Intelligent management system and method for lead bonding machine fusing cloud-fog-edge-end collaboration

Abstract

The invention discloses a cloud-fog-edge-end cooperation fused intelligent management system for a lead bonding machine, and the system comprises an equipment layer which comprises a multi-source sensor array integrated with or carried by the lead bonding machine, a physical execution unit, and a microsecond closed-loop controller. The multi-source sensor array is used for collecting real-time multi-physical field data in a bonding process, the physical execution unit is used for executing a bonding operation, and the microsecond closed-loop controller is used for responding to a control instruction and driving the physical execution unit; the edge calculation layer is deployed on a lead bonding machine and performs microsecond-level communication with the equipment layer; the fog computing layer is deployed in a local network and is used for performing plant-level collaboration on data and models of the at least two edge computing layers; and the cloud computing layer is deployed in a remote data center and is used for globally aggregating and analyzing the data and the model of the at least one fog computing layer and providing cross-domain knowledge service and full-life-cycle management, so that a high-reliability and self-adaptive bonding environment is favorably provided for semiconductor packaging.

Description

Technical Field

[0001] This application relates to the field of industrial internet technology, and in particular to an intelligent management system and method for wire bonding machines that integrates cloud-fog-edge-device collaboration. Background Technology

[0002] As the semiconductor industry moves towards higher integration and smaller dimensions, wire bonding, as a core process in chip packaging, directly impacts the performance and reliability of electronic products in terms of its quality and efficiency. Compared to traditional general-purpose industrial control systems, wire bonding requires micron-level positioning accuracy and millisecond-level action response, while simultaneously controlling multiple physical field parameters such as mechanical pressure, ultrasonic power, and temperature field. The precision requirements far exceed those of conventional industrial scenarios.

[0003] The wire bonding industry currently faces numerous technical challenges, the most significant being the difficulty of quality control. Globally, packaging plants scrap chips worth hundreds of millions of dollars annually due to poor solder joints. The wire bonding process involves the coupling of multiple physical fields, including mechanical pressure, ultrasonic vibration, and temperature. Traditional monitoring systems analyze only single physical quantities, failing to effectively process multi-physics coupled data and making it difficult to accurately predict process results and equipment status. Equipment generates thousands of process data points per second (such as solder joint deformation, wire arc curvature, and gold ball diameter), but a large amount of this data remains unutilized.

[0004] Second, real-time control is insufficient. Traditional industrial cloud platforms have data sampling frequencies of seconds or higher, which is insufficient to meet the real-time data acquisition requirements of millisecond-level bonding actions during wire bonding. This results in the inability to accurately and timely capture and provide online warnings for key process parameters, such as solder ball deformation and micro-cracks in the bonding tool. Most cloud platforms adopt a centralized data processing model, where equipment data is transmitted to the cloud computing layer for analysis and feedback, resulting in significant latency and failing to meet the real-time control requirements of the equipment.

[0005] Third, the problem of data silos is severe. Existing wire bonding systems suffer from cross-platform data interoperability barriers due to incompatible proprietary communication protocols used by equipment manufacturers. Centralized data processing architectures require raw process parameters to be uploaded to the cloud for analysis, posing a risk of leakage of core process details. Privacy concerns prevent the sharing of bonding experience data across multiple production sites, resulting in isolated process optimization strategies and hindering global collaborative optimization.

[0006] Fourth, the level of automation is low. Optimization of process parameters relies heavily on on-site engineer operations, and key constraints such as material thermal expansion coefficients and equipment positioning accuracy lack digital correlation models. Optimal parameter combinations for similar equipment under similar operating conditions cannot be reused across production lines, leading to significantly extended debugging cycles. Personnel turnover causes a break in the transmission of process experience, with new engineers relying heavily on trial and error when debugging equipment, resulting in low efficiency. For the same model of bonding machine, parameter drift occurs due to assembly tolerances and component wear, and traditional solutions cannot provide adaptive compensation.

[0007] Fifth, equipment maintenance is lagging behind. Equipment maintenance relies on fixed-cycle upkeep or reactive repairs, lacking early warning mechanisms for progressive failures such as the propagation of microcracks in the cutting blades and the degradation of transducer performance. Traditional monitoring systems are limited by data processing capabilities and cannot build failure prediction models based on multi-source sensor time-series data. Sudden failures cause unplanned production line downtime, and the time spent on repairs significantly impacts delivery cycles.

[0008] Existing technologies employ two typical approaches: one is based on a general industrial IoT architecture, aggregating device data through edge gateways and uploading it to the cloud computing layer for processing. While this enables basic status monitoring, network transmission latency between the cloud computing layer and the device fails to meet the real-time response requirements for motion control and energy regulation in the bonding process. The other approach uses locally embedded computing units on the device, ensuring the real-time performance of the core control loop, but is limited by hardware resources that cannot support intelligent algorithms requiring high computing power. Both solutions exhibit structural flaws: the cloud computing layer solution suffers from real-time control failure due to excessively long control links, while the local solution suffers from weakened intelligent decision-making capabilities due to computing power bottlenecks. Therefore, there is an urgent need to develop an intelligent management system and method for wire bonding machines to increase the data processing accuracy and bonding quality during wire bonding. Summary of the Invention

[0009] One objective of this application is to provide an intelligent management system and method for wire bonding machines that integrates cloud-fog-edge-device collaboration, in order to solve the intelligent challenges under the triple technical constraints of "microscale, multi-physical field coupling, and high reliability" unique to the field of semiconductor wire bonding.

[0010] One objective of this application is to provide an intelligent management system and method for wire bonding machines that integrates cloud-fog-edge-device collaboration, thereby solving the long-standing problems in the wire bonding field, such as the contradiction between real-time control and global optimization, the dilemma of collaboration among multi-source heterogeneous devices, and the inefficiency and waste caused by reliance on experience.

[0011] To achieve the above objectives, this application provides an intelligent management system for wire bonding machines that integrates cloud-fog-edge-device collaboration, comprising: a device layer, including a multi-source sensor array, a physical execution unit, and a microsecond-level closed-loop controller integrated or mounted on the wire bonding machine; the multi-source sensor array is used to collect real-time multi-physics field data of the bonding process; the physical execution unit is used to perform bonding operations; and the microsecond-level closed-loop controller is used to respond to control commands and drive the physical execution unit; and an edge computing layer, deployed on the wire bonding machine, communicating with the device layer at the microsecond level; the edge computing layer includes an edge data processing and modeling module, used to process the real-time multi-physics field data of the wire bonding machine. The system processes multi-physics field data and constructs a local model representing the bonding process; a real-time control and decision-making module is used to make millisecond-level intelligent decisions based on the local model and generate real-time control commands to be sent to the microsecond-level closed-loop controller; an edge learning module is used to train the model locally using confidential process data and only upload updated model parameters; a fog computing layer is deployed on the local network and is used to perform plant-level collaboration on data and models from at least two edge computing layers; a cloud computing layer is deployed in a remote data center and is used to perform global aggregation and analysis on data and models from at least one fog computing layer, and provide cross-domain knowledge services and full lifecycle management.

[0012] In some embodiments, the real-time control and decision-making module of the edge computing layer further includes a bonding process emergency control submodule. The bonding process emergency control submodule is used to monitor abnormal operating conditions in the bonding process in real time. When a preset abnormal operating condition is detected, it performs a microsecond-level emergency shutdown at the device layer and automatically backtracks and stores multi-source sensor data before the abnormal operating condition occurs for source tracing analysis.

[0013] In some embodiments, the edge data processing and modeling module of the edge computing layer includes a multiphysics coupling engine submodule. The multiphysics coupling engine submodule fuses the multiphysics data through a high-precision time synchronization mechanism and dynamically calculates and outputs real-time predicted values ​​of key indicators of solder joint quality based on a physical information neural network model, so as to form a closed-loop control for adaptive adjustment of process parameters.

[0014] In some embodiments, the real-time control and decision-making module of the edge computing layer includes a vision-motion control linkage submodule. This submodule, based on a solder joint feature-motion control parameter correlation model, predicts the quality risk of the next solder joint based on the changing trends of current and historical solder joint image features, and generates pre-adjustment instructions for motion control parameters before defects occur, thereby proactively preventing bonding quality problems. The edge computing layer also pre-sets a baseline parameter set for rapid startup when new equipment is connected. The fog computing layer maintains a debugging parameter pool, which aggregates optimized process parameters of equipment already operating stably within the plant. When new equipment is connected, the system automatically migrates the optimized process parameters that best match the initial operating conditions of the new equipment from the debugging parameter pool, and performs compensation fine-tuning based on the unique deviations of the new equipment, thereby achieving rapid and high-precision debugging.

[0015] In some embodiments, the fog computing layer includes a factory-level intelligent application module, which includes a rapid process design engine submodule. The rapid process design engine submodule automatically generates a recommended process parameter set by inputting chip design parameters and combining a process rule base with a machine learning model; and uses digital twin technology to pre-enact the bonding process in a virtual environment to detect and avoid physical risks, thereby optimizing the recommended process parameter set.

[0016] In some embodiments, the system further includes a resilient activation mechanism configured to: dynamically adjust the system architecture based on the number of wire bonding machines or the complexity of the process; when a first preset condition is met, the edge computing layer is directly connected to the cloud computing layer; when a second preset condition is met, the fog computing layer is activated and deployed to enable factory-level collaboration between the edge computing layer and the cloud computing layer; the data flow between the various layers of the system adopts a layered optimized communication protocol stack: the device layer and the edge computing layer use industrial real-time Ethernet or high-speed fieldbus protocols; the edge computing layer and the fog computing layer use lightweight message queue protocols or OPC UAPubSub mode; the fog computing layer and the cloud computing layer use enterprise-level messaging protocols or HTTPS RESTful APIs.

[0017] In some embodiments, the cloud computing layer includes a global process knowledge and optimization module, which is configured to: construct and maintain a three-dimensional dynamic knowledge graph of materials, equipment, and processes; aggregate model parameters uploaded by each fog computing layer to continuously evolve the knowledge graph through a federated learning architecture without accessing the original confidential process data; and, based on the knowledge graph, achieve rapid adaptation of process parameters under new material or equipment scenarios through transfer learning; the cloud computing layer also includes an equipment lifecycle management module, which is configured to: construct a joint prediction model of equipment health and process stability, dynamically correlate the predicted health status of key equipment components with process parameters, and issue adaptive process parameter compensation instructions; and construct a performance degradation analysis and design feedback model, feeding back anonymized faults and process data from multiple plants to the equipment R&D and design end to form an optimization closed loop of "use-design".

[0018] To achieve the above objectives, this application provides an intelligent management method for wire bonding machines that integrates cloud-fog-edge-device collaboration, comprising the following steps: at the device layer, multi-physics field data of the bonding process is collected in real time using a multi-source sensor array, and control commands from the edge computing layer are executed through a microsecond-level closed-loop controller; at the edge computing layer, the multi-physics field data is processed and modeled in real time, and millisecond-level intelligent decisions are made based on the local model to generate real-time control commands, and only model parameter updates are uploaded after training the model using local confidential process data; at the fog computing layer, model parameter updates from at least two of the edge computing layers are aggregated and collaboratively optimized at the plant level; at the cloud computing layer, the aggregated model from at least one of the fog computing layers is globally trained and analyzed, and cross-domain knowledge services and full lifecycle management are provided.

[0019] In some embodiments, the step of making intelligent decisions in the edge computing layer further includes: real-time monitoring of abnormal operating conditions during the bonding process; triggering the device layer to perform microsecond-level emergency shutdown when a preset severe abnormality is detected; and automatically backtracking and storing multi-source sensor data before the abnormality occurs. The method also includes a zero-sample debugging step: when a new device is connected, a set of reference parameters pre-installed in the edge computing layer is first loaded to achieve rapid startup; then, optimized process parameters matching the initial operating conditions of the new device are intelligently retrieved and migrated from the debugging parameter pool deployed in the fog computing layer; finally, the migrated parameters are compensated and fine-tuned in combination with the unique deviations of the new device to complete high-precision debugging.

[0020] In some embodiments, the collaborative optimization step in the fog computing layer includes a rapid process design step: receiving chip design parameters as input and automatically generating a preliminary recommended process parameter set; using digital twin technology to pre-simulate the bonding process using the recommended process parameter set in a virtual environment; detecting and correcting potential physical risks based on the pre-simulation results to output an optimized process parameter set; the global training and analysis step in the cloud computing layer includes a knowledge graph construction and application step: aggregating model parameters uploaded from each fog computing layer through a federated learning architecture to continuously evolve a three-dimensional dynamic knowledge graph of materials, equipment, and processes without touching the original confidential process data; when new materials or equipment are introduced, performing relational reasoning and similarity matching based on the knowledge graph, and combining transfer learning to generate initial process parameters for rapid adaptation of the edge computing layer.

[0021] Compared with the prior art, the beneficial effects of this application are as follows: (1) This application integrates a cloud-fog-edge-device collaborative intelligent management system for wire bonding machines. Different levels are set up for different purposes based on the characteristics of wire bonding in four layers. For example, a dedicated sensor for wire bonding is set up in the equipment layer to increase the sensor's sampling period, thereby increasing the amount of sensor data measured. An edge computing layer with multiple modules processes the real-time multiphysics data and makes millisecond-level intelligent decisions based on the local model. The fog computing layer integrates data and models from multiple edge computing layers for factory-level collaboration. The cloud computing layer performs global aggregation and analysis, and provides cross-domain knowledge services and full lifecycle management, thereby improving the data processing accuracy and speed of the intelligent management system for wire bonding machines, and further enhancing the security of data processing. Attached Figure Description

[0022] Figure 1 This is an exemplary architecture diagram of a wire bonding machine intelligent management system that integrates cloud, fog, edge, and terminal collaboration, as described in this application.

[0023] Figure 2 This is an architectural diagram of an exemplary device layer in this application.

[0024] Figure 3 This is a flowchart of an exemplary edge computing layer in this application.

[0025] Figure 4 This is an architecture diagram of an exemplary edge computing layer in this application.

[0026] Figure 5 This is a flowchart illustrating the workflow of an exemplary fog computing layer's elastic activation mechanism in this application.

[0027] Figure 6This is a process-driven hierarchical architecture diagram of an exemplary wire bonding machine intelligent management system in this application.

[0028] Figure 7 This is an architecture diagram of an exemplary fog computing layer in this application.

[0029] Figure 8 This is a diagram of a multi-protocol compatible architecture with a three-layer decoupling design for an exemplary wire bonding machine intelligent management system in this application.

[0030] Figure 9A This application describes the construction process of an exemplary three-dimensional dynamic knowledge graph analysis model of materials, equipment, and processes.

[0031] Figure 9B This is an exemplary three-dimensional dynamic knowledge graph analysis model of materials, equipment, and processes in this application.

[0032] Figure 10 This is a flowchart of an exemplary process parameter transfer learning submodule in this application.

[0033] Figure 11 This is a flowchart of an exemplary global process parameter intelligent optimization submodule in this application.

[0034] Figure 12 This is a flowchart of an exemplary health and stability prediction submodule in this application.

[0035] Figure 13 This is a flowchart of an exemplary performance degradation analysis and design feedback submodule in this application.

[0036] Figure 14 This is a flowchart of an exemplary after-sales resource and cost optimization submodule in this application.

[0037] Figure 15 This is a flowchart illustrating an exemplary joint prediction model for equipment health and process stability in this application.

[0038] Figure 16 This is a flowchart of an exemplary global quality and process control module in this application.

[0039] Figure 17 This is a flowchart illustrating an exemplary cross-plant production and energy management module in this application.

[0040] Figure 18 This is a framework diagram of an exemplary wire bonding machine intelligent management system in this application.

[0041] Figure 19 This is a flowchart of an exemplary real-time data acquisition module in this application.

[0042] Figure 20 This is a flowchart of the workflow of an exemplary device status monitoring module in this application.

[0043] Figure 21 This is a flowchart of the workflow of an exemplary remote control module in this application.

[0044] Figure 22 This is a flowchart of an exemplary data analysis module in this application.

[0045] Figure 23 This is a flowchart of an exemplary alarm system module in this application.

[0046] Figure 24 This is a flowchart of an exemplary process parameter management module in this application.

[0047] Figure 25 This is a flowchart of an exemplary historical data query module in this application.

[0048] Figure 26 This is a flowchart of an exemplary user management module in this application.

[0049] Figure 27 This is an architectural diagram of an exemplary cloud computing layer in this application.

[0050] In the diagram: 1. Equipment Layer; 11. Multi-source Sensor Array; 111. Mechanical Motion Sensing Unit; 112. Ultrasonic Energy Sensing Unit; 113. Temperature Field Sensing Unit; 114. Visual Sensing Unit; 115. Environmental Monitoring Unit; 12. Microsecond-level Closed-loop Controller; 13. Physical Execution Unit; 131. XY Platform Drive Mechanism; 132. Z-axis Actuator; 133. Ultrasonic Transducer; 134. Heating Mechanism; 2. Edge Computing Layer; 21. Edge Data Processing and Modeling Module; 211. Multi-source Data Real-time Processing Submodule; 212. Multi-physics Coupling Engine Submodule; 213. Lightweight AI Quality Inspection Submodule; 22. Real-time Control and Decision Module; 221. Vision-Motion Control Linkage Submodule; 222. Bonding Process Emergency Control Submodule; 23. Edge Learning Module; 3. Fog Computing Layer; 31. Fog Layer Data and Communication Management Module; 311. Plant-level Data Aggregation The following modules are included in the cloud computing layer: 31. Integration and Processing Submodule; 312. Multi-protocol Gateway Submodule; 313. Security Agent Submodule; 32. Plant-level Collaborative Learning and Model Management Submodule; 33. Plant-level Intelligent Application Module; 331. Rapid Process Design Engine Submodule; 332. Production Line-level Quality and Process Control Submodule; 333. Production Line-level Production and Energy Consumption Management Submodule; 4. Cloud Computing Layer; 41. Global Federated Learning Aggregation Module; 42. Global Process Knowledge and Optimization Module; 421. Knowledge Graph Construction and Management Submodule; 422. Process Parameter Transfer Learning Submodule; 423. Global Process Parameter Intelligent Optimization Submodule; 43. Equipment Lifecycle Management Module; 431. Health and Stability Prediction Submodule; 432. Performance Degradation Analysis and Design Feedback Submodule; 433. After-sales Resource and Cost Optimization Submodule; 44. Global Quality and Process Control Module; 45. Cross-Plant Production and Energy Consumption Management Module. Detailed Implementation

[0051] The present application will be further described below with reference to specific embodiments. It should be noted that, without conflict, the various embodiments or technical features described below can be arbitrarily combined to form new embodiments.

[0052] It should be noted that the terms "first," "second," etc., in the specification and claims of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0053] The terms “comprising” and “having”, and any variations thereof, in the specification and claims of this application are intended to cover non-exclusive inclusion, that is, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but includes other steps or units that are not explicitly listed or that are inherent to such process, method, product, or apparatus.

[0054] This application provides an intelligent management system for wire bonding machines that integrates cloud-fog-edge-device collaboration to resolve the contradiction between real-time control and global optimization in wire bonding processes. It constructs a hierarchical, time-scale collaborative architecture, such as... Figure 1 As shown, it includes a device layer 1 with a response time of microseconds, an edge computing layer 2 with a response time of milliseconds, a fog computing layer 3 with a response time of hundreds of milliseconds, and a cloud computing layer 4 with a response time of seconds, forming a closed loop of the entire life cycle from perception and control to optimization and redesign.

[0055]

Equipment Layer 1

[0056] In one specific embodiment, since the wire bonding machine generates thousands of high-frequency data points per second, the multi-source sensor array 11 is used to collect real-time multi-physics field data during the bonding process from all directions and at high frequency. Specifically, it includes a mechanical motion sensing unit 111, an ultrasonic energy sensing unit 112, a temperature field sensing unit 113, a visual sensing unit 114, and an environmental detection unit 115, thereby monitoring the original physical signals throughout the wire bonding process and realizing real-time perception and data acquisition of the operating status of the device layer 1.

[0057] The mechanical motion sensing unit 111 can be a high-precision grating ruler or encoder mounted on the XY platform and Z axis, used to acquire micron-level displacement, velocity, and acceleration information in real time to monitor the motion trajectory of the chopping tool. The mechanical motion sensing unit 111 may include an XY motion platform position / velocity sensor and a Z-axis position sensor.

[0058] The ultrasonic energy sensing unit 112 is used to monitor the dynamic waveform of the power, frequency and amplitude output of the ultrasonic transducer in real time, and the sampling frequency can reach more than 10kHz.

[0059] The temperature field sensing unit 113 can be a non-contact infrared thermal imager or a thermocouple, used to acquire the transient temperature distribution of the bonding region.

[0060] The vision sensing unit 114 can be a high-resolution, high-frame-rate (e.g., ≥30fps) industrial camera equipped with a high-magnification lens and a professional imaging system, used to acquire pad positioning images, solder ball morphology images, and solder joint quality images after bonding.

[0061] The environmental monitoring unit 115 is used to monitor environmental parameters such as temperature and humidity in the workshop where the equipment is located, providing a basis for process compensation.

[0062] In one specific embodiment, the physical execution unit 13 is a mechanism that implements the bonding action and is responsible for executing instructions from the controller. Specifically, it includes an XY platform drive mechanism 131, a Z-axis actuator 132, an ultrasonic transducer 133, and a heating mechanism 134.

[0063] The XY platform drive mechanism 131 can be driven by a servo motor or a linear motor to carry out rapid and precise positioning of chips or substrates in the horizontal plane.

[0064] Z-axis actuator 132 is typically driven by a voice coil motor and is used to control the vertical up-and-down movement of the chopping blade and to apply precise bonding pressure.

[0065] The ultrasonic transducer 133 is used to apply controllable ultrasonic energy at the bonding interface to promote diffusion bonding between metal atoms.

[0066] The heating mechanism 134 is used to preheat the substrate to achieve the appropriate bonding process temperature.

[0067] It should be understood that the microsecond-level closed-loop controller 12 is the core of the device layer 1 control, typically composed of an FPGA (Field-Programmable Gate Array) or DSP (Digital Signal Processor) with a main frequency of no less than 500MHz. It is directly connected to the multi-source sensor array 11 and the physical execution unit 13, forming an ultra-low latency hardware control loop with a response time ≤100μs. Its main functions are: to receive and parse control commands (such as motion trajectory correction and process parameter fine-tuning) from the upper layer (edge ​​computing layer 2), and to combine this with real-time feedback from multiple sensing units in the multi-source sensor array 11 to precisely drive the physical execution unit 13 to complete actions, achieving motion trajectory tracking, multi-physics field coordination, and basic safety protection. It not only forms a bidirectional data interaction channel between the physical execution unit 13 and the microsecond-level closed-loop controller 12, but also provides data with high reliability and precise execution capabilities for data optimization in the fog computing layer 3 and the cloud computing layer 4.

[0068] Edge computing layer 2 Edge computing layer 2 is deployed in the embedded industrial computer or edge AI box of each wire bonding machine. It is the core of realizing real-time intelligent decision-making and local data processing. It communicates with device layer 1 at the microsecond level to ensure deterministic transmission of control commands.

[0069] like Figure 3As shown, edge computing layer 2 can achieve: (1) real-time preprocessing of data collected by device layer 1, including data denoising, format normalization, and feature extraction; (2) automatic correction of pad offset and system motion control based on feedback from visual sensors, realizing real-time monitoring of wire bonding machine; (3) during force / position switching, when there is no force sensor to measure bonding pressure, the lightweight artificial intelligence built by edge computing layer 2 can estimate the bonding contact force through a large amount of data and laboratory calibration data, and determine the force / position switching criteria; (4) setting up edge federated learning nodes based on the data leakage risk level (such as sensitive data involving product yield parameters). (5) The device runtime is non-sensitive data) or user-defined rules are used to classify the data for sensitivity. Sensitive data completes model training locally and only uploads parameter update information to fog computing layer 3. Non-sensitive data can be directly uploaded to fog computing layer 3. (6) Simultaneously build multi-sensor fusion for time alignment, construct a multi-physics coupling engine, so as to realize the dynamic optimization module of the bonding process, develop a microsecond-level anomaly detection algorithm to achieve adaptive adjustment of process parameters within 10ms, such as the temperature compensation value required when using bonding wires of different materials, and run a lightweight YOLOvx model to perform real-time quality inspection of solder ball deformation. (7) Generate real-time control decisions and send them to device layer 1 for execution.

[0070] It is worth mentioning that the main reason for adopting federated learning is that chip packaging process parameters are core corporate secrets. Federated learning can avoid the leakage of raw data, and the data distribution of equipment from different manufacturers (such as copper wire bonding machines and gold wire bonding machines) varies greatly. The localized training of federated learning can be compatible with heterogeneous data.

[0071] like Figure 4 As can be seen, the edge computing layer 2 mainly includes an edge data processing and modeling module 21, a real-time control and decision-making module 22, and an edge learning module 23. The edge data processing and modeling module 21, as the core of the edge computing layer 2, is mainly used to transform the massive amounts of raw sensor data from the device layer 1 into structured and information-rich feature data streams, and on this basis, build localized models for real-time analysis and prediction. This module provides high-quality input to the real-time control and decision-making module 22, which is a prerequisite for achieving millisecond-level intelligent decision-making. The real-time control and decision-making module 22 is the "decision-making brain" of the edge computing layer 2, responsible for making millisecond-level intelligent decisions based on the local model and real-time data generated by the edge data processing and modeling module 21, and generating real-time control commands to be sent to the microsecond-level closed-loop controller 12 of the device layer 1. This module ensures that the wire bonding process is fast and accurate.

[0072] The edge data processing and modeling module 21 includes a multi-source data real-time processing submodule 211, a multi-physics coupling engine submodule 212, and a lightweight AI quality inspection submodule 213.

[0073] The multi-source data real-time processing submodule 211 is responsible for receiving massive amounts of raw data from device layer 1, performing preliminary filtering, noise reduction, unit normalization, format conversion, and microsecond-level timestamp alignment, etc., to provide a high-quality data foundation for upper-level analysis.

[0074] The multiphysics coupling engine submodule 212 receives a time-aligned, feature-rich multidimensional data stream processed by the multi-source data real-time processing submodule 211, aiming to reveal the complex coupling relationships between various physical fields during wire bonding. Specifically, to address the challenge of resolving multiphysics coupling relationships using traditional techniques, the multiphysics coupling engine submodule 212 fuses feature data from multiple physical fields, including mechanical, acoustic, thermal, and visual fields, in real time through a high-precision time synchronization mechanism. In a preferred embodiment, the multiphysics coupling engine submodule 212 is based on a lightweight Physical Information Neural Network (PINN) model. This model not only learns correlations from the data but also embeds the physical equations of the bonding process (such as heat conduction equations and vibration equations) as regularization terms or constraints. This enables the engine to dynamically and accurately calculate and output real-time predicted values ​​for key solder joint quality indicators (such as solder joint deformation, gold ball diameter, and interface bonding strength) within ≤5ms. Based on the real-time comparison of this predicted value with the process threshold, the system can immediately generate fine-tuning instructions for ultrasonic power or bonding pressure, forming a millisecond-level response process parameter adaptive adjustment closed loop, and realizing online quality intervention.

[0075] The lightweight AI quality inspection submodule 213 utilizes machine vision for real-time online quality inspection. To achieve efficient inference on edge devices with limited computing power, this module runs a deeply optimized lightweight AI model (e.g., YOLOvx, lightweight Vision Transformer, or MobileNetV3-SSD). Specifically, the lightweight AI quality inspection submodule 213 performs real-time quality inspection on solder joint images acquired by the vision sensing unit 114, determining whether the solder joints have morphological defects (such as cracks, misalignment, dimensional abnormalities, etc.) within 10ms, achieving 100% online data inspection.

[0076] The real-time control and decision-making module 22 includes a vision-motion control linkage submodule 221 and a bonding process emergency control submodule 222.

[0077] To address the pain point of yield loss caused by passively adjusting motion control parameters only after solder joint quality defects appear in existing technologies, and the risk of insufficient bonding accuracy due to the lack of coordinated control when the vision, motion control, and structural subsystems operate independently, the vision-motion control linkage submodule 221 aims to improve bonding quality. It intervenes before defects occur by building a deep collaboration between visual perception and motion control.

[0078] Specifically, the vision-motion control linkage submodule 221 embeds a "solder joint feature-motion control parameter association model" in the edge computing layer 2. This model is trained based on historical process data of wire bonding. Its inputs include, but are not limited to: (1) core feature parameters of the solder joint, which are collected by the vision sensing unit 114 and extracted by the edge data processing and modeling module 21, such as the diameter of the gold ball, the arc height of the wire, and the flatness of the bonding interface; (2) current motion control parameters, such as bonding dwell time, motion control movement speed, and ultrasonic power output value; (3) material characteristic parameters, such as the material of the lead wire (gold wire / copper wire), hardness, and ductility. Its output is a dynamic mapping relationship of "solder joint feature change trend - motion control parameter adjustment threshold".

[0079] During training, the model continuously refines its mapping logic using iterative optimization algorithms such as gradient descent. This ensures that when a solder joint feature shows a trend approaching the process threshold boundary, the model can accurately output the corresponding adjustment direction and numerical range of the motion control parameters. For example, when the gold ball diameter feature data shows "continuously approaching the lower limit value," the model can output an adjustment suggestion of "extending the bonding dwell time by 0.1ms-0.3ms." Furthermore, the model supports dynamic updates of parameter weights based on different bonding scenarios (gold / copper wire bonding, different chip package specifications).

[0080] During the production process, the pre-adjustment of motion control parameters is carried out according to the following procedure: (1) High-frequency acquisition: The visual sensing unit 114 acquires the feature data of each completed solder joint in real time at a high frame rate (e.g., sampling frequency ≥ 100fps); (2) Trend analysis and risk prediction: After each solder joint is bonded, the characteristic data of the solder joint is immediately input into the "solder joint feature-operation control parameter correlation model". The model not only analyzes the state of a single solder joint, but more importantly, it analyzes the historical characteristic change pattern of multiple consecutive solder joints. For example, the model detected that the diameters of the gold balls of three consecutive solder joints were 25μm, 24.5μm and 24μm, respectively, showing a clear trend of gradually approaching the lower limit of the 23μm process. (3) Generate pre-adjustment instructions for early intervention: Based on the above trends, the model predicts the quality risk of the next solder joint. For example, when the model detects that the diameter of the gold ball shows an obvious trend of gradually approaching the lower limit of the 23μm process, it predicts that the diameter of the gold ball of the next solder joint is likely to be lower than the lower limit of the process, and there is a quality risk of poor soldering. The corresponding operation and control parameters are immediately generated as pre-adjustment instructions. (4) Early intervention: This instruction is directly transmitted to the vision-motion control linkage submodule 221, thereby adjusting relevant parameters in advance before the next solder joint bonding is performed. For example, when it is predicted that the small diameter of the gold ball is caused by insufficient bonding time due to excessive motion control speed, the vision-motion control linkage submodule 221 automatically increases the bonding dwell time from 1.0ms to 1.2ms, while keeping other parameters such as ultrasonic power and bonding pressure stable, thereby achieving "early intervention before defects occur", which is fundamentally different from the passive mode of waiting for defects to appear and then stopping the machine for adjustment in the traditional solution.

[0081] For complex bonding scenarios with dense pads and high stress sensitivity, such as ball grid array (BGA) packages, edge computing layer 2 can link with device layer 1 to form a three-in-one control system of "vision-motion control-structure". The specific mechanism is as follows: (1) Visual guidance: The visual sensing unit 114 of the device layer 1 acquires the pad array data of the BGA package through 3D visual scanning technology (such as laser triangulation) and generates a 3D pad distribution map containing the pad position, spacing and edge stress concentration area. The preferred bonding area and the avoidance area are marked on the 3D pad distribution map. The edge stress concentration area, also known as the avoidance area, is the area within 5μm from the edge of the pad where the stress value is higher than 30% of the center area of ​​the pad. (2) Operation and control execution: After receiving the 3D image, the vision-operation and control linkage submodule 221 uses a path planning algorithm (such as A) to perform the operation. The algorithm generates an optimal bonding path that actively avoids avoidance zones located at the edges of the pads and optimizes the bonding order between priority bonding zones in adjacent pads (e.g., progressing from the center of the array to the edge) to reduce the impact of device vibration on the bonded solder joints. (3) Stable structure: While issuing path planning instructions, the vision-motion control linkage submodule 221 simultaneously transmits path change information (such as the number of path inflection points or the distance of movement) to the physical execution unit 13 (such as the fixture control mechanism) of the device layer 1. The fixture control mechanism dynamically adjusts the support force of the fixture based on this information (for example, when there are many path inflection points, the clamping force of the fixture is increased from 50N to 60N to reduce the risk of large-amplitude vibration when the equipment moves), so as to suppress the risk of reduced welding point accuracy caused by equipment vibration due to changes in motion control path, and finally achieve a dual improvement in bonding quality and efficiency in complex scenarios.

[0082] In response to the high-risk nature of wire bonding, where "broken bonds or poor soldering can lead to chip scrapping", the bonding process emergency control submodule 222 has constructed an emergency and fault-tolerant mechanism that is different from general edge control and is specific to the bonding process, to ensure high reliability of the production process.

[0083] First, the bonding process emergency control submodule 222 can realize emergency cutoff and backtracking of process anomalies to deal with sudden anomalies that pose a serious threat to product quality or equipment safety. The specific workflow includes: (1) Real-time monitoring of key abnormal conditions during the bonding process, such as sudden drop in ultrasonic power, solder joint deformation exceeding the process threshold, abnormal Z-axis movement of the cleaver, etc. (2) When the above-preset serious abnormality is detected, the bonding process emergency control submodule 222 first performs an instantaneous assessment of the abnormality risk level through the built-in "process risk priority judgment logic". Based on the assessment result, if it is determined to be a high-risk event, the physical execution unit 13 of the equipment layer 1 is immediately triggered to perform a microsecond-level emergency shutdown operation to prevent the defect from expanding or causing equipment damage. (3) Upon triggering the shutdown, the bonding process emergency control submodule 222 automatically backtracks, extracts, and stores various data transmitted by the multi-source sensor array 11 within 10ms before the anomaly occurs, including but not limited to ultrasonic energy change trends, bonding area temperature curves, and cutter movement trajectories, generating a process anomaly tracing report containing changes in key parameters before and after the anomaly. This provides engineers with crucial evidence to quickly and accurately locate the root cause of the problem, far superior to traditional solutions that only execute a single alarm action.

[0084] Secondly, the bonding process emergency control submodule 222 allows for bonding process fault-tolerant algorithms, designed to intelligently handle typical minor process deviations, avoid unnecessary downtime, and improve production continuity. Its workflow includes: (1) The bonding process fault tolerance algorithm can identify typical minor abnormal conditions that occur during the bonding process; (2) Based on conventional position correction, for example, when the pads are slightly offset during gold wire bonding, the bonding process fault tolerance algorithm can combine the characteristics of the bonding material (such as the ductility of the gold wire) and the current process parameters (such as ultrasonic power) to calculate and output a multi-parameter collaborative adjustment compensation strategy in real time through the built-in model. For example, for a small pad offset, the bonding process fault tolerance algorithm can simultaneously output a precise pad position offset compensation amount and a precise ultrasonic power fine-tuning value.

[0085] It is worth mentioning that this bonding process fault-tolerant algorithm, which uses multi-parameter collaborative adjustment, reduces the risk of new process defects (such as excessive lead stress) that may be caused by over-adjustment due to a single parameter (such as adjusting only the position). It enhances the fault tolerance capability of the general control scheme in edge computing layer 2 for the bonding process, effectively reduces the chip scrap rate under high-risk conditions, and improves the robustness of the overall wire bonding system.

[0086] Edge learning module 23 is a core component of the distributed federated learning architecture. It aims to achieve collaborative training and evolution of models while ensuring the absolute security of core process data. Its function is mainly achieved through the following steps: (a) Data Sensitivity Classification: The edge learning module 23 classifies the data processed by the edge data processing and modeling module 21 according to preset rules or user-defined strategies. For example, process parameters that are directly related to product yield and involve the core competitiveness of enterprises (such as ultrasonic power curves, bonding pressure, and bonding time of specific materials) and high-resolution solder joint images are classified as confidential process data (i.e., sensitive data), while equipment running time, ambient temperature and humidity, vibration data of non-critical components are classified as non-confidential data. (b) Local Model Training and Fine-tuning: The edge learning module 23 uses locally stored sensitive process data to train or periodically fine-tune various intelligent models deployed in the edge computing layer 2. For example, it can use local solder joint images and process parameter data to optimize the model of the lightweight AI quality inspection submodule 213, or iterate the associated model in the vision-operation control linkage submodule 221. This localized training method not only ensures that sensitive data does not leave the equipment, but also enables the model to adapt highly to the specific operating conditions of the current equipment (such as parameter drift caused by component wear), effectively solving the problem of large differences in data distribution between different equipment (such as copper wire bonding machines and gold wire bonding machines). (c) Secure parameter upload: After completing one round of local training, the edge learning module 23 only extracts the updated model parameters (such as the weights and biases of the neural network) or gradient information. This parameter information is encrypted before being uploaded and then sent to the plant-level collaborative learning and model management module 32 of the fog computing layer 3 in the form of encrypted gradients. This "parameter-only, non-data" upload mechanism fundamentally eliminates the risk of leakage of original process data and significantly reduces the network transmission bandwidth usage. (d) Model reception and update: The edge learning module 23 is configured to receive aggregated and optimized model parameters from the upper layer (fog computing layer 3 or cloud computing layer 4). After receiving new global or factory-level model parameters, the edge learning module 23 will use them to update the local model, thereby integrating the model parameters of multiple devices or even multiple factories into the local model, realizing collaborative intelligence under the premise of protecting privacy.

[0087]

Fog Computing Layer 3

[0088] Fog computing layer 3 can achieve: (1) regional collaboration and access, serving as a bridge between edge computing layer 2 and cloud computing layer 4, enabling cross-device data and model collaboration among multiple heterogeneous devices in the factory area, and breaking the data silo phenomenon through multi-protocol gateways; (2) local federated aggregation, aggregating model parameter updates from multiple edge computing layers 2, and generating a factory-level model with better performance under the premise of protecting data privacy; (3) carrying high-level intelligent applications, including a rapid process design engine for new product introduction, production line-level quality process control and production energy consumption management; (4) dynamically deploying according to production scale through an elastic activation mechanism, and providing key parameter set support for the "zero-sample debugging" of new equipment.

[0089] In some embodiments, the edge computing layer 2 pre-sets a baseline parameter set for rapid startup when a new device is connected; the fog computing layer 3 maintains a debugging parameter pool that aggregates optimized process parameters of equipment already stably operating within the plant; when a new device is connected, the baseline parameter set is loaded first to achieve rapid startup, enabling operation within ten minutes of power-on. Subsequently, the system automatically collects the initial bonding data of the new device, and through transfer learning and similarity matching algorithms, automatically transfers the optimized process parameters that best match the initial operating conditions of the new device from the debugging parameter pool, and performs compensation and fine-tuning based on the unique deviations of the new device to achieve rapid and high-precision debugging.

[0090] In one implementation, the system includes a flexible activation mechanism, such as Figure 5 As shown, this mechanism dynamically adjusts the system architecture based on factors such as the number of wire bonding machines, daily output, or process complexity within the factory. Specifically, when the first preset condition is met, the edge computing layer 2 is directly connected to the cloud computing layer 4; when the second preset condition is met, the fog computing layer 3 is activated and deployed to enable factory-level collaboration between the edge computing layer 2 and the cloud computing layer 4.

[0091] Furthermore, this elastic activation mechanism is configured to activate only a single fog node to handle plant-level collaborative optimization when the second preset condition is met but the third preset condition is not. In other words, for small-scale production line applications, the fog computing layer 3 collaborative function is deployed to high-performance edge nodes (such as NVIDIA Jetson AGX); for ultra-large-scale deployments, regional cloud nodes are added above the fog computing layer 3 to achieve hierarchical federated aggregation. When both the second and third preset conditions are met, a distributed fog cluster is enabled to achieve broader regional collaborative control.

[0092] In some embodiments, the first preset condition may be that the number of devices N ≤ 3, or that the device is in a research and development laboratory or a small-batch production scenario. In this case, fog computing layer 3 is disabled, and edge computing layer 2 is directly connected to cloud computing layer 4, which is to say, an "edge-cloud" two-level architecture is adopted. The computing power allocation of edge nodes is dynamically adjusted, such as by adding edge cluster load balancing to achieve small-scale device collaboration. However, its response speed and collaboration efficiency are weaker than the solution when fog computing layer 3 is enabled. The second preset condition may be that the number of devices is N+1 to M, such as 4 to 10, and the daily output is ≥ 100,000 solder joints or mixed production ≥ 2 types of packaging. In this case, a single fog computing node is activated. The third preset condition may be that the number of devices is > M, such as more than 10, or there is a need for cross-factory collaboration. In this case, a distributed fog cluster is enabled to achieve regional collaborative control over a larger area. Among these, parameters such as the number of devices, daily output, and mixed production packaging types can be collaboratively determined based on production costs, preparation efficiency requirements, and specific user requirements.

[0093] Among them, such as Figure 6 As can be seen, this application proposes a process-driven layered architecture. In the equipment layer 1, microsecond-level real-time control is achieved, primarily handling the control of bonding parameters such as temperature, motion control algorithms, and ultrasonic power. In the edge computing layer 2, millisecond-level optimization is used, mainly constructing equipment health assessments such as solder joint quality detection models. In the fog computing layer 3, 100ms-level collaboration is implemented, achieving multi-machine parameter synchronization, spare parts early warning, and parameter drift compensation. Furthermore, the fog computing layer 3 employs a dynamic activation mechanism based on the number of workstations and the amount of bonding. In the cloud computing layer 4, second-level analysis is performed, conducting long-term process analysis to provide process optimization suggestions and lifetime predictions, ultimately achieving layered quality control.

[0094] like Figure 7 It can be seen that the fog computing layer 3 includes a fog layer data and communication management module 31, a factory-level collaborative learning and model management module 32, and a factory-level intelligent application module 33.

[0095] The fog layer data and communication management module 31, serving as the data and communication hub of the fog computing layer 3, is a key infrastructure for achieving plant-level equipment collaboration and breaking down data silos. Its core function is to solve the access challenges of multi-source heterogeneous devices, enabling effective aggregation and secure data flow within the plant area, and providing a unified, secure, and reliable data foundation for upper-layer collaborative learning and intelligent applications. This module ensures that the data stream from the edge computing layer 2 is processed in a standardized manner, while also guaranteeing the stability and security of the communication links between the edge computing layer 2 and the cloud computing layer 4.

[0096] The plant-level collaborative learning and model management module 32 receives model parameter updates uploaded by multiple edge learning modules 23 and uses the FedAvg algorithm for local model aggregation. To improve model convergence under non-independent and identically distributed data, more advanced aggregation algorithms such as FedProx (adding proximal term constraints) or SCAFFOLD (gradient correction mechanism) can also be selected. The aggregated plant-level model can be distributed to each edge computing layer 2 for updates, or uploaded to the cloud computing layer 4 to participate in global aggregation.

[0097] The factory-level intelligent application module 33 is the core value realization unit of the fog computing layer 3. By utilizing the data and models aggregated, processed and enhanced by the fog layer data and communication management module 31 and the factory-level collaborative learning and model management module 32, it provides the factory with advanced intelligent services directly related to production efficiency, quality, cost and flexibility.

[0098] The fog layer data and communication management module 31 includes a plant-level data aggregation and processing submodule 311, a multi-protocol gateway submodule 312, and a security agent submodule 313.

[0099] The plant-level data aggregation and processing submodule 311 is used to collect non-confidential data and model parameter updates from multiple edge computing layers 2 within the plant area, and to perform preliminary feature extraction and analysis.

[0100] The multi-protocol gateway submodule 312 is designed to address the incompatibility of communication protocols between devices from different manufacturers, such as... Figure 8 As shown, this submodule has built-in parsing and conversion capabilities for various mainstream industrial protocols (such as the CANopen-based Hesse HBN protocol, KS Modbus RTU, and the domestic EtherCAT). Through standardized information models such as OPC UA or MTConnect, it unifies heterogeneous device data into a standard format. Alternatively, it can extend support for protocols such as Profinet and EtherNet / IP through a general OPC UA gateway, or deploy protocol conversion middleware (such as Apache PLC4X) to achieve plug-and-play data access for devices from multiple vendors. If the OPC UA / MTConnect protocol is not used, it can be adapted to the proprietary bonding machine protocols of specific vendors through customized API interfaces.

[0101] The security agent submodule 313 adopts technologies such as integrated zero-knowledge proof verification mechanism to perform security verification and threat isolation for internal and external communications, and builds a dual protection system for privacy protection and external threat isolation.

[0102] The factory-level intelligent application module 33 includes a rapid process design engine submodule 331, a production line-level quality and process control submodule 332, and a production line-level production and energy consumption management submodule 333.

[0103] To address the industry pain points of high reliance on manual experience, long trial-and-error cycles, and significant material waste during new product introduction, the Rapid Process Design Engine submodule 331 provides an automated, model-based process design solution. Its workflow is as follows: First, quantifiable chip design parameters (e.g., chip size, pad layout, lead material type, wire diameter, etc.) are input into the Rapid Process Design Engine submodule 331. Then, the Rapid Process Design Engine submodule 331 automatically generates a complete and directly deployable set of recommended process parameters through an embedded process rule base built from historical data and expert knowledge, and a machine learning model (this model can be a local optimization model aggregated by the plant-level collaborative learning and model management module 32, or a global model issued by the cloud computing layer 4). This parameter set includes not only basic process parameters (such as bonding pressure, ultrasonic power, and bonding time), but also precise cutting tool motion trajectories and time-varying ultrasonic power curves generated for specific chip layouts.

[0104] A key feature of the rapid process design engine submodule 331 is its deep integration of digital twin technology. After generating recommended process parameters, instead of directly sending them to the physical equipment, it first constructs a high-fidelity digital twin synchronized with the physical wire bonding machine in a virtual environment. The rapid process design engine submodule 331 uses this digital twin to pre-simulate the entire wire bonding dynamic process using the recommended parameters. Through the built-in physical simulation engine, the system can calculate and detect various potential risks that may occur during the virtual bonding process in real time, such as collisions between wires, interference between the cleaver and the chip surface, loop collapse, and wire breakage. Based on this, it automatically corrects and optimizes the recommended process parameters until the pre-simulation process safely meets the preset quality indicators.

[0105] Through this closed-loop "design-simulation-optimization" process, the rapid process design engine submodule 331 significantly shortens the process development and verification process that traditionally relies on senior engineers for hours or even days to an automated intelligent decision-making process in minutes. This improves the production line's ability to respond quickly to new products and new packaging requirements, achieving first-time excellence in new processes while reducing physical risks in the wire bonding process.

[0106] In other words, compared to the real-time quality inspection of individual solder joints in the edge computing layer 2, the production line-level quality and process control submodule 332 focuses on macro-level quality management and process control at the production line level. The production line-level quality and process control submodule 332 utilizes massive amounts of data (including solder joint images, sensor time-series data, process parameters, etc.) aggregated from multiple edge nodes to construct a more complex and accurate quality prediction and diagnostic model than the edge layer model, thereby improving the accuracy and reliability of the recommended process parameters provided.

[0107] It is understandable that the production line-level quality and process control submodule 332 can analyze the overall quality trend of the production line in real time. For example, through the visual sensing unit 114, it analyzes the image features of bonding points (such as the shape of gold balls and the flatness of the bonding interface) in real time. Various sensors analyze data such as pressure fluctuations and temperature curves to predict product quality and thus detect potential defects such as cold solder joints and cracks in advance. In other words, the production line-level quality and process control submodule 332 can identify systematic quality fluctuations caused by specific batches of materials, specific environmental changes, or parameter drifts common to multiple devices. It can detect potential systematic defects such as cold solder joints and cracks in advance, issue warnings to engineers, and even automatically trigger production line-level process parameter adjustments or equipment calibration commands to achieve closed-loop control of the production process and ensure long-term stability of product quality.

[0108] In one specific embodiment, the intelligent management system for wire bonding machines also includes a dynamic matching library of material batches and optimal process parameters. This library stores the mapping relationship between different batches of wire bonding materials (gold wire, copper wire) and their corresponding optimal process parameters. Its working mechanism is as follows: (1) When the factory changes to a new batch of lead wire materials (such as gold wire or copper wire), the edge computing layer 2 collects the basic physical parameters of the new batch of materials through the sensors (such as hardness sensor or ductility tester) mounted on the equipment layer 1 or by manual input by the operator, and uploads the parameters to the fog computing layer 3 in encryption. (2) After receiving the material parameters, the fog computing layer 3 uses a built-in similarity matching algorithm (such as the K-nearest neighbor algorithm) to search for historical batches of materials that are highly similar to the new batch of material parameters (e.g., ≥90%) in the dynamic matching library. (3) Once a match is successful, the system will extract the verified optimal process parameters corresponding to the historical batch of materials. For example, if a copper wire batch with similar hardness is found, an ultrasonic power of 85W and a bonding pressure of 12N should be used. The optimal process parameters will be sent to the edge computing layer 2. (4) Edge computing layer 2 performs trial bonding verification based on the parameters issued. If the quality meets the requirements (e.g., solder joint qualification rate ≥ 99.5%), the parameters are applied directly. If not, fine-tuning is performed, and the optimal parameters after final verification are fed back to fog computing layer 3 for updating the dynamic matching library.

[0109] The production line-level production and energy consumption management submodule 333 is responsible for the intelligent scheduling of production tasks and the refined management of energy consumption for the entire wire bonding machine cluster in the plant. In other words, regarding production scheduling, the production line-level production and energy consumption management submodule 333 comprehensively considers order requirements (such as product type, quantity, and delivery date), real-time equipment status (such as health level and whether it is under maintenance), process parameters, and material inventory to establish a production scheduling optimization model. Through this model, the system can rationally arrange the sequence of production tasks and their allocation among different machines, aiming to achieve multiple optimization goals, such as minimizing changeover time, maximizing overall equipment efficiency (OEE), and ensuring on-time order delivery.

[0110] In terms of energy management, the production line-level production and energy management submodule 333 combines energy monitoring data collected from various devices to analyze the energy consumption characteristics of the equipment under different operating conditions (such as different bonding speeds and different process parameters). Based on these analyses, the system can optimize the production process and equipment operating parameters to reduce the energy cost per unit product. For example, by dynamically adjusting the start-up and shutdown sequence and operating speed of multiple bonding machines, peak shaving and valley filling of electricity load can be achieved while meeting delivery cycle requirements, thereby reducing the factory's peak power demand and overall energy consumption.

[0111] [Cloud Computing Layer 4] Cloud layer 4 uses cloud server clusters as hardware carriers and is deployed in remote data centers of equipment manufacturers or third-party cloud service providers. It is the global intelligent hub of the system and provides global computing, storage and analysis capabilities for the wire bonding system.

[0112] like Figure 27 It can be seen that the cloud computing layer 4 includes a global federated learning aggregation module 41, a global process knowledge and optimization module 42, an equipment lifecycle management module 43, a global quality and process control module 44, and a cross-plant production and energy consumption management module 45.

[0113] The Global Federated Learning Aggregation Module 41 is the core of global intelligent decision-making and the command center for distributed learning in the entire cloud-fog-edge-device collaborative intelligent management system. It is responsible for coordinating, managing, and executing global model training tasks across multiple physical locations (e.g., factories in different cities). In other words, the Global Federated Learning Aggregation Module 41 receives model parameters that have been locally aggregated and updated from one or more fog computing layers 3, performs higher-level global model aggregation and training, and ultimately generates a globally optimized model with industry-wide applicability and high accuracy.

[0114] The Global Process Knowledge and Optimization Module 42 is the system's "process brain," designed to overcome the core challenges in wire bonding, such as data silos caused by the complex interplay of materials, equipment, and process parameters, and the difficulty in digitizing and scalably reusing expert experience. This module transforms massive, scattered process data into a structured, reasonable, and continuously evolving global knowledge system, and provides intelligent process optimization and rapid adaptation capabilities based on this.

[0115] In some embodiments, the global federated learning aggregation module 41 is implemented as follows: I. Coordination and Scheduling of Training Tasks The global federated learning aggregation module 41 actively initiates and manages the entire lifecycle of federated learning. In a typical training process, it performs the following operations: selects or generates an initial global model (e.g., a model pre-trained based on historical data, or a randomly initialized model); broadcasts the current global model and the configuration of the current training task (e.g., training epochs, learning rate, and other hyperparameters) to all participating fog computing layers 3; sets a time window or waiting condition to receive local aggregation model parameter updates uploaded by each fog computing layer 3.

[0116] II. Secure Global Parameter Aggregation The global federated learning aggregation module 41 aggregates "wisdom crystallization" from multiple factories (represented by fog computing layers 3) while strictly protecting data privacy. In a preferred embodiment, the global federated learning aggregation module 41 employs a federated averaging algorithm and performs a weighted average of the received model parameter updates based on the number of devices or the amount of data represented by each fog computing layer 3, thereby calculating the next generation of global model parameters. To achieve the highest level of security, this module may rely on key technologies such as homomorphic encryption and differential privacy.

[0117] III. Global Model Update and Distribution After aggregation, the global federated learning aggregation module 41 generates a better-performing global model, which is then safely distributed back to all participating fog computing layers 3 so that they can be further distributed to their respective edge computing layers 2 to guide the next round of local training or be directly applied to production.

[0118] IV. Model Version Management and Performance Monitoring To ensure the stability and traceability of the system, the global federated learning aggregation module 41 also has the following functions: (1) The system will automatically record and manage each version of the global model, including its training configuration, participants, performance metrics and other metadata, which makes it easy for the system to roll back to the previous stable version or to perform A / B testing on different versions of the model; the global federated learning aggregation module 41 will also continuously monitor the performance metrics of the global model (such as accuracy on the validation set, loss function value, etc.) to determine whether the model has converged, which provides a basis for automatically adjusting the learning rate and deciding when to stop training, thus realizing the automation and intelligence of the training process.

[0119] In some embodiments, such as Figure 27 It can be seen that the global process knowledge and optimization module 42 includes a knowledge graph construction and management sub-module 421, a process parameter transfer learning sub-module 422, and a global process parameter intelligent optimization sub-module 423.

[0120] The core function of the knowledge graph construction and management submodule 421 is to construct and continuously evolve a three-dimensional dynamic knowledge graph of "materials-equipment-process". This knowledge graph is not a simple database, but a semantic network that depicts deep, implicit relationships between multi-dimensional elements. For example... Figure 9A and Figure 9B As shown, the nodes of this knowledge graph encompass material properties (such as lead type, wire diameter, melting point, hardness, ductility, resistivity, etc.), equipment attributes (such as equipment model, static positioning accuracy, dynamic aging curve, frequency response characteristics of ultrasonic transducers, wedge model, etc.), process parameter sets (such as specific parameter combinations such as bonding pressure, ultrasonic power, bonding time, temperature, and arc control parameters), and production process parameters (such as solder joint image features including gold ball diameter and interface flatness, motion control trajectory data, and corresponding quality results). Through this structured "material-equipment-process" three-dimensional dynamic knowledge graph, without aggregating the original sensitive data from each factory, only the graph feature vectors and model parameters generated by local training are aggregated, thereby continuously evolving and enriching this knowledge graph at the global level. Based on this system, intelligent parameter transfer learning is realized across materials and equipment. That is, when dealing with new materials (such as switching from gold wire to copper wire) or new equipment models, the system automatically maps the baseline process parameters by performing similarity matching and relationship reasoning in the spectrum, and superimposes compensation curves based on physical laws (such as temperature compensation for the higher oxidation of copper wire). Alternatively, the calibration parameters of high-precision new equipment can be back-inferred through the aging model and transferred to the old equipment to compensate for its accuracy drift. Ultimately, the debugging cycle of new processes is shortened from several hours of traditional manual trial and error to minutes.

[0121] The secure construction and continuous evolution of this knowledge graph are achieved through a federated learning architecture. The edge computing layer 2 and fog computing layer 3 of each factory are trained on local data to extract feature vectors or model parameters (i.e., digital representations of knowledge) of the graph, and then encrypted and uploaded. The knowledge graph construction and management submodule 421 only aggregates these encrypted parameters to update the global knowledge graph, thereby gathering industry-wide wisdom without accessing any original sensitive data, and continuously enriching and improving the knowledge graph.

[0122] The core function of the process parameter transfer learning submodule 422 is to apply knowledge graphs to achieve rapid adaptation of process parameters in new scenarios, solving the problems of time-consuming and material-intensive traditional trial-and-error debugging. Its workflow is as follows: Figure 10 As shown in (1) Scene recognition and knowledge retrieval, when the factory introduces new materials (e.g., switching from mature gold wire process to copper wire process) or new models of equipment, the system first identifies the key features of the new scene. Then, this submodule performs similarity matching and relational reasoning in the knowledge graph to find the historical success case that is closest to the new scene; (2) Benchmark parameter generation, the system extracts the verified set of process parameters from the matched knowledge graph nodes as the benchmark parameters or initial values ​​of the new scene; (3) Intelligent compensation and correction, this submodule will use the deep relationships in the knowledge graph for intelligent compensation. For example, when switching from gold wire to copper wire, the relationship in the knowledge graph indicates that "copper wire is better than gold wire". "The line is more prone to oxidation". The system will automatically overlay a compensation curve based on physical laws, such as increasing the bonding temperature or adjusting the protective atmosphere parameters accordingly. Another example is to transfer the calibration parameters of a high-precision new device to an old device after back-engineering its aging model in the knowledge graph, in order to compensate for the accuracy drift caused by long-term use. (4) Distribution and local fine-tuning: The generated initial optimization parameters are distributed to the edge computing layer 2. The edge computing layer 2 uses a small amount of sample data collected locally for fine-tuning. For example, the bottom feature extraction layer of the pre-trained model is frozen, and only the top classification layer is trained, so as to quickly converge and realize the combination of "global knowledge in the cloud" and "local personalized needs".

[0123] On the other hand, the process parameter transfer learning submodule 422 utilizes a transfer learning scheduling strategy to map the optimal process parameters between similar equipment, achieving dynamic production scheduling reconstruction. Based on a federated learning framework, it integrates anonymized production data and process experience from different enterprises to train a process parameter optimization model. Specifically, a general model is pre-trained in the process knowledge graph constructed in the cloud computing layer 4, and a feature-based transfer method is used to extract common features of process parameters across equipment. When new equipment or materials are added, the model parameters of similar operating conditions in the graph are used as initial values, and fine-tuning is performed using a small number of local samples in the edge computing layer 2 to achieve rapid adaptation of process parameters in new scenarios. This solves the problem of time-consuming traditional trial-and-error debugging, shortens the debugging cycle, and reduces material waste. A general process model is generated through cloud-based federated aggregation as a transfer source model; when new equipment is connected, this model is loaded and fine-tuned using multiple local samples (freezing the underlying feature layer) to achieve rapid zero-sample adaptation across equipment.

[0124] The function of the global process parameter intelligent optimization submodule 423 is to actively explore and discover the globally optimal combination of process parameters. For example... Figure 11 and Figure 27 As shown, this submodule employs advanced algorithms such as deep reinforcement learning to model the wire bonding process as a reinforcement learning environment. Deep reinforcement learning learns the optimal strategy (i.e., the rule for selecting the best process parameters under specific conditions) by "exploring" (trying different combinations of process parameters) within this environment and learning from the "rewards" obtained (such as comprehensive indicators like product yield, production efficiency, and energy consumption). This optimization process is driven by massive, multi-dimensional global production data aggregated through federated learning from cloud computing layer 4, ensuring the universality and robustness of the optimization results. The optimization results are not a fixed set of parameters, but a dynamic recommendation model that can generate suggestions for the optimal combination of process parameters under the current conditions in real time, based on specific input conditions (such as chip type, wire batch, and substrate material). Better combinations of process parameters discovered by this submodule will be verified and added back to the knowledge graph as new high-quality nodes, thus forming a virtuous cycle of "knowledge application - discovery of new knowledge - knowledge accumulation," continuously improving the intelligence level of the entire system.

[0125] The equipment lifecycle management module 43 achieves a closed-loop data system for the entire lifecycle of the wire bonding machine, from R&D and design to production, use, maintenance, and disposal. It employs a two-way data collaboration mechanism between the manufacturer and the user, and a full-stage data collaboration mechanism encompassing R&D, production, use, and disposal. In other words, during the R&D phase, data feedback optimizes equipment design (e.g., transducer frequency response); during the production phase, adaptive optimization of process parameters is achieved; during the use phase, predictive maintenance and remote diagnostics are implemented; and during the maintenance phase, spare parts prediction and maintenance knowledge base construction are achieved. To address the technical problems in existing wire bonding machine R&D processes, such as reliance on single-environment laboratory testing or lack of real-world multi-scenario operating condition data leading to insufficient adaptability between equipment design and actual application scenarios, [further details are needed]. Figure 12 , Figure 13 and Figure 14 As shown, the equipment lifecycle management module 43 further includes a health and stability prediction submodule 431, a performance degradation analysis and design feedback submodule 432, and an after-sales resource and cost optimization submodule 433.

[0126] The health and stability prediction submodule 431 constructs a joint prediction model for "equipment health - process stability", such as... Figure 12 and Figure 15 As shown, this model uses a temporal deep learning network (such as LSTM or Transformer) as its core, comprehensively analyzing historical equipment operation data and real-time sensor readings to accurately predict the remaining lifespan and health status of key components (such as blade wear and ultrasonic transducer performance degradation). Furthermore, the health and stability prediction submodule 431 dynamically correlates the predicted equipment health status indicators with real-time acquired process parameters (such as bonding pressure and ultrasonic energy transfer efficiency), thereby quantifying and diagnosing the drift patterns of process parameters caused by equipment performance degradation. Based on this mapping relationship, the system proactively sends adaptive process parameter compensation instructions to the edge computing layer 2 (e.g., increasing the ultrasonic power setting value proportionally based on the transducer efficiency degradation curve), thereby maintaining the quality stability of the bonding process even under the premise of component performance degradation, effectively extending the effective service life of the equipment, and realizing a shift from passive maintenance to proactive health management, thus achieving accurate prediction and proactive management of key components.

[0127] Performance degradation analysis and design feedback submodule 432, such as Figure 13 and Figure 27As shown, the performance degradation analysis and design feedback submodule 432 integrates operating data from different models and service years of equipment to establish a performance degradation curve model. By analyzing the differences between actual operating data and design standards, key factors affecting the long-term stability of the equipment are identified. These insights are stored in an anonymized and aggregated form in an anonymized fault database that integrates operating data of wire bonding machines from different users. Specifically, this includes, but is not limited to, lifespan data of core components (such as cutters and ultrasonic transducers) in multiple factory scenarios (e.g., the average lifespan and failure frequency of a certain model of cutter in 10 factories with different production environments), and correlation data between core component failures and operating environment parameters (temperature, humidity, production load) (e.g., the correspondence between the number of ultrasonic transducer failures and the workshop ambient temperature). All data is anonymized before being uploaded to the cloud computing layer 4, removing sensitive information such as customer identification and factory location, and is fed back to the R&D department of the equipment manufacturer to form an optimization closed loop of "use-design". Equipment manufacturers' R&D departments can access this anonymized fault database to obtain real-world operating data across multiple scenarios, replacing traditional single-laboratory environment test data. This data can be used for design optimization of next-generation wire bonding machines. For example, if the database shows that "a certain type of bonding machine experiences an ultrasonic power drift rate of 5% in a workshop with an ambient temperature >30℃, significantly higher than the 1% drift rate at room temperature," the R&D department can specifically improve the heat dissipation structure of the ultrasonic transducer for that model of bonding machine (such as adding micro-heat sinks or optimizing the heat dissipation duct), avoiding design defects caused by laboratory room temperature tests not covering high-temperature operating conditions. For instance, if data shows that a certain type of bonding machine experiences a significantly increased ultrasonic power drift rate in a high-temperature workshop, the R&D department can improve its heat dissipation structure accordingly.

[0128] Meanwhile, a process parameter-equipment structure mapping model can be constructed on cloud computing layer 4. This model takes into account process optimization data from multiple factories, specifically including data on parameter adjustments for different bonding processes (such as copper wire bonding and gold wire bonding) by different factories (e.g., increasing ultrasonic power by 15% to meet bonding quality requirements in copper wire bonding), and data on the impact of these parameter adjustments on the wear of equipment structural components (such as fixtures and guide rails) (e.g., fixture wear rate increases by 30% after increasing ultrasonic power). The model algorithm establishes the correlation between "process parameter requirements and equipment structure wear." The R&D department can adjust the equipment structure design based on this model. For example, for the process requirement of higher ultrasonic power in copper wire bonding leading to faster fixture wear, the original fixture material can be optimized from stainless steel to ceramic material, realizing "process requirements driving equipment structure design" and improving the adaptability and durability of the next-generation bonding machine for different bonding processes.

[0129] After-sales resource and cost optimization submodule 433, such as Figure 14 and Figure 27As shown, by combining fault prediction results and equipment geographical distribution information, operations research optimization algorithms (such as genetic algorithms and simulated annealing algorithms) are used to intelligently schedule maintenance personnel and spare parts inventory, thereby reducing after-sales service costs and improving response efficiency.

[0130] like Figure 27 It can be seen that the cloud computing layer 4 also includes a global quality and process control module 44 and a cross-plant production and energy consumption management module 45, which are global versions of the corresponding modules in the fog computing layer, such as... Figure 16 and Figure 17 As shown, the global quality and process control module 44 utilizes machine vision and sensor data to construct a quality prediction model, analyzes bonding point characteristics and process parameters in real time, predicts quality defects, and triggers parameter adjustments or equipment calibration to achieve closed-loop production control. The cross-plant production and energy management module 45 integrates order requirements, equipment status, and process parameters to establish a production scheduling optimization model and rationally arrange production tasks. Combined with energy consumption monitoring data, it optimizes equipment operating parameters and reduces energy costs. In other words, the global quality and process control module 44 and the cross-plant production and energy management module 45 utilize macro-level data from across factories to conduct higher-dimensional quality analysis, trend prediction, and group-level collaborative scheduling of production resources.

[0131] To address the pain point of long deployment and debugging cycles for new equipment, the intelligent management system of this application provides a zero-sample debugging mechanism based on hierarchical parameter reuse, which includes: Edge layer fast start: Edge computing layer 2 pre-sets a set of reference parameters for each new device based on the factory calibration data of the same batch of devices. The reference parameter set includes the minimum parameter combination to realize the basic bonding function. When a new device is connected, it can directly call the parameter set to complete the initialization and achieve "power on and run" within 10 minutes. Fog Computing Layer 3 High-Precision Debugging: Fog Computing Layer 3 maintains a "debugging parameter pool" that aggregates the optimized parameters of all stable-running equipment of the same model within the factory area. After the new equipment completes its initial debugging, Edge Computing Layer 2 uploads a small amount of trial bonding data to Fog Computing Layer 3. Fog Computing Layer 3 uses a parameter similarity matching algorithm to call the optimal parameter combination with the highest adaptability from the parameter pool, and combines it with transfer learning to generate compensation coefficients (such as assembly tolerance correction values) for the new equipment, and sends them to Edge Computing Layer 2 for fine-tuning. High-precision debugging can be completed within 30 minutes.

[0132] For batch debugging scenarios involving thousands of devices, Fog Computing Layer 3 can enable a cluster debugging mode. This means that after the first device is debugged, the resulting optimized parameters are anonymized and added to a parameter pool. During subsequent device debugging, the Fog Computing Layer directly calls these parameters for adaptation correction, reducing the debugging time for a single device to within 15 minutes, with a debugging accuracy deviation of ≤0.5%. Throughout the process, the original process parameters only flow between the plant's edge layer and the Fog Computing Layer, reducing the risk of data leakage in the cloud and enabling efficient collaborative debugging of batch devices through plant-level parameter aggregation, thus solving the debugging efficiency bottleneck during large-scale deployments.

[0133] In some embodiments, the data flow between different levels of this system adopts a layered optimized communication protocol stack. The data flow of the four-level cloud-edge-device framework is as follows: Figure 18 As shown. The uplink data flow is as follows: Sensor data from device layer 1 is extracted by edge computing layer 2, aggregated in the form of encrypted gradients by fog computing layer 3, and finally uploaded to cloud computing layer 4; Downlink control flow: Global model parameters from cloud computing layer 4 are distributed to edge computing layer 2 via fog computing layer 3, generating optimized control commands which are then sent to the actuators of device layer 1.

[0134] Specifically, in some embodiments, the communication between device layer 1 and edge computing layer 2 employs industrial real-time Ethernet (preferably EtherCAT or PROFINET IRT) or high-speed fieldbus (such as CANopen FD) to meet the extremely low latency and deterministic communication requirements of the μs-level control loop. It is worth noting that the uplink data of device layer 1 refers to the raw, high-frequency, low-latency data such as position, temperature, and visual data generated by the multi-source sensor array 11. Real-time control commands sent from edge computing layer 2 to device layer 1 include motion trajectory correction, emergency stop, and parameter fine-tuning.

[0135] The edge computing layer 2 and fog computing layer 3 communicate using a lightweight message queue protocol (preferably MQTT over TLS 1.3, QoS 1 or 2) or OPC UA PubSub mode to suit the low-bandwidth, asynchronous communication scenarios within the factory network. MQTT / OPC UA PubSub offers advantages such as publish / subscribe mechanisms, low bandwidth overhead, support for asynchronous communication and reconnection after disconnection, making it more suitable for the factory network environment between edge computing layer 2 and fog computing layer 3. TLS (Transport Layer Security) ensures transmission security, and OPC UA PubSub also provides a built-in information model for structured data transmission. Notably, the uplink data from edge computing layer 2 includes feature data, model gradient updates, non-sensitive data, and alarm information. Fog computing layer 3 sends optimized model parameters, collaborative scheduling instructions, and process templates to edge computing layer 2.

[0136] Between the fog computing layer 3 and the cloud computing layer 4, reliable enterprise-grade messaging protocols (such as AMQP 1.0 over TLS) or HTTPS RESTful APIs are used to ensure high reliability and strong security for transmitting critical business data and model parameters across the wide area network. Notably, the uplink data from the fog computing layer 3 includes locally aggregated model parameters, fused feature data, summarized alarms, and performance reports. The cloud computing layer 4 sends data to the fog computing layer 3, including globally optimized model parameters, process parameter recommendation strategies, remote control commands, and software update packages. The downlink control flow uses the same protocol as the uplink to ensure reliable, secure, and efficient delivery of commands.

[0137] The intelligent management system of this application achieves intelligent coverage of the entire wire bonding process lifecycle through a layered architecture of cloud, fog, and edge. Specifically, for the coordination of "vision-operation control-structure," the intelligent management system of this application designs a two-layer mechanism that is interconnected yet has different focuses: (1) The collaborative mechanism located in the edge computing layer 2 focuses on real-time quality control in the production process. It uses high-frequency sensor data to predict the quality trend of a single solder joint in milliseconds and uses the vision-motion control linkage submodule 221 to pre-adjust parameters, thereby achieving "real-time interception" and "proactive prevention" of quality defects.

[0138] (2) The collaborative mechanism located in the fog computing layer 3 focuses on process planning before the production of new products. Based on the chip design parameters, it automatically generates and verifies the complete process scheme in the virtual space through the digital twin technology of the rapid process design engine submodule 331, realizing the "minute-level deployment" and "first-time excellence" of the new process, fundamentally reducing the dependence on on-site debugging.

[0139] By setting up the aforementioned two-layer mechanism, one real and one virtual, one fast and one intelligent, together they constitute a closed-loop intelligent system covering the entire process from process design to production execution.

[0140] The intelligent management system of this application realizes a closed-loop data system covering the entire lifecycle of wire bonding machines, from R&D, production, use, maintenance to scrapping. During the R&D phase, data-driven design optimization is achieved through an anonymized fault database and a process-structure mapping model. During the production phase, a dynamic matching library of material batches and optimal process parameters can be built in the fog computing layer 3. When a new batch of materials is replaced, the parameters of the new material are collected through the edge layer, and the optimal process parameters of similar historical batches are retrieved from the matching library in the fog computing layer and distributed, achieving collaborative optimization of materials and processes. During the scrapping phase, a residual value assessment and tiered utilization scheme for core components is also constructed in the equipment lifecycle management module 43. The remaining value of core components (such as ultrasonic transducers) in scrapped equipment is assessed through a lifespan-performance degradation model. If the conditions for degrading use are met, the components are matched to users with low-precision packaging requirements through the platform, achieving resource recycling.

[0141] The intelligent management system of this application provides a complete set of management functions, such as... Figures 19 to 26 As shown, it includes: a real-time data acquisition module, an equipment status monitoring module, a remote control module, a data analysis module, an alarm system module, a process parameter management module, a historical data query module, and a user management module. These modules work together in a cloud-fog-edge-device architecture to provide users with comprehensive monitoring, management, and traceability capabilities.

[0142] (1) The real-time data acquisition module is equipped with sensors that collect data at a preset frequency and add data-sensitive tags. The edge computing layer 2 performs preliminary filtering, noise reduction, and format conversion on the data to ensure data accuracy and consistency, such as the fog computing layer 3 forwarding the data. Figure 19 The data processing flow is shown.

[0143] (2) The equipment status monitoring module is suitable for building an equipment status monitoring model based on local data in the edge computing layer 2 to judge the equipment operating status in real time; the fog computing layer 3 analyzes the overall operating trend by aggregating equipment data within the region. Combined with the global fault diagnosis knowledge obtained by federated learning, it accurately locates the fault type and cause, triggers hierarchical alarms, and the monitoring logic is as follows: Figure 20 As shown.

[0144] (3) Remote control module: Users initiate control commands through the cloud computing layer 4 platform, which are then executed through the edge computing layer 2. This covers operations such as process parameter adjustment, equipment start / stop, and calibration. Encrypted communication and two-factor authentication are used to ensure command security, and the equipment execution status is fed back in real time. It also supports the transmission of model parameter update commands. The control flow is as follows: Figure 21As shown, after logging into the cloud computing layer 4 platform, users select the target bonding machine in the factory's digital twin map and adjust the temperature / pressure / ultrasonic power values ​​or trigger start / stop and calibration commands through the parameter control panel. After the system enforces two-factor authentication (such as dynamic token combined with fingerprint verification), the command is sent to the edge computing layer 2 for execution via an encrypted channel, and the operation status is fed back to the platform's progress bar and logs in real time. Engineers can remotely intervene in faulty equipment to improve efficiency and avoid production line downtime.

[0145] (4) Data analysis module: Edge-fog-cloud collaborative modeling is implemented based on a federated learning framework. For process optimization, each enterprise's edge devices train models locally, fog computing layer 3 aggregates parameters, and cloud computing layer 4 generates a globally optimal process parameter recommendation model. For equipment fault prediction, multi-source data is integrated to train a high-precision prediction model, and the training effect and data contribution are displayed through a visual interface. The analysis architecture is as follows: Figure 22 As shown.

[0146] (5) Alarm System Module: Based on fault diagnosis knowledge optimized by federated learning, alarm rules and thresholds are dynamically adjusted. A tiered notification strategy (internal system notification or external notification via SMS / email) is adopted according to the severity of the fault and data sensitivity. The alarm mechanism is as follows: Figure 23 As shown.

[0147] (6) Process Parameter Management Module: This module combines a global process optimization model to provide users with accurate parameter recommendations. It supports user-defined parameter templates, and users participating in federated learning can automatically synchronize optimization suggestions and use the adjusted data for local model iteration. The management process is as follows: Figure 24 As shown.

[0148] (7) Historical Data Query Module: Supports multi-dimensional querying of equipment operation data, process parameters, alarm records, etc., by time, equipment number, product batch, etc. It specifically provides data querying related to federated learning (model training logs, parameter update records, data contribution), and displays it in diverse formats such as tables and charts. The query function includes... Figure 25 As shown, users can retrieve historical process parameters, alarm records, and federated learning data contributions through multi-dimensional filters (time / equipment ID / product batch); the results are visualized in the form of time series curves, defect heatmaps, etc., and it supports exporting federated audit reports containing model training versions and parameter update trajectories. It can complete the correlation analysis of tens of millions of data within 10 seconds, improving traceability efficiency by 80%.

[0149] (8) User Management Module: To implement user registration, login, and hierarchical permission management, new federated learning-specific permissions (data upload, model download, training participation) have been added. User operation logs are recorded for easy auditing and traceability. The permission management architecture is as follows: Figure 26As shown. Administrators set user permission levels (e.g., Operator / Engineer / Administrator) in the role configuration interface, and specify the detailed rules for each user permission level: for example, an "Operator" user can only view real-time data and basic parameters; an "Engineer" user can adjust process parameters and participate in local model training; and an "Administrator" user can configure federated learning strategies and cross-plant collaboration permissions. An approval process for permission changes is added (e.g., multi-level approval or automatic activation conditions), and users participating in federated learning are individually authorized to upload data and download models. The system automatically records user logins, key operations, and model call behaviors to a blockchain audit database to ensure traceability. This achieves fine-grained permission control for cross-enterprise collaboration, meeting ISO 27001 compliance requirements.

[0150] This application constructs a cloud-fog-edge-device collaborative fusion architecture, consisting of "real-time data acquisition at the device end, real-time control at the edge, dynamic collaboration at fog computing layer 3, and intelligent decision-making at cloud computing layer 4." Through layered data processing and distributed modeling architecture, it achieves microsecond-level real-time control of the bonding process at edge computing layer 2, eliminates cross-device data silos using multi-protocol gateways at fog computing layer 3, and breaks down reliance on manual experience with intelligent models at cloud computing layer 4, forming a full-stack technology closed loop. This enables local processing of sensitive data and collaborative sharing of non-sensitive data. While ensuring data privacy and security, it integrates multi-party data resources to improve the accuracy and versatility of models for equipment fault prediction and process optimization, and enhances equipment lifecycle management functions. This, in turn, improves wire bonding production efficiency, reduces enterprise operating costs, and enhances product quality stability. It provides a highly reliable, adaptive, and scalable intelligent bonding solution for semiconductor packaging.

[0151] This application addresses the unique challenges of intelligent manufacturing in the semiconductor wire bonding field, characterized by "microscale, multi-physics coupling, and high reliability." Compared to traditional general-purpose industrial control systems, wire bonding processes require micrometer-level positioning accuracy and millisecond-level action response, while simultaneously controlling multiple physical field parameters such as mechanical pressure, ultrasonic power, and temperature field. These precision requirements far exceed those of typical industrial scenarios. Furthermore, compared to traditional federated learning models, wire bonding data distribution is strongly bound to chip type, bonding material, and equipment model, necessitating federated learning guided by a process knowledge graph. Edge nodes must complete defect detection within 10ms, fundamentally differing from the universal algorithms of general frameworks. The intelligent management system in this application, compared to general-purpose industrial control models, focuses on the semiconductor packaging industry characteristics of wire bonding machines, including microsecond-level process response, strong multi-physics coupling, high sensitivity to process parameters, and a strong correlation between the lifespan of core components and the process. This involves multiple aspects, including system composition refinement, full lifecycle data closed-loop, and cross-subsystem linkage control.

[0152] The basic principles, main features, and advantages of this application have been described above. Those skilled in the art should understand that this application is not limited to the above embodiments. The embodiments and descriptions in the specification are merely the principles of this application. Various changes and modifications can be made to this application without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claims. The scope of protection claimed by this application is defined by the appended claims and their equivalents.

Claims

1. A smart management system for wire bonding machines integrating cloud-fog-edge-device collaboration, characterized in that, include: The device layer includes a multi-source sensor array, a physical execution unit, and a microsecond-level closed-loop controller integrated or mounted with the wire bonding machine. The multi-source sensor array is used to collect real-time multi-physics field data of the bonding process, the physical execution unit is used to perform the bonding operation, and the microsecond-level closed-loop controller is used to respond to control commands and drive the physical execution unit. An edge computing layer, deployed on the wire bonding machine, communicates with the device layer at the microsecond level. The edge computing layer includes: The edge data processing and modeling module is used to process the real-time multiphysics data and construct a local model representing the bonding process. The real-time control and decision-making module is used to make millisecond-level intelligent decisions based on the local model and generate real-time control commands to be sent to the microsecond-level closed-loop controller. The edge learning module is used to train models locally using classified process data and only upload updated model parameters. A fog computing layer, deployed on a local network, is used for factory-level collaboration of data and models from at least two of the edge computing layers; The cloud computing layer, deployed in a remote data center, is used to globally aggregate and analyze data and models from at least one of the fog computing layers, and to provide cross-domain knowledge services and full lifecycle management.

2. The intelligent management system for wire bonding machines according to claim 1, characterized in that, The real-time control and decision-making module of the edge computing layer also includes a bonding process emergency control submodule. The bonding process emergency control submodule is used to monitor abnormal operating conditions in the bonding process in real time. When a preset abnormal operating condition is detected, a microsecond-level emergency shutdown is performed at the device layer, and multi-source sensor data before the abnormal operating condition occurs is automatically traced back and stored for source tracing analysis.

3. The intelligent management system for wire bonding machines according to claim 1 or 2, characterized in that, The edge data processing and modeling module of the edge computing layer includes a multi-physics coupling engine submodule. The multi-physics coupling engine submodule fuses the multi-physics data through a high-precision time synchronization mechanism and dynamically calculates and outputs real-time predicted values ​​of key indicators of weld joint quality based on a physical information neural network model, so as to form a closed-loop control for adaptive adjustment of process parameters.

4. The intelligent management system for wire bonding machines according to claim 1, characterized in that, The real-time control and decision-making module of the edge computing layer includes a vision-motion control linkage submodule. This submodule, based on a solder joint feature-motion control parameter correlation model, predicts the quality risk of the next solder joint based on the current and historical image feature change trends, and generates motion control parameter pre-adjustment instructions before defects occur, thereby achieving proactive prevention of bonding quality problems. The edge computing layer also pre-sets a benchmark parameter set for rapid startup when new equipment is connected. The fog computing layer maintains a debugging parameter pool, which aggregates the optimized process parameters of equipment that is already running stably in the plant area. When new equipment is connected, the system automatically migrates the optimized process parameters that best match the initial operating conditions of the new equipment from the debugging parameter pool, and performs compensation fine-tuning based on the unique deviations of the new equipment, thereby achieving rapid and high-precision debugging.

5. The intelligent management system for wire bonding machines according to claim 1, characterized in that, The fog computing layer includes a factory-level intelligent application module, which includes a rapid process design engine submodule. The rapid process design engine submodule automatically generates a recommended set of process parameters by inputting chip design parameters and combining a process rule library and a machine learning model. The bonding process is rehearsed in a virtual environment using digital twin technology to detect and mitigate physical risks, thereby optimizing the recommended set of process parameters.

6. The intelligent management system for wire bonding machines according to claim 1, characterized in that, The system also includes a flexible activation mechanism configured to: dynamically adjust the system architecture based on the number of wire bonding machines or the complexity of the process; when a first preset condition is met, the edge computing layer is directly connected to the cloud computing layer; when a second preset condition is met, the fog computing layer is activated and deployed to enable factory-level collaboration between the edge computing layer and the cloud computing layer; the data flow between the various layers of the system adopts a layered optimized communication protocol stack: the device layer and the edge computing layer use industrial real-time Ethernet or high-speed fieldbus protocols; the edge computing layer and the fog computing layer use lightweight message queue protocols or OPC UA PubSub mode; and the fog computing layer and the cloud computing layer use enterprise-level messaging protocols or HTTPS RESTful APIs.

7. The intelligent management system for wire bonding machines according to claim 1, characterized in that, The cloud computing layer includes a global process knowledge and optimization module, which is configured to: construct and maintain a three-dimensional dynamic knowledge graph of materials, equipment and processes; Through a federated learning architecture, model parameters uploaded by each fog computing layer are aggregated to continuously evolve the knowledge graph without accessing the original classified process data. Based on the knowledge graph, transfer learning is used to quickly adapt process parameters for new materials or new equipment scenarios. The cloud computing layer also includes an equipment lifecycle management module, which is configured to: construct a joint prediction model for equipment health and process stability, dynamically correlate the predicted health status of key equipment components with process parameters, and issue adaptive process parameter compensation instructions. In addition, a performance degradation analysis and design feedback model is constructed to feed back anonymized fault and process data from multiple plants to the equipment R&D and design end, forming an optimization closed loop of "use-design".

8. A method for intelligent management of wire bonding machines integrating cloud-fog-edge-device collaboration, characterized in that, Includes the following steps: At the device layer, a multi-source sensor array is used to collect multi-physics field data of the bonding process in real time, and a microsecond-level closed-loop controller is used to execute control commands from the edge computing layer. At the edge computing layer, the multiphysics data is processed and modeled in real time, and millisecond-level intelligent decision-making is performed based on the local model to generate real-time control commands. After training the model using local classified process data, only the model parameters are uploaded for updates. In the fog computing layer, model parameter updates from at least two of the edge computing layers are aggregated and collaboratively optimized at the plant level. At the cloud computing layer, the aggregated model from at least one of the fog computing layers is globally trained and analyzed, and cross-domain knowledge services and full lifecycle management are provided.

9. The intelligent management method for wire bonding machines according to claim 8, characterized in that, The intelligent decision-making steps in the edge computing layer also include: real-time monitoring of abnormal operating conditions during the bonding process; triggering the device layer to perform microsecond-level emergency shutdown when a preset severe abnormality is detected; and automatically backtracking and storing multi-source sensor data before the abnormality occurred. The method also includes a zero-sample debugging step: when a new device is connected, a set of benchmark parameters pre-installed in the edge computing layer is first loaded to achieve rapid startup; then, optimized process parameters matching the initial operating conditions of the new device are intelligently retrieved and migrated from the debugging parameter pool deployed in the fog computing layer; finally, the migrated parameters are compensated and fine-tuned in combination with the unique deviations of the new device to complete high-precision debugging.

10. The intelligent management method for wire bonding machines according to claim 8, characterized in that, The steps for collaborative optimization in the fog computing layer include a rapid process design step: receiving chip design parameters as input and automatically generating a preliminary set of recommended process parameters; The bonding process using the recommended set of process parameters is rehearsed in a virtual environment using digital twin technology; Based on the pre-simulation results, potential physical risks are detected and corrected to output an optimized set of process parameters. The steps of global training and analysis in the cloud computing layer include the steps of constructing and applying a knowledge graph: through a federated learning architecture, the model parameters uploaded by each fog computing layer are aggregated to continuously evolve a three-dimensional dynamic knowledge graph of materials, equipment and processes without touching the original classified process data. When new materials or equipment are introduced, relational reasoning and similarity matching are performed based on the knowledge graph, and transfer learning is combined to generate initial process parameters for rapid adaptation of the edge computing layer.

Images