Automatic production line real-time quality control system based on multi-sensor fusion

Abstract

The invention relates to the field of industrial automatic quality control, and discloses an automatic production line real-time quality control system based on multi-sensor fusion. In the prior art, the problem of incomplete perception-decision-control closed loop mainly exists, the system can only realize quality detection and alarm generally, and parameter adjustment still needs manual intervention. According to the method, standardized data are obtained through the multi-source heterogeneous data sensing module; an intelligent fusion module is adopted to generate a quality state vector; the advanced prediction and intelligent decision-making of the quality trend are realized based on the prediction decision-making module; real-time closed-loop regulation and control of process parameters are completed through the precise execution module; and continuous optimization of the system is realized by means of a self-evolution optimization module. According to the invention, a complete closed loop from quality perception to control execution is established, the accuracy and automation level of quality control are obviously improved, and the reject ratio of products is effectively reduced.

Description

Technical Field

[0001] This invention relates to the field of industrial automation quality control, and discloses a real-time quality control system for automated production lines based on multi-sensor fusion. Background Technology

[0002] In current industrial automated production environments, multi-sensor technology has been widely applied in quality monitoring. However, existing technologies still have significant shortcomings in achieving truly real-time quality control. First, most systems can only collect and make simple judgments on quality parameters, lacking the ability to deeply integrate and intelligently analyze multi-source data. This leads to quality control decisions relying on single sensor data or simple threshold judgments, making it difficult to cope with quality fluctuations under complex operating conditions.

[0003] In existing quality control systems, the sensing, decision-making, and control processes are often independent, creating information transmission barriers. Quality inspection results cannot directly drive real-time parameter adjustments in production equipment, still requiring intervention based on human experience. This lag severely impacts the timeliness and accuracy of quality control. Furthermore, the lack of an effective closed-loop feedback mechanism between system modules makes true real-time control difficult to achieve.

[0004] Traditional quality control systems generally suffer from rigid models and poor adaptability. These systems cannot autonomously adjust control strategies based on changes in production conditions, nor do they possess the ability to learn and evolve from historical data. In flexible production scenarios with multiple product types and small batches, this rigid system architecture severely restricts the continuous improvement of quality control accuracy. Furthermore, the data silos formed by different production lines hinder the collaborative optimization of overall manufacturing performance. Therefore, there is an urgent need for a real-time quality control system for automated production lines based on multi-sensor fusion. Summary of the Invention

[0005] The purpose of this invention is to provide a real-time quality control system for automated production lines based on multi-sensor fusion, so as to solve the problems in the background art.

[0006] The objective of this invention can be achieved through the following technical solutions: A real-time quality control system for an automated production line based on multi-sensor fusion includes the following modules: Data sensing and acquisition module: Relying on a multi-source heterogeneous sensor array, it performs synchronous sensing and adaptive acquisition of the overall status of the production line and the processing quality of the workpiece, and outputs a standardized synchronous data stream with spatiotemporal tags; Data fusion and feature extraction module: Introduces the standardized synchronous data stream, uses a dynamic adaptive deep learning fusion model to perform multi-source information confidence assessment and deep feature extraction, and generates a state vector representing the overall quality of the product; Prediction-Decision-Control Command Generation Module: Based on the state vector representing the overall quality of the product, it completes the advanced simulation of quality trends and intelligent decision-making through a prediction-correction mechanism driven by digital twins, and outputs a precise set of parameter adjustment commands; Precision Execution and Status Feedback Module: Receives the precise parameter adjustment instruction set, manipulates the high-precision actuator to complete real-time control of process parameters and action planning, and uses integrated sensors to synchronously verify the execution effect, forming an execution status feedback signal; Source analysis and self-evolution optimization module: It gathers the state vector representing the overall quality of the product, the precise parameter adjustment instruction set, and the execution status feedback signal, and performs root cause analysis and federated learning optimization based on the spatiotemporal correlation database to achieve continuous self-evolution of the system's full-link model.

[0007] Preferably, the data sensing and acquisition module is configured as follows: A multi-source heterogeneous sensor array is used to simultaneously monitor the status of the entire production line and the quality of workpiece processing to obtain raw multi-source data streams; The original multi-source data stream is processed using an adaptive sampling strategy to dynamically adjust the acquisition frequency and output optimized acquisition data. The optimized collected data is processed using standardized methods to unify the format and filter noise, resulting in a standardized dataset. The standardized data set is bound to timestamps and spatial location information using spatiotemporal calibration technology to generate the standardized synchronous data stream with spatiotemporal tags.

[0008] Preferably, the data fusion and feature extraction module is configured as follows: The standardized synchronous data stream with spatiotemporal labels is used to perform multi-source information confidence assessment using a dynamic adaptive deep learning fusion model to obtain the weight allocation of each data source. A weighted fusion algorithm is used to process the weight allocation of each data source, complete the multi-source data fusion calculation, and output highly reliable fused data. A deep neural network is used to extract spatiotemporal features from the high-reliability fused data to obtain a deep quality feature representation. The deep quality feature representation is processed by a feature integration method to complete multi-dimensional feature fusion and generate the state vector that represents the overall quality of the product.

[0009] Preferably, the dynamic adaptive deep learning fusion model is configured as follows: The attention mechanism is used to analyze the standardized synchronous data stream with spatiotemporal labels, to complete the importance assessment of each sensor's data, and to output dynamic weight coefficients. The dynamic weight coefficients are processed using a weighted average method, and multi-source data fusion is performed to generate a weighted fusion result. Deep feature learning is performed on the weighted fusion result using a convolutional neural network to obtain a high-quality feature representation. The parameters of the dynamic adaptive deep learning fusion model are optimized using an online learning mechanism to achieve continuous improvement of the fusion model.

[0010] Preferably, the prediction-decision-control instruction generation module is configured as follows: By using the state vector representing the overall quality of the product, a digital twin model is employed to predict quality trends and obtain future quality state evolution data. The risk assessment method is used to analyze the future quality state evolution data, complete the quality deviation assessment, and output the key adjustment parameter identification results. The reinforcement learning algorithm processes the identification results of the key adjustment parameters, optimizes the control strategy, and generates the optimal control scheme. The instruction conversion module converts the optimal control scheme into machine-executable instructions, forming the precise parameter adjustment instruction set.

[0011] Preferably, the digital twin-driven prediction-correction mechanism is configured as follows: By using digital twin models based on real-time production line operation data, the production process can be simulated to obtain the virtual production line operation status; The prediction algorithm analyzes the operating status of the virtual production line, performs quality parameter evolution simulation, and outputs quality trend prediction results. The correction mechanism processes the quality trend prediction results, adjusts the control strategy, and generates optimized control parameters. By updating the parameters of the digital twin model with real-time feedback data, a predictive model with continuously improving accuracy can be achieved.

[0012] Preferably, the precise execution and status feedback module is configured as follows: Based on the precise parameter adjustment instruction set, a high-precision actuator is used to complete the process parameter adjustment and obtain the parameter adjustment result; Embedded sensor networks monitor the operating status of actuators, acquire data in real time, and output execution process data. The verification algorithm analyzes the execution process data, completes the instruction execution effect evaluation, and generates execution status verification results. The signal processing module processes the execution status verification result, generates a feedback signal, and forms the execution status feedback signal.

[0013] Preferably, the source analysis and self-evolutionary optimization module is configured as follows: Based on the state vector representing the overall quality of the product, the precise parameter adjustment instruction set, and the execution status feedback signal, data integration technology is used to complete the full-link data integration and obtain a complete production file. The data mining algorithm processes the complete production file, performs root cause analysis of quality defects, and outputs the root cause location results. The knowledge update mechanism, combined with the root cause location results, optimizes the quality control strategy and generates improved quality control rules. The federated learning framework implements multi-node collaborative training to optimize the system model and obtain a continuously evolving end-to-end model.

[0014] Preferably, the federated learning optimization is configured as follows: Local model updates are completed by using local production data for model training, and local model parameters are obtained. The local model parameters are processed using secure encryption technology, privacy protection is implemented, and encrypted model parameters are output. The federated averaging algorithm integrates the encrypted model parameters to complete global model aggregation and generate an optimized global model. The global model is optimized using a model distribution mechanism, and local model updates are performed to obtain a collaboratively optimized system model.

[0015] Preferably, the root cause analysis is configured as follows: Based on the characteristics of quality defects, a spatiotemporal correlation database is used to complete historical data queries to obtain a set of relevant process data. Anomaly detection algorithms are used to analyze the relevant process data set to identify abnormal time periods and output abnormal time windows. The abnormal time window is processed using correlation analysis to complete the mining of abnormal parameters and generate key abnormal parameters. Based on the causal reasoning model, the key abnormal parameters are analyzed to locate the root cause and generate a quality defect cause analysis report.

[0016] The beneficial effects of this invention are: This invention, through the synergistic effect of multi-source data fusion and intelligent decision-making mechanisms, enables the system to proactively predict quality trends and implement corrective measures in advance, effectively overcoming the limitations of traditional systems that can only issue alarms after the fact, and significantly improving the timeliness and accuracy of quality control.

[0017] This invention, through dynamically adjusting data fusion strategies and continuously optimizing control models, can proactively adapt to changes in production conditions and the impact of equipment aging. The distributed optimization architecture based on federated learning further breaks down data silos, enabling the system to continuously improve quality control capabilities while protecting the data privacy of each production line, greatly enhancing the system's adaptability and stability in complex and ever-changing production environments.

[0018] This invention, by constructing spatiotemporally correlated production archives, enables the system to quickly pinpoint the root causes of quality problems and feed the analysis results back to the optimization process of control strategies. This leap from single-point control to system-level optimization not only improves the efficiency of handling quality problems but also promotes continuous improvement of production processes as a whole, laying a solid foundation for establishing a preventative quality control system. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the structure of a real-time quality control system for an automated production line based on multi-sensor fusion, according to the present invention. Detailed Implementation

[0020] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0021] Example: Figure 1 As shown, an automated production line real-time quality control system based on multi-sensor fusion includes the following modules: Data sensing and acquisition module: Relying on a multi-source heterogeneous sensor array, it performs synchronous sensing and adaptive acquisition of the overall status of the production line and the processing quality of the workpiece, and outputs a standardized synchronous data stream with spatiotemporal tags; Data fusion and feature extraction module: Introduces the standardized synchronous data stream, uses a dynamic adaptive deep learning fusion model to perform multi-source information confidence assessment and deep feature extraction, and generates a state vector representing the overall quality of the product; Prediction-Decision-Control Command Generation Module: Based on the state vector representing the overall quality of the product, it completes the advanced simulation of quality trends and intelligent decision-making through a prediction-correction mechanism driven by digital twins, and outputs a precise set of parameter adjustment commands; Precision Execution and Status Feedback Module: Receives the precise parameter adjustment instruction set, manipulates the high-precision actuator to complete real-time control of process parameters and action planning, and uses integrated sensors to synchronously verify the execution effect, forming an execution status feedback signal; Source analysis and self-evolution optimization module: It gathers the state vector representing the overall quality of the product, the precise parameter adjustment instruction set, and the execution status feedback signal, and performs root cause analysis and federated learning optimization based on the spatiotemporal correlation database to achieve continuous self-evolution of the system's full-link model.

[0022] In this embodiment, relying on a multi-source heterogeneous sensor array, synchronous sensing and adaptive acquisition of the overall status of the production line and the workpiece processing quality are performed, and a standardized synchronous data stream with spatiotemporal tags is output. Specific implementation methods include: The multi-source heterogeneous sensor array is collaboratively deployed according to the production line's process layout and quality monitoring requirements, constructing a comprehensive sensing network. Visual sensors, physical quantity sensors, and acoustic sensors are deployed at key process nodes. These sensors acquire data in synchronous triggering or asynchronous streaming modes, ensuring multi-dimensional perception of the workpiece's state within the same production cycle. Visual sensors capture the workpiece's surface texture, geometric dimensions, and assembly relationships; physical quantity sensors accurately measure process parameters such as displacement, temperature, and pressure; and acoustic sensors monitor abnormal vibrations and acoustic signals generated by internal defects during processing. To achieve unified management of the data flow, each sensor data packet is assigned a globally unique serial number issued by the production line's main control system, serving as the basis for subsequent data association and spatiotemporal calibration.

[0023] To optimize system resources and capture critical data, this embodiment implements an adaptive acquisition strategy based on information entropy. The system calculates the information entropy H(X) of each sensor data stream in real time to assess the information content and uncertainty of the data. The calculation formula is as follows: Where X represents sensor data, For data values Probability estimate of occurrence. When When the sampling frequency exceeds a preset threshold, it indicates a significant fluctuation or abnormality in the production process, and the system automatically increases the sampling frequency of the sensor. Upgrade to a higher sampling rate to capture finer dynamic processes. Conversely, maintain or reduce the sampling frequency to save computing power and storage space. Immediately after data acquisition, preliminary preprocessing is performed, including processing the raw signal. Perform moving average filtering to suppress high-frequency noise: Where N is the sliding window size. This effectively improves the signal-to-noise ratio of the original data.

[0024] Due to the heterogeneity of sensors, the collected data exhibits significant differences in units, numerical ranges, and data types. This embodiment employs the Z-score normalization method to normalize the numerical data, converting it into a distribution with a mean of 0 and a standard deviation of 1. For data points from the j-th sensor... Its standardized value The calculation is as follows: ,in, and These are the mean and standard deviation of the sensor's historical data under stable operating conditions. For non-numerical data, the format is standardized by parsing all images into grayscale images or three-channel RGB pixel matrices of uniform resolution, and basic statistical features, such as the mean pixel intensity, are extracted. and standard deviation , serving as the primary descriptor for subsequent fusion.

[0025] After generating standardized data, the core step is to perform precise spatiotemporal calibration. Time calibration is based on the production line's master clock, and each data packet is stamped with a high-precision timestamp. This accuracy is down to the millisecond level. Spatial calibration, on the other hand, maps the global sequence number of the data packet to a specific physical location, i.e., a workstation identifier. And the spatial coordinates (x, y, z) of this workstation. Ultimately, each data unit is encapsulated into a structured data object. : These Arranged chronologically and spatially, these data streams collectively constitute the "standardized synchronous data stream with spatiotemporal tags." This data stream is pushed to the next module in real time via high-speed industrial Ethernet, providing a high-quality, highly consistent data foundation for subsequent in-depth analysis and intelligent decision-making.

[0026] In this embodiment, the standardized synchronous data stream is introduced, and a dynamic adaptive deep learning fusion model is used to perform multi-source information confidence assessment and deep feature extraction to generate a state vector representing the overall quality of the product. Specific implementation methods include: The system first performs a confidence assessment on the input standardized synchronous data stream, which is a prerequisite for achieving high-quality fusion. This embodiment employs an attention-based confidence assessment network, which uses the previous system state and current sensor readings as input to dynamically calculate the confidence weights of each data source. For the data from the i-th sensor at time t... Its confidence weight Calculated using the following formula: , ,in, This represents the hidden state of the system at the previous time step, where W and U are trainable weight matrices, b is the bias term, and v is the attention vector. This mechanism enables the system to automatically adjust its dependence on different sensors based on the context, reducing the weight of a sensor when it malfunctions or experiences increased noise, thereby enhancing the system's robustness.

[0027] After obtaining the confidence weights of each data source, the system performs weighted fusion. For a feature set from N sensors... The fused feature representation The calculation is as follows: This weighted fusion method not only considers the raw values ​​of each sensor data but also incorporates their reliability assessment under the current environment. To further enhance the fusion effect, this embodiment also introduces cross-sensor correlation modeling, using a graph neural network to capture the potential correlations between different sensor readings and update the confidence weights. , Where X is the sensor reading matrix and A is the attention matrix between sensors. These are trainable parameters.

[0028] Obtain fusion features Subsequently, the system performs deep feature extraction using a deep convolutional neural network. This network employs a residual connection structure to avoid the gradient vanishing problem, and its basic unit is calculated as follows: ,in This represents the convolution operation. It is an activation function. This represents the feature representation of layer l. The network progressively extracts features from low to high levels through multiple convolutional and pooling layers, capturing local correlations and spatial hierarchical structures in the data. The network employs a multi-scale feature pyramid structure, extracting features under different receptive fields and then fusing them to capture quality features of different granularities.

[0029] To capture the dynamic evolution of quality characteristics, the system introduces a gated recurrent unit (GRU) to model the feature sequence. The GRU update mechanism is as follows: , , , ,in It's an update gate. It's a door reset. This represents the hidden state at the current moment. By modeling temporal features, the system can understand the evolution trend of quality parameters, not just the instantaneous state. Finally, the final hidden state of the GRU is projected onto a fixed-dimensional quality state vector through a fully connected layer: ,in It is a state vector that characterizes the overall quality of a product, containing comprehensive quality information of the product at the current moment.

[0030] To adapt to dynamic changes in the production environment, this embodiment also incorporates an online adaptive mechanism for the model. The system continuously monitors the performance metrics of the fused model, and initiates an online fine-tuning process when performance degradation or significant environmental changes are detected. A sliding window approach is used to retain data from the most recent period, and the loss function is calculated based on this data. ,in It is a model prediction. These are actual observed values. These are model parameters. This is the regularization coefficient. By making small adjustments to the model parameters through the backpropagation algorithm, the fusion model can quickly adapt to environmental changes while avoiding catastrophic forgetting. This mechanism ensures that the system maintains high-precision fusion capabilities throughout long-term operation.

[0031] In this embodiment, based on the state vector representing the overall quality of the product, a digital twin-driven prediction-correction mechanism is used to perform advanced simulation and intelligent decision-making on quality trends, outputting a precise set of parameter adjustment instructions. Specific implementation methods include: The system constructs a digital twin model based on the fusion of physical mechanisms and data-driven approaches. This model accurately characterizes the dynamic characteristics of the production process through deep neural networks. Let the mass state vector at time k be... It includes multiple quality dimensions such as dimensional accuracy, surface quality, and geometric tolerances; the process parameter vector is... This includes adjustable parameters such as machining speed, cutting parameters, and temperature control. The digital twin model describes the quality evolution through state-space equations: ,in Let be the state transition function represented by the deep neural network. For network parameters, This is process noise. Based on the current quality status. With a preset control sequence, the digital twin model performs multi-step advance prediction to generate the quality state trajectory for the next N sampling periods. This prediction process fully considers the dynamic response of equipment, changes in material properties, and environmental disturbances, providing an accurate basis for quality trend analysis.

[0032] To improve prediction accuracy, the system employs an adaptive Kalman filter algorithm to correct the prediction results in real time. The state estimation error covariance matrix is ​​defined. Observation matrix Then the Kalman gain The calculation is as follows: ,in The time-varying observation noise covariance matrix is ​​estimated using an innovative sequence adaptive estimation method. This is achieved when actual quality observation data is obtained. Then, the predicted state is optimally corrected: .

[0033] Simultaneously, the state estimation error covariance matrix is ​​updated to ensure the stability and convergence of the filter. This adaptive correction mechanism effectively suppresses the effects of model mismatch and measurement noise, significantly improving the accuracy of state estimation.

[0034] Based on the corrected quality prediction trajectory, the system constructs a multi-objective optimization problem, comprehensively considering multiple performance indicators such as quality accuracy, control cost, and production efficiency. The finite-time cost function is defined as follows: ,in To optimize the variables, Q and R are weight matrices. For quality target value, These are the nominal process parameters. The optimization problem must satisfy actuator amplitude constraints, rate of change constraints, and quality index constraints. The interior point method is used to solve this constrained optimization problem to obtain the optimal control sequence. .

[0035] The first element of the optimal control sequence As the execution command for the current control cycle, it is distributed to each actuator via a real-time industrial network. Simultaneously, the system establishes a performance evaluation framework, quantifying the control effect through a loss function that includes prediction error and control costs. Based on the gradient of the loss function, the parameters of the digital twin model are updated using stochastic gradient descent. ,in To enable adaptive learning rates, the system also incorporates an experience playback mechanism. Historical operational data is stored in a circular buffer, and periodic sampling is used for model retraining. This online learning mechanism allows the system to continuously adapt to changes in production processes, constantly improving control accuracy and robustness, forming a self-perfecting quality control closed loop.

[0036] In this embodiment, the precise parameter adjustment instruction set is received, and a high-precision actuator is manipulated to complete real-time control of process parameters and motion planning. An integrated sensor is used to synchronously verify the execution effect, generating an execution status feedback signal. Specific implementation methods include: This system employs a fully digital multi-axis linkage servo control system, establishing a precise mathematical model of the actuator. Let the state vector of the servo motor be... , representing rotor angular displacement, angular velocity, and q-axis current, respectively. The state-space equation of the servo system is described as: Where J is the moment of inertia and B is the damping coefficient. The torque constant is Let be the back electromotive force constant, and R and L be the resistance and inductance, respectively. The system receives a set of parameter adjustment instructions. The instruction parsing module converts the commands into motion trajectory instructions for each execution axis. A seven-segment S-curve acceleration / deceleration algorithm is used for motion planning to ensure the smoothness and accuracy of the motion process. The trajectory planning algorithm is calculated as follows: Angular velocity command Generated by an S-shaped curve, it effectively suppresses mechanical shock and vibration.

[0037] To address the system's nonlinearity and time-varying parameter characteristics, an adaptive fuzzy PID controller was designed. The controller's output consists of three parts: ,in This refers to the position error. Control parameters. The fuzzy rules are adjusted online through the fuzzy inference system and take the following form: .

[0038] Meanwhile, the system introduces a feedforward compensation term to suppress external disturbances: Total control quantity This effectively improves the system's tracking accuracy and anti-interference capability.

[0039] Multiple types of sensors, including absolute linear scales, strain gauge force sensors, and infrared temperature sensors, are deployed at key locations on the actuator. Sensor data is fused using a Kalman filter. ,in Let H be the observation vector, H be the observation matrix, and Kalman gain be... Real-time calculation and updates. Execution effect verification is achieved by calculating the error norm between the actual trajectory and the commanded trajectory: ,when When the threshold is exceeded, the system automatically triggers a replanning mechanism to ensure processing accuracy.

[0040] The execution status feedback signal uses a structured data format, including a timestamp t, actuator identifier ID, and actual position. Actual speed Control torque And the error flag. The data structure for the feedback signal is defined as follows: .

[0041] Transmitted to the upper-level control system via real-time industrial Ethernet protocol, with a transmission cycle of ≤1ms and a packet loss rate of <0.001%. The system simultaneously calculates performance indicators: It is used to evaluate the operating status of actuators and the performance of control systems.

[0042] The system integrates an intelligent fault diagnosis module and establishes a health status assessment model for actuators based on deep belief networks. Fault feature vectors. Multi-dimensional signal analysis was used to extract features, including vibration spectrum characteristics, current harmonic characteristics, and temperature variation trends. The fault diagnosis decision function is as follows: ,in This is the Sigmoid activation function. When an abnormal state is detected, the system executes corresponding protective actions according to the preset safety policy, including smooth deceleration and shutdown, switching to a backup actuator, or activating the emergency braking device to ensure production safety. Simultaneously, all fault events and response measures are recorded in the system log, providing data support for subsequent maintenance and optimization.

[0043] In this embodiment, the state vector representing the overall quality of the product, the precise parameter adjustment instruction set, and the execution state feedback signal are aggregated. Root cause analysis and federated learning optimization are performed using a spatiotemporal correlation database to achieve continuous self-evolution of the system's end-to-end model. Specific implementation methods include: The system constructs a spatiotemporal relational database based on a distributed architecture, employing a hybrid storage mode combining time-series and relational databases. The database design uses a star schema, with production events as the fact table, linked to multiple dimension tables such as equipment parameters, process settings, and quality indicators. Each data unit is identified by a composite primary key: timestamp, spatial coordinates, and data source type. The system employs a sliding window-based streaming data processing mechanism, receiving data streams from various modules in real time and ensuring data quality through preprocessing operations such as data cleaning, format conversion, and missing value imputation. To establish a unified data view, the system performs multi-source data fusion, using a Kalman filter-based fusion algorithm to perform spatiotemporal alignment and confidence weighting on data from different sources and frequencies, generating high-quality, end-to-end production archives. These archives not only record "what happened," but more importantly, "when, where, and under what conditions it happened," laying the foundation for subsequent in-depth analysis.

[0044] When the system detects quality anomalies or performance degradation, it automatically triggers a root cause analysis process. The system first constructs a causal graph model of the production process, where nodes represent various process parameters, equipment status, and quality indicators, and edges represent possible causal relationships. Based on historical data, the system learns the causal graph structure through methods such as conditional independence testing. After determining the causal structure, the system uses interventional reasoning to quantify the impact of each factor on the quality problem. Specifically, for suspected root cause variables X and quality indicators Y, the causal effect is assessed by calculating the difference between the interventional distribution and the observed distribution. The system also developed a deep learning-based anomaly propagation model, which accurately locates the origin of the problem by analyzing the propagation path and temporal pattern of quality anomalies in the production line. The entire analysis process employs a multi-evidence fusion strategy, combining quantitative analysis and domain knowledge to generate an interpretable root cause analysis report and propose targeted improvement suggestions.

[0045] To achieve cross-production line knowledge sharing without leaking sensitive data, the system constructs a federated learning optimization framework. This framework comprises a central server and multiple clients. Clients train models locally using private data, only uploading model updates to the server. The server integrates model updates from multiple clients using a secure aggregation algorithm, generating an improved global model which is then distributed to all clients. To protect data privacy, the system introduces differential privacy technology during local client training, adding calibrated noise to model updates to ensure that information from individual data points cannot be reverse-engineered. Simultaneously, homomorphic encryption is used to encrypt transmitted model parameters to prevent man-in-the-middle attacks. The federated learning process uses an adaptive weight adjustment mechanism to balance the differences in data distribution among clients, ensuring that the generated global model has good generalization ability. This mechanism allows the system to fully utilize the operational experience of multiple production lines while strictly protecting the data privacy and trade secrets of each production line.

[0046] Based on root cause analysis results and knowledge gained from federated learning, the system performs online updates to the entire model. The update process employs a continuous learning strategy, utilizing elastic weight consolidation technology to retain important existing knowledge while incorporating new knowledge, effectively overcoming catastrophic forgetting. Model performance is monitored through a multi-objective evaluation system, including prediction accuracy, control effectiveness, and computational efficiency. When model performance degradation or significant environmental changes are detected, the system automatically triggers a model retraining process. The training process uses incremental learning, fine-tuning model parameters based on the latest production data to ensure the model remains synchronized with the actual production process. The system also establishes a robust version management and rollback mechanism. Each model update generates a complete update log, including update content, performance changes, and impact assessment information. If the new model performs poorly, it can quickly revert to a previous stable version, ensuring system reliability.

[0047] The system has established a complete self-evolution mechanism, forming a continuous improvement closed loop of "monitoring-analysis-optimization-verification". By monitoring key performance indicators in real time, the system can autonomously evaluate the operating status and control effectiveness of each module. Based on these evaluation results, the system not only adjusts model parameters but also optimizes data processing strategies, feature extraction methods, and even system architecture configuration. For example, when it is found that the contribution of certain sensor data to quality prediction is consistently low, the system may suggest adjusting their sampling frequency or deployment location; when new quality influencing factors are detected, the system automatically expands the feature set and updates the fusion model. All these optimization measures are verified through A / B testing and other methods to ensure their effectiveness. The system also maintains a knowledge base to record lessons learned from previous optimizations and establishes connections between various problems and solutions through knowledge graph technology, providing a reference for future optimization decisions. This self-evolution mechanism enables the system to continuously adapt to changes in production processes, continuously improve quality control capabilities, and ultimately achieve a leap from "automation" to "intelligence".

[0048] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0049] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multi-sensor fusion based real-time quality control system for an automated production line, characterized in that, Comprise the following modules: Data perception and acquisition module: relying on multi-source heterogeneous sensor array, synchronous perception and adaptive acquisition of production line global state and workpiece processing quality are performed, and standardized synchronous data stream with space-time label is output; Data fusion and feature extraction module: introducing the standardized synchronous data stream, multi-source information confidence evaluation and deep feature extraction are performed by means of dynamic adaptive deep learning fusion model, and state vector representing product comprehensive quality is generated; Prediction-decision-control instruction generation module: according to the state vector representing product comprehensive quality, quality trend advance simulation and intelligent decision are completed through prediction-correction mechanism driven by digital twin, and precise parameter adjustment instruction set is output; Precise execution and state feedback module: receiving the precise parameter adjustment instruction set, process parameter real-time regulation and action planning are completed by manipulating high-precision execution mechanism, and execution effect is verified synchronously by using integrated sensor to form execution state feedback signal; Root cause analysis and self-evolution optimization module: the state vector representing product comprehensive quality, the precise parameter adjustment instruction set and the execution state feedback signal are converged, root cause analysis and federal learning optimization are performed relying on space-time correlation database, and continuous self-evolution of system full-link model is realized.

2. The automated production line real-time quality control system based on multi-sensor fusion according to claim 1, characterized in that, The data perception and acquisition module is configured to: Multi-source heterogeneous sensor array is used to implement synchronous monitoring of production line global state and workpiece processing quality, and original multi-source data stream is obtained; Adaptive sampling strategy is used to process the original multi-source data stream, and collection frequency dynamic adjustment is completed, and optimized collection data is output; Standardization processing method is used to perform format unification and noise filtering on the optimized collection data, and standardized data set is generated; Through space-time calibration technology, the standardized data set is bound with time stamp and space position information, and the standardized synchronous data stream with space-time label is generated.

3. The automated production line real-time quality control system based on multi-sensor fusion according to claim 1, characterized in that, The data fusion and feature extraction module is configured to: Through the standardized synchronous data stream with space-time label, multi-source information confidence evaluation is performed by using dynamic adaptive deep learning fusion model, and each data source weight distribution is obtained; Weighted fusion algorithm is used to process the each data source weight distribution, multi-source data fusion calculation is completed, and high reliability fusion data is output; Deep neural network is used to implement space-time feature extraction on the high reliability fusion data, and deep quality feature representation is obtained; Through feature integration method, the deep quality feature representation is processed, multi-dimensional feature fusion is completed, and the state vector representing product comprehensive quality is generated.

4. The automated production line real-time quality control system based on multi-sensor fusion according to claim 3, characterized in that, The dynamic adaptive deep learning fusion model is configured to: The attention mechanism is used to analyze the standardized synchronous data stream with space-time label, the importance evaluation of each sensor data is completed, and dynamic weight coefficient is output; Relying on weighted average method, the dynamic weight coefficient is processed, multi-source data fusion is performed, and weighted fusion result is generated; Through convolutional neural network, deep feature learning is performed on the weighted fusion result, high quality feature representation is obtained; Online learning mechanism is used to optimize the parameters of the dynamic adaptive deep learning fusion model, and the continuously improved fusion model is realized.

5. The automated production line real-time quality control system based on multi-sensor fusion according to claim 1, characterized in that, The prediction-decision-control instruction generation module is configured to: Perform quality trend prediction by using a digital twin model on the state vector representing the comprehensive quality of the product, to obtain future quality state evolution data; Analyze the future quality state evolution data using a risk assessment method to complete quality deviation assessment and output key adjustment parameter identification results; Process the key adjustment parameter identification results using a reinforcement learning algorithm to implement control strategy optimization and generate an optimal control scheme; An instruction conversion module converts the optimal control scheme into machine executable instructions to form the precise parameter adjustment instruction set.

6. The automated production line real-time quality control system based on multi-sensor fusion according to claim 5, characterized in that, The digital twin driven prediction-correction mechanism is configured to: Complete production process simulation using a digital twin model based on real-time production line operation data to obtain virtual production line operation states; Analyze the virtual production line operation states using a prediction algorithm to perform quality parameter evolution simulation and output quality trend prediction results; Process the quality trend prediction results using a correction mechanism to complete control strategy adjustment and generate optimized control parameters; Update the digital twin model parameters through real-time feedback data to achieve a prediction model with continuously improved accuracy.

7. The automated production line real-time quality control system based on multi-sensor fusion according to claim 1, characterized in that, The precise execution and state feedback module is configured to: Complete process parameter adjustment using high-precision execution mechanisms based on the precise parameter adjustment instruction set to obtain parameter adjustment results; Monitor the execution mechanism operation state using an embedded sensor network to implement real-time data acquisition and output execution process data; Analyze the execution process data using a verification algorithm to complete instruction execution effect evaluation and generate execution state verification results; Process the execution state verification results using a signal processing module to generate feedback signals and form the execution state feedback signals.

8. The automated production line real-time quality control system based on multi-sensor fusion according to claim 1, characterized in that, The traceability analysis and self-evolution optimization module is configured to: Complete full-link data integration using data integration technology based on the state vector representing the comprehensive quality of the product, the precise parameter adjustment instruction set, and the execution state feedback signals to obtain complete production records; Process the complete production records using a data mining algorithm to implement quality defect root cause analysis and output root cause positioning results; Combine the root cause positioning results using a knowledge update mechanism to complete quality control strategy optimization and generate improved quality control rules; Implement multi-node collaborative training using a federated learning framework to complete system model optimization and obtain a continuously evolving full-link model.

9. The automated production line real-time quality control system based on multi-sensor fusion according to claim 8, characterized in that, The federated learning optimization is configured to: Update the local model using model training based on local production data to obtain local model parameters; Process the local model parameters using secure encryption technology to implement privacy protection processing and output encrypted model parameters; Integrate the encrypted model parameters using a federated averaging algorithm to complete global model aggregation and generate an optimized global model; Process the optimized global model using a model distribution mechanism to implement local model updates and obtain a collaboratively optimized system model.

10. The automated production line real-time quality control system based on multi-sensor fusion according to claim 8, characterized in that, The root cause traceability analysis is configured to: Complete historical data query using a spatio-temporal correlation database based on quality defect characteristics to obtain a related process data set; Analyze the related process data set using an anomaly detection algorithm to implement abnormal time period identification and output an abnormal time window; The abnormal parameter mining is completed by using a correlation analysis method to process the abnormal time window, and a key abnormal parameter is generated; The root cause positioning is implemented by analyzing the key abnormal parameter based on a cause-effect reasoning model, and a quality defect cause analysis report is generated.

Images