Automatic sorting system for nitrogen springs

Abstract

The invention relates to the technical field of equipment control, in particular to an automatic nitrogen spring sorting system which comprises a data acquisition module for acquiring geometric point cloud data of nitrogen springs through structured light three-dimensional scanning, acquiring pressure time sequence data through a high-precision pressure sensor, preprocessing the data and sending the data to a feature processing module. And the feature processing module generates a comprehensive quality vector and monitors data exception. And the dynamic adaptive control module performs performance simulation, performs product grading in combination with a fuzzy decision system, and dynamically adjusts the sorting priority through a particle swarm optimization algorithm. And the sorting execution module analyzes the control instruction, plans a collision-free motion track of the collaborative robot, realizes flexible grabbing through a six-dimensional force sensor and impedance control, and feeds back data. And the system evolution module adopts a federated learning framework to update the recognition model and realizes predictive maintenance through a long-short-term memory network. According to the system, full-process automation of nitrogen spring sorting is achieved, and the production efficiency and the product consistency are improved.

Description

Technical Field

[0001] This invention relates to the field of equipment control technology, and in particular to an automatic sorting system using nitrogen springs. Background Technology

[0002] Existing nitrogen spring sorting technology suffers from the following technical challenges: During nitrogen spring production, the inspection of parameters such as roundness and diameter, as well as subsequent sorting operations, heavily rely on manual labor. Operators must rely on subjective experience for visual observation and manual measurement, which easily introduces individual judgment differences and fatigue errors, leading to inconsistent inspection standards. Furthermore, the speed and accuracy of manual sorting are limited, and stability is difficult to maintain under continuous operation, resulting in prolonged production cycles and instability in product qualification rates due to human factors. For example, when inspecting cylinder roundness, operators on different shifts may have slight differences in their understanding of tolerance zones, leading to misjudgments of some critically damaged products. In the sorting stage, uneven manual handling not only slows down the overall production line speed but may also cause secondary quality issues due to bumps and knocks. All these factors collectively restrict the improvement of production efficiency and the guarantee of product batch consistency. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides an automatic sorting system for nitrogen springs, solving the technical problems of low production efficiency and poor product consistency caused by the lack of automated control in manual inspection and sorting.

[0004] To solve the above-mentioned technical problems, the specific contents of the present invention are as follows: The present invention provides an automatic sorting system using nitrogen springs, comprising a data acquisition module, a feature processing module, a dynamic adaptive control module, a sorting execution module, and a system evolution module; The data acquisition module is configured to acquire geometric point cloud data of the nitrogen spring through structured light 3D scanning and pressure time series data through a high-precision pressure sensor; perform outlier removal and downsampling processing on the geometric point cloud data, perform moving average filtering on the pressure time series data, and use an adaptive Kalman filter algorithm to perform sensor drift compensation on the processed geometric point cloud data and pressure time series data; package the drift-compensated geometric point cloud data and pressure time series data into a standardized data package and send it to the feature processing module; The feature processing module is configured to receive the standardized data packet; perform voxelization processing on the geometric point cloud data in the data packet, and input the voxelized point cloud data into a spatial encoder to extract geometric features; perform normalization processing on the pressure time series data in the data packet, and input the normalized pressure sequence data into a time series encoder to extract pressure dynamic features; input the geometric features and pressure dynamic features into a cross-attention fusion layer to generate a comprehensive quality vector, and send the comprehensive quality vector to the dynamic adaptive control module; the feature processing module also uses a local outlier factor algorithm to monitor the data distribution of the standardized data packet, and when an abnormal data distribution is detected, sends a re-detection command to the data acquisition module; The dynamic adaptive control module is configured to: receive the comprehensive quality vector; input the comprehensive quality vector into a digital twin model based on finite element analysis for performance simulation to obtain simulation results; input the simulation results and the comprehensive quality vector into a fuzzy decision system for product classification to obtain product grades; receive real-time throughput data of the production line and use a particle swarm optimization algorithm, with the product grades and real-time throughput data of the production line as input, to dynamically adjust the sorting priority and generate sorting control instructions with timestamps; and send the sorting control instructions to the sorting execution module. The sorting execution module is configured to receive the sorting control command; parse the sorting control command to obtain the target sorting destination, and plan the collision-free motion trajectory of the collaborative robot; during the process of the collaborative robot grasping the nitrogen spring along the motion trajectory, force control data is collected in real time through a six-dimensional force sensor, and the grasping posture is dynamically adjusted according to the force control data using an impedance control algorithm; after the sorting action is completed, the grasping success status and the force control data are fed back to the dynamic adaptive control module. The system evolution module is configured to obtain incremental data required for model training from the feature processing module and historical operation logs from the sorting execution module; employing a federated learning framework, it uses the incremental data and historical operation logs to perform local model training at each sorting terminal, and uploads the encrypted model parameters generated by the training to the central server for aggregation to generate global model parameters; the global model parameters are distributed to the feature processing module to update the identification model; simultaneously, it analyzes the equipment operation data in the historical operation logs through a long short-term memory network to predict the maintenance cycle and sends early warning information to the maintenance terminal.

[0005] Furthermore, in the nitrogen spring automatic sorting system of the present invention, the data acquisition module includes a structured light three-dimensional scanning unit and a pressure sensing unit; The structured light 3D scanning unit projects a grating pattern onto the surface of the nitrogen spring; an industrial camera acquires a modulated deformed grating image of the nitrogen spring surface; phase calculation is performed on the deformed grating image, and the 3D coordinates of the workpiece surface points are calculated based on the triangulation principle to generate 3D point cloud data representing the outer contour of the workpiece. The pressure sensing unit is connected to the nitrogen spring inflation valve to sample pressure data in a sealed cavity, converting the collected analog voltage signal into discrete values ​​to form time series data of the pressure decay curve.

[0006] Furthermore, in the nitrogen spring automatic sorting system of the present invention, the feature processing module includes a spatial encoder, a temporal encoder, and a cross-attention fusion layer; The spatial encoder receives voxelized point cloud data, extracts global geometric features and local shape features of the point cloud data through a self-attention mechanism, and outputs a geometric feature vector. The timing encoder receives normalized pressure sequence data, extracts dynamic response features and changing trends of the pressure data through a recurrent neural network structure, and outputs a pressure feature vector. The cross-attention fusion layer receives the geometric feature vector and the stress feature vector, calculates the correlation weight between the two feature modes, and performs a weighted summation of the geometric feature vector and the stress feature vector based on the correlation weight to generate a low-dimensional comprehensive quality vector.

[0007] Furthermore, in the nitrogen spring automatic sorting system of the present invention, the dynamic adaptive control module includes a digital twin model, a fuzzy decision system, and a particle swarm optimizer; The digital twin model uses the integrated mass vector as the initial condition and, based on the gas state equation and mechanical performance equation, simulates the stress distribution and fatigue life of a nitrogen spring under various load conditions, and outputs the performance simulation results. The fuzzy decision system receives the performance simulation results and the comprehensive quality vector, sets the membership function using a multi-level if-then rule base, classifies the nitrogen springs, and outputs the product grade classification results. The particle swarm optimizer uses production line throughput efficiency as the objective function and product grade classification results and real-time production line throughput data as inputs to dynamically calculate the optimal priority ranking of sorting instructions.

[0008] Furthermore, in the nitrogen spring automatic sorting system of the present invention, the sorting execution module includes a collaborative robot motion planner and an impedance controller; The collaborative robot motion planner receives sorting control instructions, parses the coordinates of the target workpiece and the sorting destination, and uses a fast random exploration tree algorithm to calculate the collision-free motion trajectory of the robotic arm from the current position to the target position. During the movement of the robotic arm along its trajectory, the impedance controller collects three-dimensional force and torque data in real time through a six-dimensional force sensor, simulating the robot's end effector as a spring and damping system, and dynamically adjusts the stiffness and damping parameters of the end effector based on the real-time force and torque data collected.

[0009] Furthermore, in the nitrogen spring automatic sorting system of the present invention, the system evolution module includes a federated learning framework and a long short-term memory network; The federated learning framework obtains incremental model data from the feature processing module and historical operation logs from the sorting execution module; it uses the stochastic gradient descent algorithm to train the model locally at each sorting terminal using the incremental data and historical operation logs; it uploads the encrypted model parameters generated by the training to the central server; the central server uses a weighted average algorithm to securely aggregate the encrypted model parameters to generate global model parameters; and it distributes the global model parameters to the feature processing module. The long short-term memory network collects current fluctuation data of the joint motors of the collaborative robot, vibration spectrum data of the reducer, and zero-point drift data of the force sensor; performs time-series pattern recognition on the collected equipment operation data; when the degradation trend of the equipment is identified, it calculates the remaining service life of each component; and sends an early warning message to the maintenance terminal before the remaining service life reaches a preset threshold.

[0010] Furthermore, in the nitrogen spring automatic sorting system of the present invention, when the feature processing module detects abnormal data, it performs the following steps: Calculate the local outlier factor of the newly input standardized data packet and the historical normal sample set in the multidimensional feature space; when the local outlier factor exceeds the set threshold, determine the current data as an abnormal data point; generate a re-inspection instruction including the workpiece identifier corresponding to the abnormal data point; send the re-inspection instruction to the data acquisition module through the reverse data stream; trigger the data acquisition module to re-execute the structured light 3D scan and pressure data acquisition on the current workpiece.

[0011] Furthermore, in the nitrogen spring automatic sorting system of the present invention, the dynamic adaptive control module is configured to adjust the sorting strategy as follows: The system monitors the working status of sorting stations and the length of the buffer queue in real time; it aims to minimize production line idle time and maximize the output of qualified products per unit time; it uses product grade classification results and real-time production line throughput data as inputs to dynamically calculate the optimal sorting priority through particle swarm optimization algorithm; and it updates the classification threshold and instruction sorting in the sorting decision tree based on the calculation results.

[0012] Furthermore, in the nitrogen spring automatic sorting system of the present invention, the sorting execution module is configured after completing the sorting as follows: The successful capture status flag, actual capture time, complete six-dimensional force control data sequence, and impedance control parameter adjustment record are packaged into a feedback data packet; the feedback data packet is then sent to the dynamic adaptive control module. The dynamic adaptive control module receives feedback data packets, adjusts records based on force control data and grasping parameters in the feedback data packets, and updates the grasping parameter knowledge base; it then optimizes the force control parameter settings in subsequent sorting control instructions based on the updated grasping parameter knowledge base.

[0013] Furthermore, in the nitrogen spring automatic sorting system of the present invention, the system evolution module is configured to perform predictive maintenance as follows: Collect current fluctuation data of joint motors, vibration spectrum data of reducers, and zero-point drift data of force sensors of collaborative robots; input the collected equipment operation data into a long short-term memory network for time-series pattern recognition; when the time-series pattern recognition results show a degradation trend of the equipment, calculate the remaining service life of each component; when the remaining service life reaches the preset maintenance threshold, send an early warning message to the maintenance terminal.

[0014] Beneficial effects of this invention; This invention automatically acquires geometric point cloud data and pressure time-series data of nitrogen springs through a data acquisition module, replacing existing manual observation and measurement methods, thus eliminating individual judgment differences and fatigue errors. The feature processing module generates a comprehensive quality vector through multimodal feature fusion, achieving unified detection standards. The dynamic adaptive control module utilizes a digital twin model and particle swarm optimization algorithm for real-time simulation and sorting strategy adjustment, improving the stability of production cycle time. The sorting execution module achieves compliant grasping through collaborative robots and six-dimensional force sensors, avoiding product damage and ensuring the accuracy of sorting actions. The system evolution module uses a federated learning framework and long short-term memory network to continuously optimize the model and perform predictive maintenance, maintaining system recognition accuracy and equipment reliability in the long term. All modules form a fully automated control chain through standardized data flow, significantly improving the production efficiency and batch consistency of nitrogen springs. Attached Figure Description

[0015] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on the drawings without creative effort.

[0016] Figure 1 This is a system architecture diagram of the nitrogen spring automatic sorting system of the present invention. Detailed Implementation

[0017] To make the technical solution of the present invention clearer, the present invention will be clearly and completely described below with reference to specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. The present invention provided by various embodiments will be described in detail below with reference to the accompanying drawings. To better understand the purpose of the present invention, the present invention will be described in further detail below.

[0018] Please see Figure 1 The present invention provides an automatic sorting system using nitrogen springs, comprising a data acquisition module, a feature processing module, a dynamic adaptive control module, a sorting execution module, and a system evolution module; The data acquisition module is configured to acquire geometric point cloud data of the nitrogen spring through structured light 3D scanning and pressure time series data through a high-precision pressure sensor; perform outlier removal and downsampling processing on the geometric point cloud data, perform moving average filtering on the pressure time series data, and use an adaptive Kalman filter algorithm to perform sensor drift compensation on the processed geometric point cloud data and pressure time series data; package the drift-compensated geometric point cloud data and pressure time series data into a standardized data package and send it to the feature processing module; The feature processing module is configured to receive the standardized data packet; perform voxelization processing on the geometric point cloud data in the data packet, and input the voxelized point cloud data into a spatial encoder to extract geometric features; perform normalization processing on the pressure time series data in the data packet, and input the normalized pressure sequence data into a time series encoder to extract pressure dynamic features; input the geometric features and pressure dynamic features into a cross-attention fusion layer to generate a comprehensive quality vector, and send the comprehensive quality vector to the dynamic adaptive control module; the feature processing module also uses a local outlier factor algorithm to monitor the data distribution of the standardized data packet, and when an abnormal data distribution is detected, sends a re-detection command to the data acquisition module; The dynamic adaptive control module is configured to: receive the comprehensive quality vector; input the comprehensive quality vector into a digital twin model based on finite element analysis for performance simulation to obtain simulation results; input the simulation results and the comprehensive quality vector into a fuzzy decision system for product classification to obtain product grades; receive real-time throughput data of the production line and use a particle swarm optimization algorithm, with the product grades and real-time throughput data of the production line as input, to dynamically adjust the sorting priority and generate sorting control instructions with timestamps; and send the sorting control instructions to the sorting execution module. The sorting execution module is configured to receive the sorting control command; parse the sorting control command to obtain the target sorting destination, and plan the collision-free motion trajectory of the collaborative robot; during the process of the collaborative robot grasping the nitrogen spring along the motion trajectory, force control data is collected in real time through a six-dimensional force sensor, and the grasping posture is dynamically adjusted according to the force control data using an impedance control algorithm; after the sorting action is completed, the grasping success status and the force control data are fed back to the dynamic adaptive control module. The system evolution module is configured to obtain incremental data required for model training from the feature processing module and historical operation logs from the sorting execution module; employing a federated learning framework, it uses the incremental data and historical operation logs to perform local model training at each sorting terminal, and uploads the encrypted model parameters generated by the training to the central server for aggregation to generate global model parameters; the global model parameters are distributed to the feature processing module to update the identification model; simultaneously, it analyzes the equipment operation data in the historical operation logs through a long short-term memory network to predict the maintenance cycle and sends early warning information to the maintenance terminal.

[0019] The data acquisition module acquires geometric point cloud data of the nitrogen spring using structured light 3D scanning technology. Specifically, a grating pattern is projected onto the surface of the nitrogen spring, and an industrial camera captures a deformed grating image modulated by the workpiece shape. Phase calculation is then performed on the deformed grating image, and the 3D coordinates of points on the workpiece surface are calculated based on triangulation principles, thus generating 3D point cloud data representing the workpiece's outer contour. Simultaneously, the data acquisition module acquires pressure time-series data using a high-precision pressure sensor. The pressure sensing unit is connected to the nitrogen spring inflation valve, sampling pressure data within a sealed cavity. The acquired analog voltage signal is converted into discrete values, forming a time-series data of the pressure decay curve. Next, the data acquisition module performs outlier removal and downsampling on the geometric point cloud data to remove noise and reduce data volume. A moving average filter is applied to the pressure time-series data to smooth data fluctuations. Finally, an adaptive Kalman filter algorithm is used to compensate for sensor drift in the processed geometric point cloud data and pressure time-series data, correcting measurement errors caused by environmental changes or equipment aging. Finally, the data acquisition module packages the drift-compensated geometric point cloud data and pressure time series data into a standardized data packet and sends it to the feature processing module to achieve a unified data format and facilitate transmission.

[0020] The feature processing module receives standardized data packets from the data acquisition module. First, it performs voxelization on the geometric point cloud data in the data packets, converting it into regular 3D mesh data to facilitate subsequent feature extraction. The voxelized point cloud data is then input into the spatial encoder, which extracts global geometric features and local shape features of the point cloud data using a self-attention mechanism, outputting a geometric feature vector. Simultaneously, the feature processing module normalizes the pressure time-series data in the data packets to eliminate the influence of dimensions. The normalized pressure sequence data is then input into the time-series encoder, which extracts the dynamic response features and trends of the pressure data using a recurrent neural network structure, outputting a pressure feature vector. Subsequently, the feature processing module inputs the geometric feature vector and the dynamic pressure feature vector into a cross-attention fusion layer. The cross-attention fusion layer calculates the correlation weights between the two feature modes and performs a weighted sum of the geometric and pressure feature vectors based on these correlation weights to generate a low-dimensional comprehensive mass vector. This vector comprehensively represents the dimensional accuracy and sealing performance of the nitrogen spring. The feature processing module also sends the comprehensive mass vector to the dynamic adaptive control module for subsequent decision-making. In addition, the feature processing module uses a local outlier factor algorithm to monitor the data distribution of standardized data packets. By calculating the local outlier factor of new input data and historical normal sample sets in the multidimensional feature space, when the local outlier factor exceeds the set threshold, the data distribution is determined to be abnormal, and a re-inspection instruction including the workpiece identifier is generated. This instruction is sent to the data acquisition module through the reverse data stream to trigger the re-scanning and inspection of the current workpiece, thus forming a closed loop of data quality control.

[0021] The dynamic adaptive control module receives the comprehensive quality vector from the feature processing module and inputs it into a digital twin model based on finite element analysis. Using the comprehensive quality vector as initial conditions, the digital twin model simulates the stress distribution and fatigue life of the nitrogen spring under various load conditions based on the gas state equation and mechanical performance equation, outputting performance simulation results. Subsequently, the dynamic adaptive control module inputs the performance simulation results and the comprehensive quality vector into a fuzzy decision system. The fuzzy decision system uses a multi-level if-then rule base to define membership functions, classifies the nitrogen spring, and outputs product grade classification results. Simultaneously, the dynamic adaptive control module receives real-time production line throughput data and uses a particle swarm optimization algorithm, with the product grade classification results and real-time production line throughput data as input, to dynamically calculate the optimal priority ranking of sorting instructions, aiming to minimize production line idle time and maximize the output of qualified products per unit time. Finally, the dynamic adaptive control module generates timestamped sorting control instructions and sends them to the sorting execution module, achieving synchronization and accuracy of the instructions.

[0022] The sorting execution module receives sorting control commands from the dynamic adaptive control module, parses these commands to obtain the coordinates of the target workpiece and the sorting destination, and plans a collision-free motion trajectory for the collaborative robot. The motion planner uses a fast random exploration tree algorithm to calculate the collision-free path for the robotic arm from its current position to the target position. During the collaborative robot's grasping of the nitrogen spring along the trajectory, the sorting execution module collects three-dimensional force and torque data in real time using a six-dimensional force sensor. An impedance control algorithm dynamically adjusts the grasping posture based on the real-time force control data. The impedance controller simulates the robot's end effector as a spring and damping system, dynamically adjusting the stiffness and damping parameters of the end effector based on the force and torque data to achieve compliant grasping. After the sorting action is completed, the sorting execution module feeds back the successful grasping status and force control data to the dynamic adaptive control module to optimize subsequent grasping parameters.

[0023] The system evolution module obtains incremental data for model training from the feature processing module and historical operation logs from the sorting execution module. It uses a federated learning framework to train the model locally at each sorting terminal, employing a stochastic gradient descent algorithm to incrementally train a lightweight neural network model. The generated encrypted model parameters are then uploaded to the central server. The central server uses a weighted average algorithm to securely aggregate the encrypted model parameters, generating global model parameters, which are then distributed to the feature processing module to update the recognition model, enabling continuous evolution of the system's recognition capabilities. Simultaneously, the system evolution module analyzes equipment operation data from historical operation logs using a long short-term memory network, collecting current fluctuation data from the collaborative robot's joint motors, vibration spectrum data from the reducer, and zero-point drift data from the force sensor. It performs time-series pattern recognition, and when a degradation trend is identified, it calculates the remaining service life of key components. Before the remaining service life reaches a preset threshold, it sends an early warning to the maintenance terminal, achieving predictive maintenance.

[0024] The data acquisition module of this invention includes a structured light 3D scanning unit and a pressure sensing unit. The structured light 3D scanning unit projects a grating pattern onto the surface of a nitrogen spring, and an industrial camera acquires a modulated deformed grating image of the nitrogen spring surface. Phase calculation is performed on the deformed grating image, and the 3D coordinates of points on the workpiece surface are calculated based on the triangulation principle to generate 3D point cloud data representing the outer contour of the workpiece. The pressure sensing unit is connected to the nitrogen spring inflation valve and performs pressure data sampling in a sealed cavity. The acquired analog voltage signal is converted into discrete values ​​to form time series data of the pressure decay curve. Each step achieves accurate acquisition of geometric and pressure data through optical measurement and electrical signal conversion, providing a reliable input source for subsequent feature processing.

[0025] The feature processing module of this invention includes a spatial encoder, a temporal encoder, and a cross-attention fusion layer. The spatial encoder receives voxelized point cloud data and extracts global geometric features and local shape features of the point cloud data through a self-attention mechanism, outputting a geometric feature vector. The temporal encoder receives normalized pressure sequence data and extracts dynamic response features and changing trends of the pressure data through a recurrent neural network structure, outputting a pressure feature vector. The cross-attention fusion layer receives the geometric feature vector and the pressure feature vector, calculates the correlation weight between the two feature modes, and performs a weighted summation of the geometric feature vector and the pressure feature vector based on the correlation weight to generate a low-dimensional comprehensive mass vector. Each step achieves a comprehensive characterization of the comprehensive mass of the nitrogen spring through multimodal feature fusion, providing an accurate decision basis for dynamic adaptive control.

[0026] The dynamic adaptive control module of this invention includes a digital twin model, a fuzzy decision system, and a particle swarm optimizer. The digital twin model, using a comprehensive mass vector as initial conditions, simulates the stress distribution and fatigue life of nitrogen springs under various load conditions based on the gas state equation and mechanical performance equation, outputting performance simulation results. The fuzzy decision system receives the performance simulation results and the comprehensive mass vector, uses a multi-level if-then rule base to set membership functions, classifies the nitrogen springs, and outputs product grade classification results. The particle swarm optimizer, using production line throughput efficiency as the objective function and the product grade classification results and real-time production line throughput data as input, dynamically calculates the optimal priority ranking of sorting instructions. Each step achieves real-time adjustment of the sorting strategy through simulation and optimization algorithms, improving the system's adaptability and efficiency.

[0027] The sorting execution module of this invention includes a collaborative robot motion planner and an impedance controller. The collaborative robot motion planner receives sorting control commands, analyzes the coordinates of the target workpiece and the sorting destination, and uses a fast random exploration tree algorithm to calculate the collision-free motion trajectory of the robotic arm from its current position to the target position. During the movement of the robotic arm along the motion trajectory, the impedance controller collects three-dimensional force and torque data in real time through a six-dimensional force sensor, simulating the robot end effector as a spring and damping system, and dynamically adjusts the stiffness and damping parameters of the end effector based on the real-time force and torque data. Each step achieves compliant grasping through motion planning and force control technology, avoiding product damage and ensuring the accuracy of the sorting action.

[0028] The system evolution module of this invention includes a federated learning framework and a long short-term memory network. The federated learning framework obtains incremental model data from the feature processing module and historical operation logs from the sorting execution module. It uses a stochastic gradient descent algorithm to train the model locally at each sorting terminal using the incremental data and historical operation logs. The encrypted model parameters generated by the training are uploaded to the central server. The central server uses a weighted average algorithm to securely aggregate the encrypted model parameters to generate global model parameters, which are then distributed to the feature processing module. The long short-term memory network collects current fluctuation data of the joint motors of the collaborative robot, vibration spectrum data of the reducer, and zero-point drift data of the force sensor. It performs time-series pattern recognition on the collected equipment operation data. When a degradation trend of the equipment is identified, it calculates the remaining service life of each component and sends an early warning message to the maintenance terminal before the remaining service life reaches a preset threshold. Each step achieves continuous optimization and reliability improvement of the system through distributed learning and predictive maintenance.

[0029] When the feature processing module of this invention detects abnormal data, it performs the following steps: calculating the local outlier factor of the newly input standardized data packet and the historical normal sample set in the multidimensional feature space; when the local outlier factor exceeds a set threshold, the current data is determined to be an abnormal data point; a re-inspection instruction including the workpiece identifier corresponding to the abnormal data point is generated; the re-inspection instruction is sent to the data acquisition module through the reverse data stream, triggering the data acquisition module to re-execute the structured light 3D scan and pressure data acquisition on the current workpiece; each step achieves data quality through anomaly detection and feedback mechanisms, avoiding the impact of erroneous data on subsequent processing.

[0030] When adjusting the sorting strategy, the dynamic adaptive control module of this invention is configured to: monitor the working status of the sorting station and the length of the buffer queue in real time; minimize production line idle time and maximize the output of qualified products per unit time as the optimization objective; take product grade classification results and real-time production line throughput data as input; dynamically calculate the optimal sorting priority through particle swarm optimization algorithm; and update the classification threshold and instruction sorting in the sorting decision tree according to the calculation results. Each step achieves dynamic adjustment of the sorting process through real-time optimization, thereby improving the overall efficiency of the production line.

[0031] After the sorting execution module of this invention completes sorting, it is configured to: package the successful grabbing status flag, actual grabbing time, complete six-dimensional force control data sequence, and impedance control parameter adjustment record into a feedback data packet, and send the feedback data packet to the dynamic adaptive control module; the dynamic adaptive control module receives the feedback data packet, updates the grabbing parameter knowledge base according to the force control data and grabbing parameter adjustment record in the feedback data packet, and optimizes the force control parameter settings in subsequent sorting control instructions according to the updated grabbing parameter knowledge base; each step optimizes the grabbing parameters through feedback data, realizing continuous improvement of sorting action.

[0032] When performing predictive maintenance, the system evolution module of this invention is configured to: collect current fluctuation data of the joint motors of the collaborative robot, vibration spectrum data of the reducer, and zero-point drift data of the force sensor; input the collected equipment operation data into a long short-term memory network for time-series pattern recognition; calculate the remaining service life of each component when the time-series pattern recognition results show a degradation trend of the equipment; and send an early warning message to the maintenance terminal when the remaining service life reaches a preset maintenance threshold. Each step achieves preventive maintenance through equipment status monitoring and life prediction, reducing system downtime.

[0033] This invention addresses the inefficiencies and product consistency issues caused by the lack of automation in manual inspection and sorting by constructing a fully automated intelligent sorting system. The system employs a data acquisition module to automatically acquire geometric point cloud data and pressure time-series data of nitrogen springs, replacing existing methods of manual observation and measurement, thus eliminating individual judgment differences and fatigue errors. The data acquisition module projects a grating pattern and acquires deformed images through a structured light 3D scanning unit, generating accurate point cloud data through phase calculation and triangulation. Simultaneously, a pressure sensing unit samples pressure data within a sealed cavity and converts it into a time series, avoiding subjective fluctuations inherent in manual measurements.

[0034] The feature processing module automates the processing of the collected data. It extracts features from voxelized point cloud data and normalized stress sequences using spatial and temporal encoders, and then generates a comprehensive quality vector through a cross-attention fusion layer. This process achieves intelligent fusion of multimodal data, overcoming the inconsistencies in standards encountered in manual detection. The feature processing module also employs a local outlier factor algorithm to monitor data distribution, automatically triggering re-detection upon anomaly detection, forming a closed-loop data quality control system and further ensuring the reliability of the input data.

[0035] After receiving the comprehensive quality vector, the dynamic adaptive control module simulates the performance of the nitrogen spring using a digital twin model and performs product classification using a fuzzy decision system. Simultaneously, it dynamically adjusts sorting priorities based on real-time production line throughput using a particle swarm optimization algorithm, generating timestamped control commands. This automated decision-making process replaces the uneven rhythm problem of manual sorting, achieving real-time adaptation of sorting strategies through simulation and optimization, thus improving the stability and efficiency of the production cycle.

[0036] After parsing the control commands, the sorting execution module plans the collision-free motion trajectory of the collaborative robot. During the grasping process, it collects force control data in real time through a six-dimensional force sensor and dynamically adjusts the grasping posture using an impedance control algorithm to achieve compliant grasping. After sorting is completed, the module feeds back the grasping status and force control data to the control module, forming an execution feedback loop. This automated execution mechanism avoids bumps and speed fluctuations that occur during manual handling, ensuring the accuracy and consistency of the sorting action.

[0037] The system evolution module collects incremental data and operation logs from various terminals through a federated learning framework, performs local model training and parameter aggregation, updates the recognition model of the feature processing module, and simultaneously uses long short-term memory networks to analyze equipment operation data, predict maintenance cycles, and send early warnings. This self-learning capability enables the system to continuously optimize recognition accuracy and equipment reliability, maintaining production efficiency and product consistency in the long term, and avoiding quality fluctuations caused by insufficient experience accumulation or personnel changes in manual systems.

[0038] The modules collaborate through standardized data packets, integrated quality vectors, and control commands, forming a fully automated control chain from data acquisition, processing, decision-making, execution to optimization. This integrated design eliminates bottlenecks in manual processes, achieving efficient operation of the sorting process and stable product quality output through real-time data-driven and feedback-based adjustments.

[0039] In the automated production of nitrogen springs, the system of this invention is deployed on a standard production line to process batches of cylinder components. The data acquisition module projects a grating pattern onto the workpiece surface using a structured light 3D scanning unit. After an industrial camera captures the deformed image, high-density point cloud data is generated through phase calculation and triangulation. Simultaneously, the pressure sensing unit samples the pressure decay curve within a sealed cavity, forming a time series. In actual operation, the point cloud data undergoes outlier removal and downsampling to eliminate noise introduced by surface contaminants or ambient light interference. The pressure data is smoothed by moving average filtering and combined with adaptive Kalman filtering to compensate for data offset caused by temperature drift in the sensor. For example, when the cylinder roundness parameter is detected to be at a tolerance critical point, the system may trigger a rescan command due to local point cloud anomalies, avoiding the risk of human error. The feature processing module inputs the voxelized point cloud into a spatial encoder, using a self-attention mechanism to extract global contour and local pit features. The time encoder identifies the trend changes in sealing performance from the pressure sequence, and finally generates a comprehensive quality vector through cross-attention fusion. This process can unify the detection standards in practical applications. However, it should be noted that if the surface of the workpiece is too reflective, it will affect the point cloud quality. In this case, the system will rely on pressure data as a supplementary judgment, which reflects the robustness of multimodal fusion.

[0040] Another embodiment involves a highly mixed, small-batch production environment where nitrogen springs exhibit significant specification variations, placing higher demands on sorting adaptability. The dynamic adaptive control module receives the comprehensive mass vector and drives a digital twin model to simulate stress distribution under different loads using finite element analysis. This, combined with a fuzzy decision system, classifies products. For example, for springs in critical states, the system sets soft thresholds using an if-then rule base to reduce misclassification caused by rigid classification. Simultaneously, a particle swarm optimizer monitors production line throughput in real time and dynamically adjusts sorting priorities. When the buffer queue is congested, higher-grade products are processed first to maximize the output of qualified products. The sorting execution module plans the collision-free trajectory of the collaborative robot, using a six-dimensional force sensor to provide real-time feedback of grasping force data. An impedance controller dynamically adjusts the stiffness parameters of the end effector to achieve compliant grasping. In actual testing, this design effectively avoids product collisions; however, the robotic arm's movement speed is limited by the workpiece weight, requiring acceleration constraints to be considered during planning. The system evolution module aggregates historical operation logs from each terminal through a federated learning framework to update the identification model parameters, while the long short-term memory network analyzes joint motor current fluctuation data to predict maintenance cycles. For example, when an abnormal vibration spectrum of the reducer is detected, the system sends an early warning, allowing the maintenance team to intervene before shutdown.

[0041] In the implementation of the automated nitrogen spring sorting system, the system is deployed in a standardized production workshop and integrated with existing production line equipment. The structured light 3D scanning unit of the data acquisition module is mounted on a fixed bracket above the conveyor belt. As the nitrogen spring passes through the scanning area with the conveyor belt, the scanning unit projects a grating pattern of a specific frequency onto the workpiece surface. An industrial camera acquires deformed grating images modulated by the workpiece surface morphology at a specific frame rate, and the image data is transmitted to an industrial control computer via gigabit Ethernet. The industrial control computer runs a phase calculation algorithm, calculating the 3D coordinates of each pixel based on the triangulation principle, generating 3D point cloud data. Simultaneously, the pressure sensing unit establishes a sealed connection with the nitrogen spring inflation valve via a quick connector, maintaining the connection for pressure monitoring after nitrogen filling. The pressure sensor acquires the pressure value inside the cavity at a specific sampling frequency, converting the analog voltage signal into discrete values ​​via an analog-to-digital converter, forming a pressure decay curve time series reflecting the sealing performance.

[0042] After receiving the standardized data packets from the data acquisition module, the feature processing module first performs voxelization on the geometric point cloud data, mapping the irregularly distributed point cloud onto a uniform 3D mesh. The spatial distribution features of points within each voxel are encoded into feature vectors and input into a spatial encoder based on the Transformer architecture. The spatial encoder calculates the correlation weights of different regions in the point cloud through a self-attention mechanism, extracting global geometric features including cylinder roundness and end face parallelism, while capturing local defect features such as surface pits and scratches. The pressure time series data undergoes min-max normalization to eliminate the influence of dimensions before being input into a time series encoder composed of gated recurrent units. The time series encoder analyzes the decay gradient and fluctuation pattern of the pressure sequence, outputting a dynamic feature vector characterizing the sealing performance. Both feature vectors are fed into a cross-attention fusion layer. The cross-attention fusion layer determines the contribution weight of each feature dimension to the comprehensive quality assessment by calculating the correlation matrix between geometric features and pressure features, ultimately generating a 128-dimensional comprehensive quality vector.

[0043] The dynamic adaptive control module inputs the comprehensive quality vector into a digital twin model based on finite element analysis. This model establishes a multiphysics coupled simulation environment based on the material properties and structural parameters of the nitrogen spring, simulating the stress distribution and fatigue life cycle of the spring under different working pressures. The simulation results and the comprehensive quality vector are jointly input into a fuzzy decision system, which uses a three-level if-then rule base to define the product level membership function. For example, for products with roundness deviations within the tolerance critical range, the system will combine the pressure decay rate index for comprehensive evaluation, avoiding misjudgment based on a single parameter. The particle swarm optimizer receives throughput data from the production line monitoring system in real time, dynamically calculating the sorting priority sequence with the objective function of minimizing equipment idle time and maximizing the qualified product output rate. When the backlog of workpieces in the buffer area exceeds the threshold, the system automatically increases the processing weight of higher-level products and generates sorting instructions with millisecond-level timestamps.

[0044] After receiving sorting control commands, the collaborative robot motion planner in the sorting execution module analyzes the position and orientation of the target workpiece in the coordinate system and calculates the robot arm's trajectory using a fast random exploration tree algorithm. The trajectory planning considers the kinematic constraints of each robot joint and the layout of on-site equipment, generating a collision-free seven-segment S-shaped velocity curve. During the gripping execution phase, a six-dimensional force sensor collects three-dimensional force and torque data at a specific frequency, and the impedance controller treats the robot's end effector as a spring and damping system. When the force sensor detects a deviation of the gripping force from the set value, the controller adjusts the stiffness parameters and damping coefficient of the end effector in real time based on the impedance model, achieving compliant control of the contact force. After completing the sorting action, the module packages the operation log, which includes the gripping status, actual time consumption, and six-dimensional force data sequence, into a feedback data package and sends it to the dynamic adaptive control module via an industrial IoT gateway.

[0045] The system evolution module continuously optimizes the model using a federated learning framework. Each sorting terminal periodically extracts incremental data from the feature processing module and, combined with historical operation logs from the sorting execution module, trains a lightweight neural network locally. Differential privacy technology is used to add noise during training, and the generated encrypted model parameters are uploaded to the central server via a secure transport layer protocol. The server performs a weighted average aggregation of model parameters from multiple terminals, and the updated global model parameters are then redistributed to the feature processing modules of each terminal. Simultaneously, a long short-term memory network continuously analyzes equipment operating data, including time-series data such as harmonic components of joint motor current, characteristic values ​​of reducer vibration spectrum, and zero-point drift of force sensors. When a trend of equipment performance degradation is identified, the system calculates the remaining service life based on a pre-established life prediction model and sends a warning to maintenance personnel before reaching the maintenance threshold.

[0046] During actual system operation, when encountering situations with strong surface reflection on the workpiece, structured light scanning may result in missing point cloud data. In this case, the feature processing module initiates a data quality assessment process, detecting outlier data points using a local outlier factor algorithm. When the point cloud integrity index is found to be below a threshold, the system automatically sends a re-inspection command to the data acquisition module, triggering a secondary scanning process. For pressure test data, the system performs compensation calculations by combining ambient temperature sensor readings to eliminate the impact of thermal expansion and contraction on pressure readings. This multimodal data cross-validation mechanism effectively improves the system's robustness in complex industrial environments.

[0047] The system interacts with the Manufacturing Execution System (MES) to receive production order information and provide feedback on sorting statistics. When the production line switches to different specifications of nitrogen springs, operators can load the corresponding process parameter package through the human-machine interface. The system automatically adjusts scanning parameters and gripping force settings according to the new product specifications, shortening production line changeover time. All operational data is recorded in a distributed database, providing complete data chain support for production quality traceability.

[0048] The adaptive Kalman filter algorithm is used to estimate the system state. It can adaptively adjust the filter parameters to cope with dynamic changes or changes in noise characteristics of the system. In this invention, the algorithm is used by the data acquisition module to perform sensor drift compensation on the processed geometric point cloud data and pressure time series data. By correcting the measurement errors caused by environmental changes or equipment aging in real time, the accuracy and reliability of the acquired data are guaranteed, providing high-quality input for subsequent processing.

[0049] The Local Outlier Factor (LOF) algorithm is used to detect outliers in a dataset. It identifies outliers by calculating the density ratio of each data point to its neighbors. In this invention, the LEF algorithm is used by the feature processing module to monitor the data distribution of standardized data packets. When the calculated LEF exceeds a set threshold, it is determined to be an outlier data point and a re-detection instruction is generated, triggering the data acquisition module to rescan and detect, thereby forming a closed-loop mechanism for data quality control.

[0050] Particle swarm optimization (PSO) is an optimization algorithm based on swarm intelligence. It simulates the social behavior of flocks of birds or schools of fish and finds the optimal solution by iteratively updating the position and velocity of particles. In this invention, the PSO algorithm is used by a dynamic adaptive control module to dynamically calculate the optimal priority ranking of sorting instructions based on product grade classification results and real-time production line throughput data. The goal is to minimize production line idle time and maximize the output of qualified products per unit time, thereby achieving real-time adjustment of the sorting strategy.

[0051] Impedance control algorithm is a robot control strategy that adjusts the interaction force between the robot and the environment by simulating a spring and damping system to achieve compliant operation. In this invention, the sorting execution module uses this algorithm to dynamically adjust the stiffness and damping parameters of the collaborative robot's end effector based on the force control data collected in real time by the six-dimensional force sensor. This ensures compliant grasping during the gripping of the nitrogen spring, avoids product damage, and improves sorting accuracy.

[0052] The Fast Random Exploration Tree Algorithm is an algorithm for path planning that efficiently calculates the collision-free motion trajectory of a robot through random sampling and tree structure expansion. In this invention, the algorithm is used by the sorting execution module to plan the collision-free motion trajectory of the collaborative robot from its current position to the target position, taking into account the robot's joint kinematic constraints and the layout of on-site equipment, to generate a smooth and safe path to optimize sorting efficiency.

[0053] A digital twin model is a virtual model that mirrors a physical entity through simulation technology, used to predict and optimize system behavior. In this invention, the model is used by a dynamic adaptive control module based on finite element analysis, using a comprehensive mass vector as the initial condition, to simulate the stress distribution and fatigue life of a nitrogen spring under various load conditions, and outputs performance simulation results to assist in product classification and decision-making.

[0054] The fuzzy decision system is a decision-making method based on fuzzy logic. It uses membership functions and rule bases to handle uncertainty and fuzziness, and performs classification or decision-making. In this invention, the system is used by a dynamic adaptive control module to receive performance simulation results and comprehensive quality vectors. It uses a multi-level if-then rule base to set membership functions, classifies nitrogen springs into products, and outputs product classification results to support sorting priority adjustment.

[0055] A spatial encoder is a neural network structure commonly used to process spatial data such as point clouds, extracting global and local features through a self-attention mechanism. In this invention, the encoder is used by the feature processing module to receive voxelized point cloud data, extract global geometric features and local shape features such as roundness and surface defects of the nitrogen spring, and output geometric feature vectors to prepare for multimodal fusion.

[0056] A time-series encoder is a neural network, such as a recurrent neural network, used to process time-series data and extract dynamic response features and trends. In this invention, the encoder is used by the feature processing module to receive normalized pressure sequence data, extract dynamic response features and trends of pressure data such as attenuation gradients, and output a pressure feature vector to characterize the sealing performance of the nitrogen spring.

[0057] The cross-attention fusion layer is a multimodal fusion technique that performs weighted fusion by calculating the correlation weights between different feature modes. In this invention, this layer is used by the feature processing module to receive geometric feature vectors and pressure feature vectors, calculate the correlation weights between the two feature modes, and perform weighted summation based on the weights to generate a low-dimensional comprehensive quality vector that comprehensively characterizes the dimensional accuracy and sealing performance of the nitrogen spring.

[0058] Long Short-Term Memory (LSTM) networks are a special type of recurrent neural network capable of learning long-term dependencies and used for temporal pattern recognition and prediction. In this invention, the system evolution module uses this network to analyze equipment operation data in historical operation logs, such as joint motor current fluctuations and reducer vibration spectra, to identify equipment degradation trends and calculate remaining service life. Before reaching a preset threshold, it sends early warning information to the maintenance terminal to achieve predictive maintenance.

[0059] Federated learning framework is a distributed machine learning method that allows models to be trained on local devices, with only encrypted parameters uploaded for aggregation, thus protecting data privacy. In this invention, the system evolution module uses this framework to train local models at each sorting terminal using incremental data and historical operation logs, generates encrypted model parameters, uploads them to the central server for secure aggregation, updates global model parameters, and distributes them to the feature processing module, thereby achieving continuous optimization of the system's recognition capabilities.

Claims

1. A nitrogen spring automatic sorting system, characterized in that, It includes a data acquisition module, a feature processing module, a dynamic adaptive control module, a sorting execution module, and a system evolution module; The data acquisition module is configured to acquire geometric point cloud data of the nitrogen spring through structured light 3D scanning and pressure time series data through a high-precision pressure sensor; perform outlier removal and downsampling processing on the geometric point cloud data, perform moving average filtering on the pressure time series data, and use an adaptive Kalman filter algorithm to perform sensor drift compensation on the processed geometric point cloud data and pressure time series data; package the drift-compensated geometric point cloud data and pressure time series data into a standardized data package and send it to the feature processing module; The feature processing module is configured to receive the standardized data packet; perform voxelization processing on the geometric point cloud data in the data packet; and input the voxelized point cloud data into the spatial encoder to extract geometric features. The stress time-series data in the data packet is normalized, and the normalized stress sequence data is input into the time encoder to extract stress dynamic features. The geometric features and stress dynamic features are input into the cross-attention fusion layer to generate a comprehensive quality vector, and the comprehensive quality vector is sent to the dynamic adaptive control module. The feature processing module also uses the local outlier factor algorithm to monitor the data distribution of the normalized data packet. When an abnormal data distribution is detected, a re-detection command is sent to the data acquisition module. The dynamic adaptive control module is configured to receive the comprehensive quality vector; input the comprehensive quality vector into a digital twin model based on finite element analysis for performance simulation to obtain simulation results; and input the simulation results and the comprehensive quality vector together into a fuzzy decision system for product classification to obtain product grades. The system receives real-time throughput data from the production line and uses a particle swarm optimization algorithm, taking the product grade and real-time throughput data as input, to dynamically adjust the sorting priority and generate a sorting control instruction with a timestamp; the sorting control instruction is then sent to the sorting execution module. The sorting execution module is configured to receive the sorting control command; parse the sorting control command to obtain the target sorting destination, and plan the collision-free motion trajectory of the collaborative robot; during the process of the collaborative robot grasping the nitrogen spring along the motion trajectory, force control data is collected in real time through a six-dimensional force sensor, and the grasping posture is dynamically adjusted according to the force control data using an impedance control algorithm; after the sorting action is completed, the grasping success status and the force control data are fed back to the dynamic adaptive control module. The system evolution module is configured to obtain incremental data required for model training from the feature processing module and historical operation logs from the sorting execution module; employing a federated learning framework, it uses the incremental data and historical operation logs to perform local model training at each sorting terminal, and uploads the encrypted model parameters generated by the training to the central server for aggregation to generate global model parameters; the global model parameters are distributed to the feature processing module to update the identification model; simultaneously, it analyzes the equipment operation data in the historical operation logs through a long short-term memory network to predict the maintenance cycle and sends early warning information to the maintenance terminal.

2. The nitrogen spring automatic sorting system according to claim 1, characterized in that, The data acquisition module includes a structured light 3D scanning unit and a pressure sensing unit; The structured light 3D scanning unit projects a grating pattern onto the surface of the nitrogen spring; the industrial camera captures the modulated deformed grating image on the surface of the nitrogen spring. Phase calculation is performed on the deformed grating image, and the three-dimensional coordinates of the workpiece surface points are calculated based on the triangulation principle to generate three-dimensional point cloud data representing the outer contour of the workpiece. The pressure sensing unit is connected to the nitrogen spring inflation valve to sample pressure data in a sealed cavity, converting the collected analog voltage signal into discrete values ​​to form time series data of the pressure decay curve.

3. The nitrogen spring automatic sorting system according to claim 2, characterized in that, The feature processing module includes a spatial encoder, a temporal encoder, and a cross-attention fusion layer; The spatial encoder receives voxelized point cloud data, extracts global geometric features and local shape features of the point cloud data through a self-attention mechanism, and outputs a geometric feature vector. The timing encoder receives normalized pressure sequence data, extracts dynamic response features and changing trends of the pressure data through a recurrent neural network structure, and outputs a pressure feature vector. The cross-attention fusion layer receives the geometric feature vector and the stress feature vector, calculates the correlation weight between the two feature modes, and performs a weighted summation of the geometric feature vector and the stress feature vector based on the correlation weight to generate a low-dimensional comprehensive quality vector.

4. The nitrogen spring automatic sorting system according to claim 3, characterized in that, The dynamic adaptive control module includes a digital twin model, a fuzzy decision system, and a particle swarm optimizer; The digital twin model uses the integrated mass vector as the initial condition and, based on the gas state equation and mechanical performance equation, simulates the stress distribution and fatigue life of a nitrogen spring under various load conditions, and outputs the performance simulation results. The fuzzy decision system receives the performance simulation results and the comprehensive quality vector, sets the membership function using a multi-level if-then rule base, classifies the nitrogen springs, and outputs the product grade classification results. The particle swarm optimizer uses production line throughput efficiency as the objective function and product grade classification results and real-time production line throughput data as inputs to dynamically calculate the optimal priority ranking of sorting instructions.

5. The nitrogen spring automatic sorting system according to claim 4, characterized in that, The sorting execution module includes a collaborative robot motion planner and an impedance controller; The collaborative robot motion planner receives sorting control instructions, parses the coordinates of the target workpiece and the sorting destination, and uses a fast random exploration tree algorithm to calculate the collision-free motion trajectory of the robotic arm from the current position to the target position. During the movement of the robotic arm along its trajectory, the impedance controller collects three-dimensional force and torque data in real time through a six-dimensional force sensor, simulating the robot's end effector as a spring and damping system, and dynamically adjusts the stiffness and damping parameters of the end effector based on the real-time force and torque data collected.

6. The nitrogen spring automatic sorting system according to claim 5, characterized in that, The system evolution module includes a federated learning framework and a long short-term memory network; The federated learning framework obtains incremental model data from the feature processing module and historical operation logs from the sorting execution module; it uses the stochastic gradient descent algorithm to train the model locally at each sorting terminal using the incremental data and historical operation logs; it uploads the encrypted model parameters generated by the training to the central server; the central server uses a weighted average algorithm to securely aggregate the encrypted model parameters to generate global model parameters; and it distributes the global model parameters to the feature processing module. The long short-term memory network collects current fluctuation data of the joint motors of the collaborative robot, vibration spectrum data of the reducer, and zero-point drift data of the force sensor; performs time-series pattern recognition on the collected equipment operation data; when the degradation trend of the equipment is identified, it calculates the remaining service life of each component; and sends an early warning message to the maintenance terminal before the remaining service life reaches a preset threshold.

7. The nitrogen spring automatic sorting system according to claim 6, characterized in that, When the feature processing module detects data anomalies, it performs the following steps: Calculate the local outlier factor of the newly input standardized data packet and the historical normal sample set in the multidimensional feature space; when the local outlier factor exceeds the set threshold, determine the current data as an abnormal data point; generate a re-inspection instruction including the workpiece identifier corresponding to the abnormal data point; send the re-inspection instruction to the data acquisition module through the reverse data stream; trigger the data acquisition module to re-execute the structured light 3D scan and pressure data acquisition on the current workpiece.

8. The nitrogen spring automatic sorting system according to claim 7, characterized in that, The dynamic adaptive control module is configured to adjust the sorting strategy as follows: The system monitors the working status of sorting stations and the length of the buffer queue in real time; it aims to minimize production line idle time and maximize the output of qualified products per unit time; it uses product grade classification results and real-time production line throughput data as inputs to dynamically calculate the optimal sorting priority through particle swarm optimization algorithm; and it updates the classification threshold and instruction sorting in the sorting decision tree based on the calculation results.

9. The nitrogen spring automatic sorting system according to claim 8, characterized in that, After the sorting execution module completes the sorting, it is configured as follows: The successful capture status flag, actual capture time, complete six-dimensional force control data sequence, and impedance control parameter adjustment record are packaged into a feedback data packet; the feedback data packet is then sent to the dynamic adaptive control module. The dynamic adaptive control module receives feedback data packets, adjusts records based on force control data and grasping parameters in the feedback data packets, and updates the grasping parameter knowledge base; it then optimizes the force control parameter settings in subsequent sorting control instructions based on the updated grasping parameter knowledge base.

10. The nitrogen spring automatic sorting system according to claim 9, characterized in that, The system evolution module is configured to perform predictive maintenance as follows: Collect current fluctuation data of joint motors, vibration spectrum data of reducers, and zero-point drift data of force sensors of collaborative robots; input the collected equipment operation data into a long short-term memory network for time-series pattern recognition; When the timing pattern recognition results show a trend of equipment degradation, the remaining service life of each component is calculated; when the remaining service life reaches the preset maintenance threshold, an early warning message is sent to the maintenance terminal.

Images