Rotating machine fault diagnosis method and system based on multi-signal fusion image construction

Abstract

The invention discloses a rotating machine fault diagnosis method and system based on multi-signal fusion image construction, and relates to the technical field of fault diagnosis, and the method comprises the steps: collecting a rotating machine signal, and carrying out the signal preprocessing through sliding window segmentation; a minimum mutual information method is adopted to determine time delay, a Cao nearest neighbor method is adopted to calculate similarity measurement of adjacent points and determine an embedding dimension, and phase space reconstruction is performed based on the time delay and the embedding dimension. According to the method, displacement, speed and acceleration signals are collected, the minimum mutual information method is adopted to determine the optimal delay time, the Cao nearest neighbor method is adopted to determine the optimal embedding dimension for phase space reconstruction, the dynamic evolution process of the signals is reflected, the dynamic evolution process comprises nonlinear changes and periodic behaviors, and the phase space is reconstructed. Long-term dependence and short-term fluctuation of signals are comprehensively described, local significant change points in a time-frequency graph and a recursive binary graph are detected through an SURF algorithm, and local frequency change and transient characteristics of the signals can be captured at the same time.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and in particular to a rotating machinery fault diagnosis method and system based on multi-signal fusion image construction. Background Art

[0002] Rotating machinery is an important equipment widely used in the industrial field. With the improvement of industrial automation, the reliability and stability of rotating machinery are crucial to the continuity and safety of production systems. Common rotating machinery is usually in a high-load and complex operating environment. Its mechanical parts (such as bearings, gears, etc.) are prone to wear, fatigue or failure. The resulting failure will cause the rotating machinery and related machinery to cause related failures and lead to losses. Therefore, fault diagnosis of rotating machinery is very important.

[0003] Traditional fault diagnosis technology mainly relies on the analysis method of a single signal such as a vibration signal, and processes the signal frequency and amplitude changes through Fourier transform, short-time Fourier transform or wavelet analysis. However, when processing nonlinear and non-stationary signals, it is difficult to capture the complex characteristics of the early stages of equipment failure. At the same time, with the increase in the operating speed of modern equipment and the increase in signal noise, the fault diagnosis method relying on a single signal has gradually exposed its limitations and cannot fully and accurately reflect the operating status of the machinery. In addition, the existing imaging diagnosis methods are mostly based on simple grayscale image processing, which fails to fuse multiple signals into complete image information, making it inconvenient to fully explore fault characteristics through deep learning networks, thereby reducing the efficiency of rotating machinery fault diagnosis and identification. Summary of the invention

[0004] In view of the problems existing in the above-mentioned existing rotating machinery fault diagnosis method and system based on multi-signal fusion image construction, the present invention is proposed.

[0005] Therefore, the problem to be solved by the present invention is that it is difficult to capture the complex characteristics of the early stages of equipment failure when processing nonlinear and non-stationary signals. At the same time, with the increase in the operating speed of modern equipment and the increase in signal noise, the fault diagnosis method relying on a single signal has gradually exposed its limitations and cannot fully and accurately reflect the operating status of the machinery. In addition, the existing imaging diagnosis methods are mostly based on simple grayscale image processing, which fails to fuse multiple signals into complete image information, making it inconvenient to fully explore fault characteristics through deep learning networks, thereby reducing the efficiency of rotating machinery fault diagnosis and identification.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions: a rotating machinery fault diagnosis method and system based on multi-signal fusion images, which includes:

[0007] Collect rotating machinery signals and use sliding window segmentation for signal preprocessing;

[0008] The minimum mutual information method is used to determine the time delay, the Cao's nearest neighbor method is used to calculate the similarity measure of adjacent points and determine the embedding dimension, and the phase space is reconstructed based on the time delay and the embedding dimension;

[0009] The Heaviside function is used to generate a recursive binary map, the wavelet transform is used to extract time-frequency features and generate a time-frequency map, the SURF algorithm is used to detect local significant change points in the time-frequency map and the recursive binary map, and the t-SNE algorithm is used to generate a low-dimensional feature matrix;

[0010] Calculate the cosine similarity of each pair of feature vectors, perform grayscale value mapping to output a grayscale image, and output an RGB image based on the displacement, velocity, and acceleration signals;

[0011] The CNN convolutional neural network is used to predict the fault category probability of the input RGB image, and the threshold judgment is made based on the prediction record.

[0012] As a preferred solution of the rotating machinery fault diagnosis method based on multi-signal fusion image construction described in the present invention, wherein: the rotating machinery signal is collected and the sliding window segmentation is used for signal preprocessing, including:

[0013] The displacement, velocity and acceleration of rotating machinery are collected in real time through sensors;

[0014] Sliding window segmentation is used for signal preprocessing. The sliding window length is determined based on the motion periodicity of the rotating machinery and the sampling rate of the collected signal, which is expressed as:

[0015] T=f s ·n s ·T r ;

[0016] Where T represents the sliding window length, f s represents the sampling frequency, n s Indicates the number of coverage cycles, T r represents the mechanical rotation period;

[0017] The sliding step length is determined based on the overlapping degree of each window sliding after the sliding window length. The signal data is read, and the collected displacement, velocity and acceleration data are respectively organized into arrays according to the time index, and the sliding window is applied for segmentation, which is expressed as:

[0018] S i (t k )={x i (t k ),x i (t k +1),…,x i (t k +T-1)};

[0019] Where S i (t k ) represents the signal segment of the i-th type signal in the k-th window, t k represents the kth time t, T represents the sliding window length, x i Represents the collected value of the i-th type signal.

[0020] As a preferred solution of the rotating machinery fault diagnosis method based on multi-signal fusion image construction described in the present invention, wherein: the minimum mutual information method is used to determine the time delay, the Cao's nearest neighbor method is used to calculate the similarity measure of adjacent points and determine the embedding dimension, and the phase space is reconstructed based on the time delay and the embedding dimension, including:

[0021] The minimum mutual information method is used to discretize the continuous time signal segment S(t), calculate the probability distribution of each discrete interval in the signal S(t), and calculate the entropy value of the segmented signal segment. The entropy value of the delayed signal is calculated based on the historical data traversal time delay τ.

[0022] According to the signal segment S(t) and the delayed signal S τ (t) Calculate the joint entropy and calculate the signal segment S(t) and the delayed signal S τ The mutual information of the entropy value and the joint entropy is obtained by determining the signal segment S(t) and the delayed signal S τ (t) and determine the optimal delay time τ;

[0023] Initialize the embedding dimension based on historical data and preliminarily generate the phase space vector, and use Cao's nearest neighbor method to calculate the similarity measure of adjacent points;

[0024] Gradually increase the value of the embedding dimension and recalculate the adjacent point similarity measure until the change in the calculated adjacent point similarity measure is no longer obvious, and extract the embedding dimension m at this time;

[0025] Combining the minimum delay time and the extracted embedding dimension, the final phase space vector is calculated for each time signal segment S(t), expressed as:

[0026] Y(i)={S(i),S(i+τ),S(i+2τ),…S(i+(m-1)τ)};

[0027] Where Y(i) represents the generated phase space vector, which represents the representation of the signal at time i in the embedded high-dimensional space, S(i) represents the value of the signal segment S(t) at time i, τ represents the delay time, m represents the number of embedding dimensions, S(i+τ) represents the value of the signal segment S(t) at time i+τ, and S(i+(m-1)τ) represents the value of the signal segment S(t) starting at time i after m-1 delays τ.

[0028] As a preferred solution of the rotating machinery fault diagnosis method based on multi-signal fusion image construction described in the present invention, wherein: the Heaviside function is used to generate a recursive binary image, the wavelet transform is used to extract time-frequency features and generate a time-frequency image, the SURF algorithm is used to detect local significant change points in the time-frequency image and the recursive binary image, and the t-SNE algorithm is used to generate a low-dimensional feature matrix, including:

[0029] Based on the final phase space vector, the Euclidean distance between the state representations Y(i) and Y(j) at time i and time j is calculated to form a distance matrix d(i,j), and the distance threshold r is determined based on the average Euclidean distance;

[0030] The elements of the recursive binary graph matrix are calculated using the Heaviside function Θ, expressed as:

[0031] R(i,j)=θ(rd(i,j));

[0032] Where R(i,j) represents the element of the recursive binary graph matrix, Θ represents the Heaviside step function, and r represents the distance threshold;

[0033] Visualize the recursive binary matrix and generate a recursive binary graph;

[0034] Wavelet transform is used to extract time-frequency features of all collected signals, and the time-frequency diagram of each signal is constructed;

[0035] Use the SURF algorithm to detect local significant change points in the time-frequency graph and recursive binary graph, and calculate the Hessian matrix and determinant of the image;

[0036] Based on the determinant det(H), the extreme points are determined as the key points of the SURF algorithm. For each key point, the SURF algorithm generates a fixed-length SURF descriptor as a feature vector and forms a SURF feature matrix.

[0037] The t-SNE algorithm is used to map the SURF feature vector to a low-dimensional space to generate a low-dimensional feature matrix, and the similarity between the SURF feature vectors in the matrix is ​​calculated, and the similarity between the high-dimensional SURF feature vectors is minimized based on the similarity P ij The difference in similarity with the low-dimensional SURF feature vector is taken as the goal, and iterative optimization is performed to output the feature matrix of the low-dimensional SURF feature vector.

[0038] As a preferred solution of the rotating machinery fault diagnosis method based on multi-signal fusion image construction described in the present invention, wherein: the cosine similarity of each pair of feature vectors is calculated, and grayscale value mapping is performed to output a grayscale image, and an RGB image is output based on displacement, velocity and acceleration signals, including:

[0039] Based on the feature matrix of the low-dimensional SURF feature vector, the cosine similarity S of each pair of feature vectors is calculated ij ;

[0040] The similarity S of each pair of feature vectors calculated is ij Fill it into the similarity matrix and perform grayscale value mapping, which is expressed as:

[0041] G ij =127.5·(S ij +1);

[0042] Among them G ij For S ij The grayscale value after mapping ranges from 0 to 255;

[0043] Based on all S ij Calculate the gray value G after mapping ij Get the gray value matrix G and output the grayscale image;

[0044] The grayscale images based on three different signals of displacement, velocity and acceleration are mapped to the R, G and B channels of the RGB image respectively.

[0045] As a preferred solution of the rotating machinery fault diagnosis method based on multi-signal fusion image construction described in the present invention, wherein: the prediction of the fault category probability of the input RGB image using the CNN convolutional neural network includes:

[0046] Use CNN convolutional neural network to build a deep learning model, including input layer, convolution layer, pooling layer, fully connected layer and output layer;

[0047] The input layer inputs the generated RGB image, the convolution layer uses the ReLU activation function to extract local features in the image, the pooling layer uses the maximum pooling to reduce the dimension of the feature map after convolution, the fully connected layer maps the pooled features to a high-dimensional vector, and the output layer uses the Softmax activation function to output the probability of the category;

[0048] Use the training set to determine the fault classification label, select the cross entropy loss function to calculate the difference between the category probability predicted by CNN and the actual label, use the Adam optimizer for gradient descent optimization, update the weights of the CNN model, and stop iterating if the model loss no longer decreases significantly during continuous iterations to output model parameters and update the model.

[0049] As a preferred solution of the rotating machinery fault diagnosis method based on multi-signal fusion image construction of the present invention, wherein: the threshold determination based on the prediction record includes:

[0050] Based on the trained CNN deep learning model, the newly acquired RGB image is input into the model, and the fault category and its corresponding probability predicted by the model are recorded;

[0051] Based on the continuously updated model prediction record data and the actual maintenance judgment of the maintenance personnel, the accuracy of the model prediction output is calculated, and the historical average accuracy is used as the prediction threshold. If the accuracy of the model prediction output is lower than the prediction threshold, the model is retrained using the training set.

[0052] Another object of the present invention is to provide a system for a rotating machinery fault diagnosis method based on multi-signal fusion images, which comprises:

[0053] The signal acquisition module collects signal data from the rotating mechanical system in real time and performs sliding window segmentation;

[0054] The phase space reconstruction module determines the time delay by the minimum mutual information method and the embedding dimension by Cao's nearest neighbor method to form a high-dimensional phase space vector;

[0055] The recursive binary graph generation module generates a recursive binary graph through the Heaviside function and uses wavelet transform to extract time-frequency features;

[0056] The feature extraction and dimensionality reduction module uses the SURF algorithm to detect local significant change points in the recursive binary graph and time-frequency graph, and performs dimensionality reduction processing through the t-SNE algorithm;

[0057] Image construction module, based on low-dimensional SURF feature vectors, calculates the cosine similarity of each pair of feature vectors, generates a grayscale image, and maps the grayscale image to an RGB image;

[0058] CNN model prediction module, which predicts the fault category based on the CNN model input RGB image;

[0059] The model update module records the results of each fault prediction and determines whether the model needs to be retrained based on the historical average accuracy.

[0060] A computer device comprises: a memory and a processor; the memory stores a computer program, and the processor implements the steps of the above-mentioned rotating machinery fault diagnosis method based on multi-signal fusion image construction when executing the computer program.

[0061] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the above-mentioned rotating machinery fault diagnosis method based on multi-signal fusion images.

[0062] The beneficial effects of the present invention are as follows: by collecting displacement, velocity and acceleration signals and using the minimum mutual information method to determine the optimal delay time, using Cao's nearest neighbor method to determine the optimal embedding dimension for phase space reconstruction, the dynamic evolution process of the signal is reflected, including nonlinear changes and periodic behaviors, and the long-term dependence and short-term fluctuations of the signal are fully described. The local significant change points in the time-frequency diagram and the recursive binary diagram are detected by the SURF algorithm, and the local frequency changes and transient characteristics of the signal can be captured simultaneously and the frequency distribution of the signal changing with time can be displayed, which helps to identify the time nodes of abnormalities or faults. The computational complexity is reduced by the t-SNE dimensionality reduction algorithm, making the operation of fault diagnosis more efficient. By calculating the cosine similarity and outputting the RGB image, a more comprehensive description of the fault state can be provided by combining the feature information of multiple signal sources, and rich input data is provided for the CNN model to quickly identify the fault type. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0064] Figure 1 The figure is a flow chart of a rotating machinery fault diagnosis method based on multi-signal fusion images.

[0065] Figure 2 This is a schematic diagram of the structure of a rotating machinery fault diagnosis system based on multi-signal fusion images. DETAILED DESCRIPTION

[0066] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below in conjunction with the accompanying drawings.

[0067] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0068] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor is it a separate or selective embodiment that is mutually exclusive with other embodiments.

[0069] Example 1, reference Figure 1 , which is the first embodiment of the present invention, and provides a rotating machinery fault diagnosis method based on multi-signal fusion image construction. The rotating machinery fault diagnosis method based on multi-signal fusion image construction includes:

[0070] S1, collects rotating machinery signals and uses sliding window segmentation for signal preprocessing;

[0071] Preferably, the rotating machinery signal is collected and the signal preprocessing is performed using sliding window segmentation, including:

[0072] The displacement, velocity and acceleration of rotating machinery are collected in real time through sensors. Fiber Bragg grating sensors (FBG) can be used for signal collection. Displacement data is obtained through laser interferometers or LVDTs (linear variable differential transformers), the speed of rotating machinery is directly obtained using rotary encoders or speed measuring motors, and mechanical vibration acceleration is obtained through high-precision acceleration sensors.

[0073] Sliding window segmentation is used for signal preprocessing. The sliding window length is determined based on the motion periodicity of the rotating machinery and the sampling rate of the collected signal, which is expressed as:

[0074] T=f s ·n s ·T r ;

[0075] Where T represents the sliding window length, f s represents the sampling frequency, n s Indicates the number of cycles covered, which means the number of cycles that the window length can cover. Usually 2 to 5 cycles are selected. r Indicates the mechanical rotation period, for example 0.02 seconds;

[0076] The sliding step length is determined based on the overlapping degree of each window sliding after the sliding window length. The signal data is read, and the collected displacement, velocity and acceleration data are respectively organized into arrays according to the time index, and the sliding window is applied for segmentation, which is expressed as:

[0077] S i (t k )={x i (t k ),x i (t k +1),…,x i (t k +T-1)};

[0078] Where S i (t k ) represents the signal segment of the i-th type signal in the k-th window, t krepresents the kth time t, T represents the sliding window length, x i Represents the collected value of the i-th type signal.

[0079] Real-time signal acquisition is performed through high-precision sensors, such as fiber grating sensors (FBG), laser interferometers, rotary encoders, accelerometers, etc., to ensure that the acquired signals have high precision and high sampling rate, so as to accurately reflect the state of the rotating machinery, especially the displacement, velocity and acceleration signals, which can respectively capture the movement amplitude, velocity change and vibration characteristics of the machinery, and provide rich fault diagnosis information. By using sliding window segmentation for signal preprocessing, the appropriate window length can be determined based on the periodic characteristics of the rotating machinery. The application of sliding windows can smooth the real-time acquired signals to avoid interference caused by instantaneous fluctuations. Using sliding windows to segment signals helps to divide the data collected for a long time into smaller segments, so that each segment can be analyzed independently. The preprocessing method based on sensor signal acquisition and sliding window segmentation helps to assist in the analysis and judgment of the acquired signals.

[0080] S2, the minimum mutual information method is used to determine the time delay, the Cao's nearest neighbor method is used to calculate the similarity measure of adjacent points and determine the embedding dimension, and the phase space is reconstructed based on the time delay and embedding dimension;

[0081] Preferably, the time delay is determined by the minimum mutual information method, the Cao's nearest neighbor method is used to calculate the similarity measure of adjacent points and determine the embedding dimension, and the phase space is reconstructed based on the time delay and the embedding dimension, including:

[0082] The minimum mutual information method is used to discretize the continuous time signal segment S(t), that is, the signal value range is divided into N intervals, each interval corresponds to a signal value s i , calculate the probability distribution of each discrete interval in the signal S(t), that is, calculate the number of times the signal value falls into interval i divided by the total number of signal samples as the signal occurrence probability P(s i ), and calculate the entropy value of the segmented signal segment, and calculate the entropy value of the delayed signal according to the historical data traversal time delay τ, which is expressed as:

[0083]

[0084] where s i represents the value of the ith interval of the signal segment S(t), P(s i ) represents the probability of the ith interval after the signal segment S(t) is discretized, N represents the total number of intervals, H(S) represents the entropy value of the signal segment S(t), H(S τ ) represents the delayed signal S τ (t), where S τ(t) = S (t + τ), P (s τ,i ) represents the delayed signal S τ (t) The probability of the i-th interval after discretization;

[0085] According to the signal segment S(t) and the delayed signal S τ (t) Calculate the joint entropy and calculate the signal segment S(t) and the delayed signal S τ The mutual information of the entropy value and the joint entropy is obtained by determining the signal segment S(t) and the delayed signal S τ (t) and determine the optimal delay time τ, which is expressed as:

[0086]

[0087] I(τ)=H(S)+H(S τ )-H(S,S τ );

[0088] Where I(τ) represents the mutual information between the signal segment S(t) and the delayed signal Sτ(t), and H(S,Sτ) represents the mutual information between the signal segment S(t) and the delayed signal S τ The joint entropy of (t), P(s i ,s τ,j ) represents the value s in the ith interval of the signal segment S(t) i And the delayed signal S τ (t) The jth interval takes value s τ,j The joint probability of N is the total number of intervals, H(S) is the entropy value of the signal segment S(t), and H(S τ ) represents the delayed signal S τ The entropy value of (t);

[0089] Based on the historical data, the embedding dimension is initialized and the phase space vector is initially generated. The Cao's nearest neighbor method is used to calculate the similarity measure of adjacent points, which is expressed as:

[0090]

[0091] Where E1(m) and E2(m) represent the similarity measures of two adjacent points in the embedding dimension m, N represents the total number of intervals, and mτ represents the effective sample reduction due to the embedding dimension and time delay. is used to average the similarity measure of all N-mτ neighboring points, ||Y(i+m)-Y(i)|| represents the Euclidean distance between the phase space vectors Y(i+m) and Y(i), Y(i) represents the phase space vector at the embedding dimension m and the delay time τ, used to represent the vector generated by the signal segment S(t) at time i, Y(i+m) represents the phase space vector at time i+m, which is the same vector as Y(i), but starts from time i+m, representing the vector generated by the signal segment S(t) at time i+m, S(i+k·τ) represents the signal value of the kth component at time i, where k is an index from 1 to the embedding dimension m for the phase space vector;

[0092] Gradually increase the value of the embedding dimension and recalculate the adjacent point similarity measure until the change in the calculated adjacent point similarity measure is no longer obvious, and extract the embedding dimension m at this time;

[0093] Combining the minimum delay time and the extracted embedding dimension, the final phase space vector is calculated for each time signal segment S(t), expressed as:

[0094] Y(i)={S(i),S(i+τ),S(i+2τ),…S(i+(m-1)τ)};

[0095] Where Y(i) represents the generated phase space vector, which represents the representation of the signal at time i in the embedded high-dimensional space, S(i) represents the value of the signal segment S(t) at time i, τ represents the delay time, m represents the number of embedding dimensions, S(i+τ) represents the value of the signal segment S(t) at time i+τ, and S(i+(m-1)τ) represents the value of the signal segment S(t) starting at time i after m-1 delays τ.

[0096] The goal of phase space reconstruction is to characterize the dynamic behavior of one-dimensional time series signals through high-dimensional vectors. This method can effectively capture the nonlinear characteristics and complex dynamic characteristics hidden in the signal. The reasonable selection of time delay can ensure that the signal has an appropriate distribution in time, so that the reconstructed phase space vector can fully show the dynamic characteristics of the signal. The calculation of time delay τ directly determines the shape of the phase space vector, which can ensure that the trajectory after phase space reconstruction has sufficient complexity and dynamic expression in high-dimensional space. At the same time, calculating the optimal embedding dimension can capture the nonlinear characteristics of the signal and avoid noise and redundancy caused by excessive embedding.

[0097] By determining the optimal delay time τ, which is used to represent the time interval between adjacent signal points, it is ensured that the embedded high-dimensional vector can represent different states of the signal. The minimum mutual information method can quantify the dependency between the signal segment and its delayed signal, identify the minimum dependency point within the time delay range, and find the information independence of the signal at different time points, that is, find the optimal time delay, so that the signal becomes more independent in time, thereby more accurately describing the dynamic change characteristics of the signal. Compared with the traditional empirical selection of time delay method, the mutual information method can more scientifically and quantitatively determine the relationship between signal features, reduce the error caused by subjective selection of time delay, and at the same time, determine the optimal delay time by the mutual information method, which can avoid the dependency between the signal and its delayed signal being too strong, resulting in the extracted features having information redundancy and avoiding the situation where there is almost no dependency between the signal and its delayed signal.

[0098] By gradually increasing the embedding dimension and using Cao's nearest neighbor method to calculate the similarity measure of adjacent points, the complex behavior of signal fragments in high-dimensional space can be accurately identified. The selection of the optimal embedding dimension ensures that the signal can fully express its nonlinear dynamic characteristics in high-dimensional space without losing information due to too small an embedding dimension or introducing unnecessary noise due to too large an embedding dimension. Using Cao's nearest neighbor method can avoid the problem of artificially determining the embedding dimension in traditional methods. A reasonable embedding dimension can ensure that the dimensions contained in the phase space vector are both sufficient to cover the important features of the signal and not over-expanded to introduce irrelevant noise features.

[0099] For complex physical systems such as rotating machinery, their vibration signals usually exhibit complex nonlinear dynamics. Through the optimal delay time τ and the optimal embedding dimension m, the one-dimensional time series signal fragment can reflect the dynamic evolution process of the signal, including nonlinear changes and periodic behaviors, and comprehensively describe the long-term dependence and short-term fluctuations of the signal. The phase space vector generated by the optimal delay time and embedding dimension can make the signals of different fault modes show more obvious separation in the high-dimensional space, so that the signals in the normal state and the signals in the fault state will form different trajectories or patterns in the phase space, making the fault mode easier to distinguish.

[0100] S3, using the Heaviside function to generate a recursive binary map, using wavelet transform to extract time-frequency features and generate a time-frequency map, using the SURF algorithm to detect local significant change points in the time-frequency map and the recursive binary map, and using the t-SNE algorithm to generate a low-dimensional feature matrix;

[0101] Preferably, a Heaviside function is used to generate a recursive binary map, a wavelet transform is used to extract time-frequency features and generate a time-frequency map, a SURF algorithm is used to detect local significant change points in the time-frequency map and the recursive binary map, and a t-SNE algorithm is used to generate a low-dimensional feature matrix, including:

[0102] Based on the final phase space vector, the Euclidean distance between the state representations Y(i) and Y(j) at time i and time j is calculated to form a distance matrix d(i,j), and the distance threshold r is determined based on the average Euclidean distance;

[0103] The elements of the recursive binary graph matrix are calculated using the Heaviside function Θ, expressed as:

[0104] R(i,j)=θ(rd(i,j));

[0105] Where R(i,j) represents the element of the recursive binary graph matrix, Θ represents the Heaviside step function, and r represents the distance threshold;

[0106] Visualize the recursive binary matrix using image processing tools (such as Matplotlib, OpenCV, etc.) to generate a recursive binary graph;

[0107] Wavelet transform is used to extract time-frequency features of all collected signals, and the time-frequency diagram of each signal is constructed;

[0108] The SURF algorithm is used to detect local significant change points in the time-frequency graph and recursive binary graph, and the Hessian matrix and determinant of the image are calculated, which are expressed as:

[0109]

[0110] Where H(x,y) represents the Hessian matrix, which contains the second-order derivative information of the image at that point and is used to determine whether the point is a key point. xx (x,y) and L yy (x, y) represents the second-order derivative of the image at the (x, y) position along the x direction and the y direction, respectively, indicating the rate of change of the intensity of the point in the corresponding direction, L xy (x,y) represents the mixed second-order derivative of the image at the position (x,y) along the x and y directions, reflecting the interactive changes of the image in the two directions. det(H) represents the determinant of the Hessian matrix, indicating the local feature intensity of the point.

[0111] Based on the determinant det(H), the extreme points are determined as the key points of the SURF algorithm. For each key point, the SURF algorithm generates a fixed-length SURF descriptor as a feature vector and forms a SURF feature matrix.

[0112] Based on the SURF feature matrix, the t-SNE algorithm is used to calculate the similarity between each pair of SURF feature vectors, expressed as:

[0113]

[0114] P ij Represents the SURF feature vector Y i and Y j The similarity of Represents the SURF feature vector Y i and Y j The Euclidean distance of ||Y i -Y j || 2 Gaussian distribution function, σ represents the Gaussian kernel width, B represents the total number of SURF feature vectors, Represents the normalization term calculation, ensuring that the sum of all similarities is 1.

[0115] The t-SNE algorithm is used to map the SURF feature vector to a low-dimensional space to generate a low-dimensional feature matrix, and the similarity between the SURF feature vectors in the matrix is ​​calculated, and the similarity between the high-dimensional SURF feature vectors is minimized based on the similarity P ij The difference in similarity with the low-dimensional SURF feature vector is taken as the goal, and iterative optimization is performed to output the feature matrix of the low-dimensional SURF feature vector.

[0116] By using the Heaviside step function to generate a recursive binary graph, the state similarity between different time points of the signal can be clearly displayed. Point pairs with strong similarity appear as high-density points on the image, showing the periodicity and similarity of the system state. For fault detection, this image can intuitively reflect the system changes before and after the fault. By calculating the Euclidean distance between signal segments and determining the threshold, the recursive binary graph can display the nonlinear dynamic behavior of the signal. It can reflect the periodicity and chaotic behavior of the system and is very useful for fault detection of complex systems (such as rotating machinery). Wavelet transform can analyze signals at different time scales and is particularly suitable for processing non-stationary signals (such as mechanical vibration signals). It can simultaneously capture the local frequency changes and transient characteristics of the signal. The time-frequency graph shows the frequency distribution of the signal over time. This image form can better analyze the localized characteristics of complex signals and help identify the time nodes of abnormalities or faults.

[0117] The SURF (Speeded Robust Features) algorithm calculates the Hessian matrix and quickly detects local significant change points in the recursive binary graph and time-frequency graph, which can generate stable local features at a low computational cost. By cooperating with the t-SNE dimensionality reduction algorithm, the high-dimensional SURF feature vector is mapped to a low-dimensional space, and the local similarity between the feature vectors is retained as much as possible. In the low-dimensional space, the characteristics of different fault modes still maintain the differences in the high-dimensional space, so that different fault modes can be separated more clearly. At the same time, the computational complexity is reduced after dimensionality reduction, making the operation of fault diagnosis more efficient. The combination of Heaviside function to generate recursive binary graphs, wavelet transform to extract time-frequency features, SURF algorithm to detect local significant change points and t-SNE dimensionality reduction to generate low-dimensional feature matrices can effectively improve the accuracy and efficiency of fault detection. The combination of local significant feature detection and global feature extraction significantly improves the diagnostic accuracy of the model, making the system more robust and adaptable when processing nonlinear complex signals.

[0118] S4, calculating the cosine similarity of each pair of feature vectors, and performing grayscale value mapping to output a grayscale image, and outputting an RGB image based on the displacement, velocity and acceleration signals;

[0119] Preferably, the cosine similarity of each pair of feature vectors is calculated, and grayscale value mapping is performed to output a grayscale image, and an RGB image is output based on the displacement, velocity and acceleration signals, including:

[0120] Based on the feature matrix of the low-dimensional SURF feature vector, the cosine similarity of each pair of feature vectors is calculated, which is expressed as:

[0121]

[0122] Where S ij represents the similarity between the i-th and j-th features, F tS (i) and F tS (j) represent the i-th and j-th low-dimensional SURF feature vectors, respectively, ||F tS (i) || and ||F tS (j)||represents F tS (i) and F tS (j) the Euclidean norm;

[0123] The similarity S of each pair of feature vectors calculated is ij Fill it into the similarity matrix and perform grayscale value mapping, which is expressed as:

[0124] G ij =127.5·(S ij +1);

[0125] Among them Gij For S ij The grayscale value after mapping ranges from 0 to 255;

[0126] Based on all S ij Calculate the gray value G after mapping ij Get the gray value matrix G and output the grayscale image;

[0127] Based on the grayscale images of three different signals, namely displacement, velocity and acceleration, they are mapped to the R, G and B channels of the RGB image respectively. The displacement signal grayscale image is mapped to the R channel of the RGB image, the velocity signal grayscale image is mapped to the G channel of the RGB image, and the acceleration signal grayscale image is mapped to the B channel of the RGB image. The three different signals are mapped to different channels of the RGB image. The generated RGB image contains the red information of the displacement signal, the green information of the velocity signal, and the blue information of the acceleration signal.

[0128] By calculating the cosine similarity of each pair of low-dimensional SURF feature vectors, the similarity of signal features at different time points can be clearly displayed. By converting the similarity into grayscale values, an intuitive grayscale image can be generated, making the temporal trend of the signal more obvious. The grayscale value from 0 to 255 corresponds to the intensity of signal similarity. Areas with large signal differences will appear as darker areas in the image, while areas with high similarity will appear as brighter areas. In the form of grayscale images, local change points and abnormal behaviors of signals become easier to identify. Especially in fault detection scenarios, when a fault occurs, changes in local features of the signal will cause obvious abnormal patterns to appear in the grayscale image, making the originally complex signal features appear in a more concise way. When processing large-scale data, this compression method can greatly improve the calculation efficiency while retaining important feature information in the signal.

[0129] By mapping the grayscale images of three different signals, displacement, velocity and acceleration, to the R, G and B channels of the RGB image, the characteristics of multiple signals can be displayed in the same image, providing rich input data for the model. For the fault diagnosis of complex systems such as rotating machinery, combining the characteristic information of multiple signal sources can provide a more comprehensive description of the fault state, making the difference between fault signals and normal signals clear at a glance. Through color changes (differences in the intensity of red, green and blue channels), operators or researchers can intuitively observe the abnormality of the signal and the location of the fault point, facilitating rapid fault judgment.

[0130] S5, uses CNN convolutional neural network to predict the probability of fault category of input RGB image, and makes threshold judgment based on the prediction record;

[0131] Preferably, a CNN convolutional neural network is used to predict the probability of fault category of the input RGB image, including:

[0132] Use CNN convolutional neural network to build a deep learning model, including input layer, convolution layer, pooling layer, fully connected layer and output layer;

[0133] The input layer inputs the generated RGB image, the convolution layer uses the ReLU activation function to extract local features in the image, the pooling layer uses the maximum pooling to reduce the dimension of the feature map after convolution, the fully connected layer maps the pooled features to a high-dimensional vector, and the output layer uses the Softmax activation function to output the probability of the category;

[0134] Use the training set to determine the fault classification label, select the cross entropy loss function to calculate the difference between the category probability predicted by CNN and the actual label, use the Adam optimizer for gradient descent optimization, update the weights of the CNN model, and stop iterating if the model loss no longer decreases significantly during continuous iterations to output model parameters and update the model.

[0135] By building and training the CNN model, the synergy and mutual influence of multiple signals can be identified, thereby improving the detection accuracy of complex mechanical system faults. By comprehensively analyzing different signal characteristics, the CNN model has strong adaptability and robustness in dealing with various fault modes, and can quickly determine the fault type and give timely fault warnings.

[0136] Further, threshold determination is performed based on the predicted records, including:

[0137] Based on the trained CNN deep learning model, the newly acquired RGB image is input into the model, and the fault category and its corresponding probability predicted by the model are recorded;

[0138] Based on the continuously updated model prediction record data and the actual maintenance judgment of the maintenance personnel, the accuracy of the model prediction output is calculated, and the historical average accuracy is used as the prediction threshold. If the accuracy of the model prediction output is lower than the prediction threshold, the model is retrained using the training set.

[0139] By recording the fault category and probability predicted by the model for each new input RGB image, real-time monitoring of the model performance can be achieved. When the fault category predicted by the model does not match the actual maintenance result, the error source and fault type of the model can be quickly determined by comparing the feedback records of the maintenance personnel. By recording the accuracy of the model in long-term predictions and using the historical average accuracy as a reference threshold, the performance of the model can be dynamically tracked to prevent the model from degrading over time, reduce human intervention, and improve the model's adaptive ability. As the model continuously updates the prediction record data and actual maintenance feedback results, through long-term data accumulation, it can effectively reduce the model's false alarms (erroneously judging that the system is faulty) and missed alarms (missing faults), forming a data feedback closed loop, and enhancing the intelligence and data-driven capabilities of the fault diagnosis system.

[0140] Example 2, reference Figure 2 , which is the second embodiment of the present invention, and which is different from the previous embodiment, provides a system for a rotating machinery fault diagnosis method based on multi-signal fusion images, including:

[0141] The signal acquisition module collects signal data from the rotating mechanical system in real time and performs sliding window segmentation;

[0142] The phase space reconstruction module determines the time delay by the minimum mutual information method and the embedding dimension by Cao's nearest neighbor method to form a high-dimensional phase space vector;

[0143] The recursive binary graph generation module generates a recursive binary graph through the Heaviside function and uses wavelet transform to extract time-frequency features;

[0144] The feature extraction and dimensionality reduction module uses the SURF algorithm to detect local significant change points in the recursive binary graph and time-frequency graph, and performs dimensionality reduction processing through the t-SNE algorithm;

[0145] Image construction module, based on low-dimensional SURF feature vectors, calculates the cosine similarity of each pair of feature vectors, generates a grayscale image, and maps the grayscale image to an RGB image;

[0146] CNN model prediction module, which predicts the fault category based on the CNN model input RGB image;

[0147] The model update module records the results of each fault prediction and determines whether the model needs to be retrained based on the historical average accuracy.

[0148] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc., which can store program codes.

[0149] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute instructions), or in conjunction with such instruction execution systems, devices or apparatuses. For the purposes of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in conjunction with such instruction execution systems, devices or apparatuses.

[0150] More specific examples of computer-readable media (a non-exhaustive list) include the following: an electrical connection with one or more wires (electronic device), a portable computer disk case (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be a paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering or, if necessary, processing in another suitable manner, and then stored in a computer memory.

[0151] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above embodiments, a plurality of steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, any one of the following technologies known in the art or a combination thereof can be used to implement: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0152] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A rotating machinery fault diagnosis method based on multi-signal fusion images, characterized in that: include, Collect rotating machinery signals and use sliding window segmentation for signal preprocessing; The minimum mutual information method is used to determine the time delay, the Cao's nearest neighbor method is used to calculate the similarity measure of adjacent points and determine the embedding dimension, and the phase space is reconstructed based on the time delay and the embedding dimension; The Heaviside function is used to generate a recursive binary map, the wavelet transform is used to extract time-frequency features and generate a time-frequency map, the SURF algorithm is used to detect local significant change points in the time-frequency map and the recursive binary map, and the t-SNE algorithm is used to generate a low-dimensional feature matrix; Calculate the cosine similarity of each pair of feature vectors, perform grayscale value mapping to output a grayscale image, and output an RGB image based on the displacement, velocity, and acceleration signals; The CNN convolutional neural network is used to predict the probability of fault category of the input RGB image, and the threshold judgment is made based on the prediction record.

2. The rotating machinery fault diagnosis method based on multi-signal fusion image construction according to claim 1 is characterized in that: The method of collecting rotating machinery signals and performing signal preprocessing by using sliding window segmentation includes: The displacement, velocity and acceleration of rotating machinery are collected in real time through sensors; Sliding window segmentation is used for signal preprocessing. The sliding window length is determined based on the motion periodicity of the rotating machinery and the sampling rate of the collected signal, which is expressed as: T = fs·ns·Tr; Where T represents the sliding window length, fs represents the sampling frequency, ns represents the number of covered cycles, and Tr represents the mechanical rotation period; The sliding step length is determined based on the overlapping degree of each window sliding after the sliding window length. The signal data is read, and the collected displacement, velocity and acceleration data are respectively organized into arrays according to the time index, and the sliding window is applied for segmentation, which is expressed as: Si(tk)={xi(tk),xi(tk+1),…,xi(tk+T-1)}; Where Si(tk) represents the signal segment of the i-th signal in the k-th window, tk represents the k-th time t, T represents the sliding window length, and xi represents the collected value of the i-th signal.

3. The rotating machinery fault diagnosis method based on multi-signal fusion image construction according to claim 2 is characterized in that: The method of using the minimum mutual information method to determine the time delay, using Cao's nearest neighbor method to calculate the similarity measure of adjacent points and determine the embedding dimension, and reconstructing the phase space based on the time delay and the embedding dimension includes: The minimum mutual information method is used to discretize the continuous time signal segment S(t), calculate the probability distribution of each discrete interval in the signal S(t), and calculate the entropy value of the segmented signal segment. The entropy value of the delayed signal is calculated based on the historical data traversal time delay τ. Calculate the joint entropy according to the signal segment S(t) and the delayed signal Sτ(t), and calculate the mutual information between the entropy value of the signal segment S(t) and the delayed signal Sτ(t) and the joint entropy, and determine the optimal delay time τ by determining the minimum dependency between the signal segment S(t) and the delayed signal Sτ(t); Initialize the embedding dimension based on historical data and preliminarily generate the phase space vector, and use Cao's nearest neighbor method to calculate the similarity measure of adjacent points; Gradually increase the value of the embedding dimension and recalculate the adjacent point similarity measure until the change in the calculated adjacent point similarity measure is no longer obvious, and extract the embedding dimension m at this time; Combining the minimum delay time and the extracted embedding dimension, the final phase space vector is calculated for each time signal segment S(t), expressed as: Y(i)={S(i),S(i+τ),S(i+2τ),…S(i+(m-1)τ)}; Where Y(i) represents the generated phase space vector, which represents the representation of the signal at time i in the embedded high-dimensional space, S(i) represents the value of the signal segment S(t) at time i, τ represents the delay time, m represents the number of embedding dimensions, S(i+τ) represents the value of the signal segment S(t) at time i+τ, and S(i+(m-1)τ) represents the value of the signal segment S(t) starting at time i after m-1 delays τ.

4. The rotating machinery fault diagnosis method based on multi-signal fusion image construction according to claim 3 is characterized in that: The Heaviside function is used to generate a recursive binary graph, the wavelet transform is used to extract time-frequency features and generate a time-frequency graph, the SURF algorithm is used to detect local significant change points in the time-frequency graph and the recursive binary graph, and the t-SNE algorithm is used to generate a low-dimensional feature matrix. include, Based on the final phase space vector, the Euclidean distance between the state representations Y(i) and Y(j) at time i and time j is calculated to form a distance matrix d(i,j), and the distance threshold r is determined based on the average Euclidean distance; The elements of the recursive binary graph matrix are calculated using the Heaviside function Θ, expressed as: R(i,j)=θ(rd(i,j)); Where R(i,j) represents the element of the recursive binary graph matrix, Θ represents the Heaviside step function, and r represents the distance threshold; Visualize the recursive binary matrix and generate a recursive binary graph; Wavelet transform is used to extract time-frequency features of all collected signals, and the time-frequency diagram of each signal is constructed; Use the SURF algorithm to detect local significant change points in the time-frequency graph and recursive binary graph, and calculate the Hessian matrix and determinant of the image; Based on the determinant det(H), the extreme points are determined as the key points of the SURF algorithm. For each key point, the SURF algorithm generates a fixed-length SURF descriptor as a feature vector and forms a SURF feature matrix. The t-SNE algorithm is used to map the SURF feature vector to a low-dimensional space to generate a low-dimensional feature matrix, and the similarity between the SURF feature vectors in the matrix is ​​calculated. Based on the goal of minimizing the difference between the high-dimensional SURF feature vector similarity Pij and the low-dimensional SURF feature vector similarity, iterative optimization is performed to output the feature matrix of the low-dimensional SURF feature vector.

5. The rotating machinery fault diagnosis method based on multi-signal fusion image construction according to claim 4 is characterized in that: The cosine similarity of each pair of feature vectors is calculated, and grayscale value mapping is performed to output a grayscale image, and an RGB image is output based on the displacement, velocity and acceleration signals, including: Based on the feature matrix of low-dimensional SURF feature vectors, the cosine similarity Sij of each pair of feature vectors is calculated; The calculated similarity Sij of each pair of feature vectors is filled into the similarity matrix and grayscale value mapping is performed, which is expressed as: Gij = 127.5·(Sij+1); Where Gij is the grayscale value after Sij mapping, ranging from 0 to 255; Based on all Sij, the mapped grayscale values ​​Gij are calculated to obtain the grayscale value matrix G, and the grayscale image is output; The grayscale images based on three different signals of displacement, velocity and acceleration are mapped to the R, G and B channels of the RGB image respectively.

6. The rotating machinery fault diagnosis method based on multi-signal fusion image construction according to claim 5 is characterized in that: The CNN convolutional neural network is used to predict the probability of fault category of the input RGB image, including: Use CNN convolutional neural network to build a deep learning model, including input layer, convolution layer, pooling layer, fully connected layer and output layer; The input layer inputs the generated RGB image, the convolution layer uses the ReLU activation function to extract local features in the image, the pooling layer uses the maximum pooling to reduce the dimension of the feature map after convolution, the fully connected layer maps the pooled features to a high-dimensional vector, and the output layer uses the Softmax activation function to output the probability of the category; Use the training set to determine the fault classification label, select the cross entropy loss function to calculate the difference between the category probability predicted by CNN and the actual label, use the Adam optimizer for gradient descent optimization, update the weights of the CNN model, and stop iterating if the model loss no longer decreases significantly during continuous iterations to output model parameters and update the model.

7. The rotating machinery fault diagnosis method based on multi-signal fusion image construction according to claim 6 is characterized in that: The threshold determination based on the prediction record includes: Based on the trained CNN deep learning model, the newly acquired RGB image is input into the model, and the fault category and its corresponding probability predicted by the model are recorded; Based on the continuously updated model prediction record data and the actual maintenance judgment of the maintenance personnel, the accuracy of the model prediction output is calculated, and the historical average accuracy is used as the prediction threshold. If the accuracy of the model prediction output is lower than the prediction threshold, the model is retrained using the training set.

8. A system based on the rotating machinery fault diagnosis method based on multi-signal fusion image construction according to any one of claims 1 to 7, characterized in that: include, The signal acquisition module collects signal data from the rotating mechanical system in real time and performs sliding window segmentation; The phase space reconstruction module determines the time delay by the minimum mutual information method and the embedding dimension by Cao's nearest neighbor method to form a high-dimensional phase space vector; The recursive binary graph generation module generates a recursive binary graph through the Heaviside function and uses wavelet transform to extract time-frequency features; The feature extraction and dimensionality reduction module uses the SURF algorithm to detect local significant change points in the recursive binary graph and time-frequency graph, and performs dimensionality reduction processing through the t-SNE algorithm; Image construction module, based on low-dimensional SURF feature vectors, calculates the cosine similarity of each pair of feature vectors, generates a grayscale image, and maps the grayscale image to an RGB image; CNN model prediction module, which predicts the fault category based on the CNN model input RGB image; The model update module records the results of each fault prediction and determines whether the model needs to be retrained based on the historical average accuracy.

9. A computer device comprising: A memory and a processor; the memory stores a computer program, characterized in that: when the processor executes the computer program, the steps of the rotating machinery fault diagnosis method based on multi-signal fusion image construction described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the rotating machinery fault diagnosis method based on multi-signal fusion image construction described in any one of claims 1 to 7 are implemented.

Images