Hollow rotating platform positioning system based on intelligent image recognition

Abstract

The application relates to the field of automatic control and image recognition technology, and particularly discloses a hollow rotating platform positioning system based on intelligent image recognition. The system comprises a hollow rotating platform body, a high-precision encoder, a visual reference module with a specific reference pattern, a high-speed image acquisition module, a real-time error solving module and a dynamic compensation controller. The application introduces visual reference and image recognition technology, and constructs an absolute space reference system independent of the platform mechanical shaft system. By solving the sub-pixel level deviation of the image acquisition module on the rotating component relative to the fixed visual reference in real time, the system can directly measure and quantify the real-time geometric error of the platform in operation due to vibration and deformation, and can change the maintenance of positioning accuracy from relying on periodic offline calibration to online real-time perception and compensation, thereby solving the problem of accumulated error caused by invalidation of calibration parameters in the traditional method.

Description

Technical Field

[0001] This invention belongs to the field of automation control and image recognition technology, specifically relating to a positioning system for a hollow rotating platform based on intelligent image recognition. Background Technology

[0002] In the fields of industrial automation and precision manufacturing, high-precision positioning technology is the core foundation for achieving complex machining, precision assembly, and automated inspection. As a key actuator for achieving multi-angle positioning of workpieces or tools, the positioning accuracy of the rotary platform directly determines the performance ceiling of the entire system.

[0003] Hollow rotary platforms, with their unique central through-hole design, facilitate pipeline routing and are widely used in precision equipment requiring integrated vision inspection, multi-axis linkage, or internal wiring. The core technological goal of these platforms is to achieve rapid, highly repeatable angle positioning to meet the dual demands of efficiency and precision in modern intelligent manufacturing.

[0004] In existing technologies, the positioning accuracy of hollow rotary platforms is highly dependent on offline calibration at the factory and periodic maintenance calibration. Traditional calibration methods typically require the use of external high-precision measuring equipment, such as laser interferometers, to measure the platform's rotation angles one by one and input compensation parameters.

[0005] This process is not only complex and time-consuming, severely impacting the equipment's production cycle and availability, but its calibration accuracy is also difficult to maintain during actual platform operation. Because the rotary platform inevitably generates mechanical vibrations during high-speed start-up, shutdown, or continuous operation, these vibrations cause micron-level shifts in the encoder zero point or mechanical reference point inside the platform, rendering previously calibrated parameters invalid and resulting in cumulative repetitive positioning errors.

[0006] Furthermore, factors such as temperature variations in the production environment and long-term mechanical wear can further exacerbate this accuracy degradation. Therefore, achieving long-term, stable, and high-precision positioning of hollow rotary platforms under complex working conditions without relying on frequent external calibration has become a pressing technical challenge in this field. Summary of the Invention

[0007] The purpose of this invention is to provide a positioning system for a hollow rotating platform based on intelligent image recognition, so as to solve the technical contradiction in the prior art that relies on offline calibration and cannot compensate for the cumulative positioning error caused by vibration, temperature drift and wear in real time, thereby affecting the long-term operating accuracy and stability.

[0008] To achieve the above objectives, this invention provides a positioning system for a hollow rotating platform based on intelligent image recognition. The system includes a hollow rotating platform body, a high-precision encoder, a visual reference module, a high-speed image acquisition module, a real-time error calculation module, and a dynamic compensation controller.

[0009] The hollow rotating platform body has a through hole in its center for threading pipelines or installing coaxial components; its rotating shaft system is directly coupled to the drive motor to bear the load and perform rotational motion.

[0010] The high-precision encoder has its reading head fixedly mounted on the platform base, and its code disk is rigidly connected to the platform's rotation axis. It is used to measure and output the platform's first angular position data in real time. This first angular position data constitutes the system's basic positioning feedback.

[0011] The visual reference module is fixedly installed on the platform base and is strictly coaxial with the platform's rotation center axis. The module contains a reference pattern with unique geometric features, which consists of multiple concentric rings and coded marks with a specific radial distribution, the center of which is precisely aligned with the theoretical rotation center of the platform.

[0012] The high-speed image acquisition module is fixedly mounted on a rotatable part of the platform, with its optical axis perpendicular to the plane where the visual reference module is located and aligned with the central area of ​​the reference pattern. This module is used to acquire real-time images containing the reference pattern at a rate of more than 1,000 frames per second at any moment when the platform is rotating or stationary.

[0013] The real-time error calculation module has its input end connected to the output end of the high-speed image acquisition module, and is used to receive and process real-time images.

[0014] This module first preprocesses the real-time image, including grayscale conversion, noise reduction, and edge enhancement;

[0015] Subsequently, the edge contours of all concentric rings in the reference pattern are extracted using a sub-pixel level edge detection algorithm;

[0016] Next, the least squares method is used to fit circles to all the extracted edge point sets, and the center coordinates and radius of each circle in the image coordinate system are calculated.

[0017] Furthermore, the module performs a weighted average of the coordinates of all fitted circle centers to calculate the actual imaging center coordinates of the reference pattern in the current image;

[0018] Finally, the actual imaging center coordinates are compared with the theoretical center coordinates of the image, which correspond to the fixed pixel position of the ideal rotation center of the platform in the image, and the pixel offsets of the two in the X and Y axes of the image coordinate system are calculated.

[0019] The dynamic compensation controller has a first input terminal connected to the output terminal of a high-precision encoder to receive first angular position data, and a second input terminal connected to the output terminal of a real-time error calculation module to receive pixel offset data.

[0020] The controller has a built-in calibration mapping model and error compensation algorithm. The calibration mapping model is established in advance through a calibration process, which drives the platform to rotate at multiple known angles. The real-time error calculation module records the corresponding pixel offset at each angle, thereby constructing a nonlinear mapping relationship database between pixel offset and the theoretical rotation angle of the platform.

[0021] During system operation, the dynamic compensation controller executes the following process:

[0022] First, based on the currently received first angle position data, the calibration mapping model is queried to predict the expected pixel offset caused by factors such as platform mechanical deformation at the current angle;

[0023] Simultaneously, the actual pixel offset calculated in real time is compared with the expected pixel offset by a vector difference operation to obtain the instantaneous dynamic error vector;

[0024] This instantaneous dynamic error vector characterizes the real-time deviations that exceed model predictions caused by vibration, instantaneous impact, or thermal deformation;

[0025] Furthermore, the dynamic compensation controller inputs the instantaneous dynamic error vector into the error compensation algorithm; the error compensation algorithm first estimates the equivalent angular deviation of the error transmitted to the end of the load based on the platform's current angular velocity and angular acceleration through a second-order dynamic model;

[0026] Then, this equivalent angle deviation is added to the first angle position data to generate the second angle position data after real-time compensation;

[0027] Finally, the dynamic compensation controller generates a correction control signal for the drive motor based on the difference between the second angle position data and the system target position command, in order to eliminate real-time positioning errors.

[0028] As one embodiment of the present invention, in the reference pattern of the visual reference module, the encoding mark is composed of micro QR codes arranged along a specific radius or light and dark stripes with a specific width-to-width ratio.

[0029] After completing the center calculation, the real-time error calculation module further identifies and decodes the coded mark; the decoded information includes the absolute coded sequence number of the mark in the reference pattern.

[0030] The system compares the absolute encoded sequence number decoded from consecutive image frames and combines it with an image matching algorithm to achieve absolute counting and coarse positioning of the platform's rotation number. This forms redundancy and mutual verification with the incremental signal of the high-precision encoder, preventing the loss of rotation number information due to encoder pulse loss or power failure.

[0031] As one embodiment of the present invention, the construction process of the calibration mapping model adopts an adaptive learning mechanism;

[0032] Specifically, after each successful positioning movement and stabilization, the system records the actual pixel offset at the stable position and adds it as a new calibration data point to the mapping relationship database.

[0033] For newly added data points and historical data points, a neural network based on radial basis functions is used for surface fitting, and the calibration mapping model is dynamically updated.

[0034] This mechanism enables the system to automatically learn and compensate for slowly changing systematic errors caused by long-term mechanical wear, thus achieving self-evolution of the calibration model.

[0035] As one embodiment of the present invention, the error compensation algorithm further includes an error filtering and prediction unit;

[0036] This unit receives the instantaneous dynamic error vector sequence from historical time series, and first uses a Kalman filter to filter out measurement noise to obtain a clean error estimate.

[0037] Then, based on the cleanliness error estimates at the current and previous two time points, the error change trend in the next control cycle is predicted using linear extrapolation or quadratic polynomial fitting methods.

[0038] The dynamic compensation controller incorporates the predicted future error trend into the calculation of the equivalent angle deviation in advance, thereby achieving feedforward compensation and further improving the trajectory tracking accuracy of the system under high-speed continuous motion.

[0039] In one embodiment of the present invention, a coaxial light source is provided between the high-speed image acquisition module and the visual reference module;

[0040] The coaxial light source emits uniform parallel light that illuminates the surface of the visual reference module perpendicularly.

[0041] This illumination method eliminates shadow interference that may be generated by other structures on the rotating parts, ensuring that the reference pattern has high contrast and clear edge features in the image, thereby improving the measurement accuracy and robustness of the real-time error calculation module.

[0042] As one embodiment of the present invention, the system operates in a multi-rate control architecture; the architecture includes a high-speed closed loop and a low-speed closed loop.

[0043] The high-speed closed loop operates at a frequency greater than 10 kHz, and its feedback signal comes from the first angular position data of the high-precision encoder, which is mainly used to suppress high-frequency disturbances and resonances caused by the motor drive.

[0044] The low-speed closed loop operates at a frequency of 1 kHz, and its feedback signal comes from the second angle position data generated by the dynamic compensation controller. It is mainly used to compensate for the low-frequency, slow time-varying error calculated by the visual reference.

[0045] The control outputs of the two closed loops are superimposed before the current loop to form the final armature drive command.

[0046] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0047] 1. This invention creatively introduces visual reference and image recognition technology to construct an absolute spatial reference system independent of the platform's mechanical axis system. By calculating the sub-pixel-level offset of the image acquisition module on the rotating component relative to the fixed visual reference in real time, the system can directly measure and quantify the real-time geometric errors caused by vibration and deformation during platform operation. This transforms the maintenance of positioning accuracy from relying on periodic offline calibration to online real-time sensing and compensation, solving the problem of accumulated errors caused by calibration parameter failure in traditional methods.

[0048] 2. This invention achieves efficient collaboration of multi-sensor information by fusing encoder data and visual error data through a dynamic compensation controller. The system not only utilizes visual information for real-time error measurement but also, through an established calibration mapping model and error compensation algorithm, accurately converts pixel offsets in the image domain into equivalent angular deviations at the load end, generating forward-looking control corrections. This deep data fusion and proactive compensation mechanism enables the system to effectively combat the instantaneous impact of high-speed start-stop and the time-varying interference under complex operating conditions, achieving stable positioning performance with nanometer-level equivalent angular resolution.

[0049] 3. This invention possesses powerful adaptive and self-learning capabilities. The adaptive learning mechanism of the calibration mapping model allows the system to continuously accumulate data during long-term operation, automatically updating the model to track and compensate for slow-changing systematic drift caused by mechanical wear and material aging. Simultaneously, the application of error filtering and prediction units upgrades the system from passive error feedback compensation to predictive compensation including feedforward, improving dynamic response speed and trajectory tracking accuracy. These characteristics collectively ensure high accuracy and high reliability throughout the system's entire lifecycle, significantly reducing maintenance costs and downtime. Attached Figure Description

[0050] Figure 1 This is a schematic diagram of the overall technical solution architecture of the hollow rotating platform positioning system based on intelligent image recognition proposed in this invention;

[0051] Figure 2 This is a schematic diagram of the core principle framework of real-time geometric error calculation based on visual benchmarks and image recognition in this invention;

[0052] Figure 3 This is a flowchart illustrating the logical flow of the dynamic compensation controller in this invention, which integrates multi-source information for error prediction and compensation.

[0053] Figure 4 This is a schematic diagram illustrating the collaborative interaction between the high-speed closed loop and the low-speed closed loop under the multi-rate control architecture of this invention.

[0054] Figure 5 This is a schematic diagram illustrating the principle framework of the adaptive learning mechanism for the calibration mapping model in this invention. Detailed Implementation

[0055] Example 1: This example details the technical implementation of a hollow rotating platform positioning system based on intelligent image recognition. Please refer to the appendix. Figures 1 to 5 This system constitutes a complete mechatronics-optical integrated closed-loop control system. Its core lies in using a visual absolute reference independent of the mechanical axis system to perceive and compensate for the geometric errors generated by the hollow rotating platform during dynamic operation in real time, thereby achieving ultra-high precision and long-term stable positioning performance.

[0056] The basic mechanical structure of the system is a hollow rotating platform body.

[0057] The main body is integrally molded from a high-rigidity material. Its core feature is that there is a through hole in the center. The diameter of the through hole is typically 50 mm to 200 mm, and the specific size depends on the diameter of the pipeline harness or the size of the coaxially mounted component that needs to be installed in the application scenario.

[0058] The axis of the through hole is strictly aligned with the axis of rotation of the platform, with the tolerance controlled within 2 micrometers.

[0059] The platform's rotating shaft system is supported by crossed roller bearings or hydrostatic bearings to achieve extremely low rotational runout and frictional torque.

[0060] The drive motor is coupled to the platform's rotating shaft via a high-rigidity coupling or directly, and the motor rotor and the platform load disk form an integral moving part.

[0061] The load mounting surface is located on the upper surface of the platform rotating component and is equipped with a standard threaded hole array for fixing the load.

[0062] The platform base is rigidly installed on the foundation or frame using anchor bolts to ensure the basic stability of the entire system.

[0063] A high-precision encoder serves as the basic angle measurement unit of the system.

[0064] Its code disk is an optical glass incremental or absolute grating with a typical line count of 262,144 lines per revolution, providing 18-bit raw resolution.

[0065] The encoder is rigidly connected to the platform's rotating shaft via an interference fit or a special fixture, ensuring that its rotation is completely synchronized with the platform's main shaft.

[0066] The reading head is a photoelectric reading head with differential signal output, which is fixedly installed on an independent bracket on the platform base by a precision adjustment frame. The bracket is integrally machined with the base or bonded with epoxy resin to ensure long-term stability.

[0067] The air gap between the reading head and the code disk is strictly controlled within the nominal range of 0.1 mm to 0.3 mm. The high-precision encoder's integrated subdivision circuit performs electronic subdivision of the original sine and cosine signals by 4096 times, ultimately outputting the first angular position data.

[0068] The data is transmitted in real time in the form of a 32-bit signed integer via a high-speed serial peripheral interface protocol or Ethernet. Its update frequency is greater than 10 MHz, and the number of valid bits is guaranteed to be more than 24, corresponding to an angular resolution better than 0.0001 degrees.

[0069] The visual reference module is the core component for constructing an absolute spatial reference frame. Please refer to the attached document. Figure 2 The module is made of a highly stable ceramic or microcrystalline glass substrate and is rigidly and stress-free mounted on the platform base by means of three-point positioning pins and bolts with matching coefficients of thermal expansion.

[0070] During installation, a laser interferometer or a dual-frequency laser interferometer is used for precise calibration to ensure that the perpendicularity error between the substrate plane and the platform rotation axis is less than 3 arcseconds, and the coaxiality error between the center of the reference pattern on the substrate and the theoretical rotation center of the platform is less than 1 micrometer.

[0071] The reference pattern is processed on the substrate surface by photolithography or laser direct writing process. The pattern consists of multiple concentric rings and coded marks with specific radial distribution.

[0072] The typical number of concentric rings is 5 to 10, and the radial spacing between the rings increases from the inside to the outside. For example, the inner ring spacing is 0.5 mm and the outer ring spacing is 2 mm.

[0073] This design ensures that at least three rings can be clearly imaged at different imaging distances and viewing angles, providing redundant data for high-precision center fitting. The coded markers are located in specific sectors between the outermost two rings and consist of a matrix of miniature QR codes arranged along the radial direction.

[0074] Each QR code measures 200 micrometers x 200 micrometers and contains binary information with error-correcting codes. This information encodes the absolute position number of the mark in the circumferential direction, ranging from 0 to 1023, thus encoding 1024 absolute positions in the circumferential direction. The pattern surface may be coated with an anti-reflective coating or a metallic reflective coating to optimize optical contrast.

[0075] The high-speed image acquisition module is the execution unit for dynamic visual perception.

[0076] The module is encapsulated in a cylindrical metal housing and is directly fixed to the rotatable parts of the platform via a flange. Specifically, it is located below or to the side of the load mounting surface, and its mounting axis is parallel to the platform's rotation axis. The offset is precisely measured and recorded.

[0077] At the core of the module is a global shutter complementary metal-oxide-semiconductor image sensor with a typical pixel size of 3.45 micrometers and a typical resolution of 1280 pixels × 1024 pixels. A telecentric lens is positioned in front of the image sensor, with a magnification of 0.3x to 0.5x and a depth of field greater than 5 mm, ensuring that the reference pattern is always clearly imaged within the allowable axial movement range of the platform.

[0078] The lens optical axis is adjusted with six degrees of freedom during installation to be strictly perpendicular to the plane of the visual reference module substrate and aligned with the center area of ​​the reference pattern. The initial alignment deviation is calibrated and stored by the image center positioning algorithm.

[0079] Image acquisition is triggered by an internal free-running mode, operating at a fixed frame rate. The frame rate can be configured to 1000, 2000, or 5000 frames per second, and the exposure time is adaptively adjusted according to the brightness of the coaxial light source.

[0080] The acquired raw Bayer array images are preprocessed in real time by the module’s internal field-programmable gate array, including fixed-mode noise correction and black level correction, and converted into a 12-bit grayscale image stream, which is then output through the camera link or a 10 Gigabit Ethernet interface.

[0081] The real-time error calculation module is deployed in the real-time computing unit, which adopts a heterogeneous architecture of multi-core processors and field-programmable gate arrays.

[0082] The module input receives a grayscale image stream from the high-speed image acquisition module via a high-speed communication interface. The processing flow begins in the image preprocessing sub-stage.

[0083] First, the grayscale image is regularized, linearly mapping the 12-bit pixel values ​​to an 8-bit range of 0 to 255.

[0084] Noise reduction was then performed using a Gaussian filter with a kernel size of 5 pixels × 5 pixels and a standard deviation of 1.0.

[0085] Next, the Sobel operator is used for edge enhancement, the gradient of the image in the X and Y directions is calculated, and the gradient magnitude image is generated.

[0086] The preprocessed image enters the subpixel edge detection sub-stage.

[0087] This stage involves scanning each expected annular region of the baseline pattern.

[0088] The algorithm first finds the pixel with the largest gradient in the gradient magnitude image along a preset search path and uses it as the integer pixel edge point.

[0089] Within a 3-pixel × 3-pixel neighborhood of this point, the coordinates of the edge point with sub-pixel precision are solved by using quadratic polynomial interpolation of the gradient magnitude.

[0090] This process can be represented as performing a Taylor expansion of the gradient function and finding its extrema.

[0091] For each detected ring, hundreds of uniformly distributed sub-pixel edge points will be extracted to form the edge point set of the ring.

[0092] The next step is the circle center fitting calculation sub-stage. For the edge point set extracted from each ring, the least squares method is used to fit the circle. Let the equation of the circle be:

[0093] ;

[0094] in and Let the coordinates be the center of the circle. For radius, The x-axis is... The vertical axis is denoted by .

[0095] The equation is linearized as follows:

[0096] ;

[0097] in For those containing Points Given a point set, construct an overdetermined system of equations, and obtain the parameters by solving the normal equations. , , The least squares solution is used to inversely calculate the coordinates of the circle's center. and radius .

[0098] The goodness of fit of each ring is evaluated by the sum of squared residuals, and rings with excessively large residuals are discarded.

[0099] After obtaining the fitted center coordinates of all valid rings, the actual imaging center calculation sub-stage begins. This stage performs a weighted average of all fitted center coordinates.

[0100] The weights are determined based on two factors:

[0101] First, the fitting residual of the ring; the smaller the residual, the higher the weight.

[0102] Secondly, the radius of the ring is considered. The ring with the middle radius is given a higher weight because it is less affected by lens distortion and has high stability in edge detection.

[0103] The actual imaging center coordinates of the reference pattern in the current image frame are calculated by weighted averaging. .

[0104] Finally, there is the pixel offset solution stage. During system initialization, the theoretical center coordinates of the image are determined through a precise calibration process. .

[0105] This coordinate corresponds to the projection position of the center of the pattern circle of the visual reference module onto the image sensor when the platform is in an ideal mechanical zero position and without any deformation.

[0106] The real-time error calculation module calculates the difference between the actual imaging center and the theoretical center:

[0107] ；

[0108] ;

[0109] Should and These are the calculated pixel offsets, which directly reflect the two translational errors of the platform's rotating components relative to the fixed base in a plane perpendicular to the optical axis.

[0110] These data are output in floating-point form, with a synchronization period of 1 kilohertz.

[0111] Meanwhile, the real-time error calculation module runs the encoded tag recognition subprocess in parallel.

[0112] The process searches for miniature QR codes within a pre-defined sector area of ​​the image.

[0113] By performing steps such as location graphic detection, image binarization, and format information decoding, the absolute position sequence number encoded in the QR code is recovered.

[0114] The system records the sequence number decoded from the current frame and compares it with the sequence number from the previous frame.

[0115] By combining an image block matching algorithm based on normalized cross-correlation, it is possible to accurately determine whether the platform has crossed the QR code boundary, thereby achieving absolute counting of rotations and coarse positioning.

[0116] The number of revolutions is cross-checked with the zero-position signal of the high-precision encoder. If any inconsistency is found, the system alarm is triggered and the diagnostic process is started to prevent the loss of absolute position information due to encoder power failure or strong interference.

[0117] The dynamic compensation controller is the core decision-making unit for realizing intelligent compensation.

[0118] Please refer to the attached document. Figure 3 The controller runs on a separate core of the real-time computing unit, and its first input receives the first angular position data continuously output by the high-precision encoder via shared memory or a real-time bus. Its second input terminal receives the pixel offset output by the real-time error calculation module. and .

[0119] The controller's operation relies on a pre-established calibration mapping model. The construction of this model is a system initialization step. The system control platform rotates slowly from 0 degrees to 360 degrees in 0.1-degree increments, completing one full rotation, for a total of 3600 calibration points.

[0120] At each calibration point, after the platform comes to rest and stabilizes, the average value of the high-precision encoder readings is recorded as the theoretical angle. Simultaneously, record the pixel offset output by the real-time error calculation module. and The average value.

[0121] Thus, a dataset containing 3,600 data points was obtained, which describes the static geometric error mapping relationship of the platform under different theoretical perspectives caused by machining errors, assembly errors, and structural deformation.

[0122] The dataset is stored as a high-resolution lookup table, or used to train a radial basis function-based neural network model to form a calibration mapping model. , making .

[0123] During the real-time operation phase, the dynamic compensation controller executes the following continuous process.

[0124] First, perform model prediction: based on the current first angular position data. Query the calibration mapping model This yields the expected static pixel offset at the current angle. and Next, dynamic error extraction is performed: the actual pixel offset measured in real time is used. , The instantaneous dynamic error vector is calculated by vector difference with the predicted value:

[0125] ;

[0126] ;

[0127] The vector It characterizes the real-time deviations caused by dynamic factors such as vibration, instantaneous impact, and thermal deformation that exceed the model's prediction range.

[0128] Then, this instantaneous dynamic error vector is fed into the error compensation algorithm. The primary task of the error compensation algorithm is to convert the pixel error of the image plane into the equivalent angular deviation at the load end.

[0129] The algorithm incorporates a conversion model, which is obtained through calibration, and its input is the pixel offset. The output, along with the current working distance, is the translation error of the center point of the load mounting surface in a plane perpendicular to the rotation axis. The unit is micrometers. This takes into account that the load may have a non-zero installation height. This translation error will result in an equivalent tilt angle error. When the platform rotates by an angle of... At that time, the additional angular deviation caused by this tilt angle error at the end of the load. It can be calculated through geometric relationships, and its simplified model is the angular change caused by the tangential displacement at the end of the load.

[0130] The error compensation algorithm further includes an error filtering and prediction unit. This unit receives a sequence of instantaneous dynamic error vectors from the historical time series. First, a discrete-time Kalman filter is used to... Filtering is performed separately. The Kalman filter model treats the actual error as a uniformly changing process and models the measurement noise as Gaussian white noise.

[0131] The filtering process is iterative. The prediction step is based on the state estimate from the previous time step, while the update step incorporates the current measurement values, ultimately outputting a cleanliness error estimate after noise removal. and ,in This is the index for the current control cycle.

[0132] Then, based on the cleanliness error estimates at the current time and the previous two time points, the prediction unit uses a quadratic polynomial fitting method to predict the next control cycle. Error value at time and .

[0133] The dynamic compensation controller estimates the cleanliness error for the current cycle. And the predicted error value for the next cycle Both values ​​are input into the equivalent angle deviation calculation module. This module comprehensively considers the current value and trend of the error to calculate the total equivalent angle compensation value. Finally, the dynamic compensation controller generates the second angular position data. .

[0134] Should It is an angle estimate that is closer to the actual spatial position of the load after integrating the original data of the high-precision encoder and the real-time visual error compensation information.

[0135] The system operates on a multi-rate control architecture. Please refer to the appendix. Figure 4 The architecture comprises two parallel control loops. The high-speed loop is an inner loop nested with a current loop and a speed loop, operating at a frequency of 20 kHz. The feedback signal of this loop is directly derived from the first angular position data from a high-precision encoder. The instantaneous angular velocity is obtained through high-speed differential calculation, and combined with motor current feedback, the closed loop adopts proportional-integral-derivative control and a notch filter, which is specifically used to suppress motor cogging torque, current loop fluctuations and high-frequency resonance of mechanical structure, ensuring the fast response and stability of the drive system.

[0136] The low-speed closed-loop operation frequency is 1 kHz, synchronized with the update frequency of real-time error calculation and dynamic compensation. The feedback signal of this closed loop is the second angle position data output by the dynamic compensation controller. The low-speed closed-loop position controller receives the system's target position command. Calculate positional deviation .

[0137] The positional deviation is processed by the proportional-integral controller to generate a low-speed compensation torque command.

[0138] The high-frequency torque command from the high-speed closed-loop output and the low-frequency compensation torque command from the low-speed closed-loop output are vector-superimposed at the setpoint input of the current loop.

[0139] The superimposed total torque command is amplified by the current loop and finally drives the motor to generate the corresponding electromagnetic torque, thereby achieving precise and stable control of the load position.

[0140] This architecture enables high-frequency disturbances to be quickly suppressed by the high-speed loop, while low-frequency geometric errors and drifts are precisely compensated by the low-speed loop that incorporates visual information. The two work together to achieve high-precision control over a wide bandwidth.

[0141] The system's long-term accuracy is maintained thanks to the adaptive learning mechanism of the calibration mapping model. Please refer to the appendix. Figure 5 This mechanism runs as a background task.

[0142] Each time the system completes a positioning command and enters a stable state, that is, when the position error remains greater than 100 milliseconds within a preset threshold, a learning opportunity is triggered.

[0143] The system records the stable high-precision encoder angle at this time. And the corresponding, filtered actual pixel offset This data point It is sent to the dynamic database.

[0144] The database management submodule checks the spatial distance between the new data point and the historical data point. If the angular difference between the new point and the most recent historical point is less than 0.01 degrees, the value of the old point is updated with the measurement of the new point, thus refreshing the data.

[0145] If the new point is located in an area with sparse historical data, it is inserted as a new data point.

[0146] The model update submodule is activated periodically, or when a certain number of new data points are added.

[0147] This module uses a radial basis function neural network to perform surface fitting on all data points in the database.

[0148] The neural network input is an angle. The output is the predicted offset. .

[0149] The radial basis function center points are selected from a subset of data points in the database, and the weights are obtained by solving a linear least squares problem.

[0150] After fitting, new network weight parameters are generated and updated to the running calibration mapping model in a smooth and gradual manner. .

[0151] This process enables the system to automatically track and compensate for slow, time-varying systematic errors caused by mechanical wear, changes in lubrication conditions, and structural stress relaxation, thereby achieving self-evolution of the calibration model and ensuring the maintenance of accuracy throughout its lifecycle.

[0152] To ensure the extreme robustness of the visual measurement subsystem, a coaxial illumination system was integrated between the high-speed image acquisition module and the visual reference module.

[0153] The coaxial light source consists of an array of light-emitting diodes (LEDs), a diffuser, and a beam splitter. The LEDs emit high-brightness white light, which becomes a uniform surface light source after being diffused. Then, it is reflected by the beam splitter and becomes parallel light that is perpendicularly projected onto the surface of the visual reference module.

[0154] The reflected light from the reference pattern passes through the beam splitter again and enters the lens of the high-speed image acquisition module.

[0155] This lighting method eliminates directional shadows caused by obstruction from other mechanical structures on rotating parts, ensuring that the contrast between light and dark areas of the reference pattern remains consistent at any rotation angle.

[0156] The steep grayscale gradient at the edge of the ring in the image reduces the sensitivity of the sub-pixel edge detection algorithm to the threshold parameter, and improves the measurement repeatability and accuracy under different ambient light interferences.

[0157] The entire system's software runs on a deterministic real-time operating system.

[0158] Each module, including the image acquisition driver, real-time error calculation thread, dynamic compensation control thread, multi-rate control loop, adaptive learning task, and communication service, is assigned a specific priority and a fixed time slice.

[0159] Precise clock synchronization ensures stable and predictable end-to-end latency from image exposure and data processing to compensation control output, with typical total latency controlled within one low-speed control cycle, i.e., less than 1 millisecond.

[0160] System status, including original angle, compensated angle, pixel offset, dynamic error vector, model parameters, alarm codes, etc., is transmitted to the upper-level monitoring system via a non-real-time communication link for status display, data recording, and advanced maintenance diagnosis.

[0161] Example 2: This example provides another specific implementation of a hollow rotating platform positioning system based on intelligent image recognition, focusing on an efficient implementation scheme under limited computing resources, as well as optimized configuration for specific application scenarios.

[0162] In this embodiment, the basic architecture of the system remains the same as in Embodiment 1, but the pattern design of the visual reference module has been simplified.

[0163] The baseline pattern consists of three equally spaced concentric rings and a set of sector-shaped coding strips located outside the outermost ring.

[0164] The encoding stripe uses bright and dark stripes with a specific aspect ratio instead of a miniature QR code, such as radial stripes encoded with Gray code. Each sector corresponds to a unique sequence of stripe widths, and the absolute sector number can be decoded by detecting the number of transitions between light and dark stripes and their order.

[0165] This design reduces the resolution requirements of the image sensor and simplifies the complexity of the decoding algorithm, making it suitable for industrial environments where the absolute circle count accuracy is slightly lower but higher resistance to contamination is required.

[0166] Accordingly, the algorithm flow of the real-time error calculation module has been optimized accordingly.

[0167] In the image preprocessing stage, adaptive threshold segmentation is used instead of global threshold after noise reduction to cope with uneven illumination.

[0168] Instead of weighting all the rings, the center of the ring with the highest sharpness and the smallest fitting residual is selected as the actual imaging center.

[0169] This sacrifices some redundancy but significantly reduces the amount of computation.

[0170] For decoding the coded stripes, a one-dimensional projection method is used, which integrates the image pixels of the coded stripe region along the circumferential direction to obtain the gray value distribution curve along the radial direction.

[0171] By detecting the number of transitions, the transition order, and the corresponding width features of the light and dark gray levels in the curve, and matching them with the preset Gray code encoding table, the absolute number of the current sector can be quickly decoded, achieving coarse positioning of the rotation angle and counting of rotations. At the same time, it reduces the algorithm's requirements for image resolution and computing resources, and improves the decoding robustness in complex industrial environments such as dust and oil.

[0172] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0173] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A positioning system for a hollow rotating platform based on intelligent image recognition, characterized in that, include: The hollow rotating platform body has a through hole in the center, and the rotating shaft system is directly coupled to the drive motor. A high-precision encoder with a reading head fixedly mounted on the platform base and a code disk rigidly connected to the platform's rotating shaft is used to measure and output the platform's first angular position data in real time. A visual reference module is fixedly installed on the platform base and strictly coaxial with the platform's rotation center axis. The visual reference module contains a reference pattern with unique geometric features. The reference pattern consists of multiple concentric rings and coded marks with a specific radial distribution, and the center of the rings is precisely aligned with the theoretical rotation center of the platform. A high-speed image acquisition module is fixedly mounted on a rotatable part of the platform. Its optical axis is perpendicular to the plane where the visual reference module is located and aligned with the center area of ​​the reference pattern. It is used to acquire real-time images containing the reference pattern. The real-time error calculation module is connected to the output of the high-speed image acquisition module at its input end, and is used to receive and process real-time images. The dynamic compensation controller has a first input terminal connected to the output terminal of a high-precision encoder to receive first angular position data, and a second input terminal connected to the output terminal of a real-time error calculation module to receive pixel offset data; the dynamic compensation controller has a built-in calibration mapping model and error compensation algorithm; The real-time error calculation module first preprocesses the real-time image, including grayscale conversion, noise reduction, and edge enhancement; Subsequently, the edge contours of all concentric rings in the reference pattern are extracted using a sub-pixel level edge detection algorithm; Next, the least squares method is used to fit circles to all the extracted edge point sets, and the center coordinates and radius of each circle in the image coordinate system are calculated. The real-time error calculation module performs a weighted average of the coordinates of all fitted circle centers to calculate the actual imaging center coordinates of the reference pattern in the current image. Finally, the actual imaging center coordinates are compared with the theoretical center coordinates of the image to calculate the pixel offsets of the two in the X and Y axes of the image coordinate system. The calibration mapping model is established in advance through a calibration process, which drives the platform to rotate by multiple known angles, and the real-time error calculation module records the corresponding pixel offset at each angle, thereby constructing a nonlinear mapping relationship database between pixel offset and the theoretical rotation angle of the platform. During system operation, the dynamic compensation controller executes the following process: First, based on the currently received first angle position data, the calibration mapping model is queried to predict the expected pixel offset caused by the mechanical deformation of the platform at the current angle; Simultaneously, the actual pixel offset calculated in real time is compared with the expected pixel offset by a vector difference operation to obtain the instantaneous dynamic error vector; The dynamic compensation controller inputs the instantaneous dynamic error vector into the error compensation algorithm; The error compensation algorithm first estimates the equivalent angular deviation of the error propagation to the end of the load based on the platform's current angular velocity and angular acceleration using a second-order dynamic model; Then, this equivalent angle deviation is added to the first angle position data to generate the second angle position data after real-time compensation; Finally, the dynamic compensation controller generates a correction control signal for the drive motor based on the difference between the second angle position data and the system target position command.

2. The hollow rotating platform positioning system based on intelligent image recognition according to claim 1, characterized in that, In the reference pattern of the visual reference module, the encoding mark is composed of micro QR codes arranged along a specific radius or light and dark stripes with a specific width-to-width ratio. After completing the center calculation, the real-time error calculation module further identifies and decodes the coded mark; the decoded information includes the absolute coded sequence number of the mark in the reference pattern. The system achieves absolute counting and coarse positioning of the platform's rotation number by comparing the absolute encoded sequence number decoded from consecutive image frames and combining it with an image matching algorithm.

3. The hollow rotating platform positioning system based on intelligent image recognition according to claim 2, characterized in that, The calibration mapping model is constructed using an adaptive learning mechanism. Specifically, after each successful positioning movement and stabilization, the system records the actual pixel offset at the stable position and adds it as a new calibration data point to the mapping relationship database. For newly added data points and historical data points, a neural network based on radial basis functions is used for surface fitting, and the calibration mapping model is dynamically updated.

4. The hollow rotating platform positioning system based on intelligent image recognition according to claim 3, characterized in that, The error compensation algorithm further includes an error filtering and prediction unit; This unit receives the instantaneous dynamic error vector sequence from historical time series, and first uses a Kalman filter to filter out measurement noise to obtain a clean error estimate. Then, based on the cleanliness error estimates at the current and previous two time points, the error change trend in the next control cycle is predicted using linear extrapolation or quadratic polynomial fitting methods. The dynamic compensation controller incorporates the predicted future error trend into the calculation of the equivalent angle deviation in advance.

5. The hollow rotating platform positioning system based on intelligent image recognition according to claim 4, characterized in that, A coaxial light source is provided between the high-speed image acquisition module and the visual reference module; the coaxial light source emits uniform parallel light that illuminates the surface of the visual reference module perpendicularly.

6. The hollow rotating platform positioning system based on intelligent image recognition according to claim 5, characterized in that, The system operates on a multi-rate control architecture; The architecture includes a high-speed closed loop and a low-speed closed loop; the feedback signal of the high-speed closed loop comes from the first angular position data of the high-precision encoder. The low-speed closed-loop feedback signal originates from the second angle position data generated by the dynamic compensation controller. The control outputs of the two closed loops are superimposed before the current loop.

7. The hollow rotating platform positioning system based on intelligent image recognition according to claim 6, characterized in that, The model of the Kalman filter treats the real error as a process of uniform change and models the measurement noise as Gaussian white noise. The filtering process is carried out iteratively. The prediction step is based on the state estimate of the previous time step, while the update step integrates the current measurement value.

8. The hollow rotating platform positioning system based on intelligent image recognition according to claim 7, characterized in that, The high-speed closed loop employs proportional-integral-derivative control and a notch filter. The low-speed closed-loop position controller receives the difference between the system target position command and the second angle position data, and generates a compensation torque command after processing by the proportional-integral controller.

Images