Semi-supervised gas recognition method based on DBSCAN and random forest algorithm

Abstract

The application discloses a semi-supervised gas identification method based on DBSCAN and random forest algorithm, and measures organic gas indexes, pm2.5 indexes and environmental temperature data, determines whether the environment belongs to a known stable or unknown complex gas environment according to data complexity and whether the place is marked with a gas category, uses a random forest algorithm to obtain a gas identification result and a known abnormal gas warning if the environment is a known stable gas environment, and otherwise, jointly uses the random forest algorithm and a DBSCAN algorithm to perform gas classification and identification, in a labeled result of the random forest algorithm, if a certain gas category in the label is an abnormal gas, the gas category is divided into an abnormal gas, in an unlabeled result of the DBSCAN algorithm, if certain clustering data deviates from a set threshold, the certain clustering data is marked as an abnormal gas, and the gas classification results of the two are verified through coincidence degree cross verification, the specific category of the labeled result is corresponded to the unlabeled result, abnormal gas detection and gas classification are realized, and the application is suitable for local complex gas environment identification of an industrial system.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of industrial big data and machine learning, and more particularly, relates to a semi-supervised gas recognition method based on DBSCAN and random forest algorithm. BACKGROUND

[0002] With the rapid development of information technology, machine learning has appeared in many fields such as safety inspection and industrial production. Many related technologies have emerged, bringing convenience to human life. With the rapid development of manufacturing industry, industrial systems have become more and more complex. Inappropriate response to changes in the gas environment may cause product loss and even catastrophic accidents. As an effective tool to ensure production efficiency and operational safety, fault diagnosis is a field of particular interest for industrial Internet of Things. In recent years, with the development of sensing technology, monitoring systems have generated a large amount of industrial data, which has made data-driven abnormal gas recognition possible. Electronic nose can be used to detect, identify and analyze various odorous gases. As a typical application of machine olfaction, it has penetrated into various industries and is widely used in safety inspection, industrial production, environmental protection, medical diagnosis and other fields. Literature [4] shows that traditional gas recognition methods are based on features extracted using various signal processing techniques to identify different gas categories. Gas recognition systems use a relatively complex time series signal, and literature [5] considers that this signal appears due to the vibration of sensitive membrane materials caused by the adsorption of gas molecules. The sensitive membrane material, the external environmental factors, including the type and density of the gas, will all affect the generation of this signal. Literature [6] introduces the advantages and accuracy of the data collected by the current sensor array, but the existing sensor array is expensive and has a small coverage range. Literature [7] shows that although the current combination of deep learning has good gas recognition effect, the result has poor interpretability.

[0003] Related Chinese patents, application number 02111043.8 "a kind of olfactory simulation device and olfactory simulation test method", application number 02111963.5 "portable intelligent electronic nose and its preparation method", application number 0127299.3 "electronic nose alarm controller". Because of the variety of gas in industrial environment, the density changes greatly, the above several said invention used to identify the gas in industrial environment has the following problems: (1) feature extraction is simple, only stays in the time domain range to take characteristic value, some only take stable value, so it will waste a lot of useful information. (2) the environmental sensitivity of gas sensor is extremely high, the composition, temperature, humidity, oxygen change in industrial complex environment is violent, so the sensor array collects the data of the same gas day and night change greatly, not to mention the difference between winter and summer, even in different places, the oxygen partial pressure in the air environment is different, the composition is different, which will also lead to data difference, such data is difficult to further analyze, resulting in poor robustness of the recognition result. (3) a large number of labeled data are needed, only a few kinds of simple gas with large difference can be identified, which is not suitable for the identification of complex gas in industrial environment.

[0004] Related US patents, patent number: 6,496,813, patent name: "Classifying apparatus using a combination of statistical methods and neuronal networks, designed in particular for odour recognition (statistical methods and neuronal networks combined classification apparatus designed particularly for odor recognition)", the patent mainly focuses on sensor data analysis method, the statistical method and neural network data processing result used is not very good.

[0005] Electronic nose can be used to detect, identify and analyze a variety of odorous gases, as a sensor array, it has good recognition effect and fast response speed, but the disadvantage is that the price is expensive and the recognition range is relatively small. After collecting the data characteristics of the gas, the method based on DL in document [8] will rapidly decline in performance when identifying dozens or hundreds of gases. Although deep learning (Deep Learning, DL) methods are very effective, they still have serious problems in gas recognition classification. Document [9] shows that in traditional DL methods, different gas types share the same deep network for feature extraction and gas recognition. This design may be effective in single gas recognition, but it will be a problem in complex system gas detection. This is because the increasing diversity of gas environment caused by different gas mixtures will lead to the problem of unbalanced intra-class / inter-class distance in feature space. Or, the feature distance between the same gas is small and difficult to distinguish, while the feature distance between different gases is significantly different and can be easily distinguished. Therefore, in this big data environment, the inter-class distance of the features of some gases may even be smaller than the intra-class distance of some gases, which may cause the learning process to deviate from the global optimal solution. In summary, for the low-cost intelligent recognition and abnormal gas detection of complex gases in industry, the following three points need to be considered: (1) how to make full use of a small amount of labeled and a large amount of unlabeled gas data; (2) how to ensure accuracy while enhancing robustness; (3) how to achieve complex gas recognition in industrial environments due to the intra-class / inter-class distance imbalance and local minimum problems in different features of multiple gas types. The above three points are difficult to overcome in gas recognition. SUMMARY

[0006] The purpose of the present application is to overcome the deficiencies in the prior art and provide a semi-supervised gas recognition method based on DBSCAN and random forest algorithm, which can intelligently perform abnormal gas recognition of industrial systems under the condition of less labeled data.

[0007] The purpose of the present application is achieved by the following technical solutions.

[0008] The semi-supervised gas recognition method based on DBSCAN and random forest algorithm of the present application includes the following processes:

[0009] First step: measure the organic gas index, pm2.5 index and environmental temperature data in the measurement location for a period of time, establish a database, and determine whether the gas environment of the measurement location belongs to a known stable gas environment or an unknown complex gas environment according to the complexity of the data in the database and whether the gas category of the location is marked; if it is a known stable gas environment, only the second step of using the random forest algorithm for processing needs to be performed; if it is an unknown complex gas environment, the subsequent second to fourth steps are executed for processing;

[0010] Second step: using random forest algorithm, identifying the gas species in the database, completing the accurate identification of known gases in the specific scene of the test site, obtaining labeled classification results, and if the task marks a known gas as an abnormal gas, mark the gas as an abnormal gas when identifying it;

[0011] Third step: using DBSCAN algorithm, completing adaptive unsupervised machine learning of unknown gas environment gas, obtaining the best gas species distinction by adjusting the interval parameter, obtaining unlabeled classification results, and if a certain cluster data in the labeled classification result deviates from the set threshold, it is marked as an abnormal gas;

[0012] Fourth step: superimpose the labeled classification results and the unlabeled classification results with their respective abnormal gas data for coincidence verification, data with high coincidence rate is classified into the same class, the specific gas species of the labeled result is assigned to the unlabeled result, and after marking the specific gas species, the unlabeled classification result is used as the standard to complete gas classification and abnormal gas warning.

[0013] If the data is simple and the main gas category is marked in the first step, it is a known stable gas environment, and if the data is complex or the gas category is not marked, it is an unknown complex gas environment.

[0014] Compared with the prior art, the technical scheme of the present application has the beneficial effects that:

[0015] The semi-supervised gas identification method based on DBSCAN and random forest algorithm can intelligently and strategically identify abnormal gases in industrial systems.

[0016] On the one hand, the random forest algorithm and DBSCAN algorithm are used to structurally analyze the PM2.5, temperature, organic gas indicators and other data obtained by traditional sensors, fully utilizing the time domain range of gas data, and maintaining high accuracy in complex environments, and effectively identifying complex gas environment components; on the other hand, due to the use of traditional sensors, the stability is strong in complex environments, the price is low and the coverage is larger, so using this method can save more cost under the same level of identification ability. Therefore, the semi-supervised gas identification method based on DBSCAN and random forest algorithm makes it possible to intelligently solve large-scale industrial gas identification tasks. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 The flowchart of the semi-supervised gas identification method based on DBSCAN and random forest algorithm of the present application;

[0018] Figure 2 The gas detection and identification result based on the random forest algorithm;

[0019] Figure 3 The gas detection and identification result based on the DBSCAN algorithm. DETAILED DESCRIPTION

[0020] The application will be further described below with reference to the drawings.

[0021] The existing deep learning-based method usually directly calculates the most refined diagnosis result by modeling the gas information from the sensor array, ignoring the physical properties of each gas feature, and requiring a large amount of labeled data. For example, the literature

[10] gives a gas monitoring data set of a hospital system, and according to deep learning, it is easier to distinguish the number of gas types in the environment, but the interpretability is poor, and the specific situation of the gas environment cannot be described, which is not conducive to analyzing the causes of abnormal gas.

[0022] Therefore, the application explores and utilizes the multi-level relationship of various gas physical properties, efficiently and accurately extracts features to assist classification, in order to solve the problem of low-cost complex gas intelligent recognition in industry. The features are extracted by collecting and summarizing data using traditional sensors, and after judging the gas environment according to the complexity of the related data and whether the gas category of the place is marked, if it is a simple and stable environment, only the random forest algorithm is used to complete the subsequent gas classification and identification, if it is a complex unknown environment, the random forest algorithm and the DBSCAN algorithm are combined to classify and identify the gas, in the labeled result of the random forest algorithm, if a certain type of gas in the label is an abnormal gas, the gas is classified as an abnormal gas, in the unlabeled result of the DBSCAN algorithm, the gas data deviating from most of the results and not meeting the clustering condition is also marked as an abnormal gas, then the gas classification results of the two are verified by the coincidence degree cross-validation, the specific categories of the labeled results are realized corresponding to the labeled results and the unlabeled results, and finally the abnormal gas detection and gas classification are realized.

[0023] As shown in Figure 1 The semi-supervised gas identification method based on DBSCAN and random forest algorithm of the application includes the following processes:

[0024] First step: measure the related data of organic gas indicators, pm2.5 indicators, environmental temperature and the like in a place to be measured for a period of time, establish a database, analyze the data complexity (mark high complexity data if the data distribution is dispersed or the data fluctuation is large), and determine whether the gas environment of the place to be measured belongs to a known stable gas environment or an unknown complex gas environment according to the complexity of the data in the database and whether the gas category of the place is marked. If the data is simple and the main gas category is marked, it is a known stable gas environment, and only the second step of using the random forest algorithm for processing is needed. If the data is complex or the gas category is not marked, it is an unknown complex gas environment, and the subsequent second step to the fourth step are executed for processing.

[0025] In the measurement of database data, traditional sensors will be used, including organic gas monitoring sensors, PM content monitoring sensors, and environmental temperature monitoring sensors. The organic gas monitoring sensor is small in size, low in power consumption, long in service life, and has high detection sensitivity to low-concentration organic gas, which can be used to detect ammonia, hydrogen sulfide, VOC volatile gas, etc. The PM content monitoring sensor uses the principle of laser scattering to accurately monitor suspended particulate matter, with a minimum resolution particle size of 0.3 μm, small size, strong anti-interference ability, support for real-time response and continuous data collection, and the advantage of zero false alarm rate. The environmental temperature monitoring sensor has a temperature measurement accuracy of 0.5℃, a temperature measurement range of -55℃ to +125℃, and the advantages of waterproof and moisture-proof, small size. In the industrial scene gas intelligent recognition task studied in the present application, the above-mentioned traditional sensors are low in price, wide in range, and can stably and real-timely monitor organic gas, PM content, and environmental temperature in complex environment, and the collected data can provide a solid foundation for the classification result of the algorithm.

[0026] Second step: using the random forest algorithm, identifying the gas species in the database, completing the accurate identification of known gases in the specific scene of the measured location, obtaining a labeled classification result, and if the task marks a known gas as an abnormal gas, the gas is marked as an abnormal gas during identification.

[0027] The present application uses a supervised machine learning method, random forest algorithm, for known gas identification. Literature

[11] proves that the random forest algorithm has good performance in gas classification tasks. Random forest algorithm belongs to the ensemble learning method in machine learning algorithm, the main idea of ensemble learning is to combine multiple weak classifiers to get a strong classifier with better performance in all aspects, and when some weak classifiers predict errors, other weak classifiers can correct them, so it has good anti-noise ability.

[0028] Literature

[12] shows that random forest is an algorithm that integrates multiple trees through the idea of ensemble learning, and its basic unit is decision tree. Each decision tree is a classifier. For an input sample, N decision trees will have N classification results. Random forest integrates all classification voting results, and the class with the most votes is designated as the final output.

[0029] The decision tree is a tree structure, wherein each internal node represents a judgment on an attribute, each branch represents an output of a judgment result, and each leaf node represents a classification result. The generation algorithm of the decision tree includes ID3, C4.5, C5.0, CART, etc. The CART algorithm is selected, the CART is a classification and regression tree, which can be used for classification problems and regression problems, and the CART finally constructs a binary tree, each internal node has two branches, and the Gini coefficient is used to select a suitable feature for splitting at the splitting node.

[0030]

[0031] In the above formula, p k represents the probability of extracting the kth sample category, and the meaning of the Gini coefficient Gini(D) is represented as the probability of randomly extracting two sample categories different from each other from the data set D containing K categories, which reflects the purity of information. The smaller the Gini coefficient, the higher the data purity, and the easier the classification, so each node selects the feature with the smallest Gini coefficient for splitting in the process of constructing the decision tree.

[0032] The random forest algorithm can effectively solve the calculation classification. In the experiment of six gases of alcohol, cigarette, cable burning, incense, oil smoke and wood burning, the gas recognition accuracy of the random forest algorithm is 93.7%, as shown in Figure 2 After identifying the specific gas type, the gas data identification label is retained for subsequent identification and classification, and the task marks a known gas as an abnormal gas, and marks the gas as an abnormal gas when identifying the gas.

[0033] Step 3: Use the DBSCAN algorithm to complete the adaptive unsupervised machine learning of unknown gas environment, obtain the best gas type distinction by adjusting the interval parameter, and obtain the unlabeled classification result. If the clustering data in the labeled classification result deviates from the set threshold, it is marked as an abnormal gas.

[0034] The literature

[13] shows that unknown types of gases may appear in unknown scenes or known scenes. These gases do not collect relevant data through experiments and cannot be identified by supervised machine learning methods. Unsupervised learning methods can draw conclusions from unlabeled training data, so they do not need to collect data of gases to be identified in advance, and they classify according to the characteristics of the data themselves.

[0035] In the present application, gas recognition is performed by clustering. Clustering is a typical unsupervised learning method, which can divide samples of unknown categories and divide samples with high similarity into a category. Common clustering methods include partition clustering methods, density-based clustering methods, hierarchical clustering methods, and model-based clustering methods.

[0036] In gas recognition, clustering methods can be used to classify unknown gases, and then the results of classification are interpreted to associate clustering results with specific gas types. In addition, some clustering methods (such as DBSCAN) can also identify small probability gases in specific scenarios, achieving detection of abnormal gases.

[0037] DBSCAN (Density-Based Spatial Clustering of Applications with Noise) is a representative density-based clustering algorithm.

[0038] As described in document

[14] , unlike partition and hierarchical clustering methods, DBSCAN defines a cluster as the maximum set of density-connected points, which can divide areas with sufficiently high density into clusters and find clusters of arbitrary shape in a noisy spatial database.

[0039] Here, the following definitions are made to explain the DBSCAN algorithm:

[0040] E-neighborhood: the area within a given object radius E is called the E-neighborhood of the object;

[0041] Core object: if the number of sample points in the E-neighborhood of a given object is greater than or equal to MinPts, the object is called a core object;

[0042] Directly density reachable: for a sample set D, if a sample point q is in the E-neighborhood of p and p is a core object, then object q is directly density reachable from object p.

[0043] Density reachable: for a sample set D, given a sequence of sample points p1, p2, …, pn, p = p1, q = pn, if object pi is directly density reachable from pi-1, then object q is density reachable from object p.

[0044] Density connected: there is a point o in the sample set D, if object o is density reachable to objects p and q, then p and q are density connected.

[0045] It can be found that density reachable is the transitive closure of directly density reachable, and this relationship is asymmetric. Density connection is a symmetric relationship. The purpose of DBSCAN is to find the maximum set of density connected objects.

[0046] DBSCAN requires two parameters: scanning radius (eps) and minimum number of contained points (minPts). Starting with an optional unvisited point, find all nearby points within eps (including eps). If the number of nearby points is greater than or equal to minPts, the current point and its nearby points form a cluster, and the starting point is marked as visited. Then recursively process all unmarked points in the cluster in the same way to expand the cluster. If the number of nearby points is less than minPts, the point is temporarily marked as a noise point. If the cluster is sufficiently expanded, i.e. all points in the cluster are marked as visited, then the same algorithm is used to process unvisited points.

[0047] By adjusting the appropriate minPts, the DBSCAN algorithm can effectively identify abnormal dangerous gases in complex gas distribution, and adjusting the appropriate eps can obtain the correct gas species distribution, with strong adaptability, high robustness and high accuracy.

[0048] In an unknown gas environment, the algorithm divides different gas distributions in normal gases, and marks the gas species that deviates too much from other species as an abnormal dangerous gas, and marks it as a different color, achieving abnormal gas early warning. The effect is as shown in Figure 3

[0049] Step 4: The labeled classification result and the unlabeled classification result are overlapped with the respective abnormal gas data for coincidence verification. The data with the highest data coincidence rate in the classification results of the two are classified into the same class. The specific gas species of the labeled result is assigned to the unlabeled result. After marking the specific gas species, the unlabeled classification result is used as the standard to complete gas classification. The classification result can be displayed in the form of a table or a graph, and abnormal gas early warning is completed.

[0050] In summary, after determining the type of the gas environment, the present application extracts features from traditional sensor data and uses them as the basis for complex gas identification. Then, according to the complexity of the gas environment, in a known stable gas environment, only the random forest algorithm is used for identification to obtain the gas identification result and known abnormal gas early warning. In an unknown complex gas environment, the random forest algorithm and the DBSCAN algorithm are combined for gas classification and identification. In the labeled result of the random forest algorithm, if a certain type of gas in the label is an abnormal gas, then this type of gas is classified as an abnormal gas. In the unlabeled result of the DBSCAN algorithm, the gas data that deviates from most results and does not meet the clustering conditions is also marked as an abnormal gas. Subsequently, the gas classification results of the two are cross-verified by coincidence, the specific category of the labeled result is matched with the unlabeled result, and finally abnormal gas detection and gas classification are achieved.

[0051] ​Although the functions and working processes of the present application are described above in combination with the drawings, the present application is not limited to the specific functions and working processes described above, and the specific embodiments described above are merely illustrative rather than limiting, and those of ordinary skill in the art can make many forms under the inspiration of the present application without departing from the purpose of the present application and the scope protected by the claims, which all belong to the protection of the present application.

[0052] References

[0053] [1] Zee, Frank, and Jack W. Judy. "Micromachined polymer-based chemical gas sensor array." Sensors and Actuators B: Chemical 72.2 (2001): 120-128.

[0054] [2] Nylander, Claes, Bo Liedberg, and Tommy Lind. "Gas detection by means of surface plasmon resonance." Sensors and Actuators 3 (1982): 79-88.

[0055] [3] Lu, Ganhua, Leonidas E. Ocola, and Junhong Chen. "Gas detection using low-temperature reduced graphene oxide sheets." Applied Physics Letters 94.8 (2009): 083111.

[0056] [4] Khalaf, Walaa, Calogero Pace, and Manlio Gaudioso. "Gas detection via machine learning." Int. J. Comput. Electron. Autom. Control Inf. Eng 2.1 (2008): 61-65.

[0057] [5] Tanaka, Y., T. Nakamoto, and T. Moriizumi. "Study of highly sensitive smell sensing system using gas detector tube combined with optical sensor." Sensors and Actuators B: Chemical 119.1 (2006): 84-88.

[0058] [6] Matz, G., T. Hunte, and W. Schroeder. "Hand-portable gas-detector array (GDA) for rapid field detection and identification of chemical threat." Field Analytical Chemistry & Technology 4.4 (2000): 195-203.

[0059] [7] Peng, Pai, et al. "Gas classification using deep convolutional neural networks." Sensors 18.1 (2018): 157.

[0060] [8] Liu, Qihe, et al. "Gas recognition under sensor drift by using deep learning." International Journal of Intelligent Systems 30.8 (2015): 907-922.

[0061] [9] Wang, Syuan-He, et al. "Using a hybrid deep neural network for gas classification." IEEE Sensors Journal 21.5 (2020): 6401-6407.

[0062]

[10] Kumar, Jambi Ratna Raja, Rahul K. Pandey, and Biplab K. Sarkar. "Pollutant gases detection using the machine learning on benchmark research datasets." Procedia Computer Science 152 (2019): 360-366.

[0063]

[11] Wei, Guangfen, et al. "An effective gas sensor array optimization method based on random forest." 2018 IEEE SENSORS. IEEE, 2018.

[0064]

[12] Biau, Gérard, and Erwan Scornet. "A random forest guided tour." Test 25.2 (2016): 197-227.

[0065]

[13] Zhang, Yulu, et al. "Proposal of unsupervised gas classification by multimode microresonator." IEEE Photonics Journal 13.2 (2021): 1-11.

[0066]

[14] Khan, Kamran, et al. "DBSCAN: Past, present and future." The fifth international conference on the applications of digital information and web technologies (ICADIWT 2014). IEEE, 2014.

Claims

1. A semi-supervised gas identification method based on DBSCAN and random forest algorithms, characterized in that, Includes the following processes: Step 1: Measure organic gas, PM2.5, and ambient temperature data at the test location over a period of time, establish a database, and determine whether the gas environment at the test location is a known stable gas environment or an unknown complex gas environment based on the complexity of the data in the database and whether the gas category has been labeled at the location. If it is a known stable gas environment, only Step 2 using the random forest algorithm is required; if it is an unknown complex gas environment, proceed with Steps 2 through 4. Step 2: Use the random forest algorithm to identify the types of gases already in the database, complete the accurate identification of known gases in the gas environment of the test location, and obtain labeled classification results. If the task marks a known gas as an anomalous gas, then the gas will be marked as an anomalous gas when it is identified. Step 3: Use the DBSCAN algorithm to perform adaptive unsupervised machine learning on the unknown gas environment. By adjusting the interval parameter, the best gas species distinction is obtained, and the unlabeled classification result is obtained. If a cluster data in the labeled classification result deviates from the set threshold, it is marked as an abnormal gas. Step 4: Verify the overlap between the labeled and unlabeled classification results and their respective abnormal gas data. Data with high overlap rates are grouped into the same category. Assign the specific gas types to the unlabeled results based on the labeled results. After marking the specific gas types, use the unlabeled classification results as the standard to complete the gas classification and complete the abnormal gas warning.

2. The semi-supervised gas identification method based on DBSCAN and random forest algorithms according to claim 1, characterized in that, If the data in the first step is simple and the main gas categories have been labeled, it is a known stable gas environment; if the data is complex or the gas categories have not been labeled, it is an unknown complex gas environment.

Images