A physical driving measurement method for monocular three-dimensional dynamic displacement of a rotating machine

Abstract

The application provides a physical driving measurement method for monocular three-dimensional dynamic displacement of a rotating machine, and belongs to the technical field of computer vision and industrial state monitoring; a high-speed dynamic vision acquisition system is constructed, and a time sequence video stream is acquired by using a cooperative marker; a "Gaussian + motion blur" composite gradient model taking a motion blur width as an endogenous variable is constructed, and a nonlinear optimization algorithm is used to jointly solve sub-pixel edge coordinates and a blur degree; in combination with time sequence robust filtering and monocular imaging geometric principles, pixel displacement of radial X and Y axes and axial Z axis is decoupled by using marker width variation and edge displacement; finally, three-way physical vibration waveforms are output by using a calibration conversion factor. The application starts from a physical imaging mechanism, adaptively eliminates blur influence, and realizes three-way vibration synchronous measurement with micron-level precision by using only a single camera, so that hardware cost and deployment difficulty are reduced.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of computer vision and industrial condition monitoring, and in particular relates to a physical driving measurement method for monocular three-dimensional dynamic displacement of rotating machinery. Background Technology

[0002] Rotating machinery (such as steam turbines, aero engines, centrifugal compressors, and high-end CNC machine tool spindles) are core power components of modern industrial equipment. As the heart of rotating machinery, the rotor system's vibration state during operation is a crucial basis for assessing equipment health and diagnosing faults (such as imbalance, misalignment, rubbing, and cracks). In particular, the complex motion patterns of the rotor in three-dimensional space (including radial horizontal / vertical vibration and axial movement) contain rich dynamic information.

[0003] Currently, the mainstream methods for obtaining rotor vibration and displacement data in industry mainly rely on contact or near-field sensors such as eddy current displacement sensors (ECS) and laser Doppler vibration meters (LDV). However, these traditional methods have revealed significant limitations in precision and extreme operating condition monitoring: First, contact sensors suffer from "added mass effect," which alters the natural frequency of lightweight, thin-walled, or micro rotors, leading to measurement distortion; second, eddy current sensors have extremely high requirements for the installation environment, typically requiring a near-field distance of less than 2 mm, making them difficult to deploy in high-temperature, high-pressure, or confined spaces (such as inside an aero-engine); third, to achieve simultaneous measurement in the X, Y, and Z axes, an orthogonal array composed of multiple single-axis sensors must be constructed, resulting in high hardware costs and extremely high system integration complexity.

[0004] With the development of machine vision technology, camera-based non-contact measurement has become a hot topic. However, existing monocular vision methods face two major technical bottlenecks: first, the "dimensional collapse" problem. Due to the perspective projection principle of monocular imaging, depth information is lost during the imaging process, making it impossible to directly perceive depth displacement along the optical axis (i.e., the radial X-axis vibration of the rotor); second, the accuracy failure caused by "high-speed motion blur". In rotating machinery monitoring scenarios, the rotor linear velocity is extremely high. Limited by the physical exposure time of industrial cameras, the rotor surface texture or edges will produce non-negligible displacement within the integration time, resulting in motion blur. Existing sub-pixel edge localization algorithms (such as Canny interpolation, gray-scale centroid method, and static Gaussian fitting) are all based on the "quasi-static" assumption, which will produce severe model mismatch under high-speed blur conditions, resulting in localization errors exceeding 20%, failing to meet the needs of micron-level precision diagnosis. Summary of the Invention

[0005] To address the aforementioned problems, this invention aims to resolve the technical pain points, particularly the inability of monocular vision to perceive depth displacement and the decreased measurement accuracy caused by high-speed motion blur. Instead of relying solely on data-driven image processing, this invention proposes a Physics-informed Measurement Method based on physical imaging mechanisms. By establishing a composite gradient mathematical model incorporating optical and motion blur, and combining it with geometric projection constraints, high-precision, synchronous, and non-contact measurement of the radial (X / Y) and axial (Z) triaxial vibration displacement of a high-speed rotating rotor can be achieved using only a single low-cost industrial camera.

[0006] This invention provides a physical driving measurement method for monocular three-dimensional dynamic displacement of rotating machinery, comprising the following processes:

[0007] S1, based on cooperative markers with known physical dimensions arranged on the surface of a rotating machinery rotor, uses an industrial camera to acquire a time-series video stream containing rotor vibration characteristics;

[0008] S2, perform image preprocessing on the acquired time-series video stream, obtain the region of interest and enhance the edge gradient features to obtain a high-quality rotor edge grayscale image, and extract gradient information in a specific direction;

[0009] S3, based on gradient information, constructs a physical-driven Gaussian and motion blur composite gradient model, using the motion blur width as an endogenous variable to describe the gray-level gradient distribution at the edge under high-speed rotation;

[0010] S4. Based on the composite gradient model, a nonlinear optimization algorithm is used to jointly iteratively solve the gradient profile of the rotor edge to calculate the sub-pixel edge coordinate sequence containing position information and ambiguity information.

[0011] S5, perform time-series robust filtering on the sub-pixel edge coordinate sequence to remove outliers and smooth out jitter, and obtain a high-precision edge position time sequence.

[0012] S6. A multi-directional vibration decoupling model is established based on the monocular imaging geometry principle. Using the high-precision edge position timing, the rotor's radial vibration is calculated separately. Axial and radial Shaft and axial direction Pixel displacement along the axis;

[0013] S7 uses the conversion factor obtained from physical scale calibration to convert the pixel displacement into physical spatial displacement and outputs a time-synchronized triaxial vibration waveform.

[0014] Preferably, the specific steps of S2 are as follows:

[0015] The acquired color video stream is converted into a grayscale image, and a Gaussian filter is applied to smooth the image.

[0016] A contrast-limited adaptive histogram equalization algorithm is used to enhance local contrast: the image is divided into several sub-blocks, and the gray-level histogram of each sub-block is calculated. Then, the histogram is cropped to obtain the cropped histogram. ;

[0017] Constructing the cumulative distribution function of the clipping histogram And map the pixels according to the following formula:

[0018] ;

[0019] in, This represents the enhanced pixel grayscale value; Represents the pixel grayscale value in the original image; Represents the coordinates of a pixel; This represents the lower limit of the target's enhanced grayscale range; This indicates the upper limit of the target enhanced grayscale range; This represents the cumulative distribution function of the current sub-block; This indicates the total number of pixels within the current sub-block;

[0020] The Sobel operator is used to calculate the gradient in the horizontal and vertical directions respectively, and the gradient information in specific directions is extracted.

[0021] Preferably, the mathematical expression of the physical-driven Gaussian and motion-fuzzy composite gradient model constructed based on gradient information is as follows:

[0022] ;

[0023] in, Indicates pixel coordinates Theoretical value of the edge gradient at the location; This represents the vector of model parameters to be solved; The magnitude coefficient of the gradient; Represents the Gaussian error function; Indicates the pixel coordinate position along the gradient direction; Indicates the center position of the subpixel edge; Indicates the equivalent motion blur width; The standard deviation representing optical blur; This represents the baseline offset of the gradient.

[0024] Preferably, the specific process of S4 is as follows:

[0025] Extract within the region of interest The original gradient profile perpendicular to the edge direction will be the first... The original gradient profile is denoted as Calculate its average gradient profile The calculation formula is:

[0026] ;

[0027] in, This represents the averaged gradient profile. This indicates the total number of profiles used for averaging; Indicates the first Original gradient profile;

[0028] Construct the objective function of the sum of squared residuals between the observed data and the theoretical model, and solve for the optimal parameters:

[0029] ;

[0030] in, This represents the optimized estimate of the optimal parameter vector. This represents the parameter variables that minimize the objective function; Indicates the index of the sampling point; Indicates at the sampling point The observed average gradient value at the location; Indicates at the sampling point The model parameters are used to process the parameters. The calculated theoretical gradient value;

[0031] The objective function is iteratively optimized using the Levenberg-Marquardt nonlinear least squares optimization algorithm until the residuals converge, and the optimal solution vector is obtained. Extracting the optimal sub-pixel edge position .

[0032] Preferably, the specific process of S5 is as follows:

[0033] Obtain the time series of sub-pixel edge positions and calculate the median absolute deviation of the series;

[0034] For time Calculate the statistical residual at the edge position. ,like If the value exceeds the preset threshold, it is determined to be an abnormal frame;

[0035] The proportion of abnormal frames within the statistical sliding window ,like Greater than the preset ratio threshold Then the median filter formula is used to calculate the time. final coordinates :

[0036] ;

[0037] like Less than or equal to the preset ratio threshold Then the moving average filter formula is used to calculate the time. final coordinates :

[0038] ;

[0039] in, Indicates the time after filtering. The edge coordinates; This indicates the median-taking operation; This represents the original value of the subpixel edge position before filtering; This represents half the length of the median filter window; This represents half the length of the mean filter window; This indicates the index offset within the window.

[0040] Preferably, in step S6, the radial direction is calculated. The specific steps for axial, or vertical, displacement are as follows:

[0041] The upper edge of the rotor body is selected as the feature region, and the vertical coordinates of the top edge of the rotor in the current frame are calculated using the sub-pixel edge localization method. ;

[0042] Set the edge position of the initial reference frame to Calculate the radial direction according to the following formula Physical vibration displacement of the shaft :

[0043] ;

[0044] in, Indicates radial The physical vibration displacement of the shaft; Indicates the vertical pixel-to-physical scale conversion factor; This represents the sub-pixel coordinates of the top edge measured at the current moment; Indicates the reference coordinates.

[0045] Preferably, in step S6, the axial direction is calculated. The specific steps for horizontal movement of the axis are as follows:

[0046] The vertical edge of the rotor body side is selected as the feature region, and the horizontal coordinates of the rotor side edge in the current frame are calculated using the sub-pixel edge localization method. ;

[0047] Set the edge position of the initial reference frame to Calculate the axial direction according to the following formula Physical displacement of the shaft :

[0048] ;

[0049] in, Indicates axial direction The physical displacement of the shaft; Indicates the horizontal pixel-to-physical scale conversion factor; This represents the sub-pixel coordinates of the side edge measured at the current moment; Indicates the reference coordinates.

[0050] Preferably, in step S6, the radial direction is calculated. The specific steps for the axis, i.e., the depth, displacement are as follows:

[0051] Using the sub-pixel edge localization method, the left edge coordinates of the cooperative markers on the rotor end face are extracted respectively. and right edge coordinates ;

[0052] Calculate the instantaneous pixel width of the marker in the current frame. ;

[0053] Obtain the pixel width of the marker in the initial reference state. Calculate the radial direction according to the following formula Physical displacement of the axis :

[0054] ;

[0055] in, Indicates radial The physical displacement of the shaft; Indicates the linear scale conversion factor in the depth direction; This indicates the pixel width of the marker measured at the current moment; This represents the pixel width of the marker in the initial reference state;

[0056] The depth direction linear scale conversion factor satisfy ,in This represents the object distance in the initial state.

[0057] Preferably, the conversion factor obtained by physical scale calibration in step S7 includes the following specific steps:

[0058] The checkerboard calibration method was used to capture precise checkerboard images at different object distances;

[0059] By establishing a mapping relationship between the pixel coordinate system and the world coordinate system through corner detection, the vertical pixel-physical scale conversion factor is obtained through fitting. and horizontal pixel-to-physical scale conversion factor ;

[0060] Through multi-point object distance calibration experiments, the relationship between pixel width change and depth displacement was fitted to obtain the depth direction scale conversion factor. .

[0061] Compared with the prior art, the present invention has the following beneficial effects:

[0062] 1. Breakthrough in overcoming the limitations of high-speed motion blur: This invention no longer relies solely on data-driven image processing, but instead starts from the physical imaging mechanism, adaptively learning and subtracting blur effects through a composite gradient model, significantly improving edge positioning accuracy under high-speed rotation conditions;

[0063] 2. Achieved monocular 3D decoupling: By cleverly utilizing the geometric projection characteristics of cooperative markers, the problem of monocular vision being unable to perceive depth displacement was solved. High-precision, synchronous, and non-contact measurement of three-dimensional vibration displacement in the radial X, radial Y, and axial Z directions can be achieved with just a single industrial camera, which greatly reduces hardware costs and on-site deployment difficulty.

[0064] 3. High robustness: Through image enhancement, robust profile fitting and adaptive temporal filtering, it effectively suppresses lighting interference and random noise in industrial environments. Attached Figure Description

[0065] To more clearly illustrate the technical solutions of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the following description is only one embodiment of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0066] Figure 1 This is a schematic diagram of the triaxial vibration displacement feature extraction method based on video image analysis of the present invention, showing the overall system architecture and data flow.

[0067] Figure 2 This is a schematic diagram of the rotor triaxial vibration displacement measurement principle based on monocular vision of the present invention, which shows the core technical route from image acquisition to physical displacement output.

[0068] Figure 3 This is a flowchart of the subpixel edge localization method of the present invention, which shows in detail the logical steps of preprocessing, gradient calculation and model fitting.

[0069] Figure 4 This is a schematic diagram of the triaxial vibration component calculation process based on edge coordinate extraction in this invention, illustrating the logical relationship from sub-pixel coordinates to X / Y / Z triaxial decoupling.

[0070] Figure 5 This is a schematic diagram of the radial X-direction vibration displacement feature extraction method based on video image analysis according to the present invention, which illustrates the principle of calculating depth displacement by utilizing the change in the width of the marker.

[0071] Figure 6 This is a schematic diagram of the radial Y-direction vibration displacement feature extraction method based on video image analysis according to the present invention, which illustrates the principle of calculating vertical displacement using the top edge.

[0072] Figure 7 This is a schematic diagram of the axial Z-direction vibration displacement feature extraction method based on video image analysis of the present invention, which illustrates the principle of calculating axial displacement using the side edge.

[0073] Figure 8 The following are simulation and ROI extraction effect diagrams in the embodiments of the present invention; wherein, (a) is a time sequence image of rotor multi-directional vibration simulation; (b) is a schematic diagram of ROI selection and sub-pixel edge extraction of the upper edge of the rotor; (c) is a schematic diagram of ROI selection and sub-pixel edge extraction of the left edge of the rotor; and (d) is a schematic diagram of ROI selection and left and right sub-pixel edge extraction of the green marker of the rotor.

[0074] Figure 9 This is a schematic diagram of the physical scale calibration experimental device and process in this invention; wherein, (a) is a physical picture of the calibration experimental platform, and (b)-(d) are checkerboard calibration images at different object distances of 265mm, 315mm and 365mm respectively.

[0075] Figure 10 Lateral dimension factor in embodiments of the present invention Follow The experimental data on the change of directional object distance and the comparison plot with the linear fitting are used to verify the rationality of the linear scale transformation. Detailed Implementation

[0076] The present invention will be further described below with reference to embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0077] Please see Figure 1 and Figure 2The method provided by this invention mainly includes the entire process from information acquisition, image preprocessing, sub-pixel positioning to three-way decoupling and output.

[0078] S1. Construction of a High-Speed ​​Dynamic Vision Acquisition System

[0079] like Figure 1 As shown, a high-resolution industrial camera (such as a CMOS sensor) and a low-distortion fixed-focus lens are selected. The camera is fixed on a precision adjustment bracket, and its posture is adjusted so that the camera's optical axis is strictly perpendicular to the central axis of the rotor side, ensuring that the geometric projection relationship of the imaging plane is simplified.

[0080] The key lies in the placement of the cooperative targets: to address the problem of poor depth observability in monocular vision, targets with known physical geometric dimensions (width) are attached to the rotor end face. High-contrast markers (e.g.) Figure 8 (The green rectangular block against a black background shown in (d)). This marker serves as the physical reference for calculating the radial X-axis (depth direction) displacement.

[0081] Set the Region of Interest (ROI): such as Figure 8 As shown in (b)-(d), to capture high-frequency vibration characteristics, it is necessary to overcome the full-frame frame rate limitation of the camera. By setting the ROI, only specific pixel rows / columns (e.g., a 928x680 pixel area) containing the rotor edge contour and markers are read. This significantly reduces the data transmission bandwidth, increasing the effective sampling frame rate to 120 fps or higher. Simultaneously, a high-brightness, flicker-free ring LED light source is used for coaxial illumination to suppress specular reflections.

[0082] S2, Image Preprocessing and Gradient Field Construction

[0083] Combination Figure 3 The process shown involves fine preprocessing of the acquired images:

[0084] First, the acquired color RGB video stream is converted to a grayscale image, and a Gaussian filter (such as...) is applied. nuclear, The image is smoothed to suppress sensor thermal noise.

[0085] Secondly, the Limiting Contrast Adaptive Histogram Equalization (CLAHE) algorithm is employed. The image is divided into several non-overlapping sub-blocks, and the gray-level histogram of each sub-block is calculated. Then, the histogram is cropped to obtain the cropped histogram. Construct the cumulative distribution function of the clipping histogram. And map the pixels according to the following formula:

[0086] ;

[0087] in, This represents the enhanced pixel grayscale value; Represents the pixel grayscale value in the original image; Represents the coordinates of a pixel; and These represent the lower and upper limits of the target enhanced grayscale range, respectively; This represents the cumulative distribution function of the current sub-block; This indicates the total number of pixels within the current sub-block. This step significantly enhances the grayscale transition characteristics of the rotor edge region.

[0088] Finally, the Sobel operator is used to extract gradient information in a specific direction. Figure 8 The rotor side edge (measured along the Z-axis) in (c) and Figure 8 Calculate the horizontal gradient of the left and right edges of the marker in (d) (measured along the X-axis). ;for Figure 8 The vertical gradient is calculated at the top edge of the rotor in (b) (measured along the Y-axis). .

[0089] S3-S4, Physically Driven Subpixel Edge Localization (PIMSM)

[0090] like Figure 3 As shown, this is the core algorithm of the present invention. To address the trapezoidal blurring of edges caused by high-speed rotation, a composite gradient model of "Gaussian + motion blur" is constructed.

[0091] Assuming the edges exhibit a Gaussian gradient distribution in a static state, and that the rotor edges sweep across at a uniform speed during the exposure time to form a rectangular blur kernel, the actual observed gradient is the result of the convolution of these two gradients. The mathematical expression derived in this invention is as follows:

[0092] ;

[0093] in, Indicates pixel coordinates Theoretical value of the edge gradient at the location; This represents the vector of model parameters to be solved; Indicates the gradient magnitude; Indicates the center position of the subpixel edge; This represents the equivalent motion fuzziness width (i.e., the degree of fuzziness to be solved). Indicates the standard deviation of optical blur; This indicates the gradient baseline offset.

[0094] Employing a "profile fitting" strategy: extracting within the ROI Calculate the average gradient profile of the original gradient profile perpendicular to the edge direction. To suppress random noise.

[0095] Construct the residual sum of squares (SSR) objective function and perform iterative optimization using the Levenberg-Marquardt (LM) nonlinear least squares optimization algorithm:

[0096] ;

[0097] The algorithm automatically finds the best (Location) and (Blur width) enables joint estimation of blur and position, outputting the optimal sub-pixel edge position. .

[0098] S5, Time-series robust filtering

[0099] Obtaining time series of edge locations Calculate the median absolute deviation (MAD) of the sequence. If at time... Statistical residuals If it is, then it is determined to be an abnormal frame.

[0100] The proportion of abnormal frames within the statistical sliding window .like Greater than the preset threshold (e.g., 5%), then median filtering is used: Otherwise, use moving average filtering: This step ensures the smoothness and robustness of the measurement results over time.

[0101] S6. Specific implementation of triaxial vibration decoupling

[0102] Figure 4 The process of solving the three-dimensional vibration components based on edge coordinate extraction is demonstrated. This invention utilizes geometric constraints to decouple two-dimensional image information into three-dimensional spatial displacement.

[0103] 1. Calculation of radial Y-axis (vertical displacement):

[0104] Combination Figure 6 and Figure 8 In (b), the upper edge of the rotor body is used as a feature. Since the camera optical axis is perpendicular to the side of the rotor, the vertical vibration of the rotor is directly linearly mapped to the vertical pixel displacement of the edge in the image.

[0105] The algorithm in step S4 is used to calculate the vertical sub-pixel coordinates of the top edge of the current frame. Set the edge position of the initial reference frame to... Then radial Physical vibration displacement of the shaft The calculation is as follows:

[0106] ;

[0107] in, Vertical pixel-to-physical scale conversion factor (unit: ( / pixel), obtained through calibration.

[0108] 2. Calculation of the Z-axis (horizontal movement):

[0109] Combination Figure 7 and Figure 8 In (c), the vertical edges of the rotor's side are used as features. The axial movement of the rotor is directly linearly mapped to the horizontal pixel displacement of the edges in the image.

[0110] Calculate the horizontal subpixel coordinates of the side edges of the current frame. Set the edge position of the initial reference frame to... Then axial Physical displacement of the shaft The calculation is as follows:

[0111] ;

[0112] in, This is the horizontal pixel-to-physical scale conversion factor, obtained through calibration.

[0113] 3. Calculation of radial X-axis (depth displacement):

[0114] Combination Figure 5 and Figure 8 In the middle (d), this is the core difficulty of monocular measurement. This invention utilizes the "near is larger, far is smaller" principle of perspective imaging.

[0115] Extract the left edge coordinates of the cooperative markers respectively and right edge coordinates Calculate the instantaneous pixel width of the marker in the current frame. .

[0116] Let the initial object distance be... The pixel width of the marker in the initial reference state is When the rotor undergoes forward and backward displacement When the physical displacement is calculated, the formula is:

[0117] ;

[0118] in, This represents the linear scale transformation factor in the depth direction. This formula shows that by detecting minute changes in the width of a marker in an image, the vibrational displacement in the depth direction can be inferred. .

[0119] S7, Physical Scale Calibration

[0120] To obtain the formula above The factor is used to convert pixel displacement into physical displacement. In this embodiment, the checkerboard calibration method is used.

[0121] Combination Figure 9 As shown, a calibration experimental platform was constructed. A precision checkerboard pattern was fixed to the moving stage, and the camera was fixed to the support. The position of the moving stage was changed, and the camera was mounted at different object distances (e.g., ...). Figure 9 The checkerboard images were taken at positions of 265mm, 315mm, and 365mm as shown in (b)-(d).

[0122] The PIMSM algorithm is used to extract the pixel distances of the checkerboard corners or edges. Combined with the known physical dimensions of the checkerboard (e.g., 5mm), a mapping relationship between the pixel coordinate system and the world coordinate system is established, and the vertical pixel-physical scale transformation factor is obtained by fitting the data. (For the Y-axis) and horizontal pixel-to-physical scale conversion factor (For the Z-axis).

[0123] Combination Figure 10 As shown, for depth calibration along the radial X-axis, the relationship between pixel width change and depth displacement was obtained through multi-point object distance fitting experiments. Figure 10 As shown in the data points, the size factor exhibits a non-linear change with increasing object distance (consistent with the "nearer objects appear larger, farther objects appear smaller" imaging principle). However, within the range of minute vibrations (e.g., ...), the size factor remains constant. Figure 10 The curve (within the magnified 319mm-321mm range) can be approximated as a straight line. Based on this, the present invention employs a linear approximation formula. ,in The slope of this linear region is used to simplify the calculation process while ensuring micrometer-level accuracy and ensuring the physical consistency of the measurement.

[0124] Through the above steps, this invention can output, using only a single video stream, such as... Figure 4 The time-synchronized X, Y, and Z-axis physical vibration waveforms shown at the bottom are used for subsequent fault diagnosis and analysis.

[0125] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0126] While the specific embodiments of the present invention have been described above, they are not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A physical driving measurement method of monocular three-dimensional dynamic displacement of a rotating machine, characterized in that, The method comprises the following steps: S1, based on the cooperation marker with known physical size arranged on the surface of the rotating machine rotor, a time sequence video stream containing rotor vibration characteristics is collected by an industrial camera; S2, the collected time sequence video stream is pre-processed to obtain a region of interest and enhance edge gradient features, thereby obtaining a high-quality rotor edge gray image, and gradient information in a specific direction is extracted; S3, a physically driven Gaussian and motion blur compound gradient model is constructed based on the gradient information, and the motion blur width is taken as an endogenous variable to describe the gray gradient distribution of the edge under high-speed rotation; S4, based on the compound gradient model, a nonlinear optimization algorithm is used to jointly iteratively solve the gradient profile of the rotor edge, and a sub-pixel edge coordinate sequence containing position information and blur information is solved; S5, the sub-pixel edge coordinate sequence is subjected to time sequence robust filtering processing to eliminate abnormal values and smooth jitter, thereby obtaining a high-precision edge position time sequence; S6, based on monocular imaging geometry principle, a multi-direction vibration decoupling model is established, and the high-precision edge position timing is used to respectively calculate the pixel displacement of the rotor in the radial direction axial direction, radial direction axial direction, radial direction and axial direction axial direction S7, a conversion factor obtained by physical scale calibration is used to convert the pixel displacement into physical space displacement, and time-synchronized three-dimensional vibration waveforms are output.

2. The physical driving measurement method of monocular three-dimensional dynamic displacement of a rotating machine according to claim 1, characterized in that: The specific steps of S2 are as follows: The collected color video stream is converted into a gray image, and a Gaussian filter is applied to smooth the image; Adopting the limited contrast self-adapting histogram equalization algorithm to enhance the local contrast: dividing the image into several sub-blocks, calculating the gray scale histogram of each sub-block and cutting, obtaining the cut histogram ; Constructing a cumulative distribution function of a histogram and mapping the pixels according to the following equation: ； wherein, represents the enhanced pixel gray value; represents the pixel gray value in the original image; represents the coordinates of the pixel; represents the lower limit value of the target enhanced gray interval; represents the upper limit value of the target enhanced gray interval; represents the cumulative distribution function of the current sub-block; represents the total number of pixels within the current sub-block; The Sobel operator is used to calculate the horizontal direction gradient and the vertical direction gradient respectively, and the gradient information in a specific direction is extracted.

3. The physical drive measurement method of monocular three-dimensional dynamic displacement of a rotating machine according to claim 1, characterized in that: The mathematical expression of the physically driven Gaussian and motion blur compound gradient model based on the gradient information is as follows: ； wherein, denotes the theoretical value of the edge gradient at the pixel coordinate ; denotes the model parameter vector to be solved; denotes the magnitude coefficient of the gradient; denotes the Gaussian error function; denotes the pixel coordinate position along the gradient direction; denotes the center position of the sub-pixel edge; denotes the equivalent motion blur width; denotes the standard deviation of the optical blur; denotes the baseline offset of the gradient.

4. The physical drive measurement method of monocular three-dimensional dynamic displacement of a rotating machine according to claim 1, characterized in that: The specific process of S4 is as follows: extracting The original gradient profile in the column direction perpendicular to the edge direction is denoted as The original gradient profile in the column direction perpendicular to the edge direction is denoted as The average gradient profile of the original gradient profile in the column direction perpendicular to the edge direction is calculated as The calculation formula is: ； wherein, represents the averaged gradient profile; represents the total number of profiles for averaging; represents the first original gradient profile; A residual sum of squares objective function between observation data and a theoretical model is constructed, and optimal parameters are solved: ； wherein, represents the optimized optimal parameter vector estimate; represents the parameter argument that minimizes the objective function; represents the index of the sampling point; represents the observed average gradient value at the sampling point ; represents the theoretical gradient value calculated from the model parameters at the sampling point ; The Levenberg-Marquardt nonlinear least squares optimization algorithm is used to iteratively optimize the objective function until the residual converges, and the optimal subpixel edge position is extracted from the optimal solution vector .​ 5. The physical drive measurement method of monocular three-dimensional dynamic displacement of a rotating machine according to claim 1, characterized in that: The specific process of S5 is as follows: The time sequence of the sub-pixel edge position is obtained, and the median absolute deviation of the sequence is calculated; For the edge position of time point , its statistical residual error is calculated, and if is greater than a preset determination threshold, it is determined as an abnormal frame. The proportion of abnormal frames in the statistical sliding window If The proportion is greater than a preset proportion threshold The final coordinates of the moment are calculated by using a median filter formula : ； If is less than or equal to a preset proportion threshold , then the final coordinate of the moment is calculated by using a sliding mean filtering formula : ； wherein, represents the edge coordinates after filtering; represents the edge coordinates after filtering; represents the median operation; represents the sub-pixel edge position raw value before filtering; represents half of the median filter window length; represents half of the mean filter window length; represents the index offset within the window.

6. A physical drive measurement method of monocular three-dimensional dynamic displacement of a rotary machine according to claim 1, characterized in that: In the S6, the radial The specific steps of the vertical displacement are as follows: The upper edge of the rotor body is selected as a feature region, and a sub-pixel edge positioning method is used to calculate the vertical coordinate of the rotor top edge of the current frame ; The edge position of the initial reference frame is set as The radial physical vibration displacement of the shaft is calculated according to the following formula : ； wherein, represents a radial physical vibration displacement of the shaft; represents a longitudinal pixel-physical scale conversion factor; represents a top edge sub-pixel coordinate measured at the current time instant; represents a reference datum coordinate.

7. A physical drive measurement method of monocular three-dimensional dynamic displacement of a rotating machine according to claim 1, characterized in that: In the S6, the axial direction is solved The specific steps of the horizontal movement are as follows: The vertical edge of the rotor body side surface is selected as a feature region, and the horizontal coordinate of the rotor side surface edge of the current frame is calculated by using the sub-pixel edge positioning method ; The edge position of the initial reference frame is set as The axial physical displacement of the shaft is calculated according to the following formula: ； wherein, represents an axial physical runout displacement of the shaft; represents a lateral pixel-physical scale conversion factor; represents a side edge sub-pixel coordinate measured at the current time instant; represents a reference datum coordinate.

8. A physical drive measurement method of monocular three-dimensional dynamic displacement of a rotary machine according to claim 1, characterized in that: In the S6, the radial The specific steps of the axial, i.e. depth, forward and backward displacement are as follows: By using the sub-pixel edge positioning method, the left edge coordinate and the right edge coordinate of the cooperation marker on the rotor end face are extracted respectively and ; calculating the instantaneous pixel width of the marker in the current frame ; Acquiring a pixel width of the marker in the initial reference state , the radial physical displacement of the axis is calculated according to the following formula : ； wherein, represents a radial physical displacement of the axis; represents a depth direction linear scale conversion factor; represents a pixel width of the marker measured at the current time; represents a pixel width of the marker in the initial reference state; The depth direction linear scale conversion factor satisfies wherein denotes the object distance in the initial state.

9. A physical drive measurement method of monocular three-dimensional dynamic displacement of a rotating machine according to claim 1, characterized in that: The conversion factor obtained by physical scale calibration in S7 comprises the following specific steps: A chessboard calibration method is used to shoot a precision chessboard image at different object distances; By establishing a mapping relationship between the pixel coordinate system and the world coordinate system through corner detection, the vertical pixel-physical scale conversion factor is obtained through fitting. and horizontal pixel-to-physical scale conversion factor ; The relationship between the change of pixel width and depth displacement is fitted through a multi-point distance calibration experiment to obtain a depth direction scale conversion factor .

Images