A dynamic correction system for a filling machine robotic arm with multi-mode vision servo compensation

Abstract

This invention relates to the field of automated control technology for filling equipment, and discloses a dynamic correction system for a filling machine robotic arm with multi-mode visual servo compensation. The system includes a detection module, a container defect classification model, an adaptive parameter learning module, a prediction algorithm module, an environmental interference correction module, an attitude adjustment module, a compensation effect evaluation and iterative optimization module, and a multi-machine collaborative scheduling module. Through multispectral dynamic imaging adaptation, laser focusing depth closed-loop control, confidence-weighted data fusion, reinforcement learning-based parameter optimization, extended Kalman filter-enhanced trajectory prediction, steam concentration graded correction, and Bayesian network-based error tracing, the system achieves high-precision detection and dynamic compensation for various types of containers.

Description

Technical Field

[0001] This invention relates to the field of automated control technology for filling equipment, and in particular to a dynamic correction system for a filling machine robotic arm with multi-mode vision servo compensation. Background Technology

[0002] With the rapid development of intelligent manufacturing and industrial automation technologies, filling equipment is increasingly widely used in industries such as food and beverage, pharmaceuticals, and chemicals. As the core execution unit of a filling production line, the positioning accuracy and dynamic response capability of the robotic arm directly affect filling quality and production efficiency. In actual filling processes, factors such as container position deviation on the conveyor belt, differences in container shape, and environmental interference can all cause positional deviations between the robotic arm's end effector and the container opening, thus affecting the accuracy and stability of the filling process.

[0003] Current dynamic correction technologies for robotic arms in filling machines largely rely on single industrial cameras or laser displacement sensors for container position detection. However, single sensors are difficult to adapt to containers of different materials and shapes, such as transparent, colored, and irregularly shaped containers, resulting in detection accuracy being significantly affected by differences in container characteristics. For example, edge detection of transparent containers is challenging under normal lighting conditions; the reflectivity of colored glass containers differs significantly from that of transparent containers; and the curved surface structure of irregularly shaped metal containers causes variations in laser reflection angles. These factors make it difficult for traditional single-sensor methods to achieve high-precision and highly adaptable position deviation capture.

[0004] Furthermore, steam interference during the filling process is a significant factor affecting the accuracy of optical detection. In high-temperature filling scenarios, steam forms an optical interference layer between the camera lens and the detection area, leading to image blurring and laser signal attenuation, thus distorting the detection data. Existing technologies mostly employ fixed-parameter optical calibration methods, which struggle to dynamically adapt to real-time changes in steam concentration, causing detection accuracy to fluctuate with environmental conditions.

[0005] In trajectory prediction, existing technologies mostly employ fixed models to predict container trajectories, making it difficult to dynamically adapt to changes in environmental disturbances and container motion characteristics. Fixed models cannot adjust prediction parameters based on real-time detection data, and prediction accuracy significantly decreases when the container's speed changes or is subjected to external disturbances. Furthermore, traditional trajectory prediction methods lack segmented processing mechanisms for different time scales, making it difficult to balance the real-time nature of short-term predictions with the accuracy of long-term predictions.

[0006] In multi-machine collaborative scenarios, the robotic arms of multiple filling machines need to operate simultaneously within a shared workspace. Existing technologies lack effective motion coordination mechanisms, which easily leads to conflicts in the robotic arm movement trajectories, increasing the risk of collisions and reducing filling efficiency. Traditional collaborative scheduling methods mostly employ static strategies of time or space partitioning, making it difficult to dynamically optimize based on the real-time status of each robotic arm, resulting in resource waste and efficiency bottlenecks.

[0007] Therefore, there is an urgent need for a dynamic correction system for the robotic arm of a filling machine that can adapt to multiple types of containers, resist environmental interference, achieve accurate trajectory prediction, and support multi-machine collaboration, so as to improve the accuracy, stability and efficiency of the filling process. Summary of the Invention

[0008] To address the shortcomings of existing technologies, this invention provides a dynamic correction system for a filling machine robotic arm with multi-mode vision servo compensation, aiming to solve the following technical problems: First, existing systems have poor adaptability to detecting various types of containers, resulting in insufficient accuracy in capturing position deviations; second, in complex environmental interference and multi-machine collaborative scenarios, the accuracy of container trajectory prediction is low and the robotic arm movement is prone to conflict, affecting the dynamic correction effect and overall filling efficiency.

[0009] The present invention provides a dynamic correction system for a filling machine robotic arm with multi-mode vision servo compensation, including a detection module, a container defect classification model, an adaptive parameter learning module, a prediction algorithm module, an environmental interference correction module, an attitude adjustment module, a compensation effect evaluation and iterative optimization module, and a multi-machine collaborative scheduling module.

[0010] The detection module is deployed in the filling head, integrating an industrial camera and a laser displacement sensor to capture container position deviations in real time. The industrial camera employs multispectral dynamic imaging adaptation technology, switching spectral bands in real time based on the container material, and integrates a dynamic exposure-reflection suppression unit to suppress reflection interference from steam and liquid surfaces. The laser displacement sensor adjusts the laser beam focusing depth in real time based on the container's aperture characteristics, forming a closed-loop control of the focusing depth through image and laser data fusion. The data fusion processing unit uses a spatiotemporal synchronous weighted fusion algorithm, dynamically allocating weights based on data reliability to form a confidence-weighted fusion result.

[0011] The container defect classification model receives container appearance image data transmitted by the detection module, which is used to identify and remove deformed containers, and record feature data such as defect type, defect location and defect quantification parameters.

[0012] The adaptive parameter learning module is used to optimize system parameters based on container features and feedback. It includes a container feature-parameter mapping library, a transfer learning parameter adaptation unit, and a real-time parameter optimization unit. The real-time parameter optimization unit builds a parameter optimization model based on reinforcement learning, uses accuracy and efficiency compensation as the reward function, and dynamically updates parameters using gradient descent.

[0013] The prediction algorithm module is used to predict the container's motion trajectory, including a multi-factor fusion prediction unit, an extended Kalman filter enhancement unit, and a short-term-long-term trajectory segmentation prediction unit. The extended Kalman filter enhancement unit distinguishes between steam and vibration interference types through a noise feature identification unit and updates the noise covariance matrix online based on the interference intensity. The short-term-long-term trajectory segmentation prediction unit adaptively divides the time window based on the container's motion stability; the short-term trajectory is predicted using the extended Kalman filter algorithm, while the long-term trajectory is supplemented by a long short-term memory neural network.

[0014] The environmental interference correction module is used to correct the interference of steam on optical detection during the filling process. It divides the steam concentration into multiple levels, establishes a mapping relationship between steam concentration and optical parameters, and generates an interference compensation coefficient.

[0015] The attitude adjustment module is used to achieve dynamic pose compensation of the filling head through a six-degree-of-freedom robotic arm. It receives the predicted trajectory data from the prediction algorithm module and generates the angle and displacement compensation amounts for each axis.

[0016] The compensation effect evaluation and iterative optimization module is used to evaluate the compensation accuracy and optimize the upstream module in reverse. It includes a 3D detection unit for compensation accuracy, a multi-module error tracing unit, and an optimization instruction generation unit. The multi-module error tracing unit locates the error source based on a Bayesian network.

[0017] The multi-arm collaborative scheduling module is used to coordinate the movement of multiple robotic arms to avoid conflicts. It includes a multi-arm motion space modeling unit, a conflict prediction and trajectory replanning unit, and a collaborative efficiency optimization unit.

[0018] This invention has the following beneficial effects: First, it significantly improves the detection and dynamic compensation accuracy for various types of containers. Through multispectral dynamic imaging adaptation of industrial cameras, closed-loop control of focusing depth of laser displacement sensors, and data fusion processing, the system can accurately capture the positional deviations of containers of different materials and diameters. Combined with the real-time optimization of the adaptive parameter learning module, it greatly enhances the adaptability to diverse containers. Second, it improves the accuracy of trajectory prediction and the stability of multi-machine collaboration in complex environments. The environmental interference correction module offsets the influence of steam on detection through steam concentration classification and optical parameter linkage calibration. The prediction algorithm module improves trajectory prediction accuracy through multi-factor fusion and extended Kalman filter enhancement technology. The multi-machine collaborative scheduling module avoids robot arm movement conflicts through spatial modeling and conflict resolution, thereby improving the overall stability and efficiency of the filling process. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall architecture of the system of the present invention; Figure 2 This is a detailed structural diagram of the detection module; Figure 3This is a schematic diagram of the prediction algorithm module; Figure 4 This is a schematic diagram of the multi-machine collaborative scheduling module; Figure 5 This is a schematic diagram of the compensation effect evaluation and iterative optimization module; Figure 6 This is a cumulative distribution chart of detection deviations under different steam grades; Figure 7 This is a schematic diagram illustrating the impact of steam grade on compensation accuracy and prediction error. Detailed Implementation

[0020] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0021] like Figure 1 As shown, the multi-mode vision servo compensation system provided by the present invention includes a detection module 1, a container defect classification model 2, an adaptive parameter learning module 3, a prediction algorithm module 4, an environmental interference correction module 5, an attitude adjustment module 6, a compensation effect evaluation and iterative optimization module 7, and a multi-machine collaborative scheduling module 8. Each module forms a closed-loop control through data interaction to realize the dynamic correction of the six-degree-of-freedom robotic arm of the filling machine.

[0022] In the technical solution of this invention, a deeply coupled collaborative working mechanism is formed among the modules. Detection module 1, as the front end of data acquisition, provides basic position information for all subsequent modules; container defect classification model 2 performs quality screening on the detection data to ensure the validity of the data entering the subsequent processing flow; adaptive parameter learning module 3 dynamically adjusts system parameters based on detection feedback, forming a parameter optimization closed loop with detection module 1; prediction algorithm module 4 performs trajectory prediction based on detection data and environmental compensation coefficients, providing a decision basis for attitude adjustment module 6; environmental interference correction module 5 monitors the environmental state in real time and outputs compensation coefficients, forming a linkage calibration with detection module 1 and prediction algorithm module 4; attitude adjustment module 6 executes the final compensation action and feeds back the execution result to compensation effect evaluation and iterative optimization module 7; compensation effect evaluation and iterative optimization module 7 evaluates the compensation accuracy and feeds back optimization instructions to upstream modules; multi-machine collaborative scheduling module 8 coordinates the movement of multiple robotic arms to ensure optimal global efficiency.

[0023] From a data flow perspective, the container position deviation data output by detection module 1 flows to container defect classification model 2, adaptive parameter learning module 3, and prediction algorithm module 4, respectively. The qualified signal and defect feature data output by container defect classification model 2 are transmitted to attitude adjustment module 6 and adaptive parameter learning module 3, respectively. The optimized parameters output by adaptive parameter learning module 3 are transmitted to detection module 1 and prediction algorithm module 4, respectively. The optical parameters and interference compensation coefficients output by environmental interference correction module 5 are transmitted to detection module 1 and prediction algorithm module 4, respectively. The predicted trajectory data output by prediction algorithm module 4 is transmitted to attitude adjustment module 6. The compensated trajectory data and real-time position data output by attitude adjustment module 6 are transmitted to compensation effect evaluation and iterative optimization module 7 and multi-machine collaborative scheduling module 8, respectively. The error source tracing results and efficiency data output by compensation effect evaluation and iterative optimization module 7 are transmitted to adaptive parameter learning module 3 and multi-machine collaborative scheduling module 8, respectively. The collaborative instructions output by multi-machine collaborative scheduling module 8 are transmitted to each attitude adjustment module 6.

[0024] like Figure 2 As shown, detection module 1 is deployed on the filling head, integrating an industrial camera unit, a laser displacement sensor unit, and a data fusion processing unit to capture container position deviations in real time. Detection module 1 is the sensing core of the entire system, and its detection accuracy directly determines the working effect of subsequent modules.

[0025] The industrial camera unit employs multispectral dynamic imaging adaptation technology, switching spectral bands based on the spectral reflectance characteristics of the container material. In a preferred embodiment of the invention, the 550nm band is selected for transparent PET containers, where the reflectance of transparent materials is low and the imaging clarity is high; the 650nm band is selected for colored glass containers, as this band can penetrate the pigment layer of colored glass, reducing color interference with imaging; and the 780nm near-infrared band is selected for irregularly shaped metal containers, where the reflectance of metal materials is stable, allowing for accurate extraction of irregular contours. The band switching response time is less than 10ms, ensuring real-time requirements. The industrial camera unit also integrates a dynamic exposure-reflection suppression unit. Through an image analysis module, the acquired image is segmented to identify the area of ​​reflective regions. When the proportion of reflective regions exceeds a preset reflective threshold, the system automatically increases the exposure time and improves the contrast. This adjustment process continues until the proportion of reflective regions falls below a preset target value. Preferably, the preset reflectivity threshold is 10%, the preset target value is 5%, the exposure time base value is 10ms, the contrast base value is 50, the single adjustment increment of the exposure time is 5ms, and the single adjustment increment of the contrast is 10%.

[0026] The laser displacement sensor unit adjusts the laser beam focusing depth in real time based on the container's aperture characteristics to ensure the laser signal accurately reaches the container's edge and is stably reflected. In a preferred embodiment of the invention, the laser displacement sensor unit extracts the container's aperture edge contour using the Canny edge detection algorithm and calculates the contour diameter as the container's aperture data. Canny edge detection employs a dual-threshold strategy, with a low threshold of 50 and a high threshold of 150, ensuring edge continuity while suppressing noise interference. The system automatically sets the focusing depth according to a preset mapping relationship between the container's aperture and the focusing depth. Preferably, the focusing depth is set to 20mm when the aperture is 50mm, and increases by 2mm for every 10mm increase in aperture. Simultaneously, the laser displacement sensor unit forms a closed-loop control of the focusing depth, comparing the displacement deviation data output by the laser displacement sensor with the target deviation. If the deviation exceeds a preset deviation threshold, the focusing depth is adjusted, repeating this process until the deviation is less than the preset deviation threshold. Preferably, the target deviation is 0.1mm, the preset deviation threshold is 0.5mm, the focusing depth increment for each adjustment is 0.1mm, and the closed-loop control response frequency is 100Hz.

[0027] The data fusion processing unit employs a spatiotemporal synchronization weighted fusion algorithm to achieve precise fusion of industrial camera data and laser displacement sensor data. First, the unit synchronizes the clocks of the industrial camera and laser displacement sensor using a precise time protocol, achieving a synchronization accuracy of 1μs to avoid data misalignment caused by timestamp discrepancies. Then, the unit calculates the confidence scores of the two types of data. The confidence score for the industrial camera data is calculated based on the image entropy value, which reflects the richness and clarity of the image information. The confidence score calculation formula is: , in: This represents the confidence level of industrial camera data, with a value ranging from 0 to 1. The current image entropy value is calculated using the Shannon entropy method. The smaller the entropy value, the clearer the image and the more concentrated the information. The maximum entropy value in historical statistics is used as the normalization benchmark. Preferably, The statistical analysis based on 1000 sets of clear image samples determined the value to be 8.

[0028] The confidence level of laser displacement sensor data is calculated based on the variance of the displacement signal. Variance reflects the stability and reliability of the signal. The confidence level calculation formula is as follows: , in: The confidence level of the laser displacement sensor data, with a value ranging from 0 to 1; The variance of the current displacement signal is calculated using the sliding window method, with a window size of 20 sampling points. The smaller the variance, the more stable the signal. The maximum variance in historical statistics is used as the normalization benchmark. Preferably, The area was determined to be 0.01 mm² based on 1000 sets of stable signal samples.

[0029] Finally, the positional deviation data is calculated using a weighted fusion formula, which is: , in: The data represents the merged positional deviation, in mm. The position data output by the industrial camera is obtained through the transformation from image coordinates to world coordinates; The position data output by the laser displacement sensor; These are the weighting coefficients for industrial camera data; These are the weighting coefficients for the laser displacement sensor data. The formula for calculating the weighting coefficients is: , in: For the first Weights of class data; For the first Confidence level of class data; The total number of data types, in this embodiment . Specifically, , By quantifying data reliability through confidence levels, higher-confidence data is assigned higher weights to ensure the accuracy of the fusion results. For example, when... , hour, , The fused data focuses more on the stable signal from the laser displacement sensor.

[0030] The container defect classification model 2 receives container appearance image data transmitted by the detection module 1, enabling automatic identification and rejection of defective containers to prevent them from entering the subsequent filling process and affecting product quality. The container defect classification model 2 includes a defect data input unit, a defect identification algorithm unit, a defect judgment and rejection control unit, and a feature data recording unit.

[0031] The defect data input unit is responsible for receiving the container appearance image data output by the detection module 1 and performing preprocessing operations, including image size normalization, grayscale transformation and histogram equalization, to ensure the consistency of image quality input to the defect recognition algorithm unit.

[0032] The defect identification algorithm unit employs a convolutional neural network architecture. The input layer receives a container appearance image at a preset resolution. Multiple convolutional layers progressively extract image features, including edge features, texture features, and shape features. The pooling layer uses max pooling to reduce feature dimensionality while retaining key information. The fully connected layer maps the extracted high-dimensional features to the probability output of whether the container is qualified or defective. Preferably, the input image resolution is 256×256 pixels. The network uses three convolutional layers with kernel sizes of 3×3, 5×5, and 3×3, respectively. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function. The pooling kernel size is 2×2. The network training uses a cross-entropy loss function and the Adam optimizer, with an initial learning rate of 0.001 and 100 training epochs.

[0033] The defect judgment and rejection control unit makes decisions based on the defect probability output by the defect identification algorithm unit. When the defect probability exceeds a preset defect threshold, the defect judgment and rejection control unit outputs a high-level rejection signal, controlling the pneumatic pusher next to the conveyor belt to push the defective container into the rejection channel. The response time of the pneumatic pusher is less than 50ms. When the defect probability is lower than a preset pass threshold, the defect judgment and rejection control unit outputs a high-level pass signal to the attitude adjustment module 6, indicating that the container can enter the filling process. When the defect probability is between the preset pass threshold and the preset defect threshold, the system marks the container as awaiting re-inspection for secondary inspection and confirmation. Preferably, the preset defect threshold is 90%, and the preset pass threshold is 10%.

[0034] The feature data recording unit records the feature information of all detected defective containers, including defect type, defect location, and defect quantification parameters. Defect types are divided into three main categories: deformation, breakage, and stains. Defect locations are divided into three areas: the bottle body, the bottle mouth, and the bottle bottom. Defect quantification parameters include specific values ​​such as deformation amount, breakage area, and stain area. The feature data recording unit transmits the above feature data to the adaptive parameter learning module 3 to optimize detection parameters and improve defect recognition accuracy.

[0035] The adaptive parameter learning module 3 is the core module for system parameter optimization. It receives detection accuracy feedback data from the detection module 1, defect feature data from the container defect classification model 2, and error source tracing results from the compensation effect evaluation and iterative optimization module 7. It then uses machine learning methods to adaptively adjust the system parameters. The adaptive parameter learning module 3 includes a container feature-parameter mapping library, a transfer learning parameter adaptation unit, and a real-time parameter optimization unit.

[0036] The container feature-parameter mapping library stores the correspondence between feature data of different types of containers and optimal detection parameters, enabling rapid parameter configuration for new containers. The library extracts 3D features of containers using point cloud processing technology. Based on point cloud data of the container surface acquired by a laser displacement sensor, principal component analysis is used to generate a 3D feature model of the container. The extracted features include three dimensions: size, material, and shape. Size features include container height and diameter with an accuracy of 0.1 mm; material features distinguish between PET, glass, and metal based on spectral reflectance; shape features quantify the cylindrical or irregular shape of the container outline using Fourier descriptors. The container feature-parameter mapping library is associated with the corresponding optimal detection parameters to form a feature-parameter mapping dataset. When the similarity between a new container feature and a certain type of container feature in the mapping library exceeds a preset similarity threshold, the system directly calls the detection parameters for that type of container without relearning. Preferably, the preset similarity threshold is 90%.

[0037] The transfer learning parameter adaptation unit employs the K-nearest neighbor algorithm to achieve rapid parameter transfer and adaptation when no perfectly matching container type exists in the mapping library. The transfer learning parameter adaptation unit calculates the Euclidean distance between the new container features and all container features in the mapping library. The distance calculation formula is as follows: , in: The feature distance is the smaller the distance, the more similar the containers are. For the new container's first Each feature value is standardized. For the container in the mapping library One eigenvalue; For feature dimensions. Preferably, The five feature dimensions are height (mm), diameter (mm), reflectivity (dimensionless, range 0-1), shape factor (dimensionless, range 0-1), and wall thickness (mm). Feature similarity is quantified using Euclidean distance, and the closest feature is selected. One container is used as a reference sample. Preferably, The transfer learning parameter adaptation unit takes this... The average detection parameters of the reference containers are used as the initial parameters, and then the parameters are fine-tuned according to the characteristic differences between the new container and the reference containers. For example, if the diameter of the new container is 5mm larger than the average diameter of the reference containers, the laser focusing depth is adjusted from the initial value of 20mm to 20.5mm.

[0038] The real-time parameter optimization unit constructs a parameter optimization model based on reinforcement learning, enabling online adaptive adjustment of detection parameters. In the reinforcement learning framework, the agent represents the set of parameters to be optimized, including camera exposure time, laser focusing depth, and the noise covariance of the prediction algorithm; the environmental state represents the current detection accuracy and efficiency; and the action represents the parameter adjustment amount. The reward function comprehensively considers both compensation accuracy and compensation efficiency, and its calculation formula is: , in: The reward value ranges from... and Weighted sum and decision; To compensate for precision, the value ranges from 0 to 1, with a larger value indicating higher precision; To compensate for efficiency, the number of compensations completed per unit time is normalized and measured. For precision weighting coefficients; Let be the efficiency weighting coefficient, and satisfy The constraints. Preferably, in a high-precision priority scenario. , In scenarios where high efficiency is prioritized , .

[0039] Compensation accuracy The calculation formula is: , in: This refers to the actual positional deviation, i.e., the residual deviation between the filling head and the container opening after compensation. The maximum permissible deviation is used as the normalization benchmark. Preferably, .

[0040] Compensation efficiency The calculation formula is: , in: The number of compensation tests completed per unit of time; The maximum number of detection compensations per unit time is used as the normalization benchmark. Preferably, times per minute.

[0041] The real-time parameter optimization unit updates parameters using the gradient descent method. The parameter update formula is as follows: , in: The parameter vector to be optimized contains multiple parameter components such as exposure time and depth of focus. for The parameter value at time; for The updated parameter value at any given time; The learning rate controls the step size for parameter updates; for Time-based reward function For parameters The gradient is calculated using numerical differentiation. Preferably, the learning rate... The initial value is 0.1, and it decreases by 0.01 every 100 iterations until it reaches the lower limit of 0.01. The parameters are updated along the negative direction of the reward function gradient, and the learning rate controls the update step size to avoid parameter oscillations and ensure that the parameters converge to the optimal value. The optimized parameters are transmitted to detection module 1 to update detection parameters, and to prediction algorithm module 4 to update prediction parameters.

[0042] like Figure 3 As shown, prediction algorithm module 4 is the core module for system trajectory prediction. It receives the fused position deviation data from detection module 1, the optimized parameters from adaptive parameter learning module 3, and the interference compensation coefficients from environmental interference correction module 5, and outputs the prediction result of the container's future trajectory. Prediction algorithm module 4 includes a multi-factor fusion prediction unit, an extended Kalman filter enhancement unit, and a short-term-long-term trajectory segmentation prediction unit.

[0043] The multi-factor fusion prediction unit integrates multi-dimensional container dynamic features to improve the accuracy and robustness of trajectory prediction. The unit acquires container weight data through a weighing sensor, as weight affects the container's inertial characteristics on the conveyor belt; and acquires container velocity data through an encoder, as velocity directly determines the container's positional change trend. Preferably, the weighing sensor has a sampling frequency of 100Hz, a measuring range of 0-500g, and an accuracy of ±0.1g; the encoder has a sampling frequency of 500Hz and a resolution of 0.01mm. The multi-factor fusion prediction unit uses the Pearson correlation coefficient to calculate the correlation between each feature and the trajectory deviation, which is used to determine the weight of each feature in the prediction model. The correlation coefficient calculation formula is: , in: The Pearson correlation coefficient ranges from -1 to 1, with a larger absolute value indicating a stronger correlation. For the feature data of the first Each sample value; For the trajectory deviation data of the first Each sample value; The sample mean of the feature data; This represents the sample mean of the trajectory deviation data; The sample size is specified. Feature weights are assigned based on correlation: a weight of 0.4 when the absolute value of the correlation coefficient is greater than 0.8, a weight of 0.3 when the absolute value of the correlation coefficient is between 0.5 and 0.8, and a weight of 0.1 when the absolute value of the correlation coefficient is less than 0.5. The fused comprehensive feature data is then output as input to the subsequent prediction unit.

[0044] The Extended Kalman Filter (EKF) enhancement unit optimizes the algorithm for interference characteristics in the filling scenario, improving its adaptability by updating the noise covariance matrix online. The EKF enhancement unit includes a noise feature identification unit that distinguishes between steam interference and mechanical vibration interference using spectral analysis. Steam interference is characterized by predominantly low-frequency components, primarily distributed in the 0-10Hz range; mechanical vibration interference exhibits periodic components at specific frequencies, typically related to equipment rotation speed. Based on the identified interference type and intensity, the EKF enhancement unit updates the noise covariance matrix online. The update formula for the process noise covariance is: , The update formula for the measured noise covariance is: , in: for The process noise covariance matrix at time step reflects the uncertainty in the system's state transition process; for The process noise covariance matrix at time step; for The measurement noise covariance matrix at any given time reflects the uncertainty in the sensor measurement process; for The measurement noise covariance matrix at time point; The process noise interference intensity coefficient is calculated by the noise feature identification unit based on the interference type and interference intensity. The noise interference intensity coefficient is also output by the noise feature recognition unit to measure it; It is an identity matrix with the same dimensions as the state vector. Preferably, It is a 2×2 identity matrix, corresponding to the two state variables of position and velocity; steam concentration levels 1-5 correspond to... The values ​​are 0.001, 0.002, 0.003, 0.004, and 0.005.

[0045] Initial process noise covariance matrix and the initial measurement noise covariance matrix The settings are as follows: , , The initial values ​​mentioned above are determined based on statistical analysis of the noise characteristics of the system under normal operating conditions. The increments of the covariance matrix from the previous moment, which are positively correlated with the interference intensity, enable the extended Kalman filter algorithm to dynamically adapt to changes in environmental interference, thus improving the robustness of the prediction.

[0046] The short-term-long-term trajectory segmentation prediction unit adaptively divides the prediction time window and selects the corresponding prediction algorithm based on the stability characteristics of the container motion. Stability is determined based on velocity variance. When the velocity variance is less than a preset stability threshold, the state is considered stable, and the prediction time window is set to a longer duration to improve the prediction range. When the velocity variance is not less than the preset stability threshold, the state is considered unstable, and the prediction time window is set to a shorter duration to ensure real-time prediction. Preferably, the preset stability threshold is 0.01 mm² / s², the window for stable states is set to 1 s, and the window for unstable states is set to 0.3 s.

[0047] Short-term trajectory prediction employs an extended Kalman filter algorithm, suitable for prediction scenarios with a time window less than 0.5s, offering advantages such as high computational efficiency and good real-time performance. Long-term trajectory prediction is supplemented by a long short-term memory neural network, suitable for prediction scenarios with a time window of not less than 0.5s, capable of capturing the long-term trend of container motion. The long short-term memory neural network takes as input a sequence of position data from a preset number of past windows and outputs a predicted trajectory data for a preset future time period. Preferably, the input is position data from the past 5 windows, and the output is a predicted trajectory data for the next 0.5s. The fusion weights of the two prediction results are dynamically adjusted based on stability. In a stable state, the extended Kalman filter weight is 0.7, and the long short-term memory neural network weight is 0.3; in an unstable state, both weights are 0.5. After fusion, the predicted motion trajectory data of the container within a preset future time range is output to the attitude adjustment module 6. Preferably, the preset time range is 0.1-1s.

[0048] The environmental interference correction module 5 is specifically designed to address steam interference issues in the filling process. It receives the optical imaging parameters from the detection module 1 and the steam concentration signal from the filling environment, and outputs the corrected optical parameters and interference compensation coefficients. The environmental interference correction module 5 includes an optical parameter monitoring unit, a steam interference identification unit, and an optical parameter correction unit.

[0049] The optical parameter monitoring unit collects the exposure time, contrast, and gain parameters of the industrial camera in real time to monitor the operating status of the optical system. Preferably, the sampling frequency is 10Hz, and the collected parameter data is transmitted to the steam interference identification unit for analysis and processing.

[0050] The steam interference identification unit employs a dual detection mechanism to determine the steam concentration level. The first detection directly acquires ambient humidity data via a humidity sensor deployed at a predetermined distance below the filling head. Preferably, the deployment distance is 5cm, the humidity sensor's measurement range is 0-100%RH, its accuracy is ±2%RH, and its sampling frequency is 20Hz. The second detection analyzes the gray-level co-occurrence matrix contrast of the industrial camera image to calculate the image's haze characteristics. The haze calculation formula is: Haze = 1 - Image Contrast / Standard Contrast. When the image's gray-level co-occurrence matrix contrast is less than 50, it is determined to be high-concentration steam. The steam interference identification unit combines the results of the two detections to classify the steam concentration into 5 levels. Preferably, the steam concentration level classification standards are as follows: Level 1 corresponds to 30%-40%RH, Level 2 corresponds to 40%-50%RH, Level 3 corresponds to 50%-60%RH, Level 4 corresponds to 60%-80%RH, and Level 5 corresponds to 80%-90%RH. When the haze is greater than 0.3 and the humidity is greater than 60%RH, it is classified as Level 4; when the haze is greater than 0.5 and the humidity is greater than 80%RH, it is classified as Level 5.

[0051] The optical parameter correction unit stores a hierarchical mapping library of steam concentration levels and optimal optical parameters. Based on the concentration level output by the steam interference identification unit, it queries the corresponding optical parameters and outputs them to the detection module 1. Preferably, the optical parameters for each level are set as follows: Level 1 corresponds to an exposure time of 10ms, a contrast ratio of 50, and a gain of 1.0; Level 2 corresponds to an exposure time of 12ms, a contrast ratio of 55, and a gain of 1.0; Level 3 corresponds to an exposure time of 15ms, a contrast ratio of 65, and a gain of 1.1; Level 4 corresponds to an exposure time of 18ms, a contrast ratio of 75, and a gain of 1.1; and Level 5 corresponds to an exposure time of 20ms, a contrast ratio of 80, and a gain of 1.2.

[0052] The environmental interference correction module 5 simultaneously generates interference compensation coefficients based on the steam concentration level. This is used by prediction algorithm module 4 to update the noise covariance matrix. Preferably, levels 1-5 correspond to... The values ​​are 0.001, 0.002, 0.003, 0.004, and 0.005. The interference compensation coefficients are transmitted to the detection module 1 for dynamic adjustment of optical parameters and to the prediction algorithm module 4 for online updating of the noise covariance matrix.

[0053] The environmental interference correction module 5 also includes a real-time calibration execution unit, responsible for handling the time lag issue during parameter correction. Based on historical data analysis, there is a fixed time delay from the issuance of the calibration command to the actual effect of the parameters. The real-time calibration execution unit compensates for this delay by sending the calibration command in advance. Preferably, the time delay is 0.05s. After parameter correction, the real-time calibration execution unit verifies the calibration effect by calculating the contrast of the calibrated image. If the contrast is greater than a preset contrast threshold, the calibration is considered qualified; otherwise, the parameters are readjusted. Preferably, the preset contrast threshold is 70.

[0054] like Figure 6 As shown, at different steam levels, the cumulative distribution of detection deviations in the experimental group with environmental correction and the control group without environmental correction showed significant differences. As the steam level increased from level 1 to level 5, the mean detection deviation of the control group rapidly increased from 0.15 mm to 0.72 mm, while the mean detection deviation of the experimental group only slowly increased from 0.12 mm to 0.23 mm, verifying the effectiveness of the environmental interference correction module. Figure 7 As shown, with the increase of steam level, the average compensation accuracy and the average prediction error show a correlated trend. At steam level 1, the average compensation accuracy is 0.12 mm and the average prediction error is 0.11 mm. At steam level 5, the average compensation accuracy increases to 0.23 mm and the average prediction error increases to 0.35 mm. Through the dynamic adjustment of optical parameters and the transmission of interference compensation coefficient by the environmental interference correction module, the accuracy decay under high steam concentration is effectively controlled. The distribution of detection deviation and prediction error under different steam levels shows a clustering characteristic. The lower the steam level, the more concentrated the data points are near the origin of the coordinate system. The higher the steam level, the more the data points spread to the upper right. This distribution characteristic indicates that there is a positive correlation between detection deviation and prediction error, providing a data basis for the parameter optimization of the environmental interference correction module.

[0055] The attitude adjustment module 6 is the core module at the system execution level. It receives the predicted trajectory data from the prediction algorithm module 4, the pass / fail signal from the container defect classification model 2, and the collaborative instructions from the multi-machine collaborative scheduling module 8, and controls the six-degree-of-freedom robotic arm to perform dynamic pose compensation. The attitude adjustment module 6 includes a multi-module data integration unit, a six-axis robotic arm control unit, and a pose compensation execution unit.

[0056] The multi-module data integration unit is responsible for integrating heterogeneous data from different modules into control data in a unified format. Predicted trajectory data is input in time-series format, with each sampling point containing a three-dimensional coordinate. Preferably, the sampling interval is 0.1 s. A pass signal is a high-level active digital signal, indicating that the current container has passed defect detection and can enter the filling process. Coordination instructions include maximum speed limits and motion boundary constraints for each axis. The multi-module data integration unit first verifies the validity of the pass signal. If no pass signal is received, no compensation instruction is generated, and the robotic arm maintains its current position waiting for the next container. If a pass signal is received, the unit combines the predicted trajectory data and coordination instructions to generate control data in a unified format.

[0057] The six-axis robotic arm control unit employs a dual-loop PID control architecture, with an outer position loop and an inner velocity loop, enabling precise control of each axis of the robotic arm. Based on the control data output from the multi-module data integration unit, the six-axis robotic arm control unit calculates the target position and target velocity of each axis, and calculates the driving torque of each axis through a PID control algorithm. Preferably, the PID parameters of the position loop are set to... , , Speed ​​loop PID parameters are set to , , The position loop controls the spatial positioning accuracy of the robotic arm's end effector, while the speed loop controls the smoothness of the motion process. The control cycle of the six-axis robotic arm control unit is 1ms, ensuring real-time performance during high-speed movements.

[0058] The pose compensation execution unit converts the compensation amount calculated by the six-axis robotic arm control unit into drive signals for each joint, controlling the servo motors to perform compensation actions. Simultaneously, the pose compensation execution unit records trajectory data for each compensation, including the angle, displacement, and execution time of each axis. This recorded data is transmitted to the compensation effect evaluation and iterative optimization module 7 for accuracy assessment, and the real-time position data of the robotic arm's end effector is transmitted to the multi-machine collaborative scheduling module 8 for collaborative scheduling.

[0059] like Figure 5 As shown, the compensation effect evaluation and iterative optimization module 7 is the core module of the system closed-loop optimization. It receives the compensation trajectory data from the attitude adjustment module 6, the position deviation data from the detection module 1, and the predicted trajectory data from the prediction algorithm module 4, evaluates the compensation accuracy, and generates optimization instructions. The compensation effect evaluation and iterative optimization module 7 includes a three-dimensional detection unit for compensation accuracy, a multi-module error tracing unit, and an optimization instruction generation unit.

[0060] The compensation accuracy 3D detection unit uses a high-speed camera to record the actual trajectory of the filling point and compares it with the target trajectory to calculate the 3D deviation. Preferably, the high-speed camera has a frame rate of 1000fps and a resolution of 1920×1080, extracting one 3D coordinate point every 1ms. The 3D deviation is calculated using Euclidean distance, with the following formula: , in: This refers to the three-dimensional deviation, in mm. The actual trajectory coordinates are obtained through high-speed camera image processing. The target trajectory coordinates are derived from the prediction results of prediction algorithm module 4. The deviation between the actual trajectory and the target trajectory is quantified using Euclidean distance in three-dimensional space, intuitively reflecting the compensation accuracy. If the deviation exceeds a preset deviation warning value, the error tracing process is triggered. Preferably, the preset deviation warning value is 0.2 mm.

[0061] The multi-module error tracing unit uses a Bayesian network to locate error sources, enabling accurate identification of error sources in a multi-module system. The Bayesian network nodes include three main modules: a detection module, a prediction algorithm module, and an attitude adjustment module. The edge weights represent the probability of error propagation between modules. The prior probability is set based on historical error statistics, reflecting the basic probability of each module generating errors. The conditional probability is set based on the feature analysis of error types, reflecting the correlation between specific errors and each module. The posterior probability is calculated using the following formula: , in: In order to observe error Under the conditions, module Let be the posterior probability of the error source; For module Let be the prior probability of the error source; For the module When an error source is generated, an error is produced. The conditional probability; the denominator is a normalization factor to ensure that the sum of the posterior probabilities of all modules is 1. Preferably, the prior probability of the detection module is... The prior probability of the prediction algorithm module is 0.4. The prior probability of the attitude adjustment module is 0.25. It is 0.35. For positional deviation error... Conditional probability of the detection module The conditional probability of the prediction algorithm module is 0.8. The conditional probability of the attitude adjustment module is 0.1. The value is 0.1; for trajectory deviation error Conditional probability of the detection module The conditional probability of the prediction algorithm module is 0.2. The conditional probability of the attitude adjustment module is 0.7. The value is 0.1. The error source is derived in reverse using Bayes' theorem, and the module with the highest posterior probability is selected as the main source of error.

[0062] The optimization instruction generation unit generates corresponding optimization instructions for the error sources located by the multi-module error tracing unit. If the error source is the detection module, it outputs instructions to adjust the camera exposure time or laser focusing depth; if the error source is the prediction algorithm module, it outputs instructions to adjust the noise covariance matrix or prediction time window; if the error source is the attitude adjustment module, it outputs instructions to adjust PID parameters or speed limits. The optimization instructions are transmitted to the adaptive parameter learning module 3, which performs unified parameter updates and distribution.

[0063] like Figure 4 As shown, the multi-machine collaborative scheduling module 8 is responsible for coordinating the movement of multiple filling machine robotic arms, avoiding motion trajectory conflicts, and improving overall production efficiency. The multi-machine collaborative scheduling module 8 receives real-time position data of the robotic arms from each posture adjustment module 6 and efficiency data from the compensation effect evaluation and iterative optimization module 7, including a multi-arm motion space modeling unit, a conflict prediction and trajectory replanning unit, and a collaborative efficiency optimization unit.

[0064] The multi-arm motion space modeling unit establishes a global coordinate system covering the working range of all robotic arms, achieving a unified representation of the positions of multiple robotic arms. The origin of the global coordinate system is set as the starting point of the production line, the X-axis is along the conveying direction of the production line, the Y-axis is perpendicular to the horizontal direction of the production line, and the Z-axis is vertically upward. The multi-arm motion space modeling unit divides a three-dimensional spatial mesh with a preset precision. Each mesh is marked with three-dimensional coordinates, and the position of each robotic arm's end effector in the mesh is updated in real time, marking its occupied or idle status. Preferably, the X-axis range is 0-10000mm, the Y-axis range is 0-2000mm, the Z-axis range is 0-1000mm, and the mesh precision is 1mm. For scenarios with multiple filling machines, the X-axis range can be extended to 0-20000mm, with each filling machine covering a working area of ​​approximately 5000mm.

[0065] The conflict prediction and trajectory replanning unit predicts potential collisions within a future timeframe based on the current positions and trajectories of each robotic arm. The time range for conflict prediction is a preset time length. Preferably, the preset time length is 1 second. The conflict prediction and trajectory replanning unit calculates the position sequence of each robotic arm within the preset timeframe. If the positions of two robotic arms fall into the same grid and the time difference is less than a preset time difference threshold, it is determined to be a potential conflict. Preferably, the preset time difference threshold is 0.1 seconds.

[0066] When a potential conflict is detected, the conflict prediction and trajectory replanning unit uses Algorithm A to replan the trajectory for the conflict-prone robotic arm. The cost function of Algorithm A is calculated as follows: , in: This represents the total cost, used to evaluate the merits of the path. The actual cost from the current position to the starting point is calculated using Euclidean distance. The heuristic cost from the current position to the destination is estimated using Manhattan distance calculation to accelerate the search process. The path with the minimum cost is selected as the replanned trajectory, ensuring that the robotic arm reaches the target position as quickly as possible while avoiding conflicts. The replanned trajectory data is then transmitted to the corresponding attitude adjustment module 6 for execution.

[0067] The collaborative efficiency optimization unit dynamically allocates tasks based on the compensation efficiency and accuracy data of each robotic arm, maximizing the overall production line efficiency. The unit establishes an efficiency-accuracy comprehensive evaluation model to calculate the comprehensive score of each robotic arm. The score calculation formula is as follows: , in: The score is a comprehensive score, ranging from 0 to 1. The efficiency of the robotic arm is measured in the number of compensation operations completed per unit time. The efficiency of the most efficient robotic arm among all robotic arms is used as the normalization benchmark. This represents the mean three-dimensional deviation of the robotic arm; The maximum permissible deviation is used as the normalization benchmark. Preferably, The comprehensive score considers both efficiency and accuracy, with efficiency weighted at 0.6 and accuracy weighted at 0.4. Task allocation is based on the comprehensive score, with higher-scoring robotic arms receiving more tasks. Coordination commands are transmitted to each posture adjustment module 6 to limit the robotic arm's motion parameters and task pace.

[0068] In the technical solution of this invention, multiple closed-loop interaction mechanisms are formed between the modules to achieve continuous optimization of system performance.

[0069] The detection module 1 and the adaptive parameter learning module 3 form a dynamic parameter closed-loop adjustment. The detection module 1 transmits detection accuracy feedback data to the adaptive parameter learning module 3. The adaptive parameter learning module 3 sets a detection error warning value through a detection accuracy-parameter correlation analysis unit, identifies the parameters with the greatest impact on accuracy, and limits the parameter adjustment range through a real-time parameter adjustment execution unit to prevent system oscillation. The optimized detection parameters are then transmitted back to the detection module 1 to verify whether the detection accuracy has improved after adjustment. This closed loop achieves adaptive optimization of the detection parameters.

[0070] The prediction algorithm module 4 and the compensation effect evaluation and iterative optimization module 7 form a closed-loop correction for trajectory prediction errors. The compensation effect evaluation and iterative optimization module 7 transmits the error source tracing results and prediction error characteristics to the adaptive parameter learning module 3. The adaptive parameter learning module 3 distinguishes between systematic errors and random errors through the error feature-model parameter mapping unit and generates correction rules. The online prediction model update unit fine-tunes the extended Kalman filter parameters based on the error gradient to verify model stability and avoid divergence. The optimized prediction parameters are transmitted to the prediction algorithm module 4, and this closed loop achieves continuous improvement in prediction accuracy.

[0071] The environmental interference correction module 5 and the detection module 1 form an optical-interference linkage calibration. The environmental interference correction module 5 transmits the calibrated optical parameters to the detection module 1. The detection module 1 applies the updated parameters to acquire images and feeds back the image clarity data to the environmental interference correction module 5. The environmental interference correction module 5 stores a graded parameter library through a steam concentration-optical parameter mapping unit and corrects the time lag of parameter adjustment through a real-time calibration execution unit, verifying the clarity of the calibrated detection data. This closed loop achieves environmentally adaptive optical calibration.

[0072] The attitude adjustment module 6 and the multi-arm collaborative scheduling module 8 work together to resolve multi-arm motion conflicts. The attitude adjustment module 6 transmits the real-time position data of the robotic arms to the multi-arm collaborative scheduling module 8. The multi-arm collaborative scheduling module 8 uses a multi-arm motion space modeling unit to divide a three-dimensional spatial mesh with preset precision, unifying the coordinates of each robotic arm to the global coordinate system. The conflict prediction and trajectory replanning unit calculates the time window for potential collisions and generates the optimal avoidance trajectory with the goal of minimizing energy consumption and time. Collaborative commands are transmitted to each attitude adjustment module 6, and this closed loop enables safe and efficient operation of multi-arm collaboration.

[0073] The container defect classification model 2 and the adaptive parameter learning module 3 form a defect feature-parameter linkage optimization. The container defect classification model 2 transmits defect feature data to the adaptive parameter learning module 3. The adaptive parameter learning module 3 collects historical misjudgment cases through the defect misjudgment analysis unit, extracts the features of misjudged samples, and locates the causes of misjudgment. The detection parameter targeted correction unit enhances the corresponding detection parameters based on the causes of misjudgment, and statistically analyzes the decrease in the misjudgment rate to verify the correction effect. The optimized detection parameters are transmitted to the detection module 1, indirectly affecting the input data quality of the container defect classification model 2. This closed loop achieves a continuous improvement in defect detection accuracy.

[0074] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A dynamic correction system for a filling machine robotic arm with multi-mode vision servo compensation, characterized in that, include: The detection module, deployed on the filling head, integrates an industrial camera and a laser displacement sensor to capture container position deviations in real time. The industrial camera uses multispectral dynamic imaging adaptation technology, selecting the 550nm band for transparent PET containers, the 650nm band for colored glass containers, and the 780nm near-infrared band for irregularly shaped metal containers. The laser displacement sensor adjusts the laser beam focusing depth in real time based on the container's aperture characteristics. The data fusion processing unit uses a spatiotemporal synchronous weighted fusion algorithm to calculate the confidence level based on the image entropy value of the industrial camera data and the displacement signal variance of the laser displacement sensor data, forming a confidence-weighted fusion result. The container defect classification model is used to receive container appearance image data transmitted by the detection module, identify and remove deformed containers, and record feature data such as defect type, defect location and defect quantification parameters. The adaptive parameter learning module is used to optimize system parameters based on container features and feedback; the prediction algorithm module is used to predict the container's trajectory; and the environmental interference correction module is used to correct the interference of steam on optical detection during the filling process, classifying steam concentration into multiple levels and establishing a steam concentration-optical parameter mapping relationship. The attitude adjustment module is used to realize dynamic pose compensation of the filling head through a six-degree-of-freedom robotic arm, and to receive the predicted trajectory data to generate compensation instructions; the compensation effect evaluation and iterative optimization module is used to evaluate the compensation accuracy and optimize the upstream module in reverse, and to locate the error source based on a Bayesian network containing three nodes: detection module, prediction algorithm module and attitude adjustment module through a multi-module error tracing unit. The multi-arm collaborative scheduling module is used to coordinate the movement of multiple robotic arms to avoid conflicts. It includes a multi-arm motion space modeling unit and a conflict prediction and trajectory replanning unit.

2. The system according to claim 1, characterized in that, The adaptive parameter learning module includes a container feature-parameter mapping library, a transfer learning parameter adaptation unit, and a real-time parameter optimization unit. The real-time parameter optimization unit constructs a parameter optimization model based on reinforcement learning, with compensation for accuracy and efficiency as the reward function. The prediction algorithm module includes a multi-factor fusion prediction unit, an extended Kalman filter enhancement unit, and a short-term-long-term trajectory segmentation prediction unit. The extended Kalman filter enhancement unit updates the noise covariance matrix online based on the interference intensity.

3. The system according to claim 2, characterized in that, The weight allocation of the data fusion processing unit is dynamically adjusted based on data reliability. In the weight calculation, the weight of the i-th type of data is positively correlated with the confidence level of the i-th type of data. The confidence level of the industrial camera data is calculated based on the image entropy value, and the confidence level of the laser displacement sensor data is calculated based on the displacement signal variance. In the reward function of the real-time parameter optimization unit, the sum of the accuracy weight and the efficiency weight is 1. The parameter update adopts the gradient descent method, the initial value of the learning rate is 0.1, and it decreases by 0.01 every 100 iterations until it reaches the lower limit value of 0.

01.

4. The system according to claim 2, characterized in that, The environmental interference correction module divides the steam concentration into 5 levels. Each level corresponds to a preset optimal exposure time, contrast, and gain parameter. It generates interference compensation coefficients and transmits them to the prediction algorithm module to update the noise covariance matrix.

5. The system according to claim 2, characterized in that, The extended Kalman filter enhancement unit distinguishes between steam and vibration interference types through the noise feature identification unit. The update amount of the process noise covariance and the measurement noise covariance is positively correlated with the interference intensity coefficient. The interference intensity coefficient is calculated by the noise feature identification unit based on the interference type and interference intensity. The interference intensity coefficients corresponding to steam concentration levels 1-5 are 0.001, 0.002, 0.003, 0.004, and 0.005, respectively.

6. The system according to claim 2, characterized in that, The short-term-long-term trajectory segmentation prediction unit adaptively divides the time window based on the container's motion stability. When the velocity variance is less than 0.01 mm² / s², it is determined to be in a stable motion state, and the prediction time window is 1 s. When the velocity variance is not less than 0.01 mm² / s², it is determined to be in an unstable motion state, and the prediction time window is 0.3 s. The short-term trajectory is predicted using the extended Kalman filter algorithm, and the long-term trajectory is supplemented by the long short-term memory neural network.

7. The system according to claim 1, characterized in that, The multi-module error tracing unit extracts the error features of the detection module, prediction algorithm module, and attitude adjustment module. The nodes of the Bayesian network are the detection module, prediction algorithm module, and attitude adjustment module. The weight of the edge represents the error propagation probability between the above modules. The probability that each module is an error source is calculated based on the posterior probability.

8. The system according to claim 1, characterized in that, The detection module also includes a dynamic exposure-reflection suppression unit, which suppresses reflective interference from steam and liquid surfaces by real-time correction of exposure parameters and contrast. The laser displacement sensor unit identifies the container diameter by fusing image and laser data to form a closed-loop control of focusing depth.

9. The system according to claim 1, characterized in that, The multi-arm motion space modeling unit divides the three-dimensional space grid with a preset precision and unifies the coordinates of each robotic arm to the global coordinate system. The conflict prediction and trajectory replanning unit calculates the time window of potential collisions and generates the optimal avoidance trajectory with the goal of minimizing energy consumption and time.

10. The system according to claim 2, characterized in that, The container defect classification model and the adaptive parameter learning module form a defect feature-parameter linkage optimization. The defect misjudgment analysis unit collects historical misjudgment cases and extracts misjudgment sample features. The detection parameter orientation correction unit enhances the detection parameters of the deformed edge.

Images