has changed greatly; if a large change occurs, it is considered that there is a pattern abnormality.

FIG. 3 is a flowchart of a data-driven abnormality diagnosis method according to an embodiment of the present invention, and FIG. 4 is a schematic diagram illustrating a composition of a data-driven abnormality diagnosis unit according to an embodiment of the present invention; referring to fig. 3 and 4, in an embodiment of the invention, the data-driven abnormality diagnosis unit 204 includes: a real-time construction parameter extraction unit 2041, a pattern learning model construction unit 2042, and an abnormal pattern recognition unit 2043; the real-time construction parameter extraction unit 2041 includes a time series data cleaning unit 2044 and a time window sequence construction unit 2045.

The time sequence data cleaning unit 2044 is used for supplementing a null value with a corresponding statistical value according to the actual distribution condition of the selected construction parameter part time point when the data of the selected construction parameter part time point is missing; the time series data cleaning unit 205 is configured to delete the repeated portion when the data of the selected construction parameter portion time point is repeated; the time sequence data cleaning unit is used for considering the normal change interval and the construction state of the parameters when the data of the selected construction parameter part time points have zero values; selecting construction parameters which possibly have 0 values, not processing the construction parameters, and processing the construction parameters which are not zeroed according to missing values if the 0 values are sporadic, wherein the sensor is considered to be possibly abnormal in recording; if the 0 value appears repeatedly or continuously, considering that equipment abnormality or other shield abnormality time possibly exists, temporarily retaining the abnormal value;

In order to learn the changing mode of the shield parameter time sequence data, the method needs to compare the front time sequence data segment with the rear time sequence data segment. Therefore, a plurality of time windows with equal length are intercepted from the original shield time sequence data through the time window sequence constructing unit 2045, namely, the original data set X= { X ₁,x₂,...,x_i,...,x_n } is divided into a plurality of time subsequences by adopting a time delay method, wherein X _i＝{x_i,x_i+τ,...,x_i+(m-1)τ},X_i represents an ith time sequence segment, X _i is an ith observed sample, m is an embedding dimension, and τ is a time delay factor; the obtained time window sequence data set X ₁,X₂,...,X_N represents the shield construction state at different moments, wherein N satisfies N=n- (m-1) tau;

The pattern learning model construction unit 2042 is used for constructing a pattern learning module; the constructed model learning model is divided into two parts, namely construction parameter characteristic extraction and normal construction model self-learning;

the construction parameter characteristic extraction method is threshold-free recursion diagram conversion; the thresholdless recursive graph conversion method calculates the distance of the data between the time points in each time window as shown in formula (2):

R _i,j＝||X_i-X_j i, j=1, 2,..n' formula (2)

Wherein, I and I are distance norms, and N' is the number of time points contained in the time window; converting the shield construction data from a time sequence into a two-dimensional symmetric matrix, wherein each point in the matrix represents the distance between any two time points, so that the dynamic characteristics of the shield data on time sequence are highlighted;

the method adopted by the self-learning of the normal construction mode is an extreme learning machine self-encoder. After the non-threshold recursion diagram is converted into a recursion diagram matrix of a real-time construction parameter time subsequence, the subsequence recursion matrix is used as input and output of an extreme learning machine at the same time, and an output layer parameter set value is calculated through a random input layer parameter, so that a model reconstruction parameter is obtained; the method comprises the following specific steps:

(1) Setting the self-encoder of the extreme learning machine as a single-layer neural network of M input and output neurons and L hidden-layer neurons; constructing an input P and an output Q with M consecutive real-time construction parameter recursion matrices, wherein p=q= { R ₁,R₂,...,R_M }, the training samples are therefore M is less than N;

(2) Randomly distributing orthogonalization input weights a _i and deviation b _i based on a specific continuous probability distribution function, and solving an hidden layer H (P) = [ H ₁(p),...,h_L (P) ] corresponding to an input layer P through an activation function H (·), wherein the calculation mode of neurons of each hidden layer is shown in a formula (3);

(3) After calculating hidden layer neurons, under the condition of knowing the output Q, the optimal output weight which minimizes the residual error can be obtained through a formula (4) At this time a, b,The formed model parameter set is the result of model self-learning, namely the changing mode of the shield real-time parameter in the construction period;

The abnormal pattern recognition unit 2043 is used to recognize an abnormal pattern after the normal construction pattern learning is completed; acquiring the latest construction parameters from the shield real-time data, and forming the latest time subsequence X _N+1 to be checked; after conversion to the thresholdless recursive matrix R _N+1, input a _i,b_i, In an extreme learning machine network formed by P '(P' =R _N+1), outputting a reconstruction result in a normal construction state by using a formula (5);

Where H (a, b, P ') represents the hidden layer neurons of the current matrix and Q' is the reconstructed thresholdless recursive matrix. In order to calculate the residual errors of the actual matrix P 'and the predicted matrix Q', the method adopts radial basis function calculation, and omega _N+1 represents the residual errors of the two matrices, namely the abnormal fraction of the time subsequence of the (n+1) th construction parameter, as shown in a formula (6);

The threshold delta of the anomaly score can be set by combining the change characteristics of different construction parameters and the actual construction requirements, and when omega is more than or equal to delta, the current shield construction is considered to have the anomaly related to the construction parameters.

The invention also discloses a multi-layer shield tunnel safety protection method, which comprises the following steps:

step S1, acquiring data of a setting sensor, wherein the acquired data comprise real-time shield construction parameters;

In an embodiment of the present invention, the acquired data includes a shield tail grease pressure, and the construction parameters associated therewith further include a shield tail grease injection amount, a grease set point, and a shield tail gap.

The data acquisition step comprises a real-time construction parameter extraction step, wherein common construction parameters related to the shield machine abnormality are selected from construction experience, the selected construction parameter data are extracted from shield real-time data, and the time sequence data are preprocessed and input into a mode learning model for preliminary abnormality identification.

Step S2, abnormality diagnosis, namely identifying abnormality from the selected real-time shield construction parameters; constructing a subsequence data set by data cleaning and time delay method on shield tail grease pressure data in a section of normal construction mode, inputting the subsequence data set into a mode learning model, and learning a normal change mode of shield tail grease pressure; then updating real-time data, and identifying whether the change mode of the latest data section is changed beyond the set size according to the residual values of the actual grease pressure recursion diagram matrix and the predicted grease pressure recursion diagram matrix; if the residual value is higher than a threshold value set under engineering experience, the tail oil pressure is considered to have abnormal mode at the moment;

And (3) a step of comparing the typical characteristics, namely extracting characteristic indexes and threshold values from typical cases of the tail grease leakage event, so as to match whether the data change characteristics in the tail grease pressure data accord with the abnormal typical characteristics of the tail grease leakage event.

In one embodiment of the present invention, the exemplary feature comparison step includes:

Extracting typical characteristics from the tail oil pressure and the oil injection quantity, wherein the typical characteristics comprise abrupt proportion indexes of the oil pressure and the oil injection quantity; after determining the threshold value under each construction parameter index, taking the indexes as visual basis for abnormality identification;

A typical feature identification step, namely aiming at a shield tail grease leakage event from the preprocessed core parameter time sequence data, and calculating a feature index value related to an abnormal event; when the index value exceeds a set threshold value, the tail grease leakage event is considered to be possible;

And (S4) outputting an alarm result, and alarming when abnormality is found.

In an embodiment of the present invention, the step S2 includes at least one of a threshold alarm step, a heartbeat failure alarm step, a control deviation diagnosis step, and a data-driven abnormality diagnosis step.

The threshold alarming step comprises the following steps: identifying individual equipment faults occurring in the shield tunneling machine; according to the method, distribution of recent sensing data D of target monitoring equipment is counted, the change range D _l,d_h of the normal working state of the equipment is determined after data denoising is carried out, and an upper-lower threshold D _l,d_h is set; when the value d _r recorded in real time by the sensor is below the lower bound d _l or above the upper bound d _h, the device is considered to be abnormal at this time.

The step of heartbeat fault alarming comprises the following steps: the communication abnormality between the shield machine and an external database is found; the basic form of the heartbeat protocol is: the shield PLC control system regularly transmits pulse signals at a certain frequency, and meanwhile, the intelligent tunneling control system receives and checks the continuity of heartbeat; the heartbeat detection rule is that if the current time and the last time the intelligent tunneling control system receives the heartbeat exceed the set timeout, the timeout is reached, and the heartbeat failure of the shield PLC control system is judged, namely the shield has communication faults;

The control deviation diagnosing step includes: and identifying the control risk caused by the model defect in the shield posture. The control indexes of the shield posture comprise two horizontal deviation indexes (left and right) and two elevation deviation indexes (up and down), real four control index values (the four control index values comprise horizontal deviation left b _left, horizontal deviation right b _right, elevation deviation upper b _high and elevation deviation lower b _low) are recorded in real-time monitoring data of the control indexes in the data layer, and the output values of the posture control model to the current posture control parameters of the shield are output in real-time control output parameters of the model (the output values comprise horizontal deviation left b '_left, horizontal deviation right b' _right, elevation deviation upper b '_high and elevation deviation lower b' _low); if the deviation between the model output value and the actual control index is overlarge, the control parameter value output by the model is considered to be abnormal; therefore, the difference between the model output value of each index and the actual index is calculated, and the calculation method of the difference score _m is shown in formula (1):

score _m＝|b'_m-b_m×(1±k_m) I, m ε { left, right, low, high }, k ε [0%,100% ] formula (1)

Wherein m represents deviation indexes in different directions, b _m and b' _m represent control index values and model output recommended values of the current shield, and k _m represents a permissible floating range of a corresponding control value; meanwhile, a threshold s _m is set for the difference score _m; when score _m is larger than threshold s _m, the control parameters output by the model are considered to deviate from the current shield posture, and the model has defects and needs to be adjusted;

The data-driven abnormality diagnosis step includes: acquiring construction parameters related to common abnormal events from real-time shield data, then learning a change mode of the parameters in a normal working state from a time sequence formed by the construction parameters within a period of time, and identifying whether the change mode of the latest data section is greatly changed; if a large change occurs, it is considered that there is a pattern abnormality.

Referring to fig. 3, in an embodiment of the present invention, the data-driven anomaly diagnosis step includes: a real-time construction parameter extraction step, a pattern learning model construction step and an abnormal pattern recognition step. The real-time construction parameter extraction step comprises a time sequence data cleaning step and a time window sequence construction step.

The time sequence data cleaning step comprises the following steps: when the data of the selected construction parameter part time point is missing, the null value is complemented by the corresponding statistical value according to the actual distribution condition; deleting the repeated part when the data of the selected construction parameter part time point is repeated; when the data of the selected construction parameter part time points have zero values, the normal change interval and the construction state of the parameters are considered; selecting construction parameters which possibly have 0 values, not processing the construction parameters, and processing the construction parameters which are not zeroed according to missing values if the 0 values are sporadic, wherein the sensor is considered to be possibly abnormal in recording; if the 0 value appears repeatedly or continuously, considering that equipment abnormality or other shield abnormality time possibly exists, temporarily retaining the abnormal value;

The time window sequence construction step includes: intercepting a plurality of equal-length time windows from original shield time sequence data, namely dividing an original data set X= { X ₁,x₂,...,x_i,...,x_n } into a plurality of time subsequences by adopting a time delay method, wherein X _i＝{x_i,x_i+τ,...,x_i+(m-1)τ}, X_i represents an ith time sequence segment, X _i is an ith observed sample, m is an embedded dimension, and tau is a time delay factor; the obtained time window sequence data set X ₁,X₂,...,X_N represents the shield construction state at different moments, wherein N satisfies N=n- (m-1) tau;

the pattern learning model construction step comprises constructing a pattern learning module; the constructed model learning model is divided into two parts, namely construction parameter characteristic extraction and normal construction model self-learning;

the construction parameter characteristic extraction method is threshold-free recursion diagram conversion; the thresholdless recursive graph conversion method calculates the distance of the data between the time points in each time window as shown in formula (2):

R _i,j＝||X_i-X_j i, j=1, 2,..n' formula (2)

Wherein, I and I are distance norms, and N' is the number of time points contained in the time window; converting the shield construction data from a time sequence into a two-dimensional symmetric matrix, wherein each point in the matrix represents the distance between any two time points, so that the dynamic characteristics of the shield data on time sequence are highlighted;

The method adopted by the self-learning of the normal construction mode is an extreme learning machine self-encoder. After the non-threshold recursion diagram is converted into a recursion diagram matrix of a real-time construction parameter time subsequence, the subsequence recursion matrix is used as input and output of an extreme learning machine, and an output layer parameter set value is calculated through random input layer parameters, so that parameters of model reconstruction are obtained. The method comprises the following specific steps:

(1) Setting the self-encoder of the extreme learning machine as a single-layer neural network of M input and output neurons and L hidden-layer neurons; constructing an input P and an output Q with M consecutive real-time construction parameter recursion matrices, wherein p=q= { R ₁,R₂,...,R_M }, the training samples are therefore M is less than N;

(2) Randomly distributing orthogonalization input weights a _i and deviation b _i based on a specific continuous probability distribution function, and solving an hidden layer H (P) = [ H ₁(p),...,h_L (P) ] corresponding to an input layer P through an activation function H (·), wherein the calculation mode of neurons of each hidden layer is shown in a formula (3);

(3) After calculating hidden layer neurons, under the condition of knowing the output Q, the optimal output weight which minimizes the residual error can be obtained through a formula (4) At this time a, b,The formed model parameter set is the result of model self-learning, namely the changing mode of the shield real-time parameter in the construction period; /(I)

The abnormal pattern recognition step includes: identifying an abnormal mode after the normal construction mode learning is finished; acquiring the latest construction parameters from the shield real-time data, and forming the latest time subsequence X _N+1 to be checked; after conversion to the thresholdless recursive matrix R _N+1, input a _i,b_i,In an extreme learning machine network formed by P '(P' =R _N+1), outputting a reconstruction result in a normal construction state by using a formula (5);

Wherein H (a, b, P ') represents hidden layer neurons of the current matrix, Q' is a reconstructed thresholdless recursive matrix; in order to calculate the residual errors of the actual matrix P 'and the predicted matrix Q', the method adopts radial basis function calculation, and omega _N+1 represents the residual errors of the two matrices, namely the abnormal fraction of the time subsequence of the (n+1) th construction parameter, as shown in a formula (6);

The threshold delta of the anomaly score can be set by combining the change characteristics of different construction parameters and the actual construction requirements, and when omega is more than or equal to delta, the current shield construction is considered to have the anomaly related to the construction parameters.

In one use scenario of the present invention, the shield tunnel security hierarchy employed by the present invention includes four security levels. The system monitors the real-time state of the shield machine based on the real-time data of the shield, the output of the real-time control of the model and the real-time monitoring data of the control index, and can find various complex abnormal problems including equipment faults, communication faults, control model problems, working condition mutation risks and the like after the diagnosis of the initial diagnosis and the screening of two layers; and meanwhile, the normal change under the geological change can be eliminated, or an alarm is given to the unknown abnormality, and the shutdown processing is reserved. The following will describe a procedure for diagnosing various abnormal problems in the shield tunnel security level.

Equipment failure

The equipment fault diagnosis mainly acquires key information from shield real-time data in a data layer, including sensing data of monitoring equipment and related facilities. The real-time sensing data of the equipment adopts two methods in a primary diagnosis layer to identify the abnormality: threshold alarms and data driven anomaly diagnostics. The threshold alarm can simply and directly identify abnormal values of real-time data of the equipment and output equipment fault early warning by judging whether the sensing data of the single equipment exceeds the normal range set by the threshold. The data-driven abnormality diagnosis method also identifies abnormalities from shield real-time data. By learning the variation pattern of the relevant parameter in the normal range, an abnormal variation pattern that is greatly different from the normal pattern can be identified. The data segment of the pattern exception will pass into the screening layer in an attempt to match the typical characteristics of the partially known device failure type. When the characteristics of the data segment conform to the typical characteristics of equipment faults, the abnormal change mode can be considered to be the equipment faults and is input into a result layer, and a certain equipment in the shield is warned of faults and is reminded of carrying out corresponding equipment fault processing.

Communication failure

The communication fault diagnosis obtains real-time information from shield real-time data of a data layer and real-time control output parameters of a model, and identifies abnormality through a heartbeat pulse alarm method in a primary diagnosis layer. The heartbeat pulse alarming method can monitor the continuity of the shield real-time data and the model real-time control output parameters, and if the updating of the real-time data is discontinuous or interrupted, the method can be used for initially diagnosing as abnormality and alarming. The abnormal heartbeat pulse signals found in the heartbeat pulse alarming method are transmitted into the result layer, and an alarm that communication faults exist is sent out and corresponding communication fault processing measures are prompted to be adopted.

Model defect

The detection of model defects firstly obtains real-time information from real-time monitoring of model real-time control output parameters and control indexes of a data layer, and identifies abnormality through a trend deviation alarm method in a first diagnosis layer. The trend deviation alarm method can judge the deviation between the real-time control output parameters of the model and the real-time monitoring of the control indexes, and if the deviation of the control parameters is larger, the model can be initially diagnosed as abnormal and alarm. And the alarm found in the trend deviation alarm method is transmitted into a result layer to remind that the model defect exists and the model is reserved for subsequent adjustment.

Risk of working condition

The identification of the working condition risk acquires real-time information from shield real-time data of the data layer, and an abnormality diagnosis method based on data driving in the initial diagnosis layer is adopted to identify the abnormality. The data-driven anomaly diagnostic method identifies an anomaly pattern of change of the associated parameter and attempts to match typical characteristics of known operating condition risks in the screening layer. When a large probability match can be made to the feature, it can be considered that the abnormal pattern of change may be a risk of the condition. Can be initially diagnosed as abnormal and alerted. And transmitting the matched abnormal alarm to a result layer, confirming that the working condition risk exists and reminding to take corresponding working condition risk treatment measures.

Normal construction and unknown abnormality

Data-driven anomaly diagnosis of the shield real-time data input primary diagnostic layer based on the data layer may also identify pattern anomalies that cannot match known typical operating condition risks and known typical equipment faults. The security level thus involves the flow of handling normal mode and unknown anomalies. When the abnormality cannot match the known typical working condition risk and the known typical equipment fault, the typical characteristics of the known geological change can be matched, and the normal real-time data change caused by the geological change can be found and marked; if the typical features cannot be matched with the anomalies, the typical features are used as unknown anomalies to alarm, and enter a result layer to wait for shutdown inspection.

In summary, the system and the method for protecting the safety of the multi-level shield tunnel are beneficial to timely finding out the problem of shield construction abnormality, can improve the monitoring level of data on a construction site, and help constructors to timely early warn construction abnormality and reasonably operate shield tunneling.

It should be noted that the present application may be implemented in software and/or a combination of software and hardware; for example, an Application Specific Integrated Circuit (ASIC), a general purpose computer, or any other similar hardware device may be employed. In some embodiments, the software program of the present application may be executed by a processor to implement the above steps or functions. Likewise, the software program of the present application (including the related data structures) may be stored in a computer-readable recording medium; such as RAM memory, magnetic or optical drives or diskettes, and the like. In addition, some steps or functions of the present application may be implemented in hardware; for example, as circuitry that cooperates with the processor to perform various steps or functions.

The technical features of the above-described embodiments may be arbitrarily combined, and all possible combinations of the technical features in the above-described embodiments are not described for brevity of description, however, as long as there is no contradiction between the combinations of the technical features, they should be considered as the scope of the description.

The description and applications of the present invention herein are illustrative and are not intended to limit the scope of the invention to the embodiments described above. Effects or advantages referred to in the embodiments may not be embodied in the embodiments due to interference of various factors, and description of the effects or advantages is not intended to limit the embodiments. Variations and modifications of the embodiments disclosed herein are possible, and alternatives and equivalents of the various components of the embodiments are known to those of ordinary skill in the art. It will be clear to those skilled in the art that the present invention may be embodied in other forms, structures, arrangements, proportions, and with other assemblies, materials, and components, without departing from the spirit or essential characteristics thereof. Other variations and modifications of the embodiments disclosed herein may be made without departing from the scope and spirit of the invention.

Claims (8)

1. A multi-level shield tunnel security protection system, the tunnel security protection system comprising:

The data acquisition module is used for acquiring data of the setting sensor, and the acquired data comprise real-time shield construction parameters;

The abnormality diagnosis module is used for identifying abnormality from the selected real-time shield construction parameters;

The typical characteristic comparison module is used for extracting characteristic indexes and threshold values from typical cases of the tail grease leakage event so as to match whether the data change characteristics in the tail grease pressure data accord with abnormal typical characteristics of the tail grease leakage event or not;

the alarm result output module is used for alarming after abnormality is found;

The abnormality diagnosis module comprises at least one of a threshold value alarm unit, a heartbeat fault alarm unit, a control deviation diagnosis unit and a data-driven abnormality diagnosis unit:

The threshold alarming unit is used for identifying individual equipment faults occurring in the shield tunneling machine; determining the change range D _l,d_h of the normal working state of the equipment after denoising the data by counting the distribution of the recent sensing data D of the target monitoring equipment, and setting an upper and lower threshold D _l,d_h; when the value d _r recorded by the sensor in real time is lower than the lower limit d _l or higher than the upper limit d _h, the equipment is considered to be abnormal at the moment;

The heartbeat fault alarm unit is used for finding out abnormal communication between the shield machine and an external database; the basic form of the heartbeat protocol is: the shield PLC control system regularly transmits pulse signals at a certain frequency, and meanwhile, the intelligent tunneling control system receives and checks the continuity of heartbeat; the heartbeat detection rule is that if the current time and the last time the intelligent tunneling control system receives the heartbeat exceed the set timeout, the timeout is reached, and the heartbeat failure of the shield PLC control system is judged, namely the shield has communication faults;

The control deviation diagnosis unit is used for identifying control risks caused by model defects in the shield posture; the control indexes of the shield posture comprise two horizontal deviation indexes and two elevation deviation indexes, real-time monitoring data of the control indexes in the data layer record actual four control index values, and the four control index values comprise: left horizontal deviation b _left, right horizontal deviation b _right, upper elevation deviation b _high and lower elevation deviation b _low; the model real-time control output parameters are model output values of the attitude control model on current attitude control parameters of the shield, wherein the model output values comprise a horizontal deviation left b '_left, a horizontal deviation right b' _right, an elevation deviation upper b '_high and an elevation deviation lower b' _low; if the deviation between the model output value and the actual control index is too large, the model output value is considered to be abnormal; therefore, the difference between the model output value of each control index and the actual index is calculated, and the calculation method of the difference score _m is shown in the formula (1):

score _m＝|b'_m-b_m×(1±k_m) I, m ε { left, right, low, high }, k ε [0%,100% ] formula (1)

Wherein m represents deviation indexes in different directions, b _m and b' _m represent control index values and model output recommended values of the current shield, and k _m represents a permissible floating range of a corresponding control value; meanwhile, a threshold s _m is set for the difference score _m; when score _m is larger than threshold s _m, the control parameters output by the model are considered to deviate from the current shield posture, and the model has defects and needs to be adjusted;

The data-driven abnormality diagnosis unit is used for acquiring construction parameters related to common abnormal events from real-time shield data, then learning a change mode of parameters in a normal working state from a time sequence formed by the construction parameters within a period of time, and identifying whether the change mode of the latest data section is changed in a setting manner; if the setting is changed, the mode is considered to be abnormal; the abnormality diagnosis module is used for constructing a subsequence data set from the shield tail grease pressure data in a section of normal construction mode through a data cleaning and time delay method and inputting the subsequence data set into a mode learning model to learn a normal change mode of the shield tail grease pressure; then updating real-time data, and identifying whether the change mode of the latest data section is changed beyond the set size according to the residual values of the actual grease pressure recursion diagram matrix and the predicted grease pressure recursion diagram matrix; if the residual value is higher than a threshold value set under engineering experience, the tail oil pressure is considered to have abnormal mode.

2. The multi-level shield tunnel safety protection system according to claim 1, wherein:

the data acquired by the data acquisition module comprises shield tail grease pressure, and construction parameters related to the shield tail grease pressure comprise shield tail grease injection quantity, grease set values and shield tail gaps;

The data acquisition module comprises: the real-time construction parameter extraction unit is used for selecting common construction parameters related to the abnormality of the shield machine from construction experience, extracting selected construction parameter data from shield real-time data, preprocessing time sequence data, and inputting the preprocessed time sequence data into the pattern learning model for preliminary abnormality identification.

3. The multi-level shield tunnel safety protection system according to claim 1, wherein:

The representative feature comparison module comprises:

The typical characteristic extraction unit is used for extracting typical characteristics from the tail oil pressure and the oil injection quantity, including abrupt proportion indexes of the oil pressure and the oil injection quantity; after determining the threshold value under each construction parameter index, taking the indexes as visual basis for abnormality identification;

The typical characteristic identification unit is used for aiming at the shield tail grease leakage event from the preprocessed core parameter time sequence data and calculating characteristic index values related to the abnormal event; when the index value exceeds the set threshold, it is considered that there may be a tail grease leakage event.

4. The multi-level shield tunnel safety protection system according to claim 1, wherein:

The data-driven abnormality diagnosis unit includes: the system comprises a real-time construction parameter extraction unit, a mode learning model construction unit and an abnormal mode identification unit; the real-time construction parameter extraction unit comprises a time sequence data cleaning unit and a time window sequence construction unit;

The time sequence data cleaning unit is used for supplementing the null value with the corresponding statistical value according to the actual distribution condition when the data of the selected construction parameter part time point is missing; the time sequence data cleaning unit is used for deleting repeated parts when the data of the selected construction parameter part time points are repeated; the time sequence data cleaning unit is used for considering the normal change interval and the construction state of the parameters when the data of the selected construction parameter part time points have zero values; selecting construction parameters which possibly have 0 values, not processing the construction parameters, and processing the construction parameters which are not zeroed according to missing values if the 0 values are sporadic, wherein the sensor is considered to be possibly abnormal in recording; if the 0 value appears repeatedly or continuously, considering that equipment abnormality or other shield abnormality time possibly exists, temporarily retaining the abnormal value;

The time window sequence constructing unit is used for intercepting a plurality of time windows with equal length from the original shield time sequence data, namely, a time delay method is adopted to divide an original data set X= { X ₁,x₂,...,x_i,...,x_n } into a plurality of time subsequences, wherein X _i＝{x_i,x_i+τ,...,x_i+(m-1)τ},X_i represents an ith time sequence segment, X _i is an ith observed sample, m is an embedded dimension, and tau is a time delay factor; the obtained time window sequence data set X ₁,X₂,...,X_N represents the shield construction state at different moments, wherein N satisfies N=n- (m-1) tau;

the pattern learning model construction unit is used for constructing a pattern learning module; the constructed model learning model is divided into two parts, namely construction parameter characteristic extraction and normal construction model self-learning;

the construction parameter characteristic extraction method is threshold-free recursion diagram conversion; the thresholdless recursive graph conversion method calculates the distance of the data between the time points in each time window as shown in formula (2):

R _i,j＝||X_i-X_j i, j=1, 2,..n' formula (2)

Wherein, I and I are distance norms, and N' is the number of time points contained in the time window; converting the shield construction data from a time sequence into a two-dimensional symmetric matrix, wherein each point in the matrix represents the distance between any two time points, so that the dynamic characteristics of the shield data on time sequence are highlighted;

The method adopted by the self-learning of the normal construction mode is an extreme learning machine self-encoder; after the non-threshold recursion diagram is converted into a recursion diagram matrix of a real-time construction parameter time subsequence, the subsequence recursion matrix is used as input and output of an extreme learning machine at the same time, and an output layer parameter set value is calculated through random input layer parameters, so that parameters of model reconstruction are obtained; the method comprises the following specific steps:

(1) Setting the self-encoder of the extreme learning machine as a single-layer neural network of M input and output neurons and L hidden-layer neurons; constructing an input P and an output Q with M consecutive real-time construction parameter recursion matrices, wherein p=q= { R ₁,R₂,...,R_M }, the training samples are therefore M is less than N;

(2) Randomly distributing orthogonalization input weights a _i and deviation b _i based on a specific continuous probability distribution function, and solving an hidden layer H (P) = [ H ₁(p),...,h_L (P) ] corresponding to an input layer P through an activation function H (·), wherein the calculation mode of neurons of each hidden layer is shown in a formula (3);

(3) After calculating hidden layer neurons, under the condition of knowing the output Q, the optimal output weight which minimizes the residual error can be obtained through a formula (4) At this timeThe formed model parameter set is the result of model self-learning, namely the changing mode of the shield real-time parameter in the construction period;

The abnormal pattern recognition unit is used for recognizing an abnormal pattern after the normal construction pattern learning is finished; acquiring the latest construction parameters from the shield real-time data, and forming the latest time subsequence X _N+1 to be checked; after conversion to the thresholdless recursive matrix R _N+1, input In the formed extreme learning machine network (P' =r _N+1), and outputting a reconstruction result in a normal construction state by using a formula (5);

Wherein H (a, b, P ') represents hidden layer neurons of the current matrix, Q' is a reconstructed thresholdless recursive matrix; in order to calculate the residual errors of the actual matrix P 'and the predicted matrix Q', the method adopts radial basis function calculation, and omega _N+1 represents the residual errors of the two matrices, namely the abnormal fraction of the time subsequence of the (n+1) th construction parameter, as shown in a formula (6);

The threshold delta of the anomaly score can be set by combining the change characteristics of different construction parameters and the actual construction requirements, and when omega is more than or equal to delta, the current shield construction is considered to have the anomaly related to the construction parameters.

5. The multi-layer shield tunnel safety protection method is characterized by comprising the following steps of:

A data acquisition step of acquiring data of a setting sensor, wherein the acquired data comprises real-time shield construction parameters;

an abnormality diagnosis step of identifying an abnormality from the selected real-time shield construction parameters;

A typical characteristic comparison step, namely extracting characteristic indexes and threshold values from typical cases of shield tail grease leakage events, so as to match whether data change characteristics in shield tail grease pressure data accord with abnormal typical characteristics of the shield tail grease leakage events or not;

the abnormality diagnosis step comprises at least one of a threshold value alarming step, a heartbeat fault alarming step, a control deviation diagnosis step and a data-driven abnormality diagnosis step:

The threshold alarming step comprises the following steps: identifying individual equipment faults occurring in the shield tunneling machine; according to the method, distribution of recent sensing data D of target monitoring equipment is counted, the change range D _l,d_h of the normal working state of the equipment is determined after data denoising is carried out, and an upper-lower threshold D _l,d_h is set; when the value d _r recorded by the sensor in real time is lower than the lower limit d _l or higher than the upper limit d _h, the equipment is considered to be abnormal at the moment;

The step of heartbeat fault alarming comprises the following steps: the communication abnormality between the shield machine and an external database is found; the basic form of the heartbeat protocol is: the shield PLC control system regularly transmits pulse signals at a certain frequency, and meanwhile, the intelligent tunneling control system receives and checks the continuity of heartbeat; the heartbeat detection rule is that if the current time and the last time the intelligent tunneling control system receives the heartbeat exceed the set timeout, the timeout is reached, and the heartbeat failure of the shield PLC control system is judged, namely the shield has communication faults;

The control deviation diagnosing step includes: identifying control risks caused by model defects in the shield posture; the control indexes of the shield posture comprise two horizontal deviation indexes and two elevation deviation indexes, real-time monitoring data of the control indexes in the data layer record actual four control index values, and the four control index values comprise: left horizontal deviation b _left, right horizontal deviation b _right, upper elevation deviation b _high and lower elevation deviation b _low; the model real-time control output parameters are output values of the attitude control model on current attitude control parameters of the shield, wherein the output values comprise a horizontal deviation left b '_left, a horizontal deviation right b' _right, an elevation deviation upper b '_high and an elevation deviation lower b' _low; if the deviation between the model output value and the actual control index is overlarge, the control parameter value output by the model is considered to be abnormal; therefore, the difference between the model output value of each index and the actual index is calculated, and the calculation method of the difference score _m is shown in formula (1):

score _m＝|b'_m-b_m×(1±k_m) I, m ε { left, right, low, high }, k ε [0%,100% ] formula (1)

Wherein m represents deviation indexes in different directions, b _m and b' _m represent control index values and model output recommended values of the current shield, and k _m represents a permissible floating range of a corresponding control value; meanwhile, a threshold s _m is set for the difference score _m; when score _m is larger than threshold s _m, the control parameters output by the model are considered to deviate from the current shield posture, and the model has defects and needs to be adjusted;

The data-driven abnormality diagnosis step includes: acquiring construction parameters related to common abnormal events from real-time shield data, then learning a change mode of the parameters in a normal working state from a time sequence formed by the construction parameters within a period of time, and identifying whether the change mode of the latest data section is changed in a setting manner; if the setting is changed, the mode is considered to be abnormal; in the abnormality diagnosis step, a sub-sequence data set is constructed by a data cleaning and time delay method on the shield tail grease pressure data in a section of normal construction mode and then is input into a mode learning model, and a normal change mode of the shield tail grease pressure is learned; then updating real-time data, and identifying whether the change mode of the latest data section is changed beyond the set size according to the residual values of the actual grease pressure recursion diagram matrix and the predicted grease pressure recursion diagram matrix; if the residual value is higher than a threshold value set under engineering experience, the tail oil pressure is considered to have abnormal mode.

6. The multi-layer shield tunnel safety protection method according to claim 5, wherein the method comprises the following steps:

In the data acquisition step, acquired data comprise shield tail grease pressure, and construction parameters related to the shield tail grease pressure comprise shield tail grease injection quantity, grease set values and shield tail gaps;

The data acquisition step includes: and a real-time construction parameter extraction step, namely selecting common construction parameters related to the abnormality of the shield machine from construction experience, extracting selected construction parameter data from shield real-time data, preprocessing time sequence data, and inputting the preprocessed time sequence data into a pattern learning model for preliminarily identifying the abnormality.

7. The multi-layer shield tunnel safety protection method according to claim 5, wherein the method comprises the following steps:

the representative feature comparison step includes:

Extracting typical characteristics from the tail oil pressure and the oil injection quantity, wherein the typical characteristics comprise abrupt proportion indexes of the oil pressure and the oil injection quantity; after determining the threshold value under each construction parameter index, taking the indexes as visual basis for abnormality identification;

A typical feature identification step, namely aiming at a shield tail grease leakage event from the preprocessed core parameter time sequence data, and calculating a feature index value related to an abnormal event; when the index value exceeds a set threshold value, the tail grease leakage event is considered to be possible;

The tunnel security protection method further comprises the following steps: and outputting an alarm result, namely alarming the abnormality.

8. The multi-layer shield tunnel safety protection method according to claim 5, wherein the method comprises the following steps:

The data-driven abnormality diagnosis step includes: a real-time construction parameter extraction step, a pattern learning model construction step and an abnormal pattern recognition step; the real-time construction parameter extraction step comprises a time sequence data cleaning step and a time window sequence construction step;

The time sequence data cleaning step comprises the following steps: when the data of the selected construction parameter part time point is missing, the null value is complemented by the corresponding statistical value according to the actual distribution condition; deleting the repeated part when the data of the selected construction parameter part time point is repeated; when the data of the selected construction parameter part time points have zero values, the normal change interval and the construction state of the parameters are considered; selecting construction parameters which possibly have 0 values, not processing the construction parameters, and processing the construction parameters which are not zeroed according to missing values if the 0 values are sporadic, wherein the sensor is considered to be possibly abnormal in recording; if the 0 value appears repeatedly or continuously, considering that equipment abnormality or other shield abnormality time possibly exists, temporarily retaining the abnormal value;

The time window sequence construction step includes: intercepting a plurality of equal-length time windows from original shield time sequence data, namely dividing an original data set X= { X ₁,x₂,...,x_i,...,x_n } into a plurality of time subsequences by adopting a time delay method, wherein X _i＝{x_i,x_i+τ,...,x_i+(m-1)τ},X_i represents an ith time sequence segment, X _i is an ith observed sample, m is an embedded dimension, and tau is a time delay factor; the obtained time window sequence data set X ₁,X₂,...,X_N represents the shield construction state at different moments, wherein N satisfies N=n- (m-1) tau;

the pattern learning model construction step comprises constructing a pattern learning module; the constructed model learning model is divided into two parts, namely construction parameter characteristic extraction and normal construction model self-learning;

the construction parameter characteristic extraction method is threshold-free recursion diagram conversion; the thresholdless recursive graph conversion method calculates the distance of the data between the time points in each time window as shown in formula (2):

R _i,j＝||X_i-X_j i, j=1, 2,..n' formula (2)

Wherein, I and I are distance norms, and N' is the number of time points contained in the time window; converting the shield construction data from a time sequence into a two-dimensional symmetric matrix, wherein each point in the matrix represents the distance between any two time points, so that the dynamic characteristics of the shield data on time sequence are highlighted;

The method adopted by the self-learning of the normal construction mode is an extreme learning machine self-encoder; after the non-threshold recursion diagram is converted into a recursion diagram matrix of a real-time construction parameter time subsequence, the subsequence recursion matrix is used as input and output of an extreme learning machine at the same time, and an output layer parameter set value is calculated through a random input layer parameter, so that a model reconstruction parameter is obtained; the method comprises the following specific steps:

(1) Setting the self-encoder of the extreme learning machine as a single-layer neural network of M input and output neurons and L hidden-layer neurons; constructing an input P and an output Q with M consecutive real-time construction parameter recursion matrices, wherein p=q= { R ₁,R₂,...,R_M }, the training samples are therefore M is less than N;

(2) Randomly distributing orthogonalization input weights a _i and deviation b _i based on a specific continuous probability distribution function, and solving an hidden layer H (P) = [ H ₁(p),...,h_L (P) ] corresponding to an input layer P through an activation function H (·), wherein the calculation mode of neurons of each hidden layer is shown in a formula (3);

(3) After calculating hidden layer neurons, under the condition of knowing the output Q, the optimal output weight which minimizes the residual error can be obtained through a formula (4) At this timeThe formed model parameter set is the result of model self-learning, namely the changing mode of the shield real-time parameter in the construction period;

The abnormal pattern recognition step includes: identifying an abnormal mode after the normal construction mode learning is finished; acquiring the latest construction parameters from the shield real-time data, and forming the latest time subsequence X _N+1 to be checked; after conversion to the thresholdless recursive matrix R _N+1, input In the formed extreme learning machine network (P' =r _N+1), and outputting a reconstruction result in a normal construction state by using a formula (5);

Wherein H (a, b, P ') represents hidden layer neurons of the current matrix, Q' is a reconstructed thresholdless recursive matrix; in order to calculate the residual errors of the actual matrix P 'and the predicted matrix Q', the method adopts radial basis function calculation, and omega _N+1 represents the residual errors of the two matrices, namely the abnormal fraction of the time subsequence of the (n+1) th construction parameter, as shown in a formula (6);

The threshold delta of the anomaly score can be set by combining the change characteristics of different construction parameters and the actual construction requirements, and when omega is more than or equal to delta, the current shield construction is considered to have the anomaly related to the construction parameters.