Rigid cage guide fault diagnosis method based on two-dimensional elastic space learning

Abstract

The invention discloses a rigid cage guide fault diagnosis method based on two-dimensional elastic space learning, which mainly constructs a two-dimensional elastic space learning model, so that potential global and local features of fault data can be learned, and the accuracy of rigid cage guide fault diagnosis is effectively improved. The method specifically comprises the following steps: (1) constructing a two-dimensional elastic space learning model by means of global and local manifold structures among fault sample data; (2) an analytic solution of a two-dimensional elastic space bidirectional projection matrix in the model is optimized and solved; and (3) obtaining low-dimensional spatial features of the fault sample data on the basis of two projection directions of the two-dimensional elastic space, and obtaining a classification result of fault diagnosis by using a classifier. Experimental results show that the fault diagnosis method is more effective and robust.

Description

Technical Field

[0001] The present invention relates to technical fields such as spatial learning and fault diagnosis. Specifically, it is a fault diagnosis method for rigid cage guides based on two-dimensional elastic spatial learning, which can be applied to the field of safety monitoring and fault diagnosis of mine hoisting systems. Background Art

[0002] With the acceleration of the industrialization process, modern industrial systems are becoming increasingly complex, and efficient fault diagnosis is particularly important. In a mine hoisting system, a rigid cage guide, as a guiding and supporting device for the hoisting container, is prone to defects such as wear, deformation, and fracture due to long-term high-load operation, seriously threatening the safety production of the mine. The detection methods of rigid cage guides are divided into static detection and dynamic detection. Static detection mainly relies on manual visual maintenance. The measurement process is cumbersome, the accuracy is low, and shutdown is required, reducing production efficiency. Dynamic detection relies on the dynamic output response of the system to analyze the cage guide and can more truly reflect the operating state of the system. Fault diagnosis based on vibration signals is an important dynamic detection method. However, traditional methods mainly rely on one-dimensional vibration signals and ignore the rich information potential in multi-dimensional data. By converting one-dimensional vibration signals into two-dimensional forms such as time-frequency diagrams, more abundant signal characteristics can be presented in the time-frequency domain. However, the data feature dimensions obtained from two-dimensional images are huge, and direct fault diagnosis will lead to problems such as too long calculation time and low recognition accuracy. And traditional spatial learning methods need to be vectorized before dimensionality reduction, damaging the inherent position information in the image. To address these problems, the present invention proposes a fault diagnosis method based on two-dimensional elastic spatial learning. This method constructs a global and local geometric structure preservation mechanism to ensure that both the overall data distribution relationship and local neighborhood features are retained during dimensionality reduction, and a bidirectional projection matrix is used to directly reduce the dimensionality of time-frequency images, avoiding the loss of spatial structure information caused by traditional vectorization methods and improving the accuracy and robustness of fault recognition. Summary of the Invention

[0003] In order to better discover the local and global manifold structures of data, the present invention constructs a two-dimensional elastic spatial learning model based on the spatial learning theory and theoretically derives the solution of this model, thus solving the problem that vectorized data damages the inherent position information in the image. The specific implementation steps of the present invention are as follows:

[0004] 1. Install sensors on the hoist to collect data signals. Use cage guides with different defects to collect different types of fault signals. Set the collected vibration acceleration signals as H = {x1, x2,..., x n}, and each signal x i represents the time-domain characteristics in a certain state. Use continuous wavelet transform for each vibration signal x i to generate two-dimensional time-frequency images Form the original feature set X = {X1, X2,..., X n};

[0005] 2. Construct a two-dimensional elastic space learning model through local and global structure preservation strategies. The specific steps are as follows:

[0006] (2a) Construct a local similarity matrix based on the k-nearest neighbor relationship:

[0007]

[0008] Where: N k (X i ) represents the set of k nearest neighbors of sample X i . σ is the heat kernel parameter; this definition adopts the form of the heat kernel function to ensure the non-negativity and smoothness of the similarity measure. The k-nearest neighbor constraint not only guarantees the locality of the structural representation but also improves the computational efficiency;

[0009] (2b) Construct a local structure preservation objective function:

[0010]

[0011] Where: and are the left projection matrix and the right projection matrix respectively. The physical meaning of this objective function is to ensure that sample pairs with strong local similarity in the original space maintain a relatively close distance in the reduced-dimensional space, thereby effectively retaining the local structural features of the data;

[0012] (2c) Design a distance-weighted global similarity matrix:

[0013]

[0014] This design has a dual purpose: the distance weight term explicitly emphasizes the importance of distant sample pairs and prevents over-localization during the dimensionality reduction process; while the kernel function term provides smooth weight modulation to avoid the over-strong influence of distant sample pairs;

[0015] (2d) Construct a global structure preservation objective function:

[0016]

[0017] By maximizing this objective function, it is ensured that samples with a large distance in the original space maintain a significant distance in the reduced-dimensional space, effectively avoiding the common "crowding problem" in dimensionality reduction;

[0018] (2e) Construct a unified optimization objective function:

[0019]

[0020] Among them, α is a parameter that balances the contributions of local and global structure preservation. This optimized objective function consists of two main parts: the local structure preservation term ensures the close distribution of similar samples in the reduced-dimensional space, reducing the intra-class structure distortion caused by dimensionality reduction; the global structure preservation term optimizes the overall data projection, reasonably enlarging the inter-class distance, and improving the separability of different fault modes.

[0021] 3. Optimize and solve the two-dimensional elastic space projection matrix to obtain the final low-dimensional representation of the data The specific steps are as follows:

[0022] For the projection matrix G L and G R updates, the orthogonal constraint condition needs to be considered, and it can be transformed into a generalized eigenvalue problem for solution:

[0023]

[0024] Among them, L local = D local - W local is the local Laplacian matrix, and the corresponding degree matrix L global = D global - W global is the global Laplacian matrix, and the corresponding degree matrix I d1 and I d2 are the identity matrices of d1×d1 and d2×d2 respectively, and they correspond to the orthogonal constraints of the left and right projection matrices G L and G R Through this alternating optimization process, the projection matrices G L and G R will gradually converge to the local optimal solution.

[0025] 4. Use spatial projection to obtain the low-dimensional features of the fault data, and use a classifier to classify the fault features, so as to obtain the fault diagnosis result.

[0026] The method of the present invention has the following advantages:

[0027] (1) The present invention constructs a local and global structure preservation mechanism. The local structure preservation can ensure the close distribution of similar samples in the reduced-dimensional space, reduce the intra-class structure distortion caused by dimensionality reduction, and the global structure preservation can optimize the overall data projection, reasonably enlarge the inter-class distance, and improve the separability of different fault modes.

[0028] (2) The present invention directly performs dimensionality reduction in a two-dimensional matrix space, avoiding the loss of structural information caused by traditional vectorization methods, and can obtain a more comprehensive data flow type structure, effectively improving the accuracy of fault diagnosis;

[0029] (3) Through theoretical derivation, the present invention obtains an analytical solution of the two-dimensional elastic space learning model, which can quickly obtain the low-dimensional fault features of fault test samples, thereby realizing a more accurate classification of fault types. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 is the overall framework diagram of the present invention

[0031] Figure 2 is the bar chart of the accuracy rate with the increase of the number of training samples on the rigid guide rail dataset

[0032] Figure 3 is the classification effect diagram of the confusion matrix on the rigid guide rail dataset DETAILED DESCRIPTION OF THE INVENTION

[0033] The specific implementation steps of the present invention are as follows:

[0034] 1. Set the collected vibration acceleration signal as H = {x1, x2,..., x n}, and each signal x i represents the time-domain feature under a certain state. The vibration signal is obtained by an acceleration sensor installed on the lifting container, which can directly reflect the dynamic response characteristics of the guide rail during operation. For each vibration signal x i use continuous wavelet transform, adopt the Cmor3.3 wavelet basis function, and generate a two-dimensional time-frequency image to form the original feature set X = {X1, X2,..., X n}.

[0035] 2. Calculate the distance between samples using the Frobenius norm. The distance between samples X i and X j is calculated as: Construct a local similarity matrix based on the k-nearest neighbor relationship At the same time, design a globally weighted distance similarity matrix Construct an objective function for maintaining local structure and an objective function for maintaining global structure where and are the left projection matrix and the right projection matrix respectively. By minimizing the local structure objective function and maximizing the global structure objective function, it is ensured that the structural features in the original space are retained in the dimensionality reduction space.

[0036] 3. Construct an integrated structure maintenance and feature-guided optimization framework, and the unified optimization objective function is:

[0037]

[0038] Optimize the parameters through an alternating iteration strategy: fix G L Optimize G when R Fix G R Update G when L . After the iteration converges, calculate the final low-dimensional representation Use a support vector machine for the final fault classification.

[0039] The effect of the present invention is further verified by the following experiments:

[0040] The experiment was verified on the rigid guide rope dataset. The vibration signals of the rigid guide rope were collected by triaxial acceleration sensors, and the sampling frequency was 1024 Hz. The rigid guide rope dataset contains three states: joint gap, dislocation, and no fault. 400 samples were collected for each state, and a total of 1200 samples were used for the experiment. Then, the dataset was randomly divided into a training set and a test set according to a certain proportion, and continuous wavelet transform was performed using the cmor3.3 wavelet basis function to generate time-frequency images, which were input into the proposed method for training. The random experiment was repeated ten times. The average accuracy of the method of the present invention in ten random experiments on the rigid guide rope dataset reached 0.975, indicating that the method of the present invention is an effective method for diagnosing rigid guide rope faults.

Claims

1. The present invention discloses a rigid guideway fault diagnosis method based on two-dimensional elastic space learning, comprising the following steps: (1) Install sensors on the hoist to collect data signals. Using shaft guides with different defects, collect different types of fault signals. Set the collected vibration acceleration signals as H = {x1, x2,..., x n}, where each signal x i represents the time-domain characteristics in a certain state. Apply continuous wavelet transform to each vibration signal x i to generate a two-dimensional time-frequency image and form the original feature set X = {X1, X2,..., X n}; (2) Construct a two-dimensional elastic space learning model through local and global structure preservation strategies; (3) Optimize and solve the two-dimensional elastic space projection matrix to obtain the final low-dimensional representation of the data (4) Obtain the low-dimensional features of the fault data by spatial projection, and classify the fault features using a classifier to obtain the fault diagnosis result.

2. The rigid cage guide fault diagnosis method based on two-dimensional elastic space learning according to claim 1, characterized in that The step of constructing a two-dimensional elastic space learning model through local and global structure preservation strategies in step (2) comprises the following steps: (2a) Construct a local similarity matrix based on the k-nearest neighbor relationship; Where: N k (X i ) represents the k-nearest neighbor set of the sample X i , σ is the heat kernel parameter; this definition adopts the form of the heat kernel function to ensure the non-negativity and smoothness of the similarity measure. The k-nearest neighbor constraint not only guarantees the locality of the structural representation but also improves the computational efficiency; (2b) Construct a local structure preservation objective function; Wherein: and are the left projection matrix and the right projection matrix respectively. The physical meaning of this objective function is to ensure that sample pairs with strong local similarity in the original space maintain a relatively close distance in the dimensionality-reduced space, thereby effectively retaining the local structural features of the data; (2c) Design a distance-weighted global similarity matrix; This design has a dual purpose: the distance weight term explicitly emphasizes the importance of distant sample pairs and prevents excessive localization during the dimensionality reduction process; while the kernel function term provides smooth weight modulation to avoid an overly strong impact from distant sample pairs; (2d) Construct a global structure preservation objective function; By maximizing this objective function, it is ensured that samples with a large distance in the original space maintain a significant distance in the reduced-dimensional space, effectively avoiding the common "crowding problem" in dimensionality reduction; (2e) Construct a unified optimization objective function; where α is a parameter that balances the contributions of local and global structure preservation. This optimization objective function consists of two main parts: the local structure preservation term ensures the close distribution of similar samples in the reduced-dimensional space, reducing the intra-class structure distortion caused by dimensionality reduction, and the global structure preservation term optimizes the overall data projection, reasonably expanding the inter-class distance and improving the separability of different fault modes.

3. The rigid guideway fault diagnosis method based on two-dimensional elastic space learning according to claim 1, characterized in that Optimizing and solving the two-dimensional elastic space projection matrix in step (3) to obtain the final low-dimensional representation of the data including the following steps: For the projection matrix G L and G R For the update of, the orthogonal constraint condition needs to be considered and it can be transformed into a generalized eigenvalue problem for solution: where L local = D local - W local is the local Laplacian matrix, and the corresponding degree matrix L global = D global - W global is the global Laplacian matrix, and the corresponding degree matrix I d1 and I d2 are the identity matrices of d1×d1 and d2×d2 respectively, and they correspond to the orthogonal constraints of the left and right projection matrices G L and G R Through this alternating optimization process, the projection matrices G L and G R will gradually converge to the local optimal solution.

Images