Shield cutter wear prediction and early warning system based on online monitoring
By combining signal acquisition, separation, and K-means algorithm with convolutional neural network, the problem of shield cutterhead wear location was solved, achieving accurate wear assessment and timely early warning, thus reducing the risk of shield construction.
Patent Information
- Application Number
- CN202610055657.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-16
- Publication Date
- 2026-02-17
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Existing technology cannot accurately locate cutter wear at different radial positions and circumferential angles on the shield cutterhead, causing early warning information to lose its precise guidance value, making it impossible to identify local risks in the early stage, which may lead to a chain of failures and delays in project progress.
Vibration signals and temperature data are acquired through a signal acquisition module, high-frequency components are separated using a signal separation module, local abnormal areas are identified using a K-means algorithm, and the degree of wear is assessed using a convolutional neural network to generate targeted early warning signals.
It enables accurate assessment and failure prediction of wear in local areas of the tunnel boring machine cutterhead, improves the accuracy of anomaly detection and the timeliness of early warning, and reduces operational risks.
Smart Images
Figure CN121542969A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of tunnel boring machine data processing technology, and more specifically, to a tunnel boring machine cutter wear prediction and early warning system based on online monitoring. Background Technology
[0002] Shield tunneling is a core component of tunnel engineering. The various cutting tools installed on the cutterhead interact directly with the rock and soil in complex strata. The wear condition of the cutting tools directly affects tunneling efficiency, project cost, and construction safety. Therefore, effective prediction and early warning of cutting tool wear is crucial to ensuring safe and efficient shield tunneling.
[0003] Existing technologies attempt to achieve online monitoring by installing sensors, but the core technical challenge lies in the fact that tool wear is not a uniform, overall event, but rather exhibits significantly different local behaviors at different locations in the cutterhead space. This spatial variability stems primarily from the fact that the loads and geological processes borne by tools at different radial positions and circumferential angles on the cutterhead are completely different. The working environments and wear mechanisms of the center tool, the front tool, and the edge tool are all different. However, traditional monitoring methods often treat the cutterhead or tool group as a whole for general signal analysis, lacking the ability to accurately locate the wear points. This makes it impossible to distinguish whether the abnormality is of a specific tool or the general wear of the entire cutterhead. Because the monitoring signal cannot be accurately correlated with the specific spatial location, even if the system detects abnormal vibration or temperature rise, it is difficult to determine the exact location of the failure, thus rendering the early warning information ineffective in providing precise guidance. This lack of positioning capability further raises a unique technical challenge: how to extract and identify the failure trend characteristics of tools at specific points from the mixed global signals. For example, when a certain edge area of the cutterhead encounters hard rock or isolated rocks, a few tools in that area will wear faster first, and the resulting vibration signals will mix with the signals of other normal tools on the cutterhead. If the monitoring system cannot separate these signals and correspond them to specific tools through a spatial distribution model, it will be unable to identify this local risk in the early stages, and the early warning will be delayed. By the time the wear spreads and the signal is significantly enhanced, it may have already caused cascading damage to adjacent tools, which may not only trigger chain failures such as cutterhead damage, but also seriously slow down the project progress due to unplanned downtime for tool replacement. Summary of the Invention
[0004] In order to overcome the shortcomings of the existing technology, the present invention provides a shield cutter wear prediction and early warning system based on online monitoring, which aims to solve the problems in the above-mentioned existing technology.
[0005] The technical solution adopted by this invention to solve its technical problem is: a shield tunneling cutter wear prediction and early warning system based on online monitoring, comprising: The signal acquisition module is used to acquire vibration signals and temperature data from different radial positions and circumferential angles of the cutter head, and obtain the raw signal sequence of multiple channels; The signal separation module is used to separate the high-frequency components of the original signal sequence of each channel to obtain the time-frequency feature sequence; The identifier generation module is used to obtain the amplitude from the time-frequency feature sequence, filter out the corresponding channels with amplitudes exceeding a preset threshold, and generate preliminary abnormal channel identifiers; The abnormal area confirmation module is used to determine the local abnormal area on the cutter head based on the initial coordinates of the data collected from the channel with the preliminary abnormal channel identifier on the cutter head, using the K-means algorithm. The wear degree determination module is used to acquire the original signal sequence of the channel related to the local abnormal region, obtain the time series dataset, and input it into the convolutional neural network for processing to determine the wear degree score of the local abnormal region. The failure prediction module is used to process the changes in the original signal sequence of the channel related to the local abnormal area when the wear degree score of the local abnormal area is higher than a preset warning threshold, and obtain the failure prediction score corresponding to the local abnormal area. The early warning module is used to generate targeted early warning signals based on the failure prediction score.
[0006] Preferably, in the signal acquisition module, vibration signals and temperature data are acquired from different radial positions and circumferential angles of the cutterhead by a sensor array arranged on the cutterhead of the tunnel boring machine. After weighted summation of vibration signals and temperature data at the same time stamp in the same channel, the original signal sequences of multiple channels are obtained.
[0007] Optionally, the signal separation module is used to decompose the original signal sequence of each channel by wavelet transform to obtain the high-frequency signal subsequence and low-frequency signal subsequence corresponding to each channel. This is used to perform a short-time Fourier transform on the high-frequency signal subsequence. The short-time Fourier transform uses a Hanning window function and a fixed time step to generate a time-frequency feature sequence.
[0008] Specifically, the abnormal area confirmation module is used to obtain the initial coordinate set of the data collected by the channel with preliminary abnormality identification on the tool head, and the initial coordinate set contains multiple coordinate points; In the K-means algorithm, a preset number of cluster centers are first initialized. Each coordinate point in the initial coordinate set is assigned to the nearest cluster center to obtain the cluster partitioning result. Each cluster in the cluster partitioning result is defined as a local abnormal region on the cutter head.
[0009] Preferably, the initial cluster partition is obtained by processing each coordinate point in the initial coordinate set with a preset number of cluster centers using the K-means algorithm; The position of the cluster center is updated using the mean value of the coordinate points within each cluster in the initial cluster partitioning. If the position of the cluster center changes, the coordinate points in the initial coordinate set are reassigned to the nearest cluster center, and the cluster center is updated again until the position of the cluster center no longer changes. The clustering result is then determined, and each cluster in the clustering result is defined as a local abnormal region on the cutter head.
[0010] Specifically, the wear degree determination module is used to obtain the original signal sequence of the channel related to the local abnormal region and construct a time series dataset; This is used to perform convolution operations on the time series dataset using a convolutional neural network to extract time series features, wherein the input of the convolutional neural network is the time series dataset, and the output is a first feature vector, thereby obtaining the first feature vector; This is used to perform dimensionality integration on the first feature vector through a linear transformation layer, wherein the input of the linear transformation layer is the first feature vector and the output is a second feature vector, thereby obtaining the second feature vector; This is used to calculate the wear degree score of a local abnormal region based on the second feature vector, wherein the wear degree score is obtained by summing the elements of the second feature vector after performing a dot product with a preset weight array.
[0011] It is worth noting that the failure prediction module is used to obtain the original signal sequence of the channel related to the local abnormal area where the wear degree score is higher than the preset warning threshold. The module uses a long short-term memory network to fuse the temporal dependency of the original signal sequence. The input of the long short-term memory network is the original signal sequence. The sequence dependency is processed by the gating unit of the long short-term memory network, and the output is the failure trend parameter value of the channel. This is used to calculate the failure prediction score from the failure trend parameter values of the channel. The failure prediction score is obtained by weighted summation, and the weighted summation formula is S=∑(m_i w_i), where m_i is the failure trend parameter value of the i-th channel and w_i is the preset weight of the i-th channel.
[0012] Preferably, in the early warning module, the average failure prediction score of all local abnormal areas with wear degree scores higher than a preset warning threshold is calculated, and the cutter head is classified into risks based on the average value, wherein the risk classification is divided by comparing the average value with a preset classification threshold; and a targeted early warning signal is generated based on the risk classification.
[0013] The beneficial effects of this invention are as follows: In the shield cutterhead wear prediction and early warning system based on online monitoring, it addresses the problem of abnormal vibration signals and temperature data caused by differences in radial position and circumferential angle during cutterhead operation. It integrates signal decomposition, time-frequency feature extraction, and multi-dimensional feature fusion using convolutional neural networks to solve the complex operational challenges of assessing wear levels and predicting failures in local cutterhead areas. This solution extracts high-frequency components from each channel using signal decomposition, initially locates abnormal channels using amplitude filtering, then uses clustering algorithms to determine local abnormal areas of the cutterhead, and performs feature fusion on the time-series data of relevant channels using a convolutional neural network to accurately assess the wear level score. For areas exceeding the warning threshold, this invention further predicts failure risk through sequence changes and generates targeted early warning signals. This method effectively improves the accuracy of cutterhead anomaly detection and the timeliness of early warning, significantly reducing operational risks. Attached Figure Description
[0014] Figure 1 This is a system structure diagram of a shield tunneling cutter wear prediction and early warning system based on online monitoring.
[0015] Figure 2 This is a schematic diagram of the module functions of a shield tunneling cutter wear prediction and early warning system based on online monitoring. Detailed Implementation
[0016] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings. It should be noted that these descriptions are for the purpose of aiding understanding the present invention, but do not constitute a limitation thereof. Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.
[0017] Combination Figure 1 and Figure 2 The shield tunneling cutter wear prediction and early warning system shown includes: The signal acquisition module is used to acquire vibration signals and temperature data from different radial positions and circumferential angles of the cutter head, and obtain the raw signal sequence of multiple channels; The signal separation module is used to separate the high-frequency components of the original signal sequence of each channel to obtain the time-frequency feature sequence; The identification generation module is used to obtain amplitude values from the time-frequency feature sequence, filter out corresponding channels with amplitude values exceeding a preset threshold, and generate preliminary abnormal channel identifications. The time-frequency feature sequence is a sequence of frames composed of three-dimensional information of time, frequency, and amplitude. Each frame is a fixed-length time slice, which is further divided into several grid points according to frequency. Each grid point stores an amplitude value, which reflects the energy intensity of the signal at the corresponding frequency. In principle, the amplitude can be extracted from each column of data in the time-frequency feature sequence to obtain the energy distribution of the corresponding frequency. This method can help identify hidden abnormal patterns in the signal, especially abnormal vibrations or friction phenomena that may occur during the operation of the cutter head. For the step of judging whether the amplitude exceeds the preset threshold, a reasonable threshold standard can be set, such as setting the preset threshold to 1.5 times the amplitude under normal operating conditions based on historical data analysis. The principle of this preset threshold judgment is that an excessive amplitude often means that the cutter head may be experiencing abnormal resistance or friction, requiring further attention. In this way, channels with potential problems can be quickly screened out. During the generation of initial abnormal channel identifiers, the system marks channels exceeding a preset threshold as abnormal candidates, and then assigns initial abnormal channel identifiers to these candidates. This initial screening and identification mechanism can significantly improve the efficiency of problem localization; The abnormal area confirmation module is used to determine the local abnormal area on the cutter head based on the initial coordinates of the data collected from the channel with the preliminary abnormal channel identifier on the cutter head, using the K-means algorithm. The wear degree determination module is used to acquire the original signal sequence of the channel related to the local abnormal region, obtain the time series dataset, and input it into the convolutional neural network for processing to determine the wear degree score of the local abnormal region. The failure prediction module is used to process the changes in the original signal sequence of the channel related to the local abnormal area when the wear degree score of the local abnormal area is higher than a preset warning threshold, and obtain the failure prediction score corresponding to the local abnormal area. The early warning module is used to generate targeted early warning signals based on the failure prediction score.
[0018] In the aforementioned shield tunneling cutterhead wear prediction and early warning system based on online monitoring, this system addresses the anomalies in vibration signals and temperature data caused by variations in radial position and circumferential angle during cutterhead operation. It integrates signal decomposition, time-frequency feature extraction, and multi-dimensional feature fusion using convolutional neural networks to solve the complex operational challenges of assessing wear levels and predicting failures in localized areas of the cutterhead. This solution extracts high-frequency components from each channel using signal decomposition, initially locating abnormal channels by amplitude filtering. Subsequently, clustering algorithms are used to determine localized abnormal areas on the cutterhead, and convolutional neural networks are used to fuse the time-series data of relevant channels to accurately assess wear levels. For areas exceeding warning thresholds, this invention further predicts failure risk through sequence changes, generating targeted early warning signals. This method effectively improves the accuracy of cutterhead anomaly detection and the timeliness of early warnings, significantly reducing operational risks.
[0019] Preferably, in the signal acquisition module, vibration signals and temperature data are acquired from different radial positions and circumferential angles of the cutterhead by a sensor array arranged on the cutterhead of the tunnel boring machine. After weighted summation of vibration signals and temperature data at the same time stamp in the same channel, the original signal sequences of multiple channels are obtained.
[0020] The sensor array on the cutterhead of a tunnel boring machine typically employs a multi-point distribution strategy to cover the comprehensive monitoring needs of the cutterhead. Specifically, this array includes vibration sensors and temperature sensors, which are fixed at different radial positions on the cutterhead, extending outward from the central axis to multiple layers at the edges, and evenly distributed at angles around the circumference, such as placing a sensor point every 30 degrees, thereby ensuring the comprehensiveness and representativeness of the data acquisition.
[0021] Weighted summation of vibration signals and temperature data from the same channel and time stamp is performed to fuse multi-source data and generate a more reliable signal. Specifically, a channel can be understood as a logical grouping within a sensor array. For example, vibration sensors and temperature sensors at the same radial position and angle are grouped into one channel. For data at the same time stamp, the vibration signal collected by the vibration sensor and the temperature data collected by the temperature sensor in that channel are first collected. Then, a weighted summation formula is applied, where the weights are preset, for example, a weight of 0.4 for the vibration signal and 0.6 for the temperature data. After weighted summation, a weighted processed signal is obtained. Arranging the weighted processed signal by time stamp yields the original signal sequence for that channel. After obtaining the original signal sequences from multiple channels, these sequences can serve as the basis for subsequent analysis.
[0022] Optionally, the signal separation module is used to decompose the original signal sequence of each channel by wavelet transform to obtain the high-frequency signal subsequence and low-frequency signal subsequence corresponding to each channel. This is used to perform a short-time Fourier transform on the high-frequency signal subsequence. The short-time Fourier transform uses a Hanning window function and a fixed time step to generate a time-frequency feature sequence.
[0023] Wavelet transform is a signal processing tool that can decompose a signal into sub-signals with different frequency components. By selecting appropriate wavelet basis functions, such as the Daubechies wavelet, multi-scale analysis of time series can be performed.
[0024] Specifically, wavelet transform separates the original signal sequence into high-frequency detail parts and low-frequency approximation parts through recursive decomposition. For example, applying first-level wavelet decomposition to the original signal sequence of a channel will generate high-frequency parts that represent instantaneous sudden changes, while low-frequency parts capture the overall trend. The resulting subsequence can better reveal the local anomalies of the cutterhead during the tunneling process.
[0025] After obtaining the high-frequency and low-frequency signal subsequences for each channel, the high-frequency portion typically contains more noise and transient information, requiring further analysis to extract useful features. Specifically, the high-frequency signal subsequence is used for subsequent peak capture and fault early warning. The process of performing a short-time Fourier transform on the high-frequency signal subsequence involves segmenting the sequence and using a Hanning window function to reduce spectral leakage. The Hanning window is a weighting function with a bell-shaped curve shape that smooths the window edges, thereby improving spectral resolution. A fixed time step, such as 0.5 seconds per window, ensures temporal consistency of the transform. The principle of the short-time Fourier transform combines the frequency analysis of the Fourier transform with the localization of the time window, calculating the spectrum of each segment through a sliding window to generate a time-frequency representation. In tunnel engineering, using the Hanning window and a fixed step size effectively handles the non-stationarity of the cutterhead signal. For example, when the cutterhead encounters hard rock layers, the time-frequency characteristic sequence will highlight high-frequency energy peaks.
[0026] Specifically, the abnormal area confirmation module is used to obtain the initial coordinate set of the data collected by the channel with preliminary abnormality identification on the tool head, and the initial coordinate set contains multiple coordinate points; In the K-means algorithm, a preset number of cluster centers are first initialized. Each coordinate point in the initial coordinate set is assigned to the nearest cluster center to obtain the cluster partitioning result. Each cluster in the cluster partitioning result is defined as a local abnormal region on the cutter head.
[0027] It is worth noting that the initial cluster partitioning is obtained by processing each coordinate point in the initial coordinate set using a preset number of cluster centers from the K-means algorithm. The position of the cluster center is updated using the mean value of the coordinate points within each cluster in the initial cluster partitioning. If the position of the cluster center changes, the coordinate points in the initial coordinate set are reassigned to the nearest cluster center, and the cluster center is updated again until the position of the cluster center no longer changes. The clustering result is then determined, and each cluster in the clustering result is defined as a local abnormal region on the cutter head.
[0028] The initial coordinate set typically consists of multiple coordinate points, each corresponding to a sensor channel at a specific location on the tool turret. By analyzing these coordinate points, the distribution of abnormal areas can be further located.
[0029] When initializing a preset number of cluster centers and assigning coordinate points, the number of initial cluster centers can be set to three, based on the physical structure of the cutterhead and the possible characteristics of abnormal distribution. Assuming the initial cluster centers are located in the inner, middle, and outer rings of the cutterhead, the system will assign each coordinate point to the nearest cluster center, forming the initial cluster division. For example, point A, closer to the inner ring, is assigned to the inner ring cluster center, while point B, closer to the outer ring, is assigned to the outer ring cluster center. This initial division method helps to quickly group abnormal points. The distance is calculated by using Euclidean distance between the coordinate point and the initial cluster center.
[0030] When adjusting cluster centers, the system continuously updates the cluster center position of each cluster by using the average coordinate points in the initial cluster as the new cluster center position. It then reallocates coordinate points to the new nearest cluster center, iterating until the cluster center position stabilizes. This dynamic adjustment method more accurately reflects the true distribution of anomalies. Finally, the coordinate points corresponding to each cluster where the cluster center position no longer changes are defined as a local anomaly region. The definition of these local anomaly regions provides precise location information for subsequent toolhead maintenance. This partitioning method significantly improves the targeted nature of problem troubleshooting.
[0031] Preferably, the wear degree determination module is used to obtain the original signal sequence of the channel related to the local abnormal region and construct a time series dataset; This is used to perform convolution operations on the time series dataset using a convolutional neural network to extract time series features, wherein the input of the convolutional neural network is the time series dataset, and the output is a first feature vector, thereby obtaining the first feature vector; This is used to perform dimensionality integration on the first feature vector through a linear transformation layer, wherein the input of the linear transformation layer is the first feature vector and the output is a second feature vector, thereby obtaining the second feature vector; This is used to calculate the wear degree score of a local abnormal region based on the second feature vector, wherein the wear degree score is obtained by summing the elements of the second feature vector after performing a dot product with a preset weight array.
[0032] In one possible implementation, for a identified local anomaly area on the tunnel boring machine cutterhead, it is first necessary to obtain the original signal sequences of the associated channels. By aligning and combining these sequences in time, a time-series dataset is constructed. This dataset can be a multi-dimensional matrix, where rows represent different channels and columns represent time steps, thus providing structured input data for subsequent analysis.
[0033] When performing convolutional operations on the time-series dataset using a convolutional neural network, the dataset is used as the input layer. The convolutional neural network is primarily used here to extract temporal features. It slides a one-dimensional convolutional kernel along the time dimension to capture local patterns in the sequence, such as high-frequency components of abnormal signals. In the tunnel boring machine monitoring scenario, the time-series dataset is input into a network containing multiple convolutional layers. The first layer uses a convolutional kernel with a kernel size of 5 and a stride of 1 to perform convolutional calculations on the sequence of each channel, generating feature maps. These feature maps are processed by activation functions such as ReLU, and then further downsampled through pooling layers to reduce dimensionality. Finally, the network outputs a first feature vector, which may be a 128-dimensional array encapsulating the abstract temporal features extracted from the original sequence, such as the frequency and amplitude changes of abnormal patterns.
[0034] When fusing the first feature vector, a linear transformation layer is used to achieve dimensionality integration. This linear transformation layer is essentially a fully connected layer or matrix multiplication operation. It maps the input first feature vector into a new space to fuse the weights of different features. In tool disc anomaly monitoring, if the first feature vector captures the temporal patterns of multiple channels, the linear transformation layer can use a pre-trained weight matrix to linearly combine the vectors, outputting a more compact second feature vector, such as reducing it from 128 dimensions to 64 dimensions. This process helps to integrate cross-channel information and avoid redundancy.
[0035] The process of calculating the wear score of a local abnormal area based on the second feature vector can be achieved through a dot product operation with a preset weight array. Specifically, the preset weight array is trained based on historical data, such as wear correlation coefficients corresponding to different feature dimensions. Then, each element of the second feature vector is multiplied by its corresponding weight, and all results are summed to obtain a score. In tunnel boring machine operations, this score may represent the wear degree of the outer ring area.
[0036] Specifically, the failure prediction module is used to obtain the original signal sequence of the channel related to the local abnormal area where the wear degree score is higher than the preset warning threshold, and to use a long short-term memory network to fuse the temporal dependency of the original signal sequence. The input of the long short-term memory network is the original signal sequence, the sequence dependency is processed by the gating unit of the long short-term memory network, and the output is the failure trend parameter value of the channel. This is used to calculate the failure prediction score from the failure trend parameter values of the channel. The failure prediction score is obtained by weighted summation, and the weighted summation formula is S=∑(m_i w_i), where m_i is the failure trend parameter value of the i-th channel and w_i is the preset weight of the i-th channel.
[0037] Specifically, Long Short-Term Memory (LSTM) networks are a special type of recurrent neural network specifically designed to handle long-term dependencies in sequential data. They manage information flow through gating units such as forget gates, input gates, and output gates, avoiding the vanishing gradient phenomenon found in traditional recurrent networks. In the scenario of anomaly monitoring of a tunnel boring machine (TBM) cutterhead, each original signal sequence in a local anomaly region where the wear score exceeds a preset warning threshold is used as input to the LTM network. These sequences are fed into the network; the forget gate determines which historical information is retained, the input gate controls the addition of new information, and the output gate generates the output of the current state. Through this gating mechanism, the network can capture long-term patterns in the sequence, such as the gradual decay trend of the signal, and ultimately output the failure trend parameter value corresponding to each channel, representing the expected failure level of each channel within a future time window.
[0038] The process of calculating the failure prediction score from the failure trend parameter values of the tunnels is achieved through a weighted summation formula. This formula considers the sum of the product of the trend of each tunnel in the local anomaly region and a preset weight. By summing all the elements using weighted summation, a comprehensive score is obtained to represent the failure prediction score. This score represents the overall failure risk of a local area in tunnel boring machine (TBM) operations. The technical goal of this calculation method is to provide a quantitative indicator, support automated decision-making, and reduce downtime losses caused by sudden failures in operations.
[0039] Optionally, in the early warning module, the average failure prediction score of all local abnormal areas with wear scores higher than a preset warning threshold is calculated, and the cutterhead is classified into risks based on the average value, wherein the risk classification is performed by comparing the average value with a preset classification threshold; and a targeted early warning signal is generated based on the risk classification.
[0040] This average calculation helps integrate risks from multiple regions, avoiding the impact of single-region bias on the overall assessment. In one possible implementation, the cutterhead is classified into risk levels based on this average value, determined by comparison with preset classification thresholds. For example, a threshold below 0.5 indicates low risk, 0.5 to 0.8 indicates medium risk, and above 0.8 indicates high risk. In actual operation, if the average value is 0.85, it is classified as high risk. This relates to the principle of threshold setting, which is based on statistical analysis of historical failure data to ensure that the classification reflects the true probability of failure, thereby guiding maintenance decisions. Specifically, the principle of risk classification is to map the average value to different levels. For example, low risk means the cutterhead can continue to operate, medium risk requires enhanced monitoring, and high risk requires immediate intervention. This classification optimizes resource allocation and reduces economic losses caused by unexpected downtime.
[0041] When generating targeted early warning signals based on risk classification, for example, a red alarm signal is issued for high risk, including audible prompts and on-screen displays, notifying the operator to stop the machine for inspection; a yellow signal is issued for medium risk, suggesting increased lubrication and maintenance. This signal generation process is achieved through an integrated system, with the warning content tailored to the specific classification.
[0042] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and these variations still fall within the protection scope of the present invention.
Claims
1. An online monitoring based shield cutter wear prediction and early warning system, characterized in that, The method comprises the following steps: A signal acquisition module is used to acquire vibration signals and temperature data from different radial positions and circumferential angles of the cutter head, and to obtain original signal sequences of multiple channels; A signal separation module is used to separate high-frequency components of the original signal sequences of the channels, and to obtain time-frequency feature sequences; An identification generation module is used to obtain amplitudes from the time-frequency feature sequences, to screen corresponding channels with amplitudes exceeding a preset threshold, and to generate preliminary abnormal channel identifications; An abnormal area confirmation module is used to determine local abnormal areas on the cutter head through a K-means algorithm according to initial coordinates of data collected by the channels with preliminary abnormal channel identifications; A wear degree determination module is used to obtain original signal sequences of the channels related to the local abnormal areas, to obtain time series data sets, and to input the time series data sets into a convolutional neural network for processing, so as to determine a wear degree score of the local abnormal areas; A failure prediction module is used to process changes of the original signal sequences of the channels related to the local abnormal areas if the wear degree score of the local abnormal areas is higher than a preset warning threshold, so as to obtain a failure prediction score corresponding to the local abnormal areas; A warning module is used to generate a targeted warning signal according to the failure prediction score.
2. The online monitoring based shield cutter wear prediction and early warning system according to claim 1, characterized in that: In the signal acquisition module, a sensor array arranged on the cutter head of the shield tunneling machine is used to acquire vibration signals and temperature data from different radial positions and circumferential angles of the cutter head, and the vibration signals and the temperature data of the same channel and the same time stamp are weighted and added to obtain original signal sequences of multiple channels.
3. The system according to claim 1, characterized in that: The signal separation module is used to decompose the original signal sequences of the channels through wavelet transform, to obtain high-frequency signal subsequences and low-frequency signal subsequences corresponding to the channels; The signal separation module is used to perform short-time Fourier transform on the high-frequency signal subsequences, the short-time Fourier transform adopts a Hanning window function and a fixed time step, and time-frequency feature sequences are generated.
4. The online monitoring based shield cutter wear prediction and early warning system according to claim 1, characterized in that: The abnormal area confirmation module is used to obtain an initial coordinate set of data collected by the channels with preliminary abnormal identifications on the cutter head, and the initial coordinate set contains multiple coordinate points. In the K-means algorithm, a preset number of cluster centers of the K-means algorithm is initialized, each coordinate point in the initial coordinate set is assigned to the nearest cluster center, a clustering cluster division result is obtained, and each cluster in the clustering cluster division result is defined as a local abnormal area on the cutter head.
5. The online monitoring based shield cutter wear prediction and early warning system according to claim 4, characterized in that: After each coordinate point in the initial coordinate set is processed through the preset number of cluster centers of the K-means algorithm, an initial cluster division is obtained. The mean values of the coordinate points in each cluster in the initial cluster division are used to update the positions of the cluster centers. If the positions of the cluster centers change, the coordinate points in the initial coordinate set are re-assigned to the nearest cluster centers, and the cluster centers are updated again until the positions of the cluster centers no longer change, a clustering cluster division result is determined, and each cluster in the clustering cluster division result is defined as a local abnormal area on the cutter head.
6. The online monitoring based shield cutter wear prediction and early warning system according to claim 1, characterized in that: The wear degree determination module is used to obtain original signal sequences of the channels related to the local abnormal areas, and to construct time series data sets. extracting time sequence features by performing convolution operation on the time sequence dataset using a convolutional neural network, wherein an input of the convolutional neural network is the time sequence dataset, and an output of the convolutional neural network is a first feature vector, and the first feature vector is obtained; performing dimension integration on the first feature vector by using a linear transformation layer, wherein an input of the linear transformation layer is the first feature vector, and an output of the linear transformation layer is a second feature vector, and the second feature vector is obtained; calculating a wear degree score of a local abnormal area according to the second feature vector, wherein the wear degree score is obtained by performing element-wise multiplication on the second feature vector and a preset weight array and then summing the result.
7. The online monitoring based shield cutter wear prediction and early warning system according to claim 1, characterized in that: The failure prediction module is configured to obtain a raw signal sequence of a channel related to a local abnormal area whose wear degree score is higher than a preset warning threshold, and fuse time sequence dependency of the raw signal sequence by using a long short-term memory network, wherein an input of the long short-term memory network is the raw signal sequence, time sequence dependency is processed by using a gate unit of the long short-term memory network, and an output of the long short-term memory network is a failure trend parameter value of the channel. calculating a failure prediction score from the failure tendency parameter values of the channels, wherein the failure prediction score is obtained by a weighted summation, the weighted summation formula being S =∑(m_i w_i), where m_i is the failure tendency parameter value of the i-th channel, and w_i is a preset weight of the i-th channel.
8. The online monitoring based shield cutter wear prediction and early warning system according to claim 1, characterized in that: In the early warning module, an average value of failure prediction scores of all local abnormal areas whose wear degree scores are higher than the preset warning threshold is calculated, a risk classification of the cutter head is performed according to the average value, wherein the risk classification is divided by comparing the average value with a preset classification threshold, and a targeted early warning signal is generated according to the risk classification.