Rotating machinery health state evaluation method based on image information regression analysis

Abstract

The application discloses a kind of based on image information regression analysis's rotating machinery health state evaluation method, including the vibration signal of the core component of rotating machinery to be evaluated is collected, information fusion is carried out to the time series of vibration signal composition and obtains composite signal, analysis is carried out to composite signal by taking continuous wavelet transform and obtains time-frequency image, pre-processing is carried out to time-frequency image, and compressed gray time-frequency image is obtained, the first-order image is obtained by taking intrinsic orthogonal decomposition and analyzing gray time-frequency image from the angle of spatial domain, health factor calibration of the first-order image is carried out according to the running time of rotating machinery, the corresponding model between the first-order image component and health factor is established using convolutional neural network, and the health state of unknown state rotating machinery core component to be evaluated is evaluated. Using the application can accurately obtain the health state of various rotating machinery core components, and improve the intelligent operation and maintenance level of various rotating machinery core components.

Description

Technical Field

[0001] This invention belongs to the field of mechanical equipment condition monitoring and fault diagnosis, and relates to a method for assessing the health status of rotating machinery based on image information regression analysis. Background Technology

[0002] In recent years, with the rapid development of sensor testing technology, digital signal processing, and machine learning, data-driven methods for assessing the health status of mechanical equipment have yielded fruitful results and have been widely applied in fields such as energy and power, rail transportation, and aerospace and military industries. These studies primarily focus on fault mode classification, and while they also assess the degree of fault, their essence remains primarily classification models. However, after long-term operation under certain loads, mechanical equipment inevitably degrades over time, and the stage of this degradation process is closely related to the equipment's health status. With the increasing level of information intelligence, the requirements for health monitoring of mechanical equipment are constantly rising, and aspects such as health status and lifespan prediction are becoming increasingly important. Correspondingly, condition monitoring and fault diagnosis technologies are rapidly evolving from qualitative classification diagnosis to quantitative degree identification. Therefore, achieving the transition from fault diagnosis to health assessment is the main development direction and goal of the field of condition monitoring and fault diagnosis in the near future.

[0003] In this type of technical research, the construction of high-quality health factors is typically used to achieve quantitative fault characterization and equipment health management. The rational and effective construction of health factors is a crucial foundation for assessing the degradation process of mechanical equipment and predicting its remaining life. The acquisition of health factors largely depends on abundant monitoring information and effective data analysis. In recent years, with the promotion and popularization of machine learning methods, various intelligent neural network models have been widely used in the health status monitoring of mechanical equipment. Deep network models, represented by deep learning theory, possess stronger regression and generalization capabilities, pointing the way for the development of equipment health status assessment.

[0004] The core of establishing a health status assessment model for mechanical equipment lies in data mining and analysis of monitoring signals from different stages of the equipment's progression from health to failure. This data mining extracts quantitative features that characterize the degree of mechanical failure or health status, enabling the construction of a quantitative model relating these features to the degree of degradation. Once the model is established, users can determine the corresponding degree of degradation of the mechanical equipment based on the quantitative features corresponding to the current input signal, thus achieving the health status assessment objective. Reasonable and effective quantitative features constitute effective health factors.

[0005] Rotating machinery, as an important type of mechanical equipment, is diverse, including steam turbines, gas turbines, centrifugal compressors, generators, water pumps, water turbines, fans, and electric motors. Its core components include main shafts, bearings, gears, and couplings. These components are mostly operated under high load and high speed conditions for extended periods. Failures can lead to significant economic losses and safety risks. Therefore, conducting health status assessments of key components can effectively improve equipment management and maintenance levels, and reduce the probability of sudden risks and failures. Summary of the Invention

[0006] The purpose of this invention is to provide a method for assessing the health status of rotating machinery based on image information regression analysis, which can accurately obtain the health status of various core components of rotating machinery and improve the intelligent operation and maintenance level of various core components of rotating machinery.

[0007] The technical solution adopted in this invention is a method for assessing the health status of rotating machinery based on image information regression analysis. This method includes the following steps:

[0008] Step 1: Collect vibration signals from the core components of the rotating machinery to be evaluated;

[0009] Step 2: Perform information fusion on the time series of vibration signals to obtain a synthesized signal;

[0010] Step 3: Analyze the synthesized signal using continuous wavelet transform to obtain a time-frequency image;

[0011] Step 4: Preprocess the time-frequency image to obtain a compressed grayscale time-frequency image;

[0012] Step 5: Use intrinsic orthogonal decomposition to analyze the compressed grayscale time-frequency image from the spatial domain perspective to obtain the first-order image;

[0013] Step 6: Perform health factor calibration on the first-order image based on the operating time of the rotating machinery;

[0014] Step 7: Use a convolutional neural network to establish a correspondence model between first-order image components and health factors to assess the health status of core components of rotating machinery in unknown states.

[0015] Step 1 involves acquiring the full-life-cycle vibration signals of the core components of the rotating machinery to be evaluated using an accelerometer. The acquired vibration signals include two mutually perpendicular directions, and multi-channel synchronous acquisition is employed. The corresponding vibration amplitude signals are denoted as f. x (t) and f y (t).

[0016] In step 3, let ψ(t) be a function that is defined in a local region and tends to zero outside this region, with a mean of zero. ψ(t) is called the mother wavelet, and its function expression is:

[0017]

[0018] Where w0 represents the center frequency, i represents the complex number identifier in the complex trigonometric function, and t represents time;

[0019] Let the synthesized signal f(t) to be analyzed belong to space L. 2 (R), then the continuous wavelet transform of the synthesized signal f(t) is:

[0020]

[0021] Where p is the scale, p>1 means the signal waveform shrinks, p<1 means the signal waveform stretches, q is the displacement, and f(t) is the synthesized signal. It is for ψ p,q According to the definition of the inner product in the complex frequency domain, we take its complex conjugate, ψ p,q A series of wavelet basis functions ψ are obtained by scaling and time shifting the mother wavelet ψ(t). p,q As shown in the following formula:

[0022]

[0023] Transformation result W f (p,q) represents the mother wavelet selecting a basic center frequency, then obtaining a series of different center frequencies through scaling transformation, and then obtaining a series of basis functions in different intervals through time shifting. These are then multiplied with a certain interval of the original signal and integrated. The frequency corresponding to the extreme value is the frequency contained in that interval of the original signal. After transformation, the relationship between time and frequency of the signal to be analyzed is obtained. Plotting this as an image is the time-frequency image of the continuous wavelet transform.

[0024] In step 4, the time-frequency image is preprocessed, including image size compression and grayscale conversion.

[0025] In step 5, intrinsic orthogonal decomposition is used to analyze the compressed grayscale time-frequency image from the perspective of the spatial domain, transforming a time-frequency image into multiple modal components, finding the main feature structures in the image, and obtaining the first-order image.

[0026] Step 6 specifically includes setting the health factor of the first-order image corresponding to the damaged or failed core components of the rotating machinery to 0, and setting the health factor of the first-order image corresponding to the brand-new core components of the rotating machinery to 1. These factors are then calibrated linearly with points corresponding to different operating times, according to the following formula:

[0027]

[0028] In the formula, the time t during which the equipment under test is used when the core components of the rotating machinery are damaged is t_t. max The corresponding health factor is 0; when the core components of the rotating machinery are brand new, the usage time t is 0, corresponding to a health factor of 1; the health factors for the other states are...

[0029] Step 7 specifically includes the following steps:

[0030] Step 7.1: Build a convolutional neural network model to extract health factors from the time-frequency graph of vibration signals of core components of rotating machinery;

[0031] Step 7.2: Evaluate the performance of the obtained health factors. The evaluation criteria are monotonicity and trend. Obtain the performance index of the health factors. If the performance index is low, fine-tune the parameters of the convolutional neural network and continuously iterate and optimize to improve the performance index of the health factors.

[0032] Step 7.3: Based on the collected data of different health stages of the core components of the rotating machinery, calibration is performed to obtain the health stage division on the health factor curve;

[0033] Step 7.4: Collect vibration signals from two directions of the core component of the rotating machinery under unknown state to be evaluated, fuse them, and transform them into a time-frequency image. Then, analyze and obtain a first-order image. Input the first-order image into the established convolutional neural network model to obtain health factors, and then evaluate the health status of the core component of the rotating machinery under unknown state to be evaluated.

[0034] In step 7.2, the formula for calculating the monotonicity of health factors is:

[0035]

[0036] In the formula, K represents the total number of health factors in the sequence; d / dx = X K+1 -X K The value represents the differential value of the health factor sequence; No.ofd / dx>0 and No.ofd / dx<0 represent positive and negative differential values, respectively. The monotonicity value Mon(X) of the health factor is located in the interval [0,1]. The closer its value is to 1, the better the monotonicity of the health factor.

[0037] In step 7.2, the formula for calculating the trend of health factors is:

[0038]

[0039] In the formula, It represents the average health factors over the entire life cycle; T represents the operating cycle. It is the average value of each time period within the whole life cycle. The trend value of the health factor Tre(H,T) is located in the interval [0,1]. The closer its value is to 1, the better the trend of the health factor.

[0040] The beneficial effects of this invention are:

[0041] 1) The vibration signal is converted into a time-frequency graph based on the vibration signal, giving full play to the advantages of convolutional neural network image analysis and improving the effectiveness of extracting health status features of core components of rotating machinery;

[0042] 2) The health factor assessment model construction method is based on the continuous wavelet analysis results of vibration signals, which can obtain time-frequency joint distribution information and thus capture richer and more subtle features;

[0043] 3) Intrinsic orthogonal decomposition can realize the function of image information analysis without artificial parameter construction, and provides an adaptive method for removing redundant information;

[0044] 4) Normalize the lifespan of monitored components and set a linearly normalized health factor, which is applicable to different components and corresponding working conditions.

[0045] 5) By leveraging the powerful generalization capabilities of deep learning models, image information is cleverly transformed, and nonlinear relationships are retained in the network model. This allows the nonlinear health degradation process to be forcibly labeled as a linear output relationship, and the output relationship to be simplified by a linear classifier.

[0046] 6) The method of the present invention does not rely on any empirical knowledge or empirical parameters, has wide applicability, and can be used for health status assessment of core components in various rotating machinery equipment. Attached Figure Description

[0047] Figure 1 This is a schematic diagram of the process for assessing the health status of rotating machinery based on image information regression analysis according to the present invention;

[0048] Figure 2 This is a schematic diagram of the time-frequency image preprocessing method in this invention;

[0049] Figure 3 This is a schematic diagram illustrating the principle of the nearest neighbor interpolation method in this invention;

[0050] Figure 4 This is a flowchart of the intrinsic orthogonal decomposition process in this invention. Detailed Implementation

[0051] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0052] This invention relates to a method for assessing the health status of rotating machinery based on image information regression analysis, referring to... Figure 1 This includes the acquisition of vibration signals from core components of rotating machinery, the synthesis of vibration signals, the acquisition of time-frequency images of vibration signals, the use of intrinsic mode decomposition to remove redundancy, the calibration of health factors, the construction of convolutional neural networks to conduct regression analysis, and the assessment of the health status of unknown samples.

[0053] A bearing health status assessment method based on image information regression analysis specifically includes the following steps:

[0054] Step 1: Acquire the vibration signal of the bearing under evaluation throughout its entire life cycle using a piezoelectric accelerometer. The acquired vibration signal includes two mutually perpendicular directions. Since the acquired signal is in two directions, at least two sets of accelerometers are required, employing multi-channel synchronous acquisition at a sampling frequency of 12kHz. Time alignment must be achieved throughout the process. The acquisition location needs to consider that when the bearing fails, it will generate pulse signals with short impact times, a wide frequency range, and rapid attenuation. Therefore, the vibration signal should be acquired as close as possible to the load location of the bearing to obtain a high-quality vibration signal source. In this invention, the accelerometers are installed in the horizontal and vertical directions of the bearing, and their corresponding vibration amplitude signals are denoted as f. x (t) and f y (t);

[0055] Step 2: Perform information fusion on the time series of vibration signals to synthesize them into a new signal, namely the synthesized signal f(t), calculated as shown in the following formula:

[0056]

[0057] Synthetic signals contain data change characteristics in two directions, which can comprehensively reflect the equipment status;

[0058] Step 3: Perform continuous wavelet transform analysis on the synthesized signal. Continuous wavelet transform can observe the signal characteristics in both the time and frequency domains simultaneously, overcoming the global limitation of Fourier transform.

[0059] First, let ψ(x) be a function that is defined in a local region and tends to zero outside this region, and the mean of the function is zero, that is, it satisfies the condition that... ψ(x) is called the mother wavelet. Mother wavelets include the Morlet wavelet, Daubechies wavelet, and Haar wavelet. This invention uses the Morlet wavelet, which is constructed by multiplying a complex trigonometric function by an exponential decay function. The complex trigonometric function can identify the frequency, and the decay function can guarantee its finite-time support property. The function expression is:

[0060]

[0061] w0 represents the center frequency, i represents the complex number identifier in the complex trigonometric function, t represents time, and ψ p,q A series of wavelet basis functions ψ are obtained by scaling and time shifting the mother wavelet ψ(t). p,q The specific form is shown in the following formula:

[0062]

[0063] Where p is the scale, p>1 means the signal waveform shrinks, and p<1 means the signal waveform stretches; the scale p also affects the attenuation function, thereby changing the time domain support interval, thus allowing time to be located; q is the displacement.

[0064] Let the synthesized signal f(t) to be analyzed belong to space L. 2 If (R), then the continuous wavelet transform of the signal is:

[0065]

[0066] in, It is for ψ p,q Take its complex conjugate according to the definition of the complex frequency domain inner product.

[0067] Its transformation result W f (p,q) represents Morlet selecting a fundamental center frequency, then obtaining a series of different center frequencies through scaling, followed by time shifting to obtain a series of basis functions in different intervals, and then multiplying each of these by a segment of the original signal (corresponding to the interval of the basis function) and integrating them. The frequency corresponding to the extreme value is the frequency contained in this interval of the original signal. That is, the continuous wavelet transform can be regarded as a bandpass filter, which only allows signals in the signal to be transformed whose frequency is close to the wavelet center frequency (after scaling) to pass through.

[0068] Based on the above analysis, the relationship between time and frequency of the synthesized signal to be analyzed is obtained. Plotting this relationship as an image yields the time-frequency image of the continuous wavelet transform.

[0069] Step 4: Preprocess the time-frequency image to obtain the compressed grayscale time-frequency image.

[0070] Reference Figure 2Preprocessing includes image compression and grayscale conversion. The image compression package performs geometric transformation and resampling on the image. Image geometric transformation mainly involves image reduction, which selects or processes the original data to obtain data of the desired reduced size while preserving as many original features as possible. This embodiment uses a proportionally reduced image geometric transformation. Let the original image size be M*N, and the reduced size be K1M*K2N (K1=K2<1). More specifically, let the original image be F(i,j), i=1,2,...,M,j=1,2,...,N, and the reduced image be G(x,y), x=1,2,...,k1M,y=1,2,...,k2N. The specific compression coefficients K1 and K2 are determined based on the original image size and the target image size, and are usually selected in the range of 32×32 to 256×256 pixels. In this embodiment, the time-frequency image is compressed to 128×128 pixels after geometric transformation.

[0071] Grayscale conversion involves calculating the time-frequency image in grayscale form. Compared to RGB images, a single-channel image reduces the number of channels by two-thirds, thus reducing the parameters of the convolutional neural network by two-thirds. Simultaneously, a single-channel image is more conducive to subsequent intrinsic orthogonal decomposition calculations, effectively compressing the preprocessed time-frequency information. The simplest grayscale conversion algorithm directly maps the three color information components (R, G, B) of all pixels in the color image to grayscale values. This embodiment uses the maximum value method, taking the maximum value among the R, G, and B channels of all pixels in the color image as the grayscale value of that pixel. The expression is as follows:

[0072] S(x,y)=max(R(x,y),G(x,y),B(x,y))

[0073] To further reduce image size and preserve important features as much as possible, the image is then resampled using interpolation. This embodiment employs nearest-neighbor interpolation, as illustrated in the diagram below. Figure 3 As shown, when the image is scaled down, the coordinate correspondence between the source and target images is determined as follows:

[0074] srcX=datX*(srcWidth / dstWidth)

[0075] srcY=dstY*(srcHeight / dstHeight)

[0076] Where srcX and srcY are the source image coordinates, dstX and dstY are the target image coordinates, srcWidth is the source image width, dstWidth is the target image width, srcHeight is the source image height, and dstHeight is the target image height. In this way, a pixel in the scaled-down image can correspond to a pixel in the source image, but it may be a floating-point pixel. The source image cannot directly provide the gray value corresponding to the floating-point pixel. This floating-point pixel is recorded as the sampling point. The nearest neighbor interpolation method assigns the gray value of the nearest neighbor pixel to the sampling point to the sampling point. Let the coordinates of the sampling point be (i+u,j+v), where i and j are non-negative integers, and u and v are floating-point numbers greater than 0 and less than 1. Then the gray value of the sampling point is f(i+u,j+v). If the point i+u,j+v falls in area A, i.e., u<0.5, v<0.5, then the gray value of the top-left pixel is assigned to the sampling point. Similarly, if it falls in area B, the gray value of the top-right pixel is assigned; if it falls in area C, the gray value of the bottom-left pixel is assigned; and if it falls in area D, the gray value of the bottom-right pixel is assigned. At this point, the grayscale value corresponding to each pixel of the target image has been obtained. In this embodiment, the image size is finally compressed to 64×64.

[0077] The parameters of the input layer of a convolutional neural network are closely related to the image size. Reasonable preprocessing will avoid setting too many invalid parameters, allowing the use of smaller convolutional kernels to reduce the amount of computation, and smaller inputs can prevent the model complexity from becoming too high, reduce the risk of overfitting, significantly improve the model training efficiency, and enhance the generalization ability.

[0078] Step 5: Use intrinsic orthogonal decomposition to analyze the compressed grayscale time-frequency image from the spatial domain perspective, transform a spectrum image into multiple modal components, describe the time-frequency information from different scales, and obtain a first-order image;

[0079] Intrinsic orthogonal decomposition (IOD) can identify the most significant changes in the feature structure of an image, effectively extracting multi-scale structures and facilitating the capture of hidden details in spectral images. IOD represents each image as a vector and decomposes the samples using linear algebra and other theories. After IOD analysis, a time-frequency image yields a series of sub-images, which are arranged according to their information weights. Higher-order component images typically contain the main information elements. The first-order image information obtained from IOD analysis retains key time-frequency information after continuous wavelet transform and reduces image redundancy and complexity, facilitating subsequent regression analysis.

[0080] Reference Figure 4 The specific implementation steps of intrinsic orthogonal decomposition are as follows:

[0081] 1) Let the obtained time-frequency grayscale image A be an m×n matrix. In this embodiment, m = 128, n = 128, and x is an n-dimensional vector. Projecting A directly onto the projection axis x yields an m-dimensional feature projection vector y, thus transforming the image matrix into a column vector, i.e.:

[0082] y = Ax

[0083] 2) Construct the optimal projection axis x by analyzing the distribution of the feature projection vector y, i.e., when J(x) reaches its maximum value. The calculation formula is as follows:

[0084]

[0085] Where S x Let tr(S) be the covariance matrix of the projected eigenvector y. x ) is S x The traces.

[0086] 3) Define the image covariance matrix:

[0087] C = E(A - EA) T (A-EA)

[0088] Where C is an n×n non-negative definite matrix, and n is the pixel value of each column of the image, which can be directly calculated using sample data. If there are M image samples in total, there may be differences of several orders of magnitude between data in different dimensions. If calculated directly, dimensions with a small amount of data may produce very large deviations due to floating-point calculation errors during intrinsic orthogonal decomposition, thus resulting in errors. Therefore, for all image samples A... j (j=1,2,...,M) are centered, and C can be calculated by the following formula:

[0089]

[0090] in The mean image, i.e.

[0091] Then we can get the following formula

[0092] J(X) = x T Cx

[0093] Where X is an orthogonal normalized column vector, if there exists an X such that the generalized overall scattering criterion holds, then x is called the optimal projection axis. Its physical meaning is: the overall eigenvector obtained by projecting the image matrix onto the axis has the greatest dispersion; that is, the unit eigenvector corresponding to the largest eigenvalue of the overall image scattering matrix is ​​this optimal projection axis.

[0094] 4) Solve for the eigenvalues ​​(λ1, λ2, ..., λ) of the covariance matrix C.n ) and their corresponding eigenvectors (x1, x2, ..., x n ), and λ1>λ2>...>λ n Among them, λ1 corresponds to the largest variance, λ n The corresponding variance is minimized. Based on the contribution rate and cumulative contribution rate, the principal component d is selected, and the projection axis group X = [x1, x2, ..., x...] is determined. d ].

[0095] The projections of the image matrix X onto the projection axes or groups of projection axes are:

[0096]

[0097] x i (i = 1, 2, ..., d) represents the i-th principal feature extracted from image A.

[0098] 5) The optimal projection axis set is X = [x1, x2, ..., x...]. d The selection of the number of projection axes d in the image can be expressed by the following formula:

[0099]

[0100] Where λ1, λ2, ..., λ n θ represents the n largest eigenvalues ​​of C, and θ is a pre-set threshold, which is set to 0.6 in this embodiment.

[0101] 6) The principal component extraction of the image is shown in the following formula:

[0102] y i =Ax i (i = 1, 2, ..., d)

[0103] Where y i (i = 1, 2, ..., d) are the principal components of the image matrix A, and the image reconstruction is completed. Let X = [x1, x2, ..., xd] respectively. d ], Y = [y1, y2, ..., y d X is called the optimal projection matrix, and Y is the characteristic matrix of matrix A.

[0104] Step 6: Calibrate the health factor of the first-order image based on the bearing's operating time. This includes setting the health factor of the first-order image corresponding to bearing damage or failure to 0, and setting the health factor of the first-order image corresponding to a brand-new bearing to 1. Then, calibrate the health factor with points corresponding to different operating times using a linear relationship, according to the following formula:

[0105]

[0106] In the formula, the time t during which the tested equipment is used when the bearing fails is t_t. max The corresponding health factor is 0; when the bearing is brand new, the usage time t is 0, corresponding to a health factor of 1; the health factors for the other states are... Assuming the bearing's lifespan is 22,000 seconds, then after the bearing has operated for 15,400 seconds, the label will be set to... The dataset is composed of time-frequency images and health factors. The health status of the bearing is divided into normal state, initial deterioration state, severe damage state, and critical state.

[0107] Step 7: Establish a correspondence model between first-order image components and health factors using a convolutional neural network to assess the health status of core components of rotating machinery in an unknown state. This includes the following steps:

[0108] Step 7.1: Build a convolutional neural network model to extract health factors from the time-frequency graph of vibration signals of core components of rotating machinery;

[0109] Neural networks can deeply mine hidden features of images at different scales. Since the output of a neural network represents the regression calculation results for the lifespan of mechanical equipment, and given the inevitable differences between individual components of the same type during the failure process, it is necessary to normalize the network output representing the regression calculation results to limit the lifespan of different components to the same metric. This is achieved by selecting an appropriate activation function to limit the lifespan to between 0 and 1. Through the nonlinear mapping capability of convolutional neural networks, the complex and varied spectral image information is transformed into a linear output relationship, resulting in a health factor assessment model.

[0110] The network framework of the health factor regression model is set as a convolutional neural network, which consists of convolutional layers, pooling layers, activation functions, etc. The mathematical expression of the convolutional layer operation is shown in the following formula;

[0111]

[0112] In the formula: d is the number of convolutional layers; M v The number of feature maps in the (d-1)th layer; This is the v-th output feature map of the d-th layer; This refers to the v-th convolutional kernel in the d-th layer that is associated with the i-th feature map. f(·) is the bias of the v-th convolutional kernel in the d-th layer; f(·) is the activation function.

[0113] Pooling layers mainly include max pooling layers and average pooling layers. During feature extraction, max pooling layers focus on capturing image edge and texture features and are more widely used. The mathematical expression for max pooling is:

[0114] X = f[αS(x) + b]

[0115] In the formula: X is the output, f(·) is the activation function; α is the downsampling coefficient, S(x) is the downsampling function; b is the bias term.

[0116] The function of an activation function is to activate a neuron when the stimulus it receives is greater than a threshold, causing the neuron to produce an output signal; otherwise, the neuron remains in an inhibited state. The Sigmoid function maps values ​​to the range (0,1), corresponding to a normalized lifetime. The function is continuous and smooth within its domain. The formula for the Sigmoid function is:

[0117]

[0118] Therefore, the health status assessment method for key components of rotating machinery based on image information regression proposed in this invention adopts the construction of a convolutional neural network to extract health factors. The output of the convolutional neural network will represent the regression calculation results of the mechanical equipment life.

[0119] The convolutional neural network structure consists of 3 two-dimensional convolutional layers, 3 max pooling layers, a Dropout layer, and a BatchNormalization layer. The Adam algorithm is selected as the optimizer. Except for the initial learning rate, which is set to 0.001, other parameters are set to default values. Mean squared error is selected as the error measure, and the mean absolute difference is used as the evaluation criterion for the network model. A BatchNormalization layer is used before each pooling layer for batch standardization, and the batch size is selected to be 16. The Dropout layer is used to prevent the model from overfitting, and the probability is set to 0.2. The Sigmoid activation function is selected, and the number of iterations is selected to be 30.

[0120] In this implementation, the convolutional neural network model is constructed using seven convolutional network segments. The first segment consists of one convolutional layer and one max-pooling layer, with 32 kernels of 3x3 size, a 3x3 max-pooling layer, a stride of 3, and no padding. The second segment consists of one convolutional layer and one max-pooling layer, with 64 kernels of 3x3 size, a 3x3 max-pooling layer, a stride of 3, and no padding. The third segment consists of one convolutional layer and one max-pooling layer, with 128 kernels of 3x3 size, a 2x2 max-pooling layer, a stride of 3, and no padding. The fourth segment is a flatten layer that produces a one-dimensional output vector. The fifth segment is a fully connected layer with an output size of 256. The sixth segment is a dropout layer with a probability of 0.2. The seventh segment is an output layer with an output size of 1. The convolutional neural network outputs a health factor, realizing a linear mapping relationship between image information and the health factor.

[0121] Step 7.2 evaluates the performance of the obtained health factors using monotonicity and trend as evaluation criteria. Monotonicity and trend are commonly used indicators for evaluating health factors; the larger the values, the better the performance of the health factor. Using these two evaluation criteria as constraints, the monotonicity and trend of the current health factors, or the performance indicators of the health factors, are obtained. If the performance indicators are low, the parameters of the convolutional neural network (learning rate, weight decay value, number of nodes in the fully connected layer) are fine-tuned, and iterative optimization is continuously performed to improve the performance indicators of the health factors. The fine-tuning method will employ the controlled variable method, selecting suitable parameters for the model through comparative experiments with different step sizes of the parameters to be adjusted.

[0122] The performance degradation process of critical components in rotating machinery is irreversible. Therefore, health factors reflecting the degree of performance degradation of critical components in rotating machinery should also have a monotonically increasing or monotonically decreasing characteristic. This characteristic is called monotonicity, and the formula for calculating the monotonicity of health factors is:

[0123]

[0124] In the formula, K represents the total number of health factors in the sequence; d / dx = X K+1 -X K The value represents the differential value of the health factor sequence; No.ofd / dx>0 and No.ofd / dx<0 represent positive and negative differential values, respectively. The monotonicity value Mon(X) of the health factor is located in the interval [0,1]. The closer its value is to 1, the better the monotonicity of the health factor.

[0125] The performance of core components of rotating machinery gradually degrades with increasing operating time. Therefore, the health factor values ​​characterizing the degradation process should be correlated with operating time. This correlation is called trend, and the formula for calculating the trend of the health factor is:

[0126]

[0127] In the formula, It represents the average health factors over the entire life cycle; T represents the operating cycle. It is the average value of each time period within the whole life cycle. The trend value of the health factor Tre(H,T) is located in the interval [0,1]. The closer its value is to 1, the better the trend of the health factor.

[0128] Step 7.3: Based on the collected data of different health stages of the core components of the rotating machinery, calibration is performed to obtain the health stage division on the health factor curve, so as to assess the health status of the key components of the rotating machinery according to the health factors.

[0129] Step 7.4: Collect vibration signals from the horizontal and vertical directions of the bearing to be evaluated in the unknown state, fuse them, and transform them into time-frequency images. After preprocessing the time-frequency images, analyze them to obtain first-order images. Input the first-order images into the established convolutional neural network model to obtain health factors, and then evaluate the health state of the bearing to be evaluated in the unknown state.

Claims

1. A method for evaluating the health state of a rotating machine based on image information regression analysis, characterized in that, It comprises the following steps: Step 1, collect the vibration signal of the core component of the rotating machinery to be evaluated, including collecting the vibration signal of the core component of the rotating machinery to be evaluated through an acceleration sensor in the whole life cycle, collecting the vibration signal in two mutually perpendicular directions, taking multi-channel synchronous collection, and corresponding vibration amplitude signals are respectively denoted as and ; Step 2, information fusion is performed on the time sequence constituted by the vibration signal to obtain a synthesized signal; Step 3, continuous wavelet transform is adopted to analyze the synthesized signal to obtain a time-frequency image; In step 3, first set A function defined in a local region, and tends to zero outside this region, and the mean value of the function is zero, called Mother wavelet, the function expression is: wherein denotes the center frequency, i denotes the complex number identifier in the complex trigonometric function, and t denotes the time; synthesis signal to be analyzed belongs to the space then the continuous wavelet transform of the synthesis signal is Where p is the scale, p>1 means the signal waveform shrinks, p<1 means the signal waveform stretches, and q is the displacement. For synthesized signals, , Yes According to the definition of the inner product in the complex frequency domain, its complex conjugate is taken. It is made from mother wavelet A series of wavelet basis functions are obtained by scaling and time shifting. As shown in the following formula: Transformed result The mother wavelet is selected with a base center frequency, then a series of different center frequencies are obtained by scale transformation, and a series of base functions with different intervals are obtained by time shift. Then, the extreme value corresponding to the frequency obtained by multiplying and integrating the product of the original signal in a certain interval is the frequency contained in the original signal in this interval. The relationship between time and frequency after transformation is obtained, which is plotted as an image, which is the time-frequency image of continuous wavelet transform. Step 4, the time-frequency image is preprocessed to obtain a compressed gray time-frequency image; Step 5, the compressed gray time-frequency image is analyzed from a spatial domain angle by adopting eigenvalue orthogonal decomposition to obtain a first-order image; Step 6, the health factor of the first-order image is calibrated according to the running time of the rotating machinery; The step 6 specifically comprises calibrating the health factor of the first-order image corresponding to the damage and failure of the core component of the rotating machinery as 0, calibrating the health factor of the first-order image corresponding to the brand-new core component of the rotating machinery as 1, and calibrating the health factor of the first-order image corresponding to different running moments in a linear relationship according to the following formula: In the formula, the time of using the detection device when the core part of the rotating machinery is damaged is The corresponding health factor = 0; the time of using when the core part of the rotating machinery is brand new is The corresponding health factor = 1; the health factor corresponding to the remaining state is ; Step 7, a corresponding model between the first-order image component and the health factor is established by using a convolutional neural network to realize the evaluation of the health state of the unknown state rotating machinery core component to be evaluated; The step 7 specifically comprises the following steps: Step 7.1, a convolutional neural network model is built to extract the health factor of the time-frequency image of the vibration signal of the rotating machinery core component; Step 7.2, the performance of the obtained health factor is evaluated, and the evaluation criteria are monotonicity and trend, and the performance index of the health factor is obtained, or the performance index of the health factor is low, if the performance index is low, the parameters of the convolutional neural network are fine-tuned, and the health factor performance index is continuously iteratively optimized to improve the health factor performance index; In step 7.2, the formula for calculating the monotonicity of the health factor is: K represents the number of all health factors in the sequence; Differential value of health factor sequence; And Positive differential value and negative differential value, respectively, monotonicity value of health factor Located in the interval [0, 1], the value is closer to 1, indicating that the monotonicity of the health factor is better; In step 7.2, the formula for calculating the trend of the health factor is: wherein, is the mean value of the health factor over the whole life cycle; T represents the running period, is the mean value of the health factor over the whole life cycle; T represents the running period, is located in the interval [0, 1], the value of which is closer to 1, the better the trend of the health factor; Step 7.3, the health stage division on the health factor curve is obtained by calibrating the data of the rotating machinery core component in different health stages; Step 7.4, the vibration signals of the unknown state rotating machinery core component to be evaluated in two directions are collected, fused, and transformed into time-frequency images, and then the first-order image is obtained, the first-order image is input into the established convolutional neural network model, the health factor is obtained, and then the health state of the unknown state rotating machinery core component to be evaluated is evaluated.

2. The method of claim 1, wherein, In the step 4, the time-frequency image is preprocessed, including image size compression and gray processing.

3. The method of claim 1, wherein, In the step 5, the compressed gray time-frequency image is analyzed from a spatial domain angle by adopting eigenvalue orthogonal decomposition, a time-frequency image is converted into multiple groups of modal components, the main feature structure in the image is found, and the first-order image is obtained.

Images