CN-121980243-A - Development machine cutting head fault diagnosis method based on frequency gating tree topology network

Abstract

The invention discloses a fault diagnosis method for a cutting head of a heading machine based on a frequency-gating tree topology network. The specific implementation process includes utilizing spectrum centroid of a channel to guide channel rearrangement, combining energy distribution characteristics of high-low frequency signals to perform characteristic accurate decoupling on the channel by using anisotropic convolution, and capturing semantic dynamics in real time to generate sample specific weights through a multi-branch convolution gating mechanism. The tree topology structure brings multi-view key features such as accurate coordinates, a long-range structure, global context and the like to the backbone network, a distillation strategy oriented to the tree structure is provided for fully utilizing the key features, the injected multi-view key features and the output features of the backbone network are mapped to the same semantic measurement space by utilizing a parameter sharing classification mechanism, and a semantic soft mask is obtained by means of confidence transformation. Experiments show that the invention can effectively reduce the workload of fault diagnosis and remarkably improve the diagnosis efficiency through an intelligent feature extraction mechanism.

Inventors

• SU SHUZHI
• YANG MENGYANG
• MA TIANBING
• ZHU YANMIN
• ZHANG KEXUE

Assignees

• 安徽理工大学

Dates

Publication Date

20260505

Application Date

20260123

Claims (4)

1. 1. The fault diagnosis method for the cutting head of the heading machine based on the frequency gate tree topology network comprises the following steps: (1) Collecting a vibration signal of the heading machine on a fault experiment platform of a cutting head of the heading machine; (2) Extracting multi-view key features from the signals by using a multi-view tree topology structure; (3) The frequency sensing gating module is combined to realize frequency sensing and self-adaptive reconstruction and calibration of the feature map; (4) The rich features brought by the tree structure are fully utilized by using a knowledge distillation strategy oriented to the tree structure; (5) The multi-view tree topology structure, the frequency sensing gating module, the knowledge distillation module oriented to the tree structure and the like are fused to form a frequency gating tree topology network, so that intelligent diagnosis of the cutting head fault of the heading machine is realized.
2. 2. The method for diagnosing a fault of a cutting head of a heading machine based on a frequency-gated tree topology network according to claim 1, wherein the extracting multi-view key features from signals by using a multi-view tree topology in step (2) comprises the following steps: (2a) Obtaining a base spatial feature representation using a base residual unit For input features Wherein B is the batch size, C is the number of feature channels, H is the feature map height, W is the feature map width, and the two 3X 3 convolutions are applied, and then the two BatchNorm normalization sums and one SiLU activation function are used to obtain : Wherein the method comprises the steps of A batch normalization is shown and is performed, A 3 x3 convolution is indicated and is shown, Representation of Activating the function, then And input features Making residual error, obtaining output of basic residual error unit by activating function again, The obtained basic spatial feature representation is obtained; (2b) Focusing critical bands using coordinate attention For input features Respectively performing one-time horizontal self-adaptive average pooling and longitudinal self-adaptive average pooling to obtain And , Wherein AdaptiveAvgPool d represents adaptive average pooling, (H, 1) represents pooling direction as transverse, (1, W) represents pooling direction as longitudinal, reshape represents adjusting tensor shape, then And (3) with Generating coordinate attention weights along a third dimension by stitching and through a series of operations such as 1 x 1 convolution, batch normalization, HARDSWISH activation, etc And , Wherein the method comprises the steps of The representation HARDSWISH activates the function, concat is the channel splice, split is the inverse of Concat performs the channel Split, A1 x 1 convolution is represented and, Representing SiLU an activation function followed by input of a feature and And Element-by-element multiplication to obtain a coordinate attention output , Wherein, the Element-wise multiplication representing a tape-broadcast mechanism; (2c) Expanding low-level feature receptive fields using large-kernel convolution For input features First apply Activation functions of BN and SiLU to obtain , Immediately following for Applying a single A kind of electronic device The large kernel depth separable convolution realizes single-layer aggregation of wide-area context information on the premise of not remarkably increasing computational complexity, introduces a large receptive field in the early stage of a network, and then applies batch normalization and SiLU activation functions to obtain , And then obtaining through residual error Subsequently apply it maximum pooling, And SiLU activating the function to obtain the output of the large-kernel convolution module , Wherein the method comprises the steps of Is maximum pooling; (2d) Display of decoupled global vibrations and local textures using a frequency domain decomposition module For input features Applying And SiLU activation functions to obtain basic features Immediately after that to Applying an average pooling of 7 x 7 to obtain low frequency characteristics High frequency characteristics From the following components And (3) with The residual error is subtracted to obtain the product, Wherein the method comprises the steps of Representing 7×7 average pooling, based on signal processing theory, large-scale spatial average pooling is represented as a characteristic of low-pass filter in frequency domain, its output can be regarded as low-frequency component reflecting overall trend and slow-varying vibration mode of mechanical vibration signal, the difference between basic characteristic and low-frequency component is high-frequency component, and can capture rapid change information caused by impact, friction or local defect, namely Is the information of the low frequency which is the information of the low frequency, Is high-frequency information, pair And (3) with Applying two different types of Modulating and transforming information between channels, then both and Residual is made to obtain For a pair of Then applying maximum pooling, And SiLU activate functions: Obtaining the output of the frequency domain decomposition module ; (2E) Establishing global dependencies of feature space using global context mixer For input features Two different operations are applied, first to Application of Reshape and Softmax are spatially focused At the same time pair Reshape to obtain , And (3) with Matrix multiplication is performed to obtain And remodel into Subsequent application of Layer normalization, reLU activation function Obtaining , Wherein the method comprises the steps of For the layer normalization, And input features The global view feature X is obtained by broadcast addition, and then a depth downsampling is applied to X, including maximum pooling, The number of components in the package is SiLU, Obtaining the output of the global context mixer 。
3. 3. The fault diagnosis method for the cutting head of the heading machine based on the frequency-gated tree topology network according to claim 1, wherein the step (3) is combined with the frequency-sensing gating module to realize frequency sensing and self-adaptive reconstruction and calibration of the feature map, and the steps are as follows: (3a) Design cross-frequency feature extraction module For input feature graphs Introducing spectrum centroid as physical index to quantify frequency characteristic of each channel, thereby guiding channel rearrangement to obtain , Has been ordered by spectral centroid, will Naturally divided into high frequency components Intermediate frequency component And a low frequency component Three groups of the three-dimensional space-saving optical fiber composite material, The high frequency impact signal generally contains significant transient energy, which is characterized by a rapid and dramatic change in the frequency axis direction, resulting in a vertical spike-like energy distribution, for Using a 5x 3 anisotropic convolution kernel The modeling is performed such that the model is made, The low-frequency signal has slow change and stronger time correlation, and the low-frequency signal generally presents continuous, smooth and longer-span energy distribution along a time axis in a time-frequency diagram, and the low-frequency characteristic presents horizontal distribution, and has the advantages of low-frequency signal with high frequency Using 3 x 5 anisotropic convolution kernels The modeling is performed such that the model is made, For not exhibiting significant directional bias Modeling is performed by adopting 3 x 3 isotropic convolution kernel to realize balanced texture and structure expression, For feature-specific extraction 、 And Channel splicing is carried out, and batch normalization and SiLU are applied to obtain , Wherein the method comprises the steps of Representation SiLU activates the function, then pair Application of 、SiLU、 And SiLU to obtain , And (3) with Obtaining the final output of the cross-frequency feature extraction module by making residual errors ; (3B) Design multi-branch convolution gating mechanism For input feature graphs Grouping to obtain Wherein For each group Is subjected to point-by-point convolution to carry out dimension reduction mapping to reduce the calculated amount, and then is subjected to batch normalization and activation functions to obtain intermediate features To control information flow and capture long-range dependencies, modules are configured to mediate features Three independent 3 x3 convolution branches are applied to generate reset gates (RESET GATE) respectively Update door (Update Gate) Candidate hidden state (CANDIDATE ACTIVATION) By resetting the gate Update door Balancing historical information with current candidate information To generate gating features , Wherein the method comprises the steps of Representing element-wise multiplication, the output features consist of two weighted parts, one part consisting of And Coactivate, another part is reserved by Candidate features for modulation Finally, the polymerized gating feature Generating final output candidate dynamic kernel tensor through feature processing At the same time for inputting feature images Applying adaptive average pooling, Batch normalization, geLU, reshape, and Softmax to get attention weight tensors , Wherein the method comprises the steps of For the obtained And (3) with Fusing the G group candidate kernels into unique dynamic convolution kernels in an element-by-element weighted summation mode , Wherein the method comprises the steps of Will dynamically convolve parameters Acting on The final output of the multi-branch convolution gating mechanism is obtained.
4. 4. The method for diagnosing the fault of the cutting head of the heading machine based on the frequency-gated tree topology network according to claim 1, wherein the knowledge distillation strategy facing the tree structure in the step (4) fully utilizes multi-view rich features brought by the tree structure, and combines self-distillation to improve fault discrimination capability, and the method comprises the following steps: (4a) Multi-view key features for multi-view tree topology input Output from current stage , The method comprises mapping the cross-stage tree input features and stage output features to the same semantic metric space by using a parameter sharing classification mechanism, obtaining semantic soft masks by means of confidence transformation in the space, specifically, in each feature extraction stage of the network, firstly pre-judging the input features of the stage by using a local classifier of shared weight, mapping the input features into semantic soft masks of (0, 1) intervals by confidence transformation of a Sigmoid activation function, Wherein when i is taken to be 1, Namely, is , Is that A special-purpose classifier is arranged on the main body, Is that Activating the function, then, the mask multiplying element-by-element with the predicted distribution of the output features, thereby dynamically modulating the confidence distribution of the output features, Wherein the method comprises the steps of Representing the multiplication by element, Is that The deepest classifier guides all shallow classifier decisions, meanwhile, deep classifiers in adjacent layers guide shallow classifier decisions, and all guiding directions are unidirectional, so that a knowledge distillation strategy oriented to a tree structure is realized.

Description

Development machine cutting head fault diagnosis method based on frequency gating tree topology network Technical Field The invention relates to the technical field of fault diagnosis of industrial equipment, in particular to a fault diagnosis method for a cutting head of a heading machine based on a frequency gate tree topology network, which is particularly suitable for fault diagnosis scenes of the cutting head of the heading machine. According to the invention, dynamic perception of the characteristics is realized through frequency gating modulation, and the problems of difficult extraction of fault characteristics and low fault diagnosis accuracy caused by strong background noise coupling and high non-stationary characteristics of a cutting signal of the heading machine are solved by combining the abundant multi-view key characteristics brought by a tree topology structure and a knowledge distillation strategy oriented to the tree structure. Background Along with the continuous improvement of the intelligent and complicated degree of industrial equipment, the fault diagnosis technology of the cutting head of the heading machine becomes a core means for guaranteeing the safe operation of the equipment, but the traditional method has the remarkable limitations that the fault diagnosis method of the cutting head of the traditional heading machine usually needs to manually perform feature extraction and has the remarkable defects of time consumption and high dependence on expert experience, meanwhile, the traditional convolutional neural network uses static convolutional weights to lack the transient response capability to the content change of input signals and the perceptibility to the change of input frequencies, so that the feature extraction of the cutting signals of the heading machine under the coupling of strong background noise is not facilitated, and in addition, the traditional convolutional neural network adopts a single serial hierarchical paradigm, has limited feature expression capability and restricts the fault diagnosis discrimination capability of the cutting head of the heading machine. Aiming at the problems, the invention provides a fault diagnosis method for a cutting head of a heading machine based on a frequency gating tree topology network, the invention realizes the decomposition of multi-view key features from the display of input signals by constructing a tree topology structure, simultaneously designs a frequency perception gating module to carry out frequency rearrangement and gating modulation on the multi-view key features, realizes the accurate extraction of the features, and finally provides a knowledge distillation strategy oriented to the tree structure, thereby remarkably improving the fault diagnosis capability. Disclosure of Invention Aiming at the problems of difficult extraction of fault characteristics and low fault diagnosis accuracy caused by strong background noise coupling and high non-stable characteristics of a cutting signal of the heading machine, the invention provides a fault diagnosis method for a cutting head of the heading machine based on a frequency-gated tree topology network. The specific implementation steps of the invention are as follows: 1. An EBZ260H type heading machine is taken as a prototype, a fault experiment platform of a cutting head of the heading machine is built in a 1/3 reduced proportion, and a simulated coal wall is built on the platform to collect vibration signals of the heading machine so as to reproduce actual operation working conditions of the heading machine as far as possible. For the acquired vibration signals, the time domain signals are converted into time-frequency diagrams by wavelet transformation. 2. The multi-view tree topology structure is constructed to extract multi-view key features from signals, and the steps are carried out as follows: (2a) Obtaining a base spatial feature representation using a base residual unit For input featuresWherein B is the batch size, C is the number of feature channels, H is the feature map height, W is the feature map width, and the two 3X 3 convolutions are applied, and then the two BatchNorm normalization sums and one SiLU activation function are used to obtain: Wherein the method comprises the steps ofA batch normalization is shown and is performed,A 3 x3 convolution is indicated and is shown,Representation ofActivating the function, thenAnd input featuresMaking residual error, obtaining output of basic residual error unit by activating function again, The obtained basic spatial feature representation is obtained; (2b) Focusing critical bands using coordinate attention For input featuresRespectively performing one-time horizontal self-adaptive average pooling and longitudinal self-adaptive average pooling to obtainAnd, Wherein AdaptiveAvgPool d represents adaptive average pooling, (H, 1) represents pooling direction as transverse, (1, W) represents pooling direction as longitudi