Automatic control method and system for portable air cylinder assembling platform

Abstract

The invention relates to the technical field of industrial control systems, and discloses an automatic control method and system for a portable cylinder assembly platform, and the method comprises the steps: obtaining an original image and a multi-dimensional force signal, calculating a contour coordinate and a resistance vector, generating a three-dimensional fusion tensor, and extracting initial offset data; constructing a corresponding relation matrix based on the offset data, and performing state estimation to obtain an association vector; identifying abnormal components, performing weighted fusion to generate a deviation vector, and calculating a joint angular velocity to generate an initial compensation instruction; according to the real-time contact force correction path, generating a final compensation instruction; feedback data are collected and executed, the lag trend is analyzed, and the real-time correction control quantity is output. According to the method, fusion processing of visual information and force sense information can be realized, so that the dynamic compensation capability in the assembly process is improved.

Description

Technical Field

[0001] This invention relates to the field of industrial control system technology, and in particular to an automated control method and system for a portable cylinder assembly platform. Background Technology

[0002] Achieving a high-response, highly coordinated precision control strategy for portable cylinder assembly platforms has become a key objective for the technological upgrading of industrial control systems.

[0003] Existing methods are based on image-guided assembly control using pre-defined trajectory planning. At the initial stage of the task, this method acquires contour images of the cylinder components using a camera, extracts contour features using the Canny edge detection algorithm, and then estimates the target's relative pose using the PNP attitude calculation algorithm to generate a preliminary spatial alignment path. The control system then generates a sequence of control commands based on this path and drives the actuators to perform interpolated motion to complete the docking assembly. This method decouples visual recognition from force signals, ignoring the synergistic relationship between images and forces during assembly, and cannot determine in real-time force feedback changes caused by minute attitude shifts. Furthermore, once the path is generated, it is difficult to dynamically adjust based on feedback, resulting in the actuators being unable to quickly compensate for changes in resistance or component tilting during assembly.

[0004] In summary, existing methods suffer from insufficient dynamic adaptability due to the decoupling of visual and force information. Summary of the Invention

[0005] This invention provides an automated control method and system for a portable cylinder assembly platform, which integrates visual and force information to improve dynamic compensation capabilities during the assembly process.

[0006] In a first aspect, to solve the above-mentioned technical problems, the present invention provides an automated control method for a portable cylinder assembly platform, comprising: Acquire raw images and multidimensional force signals; Based on the original image and multidimensional force signal, the contour pixel coordinate set and the resistance amplitude vector sequence are calculated, and a three-dimensional fusion tensor is generated according to the contour pixel coordinate set and the resistance amplitude vector sequence. The three-dimensional fusion tensor is then subjected to differential quantization processing to obtain preliminary offset data. Based on the preliminary offset data, interval setting and probability mapping are performed to obtain the corresponding relationship matrix. Based on the corresponding relationship matrix, state estimation and feature construction are performed to obtain the association vector. Based on the correlation vector, abnormal component identification and weight calculation are performed to obtain complementary weight factors. Based on the complementary weight factors, the correlation vector is weighted and fused to obtain a fusion deviation vector. Based on the fusion deviation vector, inverse kinematic mapping is performed to obtain the angular velocity vector of each joint, and path fitting and direction analysis are performed on the angular velocity vector to obtain the initial compensation command; Execute the initial compensation command, collect multi-dimensional real-time contact force data, perform force signal differential and state estimation correction processing on the multi-dimensional real-time contact force data to obtain the optimized motion trajectory, and perform matching degree evaluation and command generation processing on the optimized motion trajectory to obtain the final compensation command sequence. The final compensation instruction sequence is executed, execution feedback data is collected, deviation is calculated on the feedback data to obtain the dynamic lag deviation, trend prediction analysis is performed on the dynamic lag deviation to obtain the nonlinear error trend, and real-time correction output is generated based on the nonlinear error trend.

[0007] Preferably, the step of calculating a set of contour pixel coordinates and a sequence of drag amplitude vectors based on the original image and the multidimensional force signal, generating a three-dimensional fusion tensor according to the set of contour pixel coordinates and the sequence of drag amplitude vectors, and performing differential quantization processing on the three-dimensional fusion tensor to obtain preliminary offset data includes: The Sobel operator is used to calculate the gradient of the original image to obtain image gradient maps in two directions, and a gradient magnitude map is generated based on the image gradient maps. The gradient magnitude map is filtered by a preset gradient threshold to obtain a set of contour pixel coordinates; The mean and standard deviation of the torques of each group of the multidimensional force signals are calculated to form a resistance amplitude vector sequence; The contour pixel coordinate set is projected onto spatial points in a three-dimensional coordinate system to construct a spatial matrix, and the drag amplitude vector is embedded into the spatial matrix to generate a three-dimensional fusion tensor. The element-wise difference operation is performed between the three-dimensional fusion tensor and the preset standard pose tensor to obtain the difference tensor. The difference tensor is then flattened into a one-dimensional vector and normalized and its magnitude is calculated to output preliminary offset data.

[0008] Preferably, the step of setting intervals and performing probability mapping based on the preliminary offset data to obtain a corresponding relationship matrix, and then performing state estimation and feature construction based on the corresponding relationship matrix to obtain an association vector, includes: Based on the preliminary offset data, multiple interval boundaries are set, and a two-dimensional partitioned network is generated based on the interval boundaries. Based on the two-dimensional partitioning network and the preliminary offset data, joint probability statistics are performed to generate a corresponding relationship matrix; Kalman filtering estimation optimization is performed on the correspondence matrix to obtain state estimates, and vector generation is performed on the state estimates to obtain a two-dimensional state estimate vector. The two-dimensional state estimation vector is normalized and its features are constructed to obtain the correlation vector.

[0009] Preferably, the step of identifying and calculating the weights of abnormal components based on the correlation vector to obtain a complementary weight factor, and then performing weighted correction and vector fusion on the correlation vector based on the complementary weight factor to obtain a fusion deviation vector, includes: The Euclidean distance between the associated vector and the preset reference vector is calculated. If the Euclidean distance exceeds the preset deviation threshold, the dimension with the largest difference from the reference vector in the associated vector is identified as the index position of the abnormal component. Based on the index position, the corresponding associated components are extracted, and the complementary weight factor is calculated based on the associated components. Based on the complementary weight factors, a diagonal weighted matrix is ​​constructed, and weight correction and splicing are performed according to the diagonal weighted matrix and the correlation vector to obtain the fusion deviation vector.

[0010] Preferably, the step of performing inverse kinematic mapping based on the fused deviation vector to obtain the angular velocity vector of each joint, and performing path fitting and direction analysis on the angular velocity vector to obtain the initial compensation command, includes: Based on the fusion deviation vector, coordinate transformation calculation is performed to obtain three-dimensional gap change data in the local coordinate system; Based on the three-dimensional gap change data, the Jacobian matrix of the robotic arm in its current posture is constructed, and the pseudo-inverse calculation of the Jacobian matrix is ​​performed to obtain the angular velocity vector of each joint. Based on the angular velocity vector, an integration operation is performed at preset time intervals to generate a joint angle sequence, and a motion compensation path is formed by fitting the joint angle sequence. The magnitude and azimuth of the tangent vector of the initial segment of the motion compensation path are calculated, and an initial compensation command for driving the actuator is generated based on the magnitude and azimuth.

[0011] Preferably, the initial compensation command is executed, multi-dimensional real-time contact force data is collected, force signal differential and state estimation correction processing is performed on the multi-dimensional real-time contact force data to obtain an optimized motion trajectory, and matching degree evaluation and command generation processing are performed on the optimized motion trajectory to obtain a final compensation command sequence, including: Execute the initial compensation command, collect multi-dimensional real-time contact force data, and perform axis-by-axis differential processing on the multi-dimensional real-time contact force data according to a preset time interval to form a force signal deviation sequence; The force signal deviation sequence is estimated by Kalman filtering to obtain the predicted state vector; The predicted state vector is sequentially mapped to a displacement correction vector in three-dimensional space and superimposed on the coordinates of each node in the motion compensation path to obtain the optimized motion trajectory. The optimized motion trajectory is sampled at equal intervals to obtain a discrete node set. The Euclidean distance is calculated between the discrete node set and the target location to obtain the Euclidean distance, and then the Euclidean distance is normalized to obtain the matching degree. If the matching degree is lower than the preset convergence threshold, the optimized motion trajectory is updated and the driving signal is converted to obtain the final compensation instruction sequence.

[0012] Preferably, the step of executing the final compensation instruction sequence, collecting execution feedback data, calculating the deviation from the feedback data to obtain a dynamic lag deviation, performing trend prediction analysis on the dynamic lag deviation to obtain a nonlinear error trend, and generating a real-time correction output based on the nonlinear error trend includes: The execution feedback data includes: real-time position data and torque data; Based on the real-time position data and torque data, the timing alignment deviation is calculated to obtain the dynamic hysteresis deviation. Based on the dynamic lag deviation, normalization and coordination degree calculation are performed to obtain the dynamic coordination completion degree; When the dynamic coordination completion rate is less than the preset coordination threshold, the least squares method is used to predict the trend of the dynamic lag deviation to obtain the nonlinear error trend. Based on the aforementioned nonlinear error trend, reverse compensation control quantity superposition processing is performed to obtain real-time corrected output.

[0013] Secondly, the present invention provides an automated control system for a portable cylinder assembly platform, comprising: The data acquisition module is used to acquire raw images and multidimensional force signals; The preliminary offset module is used to calculate the contour pixel coordinate set and the resistance amplitude vector sequence based on the original image and the multidimensional force signal, and generate a three-dimensional fusion tensor according to the contour pixel coordinate set and the resistance amplitude vector sequence. The three-dimensional fusion tensor is then subjected to differential quantization processing to obtain preliminary offset data. The association module is used to set intervals and perform probability mapping based on the preliminary offset data to obtain a corresponding relationship matrix, and to perform state estimation and feature construction based on the corresponding relationship matrix to obtain an association vector; The fusion deviation vector module is used to identify and calculate the weights of abnormal components based on the correlation vector to obtain complementary weight factors. Based on the complementary weight factors, the correlation vector is weighted and fused to obtain a fusion deviation vector. The initial compensation command module is used to perform inverse kinematic mapping based on the fusion deviation vector to obtain the angular velocity vector of each joint, and to perform path fitting and direction analysis on the angular velocity vector to obtain the initial compensation command. The final compensation instruction module is used to execute the initial compensation instruction, collect multi-dimensional real-time contact force data, perform force signal differential and state estimation correction processing on the multi-dimensional real-time contact force data to obtain the optimized motion trajectory, and perform matching degree evaluation and instruction generation processing on the optimized motion trajectory to obtain the final compensation instruction sequence. The real-time correction output module is used to execute the final compensation instruction sequence, collect execution feedback data, perform deviation calculation on the feedback data to obtain the dynamic lag deviation, perform trend prediction analysis on the dynamic lag deviation to obtain the nonlinear error trend, and generate real-time correction output based on the nonlinear error trend.

[0014] Compared with the prior art, the present invention has the following beneficial effects: (1) The present invention fuses the contour pixel coordinates and contact force signals to generate a three-dimensional tensor, and extracts the initial offset data through difference and normalization operations. This can effectively quantify the collaborative error between the component position and the force state in the early stage of assembly, improve the system's recognition accuracy of assembly offset and initial contact anomaly, and reduce the risk of jamming caused by initial state deviation.

[0015] (2) By constructing a partitioned network from the initial offset data, performing probability mapping and state filtering operations, a two-dimensional estimation vector with directionality and amplitude structure is formed. Then, an association vector is constructed to express the attitude and force relationship, thereby improving the system's ability to fine-grained characterize different assembly states, so that the control commands can more accurately match the current assembly state.

[0016] (3) Identify the most drastically changing component in the offset vector, calculate the complementary weight factor and construct the weighted matrix, perform vector correction and fusion to obtain an error expression vector that is closer to the actual assembly state.

[0017] (4) After the end feedback acquisition is performed, the system will form the hysteresis deviation by combining the position error and the hysteresis time. The trend change is estimated by the least squares method, and the reverse compensation displacement is superimposed on the compensation command of the next cycle to realize the early correction of the dynamic deviation of the assembly process, effectively reducing the hysteresis response and the number of repeated adjustments of the actuator. Attached Figure Description

[0018] Figure 1 This is a schematic flowchart of an automated control method for a portable cylinder assembly platform provided in the first embodiment of the present invention; Figure 2 This is a schematic diagram of the automated control system structure for a portable cylinder assembly platform provided in the second embodiment of the present invention. Detailed Implementation

[0019] 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.

[0020] Reference Figure 1 The first embodiment of the present invention provides an automated control method for a portable cylinder assembly platform, comprising the following steps: S11, acquire the original image and multidimensional force signal; S12, based on the original image and multidimensional force signal, calculate the contour pixel coordinate set and the resistance amplitude vector sequence, and generate a three-dimensional fusion tensor according to the contour pixel coordinate set and the resistance amplitude vector sequence. Perform differential quantization processing on the three-dimensional fusion tensor to obtain preliminary offset data. S13, Based on the preliminary offset data, interval setting and probability mapping are performed to obtain the corresponding relationship matrix, and based on the corresponding relationship matrix, state estimation and feature construction are performed to obtain the association vector; S14. Based on the correlation vector, perform abnormal component identification and weight calculation to obtain complementary weight factors. Based on the complementary weight factors, perform weighted correction and vector fusion on the correlation vector to obtain a fusion deviation vector. S15, based on the fusion deviation vector, perform inverse kinematic mapping to obtain the angular velocity vector of each joint, and perform path fitting and direction analysis on the angular velocity vector to obtain the initial compensation command; S16, execute the initial compensation command, collect multi-dimensional real-time contact force data, perform force signal differential and state estimation correction processing on the multi-dimensional real-time contact force data to obtain the optimized motion trajectory, and perform matching degree evaluation and command generation processing on the optimized motion trajectory to obtain the final compensation command sequence. S17, execute the final compensation instruction sequence, collect execution feedback data, perform deviation calculation on the feedback data to obtain the dynamic lag deviation, perform trend prediction analysis on the dynamic lag deviation to obtain the nonlinear error trend, and generate real-time correction output based on the nonlinear error trend.

[0021] In step S11, the original image and multidimensional force signal are acquired.

[0022] It is worth noting that a CCD image sensor is installed at the end of the platform actuator, set to a frame interval of 10-50ms. The acquired raw images are transmitted to the processing unit via USB 3.0, with an image resolution of no less than 640×480 pixels. The acquisition of multi-dimensional force signals is achieved through a six-dimensional force / torque sensor installed at the end tool flange, which outputs three components in real time: Mx, My, and Mz (torque). This sensor converts the analog signal into a digital signal through an analog-to-digital converter and transmits it synchronously to the control system at a sampling frequency of over 1kHz.

[0023] In step S12, based on the original image and the multidimensional force signal, a set of contour pixel coordinates and a sequence of drag amplitude vectors are calculated. A three-dimensional fusion tensor is generated based on the set of contour pixel coordinates and the sequence of drag amplitude vectors. The three-dimensional fusion tensor is then subjected to differential quantization processing to obtain preliminary offset data, including: The Sobel operator is used to calculate the gradient of the original image to obtain image gradient maps in two directions, and a gradient magnitude map is generated based on the image gradient maps. The gradient magnitude map is filtered by a preset gradient threshold to obtain a set of contour pixel coordinates; The mean and standard deviation of the torques of each group of the multidimensional force signals are calculated to form a resistance amplitude vector sequence; The contour pixel coordinate set is projected onto spatial points in a three-dimensional coordinate system to construct a spatial matrix, and the drag amplitude vector is embedded into the spatial matrix to generate a three-dimensional fusion tensor. The element-wise difference operation is performed between the three-dimensional fusion tensor and the preset standard pose tensor to obtain the difference tensor. The difference tensor is then flattened into a one-dimensional vector and normalized and its magnitude is calculated to output preliminary offset data.

[0024] It's worth noting that the Sobel operator is used to perform first-order gradient operations in the x and y directions on the grayscale layers of the original image, obtaining gradient maps in both directions. Then, the gradient magnitude map for each pixel is obtained by taking the square root of the sum of squares. For example, consider a 3x3 pixel region in the image, whose grayscale value is: Applying the Sobel algorithm's x-direction convolution kernel [-10 1; -2 0 2; -1 0 1] to this region, we get Gx = (1×10 + 2×10 + 1×10) − (1×10 + 2×200 + 1×10) = 40 − 420 = −380. Using [1 2 1; 0 0 0; -1 -2 -1] in the y-direction, we get Gy = (1×10 + 2×10 + 1×10) − (1×10 + 2×10 + 1×10) = 40 − 40 = 0. Based on this, we can calculate the gradient magnitude of this pixel. .

[0025] Subsequently, a gradient threshold is applied to the gradient magnitude map to obtain a set of contour pixel coordinates. Specifically, the gradient threshold is set to the region where the global maximum gradient value of the image is 30%. Pixel regions in the gradient magnitude map that are greater than 30% of the maximum gradient value are output as a set of contour pixel coordinates.

[0026] Next, the multidimensional force signals are aligned with the image frame timestamps according to the sampling time. The system selects three sets of torque signals (Mx, My, and Mz) as the analysis targets, and uses the image frame period as the window width. A sliding window is used to statistically analyze the torque signals within each period. Specifically, every 10 sampling points form a window, and the mean and standard deviation are calculated, outputting a drag amplitude vector sequence. For example, if the image frame period is 40 milliseconds, and the force sensor collects 40 data points during this period, the system uses every 10 points as a sliding window to calculate the mean and standard deviation of Mx, My, and Mz in each group. If the value of Mx in a certain group is [0.12, 0.15, 0.11, ..., 0.14] N / m, then the mean of that group is 0.13 N / m, and the standard deviation is 0.012 N / m.

[0027] Next, the set of contour pixel coordinates obtained from image processing is projected onto a three-dimensional coordinate system to construct a spatially sparse matrix centered on each pixel. The aforementioned drag amplitude vector is then embedded into the corresponding coordinate layer according to the time synchronization principle, thereby forming a three-dimensional fusion tensor. For example, image processing extracts a set of edge contour pixels, such as (320,240), (321,240), and (322,241), representing the two-dimensional position of the component boundary in the image. These pixels are projected onto a three-dimensional coordinate system using camera calibration parameters. For example, a contour pixel in the image is (320,240), corresponding to the image center, with units in pixels. The camera intrinsic parameter matrix is ​​set with focal length fx=800, fy=800, and principal point coordinates (cx,cy)=(320,240), resulting in normalized camera coordinates of (0,0). If the depth value corresponding to this point is 150mm (measured by structured light), then the three-dimensional position of this point in the camera coordinate system is (0mm,0mm,150mm). Then, using the camera extrinsic parameter matrix, i.e., the rotation and translation matrix from the camera to the platform coordinate system, the point is converted into a position in the platform coordinate system, resulting in a corresponding spatial point such as (25.4mm, 10.2mm, 150.0mm). Subsequently, combined with the drag amplitude vector extracted within the frame time period, for example, corresponding to [0.12, 0.09, 0.15] N / m, these three values ​​are embedded into the force information channels of the tensor positions corresponding to the above three spatial points.

[0028] After obtaining the 3D fusion tensor, it is subjected to element-wise interpolation with the preset standard pose tensor to generate a difference tensor. This difference tensor retains the deviation value at each corresponding position in space. Subsequently, the difference tensor is flattened into a one-dimensional vector, and maximum value normalization is performed to map all its elements to the interval [0,1]. Finally, the magnitude of this vector is calculated and output as the preliminary offset data. The standard pose tensor is constructed by recording the image edge coordinates and corresponding stabilizing torque values ​​of the assembly platform under ideal alignment conditions, and averaging them after multiple samplings. This tensor serves as the reference for the fusion tensor of each frame.

[0029] It should be noted that the preset standard attitude tensor is obtained through the following offline calibration steps: The actuator end effector is controlled to move to the theoretical assembly docking pose multiple times (e.g., no less than 50 times). This pose is determined by the platform design model and is considered the ideal alignment state. After each positioning and stabilization, a frame of original image and a set of steady-state multidimensional force signals are simultaneously acquired. The Sobel operator gradient calculation and threshold filtering are performed on each frame of image to obtain the standard contour pixel coordinate set for that test. For each set of steady-state force signals, the mean values ​​of its Mx, My, and Mz components are calculated to constitute the standard drag amplitude vector for that test. The standard contour pixel coordinate set obtained from each test is then calibrated using a fixed... The camera calibration parameters are projected onto the platform coordinate system to form a three-dimensional point cloud. The spatial distribution of the point cloud for all tests is statistically analyzed, and its centroid is taken as the standard spatial point. Based on this, a basic spatial matrix is ​​constructed. The average value of the standard drag amplitude vector for each test is taken to obtain the standard drag vector. Finally, the component values ​​of the standard drag vector are used as additional channel information and filled into the data channels corresponding to each spatial point in the basic spatial matrix to form a preset standard attitude tensor. The first and second dimensions of this tensor represent the spatial position (X,Y), and the third dimension represents the information channel (for example, channel 0 is the spatial Z coordinate, and channels 1-3 are the standardized values ​​of Mx, My, and Mz).

[0030] In step S13, based on the preliminary offset data, interval setting and probability mapping are performed to obtain a correspondence matrix. Then, based on the correspondence matrix, state estimation and feature construction are performed to obtain an association vector, including: Based on the preliminary offset data, multiple interval boundaries are set, and a two-dimensional partitioned network is generated based on the interval boundaries. Based on the two-dimensional partitioning network and the preliminary offset data, joint probability statistics are performed to generate a corresponding relationship matrix; Kalman filtering estimation optimization is performed on the correspondence matrix to obtain state estimates, and vector generation is performed on the state estimates to obtain a two-dimensional state estimate vector. The two-dimensional state estimation vector is normalized and its features are constructed to obtain the correlation vector.

[0031] It is worth noting that, based on the value range of the initial offset data, multiple interval boundaries for image intensity and resistance amplitude were set, and a two-dimensional partitioned grid was established. Using image features as the horizontal axis and force characteristics as the vertical axis, intervals such as [0,0.2], [0.2,0.5], and [0.5,0.8] for image gradient intensity and [0,0.1], [0.1,0.3], and [0.3,0.6] for resistance values ​​were divided. Subsequently, the contour feature points extracted from each frame of the image were combined and statistically analyzed with the resistance amplitude vector within that time period. The number of times the image and force data simultaneously fell within each grid interval was calculated, divided by the total number of samples, and the joint probability value was obtained and filled into the corresponding relationship matrix. For example, the interval for image gradient intensity was set to [0.2,0.5], and the interval for resistance amplitude was set to [0.1,0.3]. In a certain frame of data, 100 contour pixels were extracted from the image, and the corresponding resistance vector for that frame also contained 100 sample points. Statistical results show that 24 samples simultaneously satisfy the condition that the image gradient intensity falls within [0.2, 0.5] and the corresponding resistance amplitude is between [0.1, 0.3]. Therefore, the joint probability value of this interval combination is 24 divided by the total number of samples (100), which is 0.24.

[0032] Subsequently, Kalman filtering estimation optimization is performed on the corresponding relation matrix to obtain a two-dimensional state estimation vector. Specifically, after each frame of image and force feature input, the system updates the state estimation of the current frame based on the residual between the state of the previous frame and the current observation, combined with the Kalman gain. The update formula is as follows: in, This is the state estimate for the current frame. This is the state estimate from the previous frame. Here is the Kalman gain matrix. For the observation mapping matrix, This is the image-force perception observation vector. For example, if the predicted pose offset angle in the previous frame is 2.50 degrees, and the observation value obtained in the current frame through image-force perception feature fusion is 2.90 degrees, the Kalman gain is 0.6, and the observation mapping matrix... Set to 1, the residual is first calculated to be 0.40 degrees. Then, it is multiplied by the gain to obtain a correction of 0.24 degrees. The correction is added to the original prediction value to obtain the state estimate of the current frame as 2.74 degrees. Finally, the estimated angle offset value and the drag amplitude are combined to form a two-dimensional state estimation vector.

[0033] And the observation mapping matrix It is determined by the relationship between the system state vector and the observation vector. For example, when the state vector includes attitude angle and angular velocity, while the observation only includes angle, then... Let be a row vector, indicating that only the first component of the state is extracted for comparison with the observation. If the observation and state have the same dimension and correspond one-to-one, then... It is an identity matrix.

[0034] Kalman gain matrix The calculation formula is: in, The current state error covariance matrix is... Set it to 0.1. And... The calculation formula is: ,in, The current state error covariance matrix is... The state error covariance matrix of the previous frame. Set as , For inter-frame sampling time, Set it to 0.01.

[0035] Subsequently, the two-dimensional state estimation vector is normalized and its features are constructed to obtain the correlation vector. Specifically, the angle offset value and drag amplitude are normalized to the maximum and minimum, and then the correlation vector is constructed, including the product of the angle offset value and drag amplitude, the product of the square of the angle offset value and drag amplitude, and the product of the square of the angular velocity and drag amplitude. For example, if the maximum allowable value of the attitude offset angle is 5 degrees, then when the offset angle of the current frame is 2.85 degrees, its normalization result is 0.57. If the design upper limit of the drag amplitude is 0.2 N / m, then when the current drag is 0.14 N / m, its normalization result is 0.7. The coupling term of angle offset value and drag amplitude is calculated as 0.57 * 0.70 = 0.399.

[0036] In step S14, based on the correlation vector, abnormal component identification and weight calculation are performed to obtain a complementary weight factor. Based on the complementary weight factor, the correlation vector is weighted and fused to obtain a fusion deviation vector, including: The Euclidean distance between the associated vector and the preset reference vector is calculated. If the Euclidean distance exceeds the preset deviation threshold, the dimension with the largest difference from the reference vector in the associated vector is identified as the index position of the abnormal component. Based on the index position, the corresponding associated components are extracted, and the complementary weight factor is calculated based on the associated components. Based on the complementary weight factors, a diagonal weighted matrix is ​​constructed, and weight correction and splicing are performed according to the diagonal weighted matrix and the correlation vector to obtain the fusion deviation vector.

[0037] It's worth noting that the Euclidean distance is calculated between the associated vector and the preset benchmark vector. This involves summing the squares of the differences between two vectors of the same dimension and then taking the square root. For example, if the current vector is [0.32, 0.45, 0.29] and the benchmark is [0.30, 0.40, 0.31], the distance is approximately 0.06. If this distance is greater than the preset deviation threshold of 0.05, then the system compares each item in the current vector with the benchmark, extracting the dimensions with the largest deviations as the index positions of the outlier components. For instance, if the third dimension has the largest difference, the system records the index position of that dimension as a target for subsequent correction. Next, the system compares this value with a large amount of pre-recorded normal operating data, using a Gaussian distribution fitting method to evaluate the probability density of this value under normal conditions.

[0038] Specifically, the system first calls the normal distribution dataset, performs statistical fitting on the component to obtain the mean and standard deviation, and constructs a probability density function based on these. The current outlier is then substituted into this function to calculate the density value of the distribution. The probability density function is as follows: in, For the outlier component values ​​to be calculated, This is the mean of the component in the normal data. Let the standard deviation of this component be the standard deviation of the normal data. For example, suppose the value of an outlier component is [value]. The mean value of this component is obtained from historical normal data. The standard deviation is Substituting into the probability density function, we get a probability density of 1.93. Then, we need to normalize this probability density to its maximum value. Assuming the maximum density in the historical distribution of this outlier component is 2.5, the normalized density value is 0.772. Therefore, the weighting factor for this outlier component is 1 minus this normalized density value, i.e., 1 - 0.772 = 0.228. Its value is between 0 and 1; a larger value indicates a more outlier current value. To further smooth the control, the final value needs to be multiplied by a gain coefficient less than 1 (e.g., 0.5) to prevent over-correction.

[0039] The preset baseline vector is obtained through statistical analysis of historical data under normal operating conditions. Specifically, with the device in a standard posture and without abnormal interference, multiple frames of images and force data are continuously acquired, and corresponding correlation vectors are generated according to the aforementioned method. After time alignment of the correlation vector components in the same dimension, their means are calculated separately, and the means of all dimensions together constitute the preset baseline vector. The preset deviation threshold is determined based on the statistical distribution of historical normal data. Statistical analysis of the Euclidean distance between the correlation vectors and the baseline vector corresponding to a large number of normal samples reveals that the distance distribution approximately follows a normal distribution; therefore, the mean plus twice the standard deviation is selected as the deviation threshold.

[0040] Next, a diagonal weighted matrix is ​​constructed using the weight factors of all outlier components. The diagonal elements of this matrix take the weight values ​​at the outlier dimensions, while the remaining elements are constants of 0.95, indicating a higher retention rate. This weighted matrix is ​​then multiplied with the original vector to obtain the corrected component. This corrected component is then concatenated with the original unmarked components to form the fused bias vector. For example, the original vector is [0.42, 0.25, 0.31], where the first dimension is marked as outlier, corrected to 0.16, and concatenated with the last two dimensions to form a new vector [0.16, 0.25, 0.31]. The system outputs this vector as the fused bias vector.

[0041] In step S15, inverse kinematic mapping is performed based on the fusion deviation vector to obtain the angular velocity vectors of each joint. Then, path fitting and direction analysis are performed on the angular velocity vectors to obtain the initial compensation command, including: Based on the fusion deviation vector, coordinate transformation calculation is performed to obtain three-dimensional gap change data in the local coordinate system; Based on the three-dimensional gap change data, the Jacobian matrix of the robotic arm in its current posture is constructed, and the pseudo-inverse calculation of the Jacobian matrix is ​​performed to obtain the angular velocity vector of each joint. Based on the angular velocity vector, an integration operation is performed at preset time intervals to generate a joint angle sequence, and a motion compensation path is formed by fitting the joint angle sequence. The magnitude and azimuth of the tangent vector of the initial segment of the motion compensation path are calculated, and an initial compensation command for driving the actuator is generated based on the magnitude and azimuth.

[0042] It is worth noting that the fusion deviation vector is transformed from the global coordinate system to the local coordinate system where the end effector is currently located. This transformation is based on the rotation matrix constructed from the real-time posture of the end effector. That is, by reading the current joint angle or visual recognition results, the orientation information of the end effector in three-dimensional space is obtained, and then a rotation reference for the local coordinate system is formed. In specific operation, the system multiplies the global fusion deviation vector with the transpose of this rotation matrix to complete the coordinate transformation. Taking a surgical robot as an example, if the fusion deviation vector is [0.16, 0.25, 0.31], it represents the spatial offset in the global coordinate system. After projection calculation based on the rotation matrix constructed from the current posture of the end effector, the three-dimensional gap change data in the local coordinate system is obtained, such as an x-axis offset of 0.25 mm, a y-axis offset of -0.16 mm, and a z-axis offset of 0.31 mm.

[0043] The gap variation data of the end effector in three-dimensional space is converted into the required motion amplitude of each joint of the robotic arm. To this end, a Jacobian matrix is ​​first constructed based on the current posture of the robotic arm. The construction process of the Jacobian matrix involves sequentially establishing coordinate transformation matrices from each joint to the next, following the structural order of the robotic arm's joints. These transformation matrices contain rotation and translation information. By multiplying these transformation matrices together, a homogeneous transformation matrix of the end effector relative to the base is obtained, thus yielding its position coordinates and orientation vector. After completing the forward kinematics, the partial derivatives of the end effector position with respect to each joint variable are further calculated, forming a partial derivative matrix. Each column of this matrix represents the influence of that joint on the end effector position and posture; these columns, arranged sequentially, constitute the Jacobian matrix. After obtaining the Jacobian matrix, the system performs a pseudo-inverse operation on it. The purpose is to solve for a set of minimum norm joint angular velocity solutions in redundant or underactuated systems, enabling the end effector to move in the desired direction and velocity. The pseudo-inverse calculation uses singular value decomposition (SVD) to decompose the Jacobian matrix into multiple factors, which are then reversed and recombined to solve for the optimal inverse solution. Finally, the system multiplies the gap change data with this pseudo-inverse matrix, outputting a joint angular velocity vector. For example, if the 3D gap change is [0.14, 0.22, 0.28] millimeters, the system calculates the angular velocities of the six joints to be [0.18, 0.12, 0.05, 0.02, 0.01, 0.00] radians per second. Subsequently, the joint angular velocity vector is integrated at set time intervals to generate a joint angle sequence, and a motion compensation path is constructed. Specifically, assuming the time interval is every 0.1 seconds, the discrete velocity is converted into a continuous angular trajectory through multiple integration operations, for example, forming the joint 1 angle sequence [0, 0.018, 0.036, ...]. This path, after linear interpolation fitting, will be constructed as a motion compensation path.

[0044] Finally, the motion compensation path is planned. Specifically, two adjacent key nodes are selected in the initial segment of the compensation path. The spatial coordinates of the latter node are subtracted from the spatial coordinates of the former node to obtain a set of three-dimensional values, which constitute the tangent vector. This tangent vector is then decomposed into magnitude and direction. The magnitude of the tangent vector is obtained by summing the squares of each of the three-dimensional components and taking the square root. This magnitude represents the overall displacement intensity of the compensation action in space. Simultaneously, the azimuth angle is calculated by performing inverse trigonometric function operations on the ratio of the components of the tangent vector in the plane projection direction, which is used to determine the execution direction of the compensation action. For example, if the tangent vector calculated in the initial segment of the path is [0.2, 0.35, 0.3], this vector represents the comprehensive change trend of the end effector in the three directions, with a magnitude of approximately 0.52, corresponding to the initial amplitude of the compensation action. The azimuth angle calculated based on the lateral and longitudinal components is approximately 60 degrees, corresponding to the execution direction of the compensation action. Based on this, the system will move along a 60-degree direction with a movement range of 0.52 units, as the initial compensation command output.

[0045] In step S16, the initial compensation command is executed, multi-dimensional real-time contact force data is collected, and force signal differential and state estimation correction processing is performed on the multi-dimensional real-time contact force data to obtain an optimized motion trajectory. Then, matching degree evaluation and command generation processing are performed on the optimized motion trajectory to obtain the final compensation command sequence, including: Execute the initial compensation command, collect multi-dimensional real-time contact force data, and perform axis-by-axis differential processing on the multi-dimensional real-time contact force data according to a preset time interval to form a force signal deviation sequence; The force signal deviation sequence is estimated by Kalman filtering to obtain the predicted state vector; The predicted state vector is sequentially mapped to a displacement correction vector in three-dimensional space and superimposed on the coordinates of each node in the motion compensation path to obtain the optimized motion trajectory. The optimized motion trajectory is sampled at equal intervals to obtain a discrete node set. The Euclidean distance is calculated between the discrete node set and the target location to obtain the Euclidean distance, and then the Euclidean distance is normalized to obtain the matching degree. If the matching degree is lower than the preset convergence threshold, the optimized motion trajectory is updated and the driving signal is converted to obtain the final compensation instruction sequence.

[0046] It is worth noting that the actuator is driven to perform the corresponding compensation action according to the initial compensation command, and multi-dimensional real-time contact force data is collected synchronously during the execution process. The multi-dimensional real-time contact force data is obtained by a multi-axis force sensor set at the end of the actuator. Subsequently, the multidimensional contact force data is differentially processed at 0.1-second intervals. Specifically, the force values ​​at two adjacent sampling times are subtracted axis by axis to obtain a force signal deviation sequence. Then, the deviation sequence formed by the difference in force signals from the previous time to the current time is input into the Kalman filter estimation process. Combined with the previous state estimate, the predicted state vector at the current time is calculated. Specifically, the state vector is defined as the position deviation in three-dimensional space. The state transition model adopts a constant velocity model. The observation vector is the expected position adjustment amount calculated based on the force signal deviation sequence (the force deviation can be converted into a position adjustment amount through a fixed c). The Kalman filter recursively obtains a smoother and more accurate position deviation state estimate at the current time based on the state estimate at the previous time and the force observation value at the current time, which is the predicted state vector. The predicted state vector is mapped to a displacement correction vector by directly using the state estimate (whose physical unit is a length unit, such as millimeters) as the displacement correction amount to be superimposed in the current sampling period.

[0047] Next, the spatial displacement correction vector is applied to the motion compensation path described earlier. All nodes along the path are scanned point-by-point, and the three components of the spatial correction vector are superimposed onto the X, Y, and Z coordinates of each node. Mathematically, this operation is equivalent to translating the original path along the correction direction by a small distance, but in practice, it is performed node-by-node with precise superposition. For example, if the coordinates of a node in the original path are (1.00, 2.00, 3.00), and the correction vector is (1.07, 0.76, 0.55), then the updated node coordinates are (2.07, 2.76, 3.55). After updating all nodes along the path, this updated coordinates are output as the optimized motion trajectory.

[0048] Subsequently, the optimized motion trajectory is sampled according to a preset time sampling interval, and spatial coordinate points in the path are extracted to form a discrete node set. For example, if the sampling interval is set to every 0.1 seconds and the actuator runtime is 2 seconds, then 20 spatial coordinate points will be extracted at equal intervals throughout the entire trajectory as a discrete point set.

[0049] Next, the Euclidean distance between the discrete point set and the target position is calculated, and the Euclidean distance is normalized to its maximum and minimum values ​​before being output as the matching degree. Specifically, the Euclidean distance between each node in the discrete node set and the target position is calculated, resulting in a set of distance values. The average value of this set of distance values ​​is taken as the overall position error measure of the current trajectory. This average error value is then normalized, with the normalization denominator being the maximum allowable average error set according to assembly tolerance requirements. The normalized result is the matching degree, and the closer the value is to 1, the higher the degree of match between the trajectory and the target. For example, if the current coordinates of a node are [1.068, 2.000, 0.982] and the target is [1.00, 2.00, 1.00], the corresponding Euclidean distance is 0.0703 mm. Then, normalization is performed based on the maximum and minimum distance values ​​among the nodes in each set of discrete points. Assuming the maximum value is 0.0712 mm and the minimum value is 0.0481 mm, the normalized distance of the current node is 0.961, and this value is output as the matching degree.

[0050] The preset convergence threshold is set based on statistical analysis of the matching results of a large amount of historical trajectory data. The analysis revealed that the data follows a normal distribution, with a mean of approximately 0.97 and a standard deviation of approximately 0.01. The mean plus twice the standard deviation is selected as the preset convergence threshold, i.e., 0.97 - 2 × 0.01 = 0.95. The target position is preset during the task initialization phase, achieved through manual annotation, visual recognition, and importing standard surgical postures from a clinical database. This serves as a spatial posture reference for the final destination of the end effector.

[0051] When the matching degree is lower than the preset convergence threshold of 0.95, further correction is needed to the optimized motion trajectory. Specifically, multiple frames of image data during the execution process are called, and the SIFT algorithm is used to identify high-contrast regions such as edges, corners, or textures in the image and extract their two-dimensional feature coordinates. Then, triangulation is performed based on the viewpoint differences between adjacent frames to restore the two-dimensional feature points from multiple viewpoints to points in three-dimensional space, ultimately forming a visual feature point cloud. The SIFT algorithm works by constructing a scale space in the image, detecting extreme points at different scales as potential keypoints, calculating the gradient direction and magnitude in their neighborhood, and extracting feature descriptors with rotation and scale invariance.

[0052] It should be noted that during the path update phase, in order to obtain more accurate spatial positioning feedback, the system calls multiple consecutive frames of images; it uses feature point extraction and matching algorithms (such as SIFT, ORB or AKAZE algorithms) to detect and match stable feature points in adjacent images. Using the calibrated camera parameters and these matching point pairs, it uses the structure of motion reconstruction (SFM) or visual odometry principle to calculate the relative motion of the end effector between adjacent frames, thereby generating a sparse visual feature point cloud to help correct the spatial coordinates of motion trajectory nodes.

[0053] Subsequently, the three-dimensional spatial deviation between the visual feature point cloud and the optimized motion trajectory nodes is calculated point by point, and these deviation data are organized into a spatial deviation matrix containing error components in each direction. For example, at a certain node position, if the visual feature point cloud displays an offset of 0.08 mm in the x-axis direction and -0.03 mm in the z-axis direction, then the deviation vector of that node is [0.08, 0, -0.03]. The system superimposes these deviation vectors onto the original trajectory node coordinates to form an updated trajectory node sequence.

[0054] Finally, the system converts the updated trajectory node sequence into pulse signals that the servo motor can recognize as the final compensation command sequence. For example, if a joint needs to perform a 1 mm displacement, the system converts this value into an equivalent number of encoder pulses, such as 10 pulses for every 0.01 mm, thus generating 1000 pulses for driving.

[0055] In step S17, the final compensation instruction sequence is executed, execution feedback data is collected, deviation calculation is performed on the feedback data to obtain the dynamic lag deviation, trend prediction analysis is performed on the dynamic lag deviation to obtain the nonlinear error trend, and real-time correction output is generated based on the nonlinear error trend, including: The execution feedback data includes: real-time position data and torque data; Based on the real-time position data and torque data, the timing alignment deviation is calculated to obtain the dynamic hysteresis deviation. Based on the dynamic lag deviation, normalization and coordination degree calculation are performed to obtain the dynamic coordination completion degree; When the dynamic coordination completion rate is less than the preset coordination threshold, the least squares method is used to predict the trend of the dynamic lag deviation to obtain the nonlinear error trend. Based on the aforementioned nonlinear error trend, reverse compensation control quantity superposition processing is performed to obtain real-time corrected output.

[0056] It is worth noting that the final compensation command sequence is sent to the platform actuator. During execution, real-time position and torque data fed back by the actuator are collected synchronously. The position data forms a real-time position sequence in chronological order, and the torque data forms a torque response curve based on the same time base. For example, the compensation command sequence is generated at 10-millisecond intervals, with the timestamp of a certain command being 1.000 seconds. The sampling time points of the position and torque sensors are 0.998 seconds and 1.008 seconds, respectively. Using 1.000 seconds as the reference, linear interpolation is performed on the feedback data corresponding to 0.998 seconds and 1.008 seconds to calculate the actual position and torque values ​​at 1.000 seconds. This set of interpolated feedback data is then matched with the compensation command to complete the time alignment of the command.

[0057] After alignment, the system compares the desired position given in the command with the actual feedback position, calculates the position difference time-by-time, and combines this with the difference between the command issuance time and the actual response time to obtain the dynamic hysteresis deviation. This deviation includes both the magnitude deviation in the displacement direction and the response lag in time. For example, in a certain assembly operation, the command requires the actuator to move 0.8 mm along the spindle direction within 0.5 seconds, but the feedback data shows that the actuator reaches the target position at 0.53 seconds, and the final position deviation is 0.05 mm. In this case, the system records both the time delay and the displacement deviation as the dynamic hysteresis deviation.

[0058] Subsequently, a dynamic coordination completion degree calculation was performed on the dynamic hysteresis deviation. Specifically, the end effector was set to complete a 0.6 mm displacement within 0.5 seconds. Actual feedback data showed that the completion time was 0.51 seconds, resulting in a slight hysteresis of 0.01 seconds. Simultaneously, the maximum response value recorded by the torque sensor was 0.31 N / m, showing a slight overshoot of 0.03 N / m compared to the set value of 0.28 N / m. The position hysteresis and torque deviation were normalized separately, with a maximum allowable hysteresis of 0.05 seconds and a maximum allowable torque deviation of 0.10 N / m. The normalized value for position hysteresis was then 0.01 ÷ 0.05 = 0.20, and the normalized value for torque deviation was 0.03 ÷ 0.10 = 0.30. The two normalized values ​​were then averaged with equal weights and inverted to obtain the dynamic coordination completion degree. For example, the dynamic coordination completion degree was 1 − (0.2 + 0.3) / 2 = 0.75.

[0059] When the dynamic coordination completion rate is lower than the preset coordination threshold of 0.8, the trend of the dynamic lag deviation is analyzed. Specifically, the dynamic lag deviations are arranged in chronological order to form a lag deviation sequence. The least squares method is used to perform linear fitting on the lag deviation data to obtain an approximate trend line. The slope of this fitted line can be used as an estimate of the error growth rate, thereby predicting the potential error in the next cycle. For example, if the system records the actuator's offset in the X-axis direction as 0.02, 0.03, 0.04, 0.05, and 0.06 mm in the last five cycles, the least squares fitting yields a slope of 0.01 mm per cycle, thus predicting an offset of 0.07 mm in the next cycle. This value is output as the nonlinear error trend. The least squares method operation steps are: first, list the x and y coordinates of all observed data points; then, construct the error sum of squares function; take the partial derivatives of this function with respect to the slope and intercept, and set the derivatives to zero; solving these two equations yields the slope and intercept of the fitted line.

[0060] The coordination threshold setting is based on statistical analysis of the system's historical execution data. Specifically, the system statistically analyzes a large number of historical operation records for cycles in which compensation was successfully completed, and calculates the coordination completion rate index within these cycles. The mean of the coordination completion rate is 0.92, and the standard deviation is 0.06. According to the normal distribution characteristics, approximately 95% of the samples are distributed above the mean minus twice the standard deviation, i.e., 0.92 - 2 * 0.06 = 0.8.

[0061] After obtaining the nonlinear error trend, the system converts this trend into a reverse compensation control quantity to counteract the impending dynamic error. This reverse compensation control quantity is expressed in the form of displacement or time advance and is superimposed on the current compensation command to form a real-time correction output. The real-time correction output, as a feedforward compensation quantity, is superimposed on the final compensation command sequence of the next control cycle (or the next segment). Specifically, when generating the subsequent final compensation command sequence, the system adds the reverse compensation control quantity predicted by the nonlinear error trend to the angle or speed command of the corresponding joint in advance, thereby achieving dynamic cancellation of the expected future error and completing the complete control closed loop from perception, decision-making, execution to feedforward compensation. For example, if the nonlinear error trend result shows that the error of the end effector in the X-axis direction increases at a rate of 0.01 mm per cycle, and predicts that the next cycle will produce an additional offset of about 0.07 mm, the system generates a reverse compensation control quantity, specifically adding a displacement of equal amplitude but in the opposite direction along the X-axis direction, i.e., -0.07 mm. Subsequently, this reverse displacement control quantity is directly superimposed onto the X coordinates of each node in the current compensation command sequence to form a real-time correction output, which is used to correct the system response in advance in the next cycle.

[0062] Through the above process, the system continuously corrects dynamic deviations during execution, enabling assembly actions to remain stable, coordinated, and precise in a real-world environment.

[0063] It is worth noting that, in order to balance processing speed and information density, the system performs downsampling processing on the high-frequency sampled force signal; the axis-by-axis differential processing is performed according to a preset time interval, which is preferably matched with the image frame period or the control command period, for example, set to 20ms; in step S17, the acquisition rate of the execution feedback data (position and torque) is synchronized with or higher than the control command period, for example, both are 10ms; in the alignment operation, for each command timestamp, the nearest sensor sample value before and after the timestamp is taken for interpolation calculation to eliminate the error caused by the slight asynchrony of different hardware sampling times.

[0064] Reference Figure 2 The second embodiment of the present invention provides an automated control system for a portable cylinder assembly platform, comprising: The data acquisition module is used to acquire raw images and multidimensional force signals; The preliminary offset module is used to calculate the contour pixel coordinate set and the resistance amplitude vector sequence based on the original image and the multidimensional force signal, and generate a three-dimensional fusion tensor according to the contour pixel coordinate set and the resistance amplitude vector sequence. The three-dimensional fusion tensor is then subjected to differential quantization processing to obtain preliminary offset data. The association module is used to set intervals and perform probability mapping based on the preliminary offset data to obtain a corresponding relationship matrix, and to perform state estimation and feature construction based on the corresponding relationship matrix to obtain an association vector; The fusion deviation vector module is used to identify and calculate the weights of abnormal components based on the correlation vector to obtain complementary weight factors. Based on the complementary weight factors, the correlation vector is weighted and fused to obtain a fusion deviation vector. The initial compensation command module is used to perform inverse kinematic mapping based on the fusion deviation vector to obtain the angular velocity vector of each joint, and to perform path fitting and direction analysis on the angular velocity vector to obtain the initial compensation command. The final compensation instruction module is used to execute the initial compensation instruction, collect multi-dimensional real-time contact force data, perform force signal differential and state estimation correction processing on the multi-dimensional real-time contact force data to obtain the optimized motion trajectory, and perform matching degree evaluation and instruction generation processing on the optimized motion trajectory to obtain the final compensation instruction sequence. The real-time correction output module is used to execute the final compensation instruction sequence, collect execution feedback data, perform deviation calculation on the feedback data to obtain the dynamic lag deviation, perform trend prediction analysis on the dynamic lag deviation to obtain the nonlinear error trend, and generate real-time correction output based on the nonlinear error trend.

[0065] It should be noted that the automated control system for a portable cylinder assembly platform provided in this embodiment of the invention is used to execute all the process steps of the automated control method for a portable cylinder assembly platform in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0066] This invention also provides an electronic device. The electronic device includes a processor, a memory, and a computer program stored in the memory and executable on the processor, such as an automation control program for a portable cylinder assembly platform. When the processor executes the computer program, it implements the steps described in the various embodiments of the automation control method for a portable cylinder assembly platform, for example... Figure 1 The step S11 shown. Alternatively, when the processor executes the computer program, it implements the functions of each module / unit in the above system embodiments, such as the fusion deviation vector module.

[0067] For example, the computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the electronic device.

[0068] The electronic device may be a desktop computer, laptop, handheld computer, or smart tablet, etc. The electronic device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the above components are merely examples of electronic devices and do not constitute a limitation on the electronic device. It may include more or fewer components than described above, or combine certain components, or different components. For example, the electronic device may also include input / output devices, network access devices, buses, etc.

[0069] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the electronic device, connecting all parts of the electronic device via various interfaces and lines.

[0070] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory and by calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0071] Wherein, if the modules / units integrated in the electronic device are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0072] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0073] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. An automated control method for a portable cylinder assembly platform, characterized in that, include: Acquire raw images and multidimensional force signals; Based on the original image and multidimensional force signal, the contour pixel coordinate set and the resistance amplitude vector sequence are calculated, and a three-dimensional fusion tensor is generated according to the contour pixel coordinate set and the resistance amplitude vector sequence. The three-dimensional fusion tensor is then subjected to differential quantization processing to obtain preliminary offset data. Based on the preliminary offset data, interval setting and probability mapping are performed to obtain the corresponding relationship matrix. Based on the corresponding relationship matrix, state estimation and feature construction are performed to obtain the association vector. Based on the correlation vector, abnormal component identification and weight calculation are performed to obtain complementary weight factors. Based on the complementary weight factors, the correlation vector is weighted and fused to obtain a fusion deviation vector. Based on the fusion deviation vector, inverse kinematic mapping is performed to obtain the angular velocity vector of each joint, and path fitting and direction analysis are performed on the angular velocity vector to obtain the initial compensation command; Execute the initial compensation command, collect multi-dimensional real-time contact force data, perform force signal differential and state estimation correction processing on the multi-dimensional real-time contact force data to obtain the optimized motion trajectory, and perform matching degree evaluation and command generation processing on the optimized motion trajectory to obtain the final compensation command sequence. The final compensation instruction sequence is executed, execution feedback data is collected, deviation is calculated on the feedback data to obtain the dynamic lag deviation, trend prediction analysis is performed on the dynamic lag deviation to obtain the nonlinear error trend, and real-time correction output is generated based on the nonlinear error trend.

2. The automated control method for a portable cylinder assembly platform according to claim 1, characterized in that, Based on the original image and multidimensional force signal, a set of contour pixel coordinates and a sequence of drag amplitude vectors are calculated. A three-dimensional fusion tensor is generated based on the set of contour pixel coordinates and the sequence of drag amplitude vectors. Differential quantization is then performed on the three-dimensional fusion tensor to obtain preliminary offset data, including: The Sobel operator is used to calculate the gradient of the original image to obtain image gradient maps in two directions, and a gradient magnitude map is generated based on the image gradient maps. The gradient magnitude map is filtered by a preset gradient threshold to obtain a set of contour pixel coordinates; The mean and standard deviation of the torques of each group of the multidimensional force signals are calculated to form a resistance amplitude vector sequence; The contour pixel coordinate set is projected onto spatial points in a three-dimensional coordinate system to construct a spatial matrix, and the drag amplitude vector is embedded into the spatial matrix to generate a three-dimensional fusion tensor. The element-wise difference operation is performed between the three-dimensional fusion tensor and the preset standard pose tensor to obtain the difference tensor. The difference tensor is then flattened into a one-dimensional vector and normalized and its magnitude is calculated to output preliminary offset data.

3. The automated control method for a portable cylinder assembly platform according to claim 1, characterized in that, The process of setting intervals and performing probability mapping based on the initial offset data to obtain a corresponding relationship matrix, and then performing state estimation and feature construction based on the corresponding relationship matrix to obtain an association vector, includes: Based on the preliminary offset data, multiple interval boundaries are set, and a two-dimensional partitioned network is generated based on the interval boundaries. Based on the two-dimensional partitioning network and the preliminary offset data, joint probability statistics are performed to generate a corresponding relationship matrix; Kalman filtering estimation optimization is performed on the correspondence matrix to obtain state estimates, and vector generation is performed on the state estimates to obtain a two-dimensional state estimate vector. The two-dimensional state estimation vector is normalized and its features are constructed to obtain the correlation vector.

4. The automated control method for a portable cylinder assembly platform according to claim 1, characterized in that, The step of identifying and calculating the weights of abnormal components based on the correlation vector to obtain a complementary weight factor, and then performing weighted correction and vector fusion on the correlation vector based on the complementary weight factor to obtain a fusion deviation vector, includes: The Euclidean distance between the associated vector and the preset reference vector is calculated. If the Euclidean distance exceeds the preset deviation threshold, the dimension with the largest difference from the reference vector in the associated vector is identified as the index position of the abnormal component. Based on the index position, the corresponding associated components are extracted, and the complementary weight factor is calculated based on the associated components. Based on the complementary weight factors, a diagonal weighted matrix is ​​constructed, and weight correction and splicing are performed according to the diagonal weighted matrix and the correlation vector to obtain the fusion deviation vector.

5. The automated control method for a portable cylinder assembly platform according to claim 1, characterized in that, The process involves performing inverse kinematic mapping based on the fused deviation vector to obtain the angular velocity vectors of each joint, and then performing path fitting and direction analysis on the angular velocity vectors to obtain the initial compensation command, including: Based on the fusion deviation vector, coordinate transformation calculation is performed to obtain three-dimensional gap change data in the local coordinate system; Based on the three-dimensional gap change data, the Jacobian matrix of the robotic arm in its current posture is constructed, and the pseudo-inverse calculation of the Jacobian matrix is ​​performed to obtain the angular velocity vector of each joint. Based on the angular velocity vector, an integration operation is performed at preset time intervals to generate a joint angle sequence, and a motion compensation path is formed by fitting the joint angle sequence. The magnitude and azimuth of the tangent vector of the initial segment of the motion compensation path are calculated, and an initial compensation command for driving the actuator is generated based on the magnitude and azimuth.

6. The automated control method for a portable cylinder assembly platform according to claim 5, characterized in that, The process involves executing the initial compensation command, collecting multi-dimensional real-time contact force data, performing force signal differential and state estimation correction processing on the multi-dimensional real-time contact force data to obtain an optimized motion trajectory, and then performing matching degree evaluation and command generation processing on the optimized motion trajectory to obtain the final compensation command sequence, including: Execute the initial compensation command, collect multi-dimensional real-time contact force data, and perform axis-by-axis differential processing on the multi-dimensional real-time contact force data according to a preset time interval to form a force signal deviation sequence; The force signal deviation sequence is estimated by Kalman filtering to obtain the predicted state vector; The predicted state vector is sequentially mapped to a displacement correction vector in three-dimensional space and superimposed on the coordinates of each node in the motion compensation path to obtain the optimized motion trajectory. The optimized motion trajectory is sampled at equal intervals to obtain a discrete node set. The Euclidean distance is calculated between the discrete node set and the target location to obtain the Euclidean distance, and then the Euclidean distance is normalized to obtain the matching degree. If the matching degree is lower than the preset convergence threshold, the optimized motion trajectory is updated and the driving signal is converted to obtain the final compensation instruction sequence.

7. The automated control method for a portable cylinder assembly platform according to claim 1, characterized in that, The process of executing the final compensation instruction sequence involves collecting execution feedback data, calculating the deviation from the feedback data to obtain a dynamic lag deviation, performing trend prediction analysis on the dynamic lag deviation to obtain a nonlinear error trend, and generating a real-time correction output based on the nonlinear error trend. This includes: The execution feedback data includes: real-time position data and torque data; Based on the real-time position data and torque data, the timing alignment deviation is calculated to obtain the dynamic hysteresis deviation. Based on the dynamic lag deviation, normalization and coordination degree calculation are performed to obtain the dynamic coordination completion degree; When the dynamic coordination completion rate is less than the preset coordination threshold, the least squares method is used to predict the trend of the dynamic lag deviation to obtain the nonlinear error trend. Based on the aforementioned nonlinear error trend, reverse compensation control quantity superposition processing is performed to obtain real-time corrected output.

8. An automated control system for a portable cylinder assembly platform, characterized in that, include: The data acquisition module is used to acquire raw images and multidimensional force signals; The preliminary offset module is used to calculate the contour pixel coordinate set and the resistance amplitude vector sequence based on the original image and the multidimensional force signal, and generate a three-dimensional fusion tensor according to the contour pixel coordinate set and the resistance amplitude vector sequence. The three-dimensional fusion tensor is then subjected to differential quantization processing to obtain preliminary offset data. The association module is used to set intervals and perform probability mapping based on the preliminary offset data to obtain a corresponding relationship matrix, and to perform state estimation and feature construction based on the corresponding relationship matrix to obtain an association vector; The fusion deviation vector module is used to identify and calculate the weights of abnormal components based on the correlation vector to obtain complementary weight factors. Based on the complementary weight factors, the correlation vector is weighted and fused to obtain a fusion deviation vector. The initial compensation command module is used to perform inverse kinematic mapping based on the fusion deviation vector to obtain the angular velocity vector of each joint, and to perform path fitting and direction analysis on the angular velocity vector to obtain the initial compensation command. The final compensation instruction module is used to execute the initial compensation instruction, collect multi-dimensional real-time contact force data, perform force signal differential and state estimation correction processing on the multi-dimensional real-time contact force data to obtain the optimized motion trajectory, and perform matching degree evaluation and instruction generation processing on the optimized motion trajectory to obtain the final compensation instruction sequence. The real-time correction output module is used to execute the final compensation instruction sequence, collect execution feedback data, perform deviation calculation on the feedback data to obtain the dynamic lag deviation, perform trend prediction analysis on the dynamic lag deviation to obtain the nonlinear error trend, and generate real-time correction output based on the nonlinear error trend.

Images