Shield tunneling machine operation state monitoring method and system

Abstract

The invention relates to the technical field of shield tunneling machine monitoring, solves the technical problems that in the prior art, a neural network is prone to falling into local minimum, and normal monitoring analysis is affected by unstable network performance, and particularly relates to a shield tunneling machine operation state monitoring method and system. The method comprises the steps of collecting three-dimensional tensor data of a shield tunneling machine and calculating a noise reduction tensor; according to the method, by means of a group optimization mode, the situation that the learning efficiency of the neural network is seriously affected due to the fact that the connection weight and the threshold value are repeatedly adjusted is avoided, the selection efficiency of the weight and the threshold value can be greatly improved, and the method has the advantages of being high in robustness, high in robustness and the like. Moreover, the test error is smaller, the nonlinear fitting capability is better, and the classification accuracy of the working state of the shield tunneling machine can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of shield machine monitoring, and in particular to a shield machine operation status monitoring method and system. Background Art

[0002] A shield machine is a tunnel boring machine that uses the shield method. The shield construction method is that the tunnel boring machine builds a "shield" (referring to supporting segments) of the tunnel while excavating. It is different from the open construction method. In a broad sense, shield machines can also be used in rock formations. In my country, tunnel boring machines used in soft soil formations are customarily called shield machines. During the operation of the shield machine, it is necessary to monitor the operating status of the shield machine at all times to ensure that the shield machine is working normally, prevent the occurrence of operating accidents, and perform maintenance and repairs on the shield machine in a timely manner. Existing technologies generally use artificial intelligence networks to analyze and monitor the operating status of the shield machine based on sensor data. The most common one is the BP neural network. However, the BP neural network generally solves a nonlinear problem through local search during operation. The connection weights of the network are adjusted along the direction of local improvement during the training process, which makes it easy to fall into the problem of local minimum. In addition, with different initial weights, the network converges to different local minima during the training process, which makes the network performance unstable. Summary of the Invention

[0003] In response to the shortcomings of the existing technology, the present invention provides a shield machine operation status monitoring method and system, which solves the problem in the existing technology that the neural network is prone to fall into local minima and the technical problem that the network performance is unstable and affects normal monitoring and analysis, thereby achieving the purpose of avoiding the occurrence of local minima and improving the stability of the neural network.

[0004] To solve the above technical problems, the present invention provides the following technical solution: a method for monitoring the operating status of a shield machine, the method comprising the following steps: S1. The sensors installed on the shield machine collect the three-dimensional tensor data of the shield machine in real time when it is in operation. , and based on the three-dimensional tensor data Compute the denoising tensor ; S2, based on multiple denoising tensors Calculate the optimization coefficient after iteration , and based on the optimization coefficient Compute aligned tensors ; S3. According to the alignment tensor Calculate fusion features for monitoring shield machine dynamic characteristics ; S4. Based on fusion features Calculate the winning neuron after iteration ; S5. According to the winning neuron Calculating dynamic filter values , based on dynamic filter values Obtain the optimized shield machine status monitoring model; S6. Integrate the shield machine’s fusion features The shield machine status detection results are input into the shield machine status monitoring model, and the shield machine status detection results are sent to the shield machine monitoring center.

[0005] Furthermore, in step S1, the specific implementation steps are as follows: S11. Collect 3D tensor data of shield machine ,in, Represent the number of timestamps, the number of sensors, and the number of channels respectively. The continuous time is divided into n time periods t, and the three-dimensional tensor data of each time period is calculated. The completion data value of , the calculation formula is: in, represents a tensor indicating missing values, represents the core tensor, Indicates the first factor matrix, represents the low-rank constraint coefficient, Represents the reconstruction error at the current moment calculated by the Frobenius norm; S12. Completing data values Perform multi-scale wavelet decomposition to obtain decomposition data , the calculation formula is: in, Indicates the Decomposition data; S13. From the decomposition data Obtain the first Standard deviation of layer detail coefficient , according to the standard deviation Calculating dynamic thresholds , the calculation formula is: in, Represents decomposed data The signal-to-noise ratio, and represents the adaptive parameter; S14, according to the dynamic threshold Computing a denoising tensor , the calculation formula is: wherein, denotes the i-th denoising tensor.

[0006] Further, in step S2, the specific implementation steps are as follows: S21, any two different denoising tensors and are selected from the plurality of denoising tensors , and the joint mutual information of the denoising tensors is calculated, the calculation formula is: wherein, denotes the information entropy of the denoising tensor , and the number of denoising tensors and ; S22, the balance coefficient is calculated based on the joint mutual information , the calculation formula is: wherein, denotes the i-th balance coefficient, and ; ; S23, the gradient auxiliary value is calculated according to the balance coefficient , the calculation formula is: wherein, denotes the i-th gradient auxiliary value; ; S24, the update coefficient of the i-th iteration is calculated according to the gradient auxiliary value , the calculation formula is: wherein, denotes the learning rate; ; S25, the iteration threshold is calculated according to the number , the calculation formula is: wherein, denotes the random coefficient; S26, the iteration threshold ​determine whether the iteration is ended; If , return to step S15; If , end the iteration and obtain the final optimization coefficient ; S27, calculate the alignment tensor according to the optimization coefficient , the calculation formula is: wherein, denotes the alignment tensor.

[0007] Further, in step S3, the specific implementation steps are as follows: S31, obtain the normalized tensor based on the alignment tensor by the local standardization method of the sliding window; S32, obtain the time domain feature by the sliding differential method according to the normalized tensor ; S33, calculate the frequency band energy feature according to the normalized tensor , the calculation formula is: wherein, denotes the imaginary unit, denotes the frequency component, denotes the length of the alignment time period corresponding to the alignment tensor , denotes the time corresponding to the frequency .

[0008] S34, select the sensor pair with physical correlation in the shield machine , obtain the normalized tensor and of the sensor pair and calculate the cross feature , the calculation formula is: wherein, denotes the cross feature; S35, calculate the fusion feature according to the time domain feature , the frequency band energy feature and the cross feature , the calculation formula is: wherein, denotes the th fused feature.

[0009] Further, in step S4, the implementation steps are as follows: S41, input multiple fused features as the input of the input layer, and define the initial learning rate , weight and neighborhood radius ; S42, calculate the Euclidean distance of multiple fused features , the calculation formula is: wherein, denotes the th Euclidean distance, denotes the field function, denotes the basic weight; S43, calculate the winning neuron of the competition layer according to the Euclidean distance , the expression is: wherein, denotes the th winning neuron; S44, calculate the updated learning rate , updated weight and updated neighborhood radius according to the learning rate , weight and neighborhood radius , the calculation formula is: wherein, denotes the rounding function, denotes the current learning times; S45, control the iteration of the winning neuron according to the updated learning rate ; if .001, the iteration ends, and the current winning neuron is output; if .001, return to step S41 and recalculate the winning neuron .

[0010] Further, in step S5, the specific implementation steps are as follows: S51, a plurality of winning neurons As input, the hidden value is calculated by the hidden function of the hidden layer , the calculation formula is: Wherein, indicates the number of winning neurons , indicates the weight value of the model, indicates the intercept value; S52, according to the hidden value Calculate the output value of the output layer , the calculation formula is: Wherein, indicates the number of hidden values , indicates the intercept threshold value; S53, define the expected output , the mean square error value of the output value and the expected output , the calculation formula is: Wherein, indicates the th mean square error value; S54, according to the plurality of weight values , the intercept value and the intercept threshold Calculate the optimized weight value , the optimized intercept value and the optimized intercept threshold , based on the optimized weight value , the optimized intercept value and the optimized intercept threshold Build an optimized model ; S55, according to the mean square error value Calculate the dynamic screening value , the calculation formula is: Wherein, and respectively indicate the weight coefficient, Represents the mean square error value The mean of Represents the mean square error value The standard deviation of S56, according to dynamic screening value Complete the establishment of shield machine status monitoring model; like , then the end and the shield machine status monitoring model is obtained; like , then return to step S53 to re-train the model.

[0011] Furthermore, the optimization model The expression is: in, Indicates the An optimization model.

[0012] Furthermore, in step S54, the specific implementation steps are as follows: S541, weight value , intercept value and intercept threshold Converted into multiple particles through vector encoding , based on particles The encoded value defines the initial position of each particle ,speed , individual optimal position Optimal position of the group ; S542. Construct a fitness function to guide particles to migrate to a better area , the expression is: in, and They represent the actual particle position and the expected particle position in the particle swarm, respectively. represents the logarithm of the positions of the two particles; S543, according to speed Calculate the Daizhongdi Update velocity value of each particle , the calculation formula is: in, represents the inertia weight, and They represent the acceleration coefficients, and represent random numbers, and ; S544, according to the update speed value Calculate updated position value , the calculation formula is: in, Indicates the Update position values ; S545: Update the position value Calculate the optimal position value , the calculation formula is: in, Indicates the Optimal position values; S546. Get the current number of iterations , determine whether the iteration should stop according to the current number of iterations; like 000, then return to step S542 to continue iteration; like 000, then stop the iteration and reverse encode the optimal particle at this time to obtain the optimized weight value , optimize the intercept value and optimize the intercept threshold .

[0013] The technical solution also provides a system for the shield machine operation status monitoring method, which includes: Noise reduction module, used to collect 3D tensor data of the shield machine through sensors on the shield machine , and based on the three-dimensional tensor data Compute the denoising tensor ; Alignment module for denoising multiple tensors Calculate the optimization coefficient after iteration , and based on the optimization coefficient Compute aligned tensors ; Feature module, used to align tensors Calculate fusion features for monitoring shield machine dynamic characteristics ; Competition module, used to Calculate the winning neuron after iteration ; Optimize model module to optimize the model based on the winning neuron Calculate a dynamic screening value , based on the dynamic screening value An optimized shield machine state monitoring model is obtained. A classification module is configured to input the fusion features of the shield machine into the shield machine state monitoring model to obtain a shield machine state detection result, and send the shield machine state detection result to a monitoring center of the shield machine.

[0014] By means of the above technical solutions, the present application provides a shield machine operation state monitoring method and system, which has at least the following beneficial effects: 1. The present application can more fully pre-process the three-dimensional tensor data by decomposing and completing the three-dimensional tensor data and performing noise reduction processing, and can generate a dynamic threshold value according to the changes of the three-dimensional tensor data in the noise reduction process, so that the noise-reduced tensor can more clearly reflect the accurate state of the shield machine, which is beneficial to the analysis of subsequent steps.

[0015] 2. The present application connects the competitive layer with unsupervised learning and the neural network with supervised learning in series to form a new neural network model, which avoids the local minimum value defect of the neural network in the prior art, and can also reduce the amount of learning data, so that the new neural network can successfully learn and classify and identify data without a large amount of data, greatly improving the stability of the neural network and the learning efficiency.

[0016] 3. The present application solves the problem of selection of connection weights and threshold values of the neural network in the prior art by using a group optimization method. The prior art sets the connection weights and threshold values according to experimental experience, and then adjusts the connection weights and threshold values during the training process. This method needs to adjust the connection weights and threshold values repeatedly, which seriously affects the learning efficiency of the neural network. The present application can greatly improve the selection efficiency of the weights and threshold values based on the group optimization method, and has smaller test error and better nonlinear fitting capability, which can improve the classification accuracy of the shield machine operation state. BRIEF DESCRIPTION OF DRAWINGS

[0017] The drawings described herein are used to provide further understanding of the present application, and form a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application, and do not constitute an improper limitation on the present application. In the drawings: Figure 1 A flowchart of a shield machine operation state monitoring method of the present application; Figure 2 A structural block diagram of a shield machine operation state monitoring system of the present application. DETAILED DESCRIPTION

[0018] ​In order to make the above objectives, characteristics and advantages of the present application more obvious and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments. By this, the realization process of how to apply technical means to solve technical problems and achieve technical effects can be fully understood and implemented.

[0019] Due to the problems of the neural network in the prior art being prone to local minimum and the network performance being unstable, affecting normal monitoring analysis, the embodiment proposes a shield machine operation state monitoring method, which can avoid the occurrence of the local minimum problem and improve the stability of the neural network, as shown in Figure 1 , the method comprises the following steps: S1, collecting three-dimensional tensor data of the shield machine in the working state in real time through the sensor loaded on the shield machine , and calculating a denoising tensor according to the three-dimensional tensor data ; Due to the collision between the shield machine and the rock during the working process, a large amount of noise affecting the sensor collection is generated, and then the collected data is inaccurate. In order to solve this problem, the embodiment proposes a more detailed implementation method, as follows: S11, collecting three-dimensional tensor data of the shield machine , wherein respectively represent the number of time stamps, the number of sensors and the number of channels, and the continuous time is divided into n time periods , calculating the complete data value of the three-dimensional tensor data in each time period , and the calculation formula is: , wherein represents a missing value indicating tensor, represents a core tensor, represents the th factor matrix in the time period t, represents a low rank constraint coefficient, represents the reconstruction error of the current moment calculated by the Frobenius norm; the missing value indicating tensor can be directly generated by the missing value state of the data, the core tensor and the factor matrix can be obtained by alternating least squares iteration, and the low rank constraint coefficient is obtained by cross-validation method.

[0020] S12, performing multi-scale wavelet decomposition on the complete data value to obtain decomposition data , and the calculation formula is: , wherein represents the Decomposition data; S13. From the decomposition data Obtain the first Standard deviation of layer detail coefficient , according to the standard deviation Calculating dynamic thresholds , the calculation formula is: in, Represents decomposed data The signal-to-noise ratio, and represents the adaptive parameter; S14, according to the dynamic threshold Compute the denoising tensor , the calculation formula is: in, Indicates the A denoised tensor can be generated by decomposing, completing and denoising the three-dimensional tensor data, so that the three-dimensional tensor data can be preprocessed more fully. At the same time, dynamic threshold calculation is used in the denoising process, which can generate dynamic thresholds according to the changes in the three-dimensional tensor data, so that the denoised tensor after denoising can more clearly reflect the accurate status of the shield machine, which is beneficial to the analysis of subsequent steps.

[0021] S2, based on multiple denoising tensors Calculate the optimization coefficient after iteration , according to the optimization coefficient Compute aligned tensors ; Denoising Tensor Contains a variety of data features. If you want to further analyze the operating status of the shield machine, you need to obtain these data features to perform status analysis more accurately. To solve this problem, the detailed implementation steps are as follows: S21. In multiple denoising tensors Arbitrarily select two different denoising tensors and , and calculate the denoising tensor Joint mutual information , the calculation formula is: in, Represents the denoising tensor The information entropy of Denoising Tensor and The number of; Joint mutual information refers to the amount of information that one variable contains about another variable. Information entropy is a common method for calculating uncertainty, which will not be described here; S22, based on joint mutual information Calculate the balance coefficient , the calculation formula is: in, Indicates the A balance coefficient, and ; S23, according to the balance coefficient Calculate gradient auxiliary value , the calculation formula is: in, Indicates the Gradient auxiliary value; S24, according to the gradient auxiliary value Calculate the The update coefficient of the generation , the calculation formula is: in, represents the learning rate; S25, according to quantity Calculating the iteration threshold , the calculation formula is: in, represents the random coefficient; S26, according to the iteration threshold Determine whether the iteration is finished; like , then return to step S21; like , then the iteration ends and the final optimization coefficient is obtained ; S27, according to the optimization coefficient Compute aligned tensors , the calculation formula is: in, Indicates the Aligned tensors, by denoising the tensor The preprocessing can obtain the denoising tensor Aligned tensor of , align tensors It can more accurately reflect the status of the shield machine, thereby helping subsequent steps to analyze the status of the shield machine more accurately and quickly, and improve the efficiency and accuracy of data processing.

[0022] S3. According to the alignment tensor Calculate fusion features for monitoring shield machine dynamic characteristics ; Align tensors After optimizing the accuracy and efficiency of the data, you also need to align the tensors The features are decomposed and fused features are formed In order to solve this problem, the specific methods adopted are as follows: S31, based on aligned tensors The normalized tensor is obtained by local normalization of the sliding window ; Local normalization is a method of normalizing a tensor in a local area through a sliding window, which can enhance local contrast or reduce the inconsistency of local features. Local normalization is a common tensor processing method and will not be described in detail here.

[0023] S32, according to the standardized tensor The time domain features are obtained by sliding differential method ; The sliding differentiation method is often used in time domain feature extraction to smooth signals, enhance local contrast or normalize features. When used in conjunction with the local normalization method, the differentiation results can be further processed to enhance features. The sliding differentiation method is a common tensor processing method and will not be described in detail here.

[0024] S33, according to the standardized tensor Calculate frequency band energy characteristics , the calculation formula is: in, represents the imaginary unit, represents the frequency component, Represents aligned tensors The length of the corresponding alignment time period, Indicates frequency corresponding time; S34. Select a pair of sensors with physical association in the shield machine. , get the sensor pair The normalized tensor and And calculate the cross feature , the calculation formula is: in, represents the first cross feature; S35, according to the time domain feature , frequency band energy feature and cross feature Calculate the fusion feature , the calculation formula is: Among them, represents the first fusion feature, through the acquisition of fusion feature , the multiple features of shield machine operation data can be improved, including time domain, frequency domain and cross feature, through these features, the analysis and classification of shield machine operation state can be more accurately completed, and the accuracy of shield machine monitoring analysis is improved. S4, according to the fusion feature

[0025] Calculate the winning neuron after iteration ; Because the neural network in the prior art is easy to fall into local minimum, and the network performance is unstable, which affects the normal monitoring analysis, in order to solve this problem, the embodiment proposes a more detailed implementation method, as follows: S41, multiple fusion features are taken as the input of the input layer, and the initial learning rate , weight and neighborhood radius are defined; Here, the input layer of the model is constructed first, and the key model parameters are defined based on practical experience.

[0026] S42, calculate the Euclidean distance of multiple fusion features , the calculation formula is: Among them, represents the first Euclidean distance, represents the field function, represents the basic weight; S43, according to the Euclidean distance Calculate the winning neuron of the competition layer , the expression is: Among them, represents the first winning neuron; S44, according to the learning rate represents the first winning neuron; S44, according to the learning rate ​​​, weight and neighborhood radius Calculate the updated learning rate , update weights and update neighborhood radius , the calculation formula is: in, represents the rounding function, Indicates the current number of learning times; S45, according to the updated learning rate Controlling the winning neuron iterations; like .001, the iteration ends and the current winning neuron is output ; like 001, then return to step S41 and recalculate the winning neuron Since the learning rate is a continuously decaying quantity, after multiple iterations, the learning rate is used to determine whether the iteration has ended. By connecting the competitive layer with unsupervised learning and the neural network with supervised learning in series, a new neural network model is formed, which avoids the shortcomings of local minima in the neural network in the existing technology. At the same time, it can also reduce the amount of learning data, so that the new neural network does not need a huge amount of data to successfully learn and classify and identify data, greatly improving the stability of the neural network and improving learning efficiency.

[0027] S5. According to the winning neuron Calculating dynamic filter values , based on dynamic filter values Get the optimized shield machine state monitoring model; Based on the previous step, the winning neuron needs to be For further processing, the detailed implementation steps are as follows: S51, multiple winning neurons As input, the implicit value is calculated through the implicit function of the hidden layer , the calculation formula is: in, Represents the winning neuron the number of Represents the weight value of the model, represents the intercept value; S52. According to the implicit value Calculate the output value of the output layer , the calculation formula is: in, Indicates implicit value the number of represents the intercept threshold; S53. Define expected output , calculate the output value and expected output The mean square error value , the calculation formula is: in, Indicates the mean square error value; expected output It can be set as data under the normal state of the shield machine, can be set according to the experience of the staff, or can be set according to the average of the normal values ​​of the historical data of the shield machine.

[0028] S54, according to multiple weight values , intercept value and intercept threshold Calculate optimization weight values , optimize the intercept value and optimize the intercept threshold , based on the optimized weight value , optimize the intercept value and optimize the intercept threshold Building an optimization model , the expression is: in, Indicates the An optimization model; S541, weight value , intercept value and intercept threshold Converted into multiple particles through vector encoding , based on particles The encoded value defines the initial position of each particle ,speed , individual optimal position Optimal position of the group ; S542. Construct a fitness function to guide particles to migrate to a better area , the expression is: in, and They represent the actual particle position and the expected particle position in the particle swarm, respectively. represents the logarithm of the positions of the two particles; S543, according to speed Calculate the Daizhongdi Update velocity value of each particle , the calculation formula is: in, represents the inertia weight, and They represent the acceleration coefficients, and represent random numbers, and ; S544, according to the update speed value Calculate updated position value , the calculation formula is: in, Indicates the Update position values ; S545: Update the position value Calculate the optimal position value , the calculation formula is: in, Indicates the Optimal position values; S546. Get the current number of iterations , determine whether the iteration should stop according to the current number of iterations; like 000, then return to step S542 to continue iteration; like 000, then stop the iteration and reverse encode the optimal particle at this time to obtain the optimized weight value , optimize the intercept value and optimize the intercept threshold , S55, according to the mean square error value Calculating dynamic filter values , the calculation formula is: in, and Represent the weight coefficients, Represents the mean square error value The mean of Represents the mean square error value The standard deviation of S56, according to dynamic screening value Complete the establishment of shield machine status monitoring model; like , then the end and the shield machine status monitoring model is obtained; like , then return to step S53 to retrain the model, and solve the problem of selecting connection weights and thresholds of neural networks in the prior art by utilizing the group optimization method. The prior art sets them according to experimental experience, and then adjusts the connection weights and thresholds during the training process. This method requires repeated adjustments to the connection weights and thresholds, which seriously affects the learning efficiency of the neural network. The present invention is based on a group optimization method, which can greatly improve the selection efficiency of weights and thresholds, and has smaller test errors and better nonlinear fitting capabilities, and can improve the classification accuracy of the shield machine's operating status. By connecting a competitive layer with unsupervised learning and a neural network with supervised learning in series, a new neural network model is formed, which avoids the shortcomings of local minima in the neural network in the prior art, and can also reduce the amount of learning data, so that the new neural network does not need a huge amount of data to successfully learn and classify and identify data, greatly improving the stability of the neural network and improving learning efficiency.

[0029] S6. Integrate the shield machine’s fusion features The results of shield machine status detection are input into the shield machine status monitoring model, and the shield machine status detection results include normal, shield shell stuck and hob wear.

[0030] Since the neural network in the prior art is prone to local minima and the network performance is unstable, which affects the normal monitoring and analysis, this embodiment also proposes a shield machine operation status monitoring system, which can avoid the occurrence of local minima and improve the stability of the neural network. Figure 2 As shown, the monitoring system includes a noise reduction module 1, an alignment module 2, a feature module 3, a competition module 4, an optimization model module 5 and a classification module 6.

[0031] Noise reduction module 1, used to collect the three-dimensional tensor data of the shield machine through the sensors on the shield machine , and based on the three-dimensional tensor data Compute the denoising tensor ; Alignment module 2, used to derive the noise from multiple tensors Calculate the optimization coefficient after iteration , and based on the optimization coefficient Compute aligned tensors ; Feature module 3, used to align tensors Calculate fusion features for monitoring shield machine dynamic characteristics ; Competition module 4, used to Calculate the winning neuron after iteration ; Optimize model module 5, used to calculate the winning neuron Calculating dynamic filter values , based on dynamic filter values The optimized shield machine status monitoring model is obtained; the classification module 6 is used to integrate the fusion features of the shield machine The shield machine status detection results are input into the shield machine status monitoring model, and the shield machine status detection results are sent to the shield machine monitoring center.

[0032] Those skilled in the art will appreciate that all or part of the steps in the above-described embodiment methods can be accomplished by programming related hardware. Therefore, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0033] Each embodiment in this specification is described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the same or similar parts between the embodiments. For the above embodiments, since they are basically similar to the method embodiments, the description is relatively simple. For relevant parts, refer to the partial description of the method embodiments.

[0034] The above embodiments provide a detailed introduction to the present invention. Specific examples are used herein to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core ideas. At the same time, for those skilled in the art, according to the ideas of the present invention, there may be changes in the specific implementation methods and application scopes. In summary, the contents of this specification should not be understood as limiting the present invention.

Claims

1. A shield machine operation status monitoring method, characterized in that: The method comprises the following steps: S1. The sensors installed on the shield machine collect the three-dimensional tensor data of the shield machine in real time when it is in operation. , and based on the three-dimensional tensor data Compute the denoising tensor ; S2, based on multiple denoising tensors Calculate the optimization coefficient after iteration , and based on the optimization coefficient Compute aligned tensors ; S3. According to the alignment tensor Calculate fusion features for monitoring shield machine dynamic characteristics ; S4. Based on fusion features Calculate the winning neuron after iteration ; S5. According to the winning neuron Calculating dynamic filter values , based on dynamic filter values Obtain the optimized shield machine status monitoring model; S6. Integrate the shield machine’s fusion features The shield machine status detection results are input into the shield machine status monitoring model, and the shield machine status detection results are sent to the shield machine monitoring center.

2. The state monitoring method according to claim 1, characterized in that: In step S1, the specific implementation steps are as follows: S11. Collect 3D tensor data of shield machine ,in, Represent the number of timestamps, the number of sensors, and the number of channels respectively. The continuous time is divided into n time periods t, and the three-dimensional tensor data of each time period is calculated. The completion data value of , the calculation formula is: in, represents a tensor indicating missing values, represents the core tensor, Indicates the first factor matrix, represents the low-rank constraint coefficient, Represents the reconstruction error at the current moment calculated by the Frobenius norm; S12. Completing data values Perform multi-scale wavelet decomposition to obtain decomposition data , the calculation formula is: in, Indicates the Decomposition data; S13. From the decomposition data Obtain the first Standard deviation of layer detail coefficient , according to the standard deviation Calculating dynamic thresholds , the calculation formula is: in, Represents decomposed data The signal-to-noise ratio, and represents the adaptive parameter; S14, according to the dynamic threshold Compute the denoising tensor , the calculation formula is: in, Indicates the A denoising tensor.

3. The state monitoring method according to claim 1, characterized in that: In step S2, the specific implementation steps are as follows: S21. In multiple denoising tensors Arbitrarily select two different denoising tensors and , and calculate the denoising tensor Joint mutual information , the calculation formula is: in, Represents the denoising tensor The information entropy of Denoising Tensor and the number of S22, based on joint mutual information Calculate the balance coefficient , the calculation formula is: in, Indicates the A balance coefficient, and ; S23, according to the balance coefficient Calculate gradient auxiliary value , the calculation formula is: in, Indicates the Gradient auxiliary value; S24, according to the gradient auxiliary value Calculate the The update coefficient of the generation , the calculation formula is: in, represents the learning rate; S25, according to quantity Calculating the iteration threshold , the calculation formula is: in, represents the random coefficient; S26, according to the iteration threshold Determine whether the iteration is finished; like , then return to step S21; like , then the iteration ends and the final optimization coefficient is obtained ; S27, according to the optimization coefficient Compute aligned tensors , the calculation formula is: in, Indicates the Aligned tensors.

4. The state monitoring method according to claim 1, characterized in that: In step S3, the specific implementation steps are as follows: S31, based on aligned tensors The normalized tensor is obtained by local normalization of the sliding window ; S32, according to the standardized tensor The time domain features are obtained by sliding differential method ; S33, according to the standardized tensor Calculate frequency band energy characteristics , the calculation formula is: in, represents the imaginary unit, represents the frequency component, Represents aligned tensors The length of the corresponding alignment time period, Indicates frequency corresponding time; S34. Select a pair of sensors with physical association in the shield machine. , get the sensor pair The normalized tensor and And calculate the cross feature , the calculation formula is: in, Indicates the Cross-features; S35. According to time domain characteristics , frequency band energy characteristics and cross features Calculate fusion features , the calculation formula is: in, Indicates the A fusion feature.

5. The state monitoring method according to claim 1, characterized in that: In step S4, the specific implementation steps are as follows: S41, multiple fusion features As the input of the input layer, and define the initial learning rate , weight and neighborhood radius ; S42. Calculate multiple fusion features Euclidean distance , the calculation formula is: in, Indicates the Euclidean distance, represents the domain function, represents the basic weight; S43. According to Euclidean distance Calculate the winning neuron of the competition layer , the expression is: in, Indicates the A winning neuron; S44, according to the learning rate , weight and neighborhood radius Calculate the updated learning rate , update weights and update neighborhood radius , the calculation formula is: in, represents the rounding function, Indicates the current number of learning times; S45, according to the updated learning rate Controlling the winning neuron iterations; like .001, the iteration ends and the current winning neuron is output ; like 001, then return to step S41 and recalculate the winning neuron .

6. The state monitoring method according to claim 1, characterized in that: In step S5, the specific implementation steps are as follows: S51, multiple winning neurons As input, the implicit value is calculated through the implicit function of the hidden layer , the calculation formula is: in, Represents the winning neuron the number of Represents the weight value of the model, represents the intercept value; S52. According to the implicit value Calculate the output value of the output layer , the calculation formula is: in, Indicates implicit value the number of represents the intercept threshold; S53. Define expected output , calculate the output value and expected output The mean square error value , the calculation formula is: in, Indicates the Mean square error value; S54, according to multiple weight values , intercept value and intercept threshold Calculate optimization weight values , optimize the intercept value and optimize the intercept threshold , based on the optimized weight value , optimize the intercept value and optimize the intercept threshold Building an optimization model ; S55, according to the mean square error value Calculating dynamic filter values , the calculation formula is: in, and Represent the weight coefficients, Represents the mean square error value The mean of Represents the mean square error value The standard deviation of S56, according to dynamic screening value Complete the establishment of shield machine status monitoring model; like , then the end and the shield machine status monitoring model is obtained; like , then return to step S53 to re-train the model.

7. The state monitoring method according to claim 6, characterized in that: The optimization model The expression is: in, Indicates the An optimization model.

8. The state monitoring method according to claim 6, characterized in that: In step S54, the specific implementation steps are as follows: S541, weight value , intercept value and intercept threshold Converted into multiple particles through vector encoding , based on particles The encoded value defines the initial position of each particle ,speed , individual optimal position Optimal position of the group ; S542. Construct a fitness function to guide particles to migrate to a better area , the expression is: in, and They represent the actual particle position and the expected particle position in the particle swarm, respectively. represents the logarithm of the positions of the two particles; S543, according to speed Calculate the Daizhongdi Update velocity value of each particle , the calculation formula is: in, represents the inertia weight, and They represent the acceleration coefficients, and represent random numbers, and ; S544, according to the update speed value Calculate updated position value , the calculation formula is: in, Indicates the Update position values ; S545: Update the position value Calculate the optimal position value , the calculation formula is: in, Indicates the Optimal position values; S546. Get the current number of iterations , determine whether the iteration should stop according to the current number of iterations; like 000, then return to step S542 to continue iteration; like 000, then stop the iteration and reverse encode the optimal particle at this time to obtain the optimized weight value , optimize the intercept value and optimize the intercept threshold .

9. A system applied to the shield machine operation status monitoring method according to any one of claims 1 to 8, characterized in that: The system includes: Noise reduction module (1), used to collect the three-dimensional tensor data of the shield machine through the sensors on the shield machine , and based on the three-dimensional tensor data Compute the denoising tensor ; Alignment module (2) is used to align multiple denoised tensors Calculate the optimization coefficient after iteration , and based on the optimization coefficient Compute aligned tensors ; Feature module (3), used to align tensors Calculate fusion features for monitoring shield machine dynamic characteristics ; Competition module (4) is used to Calculate the winning neuron after iteration ; Optimization model module (5) is used to optimize the winning neuron Calculating dynamic filter values , based on dynamic filter values Obtain the optimized shield machine status monitoring model; Classification module (6) is used to integrate the fusion features of shield machine The shield machine status detection results are input into the shield machine status monitoring model, and the shield machine status detection results are sent to the shield machine monitoring center.

Images