Tunneling machine cutting head fault diagnosis method based on frequency gating tree topology network

Abstract

The invention discloses a heading machine cutting head fault diagnosis method based on a frequency gating tree topology network. The specific implementation process is as follows: channel rearrangement is guided by using the spectrum centroid of a channel, accurate feature decoupling is performed on the channel by using anisotropic convolution in combination with the energy distribution characteristics of high and low frequency signals, and semantic dynamic generation of a sample specific weight is captured in real time through a multi-branch convolution gating mechanism. The tree topology structure brings multi-view key features such as accurate coordinates, a long-range structure and global context to the trunk network, and in order to make full use of the key features, a tree structure-oriented distillation strategy is provided. And mapping the injected multi-view key features and the output features of the backbone network to the same semantic metric space by using a parameter sharing classification mechanism, and obtaining a semantic soft mask by means of confidence transformation. Experiments show that by means of an intelligent feature extraction mechanism, the workload of fault diagnosis can be effectively reduced, and meanwhile the diagnosis efficiency is remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of industrial equipment fault diagnosis technology, specifically a fault diagnosis method for tunnel boring machine (TBM) cutter heads based on frequency-gated tree topology networks, particularly suitable for TBM cutter head fault diagnosis scenarios. This invention achieves dynamic perception of features through frequency-gated modulation, combining the rich multi-view key features provided by the tree topology structure with a knowledge distillation strategy oriented towards the tree structure. This addresses the problems of strong background noise coupling and highly non-stationary characteristics of TBM cutter signals, which lead to difficulties in fault feature extraction and low fault diagnosis accuracy. Background Technology

[0002] With the increasing intelligence and complexity of industrial equipment, fault diagnosis technology for tunneling machine cutting heads has become a core means to ensure the safe operation of equipment. However, traditional methods have significant limitations: traditional fault diagnosis methods for tunneling machine cutting heads usually require manual feature extraction, which is time-consuming and highly dependent on expert experience; at the same time, traditional convolutional neural networks using static convolution weights lack the ability to respond instantly to changes in the content of the input signal and the ability to perceive changes in the input frequency, which is not conducive to feature extraction of tunneling machine cutting signals under strong background noise coupling; in addition, existing convolutional neural networks adopt a single serial hierarchical paradigm, which has limited feature expression capabilities, restricting the fault diagnosis and discrimination capabilities of tunneling machine cutting heads. To address the above problems, this invention proposes a fault diagnosis method for tunneling machine cutting heads based on a frequency-gated tree topology network. This invention constructs a tree topology structure to decompose multi-view key features from the input signal display, and designs a frequency-aware gating module to perform frequency rearrangement and gating modulation on the multi-view key features to achieve accurate feature extraction. Finally, a knowledge distillation strategy oriented towards the tree structure is proposed to significantly improve fault diagnosis capabilities. Summary of the Invention

[0003] This invention addresses the challenges of strong background noise coupling and highly non-stationary characteristics in tunnel boring machine (TBM) cutting signals, which lead to difficulties in fault feature extraction and low fault diagnosis accuracy. It proposes a fault diagnosis method for TBM cutting heads based on a frequency-gated tree topology network. The specific implementation steps of this invention are as follows: 1. Using the EBZ260H tunneling machine as a prototype, a tunneling machine cutting head failure test platform was built at a 1 / 3 scale. A simulated coal wall was constructed on this platform to collect tunneling machine vibration signals, in order to reproduce the actual operating conditions of the tunneling machine as accurately as possible. For the collected vibration signals, wavelet transform was used to convert the time-domain signals into a time-frequency diagram.

[0004] 2. Constructing a multi-view tree topology to extract key features from the signal, the steps are as follows: (2a) Obtain the basic spatial feature representation using the basic residual unit. For input features Where B is the batch size, C is the number of feature channels, H is the feature map height, and W is the feature map width, two 3×3 convolutions are applied, followed by two BatchNorm normalizations and one SiLU activation function to obtain... : in Indicates batch normalization. This represents a 3×3 convolution. express Activation function, then Input features The residual is calculated, and then passed through an activation function to obtain the output of the basic residual unit. This is the obtained basic spatial feature representation; (2b) Using coordinate attention to focus on key frequency bands For input features After applying one horizontal adaptive average pooling and one vertical adaptive average pooling respectively, we obtain and , Where AdaptiveAvgPool2d represents adaptive average pooling, (H,1) indicates the pooling direction is horizontal, (1,W) indicates the pooling direction is vertical, and Reshape indicates adjusting the tensor shape. Then... and Concatenating along the third dimension and generating coordinate attention weights through a series of operations such as 1×1 convolution, batch normalization, and Hardswish activation. and , in This represents the Hardswish activation function. Concat concatenates channels, and Split is the inverse operation of Concat, splitting the channels. Represents a 1×1 convolution. This represents the SiLU activation function, followed by input features and... and Element-wise multiplication yields coordinate attention output. , in, This represents element-wise multiplication with broadcast mechanism; (2c) Expanding the receptive field of low-level features using large kernel convolution For input features Apply first The activation functions BN and SiLU are obtained , Following that Apply a of Large kernel depthwise separable convolutions achieve single-layer aggregation of wide-area contextual information without significantly increasing computational complexity. This introduces a large receptive field early in the network, followed by batch normalization and SiLU activation. , Then obtain through residuals Then apply max pooling to it. The SiLU activation function is used to obtain the output of the large kernel convolutional module. , in It is max pooling; (2d) Displaying decoupled global vibration and local texture using the frequency domain decomposition module. For input features , apply Basic features are obtained using the SiLU activation function. Immediately afterwards Applying a 7×7 average pooling method yields low-frequency features. High-frequency features Depend on and The result is obtained by subtracting the residuals. in Representing 7×7 average pooling, based on signal processing theory, large-scale spatial average pooling exhibits the characteristics of a low-pass filter in the frequency domain. Its output can be considered as low-frequency components, reflecting the overall trend and slowly varying vibration modes of mechanical vibration signals. The difference between the basic characteristics and the low-frequency components represents the high-frequency components, which can capture rapidly changing information caused by impact, friction, or local defects. Low-frequency information, For high-frequency information; and Apply two different Inter-channel information modulation and transformation are performed, followed by the two being combined. Obtaining residuals ,right Apply max pooling again, and SiLU activation function, The output of the frequency domain decomposition module is obtained. ; (2e) Establish global dependencies in the feature space using a global context mixer. For input features Apply two different operations, first to Apply Spatial attention maps are obtained by reshaping and softmax. At the same time, Reshape to obtain , and Matrix multiplication yields and reshaped into Then apply Layer normalization, ReLU activation function and get , in For layer normalization, Input features Perform broadcast summation to obtain global view features X Subsequently X Apply a depth downsampling, including max pooling, And SiLU, Get the output of the global context mixer .

[0005] 3. Construct a frequency-aware gating module to realize frequency sensing and adaptive reconstruction and calibration of feature maps. The steps are as follows: (3a) Design a cross-frequency feature extraction module For the input feature map The spectral centroid is introduced as a physical indicator to quantify the frequency characteristics of each channel, thereby guiding the channel rearrangement. , They have been sorted by spectral centroid. Naturally classified into high-frequency components Mid-frequency components and low-frequency components Three groups High-frequency impulse signals typically contain significant transient energy, which is often characterized by rapid and drastic changes along the frequency axis, forming an energy distribution resembling a vertical spike. Use 5×3 anisotropic convolution kernels Modeling, Low-frequency signals change slowly and have stronger time correlation. In time-frequency graphs, they typically exhibit a continuous, smooth, and relatively long energy distribution along the time axis. Low-frequency characteristics show a horizontal distribution. Use 3×5 anisotropic convolution kernels Modeling, For those that do not show obvious directional bias A 3×3 isotropic convolution kernel is used for modeling to achieve a balanced representation of texture and structure. For features specifically extracted , and Channel splicing is performed, and batch normalization and SiLU are applied to obtain... , in Indicates the SiLU activation function, then... Apply SiLU And SiLU , and The final output of the cross-frequency feature extraction module is obtained by performing residual calculation. ; (3b) Design a multi-branch convolutional gating mechanism For the input feature map Grouping, and then getting ,in For each group Each feature is subjected to a pointwise convolution for dimensionality reduction to decrease computation, followed by batch normalization and activation functions to obtain intermediate features. In order to control the information flow and capture long-distance dependencies, the module focuses on intermediate features. Apply three independent 3×3 convolutional branches to generate reset gates. Update Gate And candidate activation. By resetting the door and the update gate Balancing historical information with current candidate information The contribution of this feature generates gating characteristics. , in This indicates element-wise multiplication, and the output feature consists of two weighted parts: one part is composed of... and Both are activated together, while the other part is retained by... Candidate features of modulation Finally, the aggregated gating features After feature processing, the final output candidate dynamic kernel tensor is generated. Meanwhile, for the input feature map Apply adaptive average pooling, The attention weight tensor is obtained through batch normalization, GeLU, Reshape, and Softmax. , in For what was obtained and By using element-wise weighted summation, the G groups of candidate kernels are merged into a single dynamic convolution kernel. , in , dynamically convolution parameters Effect on The final output of the multi-branch convolutional gating mechanism is obtained.

[0006] 4. Construct a knowledge distillation strategy oriented towards a tree structure, fully utilizing the rich features provided by the tree structure, and combining it with self-distillation to significantly improve fault detection capabilities. The steps are as follows: (4a) Multi-view key features for multi-view tree topology input Output at the current stage , This paper utilizes a parameter-sharing classification mechanism to map cross-stage tree-structured input features and stage output features to the same semantic metric space, and obtains a semantic soft mask within this space using a confidence transformation. Specifically, in each feature extraction stage of the network, a local classifier with shared weights is first used to predict the input features of that stage, and then a sigmoid activation function is used to perform a confidence transformation to map them into a semantic soft mask in the (0,1) interval. When i is 1 That is , for Dedicated classifier yes An activation function is then applied, and subsequently, this mask is multiplied element-wise with the predicted distribution of the output features, thereby dynamically modulating the confidence distribution of the output features. in This indicates element-wise multiplication. for The final classification result is obtained by having the deepest classifier guide the decisions of all shallow classifiers, while the deepest classifiers in adjacent layers also guide the decisions of shallow classifiers. All guidance directions are unidirectional, thus realizing a knowledge distillation strategy oriented towards tree structures.

[0007] 5. By integrating modules such as multi-view tree topology, frequency-aware gating module, and knowledge distillation oriented towards tree structure, a frequency-gated tree topology network is formed to realize intelligent diagnosis of tunneling machine cutting head faults.

[0008] The method of the present invention has the following advantages: (1) Feature extraction with frequency dynamic awareness: A frequency-aware gating module is constructed. Channel rearrangement is guided by the spectral centroid, and anisotropic convolution is used to accurately decouple high and low frequency features by combining the energy distribution characteristics of high and low frequency signals. Then, through a multi-branch convolution gating mechanism, real-time semantics are captured to dynamically generate sample-specific weights. Frequency awareness and adaptive reconstruction and calibration of feature maps are realized, thereby suppressing strong background noise and effectively overcoming the feature drift dilemma under non-stationary conditions and strong background noise.

[0009] (2) Rich multi-perspective key features: Construct a tree topology, extract multi-perspective key features, and establish lateral injection channels at each key stage of the backbone network. Use parallel branches to extract multi-perspective key features such as precise coordinates, long-range structure, and global context from the original signal.

[0010] (3) Efficient distillation strategy for tree structure: The multi-view key features injected into the tree structure at each stage and the dynamic output features of the backbone network are mapped to the same semantic metric space using the parameter sharing classification mechanism. In this space, the semantic soft mask is obtained by confidence transformation. Combined with multi-level cascaded self-distillation, the accuracy of fault diagnosis of the tunneling machine cutting head is improved. Attached Figure Description

[0011] Figure 1 This is the overall framework diagram of the present invention. Figure 2 This is a classification result of the confusion matrix in the AUST tunneling machine cutter head dataset under 5dB noise. Figure 3 This is a classification result of the confusion matrix on the CWRU dataset with -5dB noise. Detailed Implementation

[0012] The specific implementation steps of this invention are as follows: 1. Data Preprocessing Data Acquisition and Standardization: Collect raw vibration signals from the cutting head of an industrial tunneling machine, calculate the mean and standard deviation of the training set signals, and perform normalization on the raw signals to improve model stability.

[0013] Sample acquisition: The time series data is segmented using a sliding window of fixed length on the original vibration signal. The window length is set to 1024 and the step size is 512 to ensure that the sample covers the complete vibration cycle.

[0014] Time-frequency image acquisition: The standardized vibration signal is converted into a two-dimensional time-frequency image through continuous wavelet transform. The cmor100-1 wavelet basis function is selected, and a fixed scale of 128 is set to generate a time-frequency representation under actual sampling frequency conditions. The absolute values ​​of the wavelet coefficients are extracted to represent signal energy. A 224×224 pixel time-frequency image is generated using contour plots, where the horizontal axis represents the time series, the vertical axis represents the frequency distribution, and the image color depth corresponds to the signal energy intensity, thus forming a two-dimensional time-frequency image that combines transient characteristics in the time domain with frequency resonance characteristics in the frequency domain.

[0015] 2. Obtain key features of the tree topology from multiple perspectives By setting up multiple heterogeneous modules, a tree-like topology is constructed to obtain key features from corresponding perspectives. Specifically, basic residual units obtain basic spatial feature representations, coordinate attention focuses on key frequency bands, large-kernel convolution expands the receptive field of low-level features, frequency domain decomposition modules reveal the decoupling of global vibrations and local textures, and a global context mixer establishes global dependencies in the feature space. The obtained multi-view key features are then laterally injected into the network.

[0016] 3. The frequency sensing gating module performs frequency sensing and feature extraction. The frequency-aware gating module uses the spectral centroid as a physical indicator to quantify the frequency characteristics of each channel. Based on the energy distribution characteristics of high and low frequency features, anisotropic convolution is used to extract features. Subsequently, the ordered features are modulated through a gating mechanism to generate convolution kernel parameters that are sensitive to the input content, and then applied to the sorted features.

[0017] 4. Distillation strategies oriented towards tree structures Unlike traditional knowledge distillation that relies on external teacher models, the tree-structure-oriented distillation strategy is a self-supervised learning paradigm based on the network's own topology. At each feature extraction stage, a local classifier with shared weights first predicts the input features for that stage, and then uses a sigmoid activation function to perform a confidence transformation, mapping them to a semantic soft mask in the (0,1) interval. Subsequently, this mask is multiplied element-wise with the predicted distribution of the output features, thereby dynamically modulating the confidence distribution of the output features. The deepest classifier guides the decisions of all shallow classifiers, while deep classifiers in adjacent layers also guide the decisions of shallow classifiers; all guidance directions are unidirectional.

[0018] The effectiveness of this invention was further verified through the following experiments: On the self-built AUST tunneling machine cutter head dataset and the publicly available CWRU dataset, the method of this invention achieved an average accuracy of 98.34% and 95.99% under -5dB high noise, respectively, significantly outperforming the comparison models (ResNet50, FasterNet, and UniRepLKNet, etc.). Under strong noise conditions of -10dB, the method of this invention still maintained an accuracy of 79.39% on the AUST dataset, and recovered to 99.75% at a signal-to-noise ratio of 0dB. Ablation experiments confirmed the necessity of the cooperation of the modules. The accuracy of the frequency-aware gating module alone on the AUST and CWRU datasets at -10dB was 75.79% and 74.16%, respectively. When further fused with a tree topology structure, the accuracy reached 77.53% and 75.50%, respectively. When further fused with a distillation strategy oriented towards a tree structure, the accuracy reached the best of 79.39% and 76.80%, verifying the absolute advantage of the feature extraction of the tunneling machine cutter head fault diagnosis method based on frequency-gated tree topology network. The above experiments further validated the fault diagnosis effectiveness of the present invention.

Claims

1. A fault diagnosis method for tunnel boring machine cutting heads based on frequency-gated tree topology networks, comprising the following steps: (1) Collect vibration signals of the tunneling machine on the tunneling machine cutting head failure test platform; (2) Extract key features from the signal from multiple perspectives using a multi-view tree topology; (3) Combine the frequency sensing gating module to realize frequency sensing and adaptive reconstruction and calibration of feature maps; (4) Use a knowledge distillation strategy oriented towards tree structures to fully utilize the rich features brought by tree structures; (5) Integrating multi-view tree topology, frequency-aware gating module and knowledge distillation module for tree structure, a frequency-gated tree topology network is formed to realize intelligent diagnosis of tunneling machine cutting head faults.

2. The method for fault diagnosis of tunnel boring machine cutting head based on frequency-gated tree topology network according to claim 1, characterized in that, Step (2) involves extracting multi-view key features from a signal using a multi-view tree topology, including the following steps: (2a) Obtain the basic spatial feature representation using the basic residual unit. For input features Where B is the batch size, C is the number of feature channels, H is the feature map height, and W is the feature map width, two 3×3 convolutions are applied, followed by two BatchNorm normalizations and one SiLU activation function to obtain... : in Indicates batch normalization. This represents a 3×3 convolution. express Activation function, then Input features The residual is calculated, and then passed through an activation function to obtain the output of the basic residual unit. This is the obtained basic spatial feature representation; (2b) Using coordinate attention to focus on key frequency bands For input features After applying one horizontal adaptive average pooling and one vertical adaptive average pooling respectively, we obtain and , Where AdaptiveAvgPool2d represents adaptive average pooling, (H,1) indicates the pooling direction is horizontal, (1,W) indicates the pooling direction is vertical, and Reshape indicates adjusting the tensor shape. Then... and Concatenating along the third dimension and generating coordinate attention weights through a series of operations such as 1×1 convolution, batch normalization, and Hardswish activation. and , in This represents the Hardswish activation function. Concat concatenates channels, and Split is the inverse operation of Concat, splitting the channels. Represents a 1×1 convolution. This represents the SiLU activation function, followed by input features and... and Element-wise multiplication yields coordinate attention output. , in, This represents element-wise multiplication with broadcast mechanism; (2c) Expanding the receptive field of low-level features using large kernel convolution For input features Apply first The activation functions BN and SiLU are obtained , Following that Apply a of Large kernel depthwise separable convolutions achieve single-layer aggregation of wide-area contextual information without significantly increasing computational complexity. This introduces a large receptive field early in the network, followed by batch normalization and SiLU activation. , Then obtain through residuals Then apply max pooling to it. The SiLU activation function is used to obtain the output of the large kernel convolutional module. , in It is max pooling; (2d) Displaying decoupled global vibration and local texture using the frequency domain decomposition module. For input features , apply Basic features are obtained using the SiLU activation function. Immediately afterwards Applying a 7×7 average pooling method yields low-frequency features. High-frequency features Depend on and The result is obtained by subtracting the residuals. in Representing 7×7 average pooling, based on signal processing theory, large-scale spatial average pooling exhibits the characteristics of a low-pass filter in the frequency domain. Its output can be considered as low-frequency components, reflecting the overall trend and slowly varying vibration modes of mechanical vibration signals. The difference between the basic characteristics and the low-frequency components represents the high-frequency components, which can capture rapidly changing information caused by impact, friction, or local defects. Low-frequency information, For high-frequency information; and Apply two different Inter-channel information modulation and transformation are performed, followed by the two being combined. Obtaining residuals ,right Apply max pooling again, And the SiLU activation function: The output of the frequency domain decomposition module is obtained. ; (2e) Establish global dependencies in the feature space using a global context mixer. For input features Apply two different operations, first to Apply Spatial attention maps are obtained by reshaping and softmax. At the same time, Reshape to obtain , and Matrix multiplication yields and reshaped into Then apply Layer normalization, ReLU activation function and get , in For layer normalization, Input features Perform broadcast summation to obtain global view features X Subsequently X Apply a depth downsampling, including max pooling, And SiLU, Get the output of the global context mixer .

3. The method for fault diagnosis of tunnel boring machine cutting head based on frequency-gated tree topology network according to claim 1, characterized in that, Step (3) involves combining the frequency sensing gating module to achieve frequency sensing and adaptive reconstruction and calibration of the feature map. The steps are as follows: (3a) Design a cross-frequency feature extraction module For the input feature map The spectral centroid is introduced as a physical indicator to quantify the frequency characteristics of each channel, thereby guiding the channel rearrangement. , They have been sorted by spectral centroid. Naturally classified into high-frequency components Mid-frequency components and low-frequency components Three groups High-frequency impulse signals typically contain significant transient energy. Its characteristics often exhibit rapid and drastic changes along the frequency axis, forming an energy distribution resembling a vertical spike. Use 5×3 anisotropic convolution kernels Modeling, Low-frequency signals change slowly and have stronger time correlation. In time-frequency graphs, they typically exhibit a continuous, smooth, and relatively long energy distribution along the time axis. Low-frequency characteristics show a horizontal distribution. Use 3×5 anisotropic convolution kernels Modeling, For those that do not show obvious directional bias A 3×3 isotropic convolution kernel is used for modeling to achieve a balanced representation of texture and structure. For features specifically extracted , and Channel splicing is performed, and batch normalization and SiLU are applied to obtain... , in Indicates the SiLU activation function, then... Apply SiLU And SiLU , and The final output of the cross-frequency feature extraction module is obtained by performing residual calculation. ; (3b) Design a multi-branch convolutional gating mechanism For the input feature map Grouping, and then getting ,in For each group Each feature is subjected to a pointwise convolution for dimensionality reduction to decrease computation, followed by batch normalization and activation functions to obtain intermediate features. In order to control the information flow and capture long-distance dependencies, the module focuses on intermediate features. Apply three independent 3×3 convolutional branches to generate reset gates. Update Gate And candidate activation. By resetting the door and the update gate Balancing historical information with current candidate information The contribution of this feature generates gating characteristics. , in This indicates element-wise multiplication, and the output feature consists of two weighted parts: one part is composed of... and Both are activated together, while the other part is retained by... Candidate features of modulation Finally, the aggregated gating features After feature processing, the final output candidate dynamic kernel tensor is generated. Meanwhile, for the input feature map Apply adaptive average pooling, The attention weight tensor is obtained through batch normalization, GeLU, Reshape, and Softmax. , in For what was obtained and By using element-wise weighted summation, the G groups of candidate kernels are merged into a single dynamic convolution kernel. , in , dynamically convolution parameters Effect on The final output of the multi-branch convolutional gating mechanism is obtained.

4. The method for fault diagnosis of tunnel boring machine cutting head based on frequency-gated tree topology network according to claim 1, characterized in that... The knowledge distillation strategy for tree structures described in step (4) fully utilizes the rich features from multiple perspectives brought by the tree structure and combines self-distillation to improve fault detection capabilities. The steps are as follows: (4a) Multi-view key features for multi-view tree topology input Output at the current stage , This paper utilizes a parameter-sharing classification mechanism to map cross-stage tree-structured input features and stage output features to the same semantic metric space, and obtains a semantic soft mask within this space using a confidence transformation. Specifically, in each feature extraction stage of the network, a local classifier with shared weights is first used to predict the input features of that stage, and then a sigmoid activation function is used to perform a confidence transformation to map them into a semantic soft mask in the (0,1) interval. When i is 1 That is , for Dedicated classifier yes An activation function is then applied, and subsequently, this mask is multiplied element-wise with the predicted distribution of the output features, thereby dynamically modulating the confidence distribution of the output features. in This indicates element-wise multiplication. for The final classification result is obtained by having the deepest classifier guide the decisions of all shallow classifiers, while the deepest classifiers in adjacent layers also guide the decisions of shallow classifiers. All guidance directions are unidirectional, thus realizing a knowledge distillation strategy oriented towards tree structures.

Images