A method for odor sensing and monitoring, and the system thereof

Abstract

This invention relates to a method and system of odor sensing and monitoring that uses a machine learning technique to predict an odor type and intensity from an odor measured by at least one gas sensor. The invention utilizes at least one panelist to evaluate and label the odor intensity of some odor samples and measured gas sensor data of the same odor that is evaluated by the panelist to train one or more odor analysis models for machine learning. The trained models are used for predicting odor type and intensity.

Description

[0001] A METHOD FOR ODOR SENSING AND MONITORING, AND THE SYSTEM THEREOF

[0002] Technical Field

[0003] The present invention relates to a method and system for odor sensing and monitoring by using gas sensors. The method and system are particularly well suited for measuring odor in an area of interest, for example, chemical production line, waste recycling facility, water treatment facility, etc.

[0004] Background Art

[0005] Odor sensing and monitoring has been increasing its importance as industrial and residential areas are getting closer as world population is increasing. In most countries, living areas expand toward industrial areas. It creates difficulties for the industries to control their odor emissions without affecting surrounding areas. Odor type and intensity are important parameters that should be monitored. Knowing only one of them is not sufficient; in particular, when a warning should be advised to people in residential areas. Therefore, it is necessary to have a system that can accurately determine type and intensity of odor emission from an industrial area. Further, the system should provide the types of detected odor and their intensities that at least comply with conventional olfactometer techniques that certified panelists are needed to evaluate them. Certified panelists are experts who perform odor sensory evaluations. The person who wants to be a certified panelist must be trained and qualified. He or she must carefully perform evaluation steps on each odor sensory to maintain qualified evaluation results. Different countries have different acceptable standards to follow. Standard methods for scenting odor samples based on American Society for Testing and Materials (ASTM) and Japanese Industrial Standard (JIS) are preferred. Portable olfactometer technique following EN 13725:2022 standard is another industrial acceptable standard. The goal of the present invention is, therefore, to provide a system that can detect odor types and intensities to the same level that of the panelists.

[0006] In W02022005407A1 publication of the same inventors, an odor detecting and monitoring system was disclosed. The system comprises an odor sensor collecting odor samples in the air and analyzing the odor compositions from said samples, a weather station, and a processing system determining the type and dispersion pattern of said odor. The system displays the odor dispersion information on a geographical image such that the users can visualize the odor dispersion pattern. In this disclosure, only one odor sensor is used to collect and analyze odor compositions. If the sensor does not respond to some odor compositions, it would not provide an accurate analysis. Moreover, the system is intended to analyze odor concentration and use it for the pattern prediction but is not intended to analyze the odor intensity. It is, thus, not possible to accompany any standardized olfactometer technique into its results.

[0007] In U.S. Pat. No. US9028751B, a system and a method for controlling an odor emitted from an odor source or a factory is disclosed. The invention uses odor sensors to detect odor emission. The sensor data are processed together with weather information to analyze the odor dispersion. If the odor intensity is found to be higher than a specified level, the system decreases its odor emission to reduce the odor intensity until it is less than the specified level. The system is intended to be used to detect known odors. It cannot respond to odors that the sensors do not support.

[0008] In U.S. Pat. No. US11112383B2 and its corresponding application, an odor detection system and method that utilize multiple sensors or an array sensor to detect an odor. At least two sensors (or sensor elements of the array) interact with odor compositions of the measured odor sample. The concentration of the odor can be measured by changing the detection sensitivity in each sensor. An information processing unit is then used to provide pattern identification based on sensor properties, such as types and the arrangement of the sensors. Even though the invention uses an array sensor response of different types of sensors to identify the detected odor type, it cannot measure or predict the intensity of the detected odor as needed.

[0009] Therefore, it is still necessary to provide a system and / or method that can detect or predict both odor types and intensities accurately. In particular, the prediction results should at least comply with conventional olfactometer techniques, without having it tested by panelists. This allows the system to be used in various factories and industries.

[0010] Brief Summary of the Invention

[0011] The present invention relates to a method for odor sensing and monitoring. The method comprises the steps of (a) creating an odor database that is used to store a set of sample odor data for training an odor analysis model, wherein the data in the set include at least an odor type, an odor intensity, a temperature, a humidity, and measured gas sensor data of at least one gas sensor from an odor data measurement; (b) creating a plurality of odor analysis models for processing measured gas sensor data together with temperature and humidity data to predict type and intensity of an odor measured by at least one gas sensor; and (c) analyzing measured gas sensor data to predict the odor type and odor intensity from an odor measured by at least one gas sensor, wherein at least one odor analysis model is used to process the measured gas sensor data together with temperature and humidity data by using a machine learning technique. The method is characterized in that the step of (a) creating an odor database, comprising at least the steps of (a-

[0012] 1) evaluating and labelling the odor type of the odor samples by at least one panelist; (a-2) evaluating and labelling the odor intensity of the odor samples by the panelist; (a-3) storing measured gas sensor data of at least one sensor that is used to measure the same odor that is evaluated by the panelist; (a-4) storing temperature and humidity data of the surrounding atmosphere when the sensor measuring the odor; (a-5) providing the set of sample odor data, wherein each set at least includes the odor type labelled by the panelist, the odor intensity labelled by the panelist, the measured gas sensor data, the temperature data, and the humidity data; and (a-6) storing the set of sample odor data in the odor database.

[0013] In another embodiment, the step of (b) creating a plurality of odor analysis models may be comprising at least the steps of (b-1) pre-processing the set of sample odor data obtained in the step (a); (b-2) transforming said pre-processed data set; (b-3) training at least one odor analysis model by using the transformed data set; (b-4) evaluating the performance of the trained odor analysis model; and (b-5) storing the trained odor analysis model and its parameters.

[0014] In another embodiment, the step of (c) analyzing measured gas sensor data may be comprising at least the steps of (c-1) storing measured gas sensor data of at least one sensor; (c-

[0015] 2) storing temperature and humidity data of the surrounding atmosphere when the sensor measuring the odor; (c-3) processing the measured gas sensor data together with the temperature and humidity data by using at least one trained odor analysis model; and (c-4) predicting the odor type and the odor intensity of the odor measured by the sensor.

[0016] In a preferred embodiment, the step of (a-2) evaluating and labelling the odor intensity of the odor samples by the panelist can be selected from (1) collecting the odor samples into an odor bag and then analyzing the odor by scenting the odor samples based on the American Society for Testing and Materials (ASTM) or the Japanese Industrial Standard (JIS) method, or (2) scenting the odor in the environment using Portable Olfactometer technique following EN 13725:2022 standard.

[0017] In another preferred embodiment, the odor analysis model may comprise an odor-type prediction model and an odor-intensity prediction model. In some embodiments, the step of (b) creating a plurality of odor analysis models, further comprises the step of monitoring the odor analysis model deviation, which allows the model to be evaluated or tuning its parameters.

[0018] In another embodiment, the method for odor sensing and monitoring may further comprise the step of (d) displaying odor monitoring results to display at least the measured gas sensor data of at least one sensor, the predicted odor type, the predicted odor intensity, the temperature, and the humidity. Additional data to be displayed can be further selected from an odor type prediction model, an odor intensity prediction model, a radar chart that represents an odor composition, a gas sensor status, a gas sensor location, local weather data, an odor diffusion pattern, and odor monitoring data history.

[0019] The present invention also relates to a system for odor sensing and monitoring. The system comprises an odor database, at least one gas sensor, a temperature sensor, a humidity sensor, a processing unit, and a user interface unit, wherein the odor database stores data that are used for training an odor analysis model and for analyzing the odor; the gas sensor measures odor samples diffused in the air in an odor monitoring area and transmits the measured gas sensor data to the processing unit via a communication mean; the temperature sensor collects temperature data of the surrounding atmosphere when the sensor measuring the odor; the humidity sensor collects humidity data of the surrounding atmosphere when the sensor measuring the odor; the processing unit performs data processing to create an odor analysis model and to analyze the odor; and the user interface unit displays data to a user or receives input from a user. The system is characterized in that the odor database is divided into at least two data sets, comprising: an odor analysis model training data set and an odor analysis model data set, wherein the odor analysis model training data set comprises an odor type and intensity labelled by at least one panelist, measured gas sensor data of at least one sensor that is used to measure the same odor samples labelled by the panelist, temperature and humidity data during the measurement, and the odor analysis model data set comprises at least one odor analysis model and its parameters. The processing unit uses at least one odor analysis model to process the measured gas sensor data together with temperature and humidity data by using a machine learning technique.

[0020] In some embodiments, the processing unit of the system may further comprise a filtering unit for filtering the measured gas sensor data and / or a normalization unit that normalizes the measured gas sensor data by using baseline data obtained from clean air measurement. In a preferred embodiment, the user interface unit comprises at least a part for displaying measured gas sensor data of at least one sensor, a part for displaying a predicted odor type data, a part for displaying a predicted odor intensity data, a part of displaying the temperature sensor data, and a part of displaying the humidity sensor data.

[0021] In another preferred embodiment, the user interface unit may further comprise a data displaying part to display the data that can be selected from an odor type prediction model, an odor intensity prediction model, a radar chart that represents odor composition, a gas sensor status, a gas sensor location, local weather data, a predicted odor diffusion pattern, and odor monitoring data history.

[0022] In some embodiments, the user interface unit further comprises a part for displaying the odor analysis model deviation monitoring, which allows the model to be evaluated or tuning its parameters.

[0023] The object of the present invention is to provide a method of odor sensing and monitoring that uses a machine learning technique to predict an odor type and intensity from an odor measured by at least one gas sensor. The method utilizes at least one panelist to evaluate and label the odor intensity of some odor samples and measured gas sensor data of the same odor that is evaluated by the panelist to train one or more odor analysis models for machine learning. The trained models are then used to analyze measured gas sensor data from at least one sensor to predict the odor type and odor intensity. A preferred step of the method includes evaluating and labelling the odor intensity of the odor samples by the panelist based on (1) collecting the odor samples into an odor bag and then analyzing the odor by scenting the odor samples based on the American Society for Testing and Materials (ASTM) or the Japanese Industrial Standard (JIS) method, or (2) scenting the odor in the environment using Portable Olfactometer technique following EN 13725:2022 standard.

[0024] Another object of the present invention is to provide a system for odor sensing and monitoring based on the method above. The system comprises an odor database, at least one gas sensor, a temperature sensor, a humidity sensor, a processing unit, and a user interface unit. The odor database stores data used for training an odor analysis model and for analyzing the odor. The gas sensor measures odor samples diffused in the air in an odor monitoring area and transmits the measured gas sensor data to the processing unit via a communication mean while the temperature and humidity sensors collect temperature and humidity data of the surrounding atmosphere respectively. The processing unit performs data processing to create an odor analysis model and to analyze the odor, wherein the system is characterized in that the odor database is divided into at least two data sets, comprising an odor analysis model training data set and an odor analysis model data set. The odor analysis model training data set comprises an odor type and intensity labelled by at least one panelist, measured gas sensor data of at least one sensor that is used to measure the same odor samples labelled by the panelist, temperature and humidity data during the measurement. And the odor analysis model data set comprises at least one odor analysis model and its parameters. The processing unit uses at least one odor analysis model to process the measured gas sensor data together with temperature and humidity data by using a machine learning technique to predict the odor type and odor intensity.

[0025] Brief Description of the Drawings

[0026] Fig- 1 shows the method for odor sensing and monitoring of this invention.

[0027] Fig- 2 shows the steps for creating an odor database.

[0028] Fig- 3 shows the steps for creating a plurality of odor analysis models.

[0029] Fig. 4 shows the steps for analyzing measured gas sensor data.

[0030] Fig. 5 shows a radar chart of odor characteristics measured by five different types of gas sensors.

[0031] Fig. 6 shows a schematic block diagram of the system for odor sensing and monitoring of this invention.

[0032] Fig. 7 shows an example of the user interface unit.

[0033] Fig. 8A shows a radar chart representing an odor characteristic.

[0034] Fig. 8B shows another radar chart representing an odor characteristic.

[0035] Fig. 8C shows another radar chart representing an odor characteristic.

[0036] Fig. 9 shows an example of prediction performance of a machine learning model.

[0037] Detailed Description of the Invention

[0038] FIG.1-2 show a method for odor sensing and monitoring of this invention. The method comprises the steps of creating an odor database 10, creating a plurality of odor analysis models 20, and analyzing measured gas sensor data 30 to predict odor type and odor intensity of an unknow odor. To analyze measured gas sensor data, an appropriate odor analysis model needs to be created. Machine learning algorithms can be trained and tested to evaluate their performance. High accuracy algorithms may be selected as the odor analysis models. It is noted that more than one model can be used in the prediction process. Predicted results based on multiple models poses higher confidential level when compared to the result from a single model.

[0039] To create the model, the odor database is used to store a set of sample odor data for training an odor analysis model. The data in the set include at least an odor type, an odor intensity, a temperature, a humidity, and measured gas sensor data of at least one gas sensor from an odor data measurement.

[0040] The accuracy of the sample odor data sets is crucial to the present invention as they are the data used to train odor analysis models. If the data sets are not accurate, the trained models will provide inaccurate predictions. On the other hand, if accurate data sets are used for training, the models will provide accurate predictions.

[0041] The present invention, preferably, uses certified panelists to create the set of sample odor data. Certified panelists are experts who perform odor sensory evaluations. Different countries may have different requirements for a person to be a certified panelist. In the present invention, certified panelists are required to evaluate odor samples. The evaluation should at least include determining the type and intensity of the odor samples. At the same time, the odor samples are also collected or measured by gas sensors to create measurement data. By this way, the measured data can be matched with the panelists’ evaluations. This matching will be further used for training an odor analysis model.

[0042] To create an odor database, the following steps are normally performed: (i) evaluating and labelling the odor type of the odor samples by at least one panelist 11, (ii) evaluating and labelling the odor intensity of the same odor samples by the panelist 12, (iii) storing measured gas sensor data of at least one sensor that is used to measure the same odor evaluated by the panelist 13, (iv) storing temperature and humidity data of the surrounding atmosphere when the sensor measuring the odor 14, (v) providing the set of sample odor data 15, and (vi) storing the set of sample odor data in the odor database 16. The set of sample odor data should include at least the odor type and odor intensity labelled by the panelist, the measured gas sensor data, and the temperature and humidity data.

[0043] The sample odor data set is analyzed by using a machine learning technique to create an analysis model that will be used to predict the odor type and odor intensity of an odor measured by at least one gas sensor. At least one odor analysis model is used to provide a prediction to the measured gas sensor data together with temperature and humidity data. Multiple models can be used to increase the prediction performance. Many possible odor analysis models can be applied to the present invention. To get an accurate model, the set of sample odor data for model training should be pre-processed properly so that the resulting model parameters match the data provided. It is thus preferable to perform the following steps when creating an odor analysis model as shown in FIG.3. The steps are (i) pre-processing the set of sample odor data 21, (ii) transforming said pre-processed data set 22, (iii) training the odor analysis model 23 by using the transformed data set, (iv) evaluating the performance of the trained odor analysis model 24, and (v) storing the trained odor analysis model and its parameters 25. Various models can be pre-processed in the same manner to have the best model for further analysis. The steps can be performed for many models at the same time, or they may be processed consecutively in a loop manner.

[0044] Examples of the pre-processing step include data cleaning, anomaly correction, outlier removal, sensor data filtering, scaling, etc. The transforming step can be data aggregation, normalization, generalization, etc.

[0045] After the models are trained to satisfy a designed prediction performance, they are ready to be used to analyze measured gas sensor data from actual measurement to predict the unknown odor type and odor intensity in a monitoring area. The analysis step uses trained models to predict the odor type and intensity without any evaluation information from the panelists. This is where the present invention is put to work in a real environment. Since the models have been trained based on evaluated information by the panelists, they are expected to provide the predictions of the odor type and intensity as accurate as the panelists do.

[0046] For the analyzing step, it is preferable to perform the steps of (i) storing measured gas sensor data of at least one sensor 31, (ii) storing temperature and humidity data of the surrounding atmosphere 32 when the sensor measuring the odor, (iii) processing the measured gas sensor data together with the temperature and humidity data by using at least one trained odor analysis model 33, and (iv) predicting the odor type and the odor intensity 34 of the odor measured by the sensor as shown in FIG.4.

[0047] Prediction performance of the present invention depends on the accuracy of the sample odor data sets stored in the database. Since the data sets are used for training the odor analysis models, the prediction accuracy of the models follows the accuracy of the prepared data sets. If the data sets are labelled correctly, the model accuracy will be high. On the other hand, if the labelled data sets have errors, the model accuracy will be low. It is well understood amongst the machine learning community that the accuracy of training data sets should be as high as possible to obtain the accurate model. Since the present invention uses panelists to evaluate and label the odor type and intensity for the training data sets, it ensures that the accuracy of the data sets is as high as standard olfactometer methods, which in turn makes the odor analysis models accurate and standardized.

[0048] Gas sensors, preferably, should be metal oxide gas sensors that are sensitive to major characteristics of an odor of interest. Preferably, the sensor should be either hydrocarbons (CxHy), hydrogen gas (H2), carbon monoxide (CO), alcohols (CxHyOH), nitrogen dioxide (NO2), nitrogen monoxide (NO), ozone (O3), ammonia (NH3), sulfur dioxide (SO2), refrigerants R134a u.a., or volatile organic compounds, or the combination thereof. More than one sensor can be used to measure odor samples to differentiate the variety of odors. Normally, the amplitude of the sensor signal or data represents the intensity of the detected odor.

[0049] FIG.5 shows a radar chart of odor characteristics measured by five different types of gas sensors. Each comer of the chart indicates the amplitude level of the sensing gas. Preferably, the measured gas sensor data are obtained from an array sensor having eight or more sensors to create a differentiable pattern on the radar chart. It is clear that by using different types of sensors an odor characteristic can be determined. The odor characteristic and the odor type and intensity labelled by the panelists are grouped together as the training data set that will be used in the odor analysis model training process.

[0050] In a preferred embodiment, the panelists evaluate and label the odor type and intensity for the training data sets by using either (1) collecting the odor samples into an odor bag and then analyzing the odor by scenting the odor samples based on the American Society for Testing and Materials (ASTM) or the Japanese Industrial Standard (JIS) method, or (2) scenting the odor in the environment using Portable Olfactometer technique following EN 13725:2022 standard. These methods or standards are acceptable worldwide for odor evaluations. Thus, the present invention is aimed to train machine learning models to achieve the same accuracy level as that of the standards. This is particularly useful to reduce erroneous predictions based on human errors or to increase the usability of the invention for odor monitoring when the panelists are not available. The training data sets can be created once and reused to train the odor analysis models as many times as possible. The training data sets may be updated from time to time to increase the accuracy of the prediction or to increase the number of detectable odor types.

[0051] In some embodiments, one odor analysis model may be selected and trained to predict both the odor type and intensity as they can be grouped together as two dependent variables. In other embodiments, preferably the odor analysis model comprises an odor-type prediction model and an odor-intensity prediction model. In such the case, one prediction model is used to predict one variable. By using two models separately, the prediction accuracy of each variable, i.e. odor type or odor intensity, is enhanced. The prediction models may be the same or they may be totally different.

[0052] Many machine learning models can be used for odor-type prediction. For example, they are Decision Trees, Random Forest, Naive Bayes, Support Vector Machine, K-Nearest Neighbor (KNN), Gradient Boosting Machine, Light Gradient Boosting, Extreme Gradient Boosting, Extremely Randomized Tree Model, Distributed Random Forest, Generalized Linear Model, Stacked Ensembles, Deep Learning, Multilayer Perceptron, Recurrent Neural Networks, Long Short-Term Memory Networks, Generative Adversarial Networks, and Autoencoders Deep Learning Algorithm. Preferably, the odor-type prediction model is Extremely Randomized Tree Model.

[0053] Similarly, other many machine learning models can be used for odor-intensity prediction. They can be Linear Regression, Logistic Regression, Polynomial Regression, Lasso Regression, Bayesian Linear Regression, Principal Components Regression, Partial Least Squares Regression, and Elastic Net Regression. Preferably, the odor-intensity prediction model is Polynomial Regression.

[0054] In some embodiments, to facilitate the odor analysis model creations it is preferable to include a monitoring step to monitor the odor analysis model deviations. By observing the model deviations, such as precision, accuracy, recall, etc., it allows us to know when the prediction performance of the models reduces, and the models need to be re-trained to enhance their accuracies.

[0055] In some embodiments, the present invention may include the step of setting warnings when the odor intensity level exceeds a threshold. To provide an all-the-time odor monitoring, the present invention should be able to alert users when a certain odor has its intensity greater than a predefined threshold. For example, odor intensity at 300-1,000 OU / m3is a threshold for stack emission while odor intensity at 15-30 OU / m3is restricted for ambient condition.

[0056] When analyzing the measured gas sensor data, the sensor data may be filtered to reduce noise in the data. Different filtering techniques may be applied depending on the sensor type and its accuracy; for example, low-pass, band-pass, or high-pass filtering. Preferably, the filtering is performed by using an exponential filter, which smooths noisy measurements with less memory usage. Appropriate filter parameters may be stored in the odor database and applied to the sensor data when needed. The parameters may also be adjustable by users to achieve the best prediction performance. Similarly, it is understandable by a person skilled in the art that other processing or signal processing techniques may be applied to the sensor data before and / or after the filtering process such as up-sampling / down-sampling or transformation.

[0057] In a preferred embodiment, the measured gas sensor data may be normalized by using baseline data obtained from clean air measurement. Clean air has no odor, and it is pure and does not contain any odor particles. Using clean air data as references is equivalent to calibrating the sensor to have zero offset with respect to clean air. It, therefore, improves the accuracy of the prediction performance of the present invention.

[0058] In some embodiments, the method for odor sensing and monitoring according to the present invention may include the step of displaying odor monitoring results 40 to display at least the measured gas sensor data of at least one sensor, the predicted odor type, the predicted odor intensity, the temperature, and the humidity of the monitored area.

[0059] In another embodiment, additional information may be included when displaying odor monitoring results as well. This information should be selected such that it provides users with an overall scenario of odor monitoring. Preferably, the information may be selected from an odor type prediction model, an odor intensity prediction model, a radar chart that represents an odor composition, a gas sensor status, a gas sensor location, local weather data, an odor diffusion pattern, and odor monitoring data history.

[0060] To accompany the method described above, an odor sensing and monitoring system 100 will be explained next. FIG.6 shows a schematic block diagram of the odor sensing and monitoring system 100 according to the present invention. The system comprises an odor database 110, at least one gas sensor 120, a temperature sensor 130, a humidity sensor 140, a processing unit 150, and a user interface unit 160. The odor database 110 stores data used for training an odor analysis model and for analyzing the measured odor in an odor monitoring area. The odor database 110 is divided into at least two data sets, comprising an odor analysis model training data set and an odor analysis model data set. The odor analysis model training data set comprises an odor type and intensity labelled by at least one panelist, measured gas sensor data of at least one sensor that is used to measure the same odor samples labelled by the panelist, temperature and humidity data during the measurement. The odor analysis model data set comprises at least one odor analysis model and its parameters. The gas sensor 120 measures odor samples diffused in the air in an odor monitoring area and transmits the measured gas sensor data to the processing unit 150 via a communication mean. The communication mean or media can be wire or wireless and it can be either local or network connection. The temperature sensor 130 collects temperature data of the surrounding atmosphere when the gas sensor 120 measures the odor samples. The humidity sensor 140 collects humidity data of the surrounding atmosphere when the gas sensor 120 measures the odor samples. The processing unit 150 performs data processing to create an odor analysis model and to analyze the odor by processing at least one odor analysis model to process the measured gas sensor data together with temperature and humidity data by using a machine learning technique. The user interface unit 160 displays data to a user or receives input from users. The user interface unit 160 may display measured data and processed data including graphical displays of the sensor data.

[0061] In a preferred embodiment, the processing unit 150 performs data processing to create an odor analysis model by having at least the steps of (i) pre-processing the odor analysis model training data set, (ii) transforming said pre-processed data set, (iii) training at least one odor analysis model, (iv) evaluating the performance of the trained odor analysis model, and (v) storing the trained odor analysis model and its parameters.

[0062] To analyze a measured odor, the processing unit performs data according to at least the steps of (i) storing the measured gas sensor data of at least one sensor 120, (ii) storing the temperature and humidity data, (iii) processing the measured gas sensor data of at least one sensor 120 together with temperature and humidity data and at least one odor analysis model from the odor analysis model data set, and (iv) predicting the odor type and intensity of the measured data.

[0063] The odor type and intensity stored in the odor database are preferably labelled by a panelist to ensure the accuracy of the odor predictions by the odor analysis models. The panelist may choose the method for labelling the odor from either (1) collecting the odor samples into an odor bag and then analyzing the odor by scenting the odor samples based on the American Society for Testing and Materials (ASTM) or the Japanese Industrial Standard (JIS) method, or (2) scenting the odor in the environment using Portable Olfactometer technique following EN 13725:2022 standard. These two methods of odor determinations are well-known and widely accepted as industrial standards.

[0064] In some embodiments, the odor analysis model may be separated into two different models, i.e. an odor-type prediction model and an odor-intensity prediction model. Using two models for dedicated prediction purposes is recommended if high prediction accuracy is needed. However, in some embodiments, there may be the case that only one prediction model is applied to predict both the odor type and intensity.

[0065] Many machine learning models can be applied for odor-type prediction. The odor-type prediction model can be selected from Decision Trees, Random Forest, Naive Bayes, Support Vector Machine, K-Nearest Neighbor (KNN), Gradient Boosting Machine, Light Gradient Boosting, Extreme Gradient Boosting, Extremely Randomized Tree Model, Distributed Random Forest, Generalized Linear Model, Stacked Ensembles, Deep Learning, Multilayer Perceptron, Recurrent Neural Networks, Long Short-Term Memory Networks, Generative Adversarial Networks, and Autoencoders Deep Learning Algorithm. Preferably, the odor-type prediction model is Extremely Randomized Tree Model.

[0066] Similarly, different machine learning models can be used for odor-intensity prediction. The odor-intensity prediction model can be selected from Linear Regression, Logistic Regression, Polynomial Regression, Lasso Regression, Bayesian Linear Regression, Principal Components Regression, Partial Least Squares Regression, and Elastic Net Regression. Preferably, the odorintensity prediction model is Polynomial Regression.

[0067] In some embodiments, the processing unit 150 may further comprise a filtering unit for filtering the measured gas sensor data. The filtering unit is designed to reduce noise in the sensor data, which in turn increases the prediction performance of the system. Different filtering units may be selected depending on the characteristic of the measured sensor data; for example, low- pass, band-pass, or high-pass filtering. Preferably, the filtering unit is an exponential filter as it requires less memory to store the data. It is understandable by a person skilled in the art that other filtering or signal processing units may be applied to the sensor data before and / or after the filtering unit such as up-sampling / down-sampling or transformation.

[0068] In some embodiments, the processing unit 150 may further comprise a normalization unit that normalizes the measured gas sensor data by using baseline data obtained from clean air measurement. Using clean air data as reference data improves the accuracy of the prediction performance of the system.

[0069] The gas sensors 120, preferably, should be metal oxide gas sensors whose measured gas sensor data are obtained. They may be sensor that capable of measuring hydrocarbons (CxHy), hydrogen gas (EE), carbon monoxide (CO), alcohols (CxHyOH), nitrogen dioxide (NO2), nitrogen monoxide (NO), ozone (O3), ammonia (NH3), sulfur dioxide (SO2), refrigerants R134a u.a., or volatile organic compounds, or the combination thereof. More than one sensor is preferred to measure odor samples. Preferably, the gas sensor is an array sensor having eight or more sensors that is capable of measuring the gas mentioned above.

[0070] In a prefer embodiment, the user interface unit 160 comprises at least a part for displaying measured gas sensor data of at least one sensor, a part for displaying a predicted odor type data, a part for displaying a predicted odor intensity data, a part of displaying the temperature sensor data, and a part of displaying the humidity sensor data. FIG. 7 shows an example of the user interface unit 160 according to this invention.

[0071] Additional information may be selected to display with the user interface unit 160 to represent useful information to users. The user interface unit 160 may comprise a data displaying part to display an odor type prediction model, an odor intensity prediction model, a radar chart that represents odor composition, a gas sensor status, a gas sensor location, local weather data, a predicted odor diffusion pattern, and / or odor monitoring data history.

[0072] In some embodiments, the user interface unit 160 may comprise a part for displaying the odor analysis model deviation monitoring to allow users to monitor and evaluate the analysis model. And the user interface unit 160 may further comprise a part for setting and displaying warnings when an odor intensity level exceeds a threshold.

[0073] EXAMPLE

[0074] The present invention has been tested to verify its accuracy by the inventors. Three array sensors, each having 10-sensor elements, were deployed in a paper production factory to monitor odor in different areas of the factory. They were (a) paper making machine, (b) biofilter, and (c) wastewater treatment areas. Each area contained different odors, each having its own characteristic, that are differentiable as shown in the radar charts in FIG. 8A-8C. The intensities of the measured odors were predicted. The intensity of the odor at the wastewater treatment area was between 1,000-2,500 OU / m3. The odor intensity at the biofilter area was between 2,000- 7,000 OU / m3. And the odor intensity at the paper making machine was 500-2,000 OU / m3. Digital data from sensors were separated into two groups. The first group was the data from the sensors and the second group was the environmental data including air temperature, humidity, sensor location, and weather information. Several machine learning models were used to validate the present invention. The processing unit performs data analysis, data cleaning, feature selection, and feature engineering for different machine learning models. The processing unit performed outlier detection to remove data abnormalities from the group. Sensor data were filtered to reduce noise and were normalized with baseline data, which was clean air measurement of that location. Then the sensor data were grouped by using hierarchical clustering techniques to remove redundancy and the important feature group of the data was selected. Machine learning models were trained and evaluated their prediction accuracies. The best model was found to be Polynomial Regression model, whose results are presented in FIG.9.

[0075] Similar validations have been performed by the inventors to verify the performance of the present invention in different factory scenarios such as a sea food factory, a fresh produce packaging factory, etc. Their results, however, are not described here to save space of this disclosure.

Claims

CLAIMS1. A method for odor sensing and monitoring, comprising the steps of:(a) creating an odor database (10) that is used to store a set of sample odor data for training an odor analysis model, wherein the data in the set include at least an odor type, an odor intensity, a temperature, a humidity, and measured gas sensor data of at least one gas sensor;(b) creating a plurality of odor analysis models (20) for processing measured gas sensor data together with temperature and humidity data to predict type and intensity of an odor measured by at least one gas sensor;(c) analyzing measured gas sensor data (30) to predict the odor type and odor intensity from an odor measured by at least one gas sensor, wherein at least one odor analysis model is used to process the measured gas sensor data together with temperature and humidity data by using a machine learning technique, characterized in that: the step of (a) creating an odor database (10), comprising at least the steps of:(a-1) evaluating and labelling the odor type of the odor samples by at least one panelist (i i);(a-2) evaluating and labelling the odor intensity of the odor samples by the panelist (12);(a-3) storing measured gas sensor data of at least one sensor that is used to measure the same odor that is evaluated by the panelist (13);(a-4) storing temperature and humidity data of the surrounding atmosphere when the sensor measuring the odor (14);(a-5) providing the set of sample odor data (15), wherein each set at least includes the odor type labelled by the panelist, the odor intensity labelled by the panelist, the measured gas sensor data, the temperature data, and the humidity data;(a-6) storing the set of sample odor data in the odor database (16).

2. The method for odor sensing and monitoring as in claim 1, wherein the step of (b) creating a plurality of odor analysis models (20) comprising at least the steps of:(b-1) pre-processing the set of sample odor data (21) obtained in the step (a);(b-2) transforming said pre-processed data set (22);(b-3) training at least one odor analysis model (23) by using the transformed data set;(b-4) evaluating the performance of the trained odor analysis model (24);(b-5) storing the trained odor analysis model and its parameters (25).

3. The method for odor sensing and monitoring as in claim 1, wherein the step of (c) analyzing measured gas sensor data (30) comprising at least the steps of(c-1) storing measured gas sensor data of at least one sensor (31);(c-2) storing temperature and humidity data of the surrounding atmosphere (32) when the sensor measuring the odor;(c-3) processing the measured gas sensor data together with the temperature and humidity data by using at least one trained odor analysis model (33);(c-4) predicting the odor type and the odor intensity (34) of the odor measured by the sensor.

4. The method for odor sensing and monitoring as in claim 1, wherein the step of (a-2) evaluating and labelling the odor intensity of the odor samples by the panelist (12) can be selected from (1) collecting the odor samples into an odor bag and then analyzing the odor by scenting the odor samples based on the American Society for Testing and Materials (ASTM) or the Japanese Industrial Standard (JIS) method, or (2) scenting the odor in the environment using Portable Olfactometer technique following EN 13725:2022 standard.

5. The method for odor sensing and monitoring as in claim 1, wherein the odor analysis model comprises an odor-type prediction model and an odor-intensity prediction model.

6. The method for odor sensing and monitoring as in claim 5, wherein the odor-type prediction model can be selected from Decision Trees, Random Forest, Naive Bayes, Support Vector Machine, K-Nearest Neighbor (KNN), Gradient Boosting Machine, Light Gradient Boosting, Extreme Gradient Boosting, Extremely Randomized Tree Model, Distributed Random Forest, Generalized Linear Model, Stacked Ensembles, Deep Learning, Multilayer Perceptron, Recurrent Neural Networks, Long Short-Term Memory Networks, Generative Adversarial Networks, and Autoencoders Deep Learning Algorithm.

7. The method for odor sensing and monitoring as in claim 5, wherein the odor-type prediction model is Extremely Randomized Tree Model.

8. The method for odor sensing and monitoring as in claim 5, wherein the odor-intensity prediction model can be selected from Linear Regression, Logistic Regression, Polynomial Regression, Lasso Regression, Bayesian Linear Regression, Principal Components Regression, Partial Least Squares Regression, and Elastic Net Regression.

9. The method for odor sensing and monitoring as in claim 5, wherein the odor-intensity prediction model is Polynomial Regression.

10. The method for odor sensing and monitoring as in claim 1, the step of (b) creating a plurality of odor analysis models (20), further comprising the step of monitoring the odor analysis model deviation.

11. The method for odor sensing and monitoring as in claim 1, further comprising the step of setting warnings when the odor intensity level exceeds a threshold.

12. The method for odor sensing and monitoring as in claim 1, wherein the step of (c) analyzing the measured gas sensor data (30) at least comprises filtering the measured gas sensor data.

13. The method for odor sensing and monitoring as in claim 12, wherein the filtering is performed by using an exponential filter.

14. The method for odor sensing and monitoring as in claim 1, wherein the step of (c) analyzing the measured gas sensor data (30) at least comprises normalizing the measured gas sensor data by using baseline data obtained from clean air measurement.

15. The method for odor sensing and monitoring as in claim 1, wherein the measured gas sensor data are obtained from a metal oxide gas sensor.

16. The method for odor sensing and monitoring as in claim 1, wherein the measured gas sensor data are the data obtained from a sensor for sensing either hydrocarbons (CxHy), hydrogen gas (H2), carbon monoxide (CO), alcohols (CxHyOH), nitrogen dioxide (NO2), nitrogen monoxide (NO), ozone (O3), ammonia (NH3), sulfur dioxide (SO2), refrigerants R134a u.a., or volatile organic compounds, or the combination thereof.

17. The method for odor sensing and monitoring as in claim 1, wherein the measured gas sensor data are obtained from an array sensor having eight or more sensors.

18. The method for odor sensing and monitoring as in claim 1, further comprising the step of (d) displaying odor monitoring results (40) to display at least the measured gas sensor data of at least one sensor, the predicted odor type, the predicted odor intensity, the temperature, and the humidity.

19. The method for odor sensing and monitoring as in claim 18, wherein additional data to be displayed can be selected from an odor type prediction model, an odor intensity prediction model, a radar chart that represents an odor composition, a gas sensor status, a gas sensor location, local weather data, an odor diffusion pattern, and odor monitoring data history.

20. An odor sensing and monitoring system, comprising an odor database (110), at least one gas sensor (120), a temperature sensor (130), a humidity sensor (140), a processing unit (150), and a user interface unit (160), wherein: the odor database (110) stores data that are used for training an odor analysis model and for analyzing the odor; the gas sensor (120) measures odor samples diffused in the air in an odor monitoring area and transmits the measured gas sensor data to the processing unit via a communication mean; the temperature sensor (130) collects temperature data of the surrounding atmosphere when the gas sensor (120) measures the odor samples; the humidity sensor (140) collects humidity data of the surrounding atmosphere when the gas sensor (120) measures the odor samples; the processing unit (150) performs data processing to create an odor analysis model and to analyze the odor; and the user interface unit (160) displays data to a user or receives input from a user; characterized in that: the odor database (110) is divided into at least two data sets, comprising: an odor analysis model training data set and an odor analysis model data set, wherein the odor analysis model training data set comprises an odor type and intensity labelled by at least one panelist, measured gas sensor data of at least one sensor that is used to measure the same odor samples labelled by the panelist, temperature and humidity data during the measurement; and the odor analysis model data set comprises at least one odor analysis model and its parameters; and the processing unit (150) uses at least one odor analysis model to process the measured gas sensor data together with temperature and humidity data by using a machine learning technique.

21. The odor sensing and monitoring system as in claim 20, wherein the processing unit (150) performs data processing to create an odor analysis model by having at least the steps of:(b-1) pre-processing the odor analysis model training data set;(b-2) transforming said pre-processed data set;(b-3) training at least one odor analysis model;(b-4) evaluating the performance of the trained odor analysis model;(b-5) storing the trained odor analysis model and its parameters.

22. The odor sensing and monitoring system as in claim 20, wherein the processing unit (150) performs data processing to analyze the odor type and intensity by having at least the steps of:(c-1) storing the measured gas sensor data of at least one gas sensor (120);(c-2) storing the temperature and humidity data;(c-3) processing the measured gas sensor data of at least one gas sensor (120) together with temperature and humidity data and at least one odor analysis model from the odor analysis model data set;(c-4) predicting the odor type and intensity of the measured data.

23. The odor sensing and monitoring system as in claim 20, wherein the odor type and intensity labelled by the panelist is obtained from (1) collecting the odor samples into an odor bag and then analyzing the odor by scenting the odor samples based on the American Society for Testing and Materials (ASTM) or the Japanese Industrial Standard (JIS) method, or (2) scenting the odor in the environment using Portable Olfactometer technique following EN 13725:2022 standard.

24. The odor sensing and monitoring system as in claim 20, wherein the odor analysis model comprises an odor-type prediction model and an odor-intensity prediction model.

25. The odor sensing and monitoring system as in claim 24, wherein the odor-type prediction model can be selected from Decision Trees, Random Forest, Naive Bayes, Support Vector Machine, K-Nearest Neighbor (KNN), Gradient Boosting Machine, Light Gradient Boosting, Extreme Gradient Boosting, Extremely Randomized Tree Model, Distributed Random Forest, Generalized Linear Model, Stacked Ensembles, Deep Learning, Multilayer Perceptron, Recurrent Neural Networks, Long Short-Term Memory Networks, Generative Adversarial Networks, and Autoencoders Deep Learning Algorithm.

26. The odor sensing and monitoring system as in claim 24, wherein the odor-type prediction model is Extremely Randomized Tree Model.

27. The odor sensing and monitoring system as in claim 24, wherein the odor-intensity prediction model can be selected from Linear Regression, Logistic Regression, Polynomial Regression, Lasso Regression, Bayesian Linear Regression, Principal Components Regression, Partial Least Squares Regression, and Elastic Net Regression.

28. The odor sensing and monitoring system as in claim 24, wherein the odor-intensity prediction model is Polynomial Regression.

29. The odor sensing and monitoring system as in claim 20, wherein the processing unit (150) further comprises a filtering unit for filtering the measured gas sensor data.

30. The odor sensing and monitoring system as in claim 29, wherein the filtering unit is an exponential filter.

31. The odor sensing and monitoring system as in claim 20, wherein the processing unit (150) further comprises a normalization unit that normalizes the measured gas sensor data by using baseline data obtained from clean air measurement.

32. The odor sensing and monitoring system as in claim 20, wherein the gas sensor (120) is a metal oxide gas sensor.

33. The odor sensing and monitoring system as in claim 20, wherein the gas sensor (120) is the sensor for sensing either Hydrocarbons (CxHy), hydrogen gas (H2), carbon monoxide (CO), alcohols (CxHyOH), nitrogen dioxide (NO2), nitrogen monoxide (NO), ozone (O3), ammonia (NH3), sulfur dioxide (SO2), refrigerants R134a u.a. or volatile organic compounds, or combination thereof.

34. The odor sensing and monitoring system as in claim 20, wherein the gas sensor (120) is an array sensor having eight or more sensors.

35. The odor sensing and monitoring system as in claim 20, wherein the user interface unit (160) comprises at least a part for displaying measured gas sensor data of at least one sensor, a part for displaying a predicted odor type data, a part for displaying a predicted odor intensity data, a part of displaying the temperature sensor data, and a part of displaying the humidity sensor data.

36. The odor sensing and monitoring system as in claim 35, wherein the user interface unit (160) further comprises a data displaying part to display the data that can be selected from an odor type prediction model, an odor intensity prediction model, a radar chart that represents odor composition, a gas sensor status, a gas sensor location, local weather data, a predicted odor diffusion pattern, and odor monitoring data history.

37. The odor sensing and monitoring system as in claim 35, wherein the user interface unit (160) further comprises a part for displaying the odor analysis model deviation monitoring.

38. The odor monitoring system according to claim 35, wherein the user interface unit (160) further comprises a part for setting and displaying warnings when an odor intensity level exceeds a threshold.

Images