Environment anomaly detection system and environment anomaly detection method

Abstract

The present disclosure relates to an environment anomaly detection system and an environment anomaly detection method. The environment anomaly detection system comprises an air conditioning device, a sensor and a computing module. The air conditioning device is used for providing air flow to an environment. The sensor is used for obtaining feature data about the air flow. The computing module is used for inputting the feature data into each of a plurality of machine learning models to obtain a plurality of first prediction labels, and inputting the first prediction labels into an integration model to obtain a second prediction label, wherein the second prediction label is used for indicating whether the air flow is abnormal.

Description

Environmental anomaly detection system and environmental anomaly detection method

[0001] This application claims priority to Chinese patent application number 202311501172.2, filed on November 13, 2023, entitled “Environmental Anomaly Detection System and Environmental Anomaly Detection Method,” and all of its contents are incorporated herein by reference. Technical Field

[0002] The present disclosure relates to systems and methods for detecting environmental anomalies using machine learning methods. Background Art

[0003] A cleanroom is a specialized laboratory or production environment primarily used in fields such as semiconductor manufacturing, biotechnology, pharmaceuticals, and precision engineering, where extremely stringent environmental control requirements are imposed. Within a cleanroom, the air conditioning system not only controls temperature but also humidity, which is crucial for ensuring successful cleanroom operations and product quality. For example, during the photolithography process, even slight temperature fluctuations can alter the properties of the photoresist, affecting the precision of the pattern on the wafer. Proper humidity prevents static electricity, which is particularly important when handling microelectronic components. Furthermore, excessively high or low humidity can interfere with certain chemical reactions, impacting product quality. The air conditioning system also removes dust and particles from the air through filtration, maintaining the cleanroom's cleanliness level. Therefore, the air conditioning system plays a critical role in cleanrooms, ensuring not only product quality but also the stability and efficiency of the production process. Detecting air flow anomalies in this environment is a crucial technical challenge.

[0004] Summary of the Invention

[0005] Embodiments of the present disclosure provide an environmental anomaly detection system for use in air quality monitoring within an environment. The environmental anomaly detection system includes an air conditioning unit, a sensor, and a computing module. The air conditioning unit provides airflow to the environment. The sensor acquires characteristic data about the airflow. The computing module is communicatively connected to the sensor and inputs the characteristic data into each of a plurality of machine learning models to obtain a plurality of first prediction labels. The computing module inputs the first prediction labels into an integrated model to obtain a second prediction label, which indicates whether the airflow is abnormal.

[0006] In some embodiments, the number of sensors is greater than 1, wherein one sensor is disposed in an air intake area of ​​the air conditioning device through which air flows into the environment, and another sensor is disposed in the environment.

[0007] In some embodiments, the characteristic data includes temperature and humidity.

[0008] In some embodiments, the calculation module is further used to obtain the outdoor temperature and the load rate of the air-conditioning equipment. The above-mentioned characteristic data also includes the outdoor temperature and the load rate.

[0009] In some embodiments, the above-mentioned machine learning model includes a clustering algorithm, a time series prediction algorithm, and a random forest algorithm.

[0010] In some embodiments, a computing module is configured to calculate a classification metric for each machine learning model during a training phase. The computing module adds a constant to each classification metric to calculate the weight of the machine learning model. The ensemble model is configured to sum the product of the machine learning model weights and the first predicted label to obtain a predicted value, and to determine whether the predicted value is greater than a threshold to generate a second predicted label.

[0011] In some embodiments, the aforementioned critical value is set to be less than or equal to 0.6.

[0012] In some embodiments, the aforementioned critical value is set to be less than or equal to 0.3.

[0013] In some embodiments, the aforementioned classification index is calculated according to the following mathematical formula.

[0014] Where S is the classification index. i, j, j k , k is a positive integer. m represents the number of samples. n represents the number of samples in the category to which the i-th sample belongs. The i-th sample and the j-th sample belong to the same category. n k Indicates the number of samples in the kth category, the number of samples in the i-th category and the number of samples in the j-th category k samples belong to different categories. distance() is used to calculate the distance between two samples.

[0015] In some embodiments, the calculation module is used to calculate the weight according to the following mathematical formula.

[0016] Where x is a positive integer, S x is the classification index of the x-th machine learning model, w x is the weight of the x-th machine learning model.

[0017] In some embodiments, the above-mentioned machine learning model has multiple model parameters, and the computing module is used to retrain the machine learning model and the integrated model based on new feature data at default time intervals to update the model parameters, classification indicators and weights.

[0018] From another perspective, embodiments of the present disclosure provide a method for detecting an environmental anomaly, applicable to an environment and an air conditioning device that provides airflow to the environment. The method, executed by a computing module, includes: obtaining characteristic data regarding the airflow provided by the air conditioning device via a sensor; inputting the characteristic data into each of a plurality of machine learning models to obtain a plurality of first prediction labels; and inputting the first prediction labels into an integrated model to obtain a second prediction label indicating whether the airflow is abnormal. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to make the above and other objects, features, advantages and embodiments of the present invention more easily understood, the accompanying drawings are now described as follows:

[0020] FIG. 1 is a schematic diagram illustrating an environmental anomaly detection system according to an embodiment.

[0021] FIG. 2 is a flow chart illustrating a method for detecting an environmental anomaly according to an embodiment.

[0022] FIG3 is a schematic diagram illustrating the operation of a machine learning model and an integrated model according to one embodiment.

[0023] 4 and 5 are schematic diagrams illustrating experimental results according to one embodiment. DETAILED DESCRIPTION

[0024] The terms “first”, “second”, etc. used herein do not particularly refer to an order or sequence, but are only used to distinguish elements or operations described with the same technical terms.

[0025] FIG1 is a schematic diagram illustrating an environmental anomaly detection system according to one embodiment. Referring to FIG1 , the environmental anomaly detection system includes an air conditioning unit 110, sensors 121-126, and a computing module 130. This system is suitable for air monitoring in an environment 140. In this embodiment, the environment 140 is a clean room, but in other embodiments, it can also be a factory, office, storage room, etc., and the present disclosure is not limited thereto. The air conditioning unit 110 is used to provide an air flow 111 to the environment 140. The air conditioning unit 110 may include a heater, a humidifier, a compressor, a fan, a cooling tower, a cold water valve, etc., but the present disclosure is not limited thereto. The sensors 121-126 may be temperature sensors, humidity sensors, atmospheric pressure sensors, etc., but the present disclosure is not limited thereto. In this embodiment, sensor 121 is located in the air intake area 150 of the air conditioning unit 110. The air flow 111 enters the environment 140 through the air intake area 150. The other sensors 122-126 are located within the environment 140, but the present disclosure does not limit the number or placement of the sensors. Computing module 130 can be a personal computer, laptop, server, industrial computer, control center, central processing unit, or any other component or electronic device with computing capabilities. Computing module 130 is communicatively connected to sensors 121-126. This communication connection can be achieved through any wired or wireless means. Computing module 130 is used to execute an environmental anomaly detection method, which will be described in detail below.

[0026] FIG2 is a flow chart illustrating an environmental anomaly detection method according to one embodiment. Referring to FIG2 , in step 201, characteristic data regarding the air flow 111 provided by the air conditioner 110 is obtained via sensors. This characteristic data, for example, is the temperature and humidity of the air flow 111. In the aforementioned embodiment, the number of at least one sensor is one, located only in the air intake area 150 of the air conditioner 110. However, in some embodiments, the number of at least one sensor may be greater than one, such as two. In this manner, one or more of these sensors may be located in the environment 140. The characteristic data may also include, for example, the outdoor temperature and the load rate of the air conditioner 110. If the corresponding sensors 122-126 are located within the environment 140, the characteristic data also includes the temperature and humidity within the environment 140. Therefore, the environmental anomaly detection system can not only detect the temperature and humidity of the air flow provided by the air conditioner 110, but also compare it to the actual conditions of the environment 140, such as the outdoor temperature, while simultaneously monitoring the load rate of the air conditioner 110. This allows for a more comprehensive anomaly detection effect that simultaneously considers both the environment 140 and the load rate of the air conditioner 110, rather than simply detecting anomalies based on the temperature and humidity of the airflow provided by the air conditioner 110. In some embodiments, feature data sensed over a period of time, such as one or more months, weeks, days, or hours, is continuously acquired. This period of time is not limited to this disclosure. The feature data sensed over a period of time is combined into a vector for subsequent processing. In some embodiments, these feature data may also be pre-processed, such as by denoising, filling in missing values, and normalizing.

[0027] This allows us to leverage machine learning algorithms and historical data from temperature and humidity sensors to build models that more accurately detect changes in air conditioning equipment. We then validate the model's accuracy and real-time performance through multiple simulated anomaly experiments, allowing us to apply it to different time periods.

[0028] Finally, a blending ensemble learning approach is used to build the final anomaly detection model. This method combines the advantages of multiple models, not only detecting outliers faster but also improving accuracy.

[0029] Next, the process of how this embodiment performs machine learning and integration model is described in detail.

[0030] FIG3 is a schematic diagram illustrating the operation of a machine learning model and an integrated model according to one embodiment. Referring to FIG2 and FIG3 , in step 202 , the feature data 310 obtained above is input into each of a plurality of machine learning models 321 - 323 to obtain a plurality of prediction labels 331 - 333 . These machine learning models 321 - 323 can be clustering algorithms, time series prediction algorithms, decision tree algorithms, multi-level neural networks, convolutional neural networks, support vector machines (SVMs), and the like. Among them, the clustering algorithm can be a density-based spatial clustering of applications with noise (DBSCAN), which can automatically divide data into several groups and classify low-density groups as abnormalities. It can effectively identify high-density areas and exclude noise points. In addition, methods such as Ordering points to identify the clustering structure (OPTICS), H-DBSCAN, Local Outlier Factor (LOF), Connectivity-Based Outlier Factor (COF), and One-Class SVM can also be used. One of the time series prediction algorithms is FBprophet, which is open sourced by Facebook. FBprophet is an additive time series decomposition technology that can decompose time series into seasonality, trend, and non-periodicity. It then uses an adaptive regression model for prediction. In addition, methods such as the Autoregressive Integrated Moving Average model (ARIMA) and Moving Average (MA) can also be used. Decision tree algorithms can use either the Isolation Forest or Random Forest algorithms. The concept of the Isolation Forest algorithm is to partition the training set into multiple subsets using random trees (binary trees). Because outliers are mostly sparsely distributed and far away from high-density groups, they are easier to segment early on, thus detecting outliers in the data. The Random Forest algorithm constructs multiple decision trees and introduces random feature sampling, meaning that each time a node is split, the optimal feature is selected from a random subset of the feature space.After all decision trees are built, classification / regression is performed on the test sample. Each decision tree makes a prediction independently, and finally a vote is taken based on the output of all trees to select the classification / regression value with the most votes.

[0031] The predicted labels 331-333 indicate whether the air flow 111 is abnormal. During the training phase, any component or parameter of the air-conditioning device 110 can be deliberately adjusted to form a true label (ground truth) representing the abnormality. For example, the heater, humidifier, and / or ice water valve in the air-conditioning device 110 can be turned off, or the target temperature and humidity can be adjusted to the abnormal range, etc. At this time, the corresponding true label is "abnormal". In this embodiment, the predicted labels 331-333 are discrete labels with a value of "1" or "0" to indicate abnormality or no abnormality. In other embodiments, the predicted labels 331-333 can also be continuous values ​​to indicate the probability of air flow abnormality. In other words, the machine learning models 321-323 can be used for classification or regression, and the present disclosure is not limited to this.

[0032] In step 203, predicted labels 331-333 are input into ensemble model 340 to generate predicted label 341, which indicates whether the airflow is abnormal. Ensemble model 340 is used to combine the outputs of multiple machine learning models 321-323 to produce a more accurate output. For example, a weight can be assigned to each machine learning model 321-323. Ensemble model 340 sums the products of the weights of the machine learning models 321-323 and the predicted labels 331-333 to generate a predicted value, as expressed in the following mathematical formula 1.

[0033] [Mathematical formula 1]

[0034] in is the predicted value. x is a positive integer. y x is the xth predicted label among the predicted labels 331-333. x is the weight corresponding to the xth machine learning model. In some embodiments, the weights w of the machine learning models 321-323 are x The weighting is determined based on the accuracy of each machine learning model 321-323, with higher-accuracy models assigned higher weights. Specifically, during the training phase, each feature data item forms a sample, which is then classified by the machine learning models 321-323 as either "abnormal" or "no abnormality." Each machine learning model calculates a classification metric based on its classification results. This classification metric is calculated, for example, according to the following mathematical formulas 2-5.

[0035] [Mathematical formula 2]

[0036] [Mathematical formula 3]

[0037] [Formula 4]

[0038] [Formula 5]

[0039] Where S is the classification index of a machine learning model. i, j, j k , k is a positive integer. m represents the number of all samples, n represents the number of samples in the category to which the i-th sample belongs, where the i-th sample and the j-th sample belong to the same category. k Indicates the number of samples in the kth category, the number of samples in the i-th category and the number of samples in the j-th category k Samples belong to different categories. distance() is used to calculate the distance between two samples. For example, it can calculate the Euler distance between the vectors formed by the feature data. In other words, Equation 5 calculates the distance between two samples belonging to different categories. A larger distance indicates that the two categories are separated more widely, and a larger calculated value b(i) indicates a better classification result. Equation 4 calculates the distance between two samples in the same category. A smaller distance indicates that the samples in the same category are more concentrated, and a smaller calculated value a(i) indicates a better classification result. Equation 3 combines the values ​​a(i) and b(i). When the value b(i) is larger and the value a(i) is smaller, the calculated value s(i) is larger, indicating a better classification result. Finally, Equation 2 averages the values ​​s(i) corresponding to all samples to form the classification metric S of a machine learning model. Each machine learning model 321-323 has its own corresponding classification metric S. A larger classification metric S indicates a better classification result.

[0040] The numerical range of the classification index S is [-1, 1]. A constant (e.g., 1) can be added to each of the classification indexes to adjust the numerical range of the classification index S to [0, 2]. This is used to calculate the weights of the machine learning models 321-323. The larger the classification index S, the larger the weights. In some embodiments, the weights are calculated as shown in the following mathematical formula 6.

[0041] [Formula 6]

[0042] Among them S x is the classification index of the x-th machine learning model. In other words, Mathematical Formula 6 first adds 1 to each classification index, and then divides it by the sum of the classification indexes corresponding to all machine learning models 321-323 to perform normalization and calculate the weight w x.

[0043] In some embodiments, the weights of the machine learning models 321-323 can be determined in an iterative manner, and the weights of the samples are adjusted in each iteration. represents the weight of the xth machine learning model at the i-th iteration. In this embodiment, the above mathematical formula 1 can be rewritten as the following mathematical formula 7.

[0044] [Formula 7]

[0045] The following uses β n Represents the weight of the nth sample, all weights β at the first iteration n Next, the training samples are input into the machine learning model to calculate the error rate of each machine learning model. This error rate is defined as the weight of the sample multiplied by the classification result, expressed as the following mathematical formula 8.

[0046] [Formula 8]

[0047] where f x (n) represents the classification result of the x-th machine learning model for the n-th training sample. When the classification is correct, f x (n)=0, when the classification is wrong, f x (n) = 1. Then calculate the weight based on the error rate As shown in the following mathematical formula 9. Finally, the weight β of the sample is updated according to the classification results of all machine learning models in all iterations n , as shown in the following mathematical formula 10.

[0048] [Formula 9]

[0049] [Formula 10]

[0050] where β′ n Represents the updated weight. It represents the classification results of all machine learning models 321-323 on the nth sample after weighting. When the classification is correct, it is set to 0, and when the classification is wrong, it is set to 1. γ is a constant. n After updating, all weights β can be n Regularize, for example, by dividing all weights β n The sum of the values ​​of , and then the next iteration can be performed (i = i + 1). In formula 8, the error rate is calculated based on the weight of the sample, so the sample with a larger weight will contribute more error rate. Then in formula 9, the error rate ε is calculated based on the error rate x To determine the weight When the error rate ε x The smaller the weight The larger the error, the larger the error. Finally, in Equation 10, the sample weights are updated. If the classification is incorrect, the weight is increased, and if the classification is correct, the weight is decreased. This allows the next iteration to focus on previously misclassified samples. Iterations can be set to a predetermined number of times, or stopped when the error rate falls below a certain value. This approach not only combines multiple machine learning models, but also combines the classification results of multiple iterations, thereby combining multiple weak classifiers into a single strong classifier.

[0051] Whether it is Mathematical Formula 2 or Mathematical Formula 7, when calculating the predicted value Later, judge the predicted value Is it greater than a critical value to generate a prediction label 341, for example, when the predicted value When the value is greater than the critical value, the prediction label 341 is set to "1" to indicate an abnormality, otherwise the prediction label 341 is set to "0". The numerical range of the critical value is [0,1]. In this embodiment, the critical value can be set according to the experimental results. Figures 4 and 5 are schematic diagrams illustrating the experimental results according to one embodiment. Please refer to Figures 4 and 5. Chart 400 is about temperature abnormality detection, and chart 500 is about humidity abnormality detection. In experiment 1, the air conditioning equipment 110 was turned off at 09:00. The above detection method can detect the abnormality within a few minutes after the equipment is turned off. For example, when the critical value is 0.1-0.3, the temperature abnormality can be detected at 09:05, when the critical value is 0.4-0.9, the temperature abnormality can be detected at 09:06, and when the critical value is 0.1-0.9, the humidity abnormality can be detected at 09:06, and so on. In Experiment 2, air conditioner 110 was shut down at 09:00; in Experiment 3, air conditioner 110 was shut down at 14:00; in Experiment 4, air conditioner 110 was shut down at 09:10; in Experiment 5, air conditioner 110 was shut down at 14:30; and in Experiment 6, air conditioner 110 was shut down at 09:10. Based on the results of Graphs 400 and 500, in some embodiments, the critical value is set to less than or equal to 0.6. All experimental results can obtain correct abnormality detection results and enter the alarm phase. Preferably, in other embodiments, the critical value is set to less than or equal to 0.3. This can detect temperature and humidity abnormalities earlier (approximately 1 to 2 minutes earlier) and enter the alarm phase earlier.

[0052] When the air flow 111 is determined to be abnormal, the computing module 130 can send a message through any human-machine interface or device, such as displaying specific text, numbers, or patterns on a display screen, emitting light from a light source, emitting a specific sound through a speaker, or sending a text message to a relevant person's mobile phone. In this way, the relevant operator can be notified of the abnormality in the air flow 111 and arrange for equipment maintenance.

[0053] Since data such as temperature and humidity may change with environmental conditions such as seasons and weather, the machine learning models 321-323 and the integrated model 340 trained above must also be adaptively adjusted. Each machine learning model has multiple model parameters, such as the weights of neurons in a neural network, and the critical values ​​for determining which branch to take in a random forest. These model parameters must be recalculated to meet the changed environmental conditions. Similarly, the classification indicators and weights in the integrated model 340 must also be recalculated. In some embodiments, the machine learning models 321-323 and the integrated model 340 above can be retrained based on new feature data at regular intervals (e.g., weeks, months, or any time) to update the model parameters, classification indicators S x With weight w x This allows us to adjust model parameters when unexpected deviations occur in the data (cold snaps, foehn winds, etc.), and recalibrate the model based on the new data. In other words, regular model retraining maintains a high level of model health, ensuring that the model can adapt to changing circumstances.

[0054] More specifically, this embodiment uses historical data from the two weeks preceding the forecast date as a training set to build the model. Because differences in model parameters can affect the accuracy of anomaly detection, nested cross-validation is used to select the optimal model parameters each time the model is updated to improve accuracy. To ensure that model accuracy is not reduced due to changes in data distribution, these machine learning models are regularly updated and retrained, recalibrating the model based on new data to achieve better performance on the new data.

[0055] Each step in FIG2 can be implemented as multiple program codes or circuits, and the present invention is not limited thereto. In addition, the method of FIG2 can be used in conjunction with the above embodiments or used alone. In other words, other steps can be added between each step in FIG2.

[0056] The environmental anomaly detection system and method disclosed in this embodiment input the first prediction labels of each machine learning model into an integrated model, and then integrate them through a blending integration method to obtain a second prediction label to indicate whether the air flow is abnormal. This is used to make predictions and achieve effective detection effects, thereby discovering and preventing environmental anomalies as early as possible.

[0057] In this way, when the environmental anomaly detection system is deployed at a manufacturing site or production line, it can achieve the following two main technical effects:

[0058] 1. Abnormal event warning: It can issue an alarm before the temperature / humidity abnormalities occur at the manufacturing site, helping the production line to make relevant preparations and take corresponding measures, thereby reducing the possibility of defective finished products.

[0059] 2. Protect the quality of finished products and materials: Most materials require specific temperature and humidity control conditions to maintain their quality. Failure to meet these standards may result in damage. Therefore, an abnormality detection mechanism allows for timely adjustments to protect the quality of finished products and materials.

[0060] Although the present invention is disclosed above with various embodiments, they are not intended to limit the present invention. Anyone with ordinary knowledge in the technical field should be able to make some changes and modifications without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention should be based on the scope defined by the appended claims.

[0061] [List of reference numerals] 110: air conditioning equipment 111: air flow 121-126: sensor 130: computing module 140: environment 150: air intake area 201-203: step 310: feature data 321-323: machine learning model 331-333, 341: prediction label 340: integrated model 400, 500: chart.

Claims

1. An environmental anomaly detection system, applied to air detection in an environment, the environmental anomaly detection system include: Air conditioning equipment for providing air flow to the environment; at least one sensor for acquiring characteristic data about the air flow; as well as a computing module, which is communicatively connected to the at least one sensor, and is used to input the feature data into each of a plurality of machine learning models to obtain a plurality of first prediction labels, The computing module inputs the plurality of first prediction labels into an integrated model to obtain a second prediction label, wherein the second prediction label is used to indicate whether the air flow is abnormal.

2. The environmental anomaly detection system according to claim 1, wherein the number of the at least one sensor is greater than 1, one of the sensors is disposed in an air intake area of ​​the air conditioning device, and the air flow enters the environment through the air intake area, Another of the sensors is disposed in the environment.

3. The environmental anomaly detection system according to claim 2, wherein the characteristic data includes temperature and humidity.

4. The environmental anomaly detection system according to claim 3, wherein the calculation module is also used to obtain the outdoor temperature and the load rate of the air-conditioning equipment, and the characteristic data also includes the outdoor temperature and the load rate.

5. The environmental anomaly detection system according to claim 1, wherein the machine learning models include a clustering algorithm, a time series prediction algorithm, and a decision tree algorithm.

6. The environmental anomaly detection system according to claim 1, wherein the computing module is used to calculate the classification index of each of the machine learning models during the training phase, The computing module adds constants to the classification indicators to calculate the weight of each of the machine learning models. The integrated model is used to sum the products of the individual weights of these machine learning models and the first prediction labels of these machine learning models to obtain a predicted value, and to determine whether the predicted value is greater than a critical value to generate the second prediction label.

7. The environmental anomaly detection system according to claim 6, wherein the critical value is set to be less than or equal to 0.

6.

8. The environmental anomaly detection system according to claim 7, wherein the critical value is set to be less than or equal to 0.

3.

9. The environmental anomaly detection system according to claim 6, wherein the classification index is calculated according to the following mathematical formula: Where S is the classification index, i, j, j k , k is a positive integer, m represents the number of samples, n represents the number of samples in the category to which the i-th sample belongs, and the i-th sample belongs to the same category as the j-th sample, n k represents the number of samples in the kth category, the i-th sample and the j-th sample k The samples belong to different categories, and distance() is used to calculate the distance between two samples.

10. The environmental anomaly detection system according to claim 9, wherein the calculation module is used to calculate the weights according to the following mathematical formula: Where x is a positive integer, S x is the classification index of the x-th machine learning model, w x is the weight of the x-th machine learning model.

11. The environmental anomaly detection system according to claim 6, wherein each machine learning model has multiple model parameters, and the computing module is used to retrain these machine learning models and the integrated model according to new feature data at default time intervals to update these model parameters, classification indicators and weights.

12. An environmental anomaly detection method, applicable to the environmental anomaly detection system according to any one of claims 1 to 11, wherein the environmental anomaly detection method include: A data acquisition step, obtaining the characteristic data of the air flow; An input step of inputting the feature data into each of the plurality of machine learning models to obtain the plurality of first prediction labels; as well as An integration step is performed to input the plurality of first prediction labels into the integration model to obtain the second prediction label, wherein the second prediction label is used to indicate whether the air flow is abnormal.

Images