Data-driven semiconductor factory equipment risk early warning method, system and device

Abstract

The application discloses a data-driven semiconductor factory equipment risk early warning method, system and equipment. The method first obtains the operation parameters, working conditions and health state records of equipment from historical data to train a personalized performance prediction model. Then, future production scheduling and weather forecasts are obtained and converted into future predicted working condition of the equipment. Next, by situation deduction, the future working condition is input into the trained model to calculate the predicted sequence of key operation parameters of the equipment, and then the operation risk is evaluated and risk warning events and their occurrence probabilities are generated. Finally, specific warning information and operation suggestions are generated and output according to the deduction results. The application realizes the transformation from passive alarm to active early warning by constructing a personalized model and actively integrating future multi-source information for forward-looking deduction.

Description

Technical Field

[0001] This invention belongs to the field of semiconductor plant management technology, and in particular relates to a data-driven method, system and equipment for early warning of risks in semiconductor plant equipment. Background Technology

[0002] In the semiconductor manufacturing industry, the stable operation of plant facilities (such as ultrapure water systems, refrigeration units, and waste gas treatment systems) is the cornerstone of ensuring production yield and continuity. Currently, the industry generally relies on Supervisory Control and Data Acquisition (SCADA) systems and Distributed Control Systems (DCS) for monitoring critical plant equipment. These systems set fixed safety threshold ranges for various key operating parameters of the equipment (such as temperature, pressure, and flow rate). When sensors detect real-time parameter values ​​exceeding these preset ranges, an alarm is triggered, notifying maintenance personnel for handling. However, this fixed threshold-based monitoring method has significant shortcomings. First, it is essentially a passive response mechanism, only issuing alarms after abnormal equipment parameters have occurred, failing to provide early warnings before failures or performance degradation occur. For modern semiconductor factories pursuing "zero downtime" and preventative maintenance, this lag can lead to production interruptions or wafer losses. Second, fixed alarm thresholds cannot reflect the true health boundaries of equipment that dynamically change during actual operation due to load variations, equipment aging, and especially drastic fluctuations in the external environment (such as seasons and weather). For example, in extreme heat, the efficiency of cooling systems naturally decreases. Using fixed thresholds will fail to identify potential risks caused by weather changes, where equipment parameters are approaching their true operating limits. Furthermore, equipment operating data, weather forecasts, and production plan data are typically stored in isolated information systems, lacking effective correlation analysis and deep reasoning. This prevents the system from answering forward-looking questions such as "What risks might the equipment face in the next few hours under specific weather and production plans?" Therefore, the main problem with existing technologies is that fixed-threshold-based monitoring systems can only issue alarms after an anomaly occurs; they cannot proactively and individually predict and warn of future operational risks by integrating multi-source dynamic information such as historical equipment performance, future weather, and production plans before an anomaly occurs. Summary of the Invention

[0003] The technical problem to be solved by this invention is to address the shortcomings of the prior art by providing a data-driven method, system, device and medium for risk early warning of semiconductor plant equipment. By integrating historical performance data of equipment to build a personalized prediction model, and actively incorporating multi-source dynamic information such as future weather and production plans for forward-looking extrapolation, the invention achieves a shift from passive alarm to proactive early warning, solving the problem that the prior art cannot perform personalized risk prediction before anomalies occur.

[0004] The first aspect of this invention discloses a data-driven method for early warning of risks in semiconductor plant equipment, comprising the following steps: Historical data acquisition steps: Obtain multiple sets of historical data records from the historical operation and maintenance database of the plant equipment. Each set of historical data records includes the measured values ​​of key operating parameters of the plant equipment at a historical moment, the historical operating conditions corresponding to the historical moment, and the historical health status identifier that characterizes the health level of the plant equipment at the historical moment. Personalized prediction model training steps: Through the device profile intelligent agent, a personalized performance prediction model is trained based on the multiple sets of historical data records. The combination of the historical operating conditions and the historical health status identifier is used as the model input features, and the measured values ​​of the key operating parameters are used as the target labels for model training. The model learns the mapping relationship from the input features to the target labels through machine learning algorithms. Future Condition Acquisition and Adaptation Steps: Acquire production scheduling information and weather forecast data for the future target period, and convert the production scheduling information into predicted load conditions for the plant equipment during the future target period, and convert the weather forecast data into predicted environmental conditions for the plant equipment during the future target period; the predicted load conditions and the predicted environmental conditions together constitute the future operating conditions of the plant equipment during the future target period. Equipment operation risk simulation steps: Through the situation simulation intelligent agent, the predicted load conditions, the predicted environmental conditions, and a preset simulation health status identifier are input into the personalized performance prediction model to calculate the predicted sequence of key operating parameters of the plant equipment in the future target period, and evaluate the equipment operation risk based on the predicted sequence of key operating parameters to generate at least one risk warning event and its corresponding probability of occurrence. Early warning decision generation and output steps: The decision-making and early warning intelligent agent generates early warning information and equipment operation suggestions based on the risk warning event and its probability of occurrence, and outputs the early warning information and equipment operation suggestions.

[0005] In the above method, the plant equipment is a semiconductor production line system composed of multiple physical devices; the personalized performance prediction model is a system-level personalized performance prediction model trained for the semiconductor production line system. In the historical data acquisition step, the acquired sets of historical data records are the historical data records of the semiconductor production line system, wherein the measured values ​​of key operating parameters at the historical moment include the measured values ​​of key operating parameters at multiple key monitoring points in the semiconductor production line system. In the personalized prediction model training step, the trained system-level personalized performance prediction model is configured to output a set of predicted values ​​of key operating parameters corresponding to the multiple key monitoring points based on the input operating conditions and health status identifiers. In the equipment operation risk simulation step, the calculated key operation parameter prediction sequence is a set of key operation parameter prediction sequences corresponding to the multiple key monitoring points.

[0006] The above method, in the personalized prediction model training step, uses machine learning algorithms to enable the model to learn the mapping relationship from input features to target labels, including the following steps: Construct a training dataset, wherein the historical operating conditions and historical health status identifiers are combined into a feature vector X=[C,H], where C represents the historical operating condition vector and H represents the historical health status identifier vector; and the corresponding measured values ​​of the key operating parameters are used as the label vector Y. Initialize a regression prediction model f(·), whose output is the predicted value of the key operating parameters; The model training iteration is performed. In each iteration, the feature vector X is input into the current model f(·) to obtain the predicted value f(X). The loss value L=L(f(X),Y) between the predicted value f(X) and the true label Y is calculated, where L(·) is a preset loss function used to measure the difference between the predicted value and the true value. Based on the loss value L, the internal parameters θ of the regression prediction model f(·) are adjusted using an optimization algorithm to reduce the loss value; Repeat the model training iterations until the preset training stopping condition is met to obtain the trained personalized performance prediction model f. θ (·), where θ is the optimized model parameter.

[0007] In the above method, during the historical data acquisition step, the historical health status identifier is a discrete classification label or continuous numerical feature generated based on the historical maintenance records, historical anomaly logs, and / or historical component replacement records of the plant equipment, used to quantitatively represent the health status of the equipment at a historical moment.

[0008] In the above method, the future condition acquisition and adaptation step involves converting production scheduling information into predicted load conditions, including: parsing the production plan and process type in the production scheduling information, calculating the resource load that the plant equipment needs to provide to support the production plan at various future time points based on a preset process load mapping relationship, and forming the predicted load conditions in time order.

[0009] In the above method, the future condition acquisition and adaptation step involves converting weather forecast data into predicted environmental conditions, including: extracting ambient temperature, relative humidity, and wind speed information from the weather forecast data and using them as external environmental parameters that directly affect the plant equipment or its auxiliary systems, thereby forming the predicted environmental conditions in time order.

[0010] The above method, in the equipment operation risk deduction step, assesses equipment operation risk based on the predicted sequence of key operating parameters, and generates risk warning events and their probability of occurrence, including: Set dynamic safety thresholds for key operating parameters of plant equipment; Performing a Monte Carlo simulation involves multiple simulation iterations, with the following sub-steps executed in each iteration: Step a: Within the range of values ​​or probability distributions of at least one of the predicted load conditions, the predicted environmental conditions, and the preset inferred health status indicators, random sampling or setting is performed to obtain a set of input condition samples for the current simulation iteration. Step b: Input the input condition sample of the current simulation iteration into the personalized performance prediction model, calculate a set of predicted values ​​of key operating parameters, and determine whether any value in the set of predicted values ​​exceeds the dynamic safety threshold. If the risk exceeds the limit, it is determined that the aforementioned risk warning event has occurred in this simulation iteration; The total number of times the risk warning event occurred during the multiple simulation iterations is counted, and the ratio of the total number of occurrences to the total number of simulations is taken as the probability of the risk warning event occurring.

[0011] In the above method, the equipment operation suggestion in the early warning decision generation and output step includes at least one of the following: an instruction to adjust the operating setpoint parameters of the plant equipment, an instruction to generate a preventive maintenance work order, or an instruction to suggest adjusting the production schedule.

[0012] A second aspect of this invention discloses a data-driven semiconductor plant equipment risk early warning system, comprising: The historical data acquisition module is used to acquire multiple sets of historical data records from the historical operation and maintenance database of the plant equipment. Each set of historical data records includes the measured values ​​of key operating parameters of the plant equipment at a historical moment, the historical operating conditions corresponding to the historical moment, and the historical health status identifier that characterizes the health level of the plant equipment at the historical moment. The personalized prediction model training module is used to train a personalized performance prediction model based on the multiple sets of historical data records through the device profile intelligent agent. The combination of the historical operating conditions and the historical health status identifier is used as the model input features, and the measured values ​​of the key operating parameters are used as the target labels for model training. The model learns the mapping relationship from the input features to the target labels through machine learning algorithms. The future conditions acquisition and adaptation module is used to acquire production scheduling information and weather forecast data for a future target period, and to convert the production scheduling information into predicted load conditions for the plant equipment in the future target period, and to convert the weather forecast data into predicted environmental conditions for the plant equipment in the future target period; the predicted load conditions and the predicted environmental conditions together constitute the future operating conditions of the plant equipment in the future target period. The equipment operation risk simulation module is used to input the predicted load conditions, the predicted environmental conditions, and a preset simulation health status identifier into the personalized performance prediction model through a situational simulation intelligent agent, calculate the predicted sequence of key operating parameters of the plant equipment in the future target period, assess the equipment operation risk based on the predicted sequence of key operating parameters, and generate at least one risk warning event and its corresponding probability of occurrence. The early warning decision generation and output module is used to generate early warning information and equipment operation suggestions based on the risk warning event and its probability of occurrence through the decision and early warning intelligent agent, and output the early warning information and equipment operation suggestions.

[0013] A third aspect of the present invention discloses an electronic device, comprising: a memory and a processor, wherein the processor and the memory are connected; The memory is used to store programs; The processor invokes a program stored in the memory to execute the method provided by the first aspect embodiment and / or any possible embodiment in combination with the first aspect embodiment.

[0014] The fourth aspect of the present invention discloses a computer-readable storage medium having a computer program stored thereon, the computer program being executed by a computer to perform the methods provided by the embodiments of the first aspect and / or any possible embodiments in combination with the embodiments of the first aspect.

[0015] Compared with existing technologies, this invention has the following advantages: Through the historical data acquisition and personalized prediction model training steps, a personalized performance prediction model is trained based on the equipment's own historical operating data. This model can learn the performance mapping relationship of the equipment under different operating conditions and health states, thus abandoning fixed and rigid alarm thresholds and realizing a personalized and dynamic definition of the equipment's safe operation window, solving the problem that rigid thresholds cannot adapt to the actual state of the equipment. Through the future condition acquisition and adaptation steps, future production schedules and weather forecast information are actively acquired and transformed, and used as future operating conditions input into the model. This dynamically and quantitatively incorporates weather as a key external variable and production plan as an internal variable into the risk assessment system, breaking down the barriers of isolated information and inability to correlate and analyze in existing technologies. Through the equipment operation risk extrapolation steps, the future operating conditions are input into the personalized model using a situational extrapolation agent to extrapolate the prediction sequence of key operating parameters for future periods. Based on this, risks are assessed and probabilistic early warning events are generated. This allows the system to achieve advanced reasoning and early warning of potential risks before anomalies actually occur, based on multi-source information fusion and simulation calculations, fundamentally transforming the equipment monitoring mode from passive response to proactive early warning. Finally, through the aforementioned early warning decision generation and output steps, the abstract deduction results are transformed into actionable early warning information and suggestions, guiding maintenance personnel to make precise and proactive interventions. In summary, the various steps described in this invention work synergistically to solve the major technical problem of existing technologies being unable to proactively and individually predict and warn of future operational risks of equipment by integrating multi-source dynamic information such as historical equipment performance, future weather, and production plans before anomalies occur.

[0016] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method in Embodiment 1 of the present invention.

[0018] Figure 2 This is a system module diagram of Embodiment 2 of the present invention. Detailed Implementation

[0019] Example 1 like Figure 1 As shown, a data-driven method for early warning of equipment risks in semiconductor plants includes the following steps: Historical data acquisition steps: Obtain multiple sets of historical data records from the historical operation and maintenance database of the plant equipment. Each set of historical data records includes the measured values ​​of key operating parameters of the plant equipment at a historical moment, the historical operating conditions corresponding to the historical moment, and the historical health status identifier that characterizes the health level of the plant equipment at the historical moment. Personalized prediction model training steps: Through the device profile intelligent agent, a personalized performance prediction model is trained based on the multiple sets of historical data records. The combination of the historical operating conditions and the historical health status identifier is used as the model input features, and the measured values ​​of the key operating parameters are used as the target labels for model training. The model learns the mapping relationship from the input features to the target labels through machine learning algorithms. Future Condition Acquisition and Adaptation Steps: Acquire production scheduling information and weather forecast data for the future target period, and convert the production scheduling information into predicted load conditions for the plant equipment during the future target period, and convert the weather forecast data into predicted environmental conditions for the plant equipment during the future target period; the predicted load conditions and the predicted environmental conditions together constitute the future operating conditions of the plant equipment during the future target period. Equipment operation risk simulation steps: Through the situation simulation intelligent agent, the predicted load conditions, the predicted environmental conditions, and a preset simulation health status identifier are input into the personalized performance prediction model to calculate the predicted sequence of key operating parameters of the plant equipment in the future target period, and evaluate the equipment operation risk based on the predicted sequence of key operating parameters to generate at least one risk warning event and its corresponding probability of occurrence. Early warning decision generation and output steps: The decision-making and early warning intelligent agent generates early warning information and equipment operation suggestions based on the risk warning event and its probability of occurrence, and outputs the early warning information and equipment operation suggestions.

[0020] During implementation, in the historical data acquisition step, the system retrieves multiple sets of historical data records from the historical operation and maintenance database of the plant equipment. Each set of records corresponds to a specific point in the past. For example, for a chiller that provides cooling water to a lithography machine, a set of historical data records includes the following information: the historical time is 2 PM on July 15, 2025; the measured values ​​of key operating parameters at this time include a chiller outlet water temperature of 7.2 degrees Celsius, a compressor current of 45 amps, and a condensing pressure of 1.5 MPa; the corresponding historical operating conditions at this time include a plant production load rate of 85% and an outdoor ambient temperature of 32 degrees Celsius; the historical health status characterizing the equipment's health level at this historical time is marked as "good," which is determined based on records of no abnormal alarms and no unplanned maintenance within the three months prior to this time. By collecting massive amounts of such historical data records, the system lays the data foundation for subsequent modeling.

[0021] In the personalized prediction model training step, the equipment profiling agent begins its work. It utilizes multiple sets of historical data records acquired in the previous step for machine learning training, aiming to obtain a personalized performance prediction model that accurately describes the performance characteristics of the specific equipment. Specifically, the agent combines historical operating conditions and health status indicators from each historical data record to form the model's input features. For example, it encodes and concatenates an 85% production load rate, an ambient temperature of 32 degrees Celsius, and a "good" health status indicator to form a feature vector. Simultaneously, it uses the measured values ​​of key operating parameters corresponding to that record—such as an outlet water temperature of 7.2 degrees Celsius and a compressor current of 45 amps—as target labels that the model needs to approximate during training. By using machine learning algorithms such as gradient boosting trees or neural networks, the model learns on a large number of such "feature-label" samples, continuously adjusting its internal parameters, and ultimately learns the complex mapping relationship between a given combination of operating conditions and health status and the predicted values ​​of key operating parameters of the equipment. This is like creating a digital twin of the refrigeration unit, which can answer the question, "What are the most likely values ​​of its key parameters under specific loads, specific weather conditions, and a certain health condition?"

[0022] In the future condition acquisition and adaptation step, the system first obtains future target time periods, such as production scheduling information for the next 72 hours, from the Manufacturing Execution System (MES), and retrieves weather forecast data for the same period from the meteorological service interface. Next, the system transforms this information into future operating conditions that the equipment can understand. For production scheduling information, the system parses the production plan and process type it contains. For example, the schedule shows that a batch of critical layer lithography processes will be performed in workshop A tomorrow morning. Based on a pre-configured process load mapping table, the system calculates the cooling load required by the refrigeration unit to support this batch of lithography production, thereby generating a sequence of predicted load conditions for future time points. This process load mapping table is a pre-defined knowledge base or database within the system, establishing a quantitative correspondence between different process steps, product types, and the support resources required by the plant system in semiconductor manufacturing. For the refrigeration unit, its core function is to provide cooling capacity; therefore, the mapping relationship mainly focuses on the correspondence between "process - cooling load".

[0023] For example, future production scheduling information obtained from the manufacturing execution system shows that from 8:00 AM to 12:00 PM tomorrow, "Product Model A" is planned to be produced in the "Lithography Area - Line A" using the "critical layer lithography" process, with a planned wafer input of 25 wafers.

[0024] The system first analyzes this scheduling information and extracts the key elements: the production equipment is "photolithography area - line A", the process type is "critical layer photolithography", the product model is "product model A", the production time is 4 hours, and the wafer load is 25 wafers.

[0025] Next, the system queries the pre-configured process load mapping table. This table contains the following core fields: Process number / name: such as "critical layer photolithography", "etching", "chemical vapor deposition", etc.

[0026] Applicable equipment types or areas: such as "photolithography area", "etching area", etc.

[0027] Product model (optional, for more granular mapping).

[0028] Baseline load factor: This represents the theoretical heat load or cooling demand per wafer per hour of operation for this process. The unit can be "kilowatts / wafer·hour" or "cooling tons / wafer·hour".

[0029] Equipment load factor: A correction factor for a specific machine or production line, used to account for energy efficiency differences caused by different equipment models and years of use.

[0030] The system performs a matching query in the mapping table based on the conditions "critical layer lithography", "lithography area - line A", and "product model A". Assume the query result is: for the "critical layer lithography" process of "product model A", the baseline load factor in the "lithography area - line A" is 0.12 refrigeration tons / wafer·hour.

[0031] The system then performs calculations. In this example, it calculates the total heat load generated by the lithography area - line A to support this production plan during the future time period, i.e., the cooling load required from the chiller: Predicted cooling load = Number of wafers fed × Production time × Baseline load factor Predicted cooling load = 25 sheets × 4 hours × 0.12 refrigeration tons / sheet·hour = 12 refrigeration tons·hour The "12 refrigeration tons per hour" indicates that the average cooling demand during this 4-hour period is 12 refrigeration tons. The system will further distribute this total load smoothly according to the production duration or decompose it into sequence data with hourly or even shorter time intervals according to the process curve model. For example, it will generate a predicted load condition sequence with a cooling demand of 3 refrigeration tons per hour from 8:00 AM to 12:00 PM. If the production schedule is continuous or multiple tasks are superimposed, the system will perform the same analysis and calculation on all tasks and accumulate them to ultimately form a predicted load condition curve that covers the future target period and changes over time.

[0032] Through this process, the system transforms abstract production plan texts or data into quantifiable and understandable future workload data for plant equipment, thus providing crucial input conditions for subsequent personalized performance prediction and risk simulation. This achieves deep integration of production information and plant operation information, and is one of the key steps in breaking down information silos and realizing accurate and forward-looking early warning.

[0033] For weather forecast data, the system directly extracts ambient temperature, relative humidity, and other data as a sequence of predicted environmental conditions that will directly affect components such as the condenser of the refrigeration unit. The predicted load conditions and predicted environmental conditions together constitute the future operating conditions describing the future operating environment of the equipment.

[0034] In the equipment operation risk simulation step, the situational simulation agent initiates the simulation. It receives the future operating conditions obtained from the previous step and sets a preset health status identifier for simulation, such as "current status," which is based on the latest equipment health assessment results. Then, the agent inputs this information into a pre-trained personalized performance prediction model. The model calculates the predicted values ​​of key operating parameters such as chiller outlet water temperature and compressor current for each hour over the next 72 hours, forming a predicted sequence of key operating parameters. To assess risk, the agent uses a Monte Carlo simulation method. It first sets dynamic safety thresholds for key parameters; for example, the upper limit of the dynamic safety threshold for outlet water temperature may be dynamically adjusted to 8 degrees Celsius based on the condensing temperature. Then, it performs thousands of random samplings within the possible fluctuation range of predicted load conditions and predicted environmental conditions. Each sampling constitutes a set of possible future scenarios input into the model, obtaining a set of predicted parameters. Next, it determines whether this set of predicted parameters exceeds the dynamic safety threshold. The system counts the number of times the limit is exceeded in all simulations. For example, if the predicted outlet water temperature exceeds 8 degrees Celsius 1,500 times in 10,000 simulations, the system will generate a risk warning event called "Refrigeration Unit Outlet Water Temperature Exceeds Limit" and calculate its probability of occurrence as 15%.

[0035] In the early warning decision generation and output step, the decision-making and early warning intelligent agent generates the final output based on the inference results. Regarding the risk of "refrigeration unit outlet water temperature exceeding the limit" assessed above and its 15% probability of occurrence, the intelligent agent generates specific early warning information, such as "It is predicted that within the next 24 hours, due to high temperatures and high-load production, the outlet water temperature of refrigeration unit No. 5 has a high probability of approaching the safety upper limit; attention is advised." Simultaneously, it generates corresponding equipment operation suggestions, such as "It is recommended to temporarily lower the refrigeration unit setpoint temperature from 7.5 degrees Celsius to 7.0 degrees Celsius," or "It is recommended to generate a preventative maintenance work order to check the cleanliness of the condenser fins," or "It is recommended to consult with the production department to stagger the scheduling of some high-load processes." Finally, the system pushes these early warning messages and equipment operation suggestions to plant engineers through the factory's visual dashboard, mobile application, or work order system, guiding them to proactively intervene and eliminate potential faults before they occur.

[0036] To further clarify, the core task of the device profiling agent is to "train a personalized performance prediction model." Its implementation is essentially a model training pipeline. In the software system, this agent is a background service program. When triggered, it executes the following process: First, it reads data from the historical maintenance database and, following the requirements of the "historical data acquisition steps," cleans and organizes structured training samples. Each sample contains historical operating conditions, historical health status identifiers, and corresponding measured values ​​of key operating parameters. Then, it initializes a machine learning model, such as a gradient boosting tree model or a neural network model. Next, it inputs these samples into the model for training, iteratively teaching the model the mapping relationship from "operating conditions + health status" to "operating parameters." After training, it saves the trained model parameters, forming the "personalized performance prediction model" for the chiller. This model is the unique "digital profile" of the device.

[0037] Taking a chiller as an example, the intelligent agent for equipment profiling read its data from every day and every hour over the past two years, totaling approximately 17,000 valid records. It used this data to train an XGBoost regression model. After training, this model can accurately predict that when the production load is 80%, the outdoor temperature is 30 degrees Celsius, and the equipment health status is marked as "good," the chiller's outlet water temperature is highly likely to be 7.1 degrees Celsius, and the compressor current will be 42 amps. This accurate predictive capability is the result of the intelligent agent's work.

[0038] The core task of the situational simulation agent is to "assess equipment operational risks and generate risk warning events and their probabilities." Its implementation is a simulation-based risk prediction engine. This agent is a computational engine module. After acquiring the future operating conditions produced by the "future condition acquisition and adaptation steps," it combines this with a preset simulation health status indicator to initiate a Monte Carlo simulation. In its implementation, it sets up a loop, for example, 10,000 times. In each loop, the program randomly samples from the predicted values ​​of future load and future weather, according to their possible error distribution (such as a normal distribution), to obtain a set of specific input values. This set of values, along with the health status indicator, is then input into the personalized performance prediction model produced by the "equipment profiling agent" to calculate a set of predicted operating parameters. The program compares these predicted values ​​with dynamic safety thresholds to determine if limits are exceeded. After 10,000 loops, the program counts the number of times operating parameters exceed limits; for example, if there are 1,200 instances, then the probability of the risk event "outlet water temperature exceeding limits" is 12%. The core implementation of this intelligent agent is the process of extrapolating various possible future scenarios and calculating their probabilities.

[0039] For the refrigeration unit, the situational simulation agent obtained hourly predicted loads and weather forecasts for the next 24 hours. Assuming the equipment's health remained constant, it began simulations. In the first simulation, it randomly assumed that the actual load at 2 PM the following day would be 5% higher than the scheduled load and the temperature would be 1 degree Celsius higher than the forecast. It then called the model and calculated the outlet water temperature to be 7.9 degrees Celsius, below the 8-degree threshold, and recorded it as "safe." In the second simulation, it randomly assumed the load was normal but the temperature was higher. The model calculated the water temperature to be 8.2 degrees Celsius, and recorded an "over-limit warning"... This process was repeated tens of thousands of times, ultimately calculating the probability of an over-limit event to be 12%. This probability result is the situational scenario it "simulated."

[0040] The core task of the decision-making and early warning agent is to "generate early warning information and equipment operation suggestions based on risk warning events and their probability of occurrence." Its implementation is a rule- and policy-based knowledge application system. This agent typically contains a rule engine or policy library. It takes the output of the "situational simulation agent" (such as "outlet water temperature exceeds limit, probability 12%)" as input. Internally, it pre-sets or loads a series of "if-then" decision rules from the knowledge base. These rules map information such as risk type, risk probability, and equipment type to specific early warning levels, text information, and operation suggestions. Then, based on the matched rules, it generates structured output. Finally, it calls the message interface to push the early warning information to the factory monitoring screen or generate work orders for equipment operation suggestions and send them to the equipment management system. This process of converting risk values ​​into executable instructions is the implementation of this agent.

[0041] Taking a chiller as an example, the decision-making and early warning agent receives a risk event: "Chiller outlet water temperature exceeds limit, probability 12%". Its built-in rule base contains a rule that states: "If the risk event is 'outlet water temperature exceeds limit' and the probability of occurrence is greater than 10% but less than 20%, the warning level is 'Caution,' and the suggestion is 'It is recommended to lower the setpoint temperature by 0.5 degrees Celsius to allow for a safety margin.'" Therefore, the agent automatically generates an early warning message: "Caution: It is expected that the outlet water temperature of chiller No. 5 will exceed the limit within the next 24 hours. It is recommended to temporarily adjust the setpoint from 7.5 degrees to 7.0 degrees." and pushes it to the plant engineer. This precise suggestion is the product of its "decision-making."

[0042] Overall, the construction and deployment of these three intelligent agents are based on a layered software architecture. The data fusion layer is responsible for extracting data from sources such as the plant monitoring system, manufacturing execution system, and meteorological service interface in real time or on a timed basis, and performing standardized processing to provide a unified input for the intelligent agents. The intelligent agent hub layer contains three core processing modules: an equipment profiling module, a situational analysis module, and a decision-making and early warning module. The human-machine interaction layer is responsible for pushing the output results of the intelligent agents to users through visual dashboards, mobile messages, or work order systems.

[0043] Specifically, the construction of each intelligent agent is essentially an automated machine learning model training pipeline. The construction process involves developing a background service program that can start according to predetermined cycles or trigger conditions. The program first extracts historical data of the target equipment from a historical maintenance database. This data includes measured values ​​of operating parameters at multiple points in time, recorded production loads and environmental conditions at those times, and health status labels generated based on maintenance records. The program cleans and aligns this data, constructing feature vectors and label vectors. Then, the program calls a machine learning framework, such as loading a gradient boosting tree regression algorithm, and trains it using the prepared data as input. During training, the algorithm continuously adjusts the model's internal parameters to minimize the error between the model's predicted operating parameter values ​​and historical true values. After training, the program serializes and saves the final model parameters into a file; this file is the device's personalized performance prediction model. For example, building an intelligent agent for a specific refrigerator in a semiconductor factory involves using its operating logs from the past two years to train a mathematical model that accurately reflects its cooling efficiency under different loads and weather conditions.

[0044] The core of the situational simulation agent is an integrated simulation engine. Developers need to build a computing service that receives future predicted operating conditions and incorporates a Monte Carlo simulation algorithm. The service's logic is as follows: First, it reads the personalized performance prediction model file generated by the device profiling agent and loads it into memory. Next, it obtains the predicted load sequence converted from the production schedule for the next 72 hours, and the predicted environmental condition sequence converted from weather forecasts, from the data fusion layer. During simulation, the program performs thousands of simulation loops. In each loop, the program randomly samples near the baseline values ​​of the predicted load and predicted environmental conditions, according to their possible error range, generating a specific set of simulated input values. These values ​​are then input into the loaded personalized performance prediction model to calculate the key operating parameters of the device under the simulated scenario. The program compares these calculated parameters with preset dynamic safety thresholds to determine whether a risk event will occur. After all simulation loops are completed, the program calculates the proportion of risk events occurring out of the total number of occurrences and outputs this as the probability of risk occurrence. For example, for the aforementioned refrigeration unit, the intelligent system can simulate tens of thousands of possible load and temperature combinations under tomorrow's high-temperature weather, and by calling its personalized model, calculate the probability that the outlet water temperature will exceed the safety line.

[0045] The decision-making and early warning intelligent agent is primarily built upon a rule engine. It is implemented as a logic processing service, internally configured with or connected to a rule knowledge base. This rule knowledge base defines the early warning levels, notification texts, and specific operational suggestions corresponding to different risk types and risk probability levels. The service's program flow is as follows: it receives the output of the situational analysis intelligent agent, including the risk event type and its probability of occurrence. Then, the program matches these two key pieces of information with the conditions in the rule knowledge base. Once a rule is matched, the corresponding action is triggered, generating structured early warning information and equipment operation suggestions. For example, the rule base might specify that when the risk event is "outlet water temperature exceeds the limit" and the probability is between 10% and 20%, the early warning level is "Caution," and the suggestion is "It is recommended to lower the refrigeration unit setpoint temperature by 0.5 degrees Celsius." The program will then automatically generate the corresponding suggestion text.

[0046] These three intelligent agents are typically deployed in a server computing environment as microservices or containerized systems. The device profiling agent is deployed as a periodic batch processing task, such as automatically starting a training task weekly when server resources are idle. The situational simulation agent can be deployed as an event-driven service, such as triggering at fixed times daily or immediately upon receiving a new major weather forecast. The decision-making and early warning agent is deployed as a resident service, continuously monitoring the message queue output by the situational simulation agent. They communicate with the data fusion layer and human-machine interaction layer through application programming interfaces or message middleware to obtain input from the factory data platform and push early warning decisions to display terminals and work order systems. Through this construction and deployment approach, the three agents can form an automated closed-loop system for data acquisition, model calculation, risk simulation, and decision generation.

[0047] In one embodiment, the plant equipment is a semiconductor production line system composed of multiple physical devices; the personalized performance prediction model is a system-level personalized performance prediction model trained for the semiconductor production line system. In the historical data acquisition step, the acquired sets of historical data records are the historical data records of the semiconductor production line system, wherein the measured values ​​of key operating parameters at the historical moment include the measured values ​​of key operating parameters at multiple key monitoring points in the semiconductor production line system. In the personalized prediction model training step, the trained system-level personalized performance prediction model is configured to output a set of predicted values ​​of key operating parameters corresponding to the multiple key monitoring points based on the input operating conditions and health status identifiers. In the equipment operation risk simulation step, the calculated key operation parameter prediction sequence is a set of key operation parameter prediction sequences corresponding to the multiple key monitoring points.

[0048] During implementation, the historical data acquisition step involves obtaining multiple sets of historical data records from the entire semiconductor production line system. For example, consider a production line system that provides ultrapure water to the entire plant. This system includes multiple connected physical devices such as raw water pumps, reverse osmosis units, ultraviolet sterilizers, and terminal filters. The system retrieves snapshots of the system's operation at various historical moments from its historical maintenance database. Each record includes not only a single parameter but also a set of measured values ​​for key operating parameters at multiple critical monitoring points within the system. For instance, one record might simultaneously include: at historical moment T, the measured outlet pressure of the raw water pump is 0.55 MPa, the measured conductivity of the permeate from the first-stage reverse osmosis unit is 1.2 microsiemens per centimeter, and the measured inlet flow rate of the ultraviolet sterilizer is 50 liters per minute. This record also includes historical operating conditions affecting the entire production line system at that historical moment, such as the plant's instantaneous pure water demand. In addition, it is also necessary to include a historical health status identifier that characterizes the overall health level of the production line system at historical time T. This identifier may be determined based on whether any equipment on the production line has recently experienced an alarm or unplanned maintenance.

[0049] In the personalized prediction model training step, the equipment profiling agent is trained based on historical data records of the aforementioned production line system, with the goal of establishing a system-level personalized performance prediction model for the entire production line system. During training, the combination of historical operating conditions and historical health status identifiers is used as the model input features. The target label for model training is the set of measured values ​​of key operating parameters at multiple key monitoring points within the production line system corresponding to these input features. Through machine learning algorithms, the model learns the complex mapping relationship from "system operating conditions and health status" to "the set of key parameters for multiple devices within the system." The finally trained system-level model can predict the operating parameters of each key monitoring point within the system based on the input system-level operating conditions and health status. For example, when the input condition is "high-purity water demand and good system health status" for a future period, the model can simultaneously output a set of predicted values ​​for multiple parameters such as raw water pump outlet pressure, reverse osmosis permeate conductivity, and UV sterilizer flow rate.

[0050] In the future condition acquisition and adaptation steps, the future production scheduling information and weather forecast data acquired by the system are also transformed with the entire production line system as the service object. For example, the production scheduling information is transformed into the predicted water production load conditions that the ultrapure water production line system needs to meet in the future target period. The weather forecast data is transformed into the predicted environmental conditions that will also affect outdoor equipment (such as raw water tanks and cooling devices) in the production line system.

[0051] In the equipment operation risk simulation step, the situational simulation agent inputs the predicted load conditions, predicted environmental conditions, and a preset simulated health status identifier into a trained system-level personalized performance prediction model. The model's output is no longer a single parameter sequence, but a set of predicted sequences of key operating parameters corresponding to multiple key monitoring points within the production line system. Subsequently, the operational risk of the entire system is assessed based on this set of predictions. For example, the system will simultaneously check whether the raw water pump outlet pressure exceeds its safety limit, whether the reverse osmosis permeate conductivity exceeds its purity limit, and whether the ultraviolet sterilizer flow rate is below its effective sterilization limit. Through methods such as Monte Carlo simulation, the probability of multiple risk warning events such as "permeate conductivity exceeding the standard" and "insufficient sterilization flow rate" occurring in the future can be assessed. These events may occur individually or in combination.

[0052] In the early warning decision generation and output step, the decision-making and early warning agent generates early warnings and suggestions based on the system-level risk simulation results. For example, if the simulation shows that "under high load, the raw water pump pressure is likely to be too high, and the reverse osmosis conductivity is likely to deteriorate simultaneously," the agent generates an early warning message: "Warning: It is expected that during future high load periods, there is a risk of coupled deterioration in the overall pressure and water quality of the ultrapure water system," and generates system-level operational suggestions, such as "Suggestions: 1. Start the backup raw water pump to share the load; 2. Prepare to perform online chemical cleaning of the reverse osmosis membrane module." This elevates early warning from monitoring individual equipment parameters to a comprehensive assessment and coordinated control of the entire production line system's operational status.

[0053] In one embodiment, the personalized prediction model training step involves using a machine learning algorithm to teach the model the mapping relationship from input features to target labels, including the following steps: Construct a training dataset, wherein the historical operating conditions and historical health status identifiers are combined into a feature vector X=[C,H], where C represents the historical operating condition vector and H represents the historical health status identifier vector; and the corresponding measured values ​​of the key operating parameters are used as the label vector Y. Initialize a regression prediction model f(·), whose output is the predicted value of the key operating parameters; The model training iteration is performed. In each iteration, the feature vector X is input into the current model f(·) to obtain the predicted value f(X). The loss value L=L(f(X),Y) between the predicted value f(X) and the true label Y is calculated, where L(·) is a preset loss function used to measure the difference between the predicted value and the true value. Based on the loss value L, the internal parameters θ of the regression prediction model f(·) are adjusted using an optimization algorithm to reduce the loss value; Repeat the model training iterations until the preset training stopping condition is met to obtain the trained personalized performance prediction model f. θ (·), where θ is the optimized model parameter.

[0054] In implementation, taking the "ultrapure water preparation system" in the semiconductor production line system as an example, in the personalized prediction model training step, in order to enable the model to learn the mapping relationship from input features to target labels through machine learning algorithms, a system-level personalized performance prediction model for the production line system needs to be constructed and trained according to the following detailed steps.

[0055] First, a training dataset is constructed. The system extracts multiple sets of historical data records from the historical operation and maintenance database of the ultrapure water preparation system. For each set of records, historical operating conditions describing the overall operational background of the system, such as the hourly pure water consumption of the entire plant at that moment, are combined and encoded with historical health status indicators characterizing the overall health level of the system, such as the "health" level determined based on recent maintenance records, to form a feature vector. This feature vector represents the "conditions" that the model needs to learn during training. Simultaneously, the measured values ​​of key operating parameters at multiple key monitoring points within the production line system at the corresponding time in this set of records, such as the raw water pump outlet pressure, the conductivity of the first-stage reverse osmosis permeate, and the differential pressure of the terminal filter, are combined into a label vector. This label vector represents the "result" that the system should exhibit under those "conditions." A large number of such "feature vector-label vector" pairings constitute the model's training dataset.

[0056] Secondly, a regression prediction model is initialized. Those skilled in the art can choose a machine learning algorithm suitable for processing structured data and possessing strong nonlinear fitting capabilities as the model foundation, such as a gradient boosting decision tree model or a deep neural network model. In terms of program implementation, this involves calling the corresponding machine learning library, such as Scikit-learn or TensorFlow, configuring the basic architecture and initial hyperparameters of the model, and creating an untrained model instance. This initial model does not possess accurate predictive capabilities before training.

[0057] Next, the model training iteratively is performed. This is a continuous optimization process. In each iteration, the training program takes a batch of samples from the dataset, such as one hundred feature vectors. The program inputs these feature vectors into the current model, which calculates and outputs the corresponding predicted values ​​based on its current internal parameters—that is, the predicted values ​​of multiple runtime parameters. Then, the program calculates the difference between these predicted values ​​and the corresponding true label vectors. This difference is quantified using a preset loss function. A commonly used loss function is mean squared error, which calculates the average of the squares of the differences between the predicted and true values. The calculated loss value represents the degree of error in the current model's prediction.

[0058] Then, based on the calculated loss value, the model's intrinsic parameters are adjusted using an optimization algorithm. Specifically, the program calculates the gradient of the loss value relative to each of the model's intrinsic parameters. This gradient indicates how much and in which direction each parameter should be adjusted to reduce the loss. Next, an optimization algorithm such as stochastic gradient descent or its variants is used to update all parameters according to the calculated gradient direction. The goal of this process is to enable the model to make predictions closer to the true label the next time it faces the same input, thereby reducing the loss value.

[0059] Finally, the above model training iterations are repeated. The training program traverses the training dataset multiple times, with each iteration containing multiple parameter updates. This process continues until a preset training stopping condition is met. Common stopping conditions include the model's performance no longer improving on independent validation datasets, or the training reaching the preset maximum number of iterations. When training stops, the model's internal parameters have been optimized to a relatively stable state. The resulting model is the trained, personalized performance prediction model for this ultrapure water preparation system. This model can understand the complex relationships between system operating conditions, health status, and the operating parameters of multiple devices. For example, when the input features are "high-purity water demand, moderate system health," the model can accurately predict a set of parameter values ​​such as pump pressure and product water conductivity, providing a reliable basis for subsequent risk simulation.

[0060] In one embodiment, in the historical data acquisition step, the historical health status identifier is a discrete classification label or continuous numerical feature generated based on the historical maintenance records, historical anomaly logs and / or historical component replacement records of the plant equipment, used to quantitatively represent the health status of the equipment at a historical moment.

[0061] In implementation, the "chemical mechanical polishing post-cleaning unit" in a semiconductor production line system will be used as an example: In the historical data acquisition step, in order to construct training data for the production line system of this cleaning unit, the system needs to generate a corresponding, quantifiable historical health status identifier for each "historical moment" in its historical data records. This identifier is calculated based on the objectively existing operation and maintenance record data of the production line system through predefined rules or algorithms.

[0062] Specifically, the system extracts three core types of data from the cleaning unit's historical database: historical maintenance records, historical anomaly logs, and historical component replacement records. Historical maintenance records include the execution time and content of planned maintenance, such as "On August 1, 2025, preventative maintenance was performed on the megasonic generator." Historical anomaly logs record alarm information triggered by the system or sub-equipment, such as "On August 15, 2025, the roller brush motor current exceeded the limit alarm, lasting for 10 minutes." Historical component replacement records indicate the replacement history of critical consumables or components, such as "On July 1, 2025, all process nozzles were replaced."

[0063] When generating historical health status identifiers, discrete classification labels can be used. The system will backtrack relevant records within a previous time window for each historical data point. For example, for the historical record of 2 PM on August 20, 2025, the system will backtrack the maintenance data from the previous 30 days. If no abnormal alarms occurred during these 30 days, one planned maintenance was performed, and the operating time of all critical components was within 80% of their design life, the system can assign the discrete classification label "Healthy" to that moment according to rules. Conversely, if multiple process pressure fluctuation alarms occurred within the previous 7 days, and a critical valve has exceeded its recommended replacement cycle, it may be labeled "Sub-healthy" or "Caution." The label levels can be set to multiple levels such as "Excellent," "Good," "Medium," and "Poor."

[0064] Another more refined approach is to generate continuous numerical features. The system can design a health index calculation model, for example, with a baseline score of 100. Points are deducted or weighted based on records within the review period: 2 points are deducted for each minor alarm, and 10 points are deducted for each serious alarm; the longer the time since the last planned maintenance, the lower the weight of the relevant items; the remaining lifespan percentage of critical components is directly used as a multiplier factor. Through calculation, a continuous value between 0 and 1 or between 0 and 100 is obtained, for example, a health index of 0.85. This value better characterizes the gradual change in the equipment's health level.

[0065] For example, when constructing training data for a system-level personalized performance prediction model for a cleaning unit, the feature vector of a historical record must include not only the operating conditions such as process flow rate and chemical concentration at that moment, but also a historical health status identifier corresponding to that moment, generated based on the aforementioned method. This identifier serves as a key input feature, inputting into the model along with the operating conditions. This allows the model to learn the deeper knowledge that "even under the same operating conditions, key system parameters will exhibit different patterns under different health states." For instance, the model can learn to distinguish "the typical difference in megasonic power between a cleaning unit with excellent health and a cleaning unit with moderate health under full load," thereby significantly improving the accuracy and personalization of model predictions and providing a more reliable basis for subsequent risk simulations. This implementation method, transforming operational experience and objective data into machine-learnable features, is a crucial foundation for achieving accurate early warning in this solution.

[0066] In one embodiment, the future condition acquisition and adaptation step, which transforms production scheduling information into predicted load conditions, includes: parsing the production plan and process type in the production scheduling information, calculating the resource load that the plant equipment needs to provide to support the production plan at various future time points based on a preset process load mapping relationship, and forming the predicted load conditions in time order.

[0067] The following explanation uses the "process cooling water system" and "fresh air conditioning unit" in a semiconductor production line system, which provide temperature and humidity control for the photolithography area, as examples.

[0068] During implementation, the system needs to perform a structured data processing and calculation workflow to transform production scheduling information into predicted load conditions. First, the system obtains future target time periods, such as production scheduling information for the next 24 hours, from the manufacturing execution system. This scheduling information is structured data, containing key fields such as which production area, what time, what process to execute, what product to produce, and the planned wafer starts. For example, a scheduling record might indicate: tomorrow morning from 8:00 AM to 12:00 PM, in lithography area 1, the critical layer lithography process for product model A will be executed, with a planned wafer starts of 25 wafers; simultaneously, tomorrow afternoon from 2:00 PM to 6:00 PM, in lithography area 2, the thin film deposition process for product model B will be executed, with a planned wafer starts of 30 wafers.

[0069] Next, the system will analyze this information, focusing on extracting the production plan and process type. The purpose of this analysis is to clarify the nature of the production activities to be carried out in each production area during each future time period. For example, the system identifies that "critical layer lithography" will be performed in lithography area 1 in the morning, while "thin film deposition" will be performed in lithography area 2 in the afternoon. Different process types have vastly different requirements for the plant support system. Critical layer lithography requires extremely high environmental temperature and humidity stability, and the lithography machine itself generates a large amount of heat; while thin film deposition may have specific requirements for the purity and temperature of the process cooling water and generate different exhaust heat loads.

[0070] Then, the system performs calculations based on a pre-configured process load mapping relationship. This process load mapping relationship is a knowledge base stored in a database or configuration file, which defines the standard demand coefficients for various plant resources per unit time and per unit output for each process type, and even down to each product model. For example, the mapping relationship table specifies that for the critical layer lithography process of product model A, each wafer requires an average of 0.15 kilowatts of cooling capacity from the process cooling water system, and simultaneously requires a fresh air conditioning unit to provide cooling and dehumidification for an air volume of 0.5 cubic meters per minute. For the thin film deposition process of product model B, the process cooling water cooling capacity requirement per wafer may be 0.08 kilowatts, and the fresh air requirement may be 0.3 cubic meters per minute.

[0071] The system queries the mapping table based on the parsed production plan and process type to obtain the corresponding load coefficient. Then, combining the planned wafer starts and production duration, it calculates the resource load required by the plant equipment at various future time points to support the production plan for that period. Taking the above schedule as an example, for the morning production in lithography zone 1, the system calculates that during the period from 8:00 AM to 12:00 PM, the cooling load required by the process cooling water system is 25 wafers multiplied by four hours, then multiplied by 0.15 kW per wafer per hour, which equals 15 kilowatt-hours; the air volume load required by the fresh air conditioning unit is 25 wafers multiplied by four hours, then multiplied by 0.5 cubic meters per wafer per hour per minute, which equals 50 cubic meters. The system then decomposes this total load into time points at hourly or even minute intervals, based on the process start-up, stabilization, and shutdown curve model, or simply by averaging, forming a time-varying predicted load condition sequence. For the afternoon production in lithography zone 2, the system performs the same calculation and superimposes the loads generated by the two zones on the timeline.

[0072] Ultimately, the system outputs a time-ordered sequence of predicted load conditions. This sequence clearly indicates the total cooling load the process cooling water system needs to handle at each moment within the next 24 hours, as well as the total air volume and corresponding heat and humidity load that the fresh air conditioning units need to process. This precisely quantified future load condition is one of the key inputs for subsequent steps in equipment performance prediction and operational risk simulation. In this way, the invention transforms the production plan into a quantifiable "work order" for plant equipment, achieving a deep integration of production information and plant operation information. This allows the early warning system to proactively assess the pressure that future production activities will bring to the plant system, thereby issuing warnings before potential problems occur. For example, the system can predict in advance whether the cooling water system may experience a rise in supply water temperature due to the total cooling demand approaching its design limit when both areas are operating at high loads simultaneously tomorrow afternoon.

[0073] In one embodiment, the future condition acquisition and adaptation step involves converting weather forecast data into predicted environmental conditions, including: extracting ambient temperature, relative humidity, and wind speed information from the weather forecast data and using them as external environmental parameters that directly affect the plant equipment or its auxiliary systems, thereby forming the predicted environmental conditions in time order.

[0074] The following explanation uses an "open cooling tower system" that provides cooling for the entire cooling system in a semiconductor factory as an example.

[0075] In implementation, the method for converting weather forecast data into predicted environmental conditions in the future conditions acquisition and adaptation step is as follows: The system automatically acquires high temporal resolution weather forecast data covering the target future period from professional meteorological service agencies through a pre-set meteorological data interface. This raw data typically contains multiple meteorological elements, such as temperature, humidity, wind speed, wind direction, air pressure, precipitation, and descriptions of weather phenomena.

[0076] The system then analyzes and extracts the weather forecast data, specifically extracting ambient temperature, relative humidity, and wind speed. This selection is based on a deep understanding of the operating principles of plant equipment, particularly cooling tower systems. Ambient temperature directly affects the cooling tower's heat dissipation efficiency and proximity, relative humidity affects the evaporative cooling effect, and wind speed directly affects the air exchange rate and drift rate within the tower. These parameters are external environmental parameters that directly affect the plant equipment or its auxiliary systems, and have a clear physical correlation with their heat exchange performance; therefore, they were selected as core elements for constructing predicted environmental conditions.

[0077] After the extraction is complete, the system processes these three pieces of information as independent time-series data streams. For example, it extracts hourly dry-bulb temperature forecasts, relative humidity forecasts, and average wind speed forecasts for the next 72 hours from weather forecast data. These numerical sequences constitute a quantitative description of the external environment for the future operation of the cooling tower system.

[0078] Next, the system aligns and encapsulates these serialized parameters to form a time-ordered, uniformly formatted set of predicted environmental conditions data. Each data point in this set corresponds to a specific future time and includes the status of key environmental parameters affecting the equipment at that moment. For example, the generated predicted environmental conditions data indicates that at 2 PM tomorrow, the environmental conditions for the cooling tower will be: temperature 35 degrees Celsius, relative humidity 65%, and wind speed 2 meters per second.

[0079] Through the above implementation methods, this step transforms macroscopic, unstructured meteorological forecast information into microscopic, structured engineering parameters that directly interface with equipment operation models. This predicted environmental condition, combined with the time-series predicted load conditions, accurately characterizes the external operating conditions faced by plant equipment in the future, enabling subsequent performance predictions and risk simulations to fully consider the dynamic impacts of environmental changes. This effectively solves the problem of traditional early warning methods neglecting weather as a key dynamic disturbance factor, significantly improving the predictability and accuracy of operational risks in semiconductor plant systems, especially those sensitive to environmental impacts, under variable weather conditions.

[0080] In one embodiment, the equipment operation risk deduction step, which assesses equipment operation risk based on a predicted sequence of key operating parameters and generates risk warning events and their probabilities of occurrence, includes: Set dynamic safety thresholds for key operating parameters of plant equipment; Performing a Monte Carlo simulation involves multiple simulation iterations, with the following sub-steps executed in each iteration: Step a: Within the range of values ​​or probability distributions of at least one of the predicted load conditions, the predicted environmental conditions, and the preset inferred health status indicators, random sampling or setting is performed to obtain a set of input condition samples for the current simulation iteration. Step b: Input the input condition sample of the current simulation iteration into the personalized performance prediction model, calculate a set of predicted values ​​of key operating parameters, and determine whether any value in the set of predicted values ​​exceeds the dynamic safety threshold. If the risk exceeds the limit, it is determined that the aforementioned risk warning event has occurred in this simulation iteration; The total number of times the risk warning event occurred during the multiple simulation iterations is counted, and the ratio of the total number of occurrences to the total number of simulations is taken as the probability of the risk warning event occurring.

[0081] The "ultrapure water circulation and distribution system" in a semiconductor production line system will be used as an example for explanation.

[0082] In implementation, the specific method for assessing equipment operation risks based on the predicted sequence of key operating parameters and generating risk warning events and their corresponding probabilities in the equipment operation risk simulation step is as follows: First, dynamic safety thresholds for key operating parameters are set for the plant equipment, namely the ultrapure water circulation distribution system. These dynamic safety thresholds are not fixed values, but rather functions or lookup tables that can change with system operating conditions, equipment health baselines, or process requirements. For example, for the key parameter characterizing water purity, "circulating water resistivity," its lower safe operating threshold can be set as a value dynamically adjusted based on the total return water flow rate and the main system product water resistivity. When the return water flow rate is high and the main product water resistivity is low, this lower threshold is adjusted upwards accordingly to reserve a larger safety margin.

[0083] Next, a Monte Carlo simulation is performed. Monte Carlo simulation is a probabilistic statistical method that approximates a numerical solution through a large number of random samples. This step specifically involves performing multiple, for example, ten thousand, simulation iterations, with the following sub-steps executed in each iteration: In step a, random sampling is performed within the range of values ​​or probability distributions of at least one of the predicted load conditions, the predicted environmental conditions, and the preset inferred health status indicators to obtain a set of input condition samples for the current simulation iteration. This is because there is uncertainty in future operating condition predictions. For example, future production load predictions obtained from the manufacturing execution system have an error range of ±5%, and temperature predictions in weather forecasts have a normal distribution error. Therefore, in each simulation, the system will randomly sample within the error distribution range of these predicted values. For example, if the predicted load at a certain future time is 100 tons per hour, its probability distribution is a normal distribution with a mean of 100 tons and a standard deviation of ±5 tons. The Monte Carlo simulation will randomly sample a value, such as 102 tons, from this distribution as the load input sample for this iteration. At the same time, the predicted environmental temperature and the inferred system health status indicators can also be randomly sampled within their possible fluctuations or set ranges.

[0084] In step b, the input condition samples of the current simulation iteration are input into the trained personalized performance prediction model to calculate a set of predicted values ​​for key operating parameters of each key monitoring point in the system under the specific future scenario. Then, it is determined whether any value in this set of predicted values ​​exceeds its corresponding dynamic safety threshold. Taking the aforementioned circulating water resistivity as an example, if the predicted resistivity calculated in this simulation is 17.9 megaohm-cm, while the dynamic safety lower limit threshold calculated based on the load, temperature, and other conditions in the current simulation sample is 18.0 megaohm-cm, then the predicted value is determined to exceed the threshold. If an exceedance occurs, it is determined that the aforementioned risk warning event has occurred in this simulation iteration, such as "circulating water resistivity is lower than the dynamic safety lower limit".

[0085] After completing a preset number of simulation iterations, such as 10,000, the total number of times the risk warning event occurred in all simulation iterations is counted. Assume that the event "circulating water resistivity is below the dynamic safety lower limit" occurred 800 times. Finally, the ratio of the total number of 800 to the total number of simulations (10,000), i.e., 8%, is taken as the probability of the risk warning event occurring in the future target time period.

[0086] This step combines the uncertainty of future operating conditions with a personalized model of equipment performance, quantifying the probability of specific risk events through probabilistic simulation. Compared to simple judgments using fixed thresholds and single-point predictions, this method can more scientifically and comprehensively assess the operational risks of equipment under complex and volatile future conditions. It provides a precise data foundation for generating differentiated early warnings and decision-making recommendations at different probability levels, significantly improving the predictability and precision of risk management in semiconductor plant systems.

[0087] In one embodiment, the equipment operation suggestion in the early warning decision generation and output step includes at least one of the following: an instruction to adjust the operating setpoint parameters of the plant equipment, an instruction to generate a preventive maintenance work order, or an instruction to suggest adjusting the production schedule.

[0088] The following explanation uses the "ultrapure water preparation production line system" in a semiconductor factory as an example.

[0089] In implementation, during the early warning decision generation and output step, the equipment operation suggestions generated and output by the decision-making and early warning intelligent agent are selected from a variety of predefined instruction types based on the assessed specific risk type, risk probability, and system operating constraints. These equipment operation suggestions include, but are not limited to, at least one of the following: instructions to adjust the operating setpoint parameters of the plant equipment, instructions to generate preventative maintenance work orders, or instructions to suggest adjusting the production schedule. These suggestions aim to proactively intervene to avoid or reduce the deduced operational risks.

[0090] Specifically, when the decision-making and early warning agent receives a risk warning event and its probability of occurrence output by the situational analysis agent, it will perform matching and decision-making based on the built-in policy rule base. For example, if the risk warning event is "there is a 15% probability that the conductivity of the permeate from the first-stage reverse osmosis unit will exceed the dynamic safety limit tomorrow afternoon," and the probability level is determined to be "medium risk," the agent, after analysis, believes that the risk is mainly caused by the expected high temperature weather leading to a decrease in reverse osmosis membrane efficiency. At this time, the agent generates an instruction to adjust the equipment operating setpoint parameters. The specific content of this instruction is: It is recommended to increase the operating pressure setpoint of the first-stage reverse osmosis unit by 5% from noon to evening tomorrow. This adjustment aims to compensate for the decrease in membrane separation efficiency caused by high temperature by increasing the driving force, thereby controlling the permeate conductivity within a safe range.

[0091] In another scenario, if the risk warning event is "the predicted vibration amplitude of the high-pressure pump has a 30% probability of exceeding the alarm threshold next week," and trend analysis reveals a continuous decline in its health status indicator, the agent determines that the risk stems from potential wear and tear on the equipment's mechanical components. In this case, the agent's recommended equipment operation is to generate a preventative maintenance work order. This order automatically generates a preventative maintenance work order containing the equipment number, recommended maintenance content, and recommended execution time window, and pushes it to the computerized maintenance management system. For example, the work order content might be: "Perform vibration detection and bearing inspection of the high-pressure pump in the ultrapure water system, recommended to be performed within 48 hours."

[0092] Furthermore, if the risk warning event is "during tomorrow's peak production period, there is a 20% probability that the water production capacity of the entire preparation system will be unable to simultaneously meet the peak demand of all workshops," the agent, through deduction, discovers that simply adjusting its own setpoints cannot resolve this systemic bottleneck. In this case, the agent generates an equipment operation suggestion that recommends adjusting the production schedule. This suggestion will generate a coordination proposal sent to the Manufacturing Execution System (MES) or production planning personnel, such as: "It is recommended that some non-critical and water-intensive production batches be rescheduled from 2 PM tomorrow to a lower-load period in the evening to mitigate the peak demand pressure on the preparation system."

[0093] Through the above implementation methods, the decision-making and early warning intelligent agent can output specific suggestions at different time dimensions and operational levels, ranging from short-term parameter fine-tuning and medium-term preventive maintenance to long-term production coordination, based on different risk scenarios. This transforms early warning information from simple alarm signals into operational instructions that drive different departments to take clear actions, thereby achieving closed-loop management from risk warning to proactive risk mitigation. This significantly improves the initiative and precision of semiconductor plant system management and effectively ensures the stable and reliable operation of critical support systems such as ultrapure water.

[0094] Example 2 like Figure 2 As shown, a data-driven semiconductor plant equipment risk early warning system includes: The historical data acquisition module is used to acquire multiple sets of historical data records from the historical operation and maintenance database of the plant equipment. Each set of historical data records includes the measured values ​​of key operating parameters of the plant equipment at a historical moment, the historical operating conditions corresponding to the historical moment, and the historical health status identifier that characterizes the health level of the plant equipment at the historical moment. The personalized prediction model training module is used to train a personalized performance prediction model based on the multiple sets of historical data records through the device profile intelligent agent. The combination of the historical operating conditions and the historical health status identifier is used as the model input features, and the measured values ​​of the key operating parameters are used as the target labels for model training. The model learns the mapping relationship from the input features to the target labels through machine learning algorithms. The future conditions acquisition and adaptation module is used to acquire production scheduling information and weather forecast data for a future target period, and to convert the production scheduling information into predicted load conditions for the plant equipment in the future target period, and to convert the weather forecast data into predicted environmental conditions for the plant equipment in the future target period; the predicted load conditions and the predicted environmental conditions together constitute the future operating conditions of the plant equipment in the future target period. The equipment operation risk simulation module is used to input the predicted load conditions, the predicted environmental conditions, and a preset simulation health status identifier into the personalized performance prediction model through a situational simulation intelligent agent, calculate the predicted sequence of key operating parameters of the plant equipment in the future target period, assess the equipment operation risk based on the predicted sequence of key operating parameters, and generate at least one risk warning event and its corresponding probability of occurrence. The early warning decision generation and output module is used to generate early warning information and equipment operation suggestions based on the risk warning event and its probability of occurrence through the decision and early warning intelligent agent, and output the early warning information and equipment operation suggestions.

[0095] The data-driven semiconductor plant equipment risk warning system provided in this embodiment has the same implementation principle and technical effects as the method embodiment in Embodiment 1. For the sake of brevity, any parts not mentioned in the system embodiment can be referred to the corresponding content in Embodiment 1.

[0096] Example 3 A computer-readable storage medium having a computer program stored thereon, the computer program being executed by a computer as described in Embodiment 1 above, which is a data-driven semiconductor plant equipment risk warning method.

[0097] Example 4 An electronic device includes: a memory and a processor, wherein the processor and the memory are connected; The memory is used to store programs; The processor calls a program stored in the memory to execute a data-driven semiconductor plant equipment risk warning method as described in Embodiment 1.

[0098] It should be noted that the electronic device mentioned may be, but is not limited to, personal computers (PCs), tablet computers, mobile internet devices (MIDs), etc.

[0099] It should be noted that processors, memory, and other components that may be present in electronic devices are electrically connected to each other, directly or indirectly, to enable data transmission or interaction. For example, processors, memory, and other components may be electrically connected to each other via one or more communication buses or signal lines.

[0100] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0101] In the several embodiments provided in this application, it should be understood that the disclosed systems and methods can also be implemented in other ways. The system embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0102] In addition, the functional modules in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0103] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, laptop, server, mobile phone, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0104] The above description is merely a preferred embodiment of the present invention and does not constitute any limitation on the present invention. Any simple modifications, alterations, or equivalent structural changes made to the above embodiments based on the technical essence of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A data-driven method for early warning of risks in semiconductor plant equipment, characterized in that, Includes the following steps: Multiple sets of historical data records are obtained from the historical operation and maintenance database of the plant equipment. Each set of historical data records includes the measured values ​​of key operating parameters of the plant equipment at a historical moment, the historical operating conditions corresponding to the historical moment, and the historical health status identifier that characterizes the health level of the plant equipment at the historical moment. By using a device profile intelligent agent, a personalized performance prediction model is trained based on the multiple sets of historical data records. The combination of the historical operating conditions and the historical health status identifier is used as the model input features, and the measured values ​​of the key operating parameters are used as the target labels for model training. The model learns the mapping relationship from the input features to the target labels through machine learning algorithms. The system acquires production scheduling information and weather forecast data for a future target period, and converts the production scheduling information into predicted load conditions for the plant equipment during the future target period, and the weather forecast data into predicted environmental conditions for the plant equipment during the future target period; the predicted load conditions and the predicted environmental conditions together constitute the future operating conditions of the plant equipment during the future target period. The situational simulation agent inputs the predicted load conditions, the predicted environmental conditions, and a preset simulation health status identifier into the personalized performance prediction model to calculate the predicted sequence of key operating parameters of the plant equipment in the future target period, and evaluates the equipment operation risk based on the predicted sequence of key operating parameters, generating at least one risk warning event and its corresponding probability of occurrence. The decision-making and early warning intelligent agent generates early warning information and equipment operation suggestions based on the risk warning events and their probability of occurrence, and outputs the early warning information and equipment operation suggestions.

2. The method according to claim 1, characterized in that, The plant equipment is a semiconductor production line system composed of multiple physical devices; the personalized performance prediction model is a system-level personalized performance prediction model trained for the semiconductor production line system. The acquired sets of historical data records are historical data records of the semiconductor production line system, wherein the measured values ​​of key operating parameters at the historical moments include the measured values ​​of key operating parameters at multiple key monitoring points in the semiconductor production line system; The trained system-level personalized performance prediction model is configured to output a set of predicted values ​​of key operating parameters corresponding to the multiple key monitoring points based on the input operating conditions and health status identifiers. The calculated key operating parameter prediction sequence is a set of key operating parameter prediction sequences corresponding to the multiple key monitoring points.

3. The method according to claim 1, characterized in that, Machine learning algorithms enable models to learn the mapping relationship from input features to target labels, including the following steps: Construct a training dataset, wherein the historical operating conditions and historical health status identifiers are combined into a feature vector X=[C,H], where C represents the historical operating condition vector and H represents the historical health status identifier vector; and the corresponding measured values ​​of the key operating parameters are used as the label vector Y. Initialize a regression prediction model f(·), whose output is the predicted value of the key operating parameters; The model training iteration is performed. In each iteration, the feature vector X is input into the current model f(·) to obtain the predicted value f(X). The loss value L=L(f(X),Y) between the predicted value f(X) and the true label Y is calculated, where L(·) is a preset loss function used to measure the difference between the predicted value and the true value. Based on the loss value L, the internal parameters θ of the regression prediction model f(·) are adjusted using an optimization algorithm to reduce the loss value; Repeat the model training iterations until the preset training stopping condition is met to obtain the trained personalized performance prediction model f. θ (·), where θ is the optimized model parameter.

4. The method according to claim 2 or 3, characterized in that, The historical health status identifier is a discrete classification label or continuous numerical feature generated based on the historical maintenance records, historical anomaly logs and / or historical component replacement records of the plant equipment, used to quantitatively represent the health status of the equipment at a historical moment.

5. The method according to claim 1, characterized in that, The process of converting production scheduling information into predicted load conditions includes: parsing the production plan and process type in the production scheduling information, calculating the resource load that the plant equipment needs to provide to support the production plan at various future time points based on a preset process load mapping relationship, and forming the predicted load conditions in time order.

6. The method according to claim 1, characterized in that, Converting weather forecast data into predicted environmental conditions includes: extracting ambient temperature, relative humidity, and wind speed information from the weather forecast data, using them as external environmental parameters that directly affect the plant equipment or its auxiliary systems, and forming the predicted environmental conditions in time order.

7. The method according to claim 1, characterized in that, Based on the predicted sequence of key operating parameters, assess the equipment operating risk and generate risk warning events and their probability of occurrence, including: Set dynamic safety thresholds for key operating parameters of plant equipment; Performing a Monte Carlo simulation involves multiple simulation iterations, with the following sub-steps executed in each iteration: Step a: Within the range of values ​​or probability distributions of at least one of the predicted load conditions, the predicted environmental conditions, and the preset inferred health status indicators, random sampling or setting is performed to obtain a set of input condition samples for the current simulation iteration. Step b: Input the input condition sample of the current simulation iteration into the personalized performance prediction model, calculate a set of predicted values ​​of key operating parameters, and determine whether any value in the set of predicted values ​​exceeds the dynamic safety threshold. If the risk exceeds the limit, it is determined that the aforementioned risk warning event has occurred in this simulation iteration; The total number of times the risk warning event occurred during the multiple simulation iterations is counted, and the ratio of the total number of occurrences to the total number of simulations is taken as the probability of the risk warning event occurring.

8. A data-driven semiconductor plant equipment risk early warning system, characterized in that, include: The historical data acquisition module is used to acquire multiple sets of historical data records from the historical operation and maintenance database of the plant equipment. Each set of historical data records includes the measured values ​​of key operating parameters of the plant equipment at a historical moment, the historical operating conditions corresponding to the historical moment, and the historical health status identifier that characterizes the health level of the plant equipment at the historical moment. The personalized prediction model training module is used to train a personalized performance prediction model based on the multiple sets of historical data records through the device profile intelligent agent. The combination of the historical operating conditions and the historical health status identifier is used as the model input features, and the measured values ​​of the key operating parameters are used as the target labels for model training. The model learns the mapping relationship from the input features to the target labels through machine learning algorithms. The future conditions acquisition and adaptation module is used to acquire production scheduling information and weather forecast data for a future target period, and to convert the production scheduling information into predicted load conditions for the plant equipment in the future target period, and to convert the weather forecast data into predicted environmental conditions for the plant equipment in the future target period; the predicted load conditions and the predicted environmental conditions together constitute the future operating conditions of the plant equipment in the future target period. The equipment operation risk simulation module is used to input the predicted load conditions, the predicted environmental conditions, and a preset simulation health status identifier into the personalized performance prediction model through a situational simulation intelligent agent, calculate the predicted sequence of key operating parameters of the plant equipment in the future target period, assess the equipment operation risk based on the predicted sequence of key operating parameters, and generate at least one risk warning event and its corresponding probability of occurrence. The early warning decision generation and output module is used to generate early warning information and equipment operation suggestions based on the risk warning event and its probability of occurrence through the decision and early warning intelligent agent, and output the early warning information and equipment operation suggestions.

9. An electronic device, characterized in that, include: A memory and a processor, wherein the processor and the memory are connected; The memory is used to store programs; The processor invokes a program stored in the memory to execute the method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, It stores a computer program, which is executed by a computer to perform the method as described in any one of claims 1-7.

Images