CN-121980424-A - Vertical shaft cage guide abnormality detection method based on multi-mode sound vibration fusion and deep learning

CN121980424ACN 121980424 ACN121980424 ACN 121980424ACN-121980424-A

Abstract

The invention discloses a vertical shaft cage guide abnormality detection method based on multi-mode sound vibration fusion and deep learning, which comprises the following steps of firstly, synchronously acquiring and preprocessing multi-mode data; the method comprises a first step of multi-modal feature extraction, a second step of cross-modal attention and depth feature fusion based on a graph neural network, a third step of time sequence modeling and anomaly detection, a fifth step of model training and optimization, and a sixth step of online monitoring and hierarchical early warning. The invention comprehensively utilizes the sensitivity of the vibration signal to the low-frequency fault of the structure and the sensitivity of the sound signal to the high-frequency transient event, thereby realizing the monitoring of the health state of the cage guide more comprehensively without dead angles.

Inventors

XIE HAIFENG
PEI WENLIANG
WEI SHUO
CUI GUANGTAO
ZHANG XUHUA
GUO YONGTAO
YIN LIPENG

Assignees

中信重工开诚智能装备有限公司

Dates

Publication Date: 20260505
Application Date: 20260211

Claims (8)

1. A vertical shaft cage guide abnormality detection method based on multi-mode sound vibration fusion and deep learning is characterized by comprising the following steps: Step one, synchronously collecting and preprocessing multi-mode data; The vibration acceleration sensor is arranged at the rigid connection position of the guide frame of the cage, the sound pressure sensor is arranged in the cage or at the top of the cage, and the vibration signals in the cage guide operation process are synchronously collected by using the multichannel synchronous data collection card with the same time standard And sound signals Time alignment is carried out on the collected signals, and the collected signals are divided into analysis sample sections with fixed duration; step two, multi-mode feature extraction, which comprises the following steps: Extracting wavelet packet energy entropy characteristics of vibration signals to form vibration characteristic vectors; Extracting the time-frequency domain manual characteristics of the sound signals and the deep acoustic characteristics of a convolution neural network based on a Mel spectrogram to form sound characteristic vectors; Step three, fusion of depth characteristics of the cross-modal attention and the graph neural network is based; Designing a fusion module comprising a bidirectional cross attention mechanism and a graph convolution network, and carrying out depth fusion on the vibration feature vector and the sound feature vector to obtain a final depth fusion feature vector ; Step four, time sequence modeling and anomaly detection; inputting the fusion characteristic vectors with continuous time sequences into a gating circulation unit network for modeling, and outputting an abnormal probability representing cage guide state abnormality; Step five, model training and optimization; Training a model by using historical normal and abnormal data, and optimizing by adopting a Focal Loss function Focal Loss; step six, online monitoring and grading early warning; the trained model is deployed to mine edge computing equipment or a cloud server, collected sound vibration data are processed in real time, and abnormal probability is calculated Setting up multi-stage early warning threshold value when Triggering primary early warning to prompt attention and observation when a plurality of windows are continuous When the device is in use, a medium-level alarm is triggered to prompt maintenance; when (when) When the rapid ascent continues and exceeds 0.95, an emergency alert is triggered suggesting an immediate shutdown check. .
2. The method for detecting abnormal conditions of a vertical shaft cage guide based on multi-modal sound-vibration fusion and deep learning of claim 1, wherein in the first step, the method comprises the following steps of Also included in the pre-treatment of (2): for sound signals Pre-emphasis processing is performed to raise the high frequency components: Wherein ; Reusing background noise collected during static period of elevator Performing spectral subtraction and noise reduction processing on the sound signal in the operation period to inhibit underground environment noise interference: Wherein, the method comprises the steps of, And Short-time magnitude spectra of the run signal and the noise signal respectively, , For the number of frequency segments of the sound signal, In order to over-subtract the factor, Is a spectral lower-limit parameter.
3. The method for detecting the abnormal condition of the vertical shaft cage guide based on the multi-mode sound vibration fusion and the deep learning according to claim 1 is characterized in that the vibration characteristic extraction in the second step is specifically as follows: For vibration signals 4-Layer wavelet packet decomposition is carried out to obtain 16 sub-band signals Calculating the energy entropy of each sub-band as a feature: In the formula (I), in the formula (II), Represent the first The energy entropy of the sub-bands, Represent the first The first of the sub-bands The energy of the individual components is such that, Represent the first The first of the sub-bands Sub-band signal expressions of the individual components; The energy entropy calculation results of the 16 sub-band signals form vibration characteristic vectors: 。
4. The method for detecting the abnormal condition of the vertical shaft cage guide based on the multi-mode sound vibration fusion and the deep learning according to claim 1 is characterized in that the sound feature extraction in the second step comprises the following steps: Step 2.1, calculating time-frequency domain manual characteristics of the sound signal after spectral subtraction and noise reduction, namely a short-time zero-crossing rate ZCR, short-time energy, a spectrum centroid SC, a spectrum roll-off point SR and a subband energy ratio SER aiming at cage guide fault characteristics, wherein the characteristics jointly form a manual characteristic vector ; ; Step 2.2, deep Acoustic feature of converting the Sound Signal into a Mel-Chart matrix Inputting a light multi-branch convolution neural network CNN, wherein the network comprises two branches which pay attention to low-frequency periodic texture and high-frequency random impact respectively, and extracting high-level abstract acoustic characteristics The formula is as follows: In the formula (I), in the formula (II), In order to activate the function, The operation of the splice is indicated and, A convolution calculation representing two branches; abstracting high-level acoustic features And manual feature vector Splicing and fusing, and then reducing dimension through a full connection layer to obtain a depth fused sound feature vector ; In the formula (I), in the formula (II), Representing a full connection layer computation function.
5. The method for detecting the abnormal condition of the vertical shaft cage guide based on the multi-mode sound vibration fusion and the deep learning according to claim 4 is characterized in that the time-frequency domain characteristic calculation in the step 2.1 is specifically as follows: Assume that Is expressed as a time domain discrete of one frame length Discrete sound signals of (a) , The frequency domain expression is , , The calculation formula of the time-frequency domain characteristics is as follows: short time zero crossing rate: In the formula (I), in the formula (II), Is a sign function, and the expression is: ; Short time energy: In the formula (I), in the formula (II), As a window function, the invention adopts a Hamming window, and n represents an nth Hamming window; Spectral centroid: ; The spectral roll-off point represents the frequency corresponding to the cumulative spectral energy reaching a certain proportion of the total energy, and is used for describing the tail characteristic of the spectrum, and is the minimum index meeting the following conditions : In the formula (I), in the formula (II), For roll-off ratio, the frequency corresponding to the spectral roll-off point is: In the formula (I), in the formula (II), Is the sampling frequency; The frequency spectrum is divided into a plurality of sub-bands by the sub-band energy ratio, the proportion of each sub-band energy to the total energy is calculated and used for describing the frequency domain distribution of the energy, and the formula is as follows: In which, in the process, For the number of spectral sub-bands of the sound signal, Is the first Energy of the sub-band.
6. The method for detecting the abnormal condition of the vertical shaft cage guide based on the multi-mode sound vibration fusion and the deep learning according to claim 5 is characterized by comprising the following steps: step 3.1, enhancing the cross attention of double channels, namely respectively calculating the attention weight of the vibration characteristic to the sound characteristic Attention weighting of sound features to vibration features ; ; Wherein, the method comprises the steps of, Generating enhanced features for the learnable weight matrix: , ; step 3.2, graph convolution fusion, namely constructing the bimodal characteristic of each time sample into a graph Wherein the node Respectively represent And Initializing the edge weight through a leachable correlation matrix, and performing graph convolution operation to aggregate cross-modal information: Wherein, the method comprises the steps of, And Respectively the first Layer and the first The graph of layers is rolled up into a matrix of nodes, In order to have a contiguous matrix of self-connections, For the matrix of degrees thereof, Is the first The layer may be a layer of trainable weights, To activate the function, after two-layer graph convolution, the node characteristics are output Will be Proceeding with Splicing to obtain final depth fusion feature vector 。
7. The method for detecting the abnormal condition of the vertical shaft cage guide based on the multi-mode sound vibration fusion and the deep learning of claim 6, wherein the fourth step is that a depth fusion characteristic sequence obtained by continuous time window is adopted Inputting the fault time sequence evolution mode into a gate control circulation unit GRU network, wherein a specific calculation formula is as follows: First calculate candidate hidden states : In the formula (I), in the formula (II), In order to activate the function, In order to conceal the layer weights, Indicating that the gate is to be reset, Indicating the final hidden state of the previous layer, Representing the multiplication of the corresponding elements, Further calculating the final hidden state : In which, in the process, Representing update gate, and finally outputting abnormal probability of current sequence by hiding state through full connection layer and Sigmoid activation function : In which, in the process, Representing the Sigmoid activation function, Representing a weight matrix of the output layer, Representing the output layer bias matrix, Indicating the final hidden state.
8. The method for detecting the abnormal condition of the vertical shaft cage guide based on the multi-mode sound vibration fusion and the deep learning according to claim 7 is characterized in that the optimization of the Focal Loss function Focal Loss in the fifth step is as follows: ; Wherein the method comprises the steps of Representing the focal point loss function, Representing a real label of the tag, And Respectively used for controlling the proportion of positive/negative samples, the value range is , Called focusing parameter, the range of values is When (1) When, the Focal Loss is degraded to cross entropy Loss, The larger the penalty is for the easily classified samples.

Description

Vertical shaft cage guide abnormality detection method based on multi-mode sound vibration fusion and deep learning Technical Field The invention relates to the technical field of mine safety monitoring and intelligent diagnosis, in particular to a vertical shaft guide abnormality detection method based on multi-mode sound vibration fusion and deep learning. Background The traditional detection method mainly relies on periodic manual inspection or monitoring by a single sensor such as a vibration sensor, and has the problems of low efficiency, strong subjectivity, incapability of real-time early warning, insensitivity to complex or early weak faults and the like. The vibration signal can effectively reflect the low-frequency integral vibration and deformation modes of the cage guide structure, but has limited capturing capability on high-frequency transient impacts such as micro cracks and particle collisions, the sound signal is rich in broadband information, especially good at capturing transient events such as high-frequency friction and impact generated at the contact surface of the cage guide and the roller, but has weaker characterization on low-frequency resonance of the structure, the traditional deep learning-based method mostly adopts single mode data or simple multi-mode characteristic splicing, and the complementary advantages of the sound vibration signal cannot be deeply fused from a physical mechanism and a data layer, so that the diagnosis precision and the robustness on early and composite faults in a complex running environment are insufficient. Disclosure of Invention The invention aims to overcome the defects of the prior art, and provides a vertical shaft guide abnormality detection method based on multi-mode sound vibration fusion and deep learning, which realizes synchronous and accurate collection and deep fusion analysis of sound vibration signals and can adaptively learn a fault time sequence evolution rule so as to realize early, accurate and interpretable early warning of guide abnormality. The technical scheme adopted by the invention is that the vertical shaft cage guide abnormality detection method based on multi-mode sound vibration fusion and deep learning comprises the following steps: Step one, synchronously collecting and preprocessing multi-mode data; The vibration acceleration sensor is arranged at the rigid connection position of the guide frame of the cage, the sound pressure sensor is arranged in the cage or at the top of the cage, and the vibration signals in the cage guide operation process are synchronously collected by using the multichannel synchronous data collection card with the same time standard And sound signalsTime alignment is carried out on the collected signals, and the collected signals are divided into analysis sample sections with fixed duration; step two, multi-mode feature extraction, which comprises the following steps: Extracting wavelet packet energy entropy characteristics of vibration signals to form vibration characteristic vectors; Extracting the time-frequency domain manual characteristics of the sound signals and the deep acoustic characteristics of a convolution neural network based on a Mel spectrogram to form sound characteristic vectors; Step three, fusion of depth characteristics of the cross-modal attention and the graph neural network is based; Designing a fusion module comprising a bidirectional cross attention mechanism and a graph convolution network, and carrying out depth fusion on the vibration feature vector and the sound feature vector to obtain a final depth fusion feature vector ; Step four, time sequence modeling and anomaly detection; inputting the fusion characteristic vectors with continuous time sequences into a gating circulation unit network for modeling, and outputting an abnormal probability representing cage guide state abnormality; Step five, model training and optimization; Training a model by using historical normal and abnormal data, and optimizing by adopting a Focal Loss function Focal Loss; step six, online monitoring and grading early warning; the trained model is deployed to mine edge computing equipment or a cloud server, collected sound vibration data are processed in real time, and abnormal probability is calculated Setting up multi-stage early warning threshold value whenTriggering primary early warning to prompt attention and observation when a plurality of windows are continuousWhen the device is in use, a medium-level alarm is triggered to prompt maintenance; when (when)When the rapid ascent continues and exceeds 0.95, an emergency alert is triggered suggesting an immediate shutdown check. Specifically, in the first step, the voice signal is transmittedAlso included in the pre-treatment of (2): for sound signals Pre-emphasis processing is performed to raise the high frequency components: Wherein ; Reusing background noise collected during static period of elevatorPerforming spectral subtraction a