Clean room management and warning detection system and method

Abstract

This invention belongs to the field of particle concentration detection technology, specifically a cleanroom control alarm detection system and method. When a particle concentration alarm is triggered, a time window containing particle concentration and at least two types of environmental auxiliary parameters is captured and input into a pre-trained meta-learning model. The model internally maintains global prototype vectors corresponding to various anomaly causes. A feature extraction network compresses the multivariate sequence into alarm feature vectors. Based on the matching degree between the alarm feature vector and each global prototype vector, at least one anomaly cause with the highest matching degree is selected as the investigation direction. This method can automatically diagnose the root cause of anomalies from multi-dimensional environmental parameters, replacing manual investigation, significantly shortening anomaly response time, improving diagnostic efficiency, reducing reliance on personnel experience, and facilitating rapid on-site handling and reducing product contamination risks.

Description

Technical Field

[0001] This invention relates to the field of particle concentration detection technology, and in particular to a cleanroom control alarm detection system and method. Background Technology

[0002] Cleanrooms (also known as dust-free rooms or purification rooms) are indispensable controlled environments in high-end industrial fields such as semiconductor manufacturing, biomedicine, and precision machinery. Their core function is to strictly control the concentration of airborne particulate matter, harmful gases, and microorganisms within specified ranges, thereby ensuring product quality and the stability of the production process. Cleanroom environmental parameters mainly include particulate matter concentration (usually measured by the number of particles with diameters ≥0.3μm and ≥0.5μm as the cleanliness level criterion), temperature, humidity, pressure difference (between indoor and outdoor areas and between different clean areas), air intake velocity, air change rate, and airflow organization. Among these, particulate matter concentration is the core indicator for measuring the operational status of a cleanroom.

[0003] In actual operation, the causes of excessive particle concentration in cleanrooms are complex and diverse. Typical causes include: leakage or aging and clogging of HEPA filters leading to decreased filtration efficiency; insufficient fresh air volume or malfunctions in the supply / exhaust system causing inadequate air exchange in the cleanroom; imbalance of positive / negative pressure differential causing the infiltration of unfiltered external air or abnormal leakage of clean air; excessive personnel entry and exit frequency or improper operation in the cleanroom (such as large movements, improper clothing, etc.) leading to the release of additional particulate matter; abnormal temperature and humidity causing changes in particulate matter settling behavior; and sensor drift or aging causing false particle counts. However, the common manifestation of these various abnormal causes is an elevated particle monitor reading, making it a challenging problem of well-posed reasoning involving multiple factors.

[0004] The current industry practice is to trigger an audible and visual alarm or remote notification when the particle monitor reading is high, and then on-site maintenance personnel, based on their experience, systematically check the above-mentioned possible links, including checking the appearance of the filters and the differential pressure gauge readings, checking the operating parameters of the fresh air system, reviewing personnel entry and exit records, and verifying the calibration status of the particle counter. This manual troubleshooting method has significant drawbacks: First, the investigation process is highly dependent on the personal experience of the maintenance personnel and their familiarity with the specific cleanroom environment. Personnel turnover leads to significant fluctuations in investigation efficiency. Secondly, the process of checking each step one by one is time-consuming. In scenarios with extremely stringent cleanliness requirements, such as pharmaceutical production or semiconductor processes, delayed identification of the root cause of the anomaly means that the product is continuously exposed to the risk of contamination during this period, which may cause batch products to be scrapped or even force the interruption of the entire batch of production. Third, on-site maintenance personnel are limited by the boundaries of information acquisition (such as only having local instrument readings at any given time), making it difficult to establish a spatiotemporal correlation analysis perspective for multi-dimensional parameters. This can easily lead to overlooking potential hazards or misjudging the cause of anomalies. Even if some cleanrooms have deployed centralized monitoring systems, these systems typically only achieve data aggregation and display and threshold alarm functions. They fail to intelligently correlate alarm signals with the diagnosis of anomaly causes, and manual intervention is still required after an alarm is triggered, thus failing to substantially solve the aforementioned pain points.

[0005] Several patented solutions for cleanroom environmental monitoring already exist in the prior art. For example, patent application CN114502962A, entitled "Particle Detector with Remote Alarm Monitoring and Control," filed by Particle Monitoring Systems Co., Ltd., discloses a method for wirelessly transmitting alarm signals from a particle detector to a remote device and transmitting user commands back to the particle detector via a copied graphical interface on the remote device. This solution addresses the need for personnel to enter the cleanroom during on-site operation and reduces the possibility of human-induced contamination. However, its technological contribution focuses on the remote transmission and control of alarm signals, without addressing the issue of how to automatically analyze the cause of the anomaly after the alarm is triggered. After an alarm is generated, experienced engineers still need to remotely determine the abnormality, essentially remaining a manual diagnostic mode, merely shifting the diagnostic location from the on-site location to a remote terminal.

[0006] For example, the patent solution with publication number CN114486656A and titled "Dynamic Environmental Monitoring System for Medical Cleanrooms" obtains the correlation between concentration parameters and environmental parameters at each sampling point by performing sequence fluctuation analysis on concentration parameter sequences and environmental parameter sequences respectively, and judges the parameter status of each area based on the correlation to achieve early warning. This solution realizes the assessment of cleanroom environmental status based on multi-parameter fluctuation and correlation analysis and has certain data analysis functions. However, its early warning logic relies on preset statistical correlation rules. The formulation and parameter adjustment of the rules require manual participation and are difficult to generalize for different cleanrooms. At the same time, this solution focuses on judging the environmental compliance status rather than tracing and diagnosing the cause of abnormalities, and still fails to provide clear troubleshooting guidance to on-site personnel after the alarm occurs.

[0007] For example, the patent publication number CN120444710A, "AI-based Cleanroom Dynamic Energy Consumption Management System and Method," introduces artificial intelligence to optimize cleanroom management, but its technical objective focuses on energy consumption management rather than attribution diagnosis of abnormal alarms. The patent publication number CN121762121A, "FFU Filter Leakage Early Warning Trigger System and Method Based on Particle Counter," designs an early warning method based on isolated forests and gradient boosting trees for the specific anomaly type of filter leakage. While this improves the diagnostic accuracy for a single anomaly type, the diagnostic scope is limited to FFU filter leakage and cannot cover other diverse causes of excessive particle concentration in cleanrooms. Furthermore, after model training, it cannot adaptively adjust according to differences in deployment environments; different users need to train their own models independently, resulting in insufficient generalization and transferability.

[0008] Based on the above analysis, the existing cleanroom monitoring technologies have the following main shortcomings: First, the investigation of abnormal causes after an alarm relies on human experience and lacks automated root cause diagnosis capabilities, resulting in low efficiency and delayed response in abnormal handling. Second, existing solutions either focus on data display and transmission, are limited to the detection of a single type of anomaly, or rely on preset rules that cannot adapt to complex and ever-changing anomaly scenarios. They lack a systematic technical solution to establish an automated mapping relationship between multi-dimensional environmental parameters and multiple anomaly causes. Third, due to significant differences in the cleanroom structure layout, equipment configuration, air conditioning system type, and personnel operation mode among different customers, the characteristic patterns of abnormal performance are also customer-specific. Existing solutions generally lack mechanisms for rapid cross-customer adaptation and continuous self-optimization, making it difficult to effectively migrate between different user environments.

[0009] Therefore, there is an urgent need for a cleanroom control alarm detection system and method that can automatically analyze the causes of anomalies, quickly adapt to differences in user environments, and continuously self-optimize online, in order to address the shortcomings of the existing technologies mentioned above. Summary of the Invention

[0010] The purpose of this invention is to provide a cleanroom control alarm detection system and method to solve the problems mentioned in the background art.

[0011] To achieve the above objectives, the present invention provides the following technical solution: a cleanroom control alarm detection method, comprising the following steps: Acquire cleanroom environmental monitoring data, which includes at least particle concentration data and at least two types of environmental auxiliary parameter data; When the particle concentration data triggers the alarm condition, a multivariate environmental monitoring data sequence within a preset time window before the alarm trigger time is extracted. The multivariate environmental monitoring data sequence includes at least the particle concentration sequence and the environmental auxiliary parameter sequence. The multivariate environmental monitoring data sequence is input into the pre-trained meta-learning model. The model maintains a set of global prototype vectors that correspond one-to-one with multiple preset anomaly cause categories. The feature extraction network in the model compresses the multivariate environmental monitoring data sequence into alarm feature vectors. Based on the matching degree between the alarm feature vectors and each of the global prototype vectors, at least one anomaly cause category with the highest matching degree is selected from the multiple preset anomaly cause categories as the anomaly investigation direction. The anomaly investigation direction is then output to the alarm response terminal to assist on-site personnel in quickly locating the abnormal link and taking targeted measures.

[0012] The cleanroom control alarm detection method of the present invention includes at least two of the following preset abnormal cause categories: filter leakage, filter blockage, insufficient fresh air volume, frequent personnel entry and exit, abnormal pressure fluctuation, temperature and humidity out of control, and sensor drift.

[0013] The cleanroom control alarm detection method of the present invention further includes a user modulation matrix maintained within the overall model; the method also includes: Assign a user-personalized vector to the overall model deployed in the target cleanroom; The prototype offset is obtained by linearly mapping the user-personalized vector using the user modulation matrix. The prototype offset is superimposed on each of the global prototype vectors to obtain a user-specific prototype vector corresponding to the target cleanroom, which replaces the global prototype vector in the calculation of the matching degree. Different cleanrooms correspond to different user-personalized vectors. The same overall model achieves adaptive matching of abnormal patterns in different cleanrooms by configuring different user-personalized vectors, thereby improving the diagnostic accuracy of abnormal causes in different cleanroom environments.

[0014] The cleanroom control alarm detection method of the present invention includes a feature extraction network and a global prototype vector obtained by the overall model during the meta-learning pre-training stage through iterative training of the outer and inner loops using multiple different cleanroom environments as meta-learning tasks. The outer loop updates the parameters of the feature extraction network, the global prototype vector, and the user modulation matrix, while the inner loop independently updates the user-personalized vector corresponding to each meta-learning task. This enables the overall model to quickly adapt to new cleanroom environments with a small number of user samples after the outer loop training is completed.

[0015] The cleanroom control alarm detection method of the present invention further includes a sub-model local self-optimization step: Once the anomaly investigation direction is confirmed on-site and the true anomaly cause label is obtained, the alarm feature vector is associated with the true anomaly cause label and stored in the local sample cache pool; When the number of confirmed samples in the local sample cache pool reaches a preset threshold, the confirmed samples are used to perform local gradient updates on the user-personalized vector, while keeping the parameters of the feature extraction network, the global prototype vector, and the user modulation matrix unchanged. This allows the user-specific prototype vector to gradually converge to a state that matches the actual abnormal pattern distribution of the target cleanroom, thereby achieving local adaptive optimization of the sub-model without the need for manual rule adjustments.

[0016] The cleanroom control alarm detection method of the present invention wherein the local self-optimization step of the sub-model is performed in an online incremental update manner. Whenever a new confirmed sample is added, the user personalized vector is updated by gradient descent once or several times for that sample, without waiting for the number of samples in the local sample cache pool to accumulate to a preset threshold, so that the sub-model has the ability to track cleanroom environment drift online in real time.

[0017] The cleanroom control alarm detection method of the present invention further includes a cloud-based continuous evolution step for the overall model: Collect confirmed sample data from multiple different cleanroom deployment points after local self-optimization, as well as user-personalized vectors corresponding to each cleanroom deployment point; A new set of meta-learning tasks is constructed using the confirmed sample data and the user-personalized vectors; The new set of meta-learning tasks is used to perform meta-learning updates on the parameters of the feature extraction network, the global prototype vector, and the user modulation matrix of the overall model, so that the optimization experience of multiple cleanroom deployment points is continuously distilled into the overall model, thereby achieving continuous evolution of the overall model's compatibility and initial rapid adaptation capability.

[0018] The cleanroom control alarm detection method of the present invention, wherein the matching degree is measured by calculating the similarity between the alarm feature vector and each of the global prototype vectors, and the confidence score corresponding to each of the preset abnormal cause categories is calculated based on the similarity; the step of selecting at least one abnormal cause category with the highest matching degree as the direction of abnormal investigation specifically includes: The preset abnormal cause categories are arranged in descending order of confidence score. A preset number of abnormal cause categories with the highest ranking are selected as the direction of abnormal investigation. Each selected abnormal cause category and its corresponding confidence score are output together, so that on-site personnel can determine the investigation priority based on the confidence level.

[0019] In addition, the present invention also provides a cleanroom control alarm detection system, comprising: The data acquisition module is used to acquire cleanroom environmental monitoring data, which includes at least particle concentration data and at least two types of environmental auxiliary parameter data. An alarm detection module, which is communicatively connected to the data acquisition module, is used to extract a multivariate environmental monitoring data sequence within a preset time window before the alarm triggering time when the particle concentration data is detected to trigger an alarm condition. The multivariate environmental monitoring data sequence includes at least a particle concentration sequence and the environmental auxiliary parameter sequence. An anomaly diagnosis module is communicatively connected to the alarm detection module. The anomaly diagnosis module has a pre-trained meta-learning model deployed inside. The overall model maintains a set of global prototype vectors that correspond one-to-one with multiple preset anomaly cause categories. The anomaly diagnosis module is configured to: use the feature extraction network in the overall model to compress the multivariate environmental monitoring data sequence into alarm feature vectors; and based on the matching degree between the alarm feature vectors and each of the global prototype vectors, select at least one anomaly cause category with the highest matching degree from the multiple preset anomaly cause categories as the anomaly investigation direction. The alarm response module is communicatively connected to the anomaly diagnosis module and is used to output the anomaly investigation direction to the alarm response terminal to assist on-site personnel in quickly locating the abnormal link and taking targeted measures.

[0020] The cleanroom control alarm detection system of the present invention further includes, in its anomaly diagnosis module: The user-personalized vector storage unit is used to store the user-personalized vector corresponding to the current target cleanroom. The user modulation matrix unit is used to store the user modulation matrix obtained by the pre-training of the total model meta-learning; The prototype modulation unit is used to linearly map the user personalized vector using the user modulation matrix to obtain the prototype offset, and then superimpose the prototype offset with each of the global prototype vectors to obtain the user-specific prototype vector corresponding to the target cleanroom, so as to replace the global prototype vector in the calculation of the matching degree. The anomaly diagnosis module is configured with different user-personalized vectors for different cleanroom deployment points. The prototype modulation unit enables adaptive matching of the same overall model to different cleanroom anomaly patterns, thereby improving the diagnostic accuracy of anomaly causes in different cleanroom environments.

[0021] Compared with the prior art, the beneficial effects of the present invention are as follows: When a particle concentration alarm is triggered, multivariate time window data containing particle concentration and at least two types of environmental auxiliary parameters is automatically extracted. Using a pre-trained meta-learning model, feature extraction and global prototype vector matching are employed to automatically select at least one anomaly cause category with the highest matching degree as the investigation direction and output it. This method replaces the existing manual, step-by-step investigation mode, achieving automated mapping and intelligent diagnosis from multi-dimensional environmental parameters to multiple anomaly causes. It significantly shortens anomaly response time, improves investigation efficiency, reduces reliance on the personal experience of maintenance personnel, and helps to quickly guide on-site handling in the early stages of an alarm, reducing product contamination risks and production interruption losses. Attached Figure Description

[0022] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0023] Figure 1 This is a flowchart of the method steps of the present invention. Detailed Implementation

[0024] The terms "first," "second," "third," and "fourth," etc., used in the specification, claims, and accompanying drawings of this invention are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or apparatuses.

[0025] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0026] "Multiple" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. The character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0027] Furthermore, the terms indicating orientation, such as "up," "down," "left," "right," "upper end," "lower end," and "longitudinal," are all based on the posture and position of the device or equipment described in this solution during normal use.

[0028] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, a clear and complete description will be provided below in conjunction with the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the protection scope of the present invention.

[0029] like Figure 1 As shown, this embodiment discloses a cleanroom control alarm detection system and method, which can be deployed in cleanroom monitoring architectures in fields such as semiconductor manufacturing and biomedicine. In a typical cleanroom, particle monitors are distributed at multiple sampling points as needed, while environmental sensors collect auxiliary parameters such as inlet air velocity, indoor-outdoor pressure difference, temperature, and humidity in real time. All data is aggregated to edge computing devices or local servers via industrial buses or IoT gateways, forming a multivariate time-series data stream.

[0030] Particle monitors typically output the number of particles ≥0.3μm and ≥0.5μm in diameter per cubic foot or per cubic meter, reported at a certain sampling period (e.g., per minute). Alarm conditions can be set such that the particle count at any sampling point exceeds a preset cleanliness level limit (e.g., the upper limit for 0.5μm particle concentration specified by ISO Class 7) for several consecutive periods; an alarm event is triggered when the condition is met.

[0031] Once an alarm event is triggered, the alarm detection module automatically extracts a multivariate monitoring data sequence within a preset time window preceding the alarm trigger time. The window length can be set to 15 minutes, meaning it includes data from 15 sampling points. The extracted sequence has a dimension of T×C, where T is the number of sampling points within the window and C is the number of feature channels, including at least a particle concentration sequence and at least two types of environmental auxiliary parameter sequences. In this embodiment, five features are used: particle concentration, inlet air velocity, pressure difference, temperature, and humidity, i.e., C=5. Each feature dimension of the sequence undergoes online standardization, such as Z-score standardization using the moving mean and standard deviation, to eliminate dimensional differences.

[0032] The anomaly diagnosis module embeds a pre-trained meta-learning model. This model consists of a feature extraction network. The feature extraction network is composed of a global prototype vector memory matrix M, a user modulation matrix W, and a set of biases b. Designed as a lightweight temporal convolutional network. Specifically, It consists of two one-dimensional convolutional layers and a ReLU activation function, each layer containing 32 convolutional kernels with a kernel size of 3 and a stride of 1; followed by an adaptive global average pooling layer, which compresses the variable-length time dimension into a fixed-length feature vector. In this embodiment, d is set to 64. That is: h = (X), Where X is a standardized T×5 matrix and h is a 64-dimensional alarm feature vector.

[0033] Global Prototype Vector Memory Matrix Where K is the total number of preset abnormal cause categories. In this embodiment, K=7, and the abnormal cause categories include: filter leakage, filter blockage, insufficient fresh air volume, frequent personnel entry and exit, abnormal pressure fluctuations, temperature and humidity out of control, and sensor drift. Each row This represents the global prototype vector of the k-th type of anomaly in the feature space, signifying the typical multivariate temporal pattern of this type of anomaly. User modulation matrix. Where p is the dimension of the personalized vector, taken as p=16, which is much smaller than d. Bias vector It provides a learnable baseline confidence bias for each category of abnormal causes.

[0034] When the system is deployed to a specific cleanroom user, a personalized vector is assigned to that user. The initial value can be set to a zero vector. During diagnosis, the anomaly diagnosis module uses the prototype modulation unit to perform a linear mapping of u through the user modulation matrix W to obtain the prototype offset δ=Wu. Then, the user-specific prototype vector m'k = mk + δ is calculated. This offset can be understood as the overall shift in the performance characteristics of the same anomaly due to differences in the physical layout, equipment characteristics, and operating modes of different cleanrooms in the feature space. This shift is compensated by a translation vector δ shared by all prototypes and determined by the user context. The user-specific prototype vector is then used to replace the global prototype vector in subsequent matching.

[0035] The degree of matching between the alarm feature vector h and each user's unique prototype m'k is measured by the inner product similarity, i.e., the similarity score. Then, the sigmoid function is used to convert it into an independent confidence probability pk for each cause of the anomaly: pk=σ(sk+bk), Where σ(·) is the Sigmoid function (“ The function's independent variable is represented by sk + bk, where bk is the bias. Multi-label independent probabilities are used instead of Softmax normalization because a single alarm may be caused by multiple factors simultaneously, such as filter micro-leakage accompanied by abnormal pressure fluctuations. The sigmoid output allows the model to highlight multiple causes simultaneously.

[0036] The {pk} values ​​are sorted in descending order, and the top 3 (e.g., Top-3) anomaly cause categories are selected as the anomaly investigation directions. The anomaly diagnosis module packages the selected cause category labels and their corresponding confidence scores and sends them to the alarm response module. The alarm response module outputs the investigation direction and confidence level on the alarm response terminal (e.g., control room screen, maintenance personnel's handheld terminal) in the form of text, graphical interface, or push message. For example, it outputs: "Possible causes of current anomaly: 1. Filter leakage (confidence 92%); 2. Abnormal pressure difference fluctuation (confidence 67%); 3. Sensor drift (confidence 31%)." On-site personnel can then prioritize checking the filter integrity and frame sealing, thereby significantly shortening the investigation time.

[0037] The aforementioned overall model requires meta-learning pre-training before use to enable rapid adaptation across cleanrooms. The pre-training phase collects alarm samples from multiple different cleanroom environments (which may come from historical data from different clients or different operational tasks divided into different time periods for the same client). Each cleanroom constitutes a meta-learning task Ti, and each task contains a small number of samples labeled with abnormal cause tags. The samples for each task are divided into a support set Si (for inner loop adaptation) and a query set Qi (for outer loop evaluation). The meta-learning pre-training process is as follows: Initialize the global parameters of the overall model Φ = {θ, M, W, b}, as well as the user-personalized vector ui (initially zero) and bias bi (each task maintains an independent copy of the bias; in this example, for simplicity, a copy of the global b is used directly). Extract a batch of tasks in each iteration.

[0038] For each task Ti in the batch: 1. Inner loop adaptation: With Φ fixed, N steps of gradient descent updates are performed on ui and bi using only the support set Si (N can be 5), and the loss function is binary cross-entropy. , in ∈{0,1} represents the true label. This is the Sigmoid output calculated based on the current UI, BI, and global parameters. Updates only adjust the UI and BI to obtain the adapted UI and BI.

[0039] 2. Outer loop evaluation: With the adapted ui and bi fixed, calculate the model's loss on the query set Qi. The gradient is calculated and accumulated, but Φ is not updated immediately.

[0040] After accumulating gradients in both the inner and outer loops for all tasks in a batch, the Adam optimizer is used to update Φ once. This is similar to the nested inner and outer loops of the Reptile meta-learning strategy, and its training effect is equivalent to guiding the global parameters Φ towards a direction that allows each task to perform well on the query set after only a few inner loop updates. After pre-training, the feature extraction network θ, the global prototype M, and the modulation matrix W of the overall model are solidified into a deployable state, possessing good initial adaptability across tasks.

[0041] When the overall model is deployed to a brand new cleanroom user, the system only needs to create a new user-personalized vector u (initially 0) for that user and can directly access the diagnostic service. During field operation, once the troubleshooting direction given by the model is confirmed by on-site personnel as the actual cause, the system stores the corresponding alarm feature vector h and the confirmed anomaly cause label y in the local sample cache pool.

[0042] When the local self-optimization startup conditions of the sub-model are met (e.g., the number of samples in the cache pool reaches a preset threshold of 20, or online incremental updates are performed immediately upon receiving each sample as described in claim 6), local parameter fine-tuning is performed. During the fine-tuning process, the feature extraction network parameters θ, the global prototype matrix M, and the modulation matrix W are all frozen. Only the user-specific vector u and the bias b are used as trainable parameters. A small number of gradient descent steps (e.g., 10 steps) are performed using samples from the cache pool, and the loss function is also binary cross-entropy. After the update, the user-specific prototype m'k = mk + Wu shifts, gradually converging to a region that matches the user's actual abnormal pattern, thus improving the accuracy of abnormality cause classification. Since only u and b with extremely low dimensions are updated each time (the total number of parameters is less than 200 floating-point numbers), the local optimization has extremely low requirements for computational and storage resources and can be completed entirely on the edge gateway or even the particle monitor's own microprocessor.

[0043] In the online incremental update mode, each new confirmed sample is immediately used to perform one or more gradient descents, enabling the sub-model to continuously adapt to the slow evolution of the cleanroom environment, such as the drift of wind speed-pressure difference characteristics caused by the increase of filter resistance over time, seasonal temperature and humidity baseline changes, etc., and to maintain high diagnostic accuracy.

[0044] Meanwhile, the cloud server periodically (e.g., weekly) collects locally optimized, anonymized data from each cleanroom deployment point, including alarm feature vectors h, confirmation labels y, and each user's current personalized vector u. This collected multi-user data is used to construct a new set of meta-learning tasks, with each user still treated as a task. Using the same meta-learning framework, the cloud server re-executes the aforementioned outer loop update based on the current global parameters, updating the feature extraction network θ, the global prototype M, the modulation matrix W, and the bias b. After the update, the cloud server distributes the new overall model to each node, completing the continuous evolution of the overall model. This mechanism allows newly deployed users to immediately benefit from the improved diagnostic performance brought by the accumulated multi-user collective experience, achieving a spiral increase in the overall model's compatibility and rapid adaptability.

[0045] This invention achieves deep personalized adaptation across different cleanrooms with extremely low user parameter requirements through a decoupling design of global prototype vectors and individual vectors. Furthermore, it endows the system with continuous self-learning and evolution capabilities through meta-learning pre-training and a local-cloud dual-layer self-optimization mechanism. These features enable the cleanroom control alarm detection method and system to be rapidly deployed and accurately diagnosed even when facing vastly different customer needs, and to continuously self-optimize with increasing usage time. This effectively solves the problems of low automation, difficulty in personalized adaptation, and outdated models that cannot keep up with environmental changes in existing technologies.

[0046] It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A method for detecting alarms in cleanroom control, characterized in that, Includes the following steps: Acquire cleanroom environmental monitoring data, which includes at least particle concentration data and at least two types of environmental auxiliary parameter data; When the particle concentration data triggers the alarm condition, a multivariate environmental monitoring data sequence within a preset time window before the alarm trigger time is extracted. The multivariate environmental monitoring data sequence includes at least the particle concentration sequence and the environmental auxiliary parameter sequence. The multivariate environmental monitoring data sequence is input into the pre-trained meta-learning model. The model maintains a set of global prototype vectors that correspond one-to-one with multiple preset anomaly cause categories. The feature extraction network in the model compresses the multivariate environmental monitoring data sequence into alarm feature vectors. Based on the matching degree between the alarm feature vectors and each of the global prototype vectors, at least one anomaly cause category with the highest matching degree is selected from the multiple preset anomaly cause categories as the anomaly investigation direction. The anomaly investigation direction is then output to the alarm response terminal to assist on-site personnel in quickly locating the abnormal link and taking targeted measures.

2. The cleanroom control alarm detection method according to claim 1, characterized in that, The preset abnormality cause categories include at least two of the following: filter leakage, filter blockage, insufficient fresh air volume, frequent personnel entry and exit, abnormal pressure fluctuation, temperature and humidity out of control, and sensor drift.

3. The cleanroom control alarm detection method according to claim 1, characterized in that, The overall model also maintains a user modulation matrix; the method further includes: Assign a user-personalized vector to the overall model deployed in the target cleanroom; The prototype offset is obtained by linearly mapping the user-personalized vector using the user modulation matrix. The prototype offset is superimposed on each of the global prototype vectors to obtain a user-specific prototype vector corresponding to the target cleanroom, which replaces the global prototype vector in the calculation of the matching degree. Different cleanrooms correspond to different user-personalized vectors. The same overall model achieves adaptive matching of abnormal patterns in different cleanrooms by configuring different user-personalized vectors, thereby improving the diagnostic accuracy of abnormal causes in different cleanroom environments.

4. The cleanroom control alarm detection method according to claim 3, characterized in that, The feature extraction network and the global prototype vector are obtained by the overall model during the meta-learning pre-training stage through iterative training of the outer and inner loops using multiple different cleanroom environments as meta-learning tasks. The outer loop is used to update the parameters of the feature extraction network, the global prototype vector, and the user modulation matrix, while the inner loop is used to independently update the user-personalized vector corresponding to each meta-learning task. This enables the overall model to quickly adapt to new cleanroom environments with a small number of user samples after the outer loop training is completed.

5. The cleanroom control alarm detection method according to claim 4, characterized in that, It also includes a local self-optimization step for the sub-model: Once the anomaly investigation direction is confirmed on-site and the true anomaly cause label is obtained, the alarm feature vector is associated with the true anomaly cause label and stored in the local sample cache pool; When the number of confirmed samples in the local sample cache pool reaches a preset threshold, the confirmed samples are used to perform local gradient updates on the user-personalized vector, while keeping the parameters of the feature extraction network, the global prototype vector, and the user modulation matrix unchanged. This allows the user-specific prototype vector to gradually converge to a state that matches the actual abnormal pattern distribution of the target cleanroom, thereby achieving local adaptive optimization of the sub-model without the need for manual rule adjustments.

6. The cleanroom control alarm detection method according to claim 5, characterized in that, The sub-model local self-optimization step is performed in an online incremental update manner. Whenever a new confirmed sample is added, the user personalized vector is updated by gradient descent once or several times for that sample, without waiting for the number of samples in the local sample cache pool to accumulate to a preset threshold, so that the sub-model has the ability to track cleanroom environment drift online in real time.

7. The cleanroom control alarm detection method according to claim 5, characterized in that, It also includes the overall model's continuous evolution steps in the cloud: Collect confirmed sample data from multiple different cleanroom deployment points after local self-optimization, as well as user-personalized vectors corresponding to each cleanroom deployment point; A new set of meta-learning tasks is constructed using the confirmed sample data and the user-personalized vectors; The new set of meta-learning tasks is used to perform meta-learning updates on the parameters of the feature extraction network, the global prototype vector, and the user modulation matrix of the overall model, so that the optimization experience of multiple cleanroom deployment points is continuously distilled into the overall model, thereby achieving continuous evolution of the overall model's compatibility and initial rapid adaptation capability.

8. The cleanroom control alarm detection method according to claim 1, characterized in that, The matching degree is measured by calculating the similarity between the alarm feature vector and each of the global prototype vectors, and a confidence score is calculated based on the similarity for each of the preset anomaly cause categories; the step of selecting at least one anomaly cause category with the highest matching degree as the direction of anomaly investigation specifically includes: The preset abnormal cause categories are arranged in descending order of confidence score. A preset number of abnormal cause categories with the highest ranking are selected as the direction of abnormal investigation. Each selected abnormal cause category and its corresponding confidence score are output together, so that on-site personnel can determine the investigation priority based on the confidence level.

9. A cleanroom control alarm detection system, characterized in that, include: The data acquisition module is used to acquire cleanroom environmental monitoring data, which includes at least particle concentration data and at least two types of environmental auxiliary parameter data. An alarm detection module, which is communicatively connected to the data acquisition module, is used to extract a multivariate environmental monitoring data sequence within a preset time window before the alarm triggering time when the particle concentration data is detected to trigger an alarm condition. The multivariate environmental monitoring data sequence includes at least a particle concentration sequence and the environmental auxiliary parameter sequence. An anomaly diagnosis module is communicatively connected to the alarm detection module. The anomaly diagnosis module has a pre-trained meta-learning model deployed inside. The overall model maintains a set of global prototype vectors that correspond one-to-one with multiple preset anomaly cause categories. The anomaly diagnosis module is configured to: use the feature extraction network in the overall model to compress the multivariate environmental monitoring data sequence into alarm feature vectors; and based on the matching degree between the alarm feature vectors and each of the global prototype vectors, select at least one anomaly cause category with the highest matching degree from the multiple preset anomaly cause categories as the anomaly investigation direction. The alarm response module is communicatively connected to the anomaly diagnosis module and is used to output the anomaly investigation direction to the alarm response terminal to assist on-site personnel in quickly locating the abnormal link and taking targeted measures.

10. The cleanroom control alarm detection system according to claim 9, characterized in that, The anomaly diagnosis module also includes: The user-personalized vector storage unit is used to store the user-personalized vector corresponding to the current target cleanroom. The user modulation matrix unit is used to store the user modulation matrix obtained by the pre-training of the total model meta-learning; The prototype modulation unit is used to linearly map the user personalized vector using the user modulation matrix to obtain the prototype offset, and then superimpose the prototype offset with each of the global prototype vectors to obtain the user-specific prototype vector corresponding to the target cleanroom, so as to replace the global prototype vector in the calculation of the matching degree. The anomaly diagnosis module is configured with different user-personalized vectors for different cleanroom deployment points. The prototype modulation unit enables adaptive matching of the same overall model to different cleanroom anomaly patterns, thereby improving the diagnostic accuracy of anomaly causes in different cleanroom environments.

Images