Logarithmic-spiral bearing roller defect detection method and system based on ultrasonic technology

Abstract

The application provides a logarithmic spiral bearing roller defect detection method and system based on ultrasonic technology, and the method comprises the following steps: acquiring an ultrasonic detection signal of a logarithmic spiral bearing roller surface and performing pretreatment to obtain a first echo signal sequence; based on the first echo signal sequence, an average value sequence, a skewness sequence, a kurtosis sequence and a first high-order cumulative quantity sequence are obtained; the first high-order cumulative quantity sequence is corrected based on the average value sequence to obtain a second high-order cumulative quantity sequence; the first echo signal sequence is optimized based on a dictionary matrix and a signal sparse coefficient vector to obtain a second echo signal sequence; a multi-dimensional feature vector sequence is obtained by fusing the sequences, and the multi-dimensional feature vector sequence is compared with a preset vector threshold to detect whether the logarithmic spiral bearing roller has defects. The application solves the beam distortion problem of traditional ultrasonic detection of logarithmic spiral bearing roller defects, and significantly improves the detection accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intelligent quality detection, in particular to a logarithmic line bearing roller defect detection method and system based on ultrasonic technology. BACKGROUND

[0002] The bearing roller is a core component of mechanical equipment that bears load, and its surface defects (such as cracks and peeling) directly affect the reliability and life of equipment operation. The outer surface of the logarithmic line bearing roller is not a simple straight line or a symmetric circular arc convexity, but a complex profile designed with a logarithmic curve. From the center of the roller to both ends, the radius of curvature changes continuously, forming a smooth transition (similar to "micro-convex" but asymmetric). In this way, the contact stress distribution of the logarithmic line bearing roller is more uniform, improving the load-carrying capacity and life, and eliminating the edge stress concentration of traditional rollers, adapting to partial load or shaft bending working conditions. It is commonly used in heavy load, impact load or easily deformed working conditions, such as wind turbine main shaft, rolling mill bearings and mining machinery, etc. Logarithmic line bearing roller defect detection is crucial because its complex profile requires extremely high machining precision, and minor defects (such as cracks, wear or shape deviation) can disrupt stress distribution, leading to early bearing failure. Timely detection can avoid equipment downtime, reduce maintenance costs, and ensure safe operation under heavy load and high-speed conditions, prolong bearing life, and ensure the reliability of critical equipment.

[0003] However, traditional ultrasonic detection methods cannot meet the high-precision requirements of logarithmic line bearing roller defect detection, mainly due to the following reasons: (1) geometric interference. The complex surface of the logarithmic line bearing roller causes the ultrasonic beam propagation path to deviate and the reflection angle to deviate, and the traditional method cannot effectively compensate for the surface effect, so the defect signal is easily masked by noise. (2) Insufficient feature sensitivity. Traditional signal processing relies on low-order statistics such as mean and variance, making it difficult to capture non-linear and non-Gaussian features caused by defects, resulting in high false negative rate.

[0004] Therefore, there is an urgent need for a solution to high-precision detection of defects in complex logarithmic line bearing rollers. SUMMARY

[0005] The present application provides a logarithmic line bearing roller defect detection method and system based on ultrasonic technology to solve the problems in the prior art.

[0006] To solve the above technical problems, the present application solves the problems by the following technical solutions:

[0007] A logarithmic line bearing roller defect detection method based on ultrasonic technology, comprising the following steps:

[0008] Obtain the ultrasonic detection signal of the logarithmic line bearing roller surface based on time series and preprocess it to obtain the first echo signal sequence;

[0009] The average value sequence, the skewness sequence, the kurtosis sequence, and the first high-order cumulant sequence are obtained based on the average value, the skewness, the kurtosis, and the high-order cumulant of the first echo signal sequence based on a preset sliding window length;

[0010] The second high-order cumulant sequence is obtained by correcting the first high-order cumulant sequence based on the average value sequence;

[0011] The second echo signal sequence is obtained by optimizing the first echo signal sequence based on the dictionary matrix and the signal sparse coefficient vector;

[0012] The multi-dimensional feature vector sequence is obtained by fusing the skewness sequence, the kurtosis sequence, the second high-order cumulant sequence, and the second echo signal sequence;

[0013] The multi-dimensional feature vector sequence is compared with a preset vector threshold to detect whether the log cylindrical bearing roller has defects.

[0014] As an implementable manner, the ultrasonic detection signal is a phased array ultrasonic detection signal, and the ultrasonic detection signal on the surface of the log cylindrical bearing roller is obtained based on a time sequence and preprocessed to obtain the first echo signal sequence, including the following steps:

[0015] The ultrasonic detection signal on the surface of the log cylindrical bearing roller is obtained based on a time sequence to obtain an initial echo signal sequence represented in a polar coordinate system;

[0016] The initial echo signal sequence is converted from the polar coordinate system to the Cartesian coordinate system to obtain the first echo signal sequence.

[0017] As an implementable manner, the first high-order cumulant sequence includes a first third-order cumulant sequence and a first fourth-order cumulant sequence, and the second high-order cumulant sequence is obtained by correcting the first high-order cumulant sequence based on the average value sequence, including the following steps:

[0018] The ultrasonic echo signal predicted value sequence is obtained based on the first third-order cumulant sequence, the first fourth-order cumulant sequence, and the average value sequence in combination with a signal prediction model; wherein the signal prediction model is that the first third-order cumulant and the first fourth-order cumulant are weighted and summed, and summed with the average value of the ultrasonic echo signal to obtain the ultrasonic echo signal predicted value;

[0019] The signal prediction error sequence is obtained by performing difference processing on the first echo signal sequence and the ultrasonic echo signal predicted value sequence;

[0020] The correction factor of the previous time step is adjusted based on a signal prediction error of the current time step, the first third-order cumulant of the current time step, and the first fourth-order cumulant of the current time step, to obtain a correction factor of the current time step, and then a correction factor sequence is obtained;

[0021] The first third-order cumulant sequence and the first fourth-order cumulant sequence are respectively corrected based on the correction factor sequence, to obtain a second third-order cumulant sequence and a second fourth-order cumulant sequence.

[0022] As an implementable manner, the correction factor of the current time step is represented as follows:

[0023]

[0024] wherein, the correction factor of the current time step , the correction factor of the previous time step , the signal prediction error of the current time step , the first third-order cumulant of the current time step , the first fourth-order cumulant of the current time step , and the learning rate

[0025] As an implementable manner, the first high-order cumulant sequence is the first third-order cumulant sequence or the first fourth-order cumulant sequence, and the first high-order cumulant sequence is corrected based on the average value sequence to obtain a second high-order cumulant sequence, including the following steps:

[0026] When the first high-order cumulant sequence is the first third-order cumulant sequence:

[0027] The first ultrasonic echo signal prediction value sequence is obtained based on the first third-order cumulant sequence and the average value sequence in combination with a third-order signal prediction model, wherein the third-order signal prediction model is that the first third-order cumulant is adjusted and summed with an ultrasonic echo signal average value to obtain a first ultrasonic echo signal prediction value;

[0028] The first signal prediction error sequence is obtained by performing difference processing on the first echo signal sequence and the first ultrasonic echo signal prediction value sequence;

[0029] The third-order correction factor of the previous time step is adjusted based on the first signal prediction error of the current time step and the first third-order cumulant of the current time step, to obtain a third-order correction factor of the current time step, and then a third-order correction factor sequence is obtained; ​​​​​

[0030] The first third-order cumulant sequence is modified based on a third-order correction factor sequence to obtain a second third-order cumulant sequence.

[0031] When the first high-order cumulant sequence is a first fourth-order cumulant sequence, the first fourth-order cumulant sequence is modified by the same method to obtain a second fourth-order cumulant sequence.

[0032] As an implementable manner, the first echo signal sequence is optimized based on a dictionary matrix and a signal sparse coefficient vector to obtain a second echo signal sequence, including the following steps:

[0033] The first echo signal sequence is divided into signal segment sequences based on a preset segment step and a preset overlap step;

[0034] The signal segment sequences are predicted by the dictionary matrix and the signal sparse coefficient vector, and the optimal dictionary matrix and the optimal signal sparse coefficient vector of each segment are obtained based on the error minimization principle;

[0035] Each signal segment sequence of each segment is reconstructed based on the optimal dictionary matrix and the optimal signal sparse coefficient vector of each segment to achieve denoising, to obtain a denoised signal segment sequence;

[0036] The denoised signal segment sequence is restored by overlap position weighted addition to obtain a denoised echo signal sequence, which is the second echo signal sequence.

[0037] As an implementable manner, the comparison of the multi-dimensional feature vector sequence with the preset vector threshold is used to detect whether the log cylindrical bearing roller has defects, including the following steps:

[0038] The multi-dimensional feature vector sequence is standardized to obtain a standard multi-dimensional feature vector sequence;

[0039] The degree of vector deviation from Gaussian distribution in the standard multi-dimensional feature vector sequence is calculated based on Mahalanobis distance to obtain a vector deviation value, which is the Mahalanobis distance of the multi-dimensional feature vector, and further to obtain a Mahalanobis distance sequence;

[0040] If there is a Mahalanobis distance greater than a preset Mahalanobis distance threshold in the Mahalanobis distance sequence, it is determined that the log cylindrical bearing roller has defects.

[0041] As an implementable manner, the log cylindrical bearing roller defect detection method based on ultrasonic technology further includes: constructing a defect detection pre-training model, training the defect detection pre-training model based on historical multi-dimensional feature vector sequences and log cylindrical bearing roller defect detection results to obtain a defect detection model, including the following steps:

[0042] Obtain historical ultrasonic detection signals of the surface of the logarithmic generator bearing roller to obtain corresponding historical multi-dimensional feature vector sequences and historical vector defect detection result value sequences, and then form a training sample data set;

[0043] Construct a defect detection pre-training model, the defect detection pre-training model comprising an optimization target model, a constraint condition model, and a classification decision model; specifically, based on a weight vector, an optimization target model is constructed; based on a weight vector, a bias term, a multi-dimensional feature vector, and a defect detection value, a constraint condition model is constructed;

[0044] Based on the optimization target model and the constraint condition model, a Lagrange model is constructed in combination with a Lagrange multiplier;

[0045] The partial derivative of the Lagrange model is taken and set to zero to obtain a partial derivative model;

[0046] Based on the training sample data set, the partial derivative model is solved to obtain an optimal weight vector and an optimal bias term;

[0047] Based on the optimal weight vector and the optimal bias term, a classification decision model is constructed;

[0048] Based on the classification decision model, the multi-dimensional feature vector is inferred to obtain a defect detection result prediction value, i.e., a defect detection model.

[0049] An ultrasonic technology-based logarithmic generator bearing roller defect detection system for implementing the method described in any of the above embodiments, the detection system comprising a signal acquisition and processing module, an initial feature value acquisition module, a high-order feature value correction module, a signal correction module, a fusion module, and a defect detection module;

[0050] The signal acquisition and processing module is configured to acquire ultrasonic detection signals of the surface of the logarithmic generator bearing roller based on a time sequence and perform preprocessing to obtain a first echo signal sequence;

[0051] The initial feature value acquisition module is configured to calculate the average value, skewness, kurtosis, and high-order cumulant of the first echo signal sequence based on a preset sliding window length to obtain an average value sequence, a skewness sequence, a kurtosis sequence, and a first high-order cumulant sequence;

[0052] The high-order feature value correction module is configured to correct the first high-order cumulant sequence based on the average value sequence to obtain a second high-order cumulant sequence;

[0053] The signal correction module is configured to optimize the first echo signal sequence based on a dictionary matrix and a signal sparse coefficient vector to obtain a second echo signal sequence;

[0054] The fusion module is configured to fuse the skewness sequence, the kurtosis sequence, the second high-order cumulative quantity sequence, and the second echo signal sequence to obtain a multi-dimensional feature vector sequence.

[0055] The defect detection module is configured to compare the multi-dimensional feature vector sequence with a preset vector threshold to detect whether the logarithmic cylindrical bearing roller has defects.

[0056] A computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the method described in any one of the preceding embodiments.

[0057] An apparatus includes a memory, a processor, and a computer program stored in the memory and running on the processor, and the processor implements the method described in any one of the preceding embodiments when executing the computer program.

[0058] The present application has the following technical effects: the present application discloses a logarithmic cylindrical bearing roller defect detection method and system based on ultrasonic technology. First, the surface of the logarithmic cylindrical bearing roller is detected by ultrasonic waves to obtain an initial echo signal sequence, and the initial echo signal sequence is converted from a polar coordinate system to a Cartesian coordinate system to eliminate the interference of the curved surface effect on the ultrasonic waves. Then, the skewness, kurtosis, and high-order cumulative quantity of the echo signal are extracted, and a sliding window mechanism is introduced to dynamically update the statistical characteristics. Further, noise is removed by sparse representation and online dictionary learning, and defect features are retained. Finally, whether the surface of the logarithmic cylindrical bearing roller has defects is determined based on the deviation value. The present application effectively solves the beam distortion problem of traditional ultrasonic waves in detecting defects of the logarithmic cylindrical bearing roller with a complex curved surface, accurately captures the defect signal, significantly reduces the missed detection rate, and improves the detection accuracy and industrial applicability. BRIEF DESCRIPTION OF DRAWINGS

[0059] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0060] Figure 1 The flowchart of the method of the present application;

[0061] Figure 2 The overall schematic diagram of the system of the present application. DETAILED DESCRIPTION

[0062] The present invention will be further described in detail below with reference to embodiments. These embodiments are illustrative of the invention and the invention is not limited thereto. Unless otherwise specified, the features in the following embodiments can be combined with each other.

[0063] Example 1:

[0064] In one embodiment, a method for detecting roller defects in logarithmic busbar bearings based on ultrasonic technology, such as... Figure 1 As shown, it includes the following steps:

[0065] S100: The ultrasonic detection signal of the logarithmic busbar bearing roller surface is obtained based on the time series and preprocessed to obtain the first echo signal sequence;

[0066] S200: Calculate the average value, skewness, kurtosis, and higher-order cumulants of the first echo signal sequence based on a preset sliding window length to obtain the average value sequence, skewness sequence, kurtosis sequence, and first higher-order cumulants sequence;

[0067] S300: The first higher-order cumulant sequence is corrected based on the average value sequence to obtain the second higher-order cumulant sequence;

[0068] S400: The first echo signal sequence is optimized based on the dictionary matrix and the sparse coefficient vector of the signal to obtain the second echo signal sequence;

[0069] S500: The skewness sequence, kurtosis sequence, second higher-order cumulant sequence, and second echo signal sequence are fused to obtain a multidimensional feature vector sequence;

[0070] S600: Compare the multidimensional feature vector sequence with a preset vector threshold to detect whether there are defects in the logarithmic busbar bearing rollers.

[0071] In another embodiment, in S100, the ultrasonic detection signal of the logarithmic busbar bearing roller surface is acquired based on the time series and preprocessed to obtain the first echo signal sequence, including the following steps:

[0072] S110: The surface of the logarithmic generatrix bearing rollers is inspected using ultrasonic testing equipment. The ultrasonic testing signals from the surface of the logarithmic generatrix bearing rollers are acquired based on a time series, resulting in an initial echo signal sequence represented in polar coordinates. If the ultrasonic testing equipment is a phased array ultrasonic testing equipment, then the ultrasonic testing signal is a phased array ultrasonic testing signal. The logarithmic generatrix curve model represented in polar coordinates is as follows: ,in, Represents polar coordinate angle (radians). Represents angles in polar coordinates The corresponding radial distance, specifically, represents the radial distance of the logarithmic spiral in polar coordinates. represents the initial radius, i.e. the radial distance when , in mm. represents the curve shape control coefficient, which is a dimensionless number, determines the expansion rate of the curve, when the curve expands outward, when the curve shrinks inward.

[0073] The initial echo signal sequence is represented in the polar coordinate system, but the ultrasonic probe of the ultrasonic detection equipment, especially the phased array ultrasonic probe, is usually arranged in a linear or rectangular array, and its position and beam control need to be described in the Cartesian coordinate system (such as probe spacing, beam deflection angle). The surface shape of the logarithmic spiral bearing roller is represented in the Cartesian coordinate system, which can be aligned with the geometric parameters (such as transmission / reception position) of the probe array, so as to facilitate the calculation of the beam incidence point and the reflection path. Moreover, when a defect is detected, the result needs to be reported on the actual physical position of the bearing roller (usually marked in Cartesian coordinates), and the Cartesian coordinates can be directly output after conversion, which is convenient for engineers to locate. Therefore, it is necessary to convert the initial echo signal sequence represented in the polar coordinate system to the echo signal sequence represented in the Cartesian coordinate system.

[0074] S120: converting the initial echo signal sequence from the polar coordinate system to the Cartesian coordinate system to obtain a first echo signal sequence, specifically:

[0075] (1) converting the logarithmic spiral curve model represented in the polar coordinate system to the Cartesian coordinate system to obtain a Cartesian logarithmic spiral curve model, and the Cartesian logarithmic spiral curve model can be represented as: , wherein represents the transverse coordinate value in the Cartesian coordinate system, represents the longitudinal coordinate value in the Cartesian coordinate system.

[0076] (2) based on the Cartesian logarithmic spiral curve model, calculating the tangent vector of the Cartesian logarithmic spiral curve (i.e. the surface of the logarithmic spiral bearing roller), to obtain the Cartesian tangent vector , the tangent vector represents the tangent direction of a point on the curve, which is obtained by derivation of and , and the derivation of and is respectively represented as: , .

[0077] , the Cartesian tangent vector is represented as: .

[0078] A normal vector of the Cartesian tangent vector is obtained, and a Cartesian normal vector is obtained by rotating the tangent vector by 90 degrees. .

[0079] The normal vector is perpendicular to the tangent vector, and the normal vector is obtained by rotating the tangent vector by 90 degrees, and the Cartesian normal vector is expressed as: .

[0080] The Cartesian normal vector is normalized to obtain a Cartesian unit normal vector The Cartesian unit normal vector is used to correct the incident and reflected directions of the ultrasonic wave, and the Cartesian unit normal vector is expressed as: .

[0081] (3) In a curved medium, the propagation speed of the ultrasonic wave will change due to the change in curvature, that is, the curvature effect causes the change in the propagation speed of the ultrasonic wave, so it is necessary to convert the ultrasonic wave propagation speed in the polar coordinate system to the Cartesian coordinate system to obtain the corrected propagation speed. The corrected propagation speed can be expressed as: wherein, represents the corrected effective propagation speed of the ultrasonic wave, and represents the effective propagation speed related to the angle , which is used to compensate the influence of the curvature effect on the propagation of the ultrasonic wave. The corrected effective propagation speed can more accurately reflect the propagation behavior of the ultrasonic wave in the curved medium, thereby reducing the calculation error of the propagation time caused by the curvature. represents the theoretical propagation speed of the ultrasonic wave in a planar medium (unit: ). represents the reference curvature radius, and the unit is mm, It can be obtained by measuring a standard part without defects, for example.

[0082] (4) Based on the Cartesian unit normal vector and the corrected propagation speed, the ultrasonic wave propagation path in the polar coordinate system is converted to the Cartesian coordinate system to obtain the corrected propagation path. The corrected propagation path can be expressed as: wherein, represents the corrected propagation path of the ultrasonic wave, which is the ultrasonic wave path considering the curvature and the normal vector, and corrects the beam deflection problem caused by the geometric shape of the curved surface, so that the ultrasonic wave can be more accurately focused on the target area. represents the round-trip time of the ultrasonic wave, which is the time difference between the emission and reception of the ultrasonic wave.

[0083] (5) Based on the Cartesian unit normal vector, an ultrasonic beam direction control signal is obtained, and the ultrasonic beam direction control signal can be expressed as: wherein, represents an ultrasonic beam direction control signal, used to adjust the beam forming of the phased array probe, to ensure that the ultrasonic waves propagate along the corrected path. represents the total number of phased array probes, represents the excitation weight of the th probe, represents the imaginary unit, represents the phase delay of the th probe, and the phase delay is used to control the beam direction.

[0084] (6) Based on the corrected propagation path and the ultrasonic beam direction control signal, the signals in the initial echo signal sequence are adjusted to obtain the adjusted echo signal sequence, which is the first echo signal sequence. Specifically, based on the corrected propagation path and the ultrasonic beam direction control signal, the signals in the initial echo signal sequence are processed in time domain (such as phase adjustment and superposition) to generate the adjusted echo signal sequence, which is the first echo signal sequence.

[0085] In another embodiment, in S200, the first echo signal sequence is calculated based on the skewness, kurtosis, high-order cumulant and average value of the preset sliding window length to obtain the skewness sequence, kurtosis sequence, first high-order cumulant sequence and average value sequence, wherein the high-order cumulant at least includes one of third-order cumulant and fourth-order cumulant, and the following steps are included:

[0086] S210: Based on the preset sliding window length , the first echo signal sequence is divided into a plurality of segment sequences to obtain a window segment sequence. The preset sliding window length is used to adjust the statistical characteristic value in real time to adapt to the non-stationary nature of the signal, such as noise fluctuation or defect mutation. The default value of the preset sliding window length can be set to 50, and the overlapping length of the segment sequences of adjacent windows is (49).

[0087] S220: The average value, skewness, kurtosis and high-order cumulant of each segment sequence in the window segment sequence are calculated respectively to obtain the average value sequence, skewness sequence, kurtosis sequence and first high-order cumulant sequence.

[0088] (1) The average value sequence is a sequence composed of the average values of the signal values of each segment sequence in the window segment sequence, and the average value sequence is represented as: , wherein, represents the preset sliding window length, represents the ultrasonic echo signal in the first echo signal sequence at time step , and represents the average value of the first ultrasonic echo signal within the preset sliding window length at time step , that is, the average value of the window segment sequence at time step the average of the segment sequence of time step the average of the segment sequence of time step and the average of the signal values of the previous time steps.

[0089] (2) Skewness is used to measure the asymmetry of the signal distribution, and the skewness sequence is represented as: wherein, denotes the skewness of time step in the skewness sequence. denotes the standard deviation of the first ultrasonic echo signal within the preset sliding window length of time step , that is, the standard deviation of the segment sequence of time step in the window segment sequence, . Skewness denotes that the signal is right-skewed (long tail on the right side), and skewness denotes left-skewed (long tail on the left side). Defect signals usually cause skewness to deviate significantly from 0, for example, cracks cause reflection signals to be asymmetric.

[0090] (3) Kurtosis is used to measure the sharpness of the signal distribution (compared with normal distribution), and the kurtosis sequence is represented as: wherein, denotes the kurtosis of time step in the kurtosis sequence, and kurtosis denotes that the signal is sharper (higher peak value) than the normal distribution, and kurtosis denotes that it is flatter. Defect signals (such as pores) cause local signal amplitude to change abruptly, thereby causing kurtosis to increase significantly.

[0091] (4) Third-order cumulant is used to capture the non-Gaussian and nonlinear phase coupling characteristics of the signal, and the third-order cumulant is sensitive to asymmetric defects (such as inclined cracks). Fourth-order cumulant is used to further reveal the high-order nonlinear dependence relationship of the signal, and can detect complex defects. If the high-order cumulant includes at least one of the third-order cumulant and the fourth-order cumulant, then the first high-order cumulant sequence includes one or more of the first third-order cumulant sequence and the first fourth-order cumulant sequence.

[0092] The first third-order cumulant sequence is represented as: , and the first fourth-order cumulant sequence is represented as: wherein, denotes the third-order cumulant of time step in the first third-order cumulant sequence, denotes the fourth-order cumulant of time step in the first fourth-order cumulant sequence.

[0093] If If the signal mutation is determined to exist, the sliding window is reset to exclude historical data interference, the current window data is emptied, and the process of accumulating the statistical quantity is restarted.

[0094] In another embodiment, in S300, the high-order cumulants include third-order cumulants and fourth-order cumulants, the first high-order cumulant sequence includes a first third-order cumulant sequence and a first fourth-order cumulant sequence, the ultrasonic echo signal is predicted based on the average value sequence and the first high-order cumulant sequence to obtain a signal prediction error, and then a correction factor is obtained, the first high-order cumulant sequence is corrected based on the correction factor to obtain a second high-order cumulant sequence, including the following steps:

[0095] S310: Based on the first third-order cumulant sequence, the first fourth-order cumulant sequence, and the average value sequence, a signal prediction model is combined to obtain an ultrasonic echo signal prediction value sequence. The signal prediction model is: the first third-order cumulant and the first fourth-order cumulant are weighted and summed, and the sum is summed with the average value of the ultrasonic echo signal to obtain the ultrasonic echo signal prediction value. The signal prediction model can be expressed as: wherein, represents the ultrasonic echo signal prediction value at the current time step , which is a signal prediction value based on historical cumulants. represents the average value of the ultrasonic echo signal at the previous time step , represents the first third-order cumulant at the previous time step , represents the first fourth-order cumulant at the previous time step . represents a weight coefficient, and are obtained by minimizing the sum of squares of error values, and by default .

[0096] S320: The first echo signal sequence and the ultrasonic echo signal prediction value sequence are subtracted to obtain a signal prediction error sequence, and the prediction error , represents a norm.

[0097] S330: Based on the signal prediction error at the current time step, the first third-order cumulant at the current time step, and the first fourth-order cumulant at the current time step, the correction factor at the previous time step is adjusted to obtain the correction factor at the current time step, and then a correction factor sequence is obtained. The correction factor at the current time step can be expressed as: wherein, represents the ultrasonic echo signal prediction value at the current time step ​a correction factor, represents a previous time step a correction factor, the correction factor dynamically adjusts according to the nonlinear characteristics of the signal. represents a current time step a signal prediction error, , represents an ultrasonic echo signal in the first echo signal sequence at a current time step , represents a first third-order cumulant at a current time step , represents a first fourth-order cumulant at a current time step . represents a learning rate, all default to 0.01, the learning rate is used to control the correction or update speed of the correction factor, prevent over-correction, avoid false triggering caused by noise. When the signal suddenly changes (such as the appearance of defects), the error increases, and the correction factor responds quickly, enhancing the amplification effect on the nonlinear characteristics.

[0098] S340: Based on the correction factor sequence, the first third-order cumulant sequence and the first fourth-order cumulant sequence are respectively corrected to obtain the second third-order cumulant sequence and the second fourth-order cumulant sequence, then the second third-order cumulant sequence is represented as: , the second fourth-order cumulant sequence is represented as: , wherein represents a third-order cumulant in the second third-order cumulant sequence at a current time step , represents a fourth-order cumulant in the second fourth-order cumulant sequence at a current time step . By adjusting the third-order cumulant through the correction factor, when , the cumulant of the asymmetric defect (such as a crack) is amplified; when , the false asymmetric response caused by noise is suppressed, thereby enhancing the nonlinear characteristics of the third-order cumulant. By adjusting the fourth-order cumulant through the correction factor, the sensitivity to high-order nonlinear defects (such as surface peeling) is enhanced. Thus, by adjusting the third-order and fourth-order cumulants through the correction factor, the sensitivity of the nonlinear signal change caused by defects is enhanced, and the nonlinear characteristics of the defect signal are highlighted.

[0099] In another embodiment, in S300, the high-order cumulant is a third-order cumulant or a fourth-order cumulant, and the first high-order cumulant sequence is a first third-order cumulant sequence or a first fourth-order cumulant sequence. Based on the average value sequence and the first high-order cumulant sequence, the ultrasonic echo signal is predicted to obtain a signal prediction error, and then a correction factor is obtained. Based on the correction factor, the first high-order cumulant sequence is corrected to obtain a second high-order cumulant sequence, comprising the following steps:

[0100] (1) Based on the first third-order cumulant sequence / first fourth-order cumulant sequence and the average value sequence, combined with the third-order signal prediction model / fourth-order signal prediction model, the first ultrasonic echo signal prediction value sequence / second ultrasonic echo signal prediction value sequence is obtained. Among them, the third-order signal prediction model / fourth-order signal prediction model is: based on the first third-order cumulant / first fourth-order cumulant, the average value of the ultrasonic echo signal, the first ultrasonic echo signal prediction value / second ultrasonic echo signal prediction value is obtained, and the third-order signal prediction model can be expressed as: , and the fourth-order signal prediction model can be expressed as: .

[0101] (2) The first echo signal sequence and the first ultrasonic echo signal prediction value sequence / second ultrasonic echo signal prediction value sequence are subtracted to obtain the first signal prediction error sequence / second signal prediction error sequence;

[0102] (3) Based on the first signal prediction error / second signal prediction error of the current time step and the first third-order cumulant / first fourth-order cumulant of the current time step, the third-order correction factor / fourth-order correction factor of the previous time step is adjusted to obtain the third-order correction factor / fourth-order correction factor of the current time step, and then the third-order correction factor sequence / fourth-order correction factor sequence is obtained. The third-order correction factor of the current time step can be expressed as: , and the fourth-order correction factor of the current time step can be expressed as: .

[0103] (4) Based on the third-order correction factor sequence / fourth-order correction factor sequence, the first third-order cumulant sequence / first fourth-order cumulant sequence is modified to obtain the second third-order cumulant sequence / second fourth-order cumulant sequence. The second third-order cumulant sequence can be expressed as: , and the second fourth-order cumulant sequence can be expressed as: , wherein and respectively represent the first ultrasonic echo signal prediction value of the current time step , the second ultrasonic echo signal prediction value of the current time step , represents a weight coefficient, represents the first signal prediction error of the current time step in the first ultrasonic echo signal prediction value sequence, , represents the second signal prediction error of the current time step in the second ultrasonic echo signal prediction value sequence, , and respectively represent the first third-order cumulant of the current time step The third-order correction factor, the current time step The fourth-order correction factor, and They represent the previous time step, respectively. The third-order correction factor, the previous time step The fourth-order correction factor, This represents the learning rate.

[0104] In another embodiment, in S400, the first echo signal sequence is optimized based on the dictionary matrix and the signal sparse coefficient vector to obtain the second echo signal sequence, including the following steps:

[0105] S410: Based on a preset segment step size and a preset overlap step size, the first echo signal sequence is divided into several segments to obtain a signal segment sequence. The first echo signal sequence is usually a long sequence, but sparse representation does not directly process long sequences. Instead, it is processed in blocks, dividing the first echo signal sequence into several segment sequences. The length of each segment sequence is a preset segment step size m. There are preset overlap step sizes between adjacent segment sequences. Each segment sequence is used as input for dictionary learning.

[0106] S420: Predict the signal segment sequence using the dictionary matrix and the signal sparse coefficient vector, and obtain the optimal dictionary matrix and the optimal signal sparse coefficient vector for each segment based on the principle of minimizing error.

[0107] Sparse representation requires only a small number of non-zero coefficients to reconstruct a signal, significantly reducing storage and computation costs while eliminating noise (minimizing errors and filtering out irrelevant information). Sparse coefficients reflect the projection of the signal onto dictionary atoms, facilitating feature reuse in subsequent classification / detection tasks. Dynamic dictionary learning, through iterative optimization, matches the dictionary matrix to data features, improving the model's ability to represent complex signals.

[0108] S430: Based on the optimal dictionary matrix and the optimal signal sparse coefficient vector for each segment, the signal segment sequence for each segment is reconstructed to achieve denoising, resulting in a denoised signal segment sequence. The denoised signal segment sequence is then represented as: ,in, Representing fragments Denoising signal segment sequence, Represents the optimal dictionary matrix. Representing fragments The optimal sparse coefficient vector of the signal.

[0109] S440: The denoised signal segment sequence is restored by weighted summation based on overlapping positions, and the resulting denoised echo signal sequence is the second echo signal sequence.

[0110] In another embodiment, in S420, the signal segment sequence is predicted using a dictionary matrix and a signal sparse coefficient vector, and the optimal dictionary matrix and the optimal signal sparse coefficient vector for each segment are obtained based on the error minimization principle, including the following steps:

[0111] (1) Initialize the dictionary matrix and the sparse coefficient vector of each segment to obtain the initial dictionary matrix and the initial sparse coefficient vector of each segment. Use the initial dictionary matrix as the dictionary matrix of the current segment. The initial dictionary matrix can be set as a DCT basis or a random orthogonal matrix.

[0112] Based on the dictionary matrix of the current segment and the initial sparse coefficient vector of the signal for the current segment, combined with the signal dictionary prediction model, the signal prediction sequence for the current segment is obtained. The signal dictionary prediction model is as follows: multiplying the dictionary matrix by the sparse coefficient vector of the signal yields the signal prediction sequence. The signal dictionary prediction model can be expressed as: ,in, This represents the signal prediction sequence. Let represent a dictionary matrix, where each column of the dictionary matrix is ​​a basis function (unit vector). If the dictionary has *b* basis functions, then the dictionary matrix... . Represents the sparse coefficient vector of the signal. , This represents the basis function activated for non-zero elements, where most elements are close to zero. This represents a noise sequence.

[0113] Based on the signal prediction sequence and the signal segment sequence of the current segment, the reconstruction error value of the current segment is obtained, and the reconstruction error value of the current segment can be expressed as: ,in, Indicates the first The reconstruction error value of each segment is the reconstruction error value of the current segment. Represents the first segment in the signal segment sequence. A sequence of signal segments, Represents the dictionary matrix of the current segment. This represents the initial sparse coefficient vector of the current segment. This represents the signal prediction sequence for the current segment. This represents the square of the L2 norm. Let L1 norm be denoted as and let represent forced sparsity. This represents the L1 regularization parameter, used to control the sparsity strength; the default value is... .

[0114] (2) Based on the principle of minimizing the current segment reconstruction error value, the initial signal sparse coefficient vector of the current segment is solved, and the optimal signal sparse coefficient vector of the current segment is obtained. In this step, the dictionary matrix is kept unchanged, and the optimal signal sparse coefficient vector is solved based on the principle of minimizing the reconstruction error value. FISTA algorithm can be used to solve it to speed up the convergence, so as to quickly solve it.

[0115] (3) The gradient descent method is used to iteratively update the dictionary matrix. In this step, the signal sparse coefficient vector is kept unchanged, which is the optimal signal sparse coefficient vector obtained in the previous step. Specifically, based on the signal segment sequence of the current segment, the optimal signal sparse coefficient vector of the current segment and the current dictionary matrix, the current segment dictionary gradient matrix is obtained. The current segment dictionary gradient matrix can be represented as: , wherein represents the dictionary gradient matrix of the current segment , , represents the optimal signal sparse coefficient vector of the current segment represents the transpose.

[0116] Based on the current segment dictionary matrix and the current segment dictionary gradient matrix, the initial dictionary matrix of the next segment is obtained. The initial dictionary matrix of the next segment can be represented as: , wherein represents the initial dictionary matrix of the next segment . represents the dictionary learning rate, which is used to control the step size of dictionary update.

[0117] To ensure the orthogonality of the basis functions in the dictionary matrix, QR decomposition is performed on the dictionary matrix after each update, which can avoid redundant basis functions and improve the stability of sparse representation. The orthogonal matrix Q obtained after QR decomposition preserves the orthogonal basis functions. Specifically, the QR decomposition of the initial dictionary matrix of the next segment is performed to obtain the orthogonal matrix, which is the dictionary matrix of the next segment. The QR decomposition can be represented as: , wherein represents the orthogonal matrix , represents the upper triangular matrix.

[0118] (4) Until the signal segment sequence is traversed or the iteration termination condition is met, the final dictionary matrix of the next segment is obtained, which is the optimal dictionary matrix, and the optimal signal sparse coefficient vector of each segment is obtained. The iteration termination condition includes one or more of the following: reaching the preset number of iterations, the reconstruction error value being lower than the preset error threshold, and the update amplitude of the dictionary matrix being lower than the preset dictionary amplitude threshold. By dynamically updating the dictionary matrix, it can adapt to the non-stationary characteristics of the signal (such as local mutation caused by defects).

[0119] In another embodiment, in S500, the skewness sequence, the kurtosis sequence, the second high-order cumulant sequence and the second echo signal sequence are fused to obtain a multi-dimensional feature vector sequence. The multi-dimensional feature vector is represented as: wherein, represents the multi-dimensional feature vector at time step of the multi-dimensional feature vector sequence, represents the skewness at time step of the skewness sequence, represents the kurtosis at time step of the kurtosis sequence, represents the third-order cumulant at time step of the second third-order cumulant sequence, represents the fourth-order cumulant at time step of the second fourth-order cumulant sequence, represents the ultrasonic echo signal at time step of the second echo signal sequence.

[0120] wherein, the skewness sequence, the kurtosis sequence and the first high-order cumulant sequence are calculated on the basis of the first echo signal sequence with a preset sliding window length, the second high-order cumulant sequence is modified on the basis of the first high-order cumulant sequence, and it is known that the lengths of data in the skewness sequence, the kurtosis sequence and the second high-order cumulant sequence are consistent, but the length of data in the second echo signal sequence is greater than that in the skewness sequence / kurtosis sequence / second high-order cumulant sequence. At this time, the signal in the second echo signal sequence consistent with the time step in the skewness sequence / kurtosis sequence / second high-order cumulant sequence can be extracted according to the time step alignment to form a third echo signal sequence. The skewness sequence, the kurtosis sequence, the second high-order cumulant sequence and the third echo signal sequence are fused to obtain a multi-dimensional feature vector sequence.

[0121] In another embodiment, in S600, the multi-dimensional feature vector sequence is compared with a preset vector threshold to detect whether the log mother bush bearing roller has defects, including the following steps:

[0122] S610: The multi-dimensional feature vector sequence is standardized to obtain a standard multi-dimensional feature vector sequence. The standard multi-dimensional feature vector can be represented as: wherein, represents the feature value of the th feature at time step of the standard multi-dimensional feature vector sequence, represents the th feature of the multi-dimensional feature vector sequence at time step eigenvalues. Represents the th element in a multidimensional eigenvector. The average of the features, Represents the th element in a multidimensional eigenvector. The standard deviation vector of each feature and It can also be calculated using offline training data.

[0123] S620: Based on Mahalanobis distance, the degree to which vectors in a standard multidimensional feature vector sequence deviate from the normal distribution is calculated. The vector deviation value is the Mahalanobis distance of the multidimensional feature vectors, thus obtaining the Mahalanobis distance sequence. The normal distribution is usually a Gaussian distribution, so the preset vector threshold is a vector that satisfies a Gaussian distribution. The Mahalanobis distance can be expressed as: ,in, Represents the time steps in the Mahalanobis distance sequence Mahalanobis distance, This represents the standard multidimensional feature vector sequence at time step eigenvectors, This represents the average value of the multidimensional feature vector sequence when there are no defects. The covariance matrix of the multidimensional eigenvector sequence when there are no defects. and It can be calculated using offline training data.

[0124] S630: If there is a Mahalanobis distance in the Mahalanobis distance sequence that is greater than a preset Mahalanobis distance threshold, then the logarithmic line bearing roller is determined to have a defect. The preset Mahalanobis distance threshold can be expressed as: ,in, This represents the preset Mahalanobis distance threshold. This represents the average historical Mahalanobis distance. The standard deviation of the historical Mahalanobis distance is represented by... and Calculated using offline training data This represents the sensitivity coefficient.

[0125] In another embodiment, the method for detecting defects in logarithmic busbar bearing rollers based on ultrasonic technology further includes: constructing a defect detection pre-training model, training the defect detection pre-training model based on historical multidimensional feature vector sequences and logarithmic busbar bearing roller defect detection results to obtain a defect detection model, including the following steps:

[0126] (1) Obtain historical ultrasonic detection signals from the surface of the logarithmic busbar bearing rollers to obtain the corresponding historical multidimensional feature vector sequence. And the sequence of deviation values ​​of historical vectors.

[0127] The historical vector defect detection result value sequence is obtained as follows: when the vector deviation value in the historical vector deviation value sequence is greater than the preset deviation threshold, a defect exists, and the corresponding historical multidimensional feature vector is obtained. Defect detection value The first preset defect value (Set to +1); when the vector deviation value in the historical vector deviation value sequence is less than or equal to the preset deviation threshold, there is no defect, and the corresponding historical multidimensional feature vector... Defect detection value The second preset defect value (Set to -1).

[0128] The historical multidimensional feature vector sequence and the corresponding defect detection result value sequence are used to form the training sample dataset. , This represents the total number of samples in the training sample dataset.

[0129] If the amount of data is large enough, a multidimensional feature vector sequence can be used to replace the historical multidimensional feature vector sequence.

[0130] (2) Construct a defect detection pre-training model. The goal of the defect detection pre-training model is to classify the input data as defective or non-defective. The classification goal is to find the optimal hyperplane such that... To maximize the class margin, the defect detection pre-trained model includes an optimization target model, a constraint model, and a classification decision model.

[0131] Based on the weight vector, an optimization objective model is constructed, which can be represented as: ,in, Represents the weight vector. This represents the regularization term. This indicates that the regularization parameter is used to balance the maximization of the margin with the classification error. Indicates time step The corresponding slack variable indicates that a small number of misclassifications are allowed, and . Indicates time step The total number. The optimization objective of the objective function is to minimize This is equivalent to maximizing the interval. .

[0132] The training sample dataset restricts the possible locations of the hyperplane through constraints, ensuring the feasibility of classification. Based on the weight vector, bias term, standard multidimensional feature vector, and defect detection results, a constraint model is constructed, which can be expressed as: ,in, Represents the time steps in a multidimensional feature vector sequence. corresponding defect detection value, denotes transposition, denotes bias term, constraint condition represents that all samples are correctly classified and located outside the interval boundary.

[0133] (3) To process the constraint condition, a Lagrange model is constructed, specifically: based on the optimization target model and the constraint condition model, a Lagrange model is constructed in combination with the Lagrange multiplier, and the Lagrange model is represented as: wherein, denotes Lagrange value, denotes Lagrange multiplier, denotes the Lagrange multiplier of the time step . In the Lagrange model, each corresponds to the constraint condition of a sample, reflecting the influence weight of the sample on the hyperplane.

[0134] The partial derivative of the Lagrange model is taken and set to zero to obtain the partial derivative model.

[0135] Based on the training sample data set, the partial derivative model is solved in combination with the SMO algorithm to obtain the optimal solution of the weight vector and the bias term, that is, the optimal weight vector and the optimal bias term .

[0136] (4) Based on the optimal weight vector and the optimal bias term, a classification decision model is constructed, and the classification decision model can be represented as: wherein, denotes optimal defect detection result prediction value, denotes the multi-dimensional feature vector of the to-be-detected logarithmic bus bearing roller surface phased array ultrasonic detection signal.

[0137] (5) Based on the classification decision model, the multi-dimensional feature vector is inferred to obtain the defect detection result prediction value, and thus the defect detection model has been obtained. In the defect detection model, the optimization target model and the constraint condition model need to be trained, and after training, the optimal weight vector and the optimal bias term are obtained. The classification decision model is constructed through the optimal weight vector and the optimal bias term, and finally the classification decision model is actually used to predict the defect detection. It should be noted that although the classification decision model is finally used to predict the defect detection, the classification decision model depends on the training of the optimization target model and the constraint condition model. The three models together constitute the defect detection pre-training model / defect detection model, and each time the training needs to be completed in cooperation with the optimization target model, the constraint condition model and the classification decision model.

[0138] Further, the trained defect detection model (mainly a classification decision model) is used to infer the ultrasonic detection signal of the logarithmic bus bearing roller surface to obtain a defect detection result prediction value,

[0139] Embodiment 2:

[0140] A logarithmic bus bearing roller defect detection system based on ultrasonic technology, as shown in Figure 2 The signal acquisition and processing module 100, the initial feature value acquisition module 200, the high-order feature value correction module 300, the signal correction module 400, the fusion module 500, and the defect detection module 600 are shown.

[0141] The signal acquisition and processing module 100 acquires the ultrasonic detection signal of the logarithmic bus bearing roller surface based on time series and performs preprocessing to obtain a first echo signal sequence.

[0142] The initial feature value acquisition module 200 calculates the average value, skewness, kurtosis, and high-order cumulant of the first echo signal sequence based on a preset sliding window length to obtain an average value sequence, a skewness sequence, a kurtosis sequence, and a first high-order cumulant sequence.

[0143] The high-order feature value correction module 300 corrects the first high-order cumulant sequence based on the average value sequence to obtain a second high-order cumulant sequence.

[0144] The signal correction module 400 optimizes the first echo signal sequence based on a dictionary matrix and a signal sparse coefficient vector to obtain a second echo signal sequence.

[0145] The fusion module 500 fuses the skewness sequence, the kurtosis sequence, the second high-order cumulant sequence, and the second echo signal sequence to obtain a multi-dimensional feature vector sequence.

[0146] The defect detection module 600 is configured to compare the multi-dimensional feature vector sequence with a preset vector threshold to detect whether the logarithmic bus bearing roller has defects.

[0147] Various changes and modifications made without departing from the spirit and scope of the present application, all equivalent technical solutions also belong to the scope of the present application.

[0148] Each embodiment in the specification is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts of each embodiment can be referred to each other.

[0149] Those skilled in the art will appreciate that embodiments of the present application can be devised for a variety of applications. It is intended that the present application be limited only by the scope of the appended claims, and it is intended that various modifications and alterations made by persons skilled in the art to the illustrative embodiments discussed herein are within the scope, spirit, and scope of the present application. In particular, it is intended that the present application cover modifications and alterations of various embodiments of the present application that are within the scope and spirit of the present application. It is intended that the following claims be construed to embrace all such alterations and modifications as fall within the scope of the appended claims.

[0150] The present application is described in reference to the drawings using flowchart illustrations and / or block diagrams of methods, terminal devices (systems), and computer program products according to the present application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing terminal devices to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal devices, create means for implementing the functions specified in the flowchart illustrations and / or block diagrams block or blocks. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams block or blocks. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams block or blocks.

[0151] These computer program instructions can also be stored in a computer- readable memory that can direct a computer or other programmable data processing terminal devices to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the function specified in the flowchart illustrations and / or block diagrams block or blocks. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams block or blocks. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams block or blocks.

[0152] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal devices to cause a series of operational steps to be performed on the computer or other programmable terminal devices to produce a computer implemented process such that the instructions which execute on the computer or other programmable terminal devices provide steps for implementing the functions specified in the flowchart illustrations and / or block diagrams block or blocks. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams block or blocks. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams block or blocks.

[0153] It is to be understood that:

[0154] It is to be understood that the terms "including", "comprising", "consisting" and "involving", as well as variations thereof, are to be construed in an open-ended fashion, and should be interpreted to mean that certain embodiments (e.g., compositions, methods, processes, etc.) include, but are not limited to, the listed steps, elements, components, etc.

[0155] Moreover, it should be noted that the specific embodiments described in the specification are illustrative only and not restrictive of the patent concept. Equivalent or similar changes or modifications made to the configurations, features, and principles described in the patent concept are included in the scope of the patent. Those skilled in the art can make various modifications or supplements to the specific embodiments described or use similar ways to replace them, as long as they do not deviate from the structure of the patent or exceed the scope defined by the claims.

Claims

1. A method for detecting defects in a logarithmic-spiral-groove bearing roller based on ultrasonic technology, characterized in that, The method comprises the following steps: Based on time series, the ultrasonic detection signal of the logarithmic generator bearing roller surface is obtained and preprocessed to obtain a first echo signal sequence; The average value sequence, the skewness sequence, the kurtosis sequence and the first high-order cumulant sequence are calculated based on the average value of the first echo signal sequence based on the preset sliding window length; The first high-order cumulant sequence is modified based on the average value sequence to obtain a second high-order cumulant sequence; The first echo signal sequence is optimized based on the dictionary matrix and the signal sparse coefficient vector to obtain a second echo signal sequence; The skewness sequence, the kurtosis sequence, the second high-order cumulant sequence and the second echo signal sequence are fused to obtain a multi-dimensional feature vector sequence; The multi-dimensional feature vector sequence is compared with a preset vector threshold to detect whether the logarithmic generator bearing roller has defects.

2. The ultrasonic-based logarithmic-spiral-bus-bar bearing-roller defect detection method according to claim 1, characterized in that, The ultrasonic detection signal is a phased array ultrasonic detection signal, and the ultrasonic detection signal of the logarithmic generator bearing roller surface is obtained based on time series and preprocessed to obtain a first echo signal sequence, which comprises the following steps: Based on time series, the ultrasonic detection signal of the logarithmic generator bearing roller surface is obtained to obtain an initial echo signal sequence represented in a polar coordinate system; The initial echo signal sequence is converted from the polar coordinate system to the Cartesian coordinate system to obtain a first echo signal sequence.

3. The ultrasonic-based logarithmic-spiral-bus-bar bearing-roller defect detection method according to claim 1, characterized in that, The first high-order cumulant sequence includes a first third-order cumulant sequence and a first fourth-order cumulant sequence, and the first high-order cumulant sequence is modified based on the average value sequence to obtain a second high-order cumulant sequence, which comprises the following steps: Based on the first third-order cumulant sequence, the first fourth-order cumulant sequence and the average value sequence, a signal prediction model is combined to obtain an ultrasonic echo signal prediction value sequence; wherein the signal prediction model is: the first third-order cumulant and the first fourth-order cumulant are weighted and summed, and the ultrasonic echo signal average value is summed to obtain the ultrasonic echo signal prediction value; The first echo signal sequence and the ultrasonic echo signal prediction value sequence are processed to obtain a signal prediction error sequence; Based on the signal prediction error of the current time step, the first third-order cumulant of the current time step and the first fourth-order cumulant of the current time step, the correction factor of the previous time step is adjusted to obtain the correction factor of the current time step, and then the correction factor sequence is obtained; The first third-order cumulant sequence and the first fourth-order cumulant sequence are modified based on the correction factor sequence to obtain a second third-order cumulant sequence and a second fourth-order cumulant sequence.

4. The ultrasonic-based logarithmic-spiral-bus-bar-roller defect detection method according to claim 3, characterized in that, The correction factor of the current time step is represented as follows: in, Indicates the current time step The correction factor Indicates the previous time step The correction factor Indicates the current time step The signal prediction error, Indicates the current time step The first and third order cumulative amount, Indicates the current time step The first and fourth order cumulative amount, This represents the learning rate.

5. The ultrasonic-based logarithmic-spiral-bus-bar bearing-roller defect detection method according to claim 1, characterized in that, The first high-order cumulant sequence is a first third-order cumulant sequence or a first fourth-order cumulant sequence, and the first high-order cumulant sequence is modified based on the average value sequence to obtain a second high-order cumulant sequence, which comprises the following steps: When the first high-order cumulant sequence is a first third-order cumulant sequence: Based on the first third-order cumulant sequence and the average value sequence, a first ultrasonic echo signal predicted value sequence is obtained by combining a third-order signal prediction model; wherein the third-order signal prediction model is: adjusting the first third-order cumulant and summing with the average value of the ultrasonic echo signal to obtain the first ultrasonic echo signal predicted value; The first echo signal sequence and the first ultrasonic echo signal predicted value sequence are subtracted to obtain a first signal prediction error sequence; Based on the first signal prediction error of the current time step and the first third-order cumulant of the current time step, the third-order correction factor of the previous time step is adjusted to obtain the third-order correction factor of the current time step, and then a third-order correction factor sequence is obtained; Based on the third-order correction factor sequence, the first third-order cumulant sequence is corrected to obtain a second third-order cumulant sequence; When the first high-order cumulant sequence is a first fourth-order cumulant sequence, the first fourth-order cumulant sequence is corrected by the same method to obtain a second fourth-order cumulant sequence.

6. The ultrasonic-based logarithmic-spiral-bus-bar bearing-roller defect detection method according to claim 1, characterized in that, The first echo signal sequence is optimized based on the dictionary matrix and the signal sparse coefficient vector to obtain a second echo signal sequence, including the following steps: Based on the preset segment step and the preset overlap step, the first echo signal sequence is divided into a plurality of segments to obtain a signal segment sequence; The signal segment sequence is predicted by the dictionary matrix and the signal sparse coefficient vector, and based on the error minimization principle, the optimal dictionary matrix and the optimal signal sparse coefficient vector of each segment are obtained; Based on the optimal dictionary matrix and the optimal signal sparse coefficient vector of each segment, the signal segment sequence of each segment is reconstructed to achieve denoising to obtain a denoised signal segment sequence; The denoised signal segment sequence is restored by overlap position weighted addition to obtain a denoised echo signal sequence, which is the second echo signal sequence.

7. The ultrasonic-based logarithmic-spiral-bus-bar bearing-roller defect detection method according to claim 1, characterized in that, The multi-dimensional feature vector sequence is compared with the preset vector threshold to detect whether the logarithmic generatrix bearing roller is defective, including the following steps: The multi-dimensional feature vector sequence is standardized to obtain a standard multi-dimensional feature vector sequence; The degree to which the vectors in the standard multi-dimensional feature vector sequence deviate from the Gaussian distribution is calculated based on the Mahalanobis distance to obtain a vector deviation value, which is the Mahalanobis distance of the multi-dimensional feature vector, and then a Mahalanobis distance sequence is obtained; If there is a Mahalanobis distance greater than the preset Mahalanobis distance threshold in the Mahalanobis distance sequence, it is determined that the logarithmic generatrix bearing roller is defective.

8. The ultrasonic-based logarithmic-spiral-bush bearing-roller defect detection method according to claim 1, characterized by Further comprising: A defect detection pre-training model is constructed, and the defect detection pre-training model is trained based on historical multi-dimensional feature vector sequences and logarithmic generatrix bearing roller defect detection results to obtain a defect detection model, including the following steps: Historical ultrasonic detection signals of the surface of the logarithmic generatrix bearing roller are obtained to obtain corresponding historical multi-dimensional feature vector sequences and historical vector defect detection result value sequences, and then a training sample data set is formed; The defect detection pre-training model comprises an optimization target model, a constraint condition model and a classification decision model; specifically, the optimization target model is constructed based on a weight vector; the constraint condition model is constructed based on the weight vector, a bias term, a multi-dimensional feature vector and a defect detection value; The Lagrange model is constructed based on the optimization target model, the constraint condition model and a Lagrange multiplier; The partial derivative model is obtained by taking the partial derivative of the Lagrange model and setting it to zero; The optimal weight vector and the optimal bias term are obtained by solving the partial derivative model based on the training sample data set; The classification decision model is constructed based on the optimal weight vector and the optimal bias term; The defect detection result prediction value is obtained by reasoning the multi-dimensional feature vector based on the classification decision model, that is, the defect detection model is obtained.

9. A system for detecting defects in a logarithmic-spiral- groove bearing roller based on ultrasonic technology, characterized in that it comprises: The method of any one of claims 1 to 8 is implemented, and the detection system comprises a signal acquisition and processing module, an initial feature value acquisition module, a high-order feature value correction module, a signal correction module, a fusion module and a defect detection module; The signal acquisition and processing module is configured to acquire an ultrasonic detection signal of a logarithmic cylindrical bearing roller surface based on a time sequence and perform preprocessing to obtain a first echo signal sequence; The initial feature value acquisition module is configured to calculate the average value, skewness, kurtosis and high-order cumulant of the first echo signal sequence based on a preset sliding window length to obtain an average value sequence, a skewness sequence, a kurtosis sequence and a first high-order cumulant sequence; The high-order feature value correction module is configured to correct the first high-order cumulant sequence based on the average value sequence to obtain a second high-order cumulant sequence; The signal correction module is configured to optimize the first echo signal sequence based on a dictionary matrix and a signal sparse coefficient vector to obtain a second echo signal sequence; The fusion module is configured to fuse the skewness sequence, the kurtosis sequence, the second high-order cumulant sequence and the second echo signal sequence to obtain a multi-dimensional feature vector sequence; The defect detection module is configured to compare the multi-dimensional feature vector sequence with a preset vector threshold to detect whether the logarithmic cylindrical bearing roller has defects.

10. A computer-readable storage medium storing a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1-9. The computer program is executed by the processor to implement the method of any one of claims 1 to 8.

11. An apparatus comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, The processor executes the computer program to implement the method of any one of claims 1 to 8.

Images