A shield cutterhead health status assessment method based on Seq2Seq

Abstract

The present application relates to a shield cutter health status assessment method based on Seq2Seq. After collecting data, the score of the time domain features is calculated for feature extraction, and the first 5% of the data is selected as a training set. All collected data is used as a test set, and the model is trained with the training set. The test set is then input into the GRU-based Seq2Seq model to output the difference, and a full life cycle degradation curve is drawn based on the difference graph. The wear state of the cutter disc is judged based on the full life cycle degradation curve. The present application can evaluate and predict the life cycle of the tool based on the tool wear mechanism to accurately control the timing of tool change.

Description

Technical Field

[0001] The present application belongs to the technical field of shield machines, and specifically relates to a Seq2Seq-based shield cutterhead health status assessment method. Background Art

[0002] The shield method is one of the main construction methods for urban underground space. During the tunneling process of the shield machine, the wear of the shield cutter is the main factor affecting the tunneling progress. Severe wear of the cutter will reduce the tunneling efficiency. When the cutter is opened and replaced, the probability of instability of the excavation face and tunnel collapse accidents will be greatly increased.

[0003] At present, shield construction generally relies on construction experience to explore the timing of shield opening and cutting. However, when encountering high-pressure cross-river and sea shield tunnels, the construction experience relied on is conservative, which greatly limits the frequency of opening and cutting. Summary of the invention

[0004] The present invention provides a shield cutter health status assessment method based on Seq2Seq, which can assess and predict the life cycle of the tool based on the tool wear mechanism, so as to accurately control the timing of tool change.

[0005] The technical solution of this application is as follows: A shield cutterhead health status assessment method based on Seq2Seq includes the following steps: S01) Collect the tunneling speed, total thrust, total contact force of the cutterhead, penetration, cutterhead torque, and thrust of each group of main thrust cylinders of the shield machine during its entire life cycle, and perform noise reduction and normalization processing on the collected data; S02) Calculate the time domain features of the collected data and use the monotonicity index as the score of each time domain feature, and extract the time domain features based on the score; S03) using all collected data as a test set, defining the first 5% of the test set in the time series as health status data, the health status data forming a training set, and the health status data corresponding to the health status; S04) Using the training set to train a GRU-based Seq2Seq model in a healthy state, the Seq2Seq model aims to minimize the reconstruction error, and the reconstruction error is defined as: ; In the formula, e represents the reconstruction error, U input 、U output Represent the input sequence and output sequence of the model respectively; The output value of the trained Seq2Seq model represents the absolute value of the difference between the input data and the health status data, and the absolute value of the difference is defined as the difference degree; S05) inputting the time domain features in the test set into the trained Seq2Seq model and outputting the difference degree; drawing a full life cycle degradation curve based on the difference degree; the abscissa of the full life cycle degradation curve represents the full life cycle, and the ordinate represents the normalized degradation index, and the normalized degradation index is defined as the ratio of the difference degree to the true value; S06) Smoothing the full life cycle degradation curve, and determining the wear state of the cutter disc according to its life cycle.

[0006] Furthermore, the time domain features in step S02) include: mean, standard deviation, peak, root mean square value, waveform index, peak index, pulse index, kurtosis index, and skewness index.

[0007] Furthermore, the reconstruction of the health state in step S04) adopts a forward sequence reconstruction method.

[0008] Due to the adoption of the above technical solution, the beneficial effects of this application are as follows: 1. This application uses the data in the early stage of the cutter disc life cycle as the health status data, and the early wear when the cutter disc is replaced as the benchmark. This benchmark includes both the initial state of the cutter disc and some of its states after healthy work, and can fully reflect the wear state of the cutter disc in actual excavation.

[0009] 2. This application uses the cutterhead data of the entire life cycle as a test set. As the cutterhead is continuously used, its wear increases, and the difference from the healthy state becomes larger and larger. This application uses this difference as the output value of the model to characterize the gap between it and the healthy state, thereby reflecting its wear state. This application is not based on brand new cutterhead data, nor is it based on the cutterhead data at a certain moment, but rather uses a dynamic process as a benchmark to predict the entire life cycle. Therefore, this application can better reflect the wear state of the cutterhead during actual excavation. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] The drawings described herein are used to provide a further understanding of the present application and constitute 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 flow chart of a Seq2Seq-based shield cutterhead health status assessment method provided in this application; Figure 2 This is the architecture diagram of the Seq2Seq model based on GRU in this application; Figure 3 This is a full life cycle degradation curve diagram of an embodiment of the present application. DETAILED DESCRIPTION

[0011] Based on the background technology, as shown in the attached Figure 1 As shown, the present application provides a shield cutter head health status assessment method based on Seq2Seq.

[0012] S01) Collect the tunneling speed, total thrust, total contact force of the cutterhead, penetration, cutterhead torque, and thrust of each group of main thrust cylinders of the shield machine throughout its life cycle, and perform noise reduction and normalization on the collected data.

[0013] Directly establishing a prediction model based on raw data will lead to deviations in the prediction results. Since the construction conditions of the shield machine are relatively complex, there will be random noise interference during the data collection process, so the data needs to be denoised.

[0014] The thrust of each group of main push cylinders refers to the thrust of main push cylinder group A, the thrust of main push cylinder group B, the thrust of main push cylinder group C, the thrust of main push cylinder group D, the thrust of main push cylinder group E, and the thrust of main push cylinder group F. Finally, the moving sliding window mean method is used to reduce noise on the original data.

[0015] S02) Calculate the time domain features of the collected data and use the monotonicity index as the score of each time domain feature, and extract the time domain features based on the score.

[0016] After data denoising, feature extraction is required to better mine the information in the collected data. This application mainly calculates 9 time domain features of each collected data, including mean, standard deviation, peak, root mean square value, waveform, peak, pulse, kurtosis, and skewness, and uses monotonicity to calculate the score of each feature, arranges the scores from high to low, and selects useful features based on the arrangement order.

[0017] S03) Using all collected data as a test set, defining the first 5% of the time series of the test set as health status data, the health status data constitute a training set, and the health status data corresponds to a health status.

[0018] It should be noted that those skilled in the art can determine the selected time series data according to the vibration signal, the cutter head temperature and the rotation speed.

[0019] S04) Using the training set to train a GRU-based Seq2Seq model in a healthy state, the Seq2Seq model aims to minimize the reconstruction error, and the reconstruction error is defined as: ; In the formula, e represents the reconstruction error, U input , U outputRepresent the input sequence and output sequence of the model respectively; The output value of the trained Seq2Seq model represents the absolute value of the difference between the input data and the health status data, and the absolute value of the difference is defined as the difference degree.

[0020] During the model training process, the health status still needs to be monitored. The monitoring method is: installing an accelerometer on the cutter disc can capture tiny vibration changes. By analyzing the vibration spectrum and using mathematical tools such as FFT (Fast Fourier Transform) to process the collected vibration signal, a vibration spectrum diagram is obtained. Different types of faults will produce characteristic peaks within a specific frequency range, thereby diagnosing the condition of the cutter disc. At the same time, by using an infrared thermal imager to perform non-contact temperature measurement on the cutter disc, abnormal hot spots can be quickly found, which helps to identify overheated components at an early stage.

[0021] Among them, the vibration signal is obtained through the vibration data obtained by the accelerometer or vibration sensor installed near the cutter head. The change of vibration mode can characterize the imbalance of wear inside the equipment. The temperature change of the cutter head during operation can reflect the health status. Excessive temperature may be a manifestation of problems such as poor lubrication and excessive load. At the same time, whether the rotation speed of the cutter head is stable can also reflect the health status of the cutter head.

[0022] As attached Figure 2 As shown in the figure, the encoder and decoder in the Seq2Seq structure are essentially two RNN networks with different structures. The RNN network is composed of a series of hidden states. h and works on variable length sequences Optional output of y The neural network composed of t The change of hidden state of RNN network Changes according to the formula: ; In the formula ，f It is a nonlinear activation function, such as Sigmoid function and long short-term memory unit.

[0023] The RNN network can be trained by predicting the next symbol in a sequence to learn the probability distribution of the sequence. When learning the probability distribution of a sequence, at each time point t The output of is the conditional distribution , use the softmax activation function to output (1-K encoding) multinomial distribution, which is expressed as follows: ; In the formula, is the weight matrix W row; For all possible symbols j=1,…,K, by combining these probabilities, the probability of sequence X can be calculated using the following formula: ; That is, by learning the conditional distribution, a new sequence of codeword samples can be directly sampled iteratively at each time step.

[0024] In the Seq2Seq structure, the encoder reads the input sequence sequentially. X The RNN for each symbol, when it reads the sequence X At each symbol of , the hidden state of the RNN will change according to the formula. After reading the end marker symbol of the sequence, the hidden unit of the RNN becomes the semantic vector of the entire input sequence C .

[0025] In the Seq2Seq structure, the decoder is another RNN, whose main function is to be able to read the hidden state after being trained. In the case of To generate the output sequence. The structure of the decoder's RNN is different from that of the encoder. The hidden state calculation formula of the decoder RNN is as follows: ; Next symbol The conditional distribution calculation formula is: ; In the formula, g is the activation function.

[0026] In this application, the Seq2Seq model based on GRU is a variant of RNN, which can capture long-term dependencies in time series and solve the gradient vanishing problem of traditional RNN when performing sequence prediction. GRU includes two gate structures, namely reset gate and update gate. The reset gate determines whether to forget the previous state, and the update gate determines whether to update the new hidden layer state.

[0027] The outputs of the reset gate and update gate are respectively and It is expressed as follows: In the formula, For the moment t Input, is the output at time t, For the moment tHidden layer output, To reset the gate output, Update gate output. Reset gate Decide whether to restore the previous state h The forget gate determines whether to update the new hidden state renew.

[0028] S05) Inputting the time domain features in the test set into the trained Seq2Seq model and outputting the difference; drawing a full life cycle degradation curve based on the difference; the abscissa of the full life cycle degradation curve represents the full life cycle, and the ordinate represents the normalized degradation index, and the normalized degradation index is defined as the ratio of the difference to the true value.

[0029] S06) Smoothing the full life cycle degradation curve, and determining the wear state of the cutter disc according to its life cycle.

[0030] After the test set is input into the model, the predicted value, i.e., the degree of difference, is output. The seq2seq sequence prediction model with both the encoder and decoder being GRU is used to construct a cutter head tool performance degradation assessment model. Based on the feature sequence extracted previously, the feature sequence value in a healthy state with a length of L is intercepted. If the input feature is N-dimensional, the input of the model is an L×N-dimensional feature sequence, which is mainly used to train the GRU-based Seq2Seq model to reconstruct the feature sequence in a healthy state and learn the intrinsic information contained in the feature parameters of the cutter head tool in a healthy state. The sequence reconstruction method adopted is forward sequence reconstruction, that is, the sequence input to the model is , then the output sequence is , represents the predicted sequence based on the input sequence. The goal of model training is to minimize the sequence reconstruction error, that is, After the model training is completed, all feature data are input into the model and the difference is output. Since the test set includes the entire life cycle of the cutter head, the degradation curve of the entire life cycle drawn based on the difference also includes the entire life cycle. Each point of the degradation curve of the entire life cycle reflects the wear state of the cutter head tool at that time.

[0031] As attached Figure 3As shown in the figure, a total of 7410 sample data were obtained through feature extraction. These 7410 sample points represent the full life cycle of the wear and degradation of the cutter head tool. The first 5% of the data is taken as the healthy data of the tool to construct the training set of the model. The learning rate of the Seq2Seq structure is set to 0.005, the training batch size is 10, the time window size is 20, the number of network layers of the GRU network used is 5, and the number of hidden layers is 64. According to the monotonicity score of the feature, the top 30 features are selected as input to complete the training of the model. The training and construction of the model are completed using the data in the healthy state, and the features of the healthy data are learned. The entire sample data is used as a test set and input into the model to determine the difference between the test set data and the training set data. In order to quantify this difference, the absolute value of the difference between the predicted value and the true value is selected as the quantitative indicator of the difference, and the output difference is the full life cycle degradation curve of the cutter head tool wear.

[0032] In the attached Figure 3 In the figure, the degradation curve contains more burrs because the original data contains more abnormal mutation points. The original curve is smoothed by the sliding mean method and the curve is normalized to obtain the smooth curve shown in the figure. Figure 3 In the figure, the horizontal axis is the feature points arranged in time series, which represents the entire life cycle of the cutter head. Due to the large amount of data, it is impossible to expand all the data on the horizontal axis, so Figure 3 In the complete life cycle degradation curve, the horizontal axis and the vertical axis correspond one to one.

[0033] In general, the wear of the tool can be divided into three stages, namely, rapid wear in the early stage, gentle wear in the middle stage, and rapid wear in the late stage. According to the performance degradation curve after smoothing, it can be seen that the fitted curve is basically divided into three stages. The early sample point period is in the early rapid wear stage, and the wear rate of the cutter head tool is fast; the middle sample point period is in the middle stage of gentle wear, and the wear rate is relatively gentle; the late sample point time is in the late stage of rapid wear, the tool is in the damage stage, and the wear amount increases sharply. The overall change law of the fitting curve is basically consistent with the actual wear of the tool. It shows that the constructed seq2seq shield machine tool degradation model can be used to reflect the actual degradation of the shield machine tool.

[0034] In this application, the GRU-based Seq2Seq model is a variant of RNN, which can capture long-term dependencies in time series and solve the gradient vanishing problem of traditional RNN when performing sequence prediction.

[0035] Anything not described in this application can be achieved by adopting or drawing on existing technologies. The above is only an embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the scope of the claims of the present application.

Claims

1. A shield cutterhead health status assessment method based on Seq2Seq, characterized in that: The following steps are involved: S01) Collect the tunneling speed, total thrust, total contact force of the cutterhead, penetration, cutterhead torque, and thrust of each group of main thrust cylinders of the shield machine during its entire life cycle, and perform noise reduction and normalization processing on the collected data; S02) Calculate the time domain features of the collected data and use the monotonicity index as the score of each time domain feature, and extract the time domain features based on the score; S03) using all collected data as a test set, defining the first 5% of the test set in the time series as health status data, the health status data forming a training set, and the health status data corresponding to the health status; S04) Using the training set to train a GRU-based Seq2Seq model in a healthy state, the Seq2Seq model aims to minimize the reconstruction error, and the reconstruction error is defined as: ； In the formula, e represents the reconstruction error, U input 、U output Represent the input sequence and output sequence of the model respectively; The output value of the trained Seq2Seq model represents the absolute value of the difference between the input data and the health status data, and the absolute value of the difference is defined as the difference degree; S05) inputting the time domain features in the test set into the trained Seq2Seq model and outputting the difference degree; drawing a full life cycle degradation curve based on the difference degree; the abscissa of the full life cycle degradation curve represents the full life cycle, and the ordinate represents the normalized degradation index, and the normalized degradation index is defined as the ratio of the difference degree to the true value; S06) Smoothing the full life cycle degradation curve, and determining the wear state of the cutter disc according to its life cycle.

2. According to a Seq2Seq-based shield cutterhead health status assessment method according to claim 1, it is characterized in that: The time domain features in step S02) include: mean, standard deviation, peak, root mean square value, waveform index, peak index, pulse index, kurtosis index, and skewness index.

3. According to a Seq2Seq-based shield cutterhead health status assessment method according to claim 1, it is characterized in that: The reconstruction of the health state in step S04) adopts a positive sequence reconstruction method.

Images