Intelligent maintenance system for sensor

The intelligent sensor maintenance system solves the problems of disconnect between testing and diagnosis, rigid calibration strategies, and lack of data support for maintenance decisions in sensor production testing and maintenance. It enables efficient root cause tracing and personalized calibration, and improves the automated diagnostic accuracy and product consistency of the sensor production line.

CN121614542APending Publication Date: 2026-03-06GUANGZHOU JINGYAN INTELLIGENT TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202511950682.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-12-23
Publication Date
2026-03-06

AI Technical Summary

Technical Problem

Traditional sensor production, testing, and maintenance methods suffer from a disconnect between testing and diagnosis, rigid calibration strategies, lack of data support for maintenance decisions, and sluggish process feedback, resulting in low efficiency and wasted resources, making it difficult to achieve a closed loop of process optimization.

Method used

Deploy a sensor intelligent maintenance system to achieve full-process data association and intelligent diagnosis through production test data aggregation, digital twin quality benchmark model, fault root cause tracing module, adaptive calibration module and production line precision repair decision module, generate personalized calibration plan and optimize production process.

Benefits of technology

It has achieved a leap from result judgment to root cause tracing, improved the level and accuracy of diagnostic automation, generated a unique lightweight and dedicated compensation model, improved product performance consistency and yield, shortened failure analysis time, and ensured long-term reliability.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121614542A_ABST
    Figure CN121614542A_ABST
Patent Text Reader

Abstract

The invention discloses an intelligent maintenance system for a sensor, and the system employs a sensor ideal digital twinborn model to calculate a performance deviation vector through the correlation of production test data and traceability information; the core lies in that a preset production knowledge graph is combined, and a machine learning model is adopted to carry out process-level root cause tracing on faults; based on a traceability result and individual deviation characteristics, the system dynamically generates a personalized optimal calibration function for each sensor, and makes an intelligent decision including software repair and precise physical repair. According to the method, the conversion from post-screening to in-process diagnosis and repair and pre-prevention optimization in sensor production is realized, the first pass yield of a production line and the product consistency are improved, and the quality cost is reduced.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of industrial automation and intelligent manufacturing technology, and more specifically, to an intelligent sensor maintenance system. Background Technology

[0002] As core sensing components in modern industrial systems, the performance consistency of sensors directly determines the quality and reliability of end products. During mass production, due to process fluctuations, material differences, and changes in equipment condition, sensor products may experience performance deviations or even failures. Traditional production testing and maintenance methods suffer from the following significant bottlenecks: First, there is a severe disconnect between testing and diagnosis. While existing automated testing equipment can efficiently collect test data, when products exceed specifications, it typically only provides a binary "pass / fail" judgment, lacking the automatic and precise identification of the root cause of failure, such as specific processes and parameters. Fault analysis heavily relies on engineer experience, resulting in low efficiency and difficulty in standardization. Second, calibration strategies are rigid and simplistic. For products with performance parameters that can be corrected through calibration, production lines generally use fixed, general compensation formulas based on large-sample statistics. This approach cannot provide tailored optimization calibration for the unique deviation patterns of each sensor, potentially masking deeper defects or failing to fully realize individual potential. Third, maintenance decisions lack data support. Whether to scrap, rework, or downgrade non-conforming products often depends on fixed rules or manual experience, failing to comprehensively consider the nature of the defect, repair costs, production line resources, and long-term reliability risks, leading to resource waste or potential quality issues. Finally, process feedback is sluggish. Test results at the end of the production line cannot be quickly and automatically correlated and fed back to specific upstream processes, making it difficult to detect and correct process problems in a timely manner, easily leading to batch losses.

[0003] Therefore, given the current production situation, there is a need for an intelligent maintenance solution for sensor production lines that can deeply integrate test data, achieve intelligent diagnosis and decision-making, and form a closed loop for process optimization. Summary of the Invention

[0004] In view of the shortcomings of the existing technology, the purpose of this invention is to provide a sensor intelligent maintenance system to solve the problems existing in the background technology.

[0005] The above-mentioned technical objective of this invention is achieved through the following technical solution: a sensor intelligent inspection and maintenance system, deployed in the testing stage of a sensor production line, for data interaction with automated testing equipment and manufacturing execution systems, comprising: The production test data aggregation module is used to synchronously collect and associate full-dimensional test sequence data from sensors through different test stations, and then fit and generate test data packages for the corresponding sensors, and synchronously bind a full-process traceability identifier containing wafer batch, process chamber number, operator ID and material batch number to the test data package. The digital twin quality benchmark model module is used to store the multi-physics coupling simulation model established for each type of sensor product. The simulation model is used to simulate the theoretical transfer function and parameter-performance mapping relationship of the entire link of the sensor from the sensing unit to the signal conditioning circuit within ideal manufacturing tolerances, and then output the standard reference value of the corresponding sensor. The fault root cause tracing module generates a high-dimensional performance residual vector of the sensor based on the data from the production test data aggregation module and the digital twin quality benchmark model module. Then, it constructs a dynamically updated manufacturing defect propagation map. Based on the manufacturing defect propagation map, it uses a graph attention network and gradient boosting tree fusion model to perform multi-hop inference and outputs a sorted list of potential defect process chains, as well as the contribution and confidence of each chain. The adaptive calibration module performs personalized calibration on sensors that meet the calibration access conditions as determined by the fault root cause tracing module. Based on the high-dimensional performance residual vector of the sensor, it dynamically selects and combines a two-stage optimization strategy from a pool of hybrid algorithms including polynomial fitting, piecewise linear interpolation, neural network compensation and physical empirical formulas to generate a lightweight dedicated compensation model and its corresponding calibration coefficient set. The production line precision repair decision and instruction generation module is integrated into a multi-objective decision optimizer. It combines the data from the fault root cause tracing module and the adaptive calibration module to generate the optimal decision with the optimization objectives of improving first-pass yield, reducing rework costs, and ensuring long-term reliability. The optimal decision is then transformed into repair coordinates and resistance values ​​that can directly drive the laser repair machine, local re-welding parameters of selective welding equipment, or batch interception and process adjustment instructions from the MES system.

[0006] Optionally, the full-dimensional test sequence data includes raw analog voltage / current signals, digital output code values, signal noise spectra acquired at multiple static and dynamic excitation points, parameter drift data under high and low temperature and mechanical stress environments, and micron-level morphological images of sensor sensitive units and packaging structures acquired by a high-resolution linear array camera or confocal microscope integrated in the test station.

[0007] Optionally, the mechanism for constructing and updating the dynamically updated manufacturing defect propagation map in the fault root cause tracing module is as follows: Initial construction phase: Import product process roadmap from MES to automatically generate process nodes; import equipment instance nodes from the equipment management module in MES; import material batch nodes from the material management module in MES; based on the historical failure analysis report library in MES, extract defect-cause pairs through natural language processing technology, and combine the review and correction of process experts to establish directed edges between relevant nodes, and assign initial conditional transition probabilities to each edge based on historical statistical frequency; During the online learning phase: A case feedback queue is set up, and each sensor case that has completed the entire process is placed in the queue. The background process periodically retrieves cases from the queue. If the defect chain inferred from the case is consistent with the root cause that is finally verified, the probability of the relevant edge is increased in the positive direction along the defect chain using the Bayesian update rule. If they are inconsistent or the repair fails, the probability of the relevant edge is decreased in the reverse direction, or an alarm for reviewing the graph structure is triggered.

[0008] Optionally, the training and workflow of the graph attention network and gradient boosting tree fusion model are as follows: Training phase: A large amount of historical sensor data and its finally confirmed root causes of defects are used as training samples; first, the graph attention network is fixed, and unsupervised pre-training is performed using the graph structure and node attributes to obtain the initial embedding representation of the nodes; then, the node embeddings are concatenated with the high-dimensional performance residual vectors of the corresponding sensors as input features of the GBDT model, and supervised training is performed with the defect root cause process chain as the label. Inference phase: For the test sensor, a high-dimensional performance residual vector is calculated. A pre-trained graph attention network is used to calculate the attention weights of all nodes in the current graph relative to the high-dimensional performance residual vector, thereby obtaining the context-aware node embedding. The node embedding and the residual vector are concatenated and input into the trained GBDT model. GBDT outputs the original prediction score for each potential defect chain. Then, the original score is converted into a normalized confidence probability through a softmax layer and sorted and output.

[0009] Optionally, the two-stage optimization strategy in the adaptive calibration module includes a second-stage multi-objective optimization process as follows: Define a fitness function F(x) = [f1(x), f2(x), f3(x)], where f1(x) is the negative of the predicted root mean square error after adopting the candidate calibration model (maximizing accuracy), f2(x) is the reciprocal of the number of model parameters, and f3(x) is the negative of the model's estimated running time on the calibration chip; Use a non-dominated sorting genetic algorithm to search for the Pareto optimal frontier in the solution space; During decision-making, a fuzzy logic selector automatically selects the final solution from the Pareto frontier based on the current production line's preference weights for accuracy, simplicity, and delay.

[0010] Optionally, the multi-objective constraint optimization decision maker in the production line precision repair decision and instruction generation module has a built-in decision mathematical model that is a mixed integer programming problem with time constraints; wherein, the decision variables are binary choice variables, the objective function is the weighted sum of the above multi-objectives, and the constraints include: repair resource capacity constraints, mutual exclusion or dependency constraints between repair actions, order delivery time window constraints, and hard constraints on the unrepairability of defects provided by the fault root cause tracing module.

[0011] Optionally, the system also includes a continuous learning and process optimization loop module, which includes a macro-analysis unit for periodically aggregating and mining all processed data and performing the following tasks: Cross-batch spatial clustering analysis identifies sensor clusters with similar residual patterns and defect root causes, and traces back their common production factors to discover new potential defect propagation paths that have not yet been included in the map. The process window monitoring performs correlation analysis and control chart plotting on the equipment parameters of key processes and the residuals of the final aggregate analysis output, and provides early warning when the parameter drift approaches the control limit; Generate optimization suggestion reports and proactively recommend adjustments to the center values ​​of process parameters or preventive maintenance plans for equipment to MES and management personnel.

[0012] A sensor intelligent repair method based on the aforementioned sensor intelligent repair system, applied to the entire production process from sensor wafer manufacturing to final testing and packaging, includes the following steps: S1. Production test data aggregation steps: Synchronously collect and associate full-dimensional test sequence data from sensors through different test stations, fit and generate test data packages for the corresponding sensors, and synchronously bind a full-process traceability identifier containing wafer batch, process chamber number, operator ID and material batch number to the test data package. S2. Digital twin benchmark comparison steps: Call the multiphysics coupling simulation model corresponding to the sensor model, input the test conditions to simulate its full-link theoretical transfer function and parameter-performance mapping relationship, and output the standard reference value; compare the sensor's measured test data package with the standard reference value to generate the sensor's high-dimensional performance residual vector; S3. Fault Root Cause Inference Steps: Input the high-dimensional performance residual vector into a graph attention network and gradient boosting tree fusion model constructed based on a dynamically updated manufacturing defect propagation graph; use this model to perform multi-hop inference in the graph, output a sorted list of potential defect process chains, as well as the contribution and confidence of each chain; S4. Adaptive calibration modeling steps: For sensors that are determined to meet the calibration access conditions after abductive reasoning, based on their high-dimensional performance residual vector, a two-stage optimization strategy is used to dynamically select and combine calibration algorithms from the hybrid algorithm pool to generate a lightweight dedicated compensation model and its corresponding calibration coefficient set. S5. Precise Repair Decision and Execution Steps: The output results of the fault root cause reasoning step and the output results of the adaptive calibration modeling step are input into a multi-objective decision optimizer. The optimizer aims to improve the first pass rate, reduce rework costs, and ensure long-term reliability. Under the conditions of meeting resource, time, and process constraints, it solves the optimal decision and transforms the decision into control instructions to drive the laser trimming machine, selective welding equipment, or manufacturing execution system. S6. Continuous Learning and Optimization Steps: Regularly aggregate and mine the data from all processed sensors, perform cross-batch spatial clustering analysis to discover new potential defect propagation paths, and conduct correlation analysis and control chart monitoring on the batch residuals of key process equipment parameters and aggregated outputs, thereby generating process optimization or equipment maintenance suggestions.

[0013] In summary, the present invention has the following beneficial effects: 1. It achieves a leap from result judgment to root cause tracing, improving the level of automation and accuracy of diagnosis; the system can automatically and accurately associate the complex performance deviations at the test end, i.e., residual vectors, with specific defective processes, equipment instances, or material batches in the preceding stage, and output defect chains with confidence, replacing fuzzy analysis that relies on human experience; at the same time, it solidifies expert experience and historical cases through a computable knowledge graph, making the diagnostic process standardized and replicable, and solving the problem of over-reliance on senior engineers; the failure analysis process that may have taken hours or even days has been shortened to minutes or even seconds, realizing online real-time diagnosis and providing the possibility for rapid decision-making and interception.

[0014] 2. It achieves a shift from empirical decision-making to data-driven global optimal decision-making; the system generates a unique, lightweight, dedicated compensation model for each sensor type, enabling fine-grained compensation for its specific, non-linear deviation patterns, thereby achieving optimal performance within the allowable hardware potential range and improving the overall product performance consistency and yield rate; through multi-objective optimization, including but not limited to accuracy, complexity, and latency, it ensures that the generated calibration model is not only accurate but also consumes low computational resources and has a fast programming speed, making it suitable for large-scale deployment on the production line without introducing additional bottlenecks; the calibration strategy generation considers the conclusions of fault attribution, and for deep defects that may affect long-term reliability (such as material aging), the system can refuse calibration to prevent potentially risky products from entering the market. Attached Figure Description

[0015] Figure 1This is a schematic diagram of the system logic of the present invention; Figure 2 This is a schematic diagram of the method steps of the present invention. Detailed Implementation

[0016] To make the objectives, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Several embodiments of the present invention are shown in the drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein.

[0017] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances. The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first" or "second" may explicitly or implicitly include one or more of that feature.

[0018] In this invention, unless otherwise expressly specified and limited, "above" or "below" a second feature can include direct contact between the first and second features, or contact between the first and second features through another feature between them. Furthermore, "above," "over," and "on top" of a second feature includes the first feature being directly above or diagonally above the second feature, or simply indicates that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" of a second feature includes the first feature being directly below or diagonally below the second feature, or simply indicates that the first feature is at a lower horizontal level than the second feature. The terms "vertical," "horizontal," "left," "right," "above," "below," and similar expressions are for illustrative purposes only and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed or operated in a specific orientation, and therefore should not be construed as limiting the invention.

[0019] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0020] This invention provides a sensor intelligent maintenance system, such as... Figure 1 As shown, it is deployed in the sensor production line testing phase for data interaction with automated testing equipment and manufacturing execution systems, including: The production test data aggregation module is used to synchronously collect and associate full-dimensional test sequence data from sensors through different test stations, and then fit and generate test data packages for the corresponding sensors, and synchronously bind a full-process traceability identifier containing wafer batch, process chamber number, operator ID and material batch number to the test data package. The digital twin quality benchmark model module is used to store the multi-physics coupling simulation model established for each type of sensor product. The simulation model is used to simulate the theoretical transfer function and parameter-performance mapping relationship of the entire link of the sensor from the sensing unit to the signal conditioning circuit within ideal manufacturing tolerances, and then output the standard reference value of the corresponding sensor. The fault root cause tracing module generates a high-dimensional performance residual vector of the sensor based on the data from the production test data aggregation module and the digital twin quality benchmark model module. Then, it constructs a dynamically updated manufacturing defect propagation map. Based on the manufacturing defect propagation map, it uses a graph attention network and gradient boosting tree fusion model to perform multi-hop inference and outputs a sorted list of potential defect process chains, as well as the contribution and confidence of each chain. The adaptive calibration module performs personalized calibration on sensors that meet the calibration access conditions as determined by the fault root cause tracing module. Based on the high-dimensional performance residual vector of the sensor, it dynamically selects and combines a two-stage optimization strategy from a pool of hybrid algorithms including polynomial fitting, piecewise linear interpolation, neural network compensation and physical empirical formulas to generate a lightweight dedicated compensation model and its corresponding calibration coefficient set. The production line precision repair decision and instruction generation module is integrated into a multi-objective decision optimizer. It combines the data from the fault root cause tracing module and the adaptive calibration module to generate the optimal decision with the optimization objectives of improving first-pass yield, reducing rework costs, and ensuring long-term reliability. The optimal decision is then transformed into repair coordinates and resistance values ​​that can directly drive the laser repair machine, local re-welding parameters of selective welding equipment, or batch interception and process adjustment instructions from the MES system.

[0021] Furthermore, the full-dimensional test sequence data includes raw analog voltage / current signals, digital output code values, signal noise spectra acquired at multiple static and dynamic excitation points, parameter drift data under high and low temperature and mechanical stress environments, and micron-level morphological images of sensor sensitive units and packaging structures acquired by a high-resolution linear array camera or confocal microscope integrated in the test station.

[0022] Furthermore, the mechanism for constructing and updating the dynamically updated manufacturing defect propagation map in the fault root cause tracing module is as follows: Initial construction phase: Import product process roadmap from MES to automatically generate process nodes; import equipment instance nodes from the equipment management module in MES; import material batch nodes from the material management module in MES; based on the historical failure analysis report library in MES, extract defect-cause pairs through natural language processing technology, and combine the review and correction of process experts to establish directed edges between relevant nodes, and assign initial conditional transition probabilities to each edge based on historical statistical frequency; During the online learning phase: A case feedback queue is set up, and each sensor case that has completed the entire process is placed in the queue. The background process periodically retrieves cases from the queue. If the defect chain inferred from the case is consistent with the root cause that is finally verified, the probability of the relevant edge is increased in the positive direction along the defect chain using the Bayesian update rule. If they are inconsistent or the repair fails, the probability of the relevant edge is decreased in the reverse direction, or an alarm for reviewing the graph structure is triggered.

[0023] Furthermore, the training and workflow of the graph attention network and gradient boosting tree fusion model are as follows: Training phase: A large amount of historical sensor data and its finally confirmed root causes of defects are used as training samples; first, the graph attention network is fixed, and unsupervised pre-training is performed using the graph structure and node attributes to obtain the initial embedding representation of the nodes; then, the node embeddings are concatenated with the high-dimensional performance residual vectors of the corresponding sensors as input features of the GBDT model, and supervised training is performed with the defect root cause process chain as the label. Inference phase: For the test sensor, a high-dimensional performance residual vector is calculated. A pre-trained graph attention network is used to calculate the attention weights of all nodes in the current graph relative to the high-dimensional performance residual vector, thereby obtaining the context-aware node embedding. The node embedding and the residual vector are concatenated and input into the trained GBDT model. GBDT outputs the original prediction score for each potential defect chain. Then, the original score is converted into a normalized confidence probability through a softmax layer and sorted and output.

[0024] Furthermore, the two-stage optimization strategy in the adaptive calibration module, specifically the multi-objective optimization process in the second stage, is as follows: Define a fitness function F(x) = [f1(x), f2(x), f3(x)], where f1(x) is the negative of the predicted root mean square error after adopting the candidate calibration model (maximizing accuracy), f2(x) is the reciprocal of the number of model parameters, and f3(x) is the negative of the model's estimated running time on the calibration chip; Use a non-dominated sorting genetic algorithm to search for the Pareto optimal frontier in the solution space; During decision-making, a fuzzy logic selector automatically selects the final solution from the Pareto frontier based on the current production line's preference weights for accuracy, simplicity, and delay.

[0025] Furthermore, the multi-objective constraint optimization decision maker in the production line precision repair decision and instruction generation module has a built-in decision mathematical model that is a mixed integer programming problem with time constraints; wherein, the decision variables are binary choice variables, the objective function is the weighted sum of the above multi-objectives, and the constraints include: repair resource capacity constraints, mutual exclusion or dependency constraints between repair actions, order delivery time window constraints, and hard constraints on the unrepairability of defects provided by the fault root cause tracing module.

[0026] Furthermore, the system also includes a continuous learning and process optimization loop module, which includes a macro-analysis unit for periodically aggregating and mining all processed data and performing the following tasks: Cross-batch spatial clustering analysis identifies sensor clusters with similar residual patterns and defect root causes, and traces back their common production factors to discover new potential defect propagation paths that have not yet been included in the map. The process window monitoring performs correlation analysis and control chart plotting on the equipment parameters of key processes and the residuals of the final aggregate analysis output, and provides early warning when the parameter drift approaches the control limit; Generate optimization suggestion reports and proactively recommend adjustments to the center values ​​of process parameters or preventive maintenance plans for equipment to MES and management personnel.

[0027] In a specific embodiment, this system is deployed in the production line testing area and interconnected with automated testing equipment, MES, and repair equipment (such as laser repair machines) via industrial Ethernet and data interfaces. The core data flow of the system is as follows: Data uplink: All data generated by physical sensors at the testing station is collected and packaged by the production test data aggregation module and sent to the system; Core processing: The data flows sequentially through the digital twin quality benchmark model module, the fault root cause tracing module, the adaptive calibration module, and the production line precise repair decision and instruction generation module, completing benchmark comparison, defect diagnosis, repair scheme generation, and decision-making in sequence; Instruction downlink: The final decision is compiled into specific instructions and sent to the execution equipment or MES; Closed-loop feedback: All processing results are collected and aggregated by the continuous learning and process optimization loop module to update the system's internal model and knowledge base, forming a closed loop.

[0028] The specific implementation process of the production test data aggregation module: This module is the starting point of the system data flow. Its core task is to transform raw test data from different physical devices, different formats, and different time points into a unified, complete, and traceable standardized data package. The module works through a distributed data acquisition architecture, which includes a test instruction coordinator, a data acquisition agent (deployed on each test device), a visual inspection unit, and a central data aggregation server. When the manufactured sensor arrives at the testing station with the carrier, the MES sends the sensor's test task code to the test instruction coordinator of this module. The coordinator parses the task code, generates a specific sequence of equipment control instructions, and broadcasts a synchronous start signal to all relevant equipment, including but not limited to electrical testers, temperature chambers, and vision systems, via a high-precision hardware clock, ensuring that all data acquisition has a unified start time reference. Subsequently, each device performs tests and collects data in parallel: raw analog voltage / current signals are sampled at a high rate, such as 1 MS / s, through a high-precision data acquisition card connected to the sensor's analog output pins; the collected data is the complete transient response waveform of the output voltage after a pressure step signal is applied; digital output code values ​​are directly read from the digital register value of the sensor's built-in ADC through a digital I / O card or communication bus; the collected data is the 12-bit digital code value representing the current pressure under steady state; signal noise spectrum is obtained by acquiring steady-state signals over a period of time using the acquisition card, and fast Fourier transform analysis is performed by the module's built-in digital signal processor; the collected data is the calculation of the noise power spectral density in the 0-100 kHz frequency band and the extraction of the total noise RMS value; parameter drift data under high and low temperatures and mechanical stress; the sensor is placed in a program-controlled temperature chamber or stress loading device to collect parameter drift data under high and low temperatures and mechanical stress. The electrical tester performs measurements at the set temperature / stress points, collecting data on the changes in the sensor's zero point and sensitivity as the temperature cycles from -40°C to 85°C. Micron-level morphology images are captured by a high-resolution confocal microscope integrated in the workstation, automatically focused and using a ring light source at a specific angle to photograph key areas, including grayscale / color images of the varistor strip, gold wire bonding points, and encapsulation sealing surface. While collecting raw data, the module's built-in time series and image feature extraction subunit performs preliminary processing in real time to compress the data volume and extract key features, including electrical signal features and image features. All raw and feature data are streamed in real time to a central data aggregation server. The server performs fitting and generation according to the following logic: timestamp alignment and data stitching. The server uses the initial acquisition T0 as a reference and adds a millisecond-level offset timestamp Δt relative to T0 to each data point. Using Δt, data from different devices describing the same test action, such as applying 50 kPa pressure at 25°C, are aligned and correlated, stitched into a record with complete context. Then, the correlated data is encapsulated into standardized data packet objects according to a preset hierarchical tree structure, and then fitted to generate test data packets. The generation and binding of traceability identifiers are integrated throughout the entire process of test preparation, execution, and encapsulation, ensuring a unique correspondence between data and the physical world. The acquisition of identifier information involves the following steps: when the sensor carrier arrives at the workstation, the reader scans the RFID or QR code on the carrier to obtain the sensor's unique serial number module, which is then transmitted via a standard interface (such as WebService / OPC). The UA module queries the MES system in real time and submits the serial number to the MES. Based on the production model, the MES returns the complete production history bound to the serial number, starting from the wafer, including the wafer batch number, a list of key process equipment numbers, the current workstation operator number, and a list of main material batch numbers. The binding occurs when the data package is generated. Simultaneously, to ensure data authenticity, the module calculates the entire SHA-256 hash value before the data package is finally sealed and digitally signs it using a production line digital certificate. This signature is stored along with the traceability identifier and the data package; any subsequent data tampering will result in signature verification failure. The production test data aggregation module, through a series of concrete and implementable operations—hardware synchronous triggering, multi-source heterogeneous data acquisition, dynamic feature extraction, spatiotemporal alignment stitching, structured encapsulation, and real-time interaction with the MES to bind tamper-proof traceability identifiers—transforms physical world testing behaviors into a complete digital twin that can be precisely processed by the intelligent system. This provides a solid and reliable data foundation for all subsequent advanced analyses.

[0029] The digital twin quality benchmark model module serves as the ideal benchmark for the system. Its core responsibility is to dynamically calculate the theoretical output value that a perfect product should have based on the specific test conditions for each sensor under test, providing an accurate benchmark for subsequent performance deviation analysis. A module is a core software component that consists of four logically and physically cooperating sub-units, including: The parametric model library is used to store the multiphysics coupling simulation models built for each type of sensor product. It is a version-managed database or file library. Specifically, it stores the model files of all defined sensor models, such as the SP-100 pressure sensor and the GT-200 gas sensor. Each model file is not a black box, but a collection of parametric mathematical equations, finite element mesh data, material property tables and circuit netlists, which together define the ideal physical behavior of the product model. The multiphysics simulation engine unit is the computational core of the module. It can parse and execute the parametric descriptions in the model library and perform coupled solutions. The engine integrates a finite element analysis solver for mechanical and thermal structural field analysis and a circuit simulator for electrical signal analysis. It also defines the data exchange interface between the two to ensure that changes in the physical field can affect the electrical properties in real time, thus realizing multiphysics coupling. The simulation parameter manager is responsible for dynamically configuring the model's input parameters and boundary conditions based on the specific sensor's test conditions before simulation. Its input consists of test condition parameters received from the production test data aggregation module, such as applied pressure values ​​and ambient temperature. Its output is a set of complete initialization configuration instructions that are recognized by the simulation engine. The standardized output unit formats and publishes the raw results output by the simulation engine. The specific processing involves extracting key quantities from the simulation results that directly correspond to the measured data in the test data package, such as theoretical output voltage and theoretical zero-point value, and organizing them into a standard reference value vector aligned with the structure of the measured data package. In the specific implementation process, taking the MEMS pressure sensor (SP-100 type) as an example, the workflow of this module is explained in detail step by step. Assume the received test conditions are: at 25°C, the sequence of applied pressure points is as follows: ; The module loads the corresponding model file from the parametric model library based on the sensor model SP-100; the simulation parameter manager reads the test conditions. The manager will Write the global material property table into the model, such as setting the Young's modulus and piezoresistive coefficient of silicon at that temperature, and then... The sequence of boundary condition loads is prepared to be applied sequentially. Multiphysics coupling simulation is performed, with the simulation engine calculating each pressure point sequentially. For example: The finite element solver calculates the deformation of the pressure-sensitive diaphragm structure of the sensor under a uniformly distributed pressure boundary condition of 50 kPa, thus obtaining the stress distribution field on the diaphragm. Based on the piezoresistive effect, the relative change of the piezoresistive strip due to stress is calculated and expressed as: ,in and These are the longitudinal and transverse piezoresistive coefficients, respectively. and This step represents the first coupling from the physical field to the electrical parameters, where the principal stress component is located at the position of the resistor strip. The circuit simulator receives signals from the mechanical field. As a result, the resistance values ​​of the four arms of the Wheatstone bridge model are updated; based on the updated circuit, the output voltage of the bridge is solved, and its transfer function is simplified to:

[0030] in, Represented as excitation voltage, Represented as nominal resistance. This is expressed as the change in resistance calculated from stress. This is the fixed gain of the subsequent amplifier circuit; this step completes the second coupling from electrical parameters to the final output signal, forming a full-link simulation. The standardized output unit collects the simulation engine's calculation results for all test points, forming a theoretical value sequence that completely corresponds to the actual test data packet. For example: At the same time, it will be based on The zero-point simulation results are used to generate theoretical zero-point values. These theoretical values ​​are encapsulated into a standard reference value vector, whose data structure is strictly mapped to the corresponding fields in the test data package generated by the production test data aggregation module, ensuring that the downstream fault root cause tracing module can perform precise comparisons point by point and item by item. The digital twin quality benchmark model module is not a static lookup table system, but a dynamic simulation calculation system based on first principles. Through its specific components such as a parameterized model library and a multiphysics simulation engine, as well as a configuration-coupled simulation-output workflow, it generates high-fidelity ideal performance curves for each sensor under test in real time. This quantitative ideal benchmark is the technological foundation for all subsequent intelligent diagnostics (calculating high-dimensional performance residuals) and personalized calibration (the objective function of calibration).

[0031] The root cause analysis module is the intelligent diagnostic center of the system. Its core responsibility is to automatically map the performance deviation of the sensor to the specific defect root cause on the production line and output a diagnostic report in an interpretable and quantifiable manner. It consists of three core sub-units, forming a closed-loop architecture of computation-knowledge-reasoning. The residual calculation unit is responsible for generating the high-dimensional performance residual vector of the sensor. It receives two precise inputs from upstream: the sensor measured data packet from the production test data aggregation module and the corresponding standard reference value vector from the digital twin quality benchmark model module. It is used to perform item-by-item, traceable comparison and differential calculation, transforming abstract non-conformities into specific, quantified deviation patterns. The manufacturing defect propagation graph database unit is a graph structure database, where nodes represent production processes, physical equipment, and material batches; directed edges represent potential propagation paths of defects; edge weights represent the conditional probability of propagation; the database is not static, and its internal edge weights are continuously updated and optimized based on system operation results through Bayesian update rules; The Graph Attention Network and Gradient Boosting Tree Fusion Model Unit is the core reasoning unit of this module. It is actually a pre-trained and executable AI model with the following capabilities: Graph structure understanding, which uses the Graph Attention Network to understand the complex topological and semantic relationships between nodes in the manufacturing defect propagation graph; and multimodal feature decision-making, which uses the Gradient Boosting Tree to fuse the structural information of the graph and the numerical information of the residuals to make a final comprehensive judgment. In the specific implementation process, the residual calculation unit is activated, and taking the MEMS pressure sensor in the aforementioned embodiment as an example, the following formulaic comparison is performed: For the i-th pressure test point, calculate the sensitivity deviation component. Calculate the zero-point temperature drift deviation component. It also integrates morphological features, with data derived from image analysis during the testing phase. This generates a structured vector R, represented as:

[0032] The vector R is the high-dimensional performance residual vector, which fully describes the specific modes of the sensor deviating from the ideal state from multiple dimensions such as sensitivity, linearity, temperature drift, and physical morphology. Graph Node Embedding (Graph Attention Network Operation): The generated vector R is fed into the fusion model of the graph attention network and gradient boosting tree. The graph attention network part of the model treats R as a query, scanning the current manufacturing defect propagation graph and calculating an attention weight for each node v in the graph. The weight represents the correlation between the node and the current residual pattern R, expressed as: ,in These are inherent attributes of node v, such as process type and equipment ID. and It is a learnable weight matrix; then, the information from all nodes is aggregated according to the attention weights to form a context-aware graph embedding vector. It condenses the spectral structure information related to the current fault; Multi-feature fusion and classification (gradient boosting tree work): The model embeds the graph into vectors. The joint feature vector is concatenated with the original residual vector R to form a joint feature vector. The joint feature vector is input into a trained gradient boosting tree model, with GBDT as the classifier. Through a series of decision trees, it learns the mapping from complex features to specific defect root causes, and then outputs a list of all predefined potential defect chains. and the corresponding original predicted score ; The model finally uses a softmax layer to convert the original scores into probabilities, represented as follows: The module then outputs a sorted list of potential defective process chains. Each item in the list explicitly includes: a description of the defect chain (e.g., bonding machine-03 -> insufficient pressure -> causing negative sensitivity bias and abnormal morphology); contribution (the percentage of the total residual vector R explained by the defect chain); and confidence level (the calculated probability). ; Online learning and updating of the diagnostic graph (closed-loop feedback): After each diagnosis is completed, the module sends the case, including the residual vector, the inferred defect chain, and the final maintenance verification result, into a case feedback queue. If the subsequent maintenance results confirm the diagnosis, the conditional probability of the defect chain is increased using Bayesian updates along all edges. If the inference is incorrect or the repair fails, the probability of the relevant edges is decreased, or the path is marked for review by process experts. This process achieves dynamic updating, enabling the system's diagnostic knowledge to continuously improve and correct itself as the production process progresses, becoming more accurate with use.

[0033] The adaptive calibration module is the system's personalized repair scheme generator. Its core responsibility is to dynamically generate an optimal, lightweight software calibration scheme for sensors that are determined to be repairable, based on their unique performance deviation patterns, rather than using a fixed formula. When determining whether a sensor is repairable, the decision is made by the root cause analysis module. After completing the core defect diagnosis reasoning, it not only outputs a "ranked list of potential defective process chains," but also simultaneously outputs a Boolean flag indicating whether the sensor is suitable for entering the calibration process. The judgment is based on its own reasoning results, mainly considering two points: Physical properties of defects: Based on knowledge graphs and AI models, determine whether the identified defect root causes (such as insufficient bonding pressure or temperature drift coefficient deviation) belong to systematic deviations that can be compensated or corrected by software algorithms; for example, hardware damage such as permanent material damage and structural fractures will be judged as uncalibrable / unrepairable. Confidence threshold: For the inferred compensable defect, the confidence level must reach a preset threshold (e.g., confidence level > 75%); this ensures that the calibration action is based on reliable diagnosis and avoids invalid calibration for misjudged "faults".

[0034] This module is the software optimization engine, consisting of four core sub-units, including: The calibration case library and hybrid algorithm pool manager are the physical carriers and managers of a hybrid algorithm pool containing multinomial fitting, piecewise linear interpolation, neural network compensation, and physical empirical formulas. Specifically, it consists of: Algorithm pool: a pre-built, scalable function library containing the above-mentioned basic calibration algorithms; Case library: a database storing historical successful calibration cases, each case recording a residual vector pattern and the effective algorithm combination and parameters that ultimately achieve the performance target. The meta-learning controller is the overall control unit of the module, responsible for implementing the two-stage optimization strategy and clearly directing the system to switch from the first stage of case retrieval to the second stage of multi-objective evolutionary optimization. The multi-objective evolutionary optimizer is the core computational unit of the second stage. Specifically, it executes the non-dominated sorting genetic algorithm (NSGA-II). Its input is the initial candidate model population received from the first stage, and its output is a Pareto optimal front, which is a set of optimal solutions that cannot be surpassed simultaneously in terms of accuracy, simplicity, and efficiency, through iterative evolution. The fuzzy logic selector, after obtaining the Pareto optimal solution set, automatically selects a final solution from many optimal solutions based on the real-time strategy preference of the current production line. It internally defines a set of if-then rules, which transforms the strategy language, such as precision priority, into mathematical operations for filtering and sorting the solution set. In the specific implementation process, the workflow strictly follows the order of condition determination, two-stage optimization, and model generation; assuming that a sensor is determined to be calibrable by the fault root cause tracing module, its high-dimensional performance residual vector R is sent to this module. In the first stage, based on case retrieval and initialization, the controller uses the current sensor residual vector R as the query condition to perform a nearest neighbor search in the calibration case library. The similarity between R and the residual vector of each historical case in the library is calculated, preferably cosine similarity, to retrieve the K most similar successful cases. The calibration algorithm combinations and parameter ranges used in these K successful cases are used as the initial population of the genetic algorithm, thus ensuring that the optimization starting point is an empirically effective region. In the second stage, multi-objective evolutionary optimization is performed. The optimizer starts iterating from the initial population of the genetic algorithm, and performs the following operations in each generation: Fitness assessment: For each candidate calibration model M in the population, calculate its fitness function vector F(M), which specifically includes: , representing accuracy, is the negative of the root mean square error of the prediction after the model has compensated for all test points represented by R; , representing simplicity, is the reciprocal of the number of model parameters; , representing efficiency, is the negative value of the estimated runtime of the computational model on the sensor microcontroller; therefore, The optimization objective is to maximize each component of vector F. Non-dominated sorting and selection: The NSGA-II algorithm is used to perform non-dominated sorting and crowding calculation on individuals in the population based on F(M). Those solutions that are not dominated by other solutions (i.e., better on at least one objective and not worse on other objectives) and are sparsely distributed are preferentially retained to form the next generation of population. Genetic operations: selection, crossover, and mutation operations are performed on the population. Mutation operations may include replacing an algorithm from a pool of mixed algorithms, adjusting the order of a polynomial, or the number of layers in a neural network. Iterative loop: Repeat the above steps until the preset number of algebras is reached, ultimately obtaining a Pareto optimal front solution set, denoted as... ; The final decision-making and model generation process involves a fuzzy logic selector that operates based on the current production line preference strategy (e.g., "accuracy > efficiency > simplicity") obtained from the MES. It quantifies the textual strategy as a weighted allocation of the three components of F(M) and... All solutions are comprehensively evaluated and scored; The output is generated, and the selector automatically selects the final model with the highest overall score. ,Should For a lightweight, proprietary compensation model, the module then extracts... The specific parameters, such as polynomial coefficients and neural network weights, form the corresponding calibration coefficient set; The output connects with the downstream, generating The model and its coefficient set are encapsulated and transmitted to the production line precision repair decision and instruction generation module. This module performs multi-objective decision-making. The prediction accuracy (f1) and efficiency (f3) are key inputs for evaluating the effectiveness of software repair options; if the decision is software repair, this set of coefficients will be compiled into a burning file; The adaptive calibration module achieves a qualitative leap from general calibration to dedicated calibration through a specific and coherent optimization workflow, including case retrieval initialization, multi-objective evolutionary search, and fuzzy strategy decision-making.

[0035] The production line precision repair decision and instruction generation module is the system's decision-making brain and execution hub. Its core responsibility is to make the best global decision by comprehensively considering the economics, resource constraints, and reliability risks of technical diagnosis and repair solutions, and to transform abstract decisions into precise instructions that can drive specific production equipment or management systems. This module is an integrated optimization and control system, consisting of three core sub-units working together: The multi-objective constrained optimization decision maker is the core component of the module. It encapsulates the defined decision mathematical model, which is a mixed integer programming problem with time constraints. It also serves as the model solving engine, integrating solvers such as CPLEX, Gurobi, or open source solvers to perform complex mathematical programming solutions. The objective and constraint configurator dynamically loads objective function weights and constraint parameters. The real-time production line data unit is used to continuously obtain dynamic context information required for decision-making from the Manufacturing Execution System (MES) and the shop floor IoT platform. The collected data includes, but is not limited to, the real-time status and queue length of each repair resource (such as whether the laser repair machine is idle), material inventory, current order delivery time window, and comprehensive cost model parameters (such as labor hour rate and material unit price). The instruction compilation and distribution unit transforms the abstract decisions output by the optimization decision maker, such as performing laser trimming, into low-level control instructions or transactional requests that downstream devices or systems can recognize and execute. It has a built-in instruction library with pre-stored instruction templates for different types of devices, such as laser trimming machines and selective welding equipment, as well as standardized API interface protocols for interacting with MES. In practice, workflow is an automated process of data aggregation, model solving, instruction compilation, and distribution execution. Multi-source input data integration: The module receives inputs from upstream modules and aligns them with real-time production data, including: Technical inputs: Received from the root cause analysis module [potential defect chain list, calibration admission criteria, defect irreparability judgment]; where the defect irreparability judgment will be directly converted into hard constraints in the decision model; Solution inputs: Received from the adaptive calibration module [lightweight dedicated compensation model, calibration coefficient set, calibration effect prediction; calibration effect is the core indicator for evaluating the quality of software repair options]; Resource and cost inputs: Obtained through the real-time production line data unit [equipment resource status, material costs, order delivery pressure]; Construct and solve the decision mathematical model, and optimize the decision maker. Based on the above input, dynamically construct a mixed integer programming problem, the specific form of which is: Decision variables: Define a set of 0-1 variables, where x1=1 represents choosing software repair and release, x2=1 represents choosing laser trimming, and x3=1 represents choosing scrapping. The constraint is x1+x2+x3=1, indicating mutual exclusion, meaning only one can be chosen. The optimization objective function is to minimize the total cost or maximize the total utility, usually a weighted sum of multiple objectives, expressed as: ,in Estimated by the increase in pass rate; and The cost model is derived from the module's built-in cost model, representing the costs of calibration and rework. This is represented as a reliability risk score, which is positively correlated with the severity of the defect root cause. , , Weights are set according to the management strategy; Constraints include, but are not limited to: resource capacity constraints, expressed as Ensure that the time taken by the selected action does not exceed the available time of the resources; action dependency constraint, represented as, for example, x2 (laser adjustment) <= x4 (opening already performed), indicating that some rework actions require a preliminary step; time window constraint, ensuring that the total time of the path selected in the decision meets the order delivery requirements; technical hard constraint: if the defect irreparability judgment in the upstream input is true, then force x3 (scrap) = 1; The solution engine calculates the above problem and outputs a set of solutions that optimize the objective function. Value, i.e., optimal decision; The instruction compilation unit makes optimal decisions. The value determines the corresponding operation: If x1=1 (software repair), call the calibration coefficient set generated by the adaptive calibration module; select the corresponding programming protocol template according to the sensor model; compile and generate a binary programming file or bus communication instruction sequence containing coefficients, checksums, and write addresses; If x2=2 (physical repair, such as laser adjustment), combine the defect location contribution provided by the fault root cause tracing module to locate the specific resistor bar; calculate the laser adjustment coordinates, pulse energy, and number of emission times based on the resistance value to be adjusted; compile these parameters into equipment-specific G-code or SECS / GEM instructions.

[0036] If x3=1 (scrap) or batch interception is involved: generate a structured message containing sensor ID, reason code, and suggested action; call the material status update and batch locking service of the MES system through a standard API (such as RESTful); before issuing the instruction, a security and feasibility check will be performed, such as checking whether the target device is online and whether the status is ready; Command distribution and execution feedback: After verification, the command is distributed to the corresponding execution terminal through the industrial network. The module also listens for the command execution result (success / failure) and uses this final result as closed-loop information to feed back to the continuous learning and process optimization loop module for updating the knowledge graph and optimizing the decision model. The production line precision repair decision and instruction generation module encodes multiple factors such as technical feasibility, economic cost, production resources, and delivery pressure into a solvable mathematical optimization problem, and is equipped with professional instruction compilation capabilities, realizing automated decision-making from multiple solution suggestions to a single optimal executable instruction.

[0037] In other embodiments, a continuous learning and process optimization loop module is also provided. The module undertakes the functions of macro analysis, experience extraction and self-evolution. Its core responsibility is to go beyond the processing of a single sensor, to mine potential patterns from massive historical data, monitor the stability of the production process, and actively output optimization suggestions to drive the continuous improvement of the manufacturing system. The module is a back-end analysis system consisting of four logical sub-units, including: The macro analysis engine is responsible for scheduling the entire analysis process, automatically triggering analysis tasks according to preset cycles (such as each shift or daily), and aggregating and mining the required full amount of data from the system database; A cross-batch clustering analyzer is used for cross-batch spatial clustering analysis to identify sensor clusters with similar residual patterns and defect root causes. The core algorithm integrates unsupervised clustering algorithms such as DBSCAN or K-means, and has the ability to trace back production factors from the clustering results. The process window monitor is used for process window monitoring. It performs correlation analysis and control chart plotting on the residuals of equipment parameters in key processes and the final aggregate analysis output. The core tools include a statistical correlation analysis library (such as calculating the Pearson correlation coefficient) and a statistical process control module (used to plot X-bar charts, P-charts, etc.). An optimization suggestion report generator is used to generate optimization suggestion reports, proactively recommending adjustments to the center values ​​of process parameters or preventive maintenance plans for equipment to MES and management personnel, and then transforming the analysis conclusions into structured and operable report templates, which are automatically pushed through the interface; In practice, the workflow is a periodic, batch data aggregation, parallel analysis, report generation, and feedback loop process. Assume the system is set to perform automatic analysis every day at midnight. The macro-analysis engine is triggered and performs the following operations: Data extraction: Extract all processed sensor data from the system database over the past 24 hours; Key inputs include: From the root cause analysis module: the high-dimensional performance residual vector (R) for each sensor and the finally adopted defect root cause; From the production test data aggregation module: the full-process traceability identifier for each sensor (wafer batch, equipment cavity, material batch); From the MES: historical logs of key process equipment parameters corresponding to the above sensor production time (such as bonding machine pressure, curing oven temperature profiles). Then, the engine performs data alignment and aggregation. Using the sensor ID as the key, the engine aligns the above data in production time order to form a cross-batch dataset for analysis. This allows for the parallel execution of three core analysis tasks, including: Task A, Clustering: The clustering analyzer extracts the high-dimensional performance residual vector R of all sensors from the dataset, performs standardization, and then runs the DBSCAN clustering algorithm. This algorithm can discover densely connected sample groups and effectively identify anomalous clusters. Core algorithm analysis: For a sample point p, if it contains at least MinPts samples in its ε-neighborhood, then p is a core object, and all densely connected core objects form a cluster. Data tracing: For each identified anomalous cluster, the analyzer performs statistical backtracking based on the common end-to-end tracing identifier shared by all sensors within the cluster. For example, it discovers… 95% of the sensors used bonder-03 and epoxy resin batch-B; Output: Conclusion: A new potential defect pattern cluster was discovered, characterized by [sensitivity negative bias > 5%, large zero-point temperature bias], which is highly suspected to be related to the combination of {bonder-03, epoxy resin batch-B}; This conclusion is used to update the knowledge graph and add new potential defect propagation paths; Task B, Residual Aggregation: The monitor processes the dataset. For example, it selects zero-point temperature drift residual as a monitoring indicator and calculates the average zero-point temperature drift residual of all sensors produced in that chamber within that hour, using the production equipment cavity and hour as the unit. Correlation analysis will Pairing the parameters with key parameters in the cavity process log (such as the curing temperature setpoint), the Pearson correlation coefficient between the two is calculated and expressed as: ,like A strong correlation indicates that the equipment parameter has a significant impact on product quality; Control chart plotting: with time as the horizontal axis and... Plot a single-value moving range control chart (I-MR chart) with the vertical axis as the y-axis; calculate and plot the upper control limit, center line, and lower control limit; for example, when 7 consecutive points appear on the same side of the center line, or when a point exceeds the control limit, a warning is triggered that the parameter drift is approaching the control limit; Task C: The report generator receives structured conclusions from Task A and Task B, integrates and analyzes them: for example, combining the suspected problem with bonder-03 in Task A and the strong correlation and upward trend between the pressure parameters and residuals of bonder-03 in Task B; generating recommendations: process parameter adjustment: based on correlation analysis, it is recommended to lower the center value of the pressure setpoint of bonder-03 by ΔP; preventive maintenance: based on the trend warning of the control chart, it is recommended to perform calibration and maintenance of the temperature sensor of curing oven-05; formatted push: the report is generated in a standard format (such as PDF, JSON) and automatically sent to preset process engineers and managers via email or MES interface.

[0038] The feedback loop and system evolution mean that the data produced by this module does not exist in isolation, but flows back to the core system to drive evolution. This includes: feedback to the root cause analysis module: submitting new defect propagation paths discovered by Task A as candidate nodes and edges to the knowledge graph for review and integration, so that the next diagnosis can identify this new pattern; and feedback to the production line precision repair decision module: using the equipment health scores monitored by Task B as new cost or risk constraints in the decision model. For example, increasing the weight of scrapping decisions for products from equipment with frequent warnings. The continuous learning and process optimization loop module performs three specific tasks—clustering, monitoring, and suggestion generation—through periodic macro-data analysis. This not only enables passive maintenance of individual products but also drives the entire production system to evolve towards an advanced form of predictive maintenance and proactive optimization.

[0039] A sensor intelligent repair method based on the aforementioned sensor intelligent repair system, applied to the entire production process from sensor wafer manufacturing to final testing and packaging, includes the following steps: S1. Production test data aggregation steps: Synchronously collect and associate full-dimensional test sequence data from sensors through different test stations, fit and generate test data packages for the corresponding sensors, and synchronously bind a full-process traceability identifier containing wafer batch, process chamber number, operator ID and material batch number to the test data package. S2. Digital twin benchmark comparison steps: Call the multiphysics coupling simulation model corresponding to the sensor model, input the test conditions to simulate its full-link theoretical transfer function and parameter-performance mapping relationship, and output the standard reference value; compare the sensor's measured test data package with the standard reference value to generate the sensor's high-dimensional performance residual vector; S3. Fault Root Cause Inference Steps: Input the high-dimensional performance residual vector into a graph attention network and gradient boosting tree fusion model constructed based on a dynamically updated manufacturing defect propagation graph; use this model to perform multi-hop inference in the graph, output a sorted list of potential defect process chains, as well as the contribution and confidence of each chain; S4. Adaptive calibration modeling steps: For sensors that are determined to meet the calibration access conditions after abductive reasoning, based on their high-dimensional performance residual vector, a two-stage optimization strategy is used to dynamically select and combine calibration algorithms from the hybrid algorithm pool to generate a lightweight dedicated compensation model and its corresponding calibration coefficient set. S5. Precise Repair Decision and Execution Steps: The output results of the fault root cause reasoning step and the output results of the adaptive calibration modeling step are input into a multi-objective decision optimizer. The optimizer aims to improve the first pass rate, reduce rework costs, and ensure long-term reliability. Under the conditions of meeting resource, time, and process constraints, it solves the optimal decision and transforms the decision into control instructions to drive the laser trimming machine, selective welding equipment, or manufacturing execution system. S6. Continuous Learning and Optimization Steps: Regularly aggregate and mine the data from all processed sensors, perform cross-batch spatial clustering analysis to discover new potential defect propagation paths, and conduct correlation analysis and control chart monitoring on the batch residuals of key process equipment parameters and aggregated outputs, thereby generating process optimization or equipment maintenance suggestions.

[0040] This invention provides an intelligent sensor maintenance system that leaps from result judgment to root cause tracing, improving the level of automation and accuracy in diagnosis. The system can automatically and accurately correlate complex performance deviations at the testing end, i.e., residual vectors, to specific defective processes, equipment instances, or material batches in the preceding stages, outputting a defect chain with confidence, replacing fuzzy analysis relying on human experience. Simultaneously, it solidifies expert experience and historical cases through a computable knowledge graph, standardizing and replicating the diagnostic process and reducing over-reliance on senior engineers. It shortens the failure analysis process, which might have taken hours or even days, to minutes or even seconds, enabling online real-time diagnosis and providing the possibility for rapid decision-making and interception. This system moves away from empiricism... The system represents a shift from decision-making to data-driven global optimal decision-making. It generates a unique, lightweight, proprietary compensation model for each sensor type, enabling refined compensation for specific, non-linear deviation patterns. This achieves optimal performance within the allowable hardware potential, improving overall product performance consistency and yield. Through multi-objective optimization, including but not limited to accuracy, complexity, and latency, the generated calibration model is ensured to be not only accurate but also computationally efficient, fast to program, and suitable for large-scale deployment on production lines without introducing additional bottlenecks. The calibration strategy considers fault attribution conclusions; for deep-seated defects that may affect long-term reliability (such as material aging), the system can refuse calibration, preventing potentially risky products from entering the market.

[0041] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A sensor intelligent maintenance system, characterized by, Deployed in the sensor production line test link, used for data interaction with automated test equipment and manufacturing execution system, including: a production test data aggregation module for synchronously collecting and correlating full-dimension test sequence data from different test stations of a sensor, fitting to generate a test data package of the corresponding sensor, and synchronously binding a full-process traceability identifier including wafer batch, process chamber number, operator ID and material batch number on the test data package; a digital twin quality benchmark model module for storing a multi-physics field coupling simulation model established for each type of sensor product, the simulation model being used to simulate a full-link theoretical transfer function and parameter-performance mapping relationship of a sensor from a sensitive unit to a signal conditioning circuit within an ideal manufacturing tolerance, and then output a standard reference value of the corresponding sensor; a fault root cause tracing module for generating a high-dimensional performance residual vector of a sensor based on data of the production test data aggregation module and the digital twin quality benchmark model module, constructing a dynamically updated manufacturing defect propagation graph, and based on the manufacturing defect propagation graph, using a graph attention network and gradient boosting tree fusion model to perform multi-hop reasoning, and then output a sorted potential defect process chain list, and contribution and confidence of each link; an adaptive calibration module for performing individual calibration on a sensor determined by the fault root cause tracing module to meet calibration admission conditions, dynamically selecting and combining from a hybrid algorithm pool including polynomial fitting, piecewise linear interpolation, neural network compensation and physical empirical formula based on a high-dimensional performance residual vector of the sensor, and then generating a lightweight dedicated compensation model and its corresponding calibration coefficient set; a production line accurate repair decision and instruction generation module integrated as a multi-objective decision optimizer, comprehensively using data of the fault root cause tracing module and the adaptive calibration module, for generating an optimal decision to improve yield, reduce repair cost and ensure long-term reliability as optimization objectives, and converting the optimal decision into repair and adjustment coordinates and resistance values of a laser repair and adjustment machine, local re-soldering parameters of a selective soldering device, or batch interception and process adjustment instructions of a MES system.

2. The sensor intelligent maintenance system of claim 1, wherein, The full-dimension test sequence data includes original analog voltage / current signals, digital output code values, signal noise spectra collected at multiple static and dynamic excitation points, parameter drift data under high and low temperature, mechanical stress environment, and micron-level topography images of sensor sensitive units and packaging structures collected by high-resolution line array cameras or confocal microscopes integrated in the test station.

3. The sensor intelligent maintenance system of claim 1, wherein, The construction and updating mechanism of the dynamically updated manufacturing defect propagation graph in the fault root cause tracing module is as follows: The initial construction stage: importing product process route map from MES to automatically generate process nodes; importing equipment instance nodes from the equipment management module in MES; importing material batch nodes from the material management module in MES; based on the historical failure analysis report library in MES, extracting defect-cause pairs through natural language processing technology, and combining with the review and correction of process experts to establish directed edges between related nodes, and assigning initial conditional transition probabilities to each edge based on historical statistical frequency; The online learning stage: a case feedback queue is set up, and each sensor case that completes the whole process is put into the queue; the background process periodically takes cases from the queue, if the defect chain traced by the inference is consistent with the root cause verified finally, then the Bayesian update rule is used to positively enhance the probability of related edges along the defect chain; if it is not consistent or the repair fails, then the probability of related edges is negatively attenuated, or a review alert of the graph structure is triggered.

4. The sensor intelligent maintenance system of claim 3, wherein, The training and workflow of the graph attention network and gradient boosting tree fusion model are as follows: The training stage: using a large amount of historical sensor data and its final confirmed defect root cause as training samples; first, fix the graph attention network, and use the graph structure and node attributes for unsupervised pre-training to obtain the initial embedding representation of the node; Then, the node embedding and the high-dimensional performance residual vector of the corresponding sensor are spliced as the input features of the GBDT model, and the defect root cause process chain is used as the label for supervised training; The reasoning stage: for the test sensor, the high-dimensional performance residual vector is calculated, the attention weight of all nodes in the current graph relative to the high-dimensional performance residual vector is calculated by using the pre-trained graph attention network, and then the context-aware node embedding is obtained; the node embedding and the residual vector are spliced and input into the trained GBDT model, and the GBDT outputs the original prediction score of each potential defect chain; then the original score is converted into normalized confidence probability through the softmax layer and sorted output.

5. The sensor intelligent maintenance system of claim 1, wherein, The two-stage optimization strategy in the adaptive calibration module, the multi-objective optimization process in the second stage is: define the fitness function F(x)=[f1(x),f2(x),f3(x)], wherein f1(x) is the negative value of the root mean square error of the prediction after adopting the candidate calibration model (maximize the accuracy), f2(x) is the inverse of the number of model parameters, and f3(x) is the negative value of the running time of the model on the calibration chip; search for the Pareto optimal front in the solution space by using the non-dominated sorting genetic algorithm; when making a decision, the fuzzy logic selector selects the final scheme from the Pareto front according to the preference weights of accuracy, simplicity and delay of the current production line.

6. The sensor intelligent maintenance system of claim 1, wherein, The multi-objective constraint optimization decision maker in the production line precision repair decision and instruction generation module has a built-in decision mathematical model that is a mixed integer programming problem with timeliness constraints; wherein the decision variable is a binary selection variable, the objective function is the weighted sum of the above multi-objective, and the constraint conditions include: repair resource capacity constraints, mutual exclusion or dependency constraints between repair actions, order delivery time window constraints, and hard constraints on defect irreparability provided by the fault root cause tracing module.

7. The sensor intelligent maintenance system of claim 1, wherein, The module also includes a continuous learning and process optimization loop, which includes a macro analysis unit for periodically aggregating and mining all processing data and performing the following tasks: Cross-batch space clustering analysis to identify sensor clusters with similar residual patterns and defect roots, and to backtrack their common production elements to discover new potential defect propagation paths that have not been included in the atlas; Process window monitoring to perform correlation analysis and control chart drawing on the device parameters of key processes and the residuals output by the final aggregated analysis, and to provide early warning when parameter drift approaches the control limit; Generating optimization recommendation reports to actively recommend process parameter center value adjustments or equipment preventive maintenance plans to the MES and management personnel.

8. A sensor intelligent maintenance method based on the sensor intelligent maintenance system according to any one of claims 1-7, characterized in that, Applied to the entire production process from sensor wafer manufacturing to final test packaging, including the following steps: S1, production test data aggregation step: synchronously collecting and correlating full-dimensional test sequence data from different test stations of the sensor, fitting to generate a test data package corresponding to the sensor, and synchronously binding a full-process traceability identifier including wafer batch, process chamber number, operator ID, and material batch number on the test data package; S2, digital twin benchmarking step: calling a multi-physical field coupled simulation model corresponding to the sensor model, inputting test conditions to simulate its full-link theoretical transfer function and parameter-performance mapping relationship, and outputting standard reference values; comparing the measured test data package of the sensor with the standard reference values to generate a high-dimensional performance residual vector of the sensor; S3, fault root cause tracing reasoning step: inputting the high-dimensional performance residual vector into a graph attention network and gradient boosting tree fusion model based on a dynamically updated manufacturing defect propagation atlas; using the model to perform multi-hop reasoning in the atlas, outputting a sorted list of potential defect process chains, and the contribution and confidence of each link; S4, adaptive calibration modeling step: for sensors that meet the calibration admission conditions after tracing reasoning, based on their high-dimensional performance residual vectors, dynamically selecting and combining calibration algorithms from a mixed algorithm pool through a two-stage optimization strategy to generate a lightweight dedicated compensation model and its corresponding calibration coefficient set; S5, precision repair decision and execution step: inputting the output results of the fault root cause tracing reasoning step and the output results of the adaptive calibration modeling step into a multi-objective decision optimizer; the optimizer aims to improve the pass rate, reduce repair costs, and ensure long-term reliability, and solves the optimal decision under the condition of meeting resource, time, and process constraints, and converts the decision into control instructions to drive the laser tuning machine, selective welding equipment, or manufacturing execution system; S6, continuous learning and optimization step: periodically aggregate mining of all processed sensor data, perform cross-batch spatial clustering analysis to discover new potential defect propagation paths, and perform correlation analysis and control chart monitoring of key process equipment parameters and aggregated output batch residuals, and then generate process optimization or equipment maintenance recommendations.