Shield tunneling machine cutterhead mud cake early warning method based on LSTM network and nuclear density estimation

Abstract

The invention provides a shield tunneling machine cutterhead mud cake early warning method based on an LSTM network and kernel density estimation, and the method comprises the steps: screening shield tunneling machine operation parameters closely related to a mud cake phenomenon, and carrying out data preprocessing, including abnormal value removal, data standardization and linear interpolation labeling; constructing a model by using an LSTM network, and capturing time sequence relevance of the data; and an early warning threshold value is determined through a KDE method, so that early warning of a mud cake forming event is realized. The early trend of mud cake formation can be effectively recognized in the shield construction process, an early warning signal is sent out in advance, and construction efficiency reduction and potential safety hazards caused by mud cake formation are avoided.

Description

Technical Field

[0001] The present invention relates to the technical field of underground shield construction, and more particularly to a shield machine cutter head mud cake warning method based on LSTM network and kernel density estimation. Background Art

[0002] With the accelerating development of the urbanization process, the construction scale of urban underground space is constantly expanding. As an efficient and safe underground tunnel excavation technology, the shield tunneling method has been widely used in projects such as subways and underground passages. However, due to the complex geological conditions during shield construction, the cutter head of the shield machine is prone to the phenomenon of mud cake formation, which brings many challenges to the construction.

[0003] Mud cake formation refers to the phenomenon that the soil cut by the cutters on the cutter head adheres to the surface of the cutter head during tunneling, gradually accumulates and solidifies into blocks. The formation of mud cakes usually occurs in geological conditions such as plastic and hard plastic clay formations, muddy siltstones, and strongly weathered rock formations, especially when tunneling in these formations. Once the mud cake is formed on the cutter head, the cutting effect of the cutters will be seriously affected, resulting in a significant decrease in the tunneling efficiency of the shield machine, and even the tunneling work may come to a complete standstill. In addition, as the coverage area of the mud cake expands, the friction between the cutter head and the excavation face increases, the temperature of the cutter head rises, and the mud cake further hardens, eventually leading to the complete coverage of the cutter head by the mud cake, seriously affecting the construction efficiency and threatening the construction safety.

[0004] With the rise of data-driven methods, machine learning technologies have been gradually applied to the research of problems during shield construction. However, the existing machine learning methods still have the following deficiencies in mud cake warning: on the one hand, most methods fail to effectively capture the temporal correlation in the data, resulting in insufficient timeliness and accuracy of the warning; on the other hand, the traditional threshold setting method often relies on manual experience, lacks scientific basis, and is easily affected by subjective factors. Summary of the Invention

[0005] The object of the invention is to propose a shield machine cutter head mud cake warning method based on LSTM network and kernel density estimation to solve the above problems existing in the above-mentioned prior art.

[0006] Technical solution: A shield machine cutter head mud cake warning method based on LSTM network and kernel density estimation specifically includes:

[0007] Step S1, feature parameter screening: Screen out the key feature parameters related to the mud cake phenomenon on the cutter head from multiple parameters collected during the operation of the shield machine;

[0008] Step S2, data preprocessing: performing data preprocessing on the selected key characteristic parameters, wherein the data preprocessing includes the following contents: obtaining stable excavation state data, detecting and removing abnormal values ​​using the 3σ criterion, and standardizing characteristic parameters of different dimensions;

[0009] Step S3, data annotation: using a linear interpolation method to perform data annotation on the mud cake formation process;

[0010] Step S4, constructing an LSTM network model: based on the preprocessed data, constructing an LSTM network model and training it to capture the temporal correlation of the data, the LSTM network model includes an input layer, an LSTM layer, a fully connected layer and an output layer;

[0011] Step S5, determining the warning threshold: using the kernel density estimation method to determine the warning threshold of the LSTM network model, and based on the warning threshold, providing a real-time warning for the mud cake phenomenon of the shield machine.

[0012] According to a further improvement of the present invention, the key characteristic parameters in step S1 include cutter head torque, cutter head rotation speed, total thrust, penetration and propulsion speed.

[0013] According to a further improvement of the present invention, the data preprocessing in step S2 includes:

[0014] Step S21, obtaining stable excavation state data: from the key characteristic parameters, according to the shield machine working states at different times recorded during the shield machine construction process, remove the data in the shield machine stop mode and assembly mode, and remove the data from the start of each ring excavation to the stable excavation of the shield machine; the method for determining the data from the start of each ring excavation to the stable excavation is: after switching from the stop state to the excavation state, the data of one minute after the start of excavation;

[0015] Step S22, removing outliers, includes:

[0016] Step S221, using the 3σ criterion to detect and remove outliers in key characteristic parameters, the specific method is: for the working data of a certain link in the shield machine excavation process, calculate its mean μ and standard deviation σ, then according to the 3σ rule, the range of outliers is μ±3σ, that is, any data point exceeding μ+3σ or below μ-3σ is regarded as an outlier and removed.

[0017] Step S222, calculate the standard deviation σ according to the Bessel formula, the method is as follows:

[0018]

[0019] In the formula, x is the measured value of the cutter head speed, n is the number of samples, if the measured value of the shield machine operation data at a certain moment x b The residual error vb , satisfying that its absolute value is greater than 3 times the standard deviation, i.e., |v b | > 3σ, then this measured value is considered an abnormal data and should be excluded;

[0020] Step S23, data standardization: The min-max standardization method is used to standardize the key feature parameters to eliminate the dimensional differences between different features, so that the data distributions of different features are within a relatively consistent range. The calculation method is as follows:

[0021]

[0022] In the formula, x is the original data, x′ is the normalized data, x min is the minimum value of the feature parameter, x max is the maximum value of the feature parameter.

[0023] According to a further improvement of the present invention, the specific steps of the data annotation method in step S3 include:

[0024] Step S31, starting point annotation: The first sampling point of the parameter data of the starting ring is marked as 0, which is the starting point of the mud cake development at this time;

[0025] Step S32, ending point annotation: The last sampling point of the parameter data of the mud cake discovery ring is marked as 1, indicating that the mud cake has been completely formed at this time;

[0026] Step S33, linear interpolation: Between the starting point and the ending point, linear interpolation is performed on all intermediate points according to time. The interpolation method is:

[0027]

[0028] In the formula, Y i is the label of the i-th sampling point starting from the starting point of the mud cake formation, which can represent the development degree of the mud cake, t i is the time of the i-th sampling point, t start is the time of the starting point, t end is the time of the ending point;

[0029] Step S34, normal data annotation: After cleaning up the data after the mud cake formation and other data where no mud cake formation event is found, all are marked as 0.

[0030] According to a further improvement of the present invention, the specific steps of calculating the warning threshold in step S5 are:

[0031] Step S51, expand each data point into a Gaussian kernel function. For a set of sample data with the same distribution: {x 1 , x 2 , …, x n}, the estimated value of its probability density function f(x) is:

[0032] In the formula, is the density estimate at the sample point x, n is the number of sample points, h is the bandwidth parameter used to control the smoothness, x i is the i-th sample point, K is the kernel function, satisfying: ∫K(u)du = 1. If the choice of the kernel function has little influence on the result of kernel density estimation, a Gaussian kernel function with good smoothness is selected:

[0033] Step S52, use integration to calculate the sum of all kernel functions to obtain the cumulative probability density

[0034] Step S53, by setting the significance level, determine the threshold L of the cumulative probability density function:

[0035] When the output value Y of the early warning model is greater than the threshold L, the model alarms, that is, it is determined that the mud cake has begun to gradually form at the current moment.

[0036] Beneficial effects: By using the LSTM network to capture time-series data and combining kernel density estimation to determine the early warning threshold, it can more accurately predict the phenomenon of cutter head mud caking, reducing false alarms and missed alarms; through preprocessing steps such as feature screening, outlier removal, and data standardization, the data quality is ensured, and the robustness and reliability of the model are improved; real-time monitoring and risk reduction: Realize real-time early warning, help operators take measures in time, avoid equipment damage and construction stagnation, and reduce construction risks and costs. Description of the Drawings

[0037] Figure 1 is the overall flowchart of the solution of the present invention.

[0038] Figure 2 is the schematic diagram of the early warning model based on the LSTM neural network of the present invention. Detailed Embodiments

[0039] In order to make the technical problems of the present invention clearer, the embodiments of the present invention will be provided below and a more comprehensive description will be given. However, the specific embodiments given here are only used to explain the present invention and are not used to limit the use environment and scope of the invention.

[0040] As Figure 1 shown, a method for early warning of cutter head mud caking of a shield machine based on an LSTM network and kernel density estimation provided by the present invention specifically includes the following steps:

[0041] Step S1, after obtaining the shield machine operation data of sampling points at different times collected by the shield machine acquisition system, the characteristic parameters closely related to the cutterhead mud cake phenomenon are screened out, including cutterhead torque, cutterhead speed, total thrust, penetration and propulsion speed.

[0042] Step S2, data preprocessing, specifically includes:

[0043] Step S21, obtaining stable excavation state data: from the original data, according to the shield machine working state at different times recorded during the shield machine construction process, remove the data in the shield machine stop mode and assembly mode, and remove the data from the start of each ring excavation to the stable excavation of the shield machine. The method for determining the data from the start of each ring excavation to the stable excavation is: after switching from the stop state to the excavation state, the data of one minute after the start of excavation;

[0044] Step S22, remove outliers: External interference, human factors, and sensor failures may cause data anomalies. Data containing outliers is not conducive to subsequent analysis and modeling. The 3σ criterion is used to detect and remove outliers. The standard deviation σ is calculated according to the Bessel formula:

[0045]

[0046] Where x is the measured value of the cutter head speed, n is the number of samples, if the measured value of the shield machine operation data at a certain moment x b The residual error v b , its absolute value is greater than 3 times the standard deviation, that is, |v b |>3σ, the measured value is considered abnormal data and should be eliminated;

[0047] Step S23, data standardization: different feature parameters have different dimensions, so their numerical ranges vary greatly. If these parameters are directly input into the early warning model, the model performance may be degraded. This is because in machine learning, the model training process is mostly based on the gradient descent algorithm. When training the model, features with larger dimensions (such as thrust) will also have larger gradients, and the adjustment amplitude of the model weights will also be larger when the model parameters are updated, resulting in features with larger values ​​dominating the model update process during the gradient descent process, while features with smaller values ​​contribute less to the model update, or are even ignored by the model;

[0048] In order to eliminate the dimensional differences between different features, the data is standardized so that the data of different features are distributed in a relatively consistent range. The min-max standardization method is used, and the calculation method is as follows:

[0049]

[0050] Where x is the original data; x' is the data after normalization; x min is the minimum value of the characteristic parameter; x max is the maximum value of the characteristic parameter.

[0051] Step S3, the specific steps of data annotation include:

[0052] Step S31, starting point annotation: Mark the first sampling point of the parameter data of the starting ring as 0, which is the starting point of the mud cake development at this time;

[0053] Step S32, ending point annotation: Mark the last sampling point of the parameter data of the mud cake discovery ring as 1, indicating that the mud cake has been completely formed at this time;

[0054] Step S33, linear interpolation: Between the starting point and the ending point, perform linear interpolation on all intermediate points according to time. The interpolation method is:

[0055]

[0056] Where: Y i is the label of the i-th sampling point starting from the starting point of the mud cake formation, which can represent the development degree of the mud cake; t i is the time of the i-th sampling point; t start is the time of the starting point; t end is the time of the ending point;

[0057] Step S34, normal data annotation: After cleaning the data of the completed mud cake and other data where no mud cake formation is found, all are marked as 0.

[0058] Step S4, construct an LSTM network: The mud cake formation warning model based on LSTM can be divided into 4 layers according to functions inside, namely the input layer, the LSTM layer, the fully connected layer, and the output layer. Input layer. The shape of the input layer is (batch_size, T, features), where batch_size is the number of input samples, (T, features) represents that each time window sample contains T time steps, and features is the feature dimension of each time step of the sample; The LSTM layer is responsible for processing the data of each time step, capturing the temporal correlation relationship of the data through hidden units, and the output shape of the LSTM is (batch_size, units_t), where units_t is the number of LSTM hidden units; The fully connected layer has two layers. The activation function of the first fully connected layer is the Relu function, which is responsible for integrating the features of the LSTM layer; The activation function of the second fully connected layer is the Sigmoid function, which connects the integrated features to the output layer to obtain the final result;

[0059] The sliding window method is used to divide the preprocessed data into time windows. The time window length T is set to 20 to generate the input data, and the corresponding output selects the label at the next time step, that is, for each data window X t ={x t ,x t+1 ,…,x t+19}, and its corresponding label is Y t+20 ; According to this strategy, the data from the 925th ring to the 930th ring is used to construct the dataset, and the dataset is divided according to the proportions of 80% and 20% for the training set and the test set respectively. 20% of the training set is used as the validation set to monitor the model performance during training; The Adam algorithm is used as the optimizer for model training, the learning rate is 0.001, the batch size is 64, the number of training iterations is set to 400, and the EarlyStopping callback function is used with patience set to 10, that is, if the validation loss does not decrease within 10 epochs, the model stops training.

[0060] Step S5, Early warning threshold calculation: Kernel Density Estimation (KDE) is a non-parametric statistical method that does not assume the distribution form of the data and only uses the information of the data itself for estimation. Using kernel density estimation to determine the threshold helps to avoid the subjective influence when setting the threshold manually, thereby improving the objectivity and accuracy of the early warning model; The basic principle of kernel density estimation is to expand each data point into a small probability density function (i.e., kernel function), and add these kernel functions to obtain the overall probability density estimation.

[0061] Using the kernel density estimation method, calculate the early warning threshold for mud cake formation. The specific steps are as follows:

[0062] Step S51, for a set of sample data with the same distribution: {x 1 ,x 2 ,…,x n}, the estimated value of its probability density function f(x) is:

[0063] In the formula, is the density estimated value at the sample point x, n is the number of sample points, h is the bandwidth parameter used to control the smoothness, x i is the i-th sample point, K is the kernel function, and it satisfies: ∫K(u)du = 1. When the choice of the kernel function has little influence on the result of kernel density estimation, the Gaussian kernel function with good smoothness is selected:

[0064] Step S52, use integration to calculate the sum of all kernel functions to obtain the cumulative probability density

[0065] In step S53, the significance level α can be set to 0.05, and the warning threshold L is obtained by solving the following equation:

[0066]

[0067] When the output value Y of the warning model is greater than the threshold L, the model alarms, that is, it is determined that the mud cake has begun to gradually form at the current moment.

[0068] In step S53, the significance level α can be set to different values according to actual needs.

[0069] The above are only the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various changes can be made to the above embodiments of the present invention. That is, all simple, equivalent changes and modifications made according to the claims and the content of the specification of the present invention application fall within the scope of the claims of the present invention patent. What is not described in detail in the present invention is conventional technical content.

Claims

1. A shield machine cutterhead mud cake early warning method based on LSTM network and kernel density estimation, characterized in that The following steps are involved: Step S1, feature parameter screening: screening out key feature parameters related to the cutterhead mud cake phenomenon from multiple parameters collected during the operation of the shield machine; Step S2, data preprocessing: performing data preprocessing on the selected key characteristic parameters, wherein the data preprocessing includes the following contents: obtaining stable excavation state data, detecting and removing abnormal values ​​using the 3σ criterion, and standardizing characteristic parameters of different dimensions; Step S3, data annotation: using a linear interpolation method to perform data annotation on the mud cake formation process; Step S4, constructing an LSTM network model: based on the preprocessed data, constructing an LSTM network model and training it to capture the temporal correlation of the data, the LSTM network model includes an input layer, an LSTM layer, a fully connected layer and an output layer; Step S5, determining the warning threshold: using the kernel density estimation method to determine the warning threshold of the LSTM network model, and based on the warning threshold, providing a real-time warning for the mud cake phenomenon of the shield machine.

2. According to claim 1, a shield machine cutter head mud cake early warning method based on LSTM network and kernel density estimation is characterized in that: The key characteristic parameters in step S1 include cutter head torque, cutter head rotation speed, total thrust, penetration and propulsion speed.

3. A shield machine cutterhead mud cake early warning method based on LSTM network and kernel density estimation according to claim 1, characterized in that: The data preprocessing in step S2 includes: Step S21, obtaining stable excavation state data: from the key characteristic parameters, according to the shield machine working states at different times recorded during the shield machine construction process, remove the data in the shield machine stop mode and assembly mode, and remove the data from the start of each ring excavation to the stable excavation of the shield machine; the method for determining the data from the start of each ring excavation to the stable excavation is: after switching from the stop state to the excavation state, the data of one minute after the start of excavation; Step S22, removing outliers, includes: Step S221, using the 3σ criterion to detect and remove outliers in key characteristic parameters, the specific method is: for the working data of a certain link in the shield machine excavation process, calculate its mean μ and standard deviation σ, then according to the 3σ rule, the range of outliers is μ±3σ, that is, any data point exceeding μ+3σ or below μ-3σ is regarded as an outlier and removed. Step S222, calculate the standard deviation σ according to the Bessel formula, the method is as follows: In the formula, x is the measured value of the cutter head speed, n is the number of samples, if the measured value of the shield machine operation data at a certain moment x b The residual error v b , its absolute value is greater than 3 times the standard deviation, that is, |v b |>3σ, the measured value is considered abnormal data and should be eliminated; Step S23, data standardization: The key feature parameters are standardized using the min-max standardization method to eliminate the dimensional differences between different features, so that the data of different features are distributed in a relatively consistent range. The calculation method is as follows: In the formula, x is the original data, x′ is the normalized data, and x m i n is the minimum value of the characteristic parameter, x max is the maximum value of the characteristic parameter.

4. The shield machine cutter head mud cake early warning method based on LSTM network and kernel density estimation according to claim 1 is characterized in that: Step S3: The data labeling method specifically includes: Step S31, marking the starting point: marking the first sampling point of the parameter data of the starting ring as 0, which is the starting point of the mud cake development; Step S32, marking the end point: marking the last sampling point of the parameter data of the mud cake discovery ring as 1, indicating that the mud cake has been completely formed at this time; Step S33, linear interpolation: between the starting point and the end point, linear interpolation is performed on all intermediate points according to time, and the interpolation method is: Where Y i is the label of the i-th sampling point starting from the mud cake starting point, which can indicate the development degree of the mud cake. i The time of the i-th sampling point, t start is the starting time, t end It is the time of the end point; Step S34, normal data labeling: after cleaning the mud cake and other data without mud cake events, all data are labeled as 0.

5. The shield machine cutter head mud cake early warning method based on LSTM network and kernel density estimation according to claim 1, characterized in that: The specific steps of calculating the warning threshold in step S5 are: Step S51, expand each data point into a Gaussian kernel function, for a set of sample data with the same distribution: {x1,x2,...,x n }, the estimated value of its probability density function f(x) is: In the formula, is the density estimate at the sample point x, n is the number of sample points, h is the bandwidth parameter used to control the degree of smoothing, x i is the i-th sample point, K is the kernel function, satisfying: ∫K(u)du=1. If the choice of kernel function has little effect on the result of kernel density estimation, a Gaussian kernel function with good smoothness is selected: Step S52, using integral to calculate the sum of all kernel functions to obtain the cumulative probability density Step S53, by setting the significance level, determine the threshold L of the cumulative probability density function: When the output value Y of the early warning model is greater than the threshold value L, the model will alarm, that is, it is determined that the mud cake has begun to gradually form at the current moment.

Images