Intelligent detection method for gear clearance of speed reducer based on machine vision

Abstract

The application discloses a machine vision-based intelligent detection method for gear clearance of a speed reducer and belongs to the technical field of machine vision detection. The method comprises the following steps: acquiring a continuous image sequence of a gear meshing area of a target speed reducer in an unloaded operation state, and generating an image entropy variation graph through dynamic tracking of a tooth profile edge; reversely positioning a contact state conversion time between a driving wheel and a driven wheel of a driving shaft according to an extreme value point of entropy value fluctuation in the entropy variation graph; extracting an angular displacement difference value and mapping the angular displacement difference value to a normal direction of a tooth surface to obtain a first meshing clearance reference value; after a rated torque is loaded, a second meshing clearance reference value is repeatedly acquired and tracked; and the first and second reference values are fused to output a dynamic meshing clearance of the target speed reducer.

Description

Technical Field

[0001] This invention relates to the field of machine vision inspection technology, specifically to a machine vision-based intelligent detection method for gear backlash in reducers. Background Technology

[0002] Gear backlash in a speed reducer directly affects the rotational accuracy, impact noise, and service life of the transmission system. Accurate detection of backlash is essential for ensuring safe equipment operation. Existing gear backlash measurement methods are mainly divided into mechanical contact methods and machine vision-based non-contact methods. Mechanical contact methods measure tooth backlash in a stopped state using feeler gauges, lead weights, or dial indicators. This process relies on manual judgment, is inefficient, and cannot achieve online detection during operation. Machine vision-based methods obtain backlash values ​​by capturing images of the gear meshing area, extracting the tooth profile edges, and calculating the distances. These methods offer the advantages of being non-contact and deployable online. However, these methods typically focus on the relative positions of tooth profile edges within a single frame of an image or directly calculate the average distance across multiple frames, lacking the ability to perform time-series analysis of the dynamic evolution of the tooth surface contact state during meshing. When gears are running continuously, the tooth surfaces of the driving gear and the driven gear undergo a periodic contact-separation-re-contact process. If the instantaneous moments when the tooth surfaces separate from and re-enter contact cannot be accurately defined, the angular displacement data used for clearance calculation cannot be captured at the correct phase node. The clearance value obtained is actually the result of a mixture of multiple states, and there is a fundamental moment drift and measurement fluctuation.

[0003] The tooth backlash of gears under no-load conditions only reflects the geometric clearance formed during manufacturing and assembly. However, in actual driving, the rated torque of the reducer causes elastic compression deformation of the tooth surface, and the temperature rise during operation also causes thermal expansion of the gears and housing, resulting in significant changes in the tooth backlash. Currently, most common visual inspection solutions operate under no-load or stable light-load conditions, lacking a fusion mechanism to incorporate deformation information under loaded conditions into the backlash model, and therefore cannot output the dynamic backlash value under real-world operating conditions. How to automatically identify the critical moment of gear contact state transition using continuous image sequences and integrate visual inspection results from both no-load and rated-load conditions has become a key technical challenge to overcome in achieving high-precision intelligent detection of gear backlash. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent detection method for gear clearance in reducers based on machine vision, so as to solve the problems of existing technologies that are difficult to automatically and accurately capture the moment of gear contact state transition and cannot integrate no-load and loaded conditions to obtain accurate dynamic clearance.

[0005] The objective of this invention can be achieved through the following technical solutions: This invention provides a machine vision-based intelligent detection method for gear backlash in a speed reducer, comprising: A continuous image sequence of the gear meshing area of ​​the target reducer under no-load operation is obtained, and the continuous image sequence is subjected to dynamic tracking processing of the tooth profile edge to generate an image entropy change map characterizing the relative displacement of the gear pair. Based on the extreme points of entropy fluctuation in the image entropy change spectrum, the moment of contact state transition between the driving wheel and the driven wheel of the target reducer is located in reverse. Based on the contact state transition time, the angular displacement difference between the driving wheel and the driven wheel of the drive shaft is extracted, and the angular displacement difference is mapped to the tooth surface normal direction to obtain the first meshing clearance reference value; After the target reducer is loaded with rated torque, the acquisition of the continuous image sequence and the dynamic tracking processing of the tooth profile edge are repeated to generate the second meshing clearance reference value under the loading state; By combining the first meshing clearance reference value and the second meshing clearance reference value, the dynamic meshing clearance of the target reducer is output.

[0006] The process of dynamically tracking the tooth profile edges of the continuous image sequence and generating an image entropy change map includes: dividing the continuous image sequence into multiple image sub-frame sets according to the acquisition time sequence, with each image sub-frame set corresponding to one tooth profile motion cycle; performing edge enhancement filtering on each frame in each image sub-frame set to extract the tooth profile edge point set of the driving wheel and the driven wheel; calculating the minimum Euclidean distance between the tooth profile edge point set of the driving wheel and the driven wheel, and arranging the minimum Euclidean distance into a distance fluctuation sequence according to time sequence; calculating the sliding window entropy value of the distance fluctuation sequence to obtain the image entropy value within each sliding window, and expanding the image entropy value along the time axis to form the image entropy change map. Therefore, through dynamic tracking and entropy representation of the tooth profile edges, the minute relative displacement of the gear pair can be converted into intuitive entropy change fluctuations, significantly improving the detection sensitivity of gap changes.

[0007] As a preferred approach, when extracting the edge point sets of the active wheel tooth profile and the driven wheel tooth profile, the image after edge enhancement filtering is binarized and segmented to separate the active wheel tooth profile region and the driven wheel tooth profile region. Within each tooth profile region, the grayscale gradient change is scanned along the radial direction, and the positions of pixels whose gradient amplitude exceeds a preset threshold are recorded, forming the active wheel tooth profile edge point set and the driven wheel tooth profile edge point set respectively. Then, sub-pixel interpolation processing is performed on the two sets of edge points to improve the coordinate accuracy of the edge points to the sub-pixel level, thereby ensuring the accuracy of subsequent distance calculations.

[0008] In acquiring a continuous image sequence, one technical solution of the present invention is as follows: An industrial camera is mounted on the housing of the target reducer, with the optical axis of the industrial camera perpendicular to the normal direction of the gear meshing area; pulse signals from the rotary encoder of the drive shaft are synchronously acquired, and the exposure trigger time of the industrial camera is aligned with the rising edge of the rotary encoder pulse signal; the acquisition frame rate of the industrial camera is set to an integer multiple of the drive shaft rotation frequency, and multiple equally spaced image frames are acquired within each tooth profile motion cycle to form the continuous image sequence. This method eliminates measurement errors introduced by motion blur and temporal asynchrony.

[0009] The specific implementation of locating the contact state transition moment based on the extreme points of entropy fluctuation in the image entropy change spectrum is as follows: Scan the image entropy change spectrum and extract the temporal positions of all local maxima and minima; pair adjacent local maxima and minima to form multiple entropy fluctuation intervals; calculate the rate of change of entropy within each interval, and locate the moment when the rate of change crosses zero as a candidate transition moment; select the moment when the rate of change jumps from positive to negative from the candidate transition moments as the first transition moment when the driving and driven gear tooth surfaces switch from a contact state to a separation state, and select the moment when the rate of change jumps from negative to positive as the second transition moment when the separation state switches to a contact state. This method accurately captures the instantaneous moments of tooth surface contact and separation through the sign change of the entropy change rate, avoiding the lag of traditional threshold judgment.

[0010] For the extraction of local extrema, a sliding detection window can be defined on the image entropy change map, with the width of the sliding detection window set to one-quarter of the tooth profile motion period. Within each sliding detection window, the entropy value of the center point is compared with the entropy values ​​of the left and right adjacent points. If the entropy value of the center point is greater than both the entropy values ​​of the left and right adjacent points, it is marked as a local maximum point; if the entropy value of the center point is less than both the entropy values ​​of the left and right adjacent points, it is marked as a local minimum point. The coordinates of each extrema point on the time axis are recorded to form the temporal position.

[0011] Preferably, the moment when the rate of change crosses zero is located by fitting a cubic spline curve of the entropy value over time within each entropy fluctuation interval and solving for the point where the derivative of the cubic spline curve is zero, thereby obtaining the high temporal resolution conversion moment of the subsampling interval.

[0012] The process of extracting the angular displacement difference and obtaining the first meshing clearance reference value based on the contact state transition time includes: acquiring the first rotation angle encoding value of the driving wheel at the first transition time, acquiring the second rotation angle encoding value of the driving wheel at the second transition time, and calculating the difference between the two to obtain the angular displacement change of the driving wheel of the drive shaft; extracting the first tooth position image of the driven wheel corresponding to the first transition time and the second tooth position image of the driven wheel corresponding to the second transition time from the continuous image sequence, and calculating the angular displacement change of the driven wheel through a phase correlation algorithm; calculating the difference between the angular displacement change of the driving wheel and the angular displacement change of the driven wheel to obtain the angular displacement difference; obtaining the angle between the tooth surface normal direction and the pitch circle tangent direction, multiplying the angular displacement difference by the base circle radius of the driving wheel, and then multiplying by the cosine value of the angle to obtain the first meshing clearance reference value. By using the fusion of rotation angle encoding and image phase correlation algorithm, synchronous and precise measurement of the angular displacement of the driving and driven wheels is achieved, and it is directly mapped to the tooth surface normal direction, making the clearance reference value closer to the geometric definition of the actual meshing clearance.

[0013] Among them, the phase correlation algorithm calculates the cross power spectrum of the first tooth position image and the second tooth position image of the driven wheel in the frequency domain, and extracts the peak coordinates as the angular displacement change of the driven wheel after performing inverse Fourier transform on the cross power spectrum. This method is not sensitive to changes in image brightness and contrast and has strong robustness.

[0014] The steps for generating the second meshing clearance reference value after the target reducer is loaded with rated torque include: applying a rated torque load to the output shaft of the target reducer and maintaining the rated torque load for more than one tooth profile motion cycle; during the maintenance of the rated torque load, re-acquiring a continuous image sequence of the loading state of the gear meshing area; performing the edge enhancement filtering and minimum Euclidean distance calculation on the continuous image sequence of the loading state to generate a loading state distance fluctuation sequence; performing the sliding window entropy value calculation on the loading state distance fluctuation sequence to generate a loading state image entropy change spectrum; relocating the contact state transition time under the loading state based on the loading state image entropy change spectrum, and re-extracting the angular displacement difference under the loading state to map and obtain the second meshing clearance reference value. Separate detection of the no-load and loaded states can reflect the changing characteristics of the clearance under different torques, making the subsequent fusion results more representative of the actual dynamic working conditions.

[0015] The process of fusing the first meshing clearance reference value and the second meshing clearance reference value and outputting the dynamic meshing clearance amount includes: multiplying the first meshing clearance reference value by an unloaded weighting coefficient to obtain an unloaded contribution component; multiplying the second meshing clearance reference value by a loaded weighting coefficient to obtain a loaded contribution component; the sum of the unloaded weighting coefficient and the loaded weighting coefficient is one; adding the unloaded contribution component and the loaded contribution component to obtain the weighted fused clearance amount; obtaining the temperature detection value of the target reducer in the current operating cycle; obtaining the thermal expansion compensation factor from a table based on the temperature detection value; and multiplying the weighted fused clearance amount by the thermal expansion compensation factor to obtain the dynamic meshing clearance amount. Through weighted fusion and thermal expansion compensation, the bias caused by single-state detection is avoided, the interference of structural deformation caused by temperature changes on clearance measurement is eliminated, and finally, an accurate and reliable dynamic meshing clearance amount is output.

[0016] The no-load weight coefficient and the loading weight coefficient can be dynamically adjusted according to the rated load rate and operating conditions of the target reducer. The ratio of the no-load weight coefficient to the loading weight coefficient is inversely proportional to the current torque load value. The greater the load, the higher the weight of the loading state.

[0017] The preferred method for obtaining temperature detection values ​​is as follows: temperature sensors are arranged on the bearing housing and housing of the target reducer to collect the bearing housing temperature value and housing temperature value in real time; the average value of the bearing housing temperature value and housing temperature value is calculated as the temperature detection value; a correspondence table between temperature ranges and thermal expansion compensation factors is established in advance, with each temperature range corresponding to a pre-calibrated thermal expansion compensation factor; according to the target temperature range where the temperature detection value is located, the corresponding thermal expansion compensation factor is read from the correspondence table; if the temperature detection value falls exactly on the boundary between two temperature ranges, the thermal expansion compensation factor is calculated using linear interpolation, thereby ensuring the continuity and accuracy of compensation.

[0018] The beneficial effects of this invention are: Image entropy variation maps are generated by dynamically tracking the tooth profile edges. The contact state transition time is located by reverse-directing the extreme points of entropy fluctuations, and the angular displacement difference is extracted and mapped to obtain the meshing clearance reference value based on this time. During processing, the continuous image sequence is divided into multiple sub-frame sets according to the tooth profile motion cycle. Edge enhancement filtering is performed on each frame image, and the tooth profile edge point sets of the driving and driven gears are extracted. The minimum Euclidean distance between the two point sets is calculated and arranged temporally as a distance fluctuation sequence. A sliding window entropy value calculation is performed on this sequence to obtain the image entropy variation map changing over time. When the gear tooth surfaces are in a stable contact state, the distance between the tooth profiles remains at a minimum, and the entropy value changes smoothly. At the instant of tooth separation and re-contact, the distance undergoes discontinuous jumps, and the entropy value exhibits sharp local extremes. By scanning the entropy change spectrum to extract all local maxima and minima, these are paired to form entropy fluctuation intervals. The rate of change of entropy within each interval is calculated, and its zero-crossing point is located. The moment when the rate of change jumps from positive to negative is selected as the first transition moment from contact to separation, and the moment when it jumps from negative to positive is selected as the second transition moment from separation to contact. This method shifts the determination of the contact state transition from the distance value itself to the abrupt change in its statistical disorder. Utilizing the amplification effect of entropy on small fluctuations, the capture of the transition moment is no longer affected by image noise and micro-motions at the tooth profile edges. At the precisely obtained transition moment, the change in the rotation angle of the driving wheel encoder is simultaneously read, and the change in the angular displacement of the driven wheel is calculated in the frequency domain using a phase correlation algorithm. The difference between the two is multiplied by the cosine of the angle between the base circle radius of the driving wheel and the tooth surface normal, projecting the angular displacement difference from the pitch circle tangent direction to the tooth surface normal direction. The resulting meshing clearance reference value directly corresponds to the tooth flank normal clearance, skipping the averaging error introduced by the inability to define the separation starting point in traditional methods, and achieving precise characterization of the no-load geometric clearance.

[0019] The meshing clearance reference values ​​under no-load and rated torque loading conditions are obtained separately to form a first reference value and a second reference value, which are then dynamically weighted and fused and thermally expanded to compensate for. After loading the rated torque, the load is maintained for more than one tooth profile motion cycle, and a continuous image sequence under the loading condition is re-acquired and the same tooth profile edge dynamic tracking process is executed to generate an entropy change map of the loading state image. The moment of contact state transition during loading is located, and the second meshing clearance reference value reflecting load deformation is extracted. During fusion, the no-load reference value is multiplied by the no-load weight coefficient, and the loaded reference value is multiplied by the loading weight coefficient. The sum of the two coefficients is one, and the ratio is inversely proportional to the current torque load value. When the torque is low, the output result focuses on the no-load geometric clearance; when it is close to the rated torque, the output result focuses on the loaded deformation clearance, so that the fused clearance amount smoothly follows the actual load state change. On this basis, the average temperature of the bearing housing and the housing is obtained as the temperature detection value. The thermal expansion compensation factor is determined by looking up a table or linear interpolation to correct the weighted fused clearance amount and offset the changes in the gear pair and housing dimensions caused by temperature changes. This integrated compensation mechanism unifies the ideal geometric clearance, load-bearing elastic deformation, and thermal effects into a single output value, enabling the final dynamic meshing clearance to accurately reflect the real-time tooth flank clearance of the reducer under different operating conditions. This overcomes the defect of single-state measurement results deviating from the actual service condition, and improves the adaptability and engineering practical value of clearance detection. Attached Figure Description

[0020] The invention will now be further described with reference to the accompanying drawings.

[0021] Figure 1 This is a flowchart of a machine vision-based intelligent detection method for gear backlash in reducers. Figure 2 This is a flowchart of entropy change analysis of gear meshing distance fluctuation sequence based on synchronous acquisition by industrial cameras; Figure 3 This is a flowchart of the positioning process for the transition of meshing gap contact state under loading conditions based on image entropy change analysis; Figure 4 This is a flowchart for calculating dynamic meshing clearance. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] See Figure 1This invention provides a machine vision-based intelligent detection method for gear backlash in a speed reducer. The method acquires a continuous image sequence of the gear meshing area of ​​the target speed reducer under no-load operation. Dynamic tracking processing of the tooth profile edges is performed on the continuous image sequence to generate an image entropy change spectrum characterizing the relative displacement of the gear pair. Based on the extreme points of entropy fluctuations in the image entropy change spectrum, the contact state transition time between the driving and driven gears of the target speed reducer is located in reverse. The angular displacement difference between the driving and driven gears is extracted based on the contact state transition time and mapped to the tooth surface normal direction to obtain a first meshing backlash reference value. After the target speed reducer is loaded with rated torque, the acquisition of the continuous image sequence and the dynamic tracking processing of the tooth profile edges are repeated to generate a second meshing backlash reference value under loaded conditions. The first and second meshing backlash reference values ​​are fused to output the dynamic meshing backlash of the target speed reducer.

[0024] In specific implementation, please refer to Figure 2 A mounting base is machined on the housing of the target reducer. The industrial camera is fixed to the housing of the target reducer via the mounting base, with the optical axis of the industrial camera perpendicular to the normal direction of the gear tooth surface in the gear meshing area. A global shutter CMOS camera with a telecentric lens is selected for the industrial camera to reduce the impact of perspective distortion on the imaging of the tooth profile edges. The pulse signal of the rotary encoder of the target reducer's drive shaft is synchronously acquired. The A-phase pulse signal of the rotary encoder is connected to the external trigger interface of the industrial camera, and the exposure trigger time of the industrial camera is aligned with the rising edge of the rotary encoder pulse signal. When the rotary encoder outputs a rising edge of a pulse, the industrial camera performs an exposure action, thereby achieving strict synchronization between image acquisition and the rotation angle of the drive shaft. The acquisition frame rate of the industrial camera is set to an integer multiple of the drive shaft rotation frequency. The drive shaft rotation frequency is obtained by dividing the pulse frequency output by the rotary encoder by the number of pulses per revolution of the encoder. The integer multiple is selected as 36 times. 36 equally spaced image frames are acquired in each tooth profile movement cycle. All image frames acquired in multiple consecutive tooth profile movement cycles are arranged in chronological order of acquisition time to form a continuous image sequence.

[0025] When performing dynamic edge tracking on a continuous image sequence, the sequence is divided into multiple image sub-frame sets according to the acquisition time sequence. Each sub-frame set corresponds to one tooth profile motion cycle. The tooth profile motion cycle is determined by dividing the time required for the drive shaft to rotate one revolution by the number of teeth on the driving gear. Each sub-frame set contains 36 images. Edge enhancement filtering is performed on each frame in each sub-frame set using a Gaussian difference filter operator. The Gaussian difference filter operator is generated by subtracting two Gaussian kernel functions with different standard deviations. The standard deviation of the first Gaussian kernel is set to 1.0 pixel, and the standard deviation of the second Gaussian kernel is set to 2.5 pixels. The Gaussian difference filter operator is convolved with the image to enhance the edge gradient response of the driving and driven gear tooth profiles.

[0026] The process of extracting the edge point sets of the driving and driven wheel tooth profiles is as follows: The image after edge enhancement filtering is binarized. The threshold for binarization is adaptively determined using the maximum inter-class variance method. Pixels with gray values ​​greater than the threshold are classified as tooth profile regions, and pixels with gray values ​​less than or equal to the threshold are classified as background regions, thus separating the driving and driven wheel tooth profile regions. Within the driving wheel tooth profile region, with the driving wheel's rotation center as the origin, the gray-level gradient is scanned pixel-by-pixel along the radial direction from the tooth tip circle radius to the tooth root circle radius. The gray-level gradient is calculated using the Sobel operator, and the positions of pixels whose gradient magnitude exceeds a preset threshold (set to 30% of the maximum gradient magnitude) are recorded. These pixel positions form the driving wheel tooth profile edge point set. Within the driven wheel tooth profile region, with the driven wheel's rotation center as the origin, the gray-level gradient is scanned pixel-by-pixel along the radial direction from the tooth tip circle radius to the tooth root circle radius, and the positions of pixels whose gradient magnitude exceeds a preset threshold are recorded, forming the driven wheel tooth profile edge point set. Subpixel interpolation is performed on the edge point sets of the active gear tooth profile and the driven gear tooth profile respectively. The subpixel interpolation process adopts the quadratic surface fitting method, which fits a quadratic surface function in the edge pixel and its eight neighborhood, solves the coordinates of the extreme points of the quadratic surface function, and improves the coordinate accuracy of the edge points to the subpixel level.

[0027] The minimum Euclidean distance between the edge point set of the driving gear tooth profile and the edge point set of the driven gear tooth profile is calculated. For each frame of the image, every sub-pixel coordinate point in the edge point set of the driving gear tooth profile is traversed, and the Euclidean distance from that point to every sub-pixel coordinate point in the edge point set of the driven gear tooth profile is calculated. The minimum Euclidean distance is taken as the minimum Euclidean distance value corresponding to that frame of the image. The minimum Euclidean distance values ​​corresponding to each frame of the image are arranged in temporal order to form a one-dimensional distance fluctuation sequence.

[0028] A sliding window entropy value is calculated for the distance fluctuation sequence. The width of the sliding window is set to 9 sampling points, and the sliding step size is set to 1 sampling point. Starting from the beginning of the distance fluctuation sequence, the information entropy value of the 9 sampling points within the sliding window is calculated. The formula for calculating the information entropy value is: ; in, Indicates the first Image entropy values ​​of a sliding window The index number of the sliding window. Values , The total number of sliding windows. It equals the total number of sampling points in the distance fluctuation sequence minus the window width plus 1; This represents the total number of intervals into which the values ​​of the distance fluctuation sequence are divided equally within the sliding window. Set to 10; The range index number. Values ; For the first The distance value within the sliding window falls into the first... The probability of the i-th interval is calculated as follows: The sliding window falls into the first The number of sampling points in each interval is divided by the width of the sliding window. The image entropy value is calculated for each sliding window. Expanding along the time axis, with the sampling time at the center of the sliding window as the horizontal axis and the image entropy value as the vertical axis... Using the vertical axis, generate an image entropy change map.

[0029] In a specific implementation, the image entropy change spectrum generated in the above embodiment is obtained. The horizontal axis of the image entropy change spectrum is the time axis, with the unit of the time axis being milliseconds, and the vertical axis is the image entropy value. When scanning the image entropy change spectrum, a sliding detection window is defined on the image entropy change spectrum. The width of the sliding detection window is set to one-quarter of the tooth profile motion cycle. The tooth profile motion cycle is determined by dividing the time required for the drive shaft to rotate one revolution by the number of teeth on the driving gear. If the number of teeth on the driving gear is 20 and the rotation frequency of the drive shaft is 10 Hz, then the tooth profile motion cycle is 5 milliseconds, and the width of the sliding detection window is set to 1.25 milliseconds. The sliding detection window moves along the time axis of the image entropy change spectrum with a step size of one sampling point each time. The sampling point time interval of the image entropy change spectrum is determined by the sliding step size when calculating the entropy value of the sliding window.

[0030] Within each sliding detection window, the sampling point corresponding to the center position of the sliding detection window is taken as the center point, and the image entropy value of the center point is denoted as... The image entropy value of the left adjacent point is denoted as The image entropy value of the adjacent point on the right is denoted as .Compare and , and The size relationship, if and If the center point is marked as a local maximum point; if and If the center point is a local minimum, then the local minimum point is marked as such. After the sliding detection window traverses the entire time axis range of the image entropy change spectrum, several local maxima and several local minima are obtained. The coordinates of each local maximum point on the time axis of the image entropy change spectrum are recorded, and the set of temporal positions of the local maxima points is denoted as […]. ,in, Indicates the first The temporal location of local maxima. Values , This represents the total number of local maxima. The coordinates of each local minimum point on the time axis of the image entropy change map are recorded, and the set of temporal locations of the local minima is denoted as . ,in, Indicates the first The temporal location of each local minimum point Values , This represents the total number of local minima. The set of temporal locations of local maxima and local minima together constitute the temporal locations of extrema.

[0031] Pair adjacent local maxima with local minima in terms of their temporal positions. The pairing rule is: on the time axis, pair a local maximum with an adjacent time coordinate. and a local minimum point Pairing is performed such that the temporal position of the local maximum point precedes the temporal position of the local minimum point, or vice versa. Each pairing forms an entropy fluctuation interval, with the starting point of the interval being the earlier extremum point in the pairing and the ending point being the later extremum point in the pairing, thus forming multiple entropy fluctuation intervals.

[0032] Calculate the rate of change of entropy within each entropy fluctuation interval. The rate of change of entropy is defined as the first derivative of the image entropy with respect to time within the entropy fluctuation interval. The moment when the rate of change crosses zero is identified as the candidate transition moment. The method for locating the moment when the rate of change crosses zero is as follows: fit a cubic spline curve of the image entropy with time within each entropy fluctuation interval. The functional expression of the cubic spline curve is: ; in, Indicates time The image entropy value obtained by fitting a cubic spline curve is given. It is a time variable within the entropy fluctuation range; This refers to the segment interval index number of the cubic spline curve. Values , This represents the total number of sampling points within the entropy fluctuation range. Indicates the first The coordinate values ​​of each sampling point on the time axis of the image entropy change map; For the first The constant term coefficients for each segmented interval; For the first The coefficients of the linear terms in each segmented interval; For the first The coefficients of the quadratic terms in each segmented interval; For the first The coefficients of the cubic terms for each segmented interval. equal to the Image entropy value of each sampling point, coefficient , and The image entropy values ​​of all sampling points within the entropy fluctuation range are obtained by solving a system of linear equations using the three-moment method. For cubic spline curve functions... Taking the first derivative yields Solve the equation The solution within the entropy fluctuation range corresponds to the time value at which the rate of change crosses zero. This time at which the rate of change crosses zero is taken as the candidate transition time.

[0033] The candidate transition moments are selected to identify the first transition moment when the driving and driven gear tooth surfaces switch from a contact state to a separation state, and the second transition moment when they switch from a separation state to a contact state. The selection method is as follows: within the entropy fluctuation range of the candidate transition moment, the rate of change values ​​of the sampling points immediately to the left and right of the candidate transition moment are extracted. If the rate of change value of the sampling point immediately to the left is positive and the rate of change value of the sampling point immediately to the right is negative, then the rate of change jumps from positive to negative, and this candidate transition moment is determined as the first transition moment; if the rate of change value of the sampling point immediately to the left is negative and the rate of change value of the sampling point immediately to the right is positive, then the rate of change jumps from negative to positive, and this candidate transition moment is determined as the second transition moment. All first transition moments constitute the first transition moment set, and all second transition moments constitute the second transition moment set.

[0034] In specific implementation, the first and second transition time sets determined in the above embodiment are obtained. The first transition time occurring within the first tooth profile movement cycle in the first transition time set is selected and denoted as... Select the set of the second transition time that immediately follows The second transition time thereafter is denoted as , and This constitutes a contact state transition pair.

[0035] First transition time Collect the first rotation angle encoding value of the drive wheel at the first transition moment. At the corresponding rising edge trigger time of the rotary encoder pulse, the cumulative pulse count value of the rotary encoder is read, and the cumulative pulse count value is multiplied by the angular resolution of the rotary encoder to obtain the first rotation angle encoding value. At the second transition time Collect the second rotation angle encoding value of the drive wheel at the second transition time. At the corresponding rising edge trigger time of the rotary encoder pulse, the cumulative pulse count value of the rotary encoder is read, and the cumulative pulse count value is multiplied by the angular resolution of the rotary encoder to obtain the second rotation angle encoding value. Calculate the first corner code value. With the second corner code value The difference is used to obtain the change in angular displacement of the drive shaft's driving wheel. , .

[0036] Extracting the first transition moment from a continuous image sequence The corresponding image of the first tooth position of the driven gear, the image of the first tooth position of the driven gear at the acquisition time and the first conversion time. The image frame with the smallest absolute time difference. Extract the second transition time. The corresponding image of the second tooth position of the driven gear, the image of the second tooth position of the driven gear at the acquisition time and the second conversion time. The image frame with the smallest absolute time difference. Within the image of the first tooth position of the driven gear, select a rectangular region centered on the rotation center of the driven gear and with a side length equal to the diameter of the driven gear tooth tip circle as the first tooth position sub-image. Select a rectangular area of ​​the same position and size within the second tooth position image of the driven gear as the second tooth position sub-image. The angular displacement change of the driven wheel is calculated using a phase correlation algorithm. This algorithm then calculates the cross-power spectrum between the first and second tooth position images of the driven wheel in the frequency domain. The calculation process involves: processing the first tooth position sub-image... Perform a two-dimensional discrete Fourier transform to obtain the first spectrum. , ,in, Image spatial domain coordinates, For frequency domain coordinates, This represents the two-dimensional discrete Fourier transform operation. For the second tooth position sub-image... Perform a two-dimensional discrete Fourier transform to obtain the second spectrum. , Calculate the first spectrum. With the second spectrum cross power spectrum : ; in, Indicates the second spectrum The complex conjugate, Represents product The modulus of cross-power spectrum. The dimension and the first tooth position sub-image They are the same size. Regarding the cross-power spectrum... Perform a two-dimensional inverse Fourier transform to obtain the phase-correlated surface. , ,in, This represents the two-dimensional inverse discrete Fourier transform operation. Phase-correlated surface. It is a sub-image of the first tooth position. Two-dimensional matrices of the same size, where each element represents the degree of phase correlation at the corresponding displacement. Extracting the phase-correlated surface. Peak coordinates Peak coordinates Phase-correlated surfaces The row and column indices of the maximum value. The peak coordinates. row index in Convert the change in angular displacement of the driven wheel to the following method: calculate the row index. Relative to phase-correlated surfaces The offset of the center row coordinates is multiplied by the angle value corresponding to a unit pixel in the image space to obtain the change in angular displacement of the driven wheel. .

[0037] Calculate the change in angular displacement of the driving wheel Change in angular displacement of the driven wheel The difference is used to obtain the angular displacement difference. , .

[0038] Obtain the angle between the tooth surface normal direction and the pitch circle tangent direction. This angle is calculated using the gear parameters of the driving and driven gears. Obtain the pressure angle of the driving gear. The national standard for involute cylindrical gears specifies a pressure angle of 20 degrees, where the angle between the tooth surface normal and the pitch circle tangent is equal to the pressure angle of the driving gear. Obtain the base circle radius of the driving wheel. Base circle radius By the pitch circle radius of the driving wheel Multiply by pressure angle The cosine value is calculated. The angular displacement difference is then used. Multiply by the base circle radius of the driving wheel Then multiply by the cosine of the included angle to obtain the reference value of the first meshing clearance. , First meshing clearance reference value Expressed in units of length, it reflects the clearance distance between the tooth surfaces of the driving gear and the driven gear in the direction normal to the tooth surfaces.

[0039] In specific implementation, please refer to Figure 3 The rated torque load is applied to the output shaft of the target reducer by installing a torque loading device on the output shaft. The torque loading device can be a magnetic powder brake or a servo motor loading platform. The loading torque value is set to the rated torque value indicated on the target reducer's nameplate via a torque controller. The rated torque load is maintained for more than one tooth profile motion cycle. The tooth profile motion cycle is determined by dividing the time required for one rotation of the drive shaft by the number of teeth on the driving gear. The time for maintaining the rated torque load is set to three times the tooth profile motion cycle to ensure that image data for at least one complete tooth profile motion cycle is acquired under stable load conditions.

[0040] During the maintenance of the rated torque load, a continuous image sequence of the loaded state in the gear meshing area is re-acquired. When acquiring the continuous image sequence under load, the installation position, optical axis direction, exposure triggering method, and acquisition frame rate of the industrial camera are kept exactly the same as those used when acquiring the continuous image sequence under no-load conditions. The optical axis of the industrial camera is perpendicular to the normal direction of the gear tooth surface in the gear meshing area. The exposure triggering time of the industrial camera is aligned with the rising edge of the rotary encoder pulse signal, and the acquisition frame rate of the industrial camera is 36 times the rotational frequency of the drive shaft. All image frames acquired during the maintenance of the rated torque load are arranged in chronological order to form a continuous image sequence under load.

[0041] Edge enhancement filtering is performed on the continuous image sequence under loading conditions. The edge enhancement filtering uses the difference of Gaussian filtering operator, which is generated by subtracting a first Gaussian kernel with a standard deviation of 1.0 pixels and a second Gaussian kernel with a standard deviation of 2.5 pixels. The difference of Gaussian filtering operator is then convolved with each frame of the continuous image sequence under loading conditions to obtain the edge enhancement image sequence under loading conditions.

[0042] Perform minimum Euclidean distance calculation on the loaded edge-enhanced image sequence. For each frame in the loaded state edge enhancement image sequence, the edge point sets of the active wheel tooth profile and the driven wheel tooth profile are extracted. The extraction method is as follows: the loaded state edge enhancement image is binarized and segmented. The threshold for binarization and segmentation is adaptively determined using the Otsu's method to separate the loaded state active wheel tooth profile region and the loaded state driven wheel tooth profile region. Within the loaded state active wheel tooth profile region, gray-level gradient changes are scanned along the radial direction. The gray-level gradient is calculated using the Sobel operator, and the positions of pixels with gradient magnitudes exceeding 30% of the maximum gradient magnitude are recorded, forming the edge point set of the loaded state active wheel tooth profile. Within the loaded state driven wheel tooth profile region, gray-level gradient changes are scanned along the radial direction, and the positions of pixels with gradient magnitudes exceeding 30% of the maximum gradient magnitude are recorded, forming the edge point set of the loaded state driven wheel tooth profile. Subpixel interpolation processing is performed on both the loaded state active wheel tooth profile edge point sets and the loaded state driven wheel tooth profile edge point sets. The subpixel interpolation processing uses a quadratic surface fitting method to improve the coordinate accuracy of the loaded state edge points to the subpixel level. Iterate through each sub-pixel coordinate point in the set of edge points of the active gear tooth profile under loading conditions, calculate the Euclidean distance from that point to each sub-pixel coordinate point in the set of edge points of the driven gear tooth profile under loading conditions, and take the minimum Euclidean distance as the minimum Euclidean distance value for the loading state image of that frame. Arrange the minimum Euclidean distance values ​​for the loading state corresponding to each frame in the continuous image sequence of loading states in temporal order to generate a loading state distance fluctuation sequence.

[0043] A sliding window entropy value was calculated for the distance fluctuation sequence under loading state. The width of the sliding window was set to 9 sampling points, and the sliding step size was set to 1 sampling point. Starting from the beginning of the distance fluctuation sequence under loading state, the information entropy value of the 9 sampling points within the sliding window was calculated. The formula for calculating the entropy value of the loading state image is: ; in, Indicates the first The entropy value of the loading state image of a sliding window. The index number of the sliding window in the loading state. Values , This represents the total number of sliding windows in the loading state. It equals the total number of sampling points from the loading state to the fluctuation sequence minus the window width plus 1; This represents the total number of equally spaced intervals within the sliding window representing the value of the load state distance fluctuation sequence. Set to 10; This is the index number of the loading status range. Values ; For the first The distance value within the loading state sliding window falls into the first... The probability of the i-th interval is calculated as follows: The first loading status sliding window fell into the first The number of sampling points in each interval is divided by the width of the loading state sliding window. The loading state image entropy value is calculated for each loading state sliding window. Using the sampling time at the center of the sliding window as the x-axis and the image entropy value in the loading state as the y-axis... Using the vertical axis, generate an entropy change map of the image under loading conditions.

[0044] The contact state transition time under loading state is relocated based on the entropy change spectrum of the loading state image. The relocation method is as follows: a sliding detection window is defined on the entropy change spectrum of the loading state image, and the width of the sliding detection window is set to one-quarter of the tooth profile motion period; within the sliding detection window, the relationship between the entropy value of the loading state image at the center point and the entropy values ​​of the loading state image at the left and right adjacent points is compared, and the local maxima and local minima of the loading state are marked; the coordinates of the local maxima and local minima of the loading state on the time axis of the entropy change spectrum of the loading state image are recorded to form the temporal position of the extreme points of the loading state; the temporal positions of adjacent local maxima and local minima of the loading state are paired to form multiple entropy value fluctuation intervals of the loading state; the rate of change of the entropy value of the loading state image within each entropy value fluctuation interval is calculated, and the moment when the rate of change crosses zero is identified as the candidate transition time of the loading state; from the candidate transition times of the loading state, the moment when the rate of change jumps from a positive value to a negative value is selected as the first transition time of the loading state, and the moment when the rate of change jumps from a negative value to a positive value is selected as the second transition time of the loading state.

[0045] The angular displacement difference under loading conditions is extracted again. The extraction method is as follows: at the first transition moment of loading state, the first rotation angle code value of the driving wheel under loading state is collected, and at the second transition moment of loading state, the second rotation angle code value of the driving wheel under loading state is collected. The difference between the first rotation angle code value and the second rotation angle code value under loading state is calculated to obtain the angular displacement change of the driving wheel under loading state. The first tooth position image of the driven wheel under loading state corresponding to the first transition moment of loading state and the second tooth position image of the driven wheel under loading state corresponding to the second transition moment of loading state are extracted from the continuous image sequence of loading state. The cross power spectrum of the first tooth position image and the second tooth position image of the driven wheel under loading state is calculated in the frequency domain using a phase correlation algorithm. After performing an inverse Fourier transform on the cross power spectrum, the peak coordinates are extracted as the angular displacement change of the driven wheel under loading state. The difference between the angular displacement change of the driving wheel under loading state and the angular displacement change of the driven wheel under loading state is calculated to obtain the angular displacement difference under loading state. The second meshing clearance reference value is obtained by multiplying the difference in angular displacement under loading by the base circle radius of the driving wheel, and then by the cosine of the angle between the tooth surface normal direction and the pitch circle tangent direction.

[0046] In specific implementation, please refer to Figure 4 The first meshing clearance reference value calculated in the above embodiment is obtained and denoted as... The second meshing clearance reference value calculated in this embodiment is obtained and denoted as... Set the no-load weight coefficient, denoted as . No-load weighting coefficient The value range is the open interval 0 to 1. Set the loading weight coefficient, denoted as... Load weight coefficients The value range is the open interval 0 to 1. (Unloaded weight coefficient) With loading weight coefficient The sum of is a value of one, that is .

[0047] No-load weighting factor With loading weight coefficient The rated load rate is dynamically adjusted based on the target reducer's rated load rate and operating conditions. The rated load rate is the ratio of the actual torque load currently borne by the target reducer to its rated torque value. During the operation of the target reducer, a torque sensor installed on the output shaft collects the current torque load value in real time. Dividing the current torque load value by the rated torque value specified on the target reducer's nameplate yields the rated load rate, denoted as [missing value]. , The value ranges from 0 to 1. (Unloaded weight coefficient) With loading weight coefficient The ratio is inversely proportional to the current torque load value, that is... Solving the two constraints simultaneously yields the no-load weight coefficient. The calculation formula is Load weight coefficients The calculation formula is When the rated load rate When it is 0.3, the no-load weighting factor The value is 0.7, and the weighting coefficient is applied. The value is 0.3; when the rated load rate When it is 0.8, the no-load weighting factor The value is 0.2, and the weighting coefficient is applied. The value is 0.8.

[0048] The first meshing clearance reference value Multiply by the no-load weighting factor The no-load contribution component is obtained. , The second meshing clearance reference value. Multiply by the loading weight coefficient , and obtain the loading contribution component , The no-load contribution component With load contribution component Add them together to obtain the weighted fusion gap. , .

[0049] To obtain the temperature detection value of the target reducer during the current operating cycle, a first temperature sensor, selected as a PT100 platinum resistance temperature sensor, is installed on the bearing housing of the target reducer. The first temperature sensor is attached to the outer surface of the bearing housing to collect the bearing housing temperature value in real time. A second temperature sensor, a PT100 platinum resistance temperature sensor, is installed on the housing of the target reducer. The second temperature sensor is attached to the outer surface of the housing in the middle, away from the bearing housing, to collect the housing temperature value in real time. Collect bearing housing temperature values ​​and shell temperature values. Body temperature value The sampling period is set to 10 seconds. The bearing housing temperature value is calculated. With shell temperature value The average value is used as the temperature detection value. , .

[0050] A table mapping temperature ranges to thermal expansion compensation factors is pre-established. Temperature measurement values ​​are then... The measurement range is divided into multiple continuous temperature intervals, each spanning 5 degrees Celsius. Specifically, the temperature intervals are: [-20℃, -15℃), [-15℃, -10℃), ..., [95℃, 100℃]. Under experimental calibration conditions, the target reducer is placed in the median temperature environment of each temperature interval. After the target reducer reaches thermal equilibrium, the actual meshing clearance of the gear pair is measured. The actual meshing clearance value is divided by the reference meshing clearance value measured at a standard temperature of 20℃ to obtain the thermal expansion compensation factor for the corresponding temperature interval. The lower and upper limits of each temperature interval, along with the corresponding thermal expansion compensation factor, are stored in a correspondence table. Each row in the correspondence table records a temperature interval and its corresponding thermal expansion compensation factor.

[0051] Based on temperature detection values The thermal expansion compensation factor is obtained by referring to the table. If the temperature measurement value... The temperature value falls within a certain temperature range in the corresponding relationship table. If the temperature reading is greater than the lower limit of the temperature range but less than the upper limit, then the thermal expansion compensation factor corresponding to that temperature range is directly read. If the temperature detection value... It falls precisely on the boundary between two adjacent temperature ranges, i.e., the temperature reading. If the median temperature of the previous temperature interval is equal to the upper limit of the next temperature interval, then the thermal expansion compensation factor is calculated using linear interpolation. The linear interpolation method involves extracting the median temperature of the previous temperature interval. and the corresponding thermal expansion compensation factor Extract the median temperature value of the next temperature range. and the corresponding thermal expansion compensation factor Thermal expansion compensation factor where temperature readings fall on the boundary The calculation formula is: ; in, The thermal expansion compensation factor is obtained by linear interpolation when the temperature detection value falls exactly on the boundary of two temperature ranges. This is the thermal expansion compensation factor corresponding to the previous temperature range. This is the thermal expansion compensation factor corresponding to the next temperature range. It is the median temperature value of the previous temperature range, calculated by summing the lower limit and upper limit of the previous temperature range and dividing by 2. The median temperature value of the next temperature range is calculated by dividing the sum of the lower limit and the upper limit of the next temperature range by 2. This represents the temperature reading during the current operating cycle.

[0052] Weighted fusion gap amount Multiply by the thermal expansion compensation factor obtained from a table lookup or linear interpolation, if the temperature measurement value Within the temperature range, the thermal expansion compensation factor is denoted as... If the temperature detection value If the temperature falls on the boundary of the temperature range, the thermal expansion compensation factor is denoted as . The dynamic meshing clearance is obtained. , ,in, Represents the final determined thermal expansion compensation factor. Dynamic meshing clearance. Expressed in units of length, it reflects the actual tooth backlash value of the target reducer under current operating conditions and temperature.

[0053] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A machine vision-based intelligent detection method for gear backlash in a speed reducer, characterized in that, include: A continuous image sequence of the gear meshing area of ​​the target reducer under no-load operation is obtained, and the continuous image sequence is subjected to dynamic tracking processing of the tooth profile edge to generate an image entropy change map characterizing the relative displacement of the gear pair. Based on the extreme points of entropy fluctuation in the image entropy change spectrum, the moment of contact state transition between the driving wheel and the driven wheel of the target reducer is located in reverse. Based on the contact state transition time, the angular displacement difference between the driving wheel and the driven wheel of the drive shaft is extracted, and the angular displacement difference is mapped to the tooth surface normal direction to obtain the first meshing clearance reference value; After the target reducer is loaded with rated torque, the acquisition of the continuous image sequence and the dynamic tracking processing of the tooth profile edge are repeated to generate the second meshing clearance reference value under the loading state; By combining the first meshing clearance reference value and the second meshing clearance reference value, the dynamic meshing clearance of the target reducer is output.

2. The intelligent detection method for gear backlash in a reducer based on machine vision according to claim 1, characterized in that, Dynamic tracking of tooth profile edges is performed on the continuous image sequence to generate an image entropy change map characterizing the relative displacement of the gear pair, including: The continuous image sequence is divided into multiple image sub-frame sets according to the acquisition time sequence, and each image sub-frame set corresponds to a tooth profile motion cycle; For each frame in each image sub-frame set, edge enhancement filtering is performed to extract the active gear tooth profile edge point set and the driven gear tooth profile edge point set; Calculate the minimum Euclidean distance between the set of edge points of the active gear tooth profile and the set of edge points of the driven gear tooth profile, and arrange the minimum Euclidean distance into a distance fluctuation sequence according to the time order; The distance fluctuation sequence is subjected to sliding window entropy calculation to obtain the image entropy value within each sliding window, and the image entropy value is expanded along the time axis to form the image entropy change map.

3. The intelligent detection method for gear backlash in a reducer based on machine vision according to claim 2, characterized in that, Based on the extreme points of entropy fluctuation in the image entropy change spectrum, the moment of contact state transition between the driving wheel and the driven wheel of the target reducer is located in reverse, including: Scan the entropy change spectrum of the image and extract the temporal positions of all local maxima and local minima. Pairing adjacent local maxima and local minima in time sequence creates multiple entropy fluctuation ranges. Calculate the rate of change of entropy value within each entropy fluctuation interval, and locate the moment when the rate of change crosses zero as the candidate transition moment; The moment when the rate of change jumps from a positive value to a negative value is selected from the candidate transition moments as the first transition moment when the driving gear tooth surface and the driven gear tooth surface switch from a contact state to a separation state. The moment when the rate of change jumps from a negative value to a positive value is selected as the second transition moment when the separation state switches to a contact state.

4. The intelligent detection method for gear backlash in a reducer based on machine vision according to claim 3, characterized in that, The moment when the rate of change crosses zero is located by fitting a cubic spline curve of the entropy value over time within each entropy fluctuation interval, and solving for the point where the derivative of the cubic spline curve is zero.

5. The intelligent detection method for gear backlash in a reducer based on machine vision according to claim 3, characterized in that, Based on the contact state transition moment, the angular displacement difference between the driving wheel and the driven wheel of the drive shaft is extracted, and the angular displacement difference is mapped to the tooth surface normal direction to obtain the first meshing clearance reference value, including: At the first conversion time, the first rotation angle code value of the drive wheel is collected, and at the second conversion time, the second rotation angle code value of the drive wheel is collected. The difference between the first rotation angle code value and the second rotation angle code value is calculated to obtain the angular displacement change of the drive wheel of the drive shaft. Extract the first tooth position image of the driven wheel corresponding to the first conversion time and the second tooth position image of the driven wheel corresponding to the second conversion time from the continuous image sequence, and calculate the angular displacement change of the driven wheel using a phase correlation algorithm; The difference between the angular displacement change of the driving wheel and the angular displacement change of the driven wheel is calculated to obtain the angular displacement difference. Obtain the angle between the tooth surface normal direction and the pitch circle tangent direction, multiply the angular displacement difference by the base circle radius of the driving wheel, and then multiply by the cosine of the angle to obtain the first meshing clearance reference value.

6. The intelligent detection method for gear backlash in a reducer based on machine vision according to claim 5, characterized in that, The phase correlation algorithm calculates the cross power spectrum of the first tooth position image and the second tooth position image of the driven wheel in the frequency domain, and extracts the peak coordinates as the angular displacement change of the driven wheel after performing an inverse Fourier transform on the cross power spectrum.

7. The intelligent detection method for gear backlash in a reducer based on machine vision according to claim 5, characterized in that, After the target reducer is loaded with rated torque, the acquisition of the continuous image sequence and the dynamic tracking processing of the tooth profile edges are repeated to generate a second meshing clearance reference value under loading conditions, including: A rated torque load is applied to the output shaft of the target reducer, and the rated torque load is maintained for a time exceeding one tooth profile motion cycle; During the period of maintenance of the rated torque load, a continuous image sequence of the loading state of the gear meshing area is reacquired; The edge enhancement filtering and minimum Euclidean distance calculation are performed on the continuous image sequence of the loading state to generate a loading state distance fluctuation sequence; The sliding window entropy value is calculated on the distance fluctuation sequence of the loading state to generate an entropy change map of the loading state image; Based on the entropy change spectrum of the loaded state image, the contact state transition time under the loaded state is repositioned, and the angular displacement difference under the loaded state is re-extracted to map the second meshing gap reference value.

8. The intelligent detection method for gear backlash in a reducer based on machine vision according to claim 7, characterized in that, By integrating the first meshing clearance reference value and the second meshing clearance reference value, the dynamic meshing clearance of the target reducer is output, including: Multiply the first meshing clearance reference value by the no-load weight coefficient to obtain the no-load contribution component, and multiply the second meshing clearance reference value by the loading weight coefficient to obtain the loading contribution component. The sum of the no-load weight coefficient and the loading weight coefficient is one. The unloaded contribution component is added to the loaded contribution component to obtain the weighted fusion gap amount; Obtain the temperature detection value of the target reducer during the current operating cycle, and obtain the thermal expansion compensation factor by looking up the table based on the temperature detection value; The dynamic meshing gap is obtained by multiplying the weighted fusion gap by the thermal expansion compensation factor.

9. The intelligent detection method for gear backlash in a reducer based on machine vision according to claim 8, characterized in that, The no-load weight coefficient and the load weight coefficient are dynamically adjusted according to the rated load rate and operating conditions of the target reducer, and the ratio of the no-load weight coefficient to the load weight coefficient is inversely proportional to the current torque load value.

10. The intelligent detection method for gear backlash in a reducer based on machine vision according to claim 1, characterized in that, The acquisition of a continuous image sequence of the gear meshing area of ​​the target reducer under no-load operation includes: An industrial camera is mounted on the housing of the target reducer, with the optical axis of the industrial camera perpendicular to the normal direction of the tooth surface in the gear meshing area; The pulse signal of the rotary encoder on the drive shaft of the target reducer is acquired synchronously, and the exposure trigger time of the industrial camera is aligned with the rising edge of the rotary encoder pulse signal. The industrial camera's frame rate is set to an integer multiple of the drive shaft's rotation frequency. Multiple equally spaced image frames are acquired within each tooth profile motion cycle to form the continuous image sequence.

Images