Two-dimensional image feature extraction method under adaptive polar coordinate mapping

Abstract

The invention discloses a two-dimensional image feature extraction method under adaptive polar coordinate mapping. The method comprises the following steps: step 1, intercepting a short-time sequence by using a short-time sliding window function; step 2, generating a time domain signal short-time matrix energy-time distribution map; step 3, positioning a key signal sequence representing fault characteristics through a peak screening algorithm; step 4, adaptively adjusting scale parameters of polar coordinate mapping, and constructing a multi-scale polar coordinate space; 5, generating a two-dimensional image under adaptive polar coordinate mapping; and step 6, eliminating the influence of rotation speed fluctuation on image features through normalization processing. And step 7, extracting spatial features by using a convolutional layer, and realizing adaptive diagnosis under a variable-speed working condition. According to the method, the fault features of the rotating equipment under the variable-speed working condition can be effectively extracted, and the features have robustness in a time-varying speed strong noise environment.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of fault diagnosis and signal processing, and particularly relates to a two-dimensional image feature extraction method under adaptive polar coordinate mapping. BACKGROUND

[0002] With the accelerated progress of the intelligent process of industrial equipment, as the core power device in the fields of energy development, transportation and high-end manufacturing, the state monitoring and fault diagnosis technology of rotating machinery has become the key support to ensure the reliability of equipment. However, the increasingly complex internal structure of rotating machinery significantly enhances the dynamic coupling effect between components, and the failure of a key component (such as a gear box, a bearing, etc.) may trigger a cascading failure reaction. According to statistical data, more than 30% of major equipment mechanical failures are caused by rotating parts, and the wind power field is particularly prominent - nearly half of the motor failures are caused by rolling bearing failure. In the industrial application scenario, the operation and maintenance challenge of offshore wind farms is extremely representative. Influenced by the special marine environment, the wind turbine generator set is subjected to multiple actions such as turbulent wind and wave load for a long time, and its operating state presents significant non-steady-state characteristics: the rotating speed presents wide fluctuations with the turbulent wind field, and the environmental noise level can reach 3-5 times that of the conventional working condition. Therefore, the weak impact characteristics of early-stage faults of rotating parts are easily completely submerged by the wideband background noise, and the nonlinear shift of fault feature frequency caused by the violent fluctuation of rotating speed makes the failure rate of traditional diagnosis models based on the assumption of fixed rotating speed as high as 60%. These technical problems have put forward innovative demands for modern fault detection methods.

[0003] Aiming at the problem of fault feature extraction under time-varying speed conditions, developing a feature extraction method with working condition self-adaptive ability has become a common demand in the field of intelligent diagnosis of rotating machinery. The existing technical system has the following technical problems: the traditional time-frequency analysis method is prone to spectrum blurring under dynamic speed fluctuation conditions, and its time-frequency resolution is subject to fixed window function parameters; the robustness is poor in strong noise environment. Taking a typical variable speed equipment, wind turbine generator, as an example, its variable speed range is 30%-120% of the rated speed, and the fault recognition accuracy of the traditional method under this working condition is seriously decreased, and there is a high false alarm rate. In recent years, although deep learning methods have made breakthroughs in state classification tasks through end-to-end training, there are still significant defects in the input layer: existing researches mostly directly use original signals without physical mechanism verification as input, which makes it difficult to trace the correlation between feature space and fault mode. This "black box" characteristic seriously restricts the engineering application of the diagnostic model in safety-sensitive scenarios. For example, Chinese patent CN202210351542.8 discloses a motor bearing fault feature extraction method, which uses Teager energy operator to realize noise reduction and feature enhancement of vibration signals. This method performs poorly under variable speed conditions and relies on experience-based parameter selection; for example, Chinese patent 202311278230.X discloses a variable speed condition rolling bearing fault feature extraction method, which constructs high-pass filter convolution coefficients and band-pass filters through Savitzky-Golay filter to realize variable speed rolling bearing fault feature extraction. However, this method needs to test and verify the filter parameters multiple times, and lacks real-time diagnosis capability. In order to solve these problems, the inventors provide a two-dimensional image feature extraction method under adaptive polar coordinate mapping. SUMMARY

[0004] The present application proposes a two-dimensional image feature extraction method under adaptive polar coordinate mapping to overcome the influence of noise interference and speed fluctuation during the operation of mechanical equipment by converting one-dimensional time domain signals into two-dimensional image features in a multi-scale polar coordinate space.

[0005] The core steps of the two-dimensional image feature extraction method under adaptive polar coordinate mapping are as follows:

[0006] Step one, segment and intercept the time domain signal using a short-time sliding window function to obtain a continuous short-time sequence;

[0007] Step two, calculate the energy distribution of each short-time sequence according to the root mean square value to generate a time-domain signal short-time matrix energy-time distribution graph;

[0008] Step three, extract the significant impact components in the energy-time distribution graph through a peak value screening algorithm, locate their corresponding time sequence positions, and obtain key signal segments representing fault characteristics.

[0009] Step four, according to the time domain characteristics of the key signal segment containing rotating machinery fault information, dynamically adjusting the scale parameter of polar coordinate mapping, taking signal amplitude as radial coordinate and time cumulative phase as angular coordinate, realizing multi-scale polar coordinate space construction;

[0010] Step five, mapping the key signal segment to polar coordinate space, determining the relative position relationship of time domain signal points in polar coordinates according to the signal characteristic mechanism and rotating equipment fault characteristics, and generating multi-scale two-dimensional image;

[0011] Step six, eliminating the influence of speed fluctuation on image features through normalization processing, and obtaining two-dimensional image features with constant speed;

[0012] Step seven, inputting the generated polar coordinate mapping feature image features into a deep learning model, extracting spatial features using convolutional layers, combining bidirectional long short-term memory networks to mine time sequence dependency, and realizing adaptive diagnosis under variable speed operating conditions.

[0013] The above-mentioned two-dimensional image feature extraction method under adaptive polar coordinate mapping has the characteristics that: in steps one to three, due to the complex operating conditions of rotating machinery, weak fault signals are easily submerged in strong noise environment, therefore, a sliding window function is introduced to slide and intercept the time domain signal into continuous short-time signals, and the rectangular window function w(k) is expressed as:

[0014]

[0015] Wherein, M is the width of the rectangular window function. By applying the root mean square value calculation method to these short-time signals, the short-time matrix energy-time distribution atlas of the time domain signal is obtained. On this basis, in order to obtain the fault information in the energy-time distribution atlas, a peak value point screening and identification method is proposed to realize the positioning of the peak value point in the time domain. The identification and extraction method of the peak component in the energy-time distribution atlas is as follows: first, calculate the mean value E(Ld) of the whole atlas as the basis for judging the peak value:

[0016]

[0017] In the formula, N Ld is the number of data points contained in the energy-time distribution atlas Ld. When an impact occurs, the signal energy increases, and a variable A(j) is proposed to judge whether the short-time matrix Ld(t j ) is a peak signal point. A(j) is defined as follows:

[0018] A(j)=Ld(t j )-E(Ld)

[0019] Threshold C l and Ch , Ld(t j ) is a necessary and sufficient condition for the peak signal point:

[0020]

[0021] By analyzing the characteristic distribution of the short-time signal corresponding to the time point of the spectral peak, a short-time signal matrix containing the most fault information is obtained, and the effectiveness of feature extraction is enhanced.

[0022] The adaptive polar coordinate mapping-based two-dimensional image feature extraction method has the characteristics that in step four, the polar coordinate mapping feature map parameters are determined according to the short-time matrix containing fault information, and the proposed method is described as follows:

[0023] For a time-domain signal X = [x1, x2,..., x n , any point x i in it needs to determine the radius component r(i) when mapped to polar coordinates, and the angle components φ(i) and Φ(i) can be determined by the following formula:

[0024]

[0025] In the formula: r(i) is the radius of x i mapped to polar coordinates; φ i (i) is the angle of the point in the jth polar coordinate dimension relative to the center line θ j counterclockwise rotation; Φ j (i) is the angle of the point in the jth polar coordinate dimension relative to the center line clockwise rotation; x max is the maximum amplitude of the signal X; x min is the minimum amplitude of the signal X; L j is the time scale parameter; θ j is the rotation angle of the center line of the jth polar coordinate dimension relative to the initial line, and N is the number of polar coordinate dimensions.

[0026] The center line rotation angle of each polar coordinate dimension is determined according to the number of polar coordinate dimensions:

[0027]

[0028] The adaptive polar coordinate mapping-based two-dimensional image feature extraction method has the characteristics that in step five, the time scale parameter L j in the polar coordinate mapping feature map determines the polar coordinate dimension position relationship, and different time scale parameters L j are selected at different polar coordinate dimensions to fully display the structural feature information in the original time-domain signal. Through analysis of the equipment structural characteristics and fault signals, the number of polar coordinate dimensions N = 5 is selected:

[0029]

[0030] wherein f s is the signal sampling frequency; f c is the rotating equipment inherent frequency; f b , f i , f o is the rotating equipment component failure through frequency. By selecting the above time scale parameters, the signal characteristics can be fully reflected. The selection scale of L1 and L2 is within one impact, and the impact characteristics of the signal itself are combined, which can fully reflect the impact response structure in the original time domain signal caused by the failure impact. The selection of L3, L4 and L5 combines the characteristic frequency of the rotating equipment component failure. When there is a certain failure, the interval generated by the impact is highly consistent with the time scale parameter, so that the signal point x(i) and the comparison point x(i±L j ) used to determine the deflection angle have the same characteristics, thereby showing the gathering characteristics at the polar coordinate dimension corresponding to the failure, so that the signals of different failures can be distinguished. By designing the time scale parameters, the sensitivity of the polar coordinate mapping feature map to different failure signal characteristics can be effectively improved.

[0031] The above adaptive polar coordinate mapping two-dimensional image feature extraction method is characterized in that: in step six, the polar coordinate mapping feature map dimension parameter is related to the rotating equipment failure through frequency, and is therefore affected by the rotating speed, but the value has a corresponding relationship with the rotating speed. For failure signals of different rotating speeds, L3, L4 and L5 also need to be changed and valued correspondingly:

[0032]

[0033] wherein L 3,n1 , L 4,n1 , L 5,n1 are time scale parameters designed for the time-varying rotating speed signal; f b,n1 , f i,n1 , f o,n1 are respectively the rotating equipment failure through frequency under the rotating speed n1. With the increase of the rotating speed of the equipment, the number of failure impacts per unit time will also increase. The time scale parameters L3, L4 and L5 reflect the periodic information of the impact generated in the signal. In order to make the corresponding polar coordinate scale characteristics consistent, the selection of these parameters should be corresponding to the rotating speed. The selection of the parameters under different rotating speeds should have the following relationship with the rotating speed:

[0034]

[0035] The advantages of the application are: (1) noise robustness: the fault impact component is enhanced by the time-domain signal short-time matrix energy-time distribution map, and the environmental noise interference is suppressed; (3) speed invariability: the speed information is fused in the polar coordinate mapping process, the scale parameter is adaptively adjusted, and the influence of speed fluctuation on image features is eliminated; (3) high computational efficiency: the sliding window segmentation processing and parallel mapping algorithm are adopted, and the computational complexity is significantly reduced; (4) diagnostic accuracy: the two-dimensional image features fuse time-space multi-dimensional information, and the deep learning model is combined to realize high-precision classification of faults.

[0036] In view of the difficulty in bearing fault diagnosis under variable speed caused by the great influence of time-domain and frequency-domain feature indexes on speed change, and the limitations of traditional time-frequency methods such as complex calculation, an adaptive polar coordinate mapping two-dimensional image feature extraction method is innovatively proposed, which realizes the conversion of time-domain signal from time domain to multi-scale polar coordinate angle domain, reconstructs one-dimensional signal into two-dimensional image, and obtains polar coordinate mapping features less affected by speed. In the construction process of the polar coordinate mapping feature map, the shape characteristics of each polar coordinate dimension are determined by the fault information of rotating equipment and the speed working condition, the adaptive selection of the parameter is realized, and the experience dependence is avoided. These advantages make the polar coordinate mapping feature map effective for variable speed fault diagnosis.

[0037] The technical solutions of the application will be further described in detail below with the help of the drawings and examples. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 It is a schematic diagram of the two-dimensional image feature extraction method under adaptive polar coordinate mapping.

[0039] Figure 2 It is a short signal sliding interception process.

[0040] Figure 3 It is a schematic diagram of screening and extracting the peak points of the energy-time sequence distribution map.

[0041] Figure 4 The bearing one-dimensional signal is converted into a polar coordinate mapping feature map.

[0042] Figure 5 The polar coordinate mapping feature maps of the outer ring fault of the rolling bearing under two speeds. DETAILED DESCRIPTION

[0043] In order to better understand the above technical solutions, the exemplary embodiments of the application will be described in detail below with the help of the drawings, which are only exemplary embodiments of the application, however, it should be understood that the application can also be realized in various forms and is not limited to the embodiments described herein. These embodiments are to enable those skilled in the art to more clearly and more thoroughly understand the application.

[0044] AsFigure 1 An adaptive polar coordinate mapping two-dimensional image feature extraction method is shown, and the specific steps are as follows:

[0045] Step one, the short-time sliding window function is used to segment and intercept the time domain signal, and the continuous short-time sequence is obtained;

[0046] Step two, the energy distribution of each short-time sequence is calculated according to the root mean square value, and the time domain signal short-time matrix energy-time distribution graph is generated;

[0047] Step three, the significant impact component in the energy-time distribution graph is extracted by the peak value screening algorithm, and the corresponding time sequence position is located to obtain the key signal segment representing the fault characteristics.

[0048] Step four, according to the time domain characteristics of the key signal segment containing the rotating machinery fault information, the scale parameter of the polar coordinate mapping is dynamically adjusted, and the multi-scale polar coordinate space is realized with the signal amplitude as the radial coordinate and the time cumulative phase as the angular coordinate;

[0049] Step five, the key signal segment is mapped to the polar coordinate space, and the relative position relationship between the time domain signal points in the polar coordinate is determined according to the signal characteristic mechanism and the rotating equipment fault characteristics, and a multi-scale two-dimensional image is generated;

[0050] Step six, the influence of speed fluctuation on image features is eliminated by normalization processing, and the two-dimensional image features with constant speed are obtained.

[0051] Step seven, the generated polar coordinate mapping feature graph features are input into the deep learning model, the spatial features are extracted by convolution layer, and the time sequence dependence relationship is mined by combining bidirectional long short-term memory network, and adaptive diagnosis under variable speed conditions is realized.

[0052] In specific implementation, in step one, the short-time time domain signal sequence will be used to obtain the energy-time distribution graph and the extraction of fault characteristics in the signal, and the sliding window width has robustness and can be adaptively adjusted according to the sampling frequency. The sliding interception process is shown in Figure 2 .

[0053] In specific implementation, in steps two to three, when the rotating equipment fails, the collision impact will occur in the fault area of the component. This impact will cause the peak value of the time domain signal, and this part of the short-time time domain signal contains more fault information. Therefore, as Figure 3 shown, the peak component of the energy-time distribution graph is extracted to track the fault information, and the short-time sequence corresponding to the collision process is analyzed to obtain the signal features reflecting the fault information. The identification and extraction method of the peak component in the energy-time distribution graph is as follows: first, calculate the mean value E(Ld) of the whole graph as the basis for judging the peak value:

[0054]

[0055] where N Ld is the number of data points contained in the energy-time profile Ld. When a shock occurs, the signal energy increases, and a variable A(j) is introduced to determine whether the short-time matrix Ld(t j ) is a peak signal point. A(j) is defined by the following equation:

[0056] A(j) = Ld(t j ) - E(Ld)

[0057] Thresholds C l and C h are introduced, and the sufficient and necessary condition for Ld(t j ) to be a peak signal point is:

[0058]

[0059] By analyzing the characteristic distribution of the short-time signal at the time corresponding to the peak point of the profile, the short-time signal matrix containing the most fault information is obtained, and the effectiveness of feature extraction is enhanced.

[0060] In the specific implementation, in steps four to five, the polar coordinate mapping feature map parameters are determined according to the short-time matrix containing fault information described above, and the proposed method is described as follows:

[0061] For a time-domain signal X = [x1, x2,..., x n ], the radius component r(i) to be determined when mapping any point x i to polar coordinates, the angle component φ(i), and Φ(i) can be determined by the following equation:

[0062]

[0063] where: r(i) is the radius of x i mapping to polar coordinates; φ i (i) is the angle of the point in the jth polar coordinate dimension relative to the center line θ j counterclockwise rotation; Φ j (i) is the angle of the point in the jth polar coordinate dimension relative to the center line clockwise rotation; x max is the maximum amplitude of the signal X; x min is the minimum amplitude of the signal X; L j is the time scale parameter; θ j is the rotation angle of the center line of the jth polar coordinate dimension relative to the initial line, and N is the number of polar coordinate dimensions.

[0064] The center line rotation angle of each polar coordinate dimension is determined according to the number of polar coordinate dimensions:

[0065]

[0066] At the same time, the time scale parameter L in the polar coordinate mapping feature map j determines the polar coordinate dimension position relationship, and different time scale parameters L are selected at different polar coordinate dimensions j , which can fully display the structural characteristic information in the original time domain signal. Through analysis of the structural characteristics and fault signals of equipment, the bearing is taken as an example, and the key components are rolling body, bearing outer ring and bearing inner ring, therefore, the polar coordinate dimension number N is selected as 5:

[0067]

[0068]

[0069] In the formula, f s is the signal sampling frequency; f c is the inherent frequency of rotating equipment; f b , f i , f o The distribution is the fault pass frequency of the rolling body, outer ring and inner ring, as shown in Figure 4 .

[0070] In the implementation, the polar coordinate mapping feature map dimension parameters described above in step six are related to the rotating equipment fault pass frequency, and therefore are affected by the rotating speed, but have a corresponding relationship with the rotating speed. For fault signals of different rotating speeds, L3, L4 and L5 also change and take values correspondingly:

[0071]

[0072] Among them, L 3,n1 , L 4,n1 , L 5,n1 are time scale parameters designed for time-varying rotating speed signals; f b,n1 , f i,n1 , f o,n1 are the rotating equipment fault pass frequencies under the rotating speed n1. With the increase of the rotating speed of the equipment, the number of fault impacts per unit time will also increase, and the time scale parameters L3, L4 and L5 reflect the periodic information of the impact in the signal. In order to make the corresponding polar coordinate scale characteristics consistent, the selection of these parameters should be corresponding to the rotating speed, according to two rotating speeds n1 and n2, as shown in Figure 5 is the polar coordinate mapping feature map of the outer ring fault of the rolling bearing under two rotating speeds.

[0073] In the step seven, the polar coordinate mapping feature map is taken as the input of the diagnosis model by constructing the convolutional neural network architecture. Based on the shape difference of the polar coordinate mapping feature map of different fault signals, the global image features in the polar coordinate mapping feature map are obtained through the forward propagation and the back propagation of the convolutional neural network. The probability of each sample belonging to a certain fault category is obtained through the classifier, and finally the diagnosis result is obtained in the output layer.

[0074] The above is only the preferred embodiment of the present application, not any limitation on the present application, any simple modification, change and equivalent structure change of the above embodiment according to the technical essence of the present application are still within the protection scope of the technical solution of the present application.

Claims

1. A two-dimensional image feature extraction method based on adaptive polar coordinate mapping, the core steps of the present application are as follows: Step one, adopt short-time sliding window function to segment and intercept time domain signal, obtain continuous short-time sequence; Step two, calculate the energy distribution of each short-time sequence according to the root mean square value, generate time domain signal short-time matrix energy-time distribution graph; Step three, extract the significant impact component in the energy-time distribution graph by peak value screening algorithm, locate its corresponding time sequence position, and obtain the key signal segment representing fault characteristics; Step four, according to the time domain characteristics of the key signal segment containing rotating machinery fault information, dynamically adjust the scale parameter of polar coordinate mapping, take signal amplitude as radial coordinate and time cumulative phase as angular coordinate, realize multi-scale polar coordinate space construction; Step five, map the key signal segment to polar coordinate space, determine the relative position relationship of time domain signal points in polar coordinates according to the signal characteristic mechanism and rotating equipment fault characteristics, and generate multi-scale two-dimensional image; Step six, eliminate the influence of speed fluctuation on image features through normalization processing, and obtain two-dimensional image features with constant speed; Step seven, input the generated polar coordinate mapping feature graph features into a deep learning model, extract spatial features using convolutional layers, combine bidirectional long short-term memory network to mine time sequence dependency, and realize adaptive diagnosis under variable speed conditions.

2. The method according to claim 1, wherein the method is characterized by: In steps one to three, the rectangular window function of sliding interception is introduced, and the time domain signal is intercepted into continuous short-time signal, the rectangular window function w(k) is expressed as: Wherein, M is the width of rectangular window function, k is the variable of rectangular window function, else is other; By applying the root mean square value calculation method to these short-time signals, the time domain signal short-time matrix energy-time distribution graph is obtained.

3. The method according to claim 2, wherein the method is a self-adaptive polar mapping down 2D image feature extraction method. The identification and extraction method of peak components in the energy-time distribution graph is as follows: First, calculate the mean value E(Ld) of the whole graph as the basis for judging the peak value: where N Ld is the number of data points contained in the energy-time profile Ld. When an impact occurs, the signal energy increases, and the variable A(j) is introduced to determine whether the short-time matrix Ld(t j ) is a peak signal point or not. A(j) is defined by the following formula: A(j) = Ld(t j )- E(Ld), Introducing threshold C l and C h , Ld(t j ) is a necessary and sufficient condition for the peak signal point By analyzing the characteristic distribution of the short-time signal corresponding to the time of the peak value point, the short-time signal matrix containing the most fault information is obtained, and the effectiveness of feature extraction is enhanced.

4. The method of claim 1, wherein the method is a self-adaptive polar mapping down 2D feature extraction method. In step four, the polar coordinate mapping feature graph parameters are determined according to the above short-time matrix containing fault information, including the following processes: For a time-domain signal X = [x1, x2, ..., x...] n ], take any point x i The radius component r(i), and the angle components φ(i) and Φ(i) that need to be determined when mapping to polar coordinates can be determined by the following formula: where: r(i) is x i radius mapped to polar coordinates; φ i (i) is the angle of the point in the jth polar dimension relative to the center line clockwise; x j angle of counterclockwise rotation; Φ j (i) is the angle of the point in the jth polar dimension relative to the center line counterclockwise; x max is the maximum amplitude of the signal X; x min is the minimum amplitude of the signal X; L j is the time scale parameter; θ j is the rotation angle of the centerline of the jth polar coordinate dimension relative to the initial line, and N is the number of polar coordinate dimensions. The center line rotation angle of each polar coordinate dimension is determined according to the number of polar coordinate dimensions:

5. The method of claim 1, wherein the method is a self-adaptive polar mapping down 2D feature extraction method. In step five, the time scale parameter L in the polar coordinate mapping feature map j The polar coordinate dimension position relationship is determined, and different time scale parameters L are selected at different polar coordinate dimensions j Through analysis of the equipment structure characteristics and the fault signal, the polar coordinate dimension number N=5 is selected where f s is the signal sampling frequency; f c is the rotating equipment natural frequency; f b , f i , f o is the distribution of the rotating equipment component failure through frequencies.

6. The method of claim 1, wherein the method is a self-adaptive polar mapping down 2D feature extraction method. In step six, the above polar coordinate mapping feature graph dimension parameters and rotating equipment fault pass frequency are related, so they will be affected by the speed, but their values have corresponding relationship with the speed. For fault signals of different speeds, L3, L4 and L5 also change value correspondingly: wherein L 3,n1 , L 4,n1 , L 5,n1 is a time scale parameter designed for the time-varying rotational speed signal; f b,n1 , f i,n1 , f o,n1 is the rotating equipment fault pass frequency under the rotational speed n1, respectively; n represents different rotational speed cases, n1 represents the rotational speed case 1; D is the bearing pitch circle diameter; d is the bearing rolling body diameter; a is the bearing contact angle; and Z is the number of bearing internal rolling bodies. The selection of parameters under different speeds should have the following relationship with the speed:

Images