A robot cable production quality management system and method

Abstract

The present application relates to the technical field of cable intelligent manufacturing, in particular to a robot-based cable production quality management system and method, which comprises: collecting multi-source production data such as spectral feature sequences, process parameter time series flow, multi-view visual images and online sensor measurement values obtained by a robot, extracting deep quality features by using an improved variational autoencoder incorporating cable production physical constraints and process rules, constructing a process state evolution path and completing three-dimensional voxel reconstruction of appearance defects. The multi-dimensional features are input into a multi-modal fusion network to obtain a comprehensive quality tensor, the potential defect probability of the future link is output by an attention mechanism defect prediction network, and a robot real-time process adjustment instruction is generated accordingly. This method can improve the feature extraction adhesion, completely present the defect morphology, realize accurate defect prediction and dynamic production parameter control, and improve the accuracy and real-time performance of cable production quality control.

Description

Technical Field

[0001] This invention relates to the field of intelligent cable manufacturing technology, and in particular to a robot-based cable production quality management system and method. Background Technology

[0002] In traditional cable production quality control, production data is often collected using a single type of sensor, data feature extraction is performed using conventional variational autoencoders, process parameters are simply stored and recorded in a time sequence, appearance defect detection relies on two-dimensional visual image recognition, quality defect prediction uses a single data analysis model, and process parameter adjustments for production line robots depend on manual experience-based instructions. These conventional quality control methods fail to consider the actual physical rules and process standards of cable production, and the data processing models are not optimized specifically for the cable production scenario.

[0003] Conventional variational autoencoders lack structured constraints on their latent space, leading to discrepancies between feature extraction results from multi-source production data and actual cable production conditions. The dynamic changes in process parameters fail to form a complete evolution path, and appearance defects are only presented in two dimensions, hindering three-dimensional reconstruction. Furthermore, effective fusion of various production data types is impossible, defect prediction cannot focus on key features, and it's difficult to output the probability of potential defects in future production stages. This also prevents the generation of precise real-time process parameter adjustment instructions for production line robots. Therefore, an improved algorithm adapted to cable production constraints is needed to extract features from multi-source production data, simultaneously performing process state modeling and three-dimensional reconstruction of appearance defects. After fusing multi-dimensional data, a dedicated predictive network should be used to analyze defect probabilities, thereby generating robot-adaptive parameter adjustment instructions. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a robot-based cable production quality management system and method.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a robot-based cable production quality management method, comprising: Acquire a set of multi-source production data collected by robots during the production process of the target cable, including raw material spectral feature sequences, process parameter time-series flow, multi-view visual images, and online sensor measurements. An improved variational autoencoder algorithm is applied to the multi-source production data set to generate a deep quality feature representation. The improved variational autoencoder algorithm provides structured constraints on the potential space of the original variational autoencoder based on the physical constraints and process rules in cable production. The process state transition model is performed on the time-series flow of the process parameters to generate the process state evolution path, and the three-dimensional defect reconstruction is performed on the multi-view visual image to generate a voxel model of cable appearance defects. The deep quality feature representation, process state evolution path and cable appearance defect voxel model are input into a multimodal fusion network to generate the comprehensive quality tensor of the cable product. The comprehensive quality tensor is analyzed by calling an attention-based quality defect prediction network, and the probability distribution of multiple potential defects of the cable product in the future preset production stage is output. Based on the probability distribution of the various potential defects, an online control strategy generator calculates and outputs a set of real-time process parameter adjustment instructions for the production line robot.

[0006] As a further aspect of the present invention, an improved variational autoencoder algorithm is applied to the multi-source production data set to generate a deep quality feature representation, including: An encoder network is constructed to encode the raw material spectral feature sequence, process parameter time flow, multi-view visual image and online sensor measurement value respectively, so as to obtain their respective initial latent variables; The corresponding initial latent variables are input into the structured latent space of the improved variational autoencoder algorithm. The structured latent space imposes constraints on the distribution and relationship of the latent variables based on the physical constraints and process rules of cable production. The constraint on the relationship includes: forcing the dimensions of latent variables representing the same physical attribute to be statistically correlated, and limiting the probability of latent variable combinations that violate known process rules. In the structured latent space, latent variables that satisfy the constraints are sampled and input into the corresponding decoder network for data reconstruction; During training, by minimizing the difference between the reconstruction error and the prior distribution of latent variables that satisfy the physical constraints and process rules, the encoder learns the deep quality feature representation that contains physical meaning and conforms to the process rules.

[0007] As a further aspect of the present invention, process state transition modeling is performed on the time-series flow of the process parameters to generate a process state evolution path, including: Identify key process parameters in the process parameter time-series flow, including temperature, pressure, extrusion speed, and tension; On the time series of the key process parameters, multiple stable intervals of the process state are segmented by the state change point detection algorithm, and the statistical characteristics of the key process parameters are calculated for each stable interval as the state representation vector of the stable interval. Based on the time sequence, the state representation vectors of consecutive stable intervals of process states are connected to form a preliminary state evolution sequence. The initial state evolution sequence is modeled using a hidden Markov model to infer the hidden process state sequence. The hidden process state sequence and its corresponding state representation vector are used together as the process state evolution path.

[0008] As a further aspect of the present invention, three-dimensional defect reconstruction is performed on the multi-view visual image to generate a voxel model of cable appearance defects, including: For each multi-view visual image, cable target segmentation and feature point extraction are performed to obtain the cable contour and surface key points under different viewpoints; By using multi-view geometry principles, the coordinates of key points on the cable surface in three-dimensional space are calculated, and a three-dimensional point cloud model of the cable surface is reconstructed. Based on the three-dimensional point cloud model, surface bumps, scratches, or contamination areas that exist compared with the standard cable three-dimensional model are detected and identified as potential appearance defects. The three-dimensional point cloud data corresponding to the potential appearance defect area is voxelized and filled into a voxel mesh with space occupation, generating the cable appearance defect voxel model containing defect location, shape and size information.

[0009] As a further aspect of the present invention, the deep quality feature representation, the process state evolution path, and the cable appearance defect voxel model are input into a multimodal fusion network to generate a comprehensive quality tensor for the cable product, including: The multimodal fusion network includes a feature alignment subnetwork, a feature interaction subnetwork, and a feature aggregation subnetwork; The feature alignment sub-network maps the temporal dimension of the deep quality feature representation, the state dimension of the process state evolution path, and the spatial dimension of the cable appearance defect voxel model to a feature space, thereby achieving scale and semantic alignment of multimodal features. The feature interaction subnetwork establishes a bidirectional attention mechanism between the deep quality feature representation, the process state evolution path, and the cable appearance defect voxel model within the feature space, simulating how process parameters affect the formation of internal quality features and appearance defects. The feature aggregation subnetwork performs weighted fusion and compression on the multimodal features after sufficient interaction, and outputs a comprehensive mass tensor that fully characterizes the cable material, process and appearance.

[0010] As a further aspect of the present invention, the step of calling an attention-based quality defect prediction network to analyze the comprehensive quality tensor and outputting the probability distribution of multiple potential defects of the cable product in a future predetermined production stage includes: The quality defect prediction network comprises a multi-head self-attention layer and a feedforward prediction network; The comprehensive quality tensor is input into the multi-head self-attention layer, which enables information exchange and weight focusing between different feature regions in the comprehensive quality tensor, highlighting the feature combinations most relevant to various defects. The features processed by the multi-head self-attention layer are input into the feedforward prediction network. The feedforward prediction network maps the input features into a multi-dimensional vector. Each dimension of the multi-dimensional vector corresponds to a type of potential defect in a pre-defined future production process, and its value represents the probability of the defect occurring. The probability values ​​corresponding to all dimensions are normalized to form the probability distribution of multiple potential defects of the cable product in the future preset production stage.

[0011] As a further aspect of the present invention, based on the probability distribution of the multiple types of potential defects, a set of real-time process parameter adjustment instructions for the production line robot is calculated and output through an online control strategy generator, including: The online control strategy generator receives the probability distribution of the multiple potential defects and the real-time status feedback of the current production line; Within the online control strategy generator, an optimization function is established with the objective of minimizing the overall probability of defect occurrence. The decision variables of the optimization function are adjustable process parameters. Using gradient-based optimization algorithms or heuristic search algorithms, under the constraints of the process parameter adjustment range and robot execution capability, the optimization function is solved to obtain a set of optimized process parameter target values; The difference between the optimized process parameter target value and the current process parameter setting value is calculated, and the difference is converted into specific adjustment actions that can be executed by the production line robot, forming the real-time process parameter adjustment instruction set.

[0012] As a further aspect of the present invention, the step of using a gradient-based optimization algorithm or a heuristic search algorithm to solve the optimization function under the constraints of the process parameter adjustment range and the robot's execution capability, to obtain a set of optimized process parameter target values, includes: The adjustment range of process parameters and the constraints of robot execution capability are encoded as the boundary of the feasible solution space of the optimization problem; If a gradient-based optimization algorithm is used, the gradient of the optimization function with respect to each process parameter is calculated, and the process parameter values ​​are iteratively updated along the gradient descent direction within the boundary of the feasible solution space until the convergence condition is met. If a heuristic search algorithm is used, multiple candidate solutions for process parameters are generated and evaluated according to preset heuristic rules within the feasible solution space. Through simulated evolution or local search strategies, the combination of process parameters that minimizes the optimization function value is iteratively searched as the target value of the optimized process parameters.

[0013] As a further aspect of the present invention, the method further includes: converting the real-time process parameter adjustment instruction set into a robot control instruction stream and sending it to the production robot at the corresponding workstation to drive it to perform adjustments to optimize the production quality of subsequent cables, specifically including: Analyze each adjustment action in the set of real-time process parameter adjustment instructions to determine its corresponding target station robot, target process parameter, and adjustment amount; Based on the target robot's model and control protocol, each adjustment action is translated into a native control command that is recognized by the underlying robot controller. The native control command includes motion instructions, speed settings, torque control, or I / O signal operations. Based on the production line's cycle time and the coordination relationship between each workstation, an execution timestamp is assigned to each native control command to ensure that adjustment actions are executed in an orderly manner without interfering with the normal production cycle time; All the original control commands with timestamps are arranged in chronological order and combined into the robot control command stream.

[0014] As a further aspect of the present invention, the present invention also includes a robot-based cable production quality management system, the system including a processor and a memory, the memory and the processor being connected, the memory being used to store programs, instructions or code, and the processor being used to run the programs, instructions or code in the memory to implement the robot-based cable production quality management method as described above.

[0015] Compared with the prior art, the advantages and positive effects of the present invention are as follows: Based on the physical constraints and process rules in cable production, the latent space of the original variational autoencoder is structurally constrained, resulting in an improved variational autoencoder algorithm. This algorithm processes a multi-source production data set, consisting of raw material spectral feature sequences, process parameter time-series flows, multi-view visual images, and online sensor measurements collected by a robot, to generate a deep quality feature representation. This processing method ensures that the data feature extraction process aligns with the physical laws and process specifications of cable production, avoiding feature deviation problems caused by the disordered latent space of the original variational autoencoder. The deep quality feature representation can fully carry the core information of the multi-source production data, eliminating the limitations of single-data-dimensional feature extraction, and ensuring a precise correspondence between quality features and the actual cable production process.

[0016] The process involves performing process state transition modeling on the time-series flow of process parameters to generate process state evolution paths, and performing 3D defect reconstruction on multi-view visual images to generate cable appearance defect voxel models. Deep quality feature representations, process state evolution paths, and cable appearance defect voxel models are input into a multimodal fusion network to generate a comprehensive quality tensor. An attention-based quality defect prediction network analyzes the comprehensive quality tensor to output the probability distribution of multiple potential defects in future pre-set production stages of the cable product. An online control strategy generator then calculates and outputs a set of real-time process parameter adjustment instructions for the production line robot. The process state evolution path fully presents the dynamic change trend of process parameters, the appearance defect voxel model reconstructs the 3D spatial morphology of defects, multimodal data fusion integrates various quality-related information, and the attention mechanism focuses on defect correlation features in the comprehensive quality tensor, accurately outputting the defect probability distribution for future production stages. Based on the probability distribution, parameter adjustment instructions adapted to the production line robot are directly generated, achieving real-time dynamic control of the production process. Attached Figure Description

[0017] Figure 1 This is a flowchart of a robot-based cable production quality management method according to the present invention; Figure 2 A flowchart for generating deep quality feature representations for an improved variational autoencoder; Figure 3 A flowchart for generating a voxel model of cable appearance defects for 3D defect reconstruction. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0019] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0020] See Figure 1 This invention provides a robot-based cable production quality management method, the specific method including: This process acquires a multi-source production data set collected by robots during the production of the target cable. This set includes raw material spectral feature sequences, process parameter time-series flows, multi-view visual images, and online sensor measurements. An improved variational autoencoder algorithm is applied to this multi-source production data set. Based on the physical constraints and process rules in cable production, structural constraints are imposed on the latent space of the original variational autoencoder to generate a deep quality feature representation. Process state transition modeling is performed on the process parameter time-series flows to generate process state evolution paths. Simultaneously, 3D defect reconstruction is performed on the multi-view visual images to generate a voxel model of cable appearance defects. The deep quality feature representation, process state evolution paths, and cable appearance defect voxel models are input into a multimodal fusion network to generate a comprehensive quality tensor for the cable product. An attention-based quality defect prediction network is invoked to analyze the comprehensive quality tensor, outputting the probability distribution of multiple potential defects in the cable product during future pre-set production stages. Based on the probability distribution of multiple potential defects, an online control strategy generator calculates and outputs a set of real-time process parameter adjustment instructions for the production line robots.

[0021] In one embodiment of the present invention, see [reference] Figure 2 An encoder network is constructed to encode the spectral feature sequence of raw materials, the time-series flow of process parameters, multi-view visual images, and online sensor measurements, respectively, to obtain their corresponding initial latent variables. These initial latent variables are then input into the structured latent space of an improved variational autoencoder algorithm. This structured latent space, based on the physical constraints and process rules of cable production, imposes constraints on the distribution and relationships of the latent variables, forcing latent variables representing the same physical attribute to be statistically correlated and limiting the probability of combinations of latent variables that violate known process rules. In the structured latent space, latent variables that satisfy the constraints are sampled and input into the corresponding decoder network for data reconstruction. During training, by minimizing the difference between the reconstruction error and the prior distribution of latent variables that satisfy the physical constraints and process rules, the encoder learns a deep quality feature representation that contains physical meaning and conforms to process rules.

[0022] In specific implementations, an encoder network is constructed to encode the raw material spectral feature sequence, the time-series process parameters, multi-view visual images, and online sensor measurements separately, obtaining their respective initial latent variables. Specifically, the raw material spectral feature sequence is encoded into spectral latent variables using a one-dimensional convolutional neural network; the time-series process parameters are encoded into process latent variables using a long short-term memory network; the multi-view visual images are encoded into visual latent variables using a two-dimensional convolutional neural network; and the online sensor measurements are encoded into sensor latent variables using a fully connected network. In some embodiments, the output dimension of the encoder network is configured according to the characteristics of each data type. For example, the dimension of the spectral latent variables is set to 32 dimensions, the dimension of the process latent variables is set to 64 dimensions, the dimension of the visual latent variables is set to 128 dimensions, and the dimension of the sensor latent variables is set to 16 dimensions to ensure the feature representation capability of different types of data.

[0023] In practical implementation, the corresponding initial latent variables are input into the structured latent space of the improved variational autoencoder algorithm. The structured latent space imposes constraints on the distribution and relationships of the latent variables based on the physical constraints and process rules of cable production. Physical constraints include the correlation rules between raw material components and process temperature, while process rules include the matching relationship between extrusion speed and tension. The constraints imposed on the relationships in the structured latent space include: forcing latent variable dimensions representing the same physical property to be statistically correlated; for example, the spectral latent variable dimension representing the melt index of the insulation material and the process latent variable dimension representing extrusion temperature are forced to have a positive correlation; and limiting the probability of latent variable combinations that violate known process rules. For example, the combination of high extrusion speed and low tension is given an extremely low prior probability because it easily leads to cable eccentricity.

[0024] In practical implementation, latent variables satisfying the constraints are sampled in the structured latent space and input into the corresponding decoder network for data reconstruction. The sampling process is carried out from the posterior distribution that satisfies physical constraints and process rules. The structure of the decoder network is symmetrical to the encoder network, including a spectral decoder, a process decoder, a vision decoder, and a sensor decoder, which respectively reconstruct output data with the same dimension as the original input data. It can be understood that the reconstruction error is calculated using the mean squared error loss function to measure the difference between the original input data and the reconstructed data.

[0025] In practical implementation, during the training process, by minimizing the difference between the reconstruction error and the prior distribution of latent variables that satisfy physical constraints and process rules, the encoder learns a deep quality feature representation that contains physical meaning and conforms to process rules; the total loss function is defined as:

[0026] in: Represents the reconstruction error term. Represents the KL divergence term. Indicates the given input data Latent variables The posterior distribution, This represents a structured prior distribution that incorporates physical constraints and process rules. It is a hyperparameter that balances the weights of the two components. It can be understood that by optimizing the total loss function through the backpropagation algorithm, the deep quality feature representation can both retain the original data information and conform to the physical laws of cable production.

[0027] In one embodiment of the present invention, key process parameters in the time-series flow of process parameters are identified. These key process parameters include temperature, pressure, extrusion speed, and tension. On the time series of these key process parameters, multiple stable process state intervals are segmented using a state change point detection algorithm. For each stable interval, the statistical characteristics of the key process parameters are calculated as the state representation vector. Based on the time sequence, the state representation vectors of consecutively occurring stable process state intervals are connected to form a preliminary state evolution sequence. A hidden Markov model is used to model the preliminary state evolution sequence, inferring a hidden process state sequence. This hidden process state sequence and its corresponding state representation vector are used together as the process state evolution path. (See also...) Figure 3 For each multi-view visual image, cable target segmentation and feature point extraction are performed to obtain the cable's contour and surface key points under different viewpoints. Using multi-view geometric principles, the coordinates of the cable surface key points in three-dimensional space are calculated to reconstruct a three-dimensional point cloud model of the cable surface. Based on the three-dimensional point cloud model, surface bumps, scratches, or contamination areas that exist compared with the standard cable three-dimensional model are detected and identified as potential appearance defects. The three-dimensional point cloud data corresponding to the potential appearance defect areas are voxelized and filled into voxel meshes with space occupation to generate a cable appearance defect voxel model containing defect location, shape, and size information.

[0028] In practical implementation, key process parameters in the time-series flow of process parameters are identified. These key parameters include temperature, pressure, extrusion speed, and tension. The time-series flow of process parameters comes from the production line monitoring system, recording minute-by-minute temperature readings, pressure sensor outputs, extruder speed, and traction device tension settings. Multiple stable intervals in the time series of key process parameters are segmented using a state change point detection algorithm. This algorithm employs a sliding window-based mean mutation detection with a window length of 10 sampling points. A state change point is identified when the mean change of any key process parameter exceeds twice its historical standard deviation. Statistical features of the key process parameters are calculated for each stable interval as the state representation vector. These statistical features include mean, variance, and autocorrelation coefficient. The mean and variance of temperature, pressure, extrusion speed, and tension constitute an 8-dimensional state representation vector. Based on the time sequence, the state representation vectors of consecutively occurring stable intervals are connected to form a preliminary state evolution sequence. If a cable production record contains 5 stable intervals, the connected preliminary state evolution sequence is a 40-dimensional vector sequence. Hidden Markov Models (HMMs) are used to model the initial state evolution sequence and infer the hidden process state sequence. The HMM defines three hidden process states: ideal state, fluctuating state, and abnormal state. The optimal state sequence is decoded using the Viterbi algorithm. The hidden process state sequence and its corresponding state representation vector are used together as the process state evolution path. The hidden process state sequence labels the state category of each stable interval, and the state representation vector records specific parameter characteristics. The combination of both provides a complete description of the process dynamics.

[0029] In some embodiments, cable target segmentation and feature point extraction are performed on each multi-view visual image to obtain the cable's contour and surface key points from different perspectives. The multi-view visual images are simultaneously captured by industrial cameras installed at the extrusion die outlet, cooling water tank inlet, and take-up reel. Cable target segmentation uses the U-Net semantic segmentation network, and feature point extraction uses the SIFT operator, extracting 200 key points per image. Using multi-view geometry principles, the coordinates of the cable surface key points in three-dimensional space are calculated to reconstruct a three-dimensional point cloud model of the cable surface. The multi-view geometry principle is based on epipolar geometry and triangulation. Assuming the camera's intrinsic and extrinsic parameters are calibrated, the pixel coordinates of the same key point are matched in at least two views, back-projected onto the world coordinate system to obtain three-dimensional coordinates, generating a dense point cloud of approximately 50,000 points. Based on the 3D point cloud model, surface irregularities, scratches, or contamination areas compared to the standard cable 3D model are detected and identified as potential appearance defects. The standard cable 3D model is a defect-free cylindrical surface designed in CAD. By calculating the distance field from the point cloud to the standard model, areas with a distance greater than 0.2mm are marked as surface irregularities, continuous concave points constitute scratches, and dissimilar colored areas are clustered as contamination. The 3D point cloud data corresponding to the potential appearance defect areas are voxelized and filled into a voxel mesh with spatial occupancy, generating a cable appearance defect voxel model containing information on defect location, shape, and size. The voxelization resolution is 0.1mm × 0.1mm × 0.1mm, the defect location is recorded by the voxel center coordinates, the shape is described by connected component labels, and the dimensions are the length, width, and height of the bounding box.

[0030] In practice, the construction of the process state evolution path and the generation of the cable appearance defect voxel model are carried out independently, and the results of both are integrated in the subsequent multimodal fusion network. The process state evolution path reflects the historical trend of process parameters, while the cable appearance defect voxel model records the external defects of the current product. The two complement each other to characterize the production process and product quality. It can be understood that the division of the stable interval of the process state depends on the sensitivity of the state change point detection algorithm. Smaller windows can capture rapid fluctuations but increase noise, while larger windows smooth fluctuations but may ignore transient anomalies. Optionally, the reconstruction accuracy of the 3D point cloud model is affected by camera resolution and matching accuracy. Higher resolution cameras and sub-pixel matching algorithms can improve point cloud density, thereby more finely characterizing defect morphology.

[0031] In one embodiment of the present invention, the multimodal fusion network includes a feature alignment subnetwork, a feature interaction subnetwork, and a feature aggregation subnetwork. The feature alignment subnetwork maps the temporal dimension of the deep quality feature representation, the state dimension of the process state evolution path, and the spatial dimension of the cable appearance defect voxel model to a feature space, thereby achieving scale and semantic alignment of the multimodal features. The feature interaction subnetwork establishes a bidirectional attention mechanism between the deep quality feature representation, the process state evolution path, and the cable appearance defect voxel model within the feature space, simulating how process parameters affect the formation of internal quality features and appearance defects. The feature aggregation subnetwork performs weighted fusion and compression on the multimodal features after sufficient interaction, outputting a comprehensive quality tensor that fully characterizes the cable material, process, and appearance state. The quality defect prediction network comprises a multi-head self-attention layer and a feedforward prediction network. The comprehensive quality tensor is input into the multi-head self-attention layer, which facilitates information exchange and weight focusing between different feature regions within the tensor, highlighting the feature combinations most relevant to various defects. The features processed by the multi-head self-attention layer are then input into the feedforward prediction network. This network maps the input features into a multi-dimensional vector, where each dimension corresponds to a pre-defined potential defect in a future production stage, with its value representing the probability of the defect occurring. The probability values ​​corresponding to all dimensions are normalized to form a probability distribution of multiple potential defects for cable products in the pre-defined future production stages.

[0032] In its implementation, the multimodal fusion network comprises a feature alignment subnetwork, a feature interaction subnetwork, and a feature aggregation subnetwork. The feature alignment subnetwork maps the temporal dimension of the deep quality feature representation, the state dimension of the process state evolution path, and the spatial dimension of the cable appearance defect voxel model to a unified feature space. The unified feature space is set to 256 dimensions to achieve scale and semantic alignment of multimodal features. The deep quality features are represented as a 16-dimensional temporal vector sequence from the improved variational autoencoder, the process state evolution path is an 8-dimensional state representation vector sequence, and the cable appearance defect voxel model is a 32×32×32 binary voxel grid. The feature alignment subnetwork uses temporal convolution to map the deep quality feature representation to 256 dimensions, uses fully connected layers to map the process state evolution path to 256 dimensions, and uses 3D convolution to flatten and map the cable appearance defect voxel model to a 256-dimensional vector, thus aligning the features of the three modalities in the same dimensional space.

[0033] In its implementation, the feature interaction subnetwork establishes a bidirectional attention mechanism within a unified feature space between deep quality feature representations, process state evolution paths, and cable appearance defect voxel models. This simulates how process parameters influence the formation of internal quality features and appearance defects. The bidirectional attention mechanism calculates the similarity matrix between any two modal features, such as the dot product of process state evolution path features and cable appearance defect voxel model features. Softmax normalization is then used to obtain attention weights, which are applied to enhance process state features strongly correlated with defects. The feature aggregation subnetwork performs weighted fusion and compression on the multimodal features after sufficient interaction, outputting a comprehensive quality tensor that fully characterizes the cable material, process, and appearance state. The weighted fusion uses learnable weight parameters to linearly weight the feature vectors of the three modalities. The compression process reduces the dimensionality of the 256×3 spliced ​​features to a 512-dimensional comprehensive quality tensor through a fully connected layer.

[0034] In some embodiments, the quality defect prediction network includes a multi-head self-attention layer and a feedforward prediction network. The comprehensive quality tensor is input into the multi-head self-attention layer, which enables information exchange and weight focusing between different feature regions in the comprehensive quality tensor, highlighting the feature combinations most relevant to various defects. The parameter configuration of the multi-head self-attention layer is shown in Table 1. Table 1: Multi-head Self-Attention Layer Parameter Configuration Table Parameter name Value Head count 8 Query / Key / Value Dimension 64 Dropout rate 0.1 In the specific implementation, the features processed by the multi-head self-attention layer are input into the feedforward prediction network. The feedforward prediction network contains two fully connected layers, with a hidden layer dimension of 256 and the activation function being ReLU. The feedforward prediction network maps the input features into a multi-dimensional vector. Each dimension of the multi-dimensional vector corresponds to a type of potential defect in a pre-defined future production stage, and its value represents the probability of the defect occurring. The pre-defined future production stages are cross-linking, sheath extrusion, and finished product testing. The corresponding potential defects include insufficient cross-linking, sheath misalignment, and insulation breakdown; therefore, the multi-dimensional vector is 3-dimensional. The probability values ​​corresponding to all dimensions are normalized to form the probability distribution of multiple potential defects of the cable product in the pre-defined future production stages. The normalization uses the Softmax function, and the formula is as follows:

[0035] in: Indicates the first The probability of occurrence of class defects, The output of the feedforward prediction network is the first Original score, This represents the k-th dimension of the original score output by the feedforward prediction network. This represents the total number of defect categories. It is understandable that multi-head self-attention layers can capture long-range dependencies across modes in the comprehensive quality tensor, such as the correlation between high-temperature fluctuations in the process state evolution path and surface charring areas in the voxel model of cable appearance defects, thereby improving the accuracy of defect prediction.

[0036] In one embodiment of the present invention, the online control strategy generator receives multiple potential defect probability distributions and real-time status feedback of the current production line. Within the online control strategy generator, an optimization function is established with the objective of minimizing the overall defect occurrence probability. The decision variables of the optimization function are adjustable process parameters. Using a gradient-based optimization algorithm or a heuristic search algorithm, the optimization function is solved under the constraints of the process parameter adjustment range and the robot's execution capability to obtain a set of optimized process parameter target values. The difference between the optimized process parameter target values ​​and the current process parameter set values ​​is calculated, and the difference is converted into specific adjustment actions that can be executed by the production line robot, forming a set of real-time process parameter adjustment instructions. The adjustment range of process parameters and the constraints of robot execution capability are encoded as the boundary of the feasible solution space of the optimization problem. If a gradient-based optimization algorithm is used, the gradient of the optimization function with respect to each process parameter is calculated. Within the boundary of the feasible solution space, the process parameter values ​​are iteratively updated along the gradient descent direction until the convergence condition is met. If a heuristic search algorithm is used, multiple candidate solutions of process parameters are generated and evaluated according to preset heuristic rules within the feasible solution space. Through simulated evolution or local search strategies, the combination of process parameters that minimizes the value of the optimization function is iteratively searched as the target value of the optimized process parameters.

[0037] In practical implementation, the online control strategy generator receives probability distributions of multiple potential defects and real-time status feedback from the current production line. The probability distributions, derived from the quality defect prediction network, include probabilities of insufficient cross-linking, sheath misalignment, and insulation breakdown. The real-time status feedback includes the extruder's current temperature, screw speed, and traction tension setpoint. Within the online control strategy generator, an optimization function is established to minimize the overall defect occurrence probability. The decision variables of this function are adjustable process parameters: extrusion temperature, extrusion speed, and traction tension. The optimization objective is to minimize the weighted sum of the probabilities of the three types of defects, with weights set according to the defect severity. Using gradient-based optimization algorithms or heuristic search algorithms, the optimization function is solved under constraints of process parameter adjustment range and robot execution capability, yielding a set of optimized process parameter target values. The process parameter adjustment range and robot execution capability constraints are determined by equipment performance; for example, the allowable adjustment range for extrusion temperature is ±5℃, the extrusion speed adjustment does not exceed 10% of its current value, and the maximum adjustment of traction tension is ±50N. The difference between the optimized target value of the process parameters and the current set value of the process parameters is calculated and converted into specific adjustment actions that can be executed by the production line robot, forming a set of real-time process parameter adjustment instructions. The difference value is the difference between the target value and the current value. For example, if the target extrusion temperature is 3°C higher than the current temperature, the adjustment action is "increase the temperature by 3°C".

[0038] In practice, the adjustment range of process parameters and the constraints of robot execution capability are encoded as the boundary of the feasible solution space of the optimization problem; the boundary is represented by the upper and lower limits of the parameters, see Table 2: Table 2: Constraint Range for Process Parameter Adjustment Process parameters lower limit upper limit unit Extrusion temperature 180 190 ℃ Extrusion speed 90 110 % traction tension 450 550 N The specific form of the optimization function is:

[0039] in: For overall defect risk, Represents a vector of decision variables (extrusion temperature, extrusion speed, traction tension). , , These are the probabilities of three types of defects given by the quality defect prediction network. , , The corresponding weights are 0.5, 0.3, and 0.2. If a gradient-based optimization algorithm is used, the gradient of the optimization function with respect to each process parameter is calculated. Within the feasible solution space boundary, the process parameter values ​​are iteratively updated along the gradient descent direction until the convergence condition is met. The gradient components are approximated using the numerical difference method, with a step size of 0.5℃ adjusted in each iteration. After updating, it is checked for boundary overflow; if overflow occurs, the value is truncated to the boundary value. The convergence condition is that the change in the objective function between two consecutive iterations is less than 1e-4. If a heuristic search algorithm is used, multiple candidate solutions for process parameters are generated and evaluated according to preset heuristic rules within the feasible solution space. Through simulated evolution or local search strategies, the combination of process parameters that minimizes the optimization function value is iteratively searched as the target value for optimization. The heuristic rule uses simulated annealing, with an initial temperature of 100℃ and a cooling coefficient of 0.95. Ten neighborhood candidate solutions are generated in each round, and the solution with the better target value is selected to update the current solution until the temperature drops below 1e-3.

[0040] In some embodiments, gradient-based optimization algorithms are suitable for scenarios with relatively smooth defect probability functions, converge quickly but are prone to getting trapped in local optima; heuristic search algorithms have strong global search capabilities and are suitable for nonlinear and complex probabilistic response surfaces, but have higher computational costs. It can be understood that the generation frequency of the process parameter adjustment instruction set is synchronized with the production cycle, and each cable triggers an optimization calculation once during a critical process stage to ensure that the adjustment instructions are timely and effective.

[0041] In one embodiment of the present invention, each adjustment action in the real-time process parameter adjustment instruction set is parsed to determine its corresponding target station robot, target process parameter, and adjustment amount; according to the target robot's model and control protocol, each adjustment action is translated into a native control command recognized by the underlying robot controller. The native control command includes motion instructions, speed settings, torque control, or IO signal operations; based on the production line's cycle time and the coordination relationship between each station, an execution timestamp is assigned to each native control command to ensure that the adjustment action is executed in an orderly manner without interfering with the normal production cycle time; all timestamped native control commands are arranged in chronological order and combined into a robot control instruction stream.

[0042] In practical implementation, each adjustment action in the real-time process parameter adjustment instruction set is analyzed to determine its corresponding target station robot, target process parameter, and adjustment amount. The real-time process parameter adjustment instruction set contains multiple independent instructions, such as "heat up the extruder at station 3 by 3°C" and "increase the tension of the traction machine at station 5 by 20N". The analysis module identifies the target station identifier, process parameter type, and numerical adjustment amount in the instruction. Based on the target robot's model and control protocol, each adjustment action is translated into a native control command recognized by the underlying robot controller. The native control commands include motion commands, speed settings, torque control, or IO signal operations. If the extruder temperature control unit of target station 3 supports the Modbus-TCP protocol, then "heat up by 3°C" is translated into a command frame that writes the temperature setpoint to the holding register address 0x3001. If the traction machine controller of target station 5 uses the EtherCAT protocol, then "increase the tension by 20N" is mapped to a numerical update of a specific object dictionary entry in the PDO mapping.

[0043] In practice, based on the production line's cycle time and the coordination relationship between each workstation, an execution timestamp is assigned to each native control command to ensure that adjustments are executed in an orderly manner without disrupting the normal production cycle time. The main cycle time of the production line is 10 seconds, and the equipment actions and process adjustments at each workstation must be completed within a fixed time window. The formula for calculating the execution timestamp is:

[0044] in: For the command execution timestamp, As the current system time base, To adjust the beat number to which the action belongs, The cycle length is 10 seconds. Timestamps are embedded in the header or scheduling queue of each native control command, and the control system triggers execution in time-stamp order. All timestamped native control commands are arranged chronologically and combined into a robot control command stream. The command stream is stored in the central controller's memory as a queue. Each element contains a timestamp, target device address, control command code, and parameter value. For example, timestamp 10:30:05 corresponds to the temperature control write command for station 3, and timestamp 10:30:08 corresponds to the tension update command for station 5.

[0045] In some embodiments, the timestamp allocation strategy considers the causal dependence of process adjustments on equipment response delays. For example, extruder temperature adjustments must be executed at least 2 seconds before extrusion speed changes to avoid material melting fluctuations. Such constraints are encoded as timestamp offset rules and automatically applied when generating the command stream. Optionally, the transmission of robot control command streams uses real-time Ethernet protocols, such as PROFINETIRT or EtherCAT, to ensure that timestamped commands are executed synchronously with microsecond-level precision, meeting the timing consistency requirements of high-speed production lines. It can be understood that commands in the command stream are strictly ordered by timestamps, allowing adjustment actions from different workstations to coordinate within a unified timing framework, avoiding process disruptions caused by execution sequence conflicts.

[0046] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A robot-based cable production quality management method, characterized in that, The method includes: Acquire a set of multi-source production data collected by robots during the production process of the target cable, including raw material spectral feature sequences, process parameter time-series flow, multi-view visual images, and online sensor measurements. An improved variational autoencoder algorithm is applied to the multi-source production data set to generate a deep quality feature representation. The improved variational autoencoder algorithm provides structured constraints on the potential space of the original variational autoencoder based on the physical constraints and process rules in cable production. The process state transition model is performed on the time-series flow of the process parameters to generate the process state evolution path, and the three-dimensional defect reconstruction is performed on the multi-view visual image to generate a voxel model of cable appearance defects. The deep quality feature representation, process state evolution path and cable appearance defect voxel model are input into a multimodal fusion network to generate the comprehensive quality tensor of the cable product. The comprehensive quality tensor is analyzed by calling an attention-based quality defect prediction network, and the probability distribution of multiple potential defects of the cable product in the future preset production stage is output. Based on the probability distribution of the various potential defects, an online control strategy generator calculates and outputs a set of real-time process parameter adjustment instructions for the production line robot.

2. The robot-based cable production quality management method according to claim 1, characterized in that, An improved variational autoencoder algorithm is applied to the multi-source production data set to generate a deep quality feature representation, including: An encoder network is constructed to encode the raw material spectral feature sequence, process parameter time flow, multi-view visual image and online sensor measurement value respectively, so as to obtain their respective initial latent variables; The corresponding initial latent variables are input into the structured latent space of the improved variational autoencoder algorithm. The structured latent space imposes constraints on the distribution and relationship of the latent variables based on the physical constraints and process rules of cable production. The constraint on the relationship includes: forcing the dimensions of latent variables representing the same physical attribute to be statistically correlated, and limiting the probability of latent variable combinations that violate known process rules. In the structured latent space, latent variables that satisfy the constraints are sampled and input into the corresponding decoder network for data reconstruction; During training, by minimizing the difference between the reconstruction error and the prior distribution of latent variables that satisfy the physical constraints and process rules, the encoder learns the deep quality feature representation that contains physical meaning and conforms to the process rules.

3. The robot-based cable production quality management method according to claim 1, characterized in that, Perform process state transition modeling on the time-series flow of the process parameters to generate process state evolution paths, including: Identify key process parameters in the process parameter time-series flow, including temperature, pressure, extrusion speed, and tension; On the time series of the key process parameters, multiple stable intervals of the process state are segmented by the state change point detection algorithm, and the statistical characteristics of the key process parameters are calculated for each stable interval as the state representation vector of the stable interval. Based on the time sequence, the state representation vectors of consecutive stable intervals of process states are connected to form a preliminary state evolution sequence. The initial state evolution sequence is modeled using a hidden Markov model to infer the hidden process state sequence. The hidden process state sequence and its corresponding state representation vector are used together as the process state evolution path.

4. The robot-based cable production quality management method according to claim 1, characterized in that, Perform three-dimensional defect reconstruction on the multi-view visual images to generate a voxel model of cable appearance defects, including: For each multi-view visual image, cable target segmentation and feature point extraction are performed to obtain the cable contour and surface key points under different viewpoints; By using multi-view geometry principles, the coordinates of key points on the cable surface in three-dimensional space are calculated, and a three-dimensional point cloud model of the cable surface is reconstructed. Based on the three-dimensional point cloud model, surface bumps, scratches, or contamination areas that exist compared with the standard cable three-dimensional model are detected and identified as potential appearance defects. The three-dimensional point cloud data corresponding to the potential appearance defect area is voxelized and filled into a voxel mesh with space occupation, generating the cable appearance defect voxel model containing defect location, shape and size information.

5. The robot-based cable production quality management method according to claim 1, characterized in that, The deep quality feature representation, process state evolution path, and cable appearance defect voxel model are input into a multimodal fusion network to generate a comprehensive quality tensor for the cable product, including: The multimodal fusion network includes a feature alignment subnetwork, a feature interaction subnetwork, and a feature aggregation subnetwork; The feature alignment sub-network maps the temporal dimension of the deep quality feature representation, the state dimension of the process state evolution path, and the spatial dimension of the cable appearance defect voxel model to a feature space, thereby achieving scale and semantic alignment of multimodal features. The feature interaction subnetwork establishes a bidirectional attention mechanism between the deep quality feature representation, the process state evolution path, and the cable appearance defect voxel model within the feature space, simulating how process parameters affect the formation of internal quality features and appearance defects. The feature aggregation subnetwork performs weighted fusion and compression on the multimodal features after sufficient interaction, and outputs a comprehensive mass tensor that fully characterizes the cable material, process and appearance.

6. The robot-based cable production quality management method according to claim 1, characterized in that, The call to the attention-based quality defect prediction network analyzes the comprehensive quality tensor and outputs the probability distribution of multiple potential defects in the cable product during future pre-set production stages, including: The quality defect prediction network comprises a multi-head self-attention layer and a feedforward prediction network; The comprehensive quality tensor is input into the multi-head self-attention layer, which enables information exchange and weight focusing between different feature regions in the comprehensive quality tensor, highlighting the feature combinations most relevant to various defects. The features processed by the multi-head self-attention layer are input into the feedforward prediction network. The feedforward prediction network maps the input features into a multi-dimensional vector. Each dimension of the multi-dimensional vector corresponds to a type of potential defect in a pre-defined future production process, and its value represents the probability of the defect occurring. The probability values ​​corresponding to all dimensions are normalized to form the probability distribution of multiple potential defects of the cable product in the future preset production stage.

7. The robot-based cable production quality management method according to claim 1, characterized in that, Based on the probability distribution of the various potential defects, an online control strategy generator calculates and outputs a set of real-time process parameter adjustment instructions for the production line robot, including: The online control strategy generator receives the probability distribution of the multiple potential defects and the real-time status feedback of the current production line; Within the online control strategy generator, an optimization function is established with the objective of minimizing the overall probability of defect occurrence. The decision variables of the optimization function are adjustable process parameters. Using gradient-based optimization algorithms or heuristic search algorithms, under the constraints of the process parameter adjustment range and robot execution capability, the optimization function is solved to obtain a set of optimized process parameter target values; The difference between the optimized process parameter target value and the current process parameter setting value is calculated, and the difference is converted into specific adjustment actions that can be executed by the production line robot, forming the real-time process parameter adjustment instruction set.

8. The robot-based cable production quality management method according to claim 7, characterized in that, The optimization function is solved using gradient-based optimization algorithms or heuristic search algorithms, under the constraints of the process parameter adjustment range and robot execution capability, to obtain a set of optimized process parameter target values, including: The adjustment range of process parameters and the constraints of robot execution capability are encoded as the boundary of the feasible solution space of the optimization problem; If a gradient-based optimization algorithm is used, the gradient of the optimization function with respect to each process parameter is calculated, and the process parameter values ​​are iteratively updated along the gradient descent direction within the boundary of the feasible solution space until the convergence condition is met. If a heuristic search algorithm is used, multiple candidate solutions for process parameters are generated and evaluated according to preset heuristic rules within the feasible solution space. Through simulated evolution or local search strategies, the combination of process parameters that minimizes the optimization function value is iteratively searched as the target value of the optimized process parameters.

9. The robot-based cable production quality management method according to claim 1, characterized in that, The method further includes: converting the real-time process parameter adjustment instruction set into a robot control instruction stream and sending it to the production robot at the corresponding workstation to drive it to perform adjustments to optimize the production quality of subsequent cables, specifically including: Analyze each adjustment action in the set of real-time process parameter adjustment instructions to determine its corresponding target station robot, target process parameter, and adjustment amount; Based on the target robot's model and control protocol, each adjustment action is translated into a native control command that is recognized by the underlying robot controller. The native control command includes motion instructions, speed settings, torque control, or I / O signal operations. Based on the production line's cycle time and the coordination relationship between each workstation, an execution timestamp is assigned to each native control command to ensure that adjustment actions are executed in an orderly manner without interfering with the normal production cycle time; All the original control commands with timestamps are arranged in chronological order and combined into the robot control command stream.

10. A robot-based cable production quality management system, characterized in that, The system includes a processor and a memory, the memory and the processor being connected. The memory is used to store programs, instructions or code, and the processor is used to run the programs, instructions or code in the memory to implement the robot cable production quality management method according to any one of claims 1-9.

Images