Method and system for training model for evaluating object freshness, method and system for evaluating object freshness, and computer program product

Abstract

This application relates to a method and system for training a model for evaluating object freshness. The method includes: respectively labeling VOC detection data, obtained by each semiconductor gas sensor, on an object of each category with a label for representing a freshness state of an object of a corresponding category; and training, based at least on a labeled VOC detection dataset, a category identification model for determining a category of an object and each freshness identification model for determining a freshness state of an object of a corresponding category. This application further relates to a method and system for evaluating object freshness, and a computer program product. According to this application, by using a two-stage training manner and an evaluation manner, each trained deep learning model can fully mine information included in VOC detection data, and focus more on a to-be-executed subtask, thereby effectively improving accuracy of predicting object freshness, and further effectively improving a freshness regulation function of a household appliance based on object freshness.

Description

[0001] METHOD AND SYSTEM FOR TRAINING MODEL FOR

[0002] EVALUATING OBJECT FRESHNESS, METHOD AND SYSTEM FOR EVALUATING OBJECT FRESHNESS, AND COMPUTER PROGRAM

[0003] PRODUCT

[0004] TECHNICAL FIELD

[0005] This application relates to the field of household appliances, and in particular, to a method for training a model for evaluating obj ect freshness, a system for training a model for evaluating object freshness, a method for evaluating object freshness, a system for evaluating object freshness, and a computer program product for at least enabling auxiliary implementation of steps of the methods according to this application.

[0006] BACKGROUND

[0007] In current household appliances (such as a refrigerator, a fresh-keeping cabinet, and a freezer) for object freshness preservation, a VOC (volatile organic compound) gas sensor made of a metal oxide is usually used for detecting a VOC gas component in a storage space, and freshness of an object placed in the storage space is determined based on the VOC gas component. However, in view of the fact that objects of different types release VOC gases of different types and concentrations as freshness decreases, object freshness determined directly based on the types and the concentrations of the VOC gases has a large deviation from an actual food freshness, which not only affects a function of predicting object freshness of a household appliance, but also affects a freshness regulation function based on the object freshness.

[0008] Therefore, how to accurately and rapidly evaluate object freshness in a household appliance becomes a technical problem to be resolved currently.

[0009] SUMMARY

[0010] An objective of this application is to provide a method for training a model for evaluating object freshness, a system for training a model for evaluating object freshness, a method for evaluating object freshness, a system for evaluating object freshness, and a computer program product, to resolve problems in the existing technology. According to a first aspect of this application, a method for training a model for evaluating object freshness is provided. The method includes:

[0011] - step SI: respectively labeling VOC detection data, obtained by each semiconductor gas sensor, on an object of each category with a label for representing a freshness state of an object of a corresponding category; and

[0012] - step S2: training, based at least on a labeled VOC detection dataset, a category identification model for determining a category of an object and each freshness identification model for determining a freshness state of an object of a corresponding category.

[0013] Based on a core idea of the method for training a model for evaluating object freshness according to this application, a training process of the model for evaluating object freshness is divided into a first training stage of training a category identification model for determining a category of an object and a second training stage of training each freshness identification model for determining a freshness state of an object of a corresponding category. Compared with the existing technology of training a single deep model using VOC detection data, a two-stage training manner enables each trained deep learning model to fully mine information included in VOC detection data, and to focus more on a to-be-executed subtask, thereby effectively improving accuracy of predicting object freshness, and further effectively improving a freshness regulation function of a household appliance based on object freshness.

[0014] According to an optional embodiment of this application, the category identification model and / or the freshness identification model may be configured as a deep learning model.

[0015] According to an optional embodiment of this application, step SI may include:

[0016] - step SI 01: performing cyclic temperature modulation in a preset temperature range on a plurality of semiconductor gas sensors, and obtaining VOC detection data, on an object of each category, of each semiconductor gas sensor by sampling, for example, using a sliding sampling window technique, in a cyclic temperature modulation period; and

[0017] - step SI 02: respectively labeling, based on a temperature feature of each semiconductor gas sensor, VOC detection data obtained by a corresponding semiconductor gas sensor in each cyclic temperature modulation period with a label for representing a freshness state of an obj ect of a corresponding category.

[0018] Different types of VOC gas detection data may be obtained by heating each semiconductor gas sensor and maintaining the temperature in different temperature ranges. In addition, by using the sliding sampling window technique, manpower and time costs can be significantly reduced, interference of a random error and a human factor on sample data can be effectively reduced, and distribution of obtained VOC detection data samples is balanced.

[0019] According to another optional embodiment of this application, the method may further include:

[0020] - step SI 1 : merging VOC detection data obtained by one semiconductor gas sensor in each cyclic temperature modulation period into VOC detection data of a dimension assigned to the semiconductor gas sensor, and assigning a label to VOC detection data of a corresponding dimension based on a labeled label, to construct a multi-dimensional labeled VOC detection dataset.

[0021] By constructing a multi-dimensional VOC detection dataset, richness of information included in the VOC detection dataset can be improved, to lay a foundation for training of a deep learning model for executing different identification tasks, thereby effectively improving the identification capability and prediction accuracy of the deep learning model.

[0022] According to another optional embodiment of this application, the method may further include:

[0023] - step S12: normalizing VOC detection data of one dimension based on a maximum VOC detection data value and a minimum VOC detection data value in the VOC detection data of the corresponding dimension.

[0024] In this manner, all elements of VOC detection data of each dimension are normalized into a value range of 0 to 1, thereby improving processing efficiency of the VOC detection data of each dimension. According to another optional embodiment of this application, the method may further include:

[0025] - step SI 3: performing feature importance analysis on VOC detection data of each dimension, and merging the VOC detection data of each dimension based on an analysis result of the feature importance, to reduce a dimension of the VOC detection dataset.

[0026] Optionally, based on the analysis result of the feature importance, a weight factor of VOC detection data assigned to a dimension with low feature importance may be assigned to the VOC detection data of the dimension with low feature importance, and a product of the VOC detection data of the dimension with low feature importance and the weight factor assigned to the dimension is merged into VOC detection data of a dimension with high feature importance, thereby reducing the dimensions of the VOC detection data, and effectively improving the efficiency of processing the VOC detection data by the deep learning model. Optionally, in particular, a weight factor of VOC detection data of a dimension with feature importance lower than a preset threshold is set to zero. Therefore, VOC detection data of an inaccurate dimension is deleted as a whole, to eliminate possible negative impacts on a detection result, thereby further improving processing efficiency and prediction accuracy of the deep learning model.

[0027] According to another optional embodiment of this application, the plurality of semiconductor gas sensors are made of same or different semiconductor materials, and the semiconductor gas sensors made of different semiconductor materials have different temperature features, so that different semiconductor gas sensors respectively present different sensitivity to different types of VOC gases in different temperature ranges.

[0028] According to a second aspect of this application, a system for training a model for evaluating object freshness is provided. The system is configured to perform the method for training a model for evaluating object freshness according to this application. The system may include the following components:

[0029] - a plurality of semiconductor gas sensors, configured to respectively obtain VOC detection data on an object of each category;

[0030] - a data labeling module, configured to respectively label VOC detection data, obtained by each semiconductor gas sensor, on an object of each category with a label for representing a freshness state of an object of a corresponding category; and

[0031] - a model training module, configured to train, based at least on a labeled VOC detection dataset, a category identification model for determining a category of an object and each freshness identification model for determining a freshness state of an object of a corresponding category.

[0032] According to another optional embodiment of this application, the system may further include a data construction module, configured to merge VOC detection data obtained by one semiconductor gas sensor in each cyclic temperature modulation period into VOC detection data of a dimension assigned to the semiconductor gas sensor, and assign a label to VOC detection data of a corresponding dimension based on a labeled label, to construct a multidimensional labeled VOC detection dataset, thereby laying a foundation for training a deep learning model using the VOC detection data.

[0033] According to a third aspect of this application, a method for evaluating object freshness is provided. The method may include: - step SI': determining, by a category identification model, a category of a to-be-identified object based on VOC detection data, obtained by semiconductor gas sensors, on the to-be- identified object; and

[0034] - step S2': evaluating a freshness state of the object based on a VOC detection dataset on the to-be-identified object by freshness identification models assigned to the determined category of the object.

[0035] Based on a core idea of the method for evaluating object freshness according to this application, an object freshness evaluation process is divided into two evaluation stages, where a category of a to-be-identified object is determined by using the category identification model in the first evaluation stage, and a freshness state of the object is evaluated by using the freshness identification models for objects of corresponding categories in the second evaluation stage. Compared with the existing technology of directly identifying object freshness based on VOC detection data, a two-stage evaluation manner enables each deep learning models to fully mine information included in VOC detection data, and to focus more on a to-be-executed subtask, thereby effectively improving accuracy of predicting object freshness of each category, and further effectively improving a freshness regulation function of a household appliance based on object freshness.

[0036] According to another optional embodiment of this application, the category identification model and / or the freshness identification model may be configured as a deep learning model.

[0037] According to another optional embodiment of this application, the method may further include:

[0038] - step Si l': merging VOC detection data obtained by one semiconductor gas sensor among a plurality of semiconductor gas sensors into VOC detection data of a dimension assigned to the semiconductor gas sensor, to construct a multi-dimensional VOC detection dataset.

[0039] By constructing a multi-dimensional VOC detection dataset, richness of information included in the VOC detection dataset can be improved, to lay a foundation for execution of a subtask of each deep learning model, thereby effectively improving accuracy of predicting object freshness.

[0040] According to another optional embodiment of this application, step Si l' may include:

[0041] - step Si l l': performing cyclic temperature modulation in a preset temperature range on a plurality of semiconductor gas sensors, and obtaining VOC detection data, on a to-be-identified object, of each semiconductor gas sensor by sampling, for example, using a sliding sampling window technique, in a cyclic temperature modulation period; and

[0042] - step SI 12': merging VOC detection data obtained by one semiconductor gas sensor in each cyclic temperature modulation period into VOC detection data of a dimension assigned to the semiconductor gas sensor, to construct a multi-dimensional VOC detection dataset.

[0043] Different types of VOC gas detection data may be obtained by heating each semiconductor gas sensor and maintaining the temperature in different temperature ranges. In addition, by using the sliding sampling window technique, manpower and time costs can be significantly reduced, interference of a random error and a human factor on sample data can be effectively reduced, and distribution of obtained VOC detection data samples is balanced.

[0044] According to another optional embodiment of this application, the method may further include:

[0045] - step S 12': normalizing VOC detection data of one dimension based on a maximum VOC detection data value and a minimum VOC detection data value of the VOC detection data of the corresponding dimension.

[0046] In this manner, all elements of VOC detection data of each dimension are normalized into a value range of 0 to 1, thereby improving processing efficiency of the VOC detection data of each dimension. According to another optional embodiment of this application, the method may further include:

[0047] - step S13': performing feature importance analysis on VOC detection data of each dimension, and merging the VOC detection data of each dimension based on an analysis result of the feature importance, to reduce a dimension of the VOC detection dataset.

[0048] Optionally, based on the analysis result of the feature importance, a weight factor of VOC detection data assigned to a dimension with low feature importance may be assigned to the VOC detection data of the dimension with low feature importance, and a product of the VOC detection data of the dimension with low feature importance and the weight factor assigned to the dimension is merged into VOC detection data of a dimension with high feature importance, thereby reducing the dimensions of the VOC detection data, and effectively improving the efficiency of processing the VOC detection data by the deep learning model.

[0049] Optionally, in particular, a weight factor of VOC detection data of a dimension with feature importance lower than a preset threshold is set to zero. Therefore, VOC detection data of an inaccurate dimension is deleted as a whole, to eliminate possible negative impacts on a detection result, thereby further improving processing efficiency and prediction accuracy of the deep learning model.

[0050] According to another optional embodiment of this application, the plurality of semiconductor gas sensors are made of same or different semiconductor materials, and the semiconductor gas sensors made of different semiconductor materials have different temperature features, so that different semiconductor gas sensors respectively present different sensitivity to different types of VOC gases in different temperature ranges.

[0051] According to another optional embodiment of this application, the method may further include:

[0052] - step S3': outputting notification information about a freshness state of the evaluated object.

[0053] By outputting the notification information, a user can learn a freshness state of an object in a storage space of a household appliance in time.

[0054] According to a fourth aspect of this application, a system for evaluating object freshness is provided. The system is configured to perform the method for evaluating object freshness according to this application. The system may include the following components:

[0055] - semiconductor gas sensors, configured to respectively obtain VOC detection data on a to-be-identified object;

[0056] - a category identification model, configured to determine a category of a to-be-identified object based on VOC detection data, obtained by each semiconductor gas sensor, on the to-be- identified object; and

[0057] - freshness identification models, configured to evaluate a freshness state of the object based on a VOC detection dataset on the to-be-identified object by the freshness identification models assigned to the determined category of the object.

[0058] According to another optional embodiment of this application, the system may further include a data construction module, configured to merge VOC detection data obtained by one semiconductor gas sensor into VOC detection data of a dimension assigned to the semiconductor gas sensor, to construct a multi-dimensional VOC detection dataset. By constructing a multi-dimensional VOC detection dataset, richness of information included in the VOC detection dataset can be improved, so that a deep learning model executing different identification tasks can mine more effective information from the multi-dimensional VOC detection dataset, thereby effectively improving the identification capability and prediction accuracy of the deep learning model. According to another optional embodiment of this application, the system may further include a notification module, configured to output notification information about a freshness state of an evaluated object, so that a user can learn a freshness state of an object in a storage space of a household appliance in time.

[0059] According to a fifth aspect of this application, a computer program product, for example, a computer-readable program carrier, is provided, including computer program instructions. When the computer program instructions are executed by a processor, the steps of the method for evaluating object freshness and the steps of the method for training a model for evaluating object freshness according to this application are at least assisted in being implemented.

[0060] BRIEF DESCRIPTION OF THE DRAWINGS

[0061] The following describes this application in more detail with reference to the accompanying drawings, to better understand the principles, features, and advantages of this application. The accompanying drawings include:

[0062] FIG. 1 shows a working flowchart of a method for training a model for evaluating object freshness according to an exemplary embodiment of this application;

[0063] FIG. 2 shows a working flowchart of a method for training a model for evaluating object freshness according to another exemplary embodiment of this application;

[0064] FIG. 3 shows a working flowchart of a method for training a model for evaluating object freshness according to another exemplary embodiment of this application;

[0065] FIG. 4 shows a working flowchart of a method for training a model for evaluating object freshness according to another exemplary embodiment of this application;

[0066] FIG. 5 shows a working flowchart of a method for training a model for evaluating object freshness according to another exemplary embodiment of this application;

[0067] FIG. 6 shows a block diagram of a system for training a model for evaluating object freshness according to an exemplary embodiment of this application;

[0068] FIG. 7 shows a working flowchart of a method for evaluating object freshness according to an exemplary embodiment of this application;

[0069] FIG. 8 shows a working flowchart of a method for evaluating object freshness according to another exemplary embodiment of this application;

[0070] FIG. 9 shows a working flowchart of a method for evaluating object freshness according to another exemplary embodiment of this application;

[0071] FIG. 10 shows a working flowchart of a method for evaluating object freshness according to another exemplary embodiment of this application;

[0072] FIG. 11 shows a working flowchart of a method for evaluating object freshness according to another exemplary embodiment of this application;

[0073] FIG. 12 shows a working flowchart of a method for evaluating object freshness according to another exemplary embodiment of this application; and

[0074] FIG. 13 shows a block diagram of a system for evaluating object freshness according to an exemplary embodiment of this application.

[0075] DETAILED DESCRIPTION

[0076] To make technical problems to be resolved, technical solutions, and beneficial technical effects of this application clearer and more comprehensible, this application is further described in detail below with reference to the accompanying drawings and a plurality of exemplary embodiments. It should be understood that specific embodiments described herein are merely used for explaining this application but are not used for limiting the protection scope of this application.

[0077] FIG. 1 shows a working flowchart of a method for training a model for evaluating object freshness according to an exemplary embodiment of this application. The following exemplary embodiments describe the method according to this application in more detail.

[0078] As shown in FIG. 1, the method may include steps SI and S2. In step SI, VOC detection data, obtained by each semiconductor gas sensor, on an object of each category is respectively labeled with a label for representing a freshness state of an object of a corresponding category.

[0079] A plurality of semiconductor gas sensors are usually disposed in a household appliance (such as a refrigerator, a freezer, and a fresh-keeping cabinet) for object freshness preservation, and are configured to respectively obtain VOC detection data on a to-be-identified object. The plurality of semiconductor gas sensors may be made of same semiconductor materials. For example, the semiconductor gas sensors are zinc oxide gas sensors having excellent gassensitive performance and wide applicability. Because temperature features of the zinc gas sensors indicate that the zinc gas sensors present different sensitivity to different types of VOC gases in different temperature ranges, corresponding VOC gas detection data may be obtained by heating each zinc gas sensor and maintaining the temperature in different temperature ranges.

[0080] Optionally, the plurality of semiconductor gas sensors may alternatively be made of different semiconductor materials, and the semiconductor gas sensors made of different semiconductor materials have different temperature features. To be specific, different semiconductor gas sensors respectively present different sensitivity to different types of VOC gases in different temperature ranges. For example, semiconductor gas sensors made of Codoped SnCh nano-sensitive materials present high sensitivity to formaldehyde and acetone in a preset temperature range.

[0081] A system 1 for training a model for evaluating object freshness shown in FIG. 6 includes, for example, a first semiconductor gas sensor 111, a second semiconductor gas sensor 112, a third semiconductor gas sensor 113, and a fourth semiconductor gas sensor 114. The semiconductor gas sensors may respectively present different sensitivity to different types of gases in different temperature ranges, thereby obtaining VOC detection data of various types and / or various concentrations of VOC gases released by a to-be-detected object. In view of the fact that when a deep learning model uses VOC detection data for training, the VOC detection data is required to have a label and VOC gases of different types and concentrations are released by objects of different categories as freshness decreases, a category and freshness of the to-be- detected object need to be labeled when the label of the VOC detection data is labeled.

[0082] The following describes step S 1 in detail with reference to a working flowchart of a method for training a model for evaluating object freshness according to another exemplary embodiment of this application shown in FIG. 2. As shown in FIG. 2, step SI may include steps SI 01 and SI 02. In step SI 01, cyclic temperature modulation in a preset temperature range is performed on a plurality of semiconductor gas sensors, and VOC detection data, on an object of each category, of each semiconductor gas sensor is obtained by sampling in a cyclic temperature modulation period.

[0083] In a test scenario set in this application, obj ects of different categories, such as a strawberry, a litchi, and a waxberry, may be respectively placed in a storage space of a household appliance. In a process of reducing freshness, VOC gases of different types and different concentrations are respectively released. A current having a modulated waveform, for example, a rectangular wave having a modulation period Tt, a maximum current value Imin, a maximum current value Imax, and a duty ratio D, may be applied to a semiconductor gas sensor in a cyclic temperature modulation period of, for example, 100°C to 120°C, to increase a temperature of the semiconductor gas sensor and maintain the temperature in a first preset temperature range. The semiconductor gas sensor is, for example, highly sensitive to a first VOC gas in the first preset temperature range, thereby obtaining a VOC time sequence response of the first VOC gas, sampling the VOC time sequence response at a fixed sampling frequency to obtain VOC time sequence response data on the first VOC gas. Next, the temperature of the semiconductor gas sensor is further increased by applying a current having a modulated waveform and is maintained in a second preset temperature range, thereby obtaining VOC time sequence response data on a second VOC gas. The rest can be deduced by analogy until the temperature of the semiconductor gas sensor traverses the entire cyclic temperature modulation period from 100°C to 120°C. Then, a next cyclic temperature modulation period is repeatedly performed after the temperature of the semiconductor gas sensor is reduced to 100°C.

[0084] Herein, each semiconductor gas sensor may be heated, and the temperature may be maintained in different preset temperature ranges, to obtain VOC time sequence response data of VOC gases of as many types as possible. The plurality of semiconductor gas sensors therein may alternatively be heated, and the temperatures are maintained in a same preset temperature range. In addition, VOC time sequence response data of some types of VOC gases is repeatedly detected, to verify credibility of types of VOC gases importantly related to a function of predicting object freshness, for example, a VOC gas generated when food spoils and a VOC gas generated by using mildew in food. For example, a sliding sampling window may be disposed on a time axis of the VOC time sequence response data.

[0085] A sampling window width Tc is equal to an integer (m) multiple of a modulation period Tt of the applied waveform, and a sliding step length is one sampling data point. The sampling window slides in a direction of a time axis from a first sampling data point of the VOC time sequence response data. Each time the sampling window slides for one step length, all sampling data in the sampling window is captured as one group of VOC time sequence response sample data, so that a plurality of groups of VOC time sequence response sample data can be obtained for each type of VOC gas, and VOC detection data of each semiconductor gas sensor is obtained by continuous resampling of the VOC time sequence response data point by point, thereby significantly reducing manpower and time costs, and effectively reducing interference of a random error and a human factor to the sample data. In addition, distribution of obtained VOC detection data samples is balanced.

[0086] In step SI 02, based on a temperature feature of each semiconductor gas sensor, VOC detection data obtained by a corresponding semiconductor gas sensor in each cyclic temperature modulation period is respectively labeled with a label for representing a freshness state of an object of a corresponding category. In a test scenario in which objects of different categories are respectively placed in a storage space, a category of a to-be-detected object may be respectively labeled, for VOC detection data obtained in the corresponding test scenario, and freshness of the to-be-detected object is respectively labeled in a process of reducing freshness of an object of a corresponding category.

[0087] In step S2, based at least on a labeled VOC detection dataset, a category identification model 22 for determining a category of an object and each freshness identification model for determining a freshness state of an object of a corresponding category are trained.

[0088] In view of the fact that objects of different types release VOC gases of different types and concentrations as the freshness decreases, object freshness determined by using a single deep learning model directly based on the types and the concentrations of the VOC gases has a large deviation from an actual food freshness. Herein, a training process of the model for evaluating object freshness is divided into a first training stage of training the category identification model 22 for determining a category of an object and a second training stage of training each freshness identification model for determining a freshness state of an object of a corresponding category, so that each trained deep learning models can fully mine information included in VOC detection data, and focus more on a to-be-executed subtask, thereby effectively improving precision of predicting a freshness state of an object of each category.

[0089] For example, a freshness identification model for determining a freshness state of an object of a corresponding category may be provided for a to-be-detected object of each category. The freshness identification model is, for example, a first freshness identification model 231 for determining a freshness state of a strawberry, a second freshness identification model 232 for determining a freshness state of a litchi, or a third freshness identification model 233 for determining a freshness state of a waxberry. The category identification model 22 and / or the freshness identification models 231, 232, 233 may be configured as a deep learning model, for example, including a convolutional neural network, a multilayer self-encoding neural network, and / or a deep confidence network, or may be configured as another type of mathematical model capable of implementing a same function.

[0090] In the existing technology, time sequence response sample data directly obtained by a semiconductor gas sensor is one-dimensional data, a deep learning model can extract less effective information from the one-dimensional time sequence response sample data, and prediction precision is low. This greatly limits application of a VOC gas sensor to a complex scenario of a household appliance. A construction process of a multi-dimensional VOC detection dataset is described in detail below with reference to a working flowchart of a method for training a model for evaluating object freshness according to another exemplary embodiment of this application shown in FIG. 3. Only differences from the embodiment shown in FIG. 1 are described below. For brevity, same steps are not described again.

[0091] As shown in FIG. 3, the method may further include step SI 1. In step SI 1, VOC detection data obtained by one semiconductor gas sensor in each cyclic temperature modulation period is merged into VOC detection data of a dimension assigned to the semiconductor gas sensor, and a label is assigned to VOC detection data of a corresponding dimension based on a labeled label, to construct a multi-dimensional labeled VOC detection dataset. In the meaning of this application, the term "merging" may be understood as: constructing one row of data elements using VOC detection data obtained by one semiconductor gas sensor in a cyclic temperature modulation period, and processing the VOC detection data of the semiconductor gas sensor one by one in the cyclic temperature modulation periods in this manner, to construct corresponding rows of data elements, where quantities of VOC detection data obtained in each cyclic temperature modulation period are equal, to be specific, quantities of data elements in each row are equal. Therefore, a matrix in which a quantity of rows is equal to a quantity of cyclic temperature modulation periods may be constructed by using these rows of data elements. The matrix is VOC detection data assigned to one dimension of the semiconductor gas sensor. Herein, data processing, including label assignment, normalization, feature importance analysis, and / or weight factor assignment, may be uniformly performed on VOC detection data of a same dimension. VOC detection data of a corresponding dimension is constructed in this manner for each semiconductor gas sensor, thereby obtaining a multi-dimensional VOC detection dataset. For example, VOC detection data of a first dimension is made of VOC detection data obtained by the first semiconductor gas sensor 111 in a plurality of cyclic temperature modulation periods, VOC detection data of a second dimension is made of VOC detection data obtained by the second semiconductor gas sensor 112 in a plurality of cyclic temperature modulation periods, VOC detection data of a third dimension is made of VOC detection data obtained by the third semiconductor gas sensor 113 in a plurality of cyclic temperature modulation periods, VOC detection data of a fourth dimension is made of VOC detection data obtained by the fourth semiconductor gas sensor 114 in a plurality of cyclic temperature modulation periods, and so on. In view of the fact that quantities of cyclic temperature modulation periods in which each semiconductor gas sensor performs VOC gas detection are equal, VOC detection data volumes of each dimension are also equal.

[0092] In a process of merging VOC detection data of different dimensions, a label may be assigned to VOC detection data of a corresponding dimension based on a label of original onedimensional VOC detection data, so that VOC detection data of each dimension obtained by merging has a corresponding label, thereby constructing a multi-dimensional labeled VOC detection dataset. The label includes a category and freshness of a to-be-detected object. To improve processing efficiency of VOC detection data of each dimension, FIG. 4 shows a working flowchart of a method for training a model for evaluating object freshness according to another exemplary embodiment of this application. The method may further include step S12. In step S12, VOC detection data of one dimension is normalized based on a maximum VOC detection data value and a minimum VOC detection data value in the VOC detection data of the corresponding dimension, so that all elements of the VOC detection data of each dimension are normalized into a value range of 0 to 1.

[0093] In the presence of a larger number of semiconductor gas sensors in the system 1, there are also many dimensions of obtained VOC detection data. VOC detection data of excessive dimensions may affect processing efficiency of the deep learning model, and VOC detection data of dimensions obtained by the semiconductor gas sensors with low detection precision may affect prediction accuracy of the deep learning model. Therefore, there is a requirement of reducing the dimensions of the VOC detection data in some application scenarios. A process of reducing a dimension of VOC detection data is described in detail below with reference to a working flowchart of a method for training a model for evaluating object freshness according to another exemplary embodiment of this application shown in FIG. 5.

[0094] As shown in FIG. 5, the method may further include step S13. In step S13, feature importance analysis is performed on VOC detection data of each dimension, and the VOC detection data of each dimension is merged based on an analysis result of the feature importance, to reduce a dimension of the VOC detection dataset.

[0095] Herein, feature importance analysis may be performed on the VOC detection data, for example, in terms of a type, a concentration, and detection precision of a VOC gas. VOC detection data with more types of VOC gas (particularly, including a VOC gas of a type importantly related to a function of predicting object freshness), a higher concentration, and / or a higher detection precision has high feature importance, while VOC detection data with fewer types of VOC gas, a lower concentration, and / or a lower detection precision has low feature importance. Based on the analysis result of the feature importance, a weight factor of VOC detection data assigned to a dimension with low feature importance is assigned to the VOC detection data of the dimension with low feature importance, and a product of the VOC detection data of the dimension with low feature importance and the weight factor assigned to the dimension is merged into VOC detection data of a dimension with high feature importance.

[0096] For example, the label of the VOC detection data of the fourth dimension has fewer types of VOC gas, a lower concentration, or a lower detection precision, and the VOC detection data of the fourth dimension may be merged into the VOC detection data of another dimension, thereby reducing the dimensions of the VOC detection data. In particular, when it is identified, according to the VOC detection data of the fourth dimension, that the VOC detection data of the fourth semiconductor gas sensor 114 obviously deviates from a normal detection concentration range, it may be determined that the feature importance of the VOC detection data of the fourth dimension is less than a preset threshold, and a weight factor of VOC detection data of a dimension with feature importance lower than a preset threshold is set to zero. Therefore, VOC detection data of an inaccurate dimension is deleted as a whole, to eliminate possible negative impacts on a detection result, thereby improving processing efficiency and prediction accuracy of the deep learning model.

[0097] According to a current embodiment of this application, a training process of the model for evaluating object freshness is divided into a first training stage of training a category identification model for determining a category of an object and a second training stage of training each freshness identification model for determining a freshness state of an object of a corresponding category. Compared with the existing technology of training a single deep model using VOC detection data, a two-stage training manner enables each trained deep learning model to fully mine information included in VOC detection data, and to focus more on a to- be-executed subtask, thereby effectively improving accuracy of predicting object freshness, and further effectively improving a freshness regulation function of a household appliance based on object freshness.

[0098] FIG. 6 shows a block diagram of a system 1 for training a model for evaluating object freshness according to an exemplary embodiment of this application. As shown in FIG. 6, the system 1 may include the following components:

[0099] - a plurality of semiconductor gas sensors, configured to respectively obtain VOC detection data on an object of each category, and including, for example, a first semiconductor gas sensor 111, a second semiconductor gas sensor 112, a third semiconductor gas sensor 113, and / or a fourth semiconductor gas sensor 114;

[0100] - a data labeling module 12, configured to respectively label VOC detection data, obtained by each semiconductor gas sensor, on an object of each category with a label for representing a freshness state of an object of a corresponding category; and

[0101] - a model training module 13, configured to train, based at least on a labeled V OC detection dataset, a category identification model 22 for determining a category of an object and each freshness identification model for determining a freshness state of an object of a corresponding category, where the freshness identification model includes, for example, a first freshness identification model 231 for determining a freshness state of a strawberry, a second freshness identification model 232 for determining a freshness state of a litchi, and / or a third freshness identification model 233 for determining a freshness state of a waxberry.

[0102] Optionally, the system 1 may further include a data construction module 14, configured to merge VOC detection data obtained by one semiconductor gas sensor in each cyclic temperature modulation period into VOC detection data of a dimension assigned to the semiconductor gas sensor, and assign a label to VOC detection data of a corresponding dimension based on a labeled label, to construct a multi-dimensional labeled VOC detection dataset.

[0103] After a deep learning model, including the category identification model and freshness identification models assigned to an object of each category, is trained by using the foregoing method, a freshness state of an object of each category in a storage space of a household appliance may be evaluated by using a method for evaluating object freshness according to an exemplary embodiment of this application shown in FIG. 7.

[0104] As shown in FIG. 7, the method may include steps SI' and S2'. In step SI', by a category identification model 22, a category of a to-be-identified object is determined based on VOC detection data, obtained by semiconductor gas sensors, on the to-be-identified object. For example, it is determined that an object in a storage space of a household appliance includes a strawberry, a litchi, and / or a waxberry.

[0105] In view of the fact that different types of objects release different types and concentrations of VOC gases as the freshness decreases, in step S2', a freshness state of the object may be evaluated based on a VOC detection dataset on the to-be-identified object by freshness identification models assigned to the determined category of the object. For example, a freshness state of the strawberry is evaluated by using the first freshness identification model 231, a freshness state of the litchi is evaluated by using the second freshness identification model 232, and a freshness state of the waxberry is evaluated by using the third freshness identification model 233.

[0106] According to a current embodiment of this application, an object freshness evaluation process is divided into two evaluation stages, where a category of a to-be-identified object is determined by the category identification model 22 in the first evaluation stage, and a freshness state of the object is evaluated by the freshness identification models for objects of corresponding categories in the second evaluation stage.

[0107] Compared with the existing technology of directly identifying object freshness based on VOC detection data, a two-stage evaluation manner enables each deep learning models to fully mine information included in VOC detection data, and to focus more on a to-be-executed subtask, thereby effectively improving accuracy of predicting object freshness of each category, and further effectively improving a freshness regulation function of a household appliance based on object freshness.

[0108] In the existing technology, time sequence response sample data directly obtained by a semiconductor gas sensor is one-dimensional data, a deep learning model can extract less effective information from the one-dimensional time sequence response sample data, and prediction precision is low. This greatly limits application of a VOC gas sensor to a complex scenario of a household appliance. A construction process of a multi-dimensional VOC detection dataset is described in detail below with reference to a working flowchart of a method for evaluating object freshness according to another exemplary embodiment of this application shown in FIG. 8. Only differences from the embodiment shown in FIG. 7 are described below. For brevity, same steps are not described again.

[0109] As shown in FIG. 8, the method may further include step S 11'. In step S 11', V OC detection data obtained by one semiconductor gas sensor among a plurality of semiconductor gas sensors is merged into VOC detection data of a dimension assigned to the semiconductor gas sensor, to construct a multi-dimensional VOC detection dataset. A construction manner of the multidimensional VOC detection dataset is the same as that of a model training stage. The following describes step Si l' in detail with reference to a working flowchart of a method for evaluating object freshness according to another exemplary embodiment of this application shown in FIG. 9.

[0110] As shown in FIG. 9, step Si l' may include steps Si l l' and SI 12'. In step Si l l', cyclic temperature modulation in a preset temperature range is performed on a plurality of semiconductor gas sensors, and VOC detection data, on a to-be-identified object, of each semiconductor gas sensor is obtained by sampling in a cyclic temperature modulation period. As already described above, currents having a modulated waveform may also be applied to semiconductor gas sensors 211, 212, 213, 214, in a cyclic temperature modulation period of, for example, 100°C to 120°C, to increase temperatures of the corresponding semiconductor gas sensors and maintain the temperatures in a first preset temperature range. The semiconductor gas sensors are, for example, highly sensitive to a first VOC gas in the first preset temperature range, thereby obtaining a VOC time sequence response of the first VOC gas, sampling the VOC time sequence response at a fixed sampling frequency to obtain VOC time sequence response data on the first VOC gas. Next, the temperature of the semiconductor gas sensor is further increased by applying a current having a modulated waveform and is maintained in a second preset temperature range, thereby obtaining VOC time sequence response data on a second VOC gas. The rest can be deduced by analogy until the temperature of the semiconductor gas sensor traverses the entire cyclic temperature modulation period from 100°C to 120°C. Then, a next cyclic temperature modulation period is repeatedly performed after the temperature of the semiconductor gas sensor is reduced to 100°C.

[0111] Herein, each semiconductor gas sensor may be heated, and the temperature may be maintained in different preset temperature ranges, to obtain VOC time sequence response data of VOC gases of as many types as possible. The plurality of semiconductor gas sensors therein may alternatively be heated, and the temperatures are maintained in a same preset temperature range. In addition, VOC time sequence response data of some types of VOC gases is repeatedly detected, to verify credibility of types of VOC gases importantly related to a function of predicting object freshness, for example, a VOC gas generated when food spoils and a VOC gas generated by using mildew in food. For example, a sliding sampling window may be disposed on a time axis of the VOC time sequence response data.

[0112] A sampling window width Tc is equal to an integer (m) multiple of a modulation period Tt of the applied waveform, and a sliding step length is one sampling data point. The sampling window slides in a direction of a time axis from a first sampling data point of the VOC time sequence response data. Each time the sampling window slides for one step length, all sampling data in the sampling window is captured as one group of VOC time sequence response sample data, so that a plurality of groups of VOC time sequence response sample data can be obtained for each type of VOC gas, and VOC detection data of each semiconductor gas sensor is obtained by continuous resampling of the VOC time sequence response data point by point, thereby significantly reducing manpower and time costs, and effectively reducing interference of a random error and a human factor to the sample data. In addition, distribution of obtained VOC detection data samples is balanced.

[0113] In step SI 12', VOC detection data obtained by one semiconductor gas sensor in each cyclic temperature modulation period is merged into VOC detection data of a dimension assigned to the semiconductor gas sensor, to construct a multi-dimensional VOC detection dataset. Herein, for example, VOC detection data obtained by the first semiconductor gas sensor 211 in a plurality of cyclic temperature modulation periods is merged into VOC detection data of a first dimension, VOC detection data obtained by the second semiconductor gas sensor 212 in a plurality of cyclic temperature modulation periods is merged into VOC detection data of a second dimension, VOC detection data obtained by the third semiconductor gas sensor 213 in a plurality of cyclic temperature modulation periods is merged into VOC detection data of a third dimension, VOC detection data obtained by the fourth semiconductor gas sensor 214 in a plurality of cyclic temperature modulation periods is merged into VOC detection data of a fourth dimension, and so on. In view of the fact that quantities of cyclic temperature modulation periods in which each semiconductor gas sensor performs VOC gas detection are equal, VOC detection data volumes of each dimension are also equal.

[0114] It should be noted that accuracy of evaluating object freshness based on one-dimensional VOC detection data by a current deep learning model is relatively low. According to test data, average accuracy of evaluating, based on one-dimensional VOC detection data, freshness of various types of foods by a deep learning sub-model trained based on the one-dimensional VOC detection data is only 26%. For example, accuracy of evaluating freshness of a strawberry is 63%, accuracy of evaluating freshness of a litchi is 46%, and accuracy of evaluating freshness of a litchi is 42%. According to a trained deep learning model of this application, average accuracy of evaluating freshness of various types of foods based on a multidimensional VOC detection dataset is 85%. For example, accuracy of evaluating freshness of a strawberry by using the trained first freshness identification model 231 based on the multidimensional VOC detection dataset is 86%, accuracy of evaluating freshness of a litchi by using the trained second freshness identification model 232 based on the multi-dimensional VOC detection dataset is 85%, accuracy of evaluating freshness of a litchi by using the trained third freshness identification model 233 based on the multi-dimensional VOC detection dataset is 75%, and so on.

[0115] According to the embodiments of this application, by constructing a multi-dimensional VOC detection dataset, richness of information included in the VOC detection dataset can be improved, to lay a foundation for training of a deep learning model for evaluating a freshness state of an object and execution of a related task, thereby effectively improving prediction accuracy of the deep learning model for the freshness state of the object.

[0116] To improve processing efficiency of VOC detection data of each dimension, FIG. 10 shows a working flowchart of a method for evaluating object freshness according to another exemplary embodiment of this application. The method may further include step S 12'. In step S 12', VOC detection data of one dimension is normalized based on a maximum VOC detection data value and a minimum VOC detection data value in the VOC detection data of the corresponding dimension, so that all elements of the VOC detection data of each dimension are normalized into a value range of 0 to 1.

[0117] In the presence of a larger number of semiconductor gas sensors in the system 2, there are also many dimensions of obtained VOC detection data. VOC detection data of excessive dimensions may affect processing efficiency of the deep learning model, and VOC detection data of dimensions obtained by the semiconductor gas sensors with low detection precision may affect prediction accuracy of the deep learning model. Therefore, there is a requirement of reducing the dimensions of the VOC detection data in some application scenarios. A process of reducing a dimension of VOC detection data is described in detail below with reference to a working flowchart of a method for evaluating object freshness according to another exemplary embodiment of this application shown in FIG. 11.

[0118] As shown in FIG. 11, the method may further include step S13'. In step S13', feature importance analysis is performed on VOC detection data of each dimension, and the VOC detection data of each dimension is merged based on an analysis result of the feature importance, to reduce a dimension of the VOC detection dataset. Herein, feature importance analysis may be performed on the VOC detection data, for example, in terms of a type, a concentration, and detection precision of a VOC gas. VOC detection data with more types of VOC gas (particularly, including a VOC gas of a type importantly related to a function of predicting object freshness), a higher concentration, and / or a higher detection precision has high feature importance, while VOC detection data with fewer types of VOC gas, a lower concentration, and / or a lower detection precision has low feature importance. Based on the analysis result of the feature importance, a weight factor of VOC detection data assigned to a dimension with low feature importance is assigned to the VOC detection data of the dimension with low feature importance, and a product of the VOC detection data of the dimension with low feature importance and the weight factor assigned to the dimension is merged into VOC detection data of a dimension with high feature importance.

[0119] For example, the label of the VOC detection data of the fourth dimension has fewer types of VOC gas, a lower concentration, or a lower detection precision, and the VOC detection data of the fourth dimension may be merged into the VOC detection data of another dimension, thereby reducing the dimensions of the VOC detection data. In particular, when it is identified, according to the VOC detection data of the fourth dimension, that the VOC detection data of the fourth semiconductor gas sensor 214 obviously deviates from a normal detection concentration range, it may be determined that the feature importance of the VOC detection data of the fourth dimension is less than a preset threshold, and a weight factor of VOC detection data of a dimension with feature importance lower than a preset threshold is set to zero. Therefore, VOC detection data of an inaccurate dimension is deleted as a whole, to eliminate possible negative impacts on a detection result, thereby improving processing efficiency and prediction accuracy of the deep learning model.

[0120] FIG. 12 shows a working flowchart of a method for evaluating object freshness according to another exemplary embodiment of this application. Only differences from the embodiment shown in FIG. 11 are described below. For brevity, same steps are not described again.

[0121] As shown in FIG. 12, the method may further include step S3'. In step S3', notification information about a freshness state of the evaluated object is output, for example, by using a notification module 25. The notification module 25 is, for example, constructed as a display screen and / or a voice device for displaying the notification information to a user, so that the user can leam a freshness state of an object in a storage space of a household appliance in time.

[0122] In addition, it should be noted that sequence numbers of steps described herein do not necessarily represent an order, but are merely reference numerals. The order may be changed according to specific situations, provided that the technical objectives of this application can be achieved.

[0123] FIG. 13 shows a block diagram of a system 2 for evaluating object freshness according to an exemplary embodiment of this application.

[0124] As shown in FIG. 13, the system 2 may include the following components:

[0125] - semiconductor gas sensors, configured to respectively obtain VOC detection data on a to-be-identified object, and including, for example, a first semiconductor gas sensor 211, a second semiconductor gas sensor 212, a third semiconductor gas sensor 213, and / or a fourth semiconductor gas sensor 214;

[0126] - a category identification model 22, configured to determine a category of a to-be- identified object based on VOC detection data, obtained by each semiconductor gas sensor, on the to-be-identified object; and

[0127] - freshness identification models, configured to evaluate a freshness state of the object based on a VOC detection dataset on the to-be-identified object by the freshness identification models assigned to the determined category of the object, where the freshness identification models include, for example, a first freshness identification model 231 for determining a freshness state of a strawberry, a second freshness identification model 232 for determining a freshness state of a litchi, and / or a third freshness identification model 233 for determining a freshness state of a waxberry.

[0128] Optionally, the system 2 may further include a data construction module 24, configured to merge VOC detection data obtained by one semiconductor gas sensor into VOC detection data of a dimension assigned to the semiconductor gas sensor, to construct a multi-dimensional VOC detection dataset.

[0129] Optionally, the system 2 may further include a notification module 25, configured to output notification information about a freshness state of an evaluated object.

[0130] It will be appreciated that the expressions "first", "second", "third", and the like are merely used for descriptive purposes, and should not be understood as indicating or implying relative importance, nor as implicitly specifying the quantity of indicated technical features.

[0131] Although specific embodiments have been described above, these embodiments are not intended to limit the scope of this application, even when only a single embodiment is described with respect to specific features. Feature examples provided in the disclosure of this application are intended to be exemplary rather than limiting, unless indicated differently. In a specific implementation, a plurality of features may be combined with each other according to actual requirements and in a technically feasible case. Various replacements, changes, and alterations may further be conceived without departing from the spirit and scope of this application.

Claims

CLAIMSWhat is claimed is:

1. A method for training a model for evaluating object freshness, the method comprising: step SI: respectively labeling VOC detection data, obtained by each semiconductor gas sensor (111, 112, 113, 114), on an object of each category with a label for representing a freshness state of an object of a corresponding category; and step S2: training, based at least on a labeled VOC detection dataset, a category identification model (22) for determining a category of an object and each freshness identification model (231, 232, 233) for determining a freshness state of an object of a corresponding category.

2. The method according to claim 1, wherein step SI comprises: step SI 01: performing cyclic temperature modulation in a preset temperature range on a plurality of semiconductor gas sensors (111, 112, 113, 114), and obtaining VOC detection data, on an object of each category, of each semiconductor gas sensor (111, 112, 113, 114) by sampling, for example, using a sliding sampling window technique, in a cyclic temperature modulation period; and step SI 02: respectively labeling, based on a temperature feature of each semiconductor gas sensor (111, 112, 113, 114), VOC detection data obtained by a corresponding semiconductor gas sensor (111, 112, 113, 114) in each cyclic temperature modulation period with a label for representing a freshness state of an object of a corresponding category.

3. The method according to claim 2, wherein the method further comprises: step Sil: merging VOC detection data obtained by one semiconductor gas sensor (111, 112, 113, 114) in each cyclic temperature modulation period into VOC detection data of a dimension assigned to the semiconductor gas sensor (l ll, 112, 113, 114), and assigning a label to VOC detection data of a corresponding dimension based on a labeled label, to construct a multi-dimensional labeled VOC detection dataset.

4. The method according to claim 3, wherein the method further comprises: step S12: normalizing VOC detection data of one dimension based on a maximum VOC detection data value and a minimum VOC detection data value in the VOC detection data of the corresponding dimension; and / or wherein the method further comprises:step SI 3: performing feature importance analysis on VOC detection data of each dimension, and merging the VOC detection data of each dimension based on an analysis result of the feature importance, to reduce a dimension of the VOC detection dataset; or addionally to step S12, wherein based on the analysis result of the feature importance, a weight factor of VOC detection data assigned to a dimension with low feature importance is assigned to the VOC detection data of the dimension with low feature importance, and a product of the VOC detection data of the dimension with low feature importance and the weight factor assigned to the dimension is merged into VOC detection data of a dimension with high feature importance, in particular, a weight factor of VOC detection data of a dimension with feature importance lower than a preset threshold being set to zero.

5. The method according to any one of claims 1 to 4, wherein the plurality of semiconductor gas sensors (111, 112, 113, 114) are made of same or different semiconductor materials, and the semiconductor gas sensors (111, 112, 113, 114) made of different semiconductor materials have different temperature features; and / or the category identification model (22) and / or the freshness identification model (231, 232, 233) are configured as a deep learning model.

6. A system (1) for training a model for evaluating object freshness, the system (1) being configured to perform the method according to any one of claims 1 to 5, wherein the system (1) comprises the following components: a plurality of semiconductor gas sensors (111, 112, 113, 114), configured to respectively obtain VOC detection data on an object of each category; a data labeling module (12), configured to respectively label VOC detection data, obtained by each semiconductor gas sensor (111, 112, 113, 114), on an object of each category with a label for representing a freshness state of an object of a corresponding category; and a model training module (13), configured to train, based at least on a labeled VOC detection dataset, a category identification model (22) for determining a category of an object and each freshness identification model (231, 232, 233) for determining a freshness state of an object of a corresponding category; or additionally, wherein the system (1) further comprises a data construction module (14), configured to merge VOC detection data obtained by one semiconductor gas sensor (111, 112, 113, 114) in each cyclic temperature modulation period into VOC detection data of a dimension assigned to the semiconductor gas sensor (111, 112, 113, 114), and assign a label to VOC detection data of a corresponding dimension based on alabeled label, to construct a multi-dimensional labeled VOC detection dataset.

7. A method for evaluating object freshness, the method comprising: step SI': determining, by a category identification model (22), a category of a to-be- identified object based on VOC detection data, obtained by semiconductor gas sensors (211, 212, 213, 214), on the to-be-identified object; and step S2': evaluating a freshness state of the object based on a VOC detection dataset on the to-be-identified object by freshness identification models (231, 232, 233) assigned to the determined category of the object.

8. The method according to claim 7, wherein the method further comprises: step Sil': merging VOC detection data obtained by one semiconductor gas sensor (211, 212, 213, 214) among a plurality of semiconductor gas sensors (211, 212, 213, 114) into VOC detection data of a dimension assigned to the semiconductor gas sensor (211, 212, 213, 214), to construct a multi-dimensional VOC detection dataset.

9. The method according to claim 8, wherein step Si l' comprises: step Sil l': performing cyclic temperature modulation in a preset temperature range on a plurality of semiconductor gas sensors (211, 212, 213, 214), and obtaining VOC detection data, on ato-be-identified object, of each semiconductor gas sensor (211, 212, 213, 214) by sampling, for example, using a sliding sampling window technique, in a cyclic temperature modulation period; and step S 112': merging VOC detection data obtained by one semiconductor gas sensor (211, 212, 213, 214) in each cyclic temperature modulation period into VOC detection data of a dimension assigned to the semiconductor gas sensor (211, 212, 213, 214), to construct a multidimensional VOC detection dataset.

10. The method according to claim 8 or 9, wherein the method further comprises: step SI 2': normalizing VOC detection data of one dimension based on a maximum VOC detection data value and a minimum VOC detection data value in the VOC detection data of the corresponding dimension; and / or wherein the method further comprises: step S13': performing feature importance analysis on VOC detection data of each dimension, and merging the VOC detection data of each dimension based on an analysis result of the feature importance, to reduce a dimension of the VOC detection dataset; or additionally to step S13, wherein based on the analysis result of the feature importance, a weight factor ofVOC detection data assigned to a dimension with low feature importance is assigned to the VOC detection data of the dimension with low feature importance, and a product of the VOC detection data of the dimension with low feature importance and the weight factor assigned to the dimension is merged into VOC detection data of a dimension with high feature importance, in particular, a weight factor of VOC detection data of a dimension with feature importance lower than a preset threshold being set to zero; and / or the plurality of semiconductor gas sensors (211, 212, 213, 214) are made of same or different semiconductor materials, and the semiconductor gas sensors (211, 212, 213, 214) made of different semiconductor materials have different temperature features; and / or the category identification model (22) and / or the freshness identification model (231, 232, 233) are configured as a deep learning model; and / or wherein the method further comprises: step S3': outputting notification information about a freshness state of an evaluated object.

11. A system (2) for evaluating object freshness, the system (2) being configured to perform the method according to any one of claims 7 to 10, wherein the system (2) comprises the following components: semiconductor gas sensors (211, 212, 213, 214), configured to respectively obtain VOC detection data on a to-be-identified object; a category identification model (22), configured to determine a category of a to-be- identified object based on VOC detection data, obtained by each semiconductor gas sensor (211, 212, 213, 214), on the to-be-identified object; and freshness identification models (231, 232, 233), configured to evaluate a freshness state of the object based on a VOC detection dataset on the to-be-identified object by the freshness identification models (231, 232, 233) assigned to the determined category of the object.

12. The system (2) according to claim 18, wherein the system (2) further comprises: a data construction module (24), configured to merge VOC detection data obtained by one semiconductor gas sensor (211, 212, 213, 214) into VOC detection data of a dimension assigned to the semiconductor gas sensor (211, 212, 213, 214), to construct a multi-dimensional VOC detection dataset; and / or a notification module (25), configured to output notification information about a freshness state of an evaluated object.

13. A computer program product, for example, a computer-readable program carrier,comprising or storing computer program instructions, wherein when the computer program instructions are executed by a processor, the steps of the methods according to any one of claims 1 to 7 and 10 to 17 are at least assisted in being implemented.

Images