CN-122020120-A - Multi-condition equipment health assessment method based on self-adaptive configuration convolution

CN122020120ACN 122020120 ACN122020120 ACN 122020120ACN-122020120-A

Abstract

The invention relates to the technical field of health evaluation of mechanical equipment, in particular to a multi-working condition equipment health evaluation method based on self-adaptive configuration convolution, which comprises the steps of obtaining multi-working condition vibration signals in the operation of the multi-working condition mechanical equipment, and carrying out normalization processing to obtain an original characteristic tensor; the method comprises the steps of obtaining low-level data representation by reducing the dimension of an original characteristic tensor through a standard convolution module, obtaining high-level data representation by deep self-adaptive feature extraction of the low-level data representation through a self-adaptive configuration convolution network constructed by a self-adaptive configuration convolution module, a nerve convergence module and a nerve divergence module, and outputting an equipment health state assessment result by assessing the high-level data representation through an assessment result output unit comprising a full connection layer. According to the invention, the number of extracted features can be dynamically determined according to the multi-working condition data, so that the spatial position distribution of the extracted features and the sampling window can be adjusted, and the calculated parameter number can be reduced.

Inventors

WANG BIAO
DAI ZHIPENG
Niu Yingcheng
XIE YUHAN
REN XIANGYU

Assignees

北京交通大学

Dates

Publication Date: 20260512
Application Date: 20260129

Claims (10)

1. The multi-condition equipment health assessment method based on the self-adaptive configuration convolution is characterized by comprising the following steps of: Acquiring multi-working condition vibration signals in the operation of multi-working condition mechanical equipment, and carrying out normalization processing to obtain an original characteristic tensor; Performing dimension reduction on the original characteristic tensor by using a standard convolution module to obtain low-level data representation; the self-adaptive configuration convolution network constructed by the self-adaptive configuration convolution module, the nerve convergence module and the nerve divergence module is utilized to carry out deep self-adaptive feature extraction on the low-level data representation, so as to obtain a high-level data representation; And evaluating the high-level data representation by using an evaluation result output unit comprising a full connection layer, and outputting an evaluation result of the health state of the equipment.
2. The multi-working condition equipment health assessment method based on the adaptive configuration convolution according to claim 1, wherein the adaptive configuration convolution network performs depth adaptive feature extraction on the input low-level data representation, and specifically is represented by the following formula: Representing a high-level representation of the data output, Representing the representation of the input low-level data, AFB # ) Representing a self-adaptive configuration convolution module D # ) Representing a nerve divergence module C% ) Representing a nerve convergence module.
3. The multi-condition equipment health assessment method based on adaptive configuration convolution according to claim 2, wherein the adaptive configuration convolution module is represented by the following expression: Wherein the method comprises the steps of In order to modify the linear cell activation function, For the batch of normalized layers, Representing the input characteristics of the adaptive configuration convolution module, Representing the convolution of an adaptive configuration to an input feature And (5) extracting output features.
4. A multi-condition equipment health assessment method based on adaptive configuration convolution according to claim 3, wherein the adaptive configuration convolution is based on input features The output characteristics extracted above are obtained by the following steps: acquiring input features Left sampling feature to the left of the sampling location And a right sampling feature to the right of the sampling location ; Separately computing left-hand sampling features And right side sampling feature Weight of (2): Wherein, the Representing the left-hand linear window interpolation weights, Representing the right-hand linear window interpolation weights, Representing the sampling position; representing the neighborhood anchor to the left of the sampling location, Representing a neighborhood anchor point on the right side of the sampling position; According to left-hand sampling characteristics Right side sampling feature Linear window interpolation of corresponding weight calculation sampling locations : Linear window interpolation from sampling locations Obtaining the self-adaptive configuration convolution on the input characteristic The output characteristics extracted above: Wherein Y represents the output characteristic of the convolution extraction of the adaptive configuration, Representing the selection of convolutions of different sizes according to the characteristics, Representing element-by-element multiplication, z represents computing input features using a deformation convolution tensor sub-network The resulting deformation convolution tensor is then used, Representing computing input features using pivot-core offset sub-networks The obtained pivot-core offset value is used for controlling the pivot-core offset, A linear window interpolation representing each sampling position, Representing a convolution operation with a convolution kernel size of 3, Indicating that the Dropout layer is to be used, Representing a hyperbolic tangent function.
5. The method for multi-condition equipment health assessment based on adaptive configuration convolution of claim 4, wherein the sampling locations are obtained by: 1) Acquiring input features based on offset-aware base network And obtaining a predicted dynamic perception scale l through a dynamic perception scale network according to the obtained offset perception base P: The offset perceptual base P is obtained by the following formula: Wherein the method comprises the steps of Representing the modified linear units of the device, Denoted as a constant set to prevent division by zero, The expression convolution operation is used, Representing input characteristics of the offset-aware base network, W P,1 and b P,1 represent weights and offsets, respectively, of a first convolutional layer in the offset-aware base network, W P,2 and b P,2 represent weights and offsets, respectively, And The first batch of normalized layer offset parameters and scaling parameters are represented respectively, And Representing a second set of normalized layer offset parameters and scaling parameters respectively, And Representing the mean and variance of the normalized layer of the first batch over a batch, And Representing the mean and variance of the second normalized layer over a lot, Activating a function for the modified linear cell; Wherein the method comprises the steps of Representation of The functions, W l,1 and b l,1 , represent the weight matrix and bias, respectively, of the first convolutional layer of the variable receptive field network, W l,2 and b l,2 represent the weight matrix and bias, respectively, of the second convolutional layer of the variable receptive field network, And The first batch of normalized layer offset parameters and scaling parameters are represented respectively, And Representing a second set of normalized layer offset parameters and scaling parameters respectively, And Representing the mean and variance of the normalized layer of the first batch over a batch, And Respectively representing the mean and variance of the second batch of normalized layers on one batch; 2) Generating a time sequence regulation matrix U according to the number S of sampling points of the input features: Where S represents the number of sampling points, and the element r i of the i-th column in U is defined as: ; 3) Constructing a base position grid based on an outer product operation : Where T represents the number of time steps of the sample, The outer product is represented by the sum of the products, Representing a step size; 4) Acquiring sampling positions according to the dynamic sensing scale l, the time regulation matrix U and the basic position grid Wherein the method comprises the steps of Representing the multiplication by element, Represents the average of the dynamic perceptual scale l.
6. The multi-condition equipment health assessment method based on adaptive configuration convolution according to claim 5, wherein the number of sampling points S is obtained by: Where epoch represents the training round, Representing the number of samples S calculated and saved when the training round reached 150, B representing the batch sample size, T representing the number of time steps, Representation of Performing downward rounding operation; The definition is as follows: Wherein the method comprises the steps of A Ai Fosen brackets, 1 if y is an odd number and 0 if y is not an odd number; Representing the maximum value in terms of a given time scale And (5) determining a time sequence adjustment factor.
7. The method for multiple condition equipment health assessment based on adaptive configuration convolution of claim 4, wherein each sampling position has a left neighborhood anchor point And right neighborhood anchor point The method comprises the following steps of: Wherein the method comprises the steps of Representing the left-hand neighborhood anchor point, Representing the right-hand neighborhood anchor point, Representation of The function of the function is that, Representing the sampling position The rounding down operation, T, represents the number of time steps.
8. The multi-condition equipment health assessment method based on adaptive configuration convolution according to claim 2, wherein the nerve convergence module specifically comprises the following expression: Wherein the method comprises the steps of Representing a convolution operation with a convolution kernel of size k, Representing the input characteristics of the nerve convergence module, Representing the training round.
9. The multi-condition equipment health assessment method based on adaptive configuration convolution according to claim 2, wherein the neural divergence module specifically comprises the following expression: Wherein, the And Two input features representing the neural divergence module are presented, Representing a transpose matrix of convolution kernel size k, Representing the training round.
10. The multi-condition equipment health assessment method based on adaptive configuration convolution according to claim 1, wherein the high-level data representation is assessed by using an assessment result output unit comprising a fully-connected layer, and in the step of outputting the equipment health status assessment result, the assessment result output unit outputs the equipment health status assessment result, which is represented by the following expression: Wherein the method comprises the steps of A high-level data representation representing the output of the adaptive configuration convolutional network, Indicating that the full-link layer is to be formed, In order to perform the flattening operation, Representing global average pooling.

Description

Multi-condition equipment health assessment method based on self-adaptive configuration convolution Technical Field The invention relates to the field of mechanical equipment health evaluation, in particular to a multi-working condition equipment health evaluation method based on self-adaptive configuration convolution. Background Mechanical equipment is a core component of a modern industrial system, and each key link of the modern industrial system depends on continuous and stable operation of the mechanical equipment in key fields of railways, aerospace, heavy machinery, precision manufacturing and the like. The operating state of the mechanical equipment and the health level thereof are directly related to production safety, operation efficiency and overall economic benefit. Therefore, it is critical to develop systematic health assessment for mechanical equipment. Based on scientific and accurate health state evaluation, maintenance personnel can reasonably make maintenance plans and accurately grasp maintenance time, so that unplanned shutdown is avoided, unnecessary maintenance expenditure is reduced, and the safety and stability of equipment operation are remarkably improved. The traditional health assessment method is mostly dependent on signal processing and feature extraction technology of a fixed structure, and is poor in performance when processing non-stable and multi-scale characteristics of vibration signals under complex working conditions, so that the identification capability of early health assessment and progressive degradation processes is limited. With the development of deep learning technology, especially the application of convolutional neural network in time sequence signal processing, a new solution path is provided for equipment health evaluation. However, the conventional convolutional neural network structure has the limitations of fixed receptive field, stiff characteristic extraction mode and the like, is difficult to adapt to dynamic changes of signals, and is difficult to adaptively capture characteristics of different scales, so that the characteristic extraction efficiency is low, and the adaptability and generalization capability of the convolutional neural network in actual industrial scenes are restricted. The above problems are well addressed by providing the ability to dynamically adjust the convolution kernel to promote model flexibility. However, in actual operation, the mechanical equipment often faces a complex environment with multiple working conditions such as rotating speed, load and the like, so that vibration signals of the mechanical equipment show strong non-stationarity and multi-scale characteristics. This characteristic makes existing methods often require learning a large number of parameters. Meanwhile, the existing method cannot realize collaborative self-adaptive adjustment of the number and the positions of sampling points, so that the method is difficult to flexibly adapt to huge scale differences of vibration characteristics under different working conditions and different degradation stages while accurately capturing key fault impact components according to time-varying characteristics of vibration signals. This structural limitation further exacerbates the computational burden, resulting in a sharp rise in overhead with an increasing number of sampling points. Therefore, there is a need for a mechanical equipment health evaluation method that can dynamically optimize the feature extraction process under multiple working conditions and simultaneously combine the calculation efficiency and the model accuracy, so as to reduce the calculation overhead generated by increasing the number of sampling points, and meet the dual demands of efficiency and performance in practical engineering. Disclosure of Invention In view of the above, the invention provides a multi-working condition equipment health evaluation method based on adaptive configuration convolution, so as to solve part of technical problems in the background art. In order to achieve the above purpose, the present invention adopts the following technical scheme: A multi-working condition equipment health assessment method based on self-adaptive configuration convolution is designed, and comprises the following steps: Acquiring multi-working condition vibration signals in the operation of multi-working condition mechanical equipment, and carrying out normalization processing to obtain an original characteristic tensor; Performing dimension reduction on the original characteristic tensor by using a standard convolution module to obtain low-level data representation; the self-adaptive configuration convolution network constructed by the self-adaptive configuration convolution module, the nerve convergence module and the nerve divergence module is utilized to carry out deep self-adaptive feature extraction on the low-level data representation, so as to obtain a high-level data representation;