An olfactory function evaluation device for human olfactory detection

Abstract

The present invention provides an olfactory function evaluation device for human olfactory detection, which belongs to the technical field of olfactory detection based on machine learning. First, representative odor substances covering different categories are collected, and odor samples of different concentrations are released through an odor release device. After air purification and environmental control to ensure the stability of the test environment, data is collected, and an olfactory function evaluation model based on multi-dimensional comprehensive analysis and judgment is designed. The reaction of the subjects after contact with the odor samples is subjected to information processing and feature extraction, and the hierarchical data processing method is used to further improve the mining depth of the feedback data to improve the evaluation accuracy. At the same time, during the actual training of the model, a model training strategy for adaptive parameter adjustment is designed. The present invention applies the optimized model to the actual olfactory function evaluation scenario, which can realize a comprehensive and accurate evaluation and analysis of the human olfactory function, and assist in judging the olfactory health status.

Description

Technical Field

[0001] The present invention belongs to the technical field of olfactory detection based on machine learning, and in particular relates to an olfactory function evaluation device for human olfactory detection. Background Art

[0002] The sense of smell is an important sensory function that plays a vital role in our daily lives. Not only does it help us identify the taste of food and enhance the eating experience, it also plays a role in social interactions, such as transmitting non-verbal information through body odor. In addition, the sense of smell can monitor harmful substances in the environment, and even in some cases, it can warn of potential health problems, such as certain diseases that may lead to loss of smell. However, environmental pollution, aging, and certain diseases can all cause damage to olfactory function. Therefore, accurately evaluating and monitoring olfactory function is very important for maintaining quality of life and detecting health problems in a timely manner.

[0003] Existing olfactory function assessment methods based on human olfactory detection mainly include the following two methods: chemical substance detection method and subjective questionnaire evaluation method.

[0004] Chemical substance detection method: This method first prepares a series of chemical odor substances with known concentrations and stable properties, and releases them into the test environment in a specific dose and manner. Subsequently, professional equipment is used to accurately monitor the subject's perception threshold of these chemical odors, that is, the minimum concentration at which the odor can be first perceived, and the recognition threshold, that is, the concentration of the specific odor that can be accurately identified. Furthermore, by utilizing the differences in the characteristics of the odors of different chemical substances, this method can analyze key functional indicators such as the subject's olfactory resolution and sensitivity.

[0005] Subjective questionnaire evaluation method: This method involves sending carefully designed olfactory-related questionnaires to the subjects, which cover questions such as preference for different odors, odor intensity, and ability to identify common odors. The subjects are required to answer based on their actual olfactory experience. Relevant researchers systematically analyze the answers to the questionnaires to determine the olfactory function of the subjects and possible abnormalities.

[0006] The subjective questionnaire assessment of olfactory function is easily affected by individual differences, resulting in inaccurate assessment and lack of quantitative indicators, which is not conducive to comparison and monitoring. The chemical substance detection method is costly, has limited odor types, and is difficult to simulate real complex odors, which affects the assessment of the ability to distinguish mixed odors. Summary of the invention

[0007] In response to the above problems, the present invention integrates multiple odor stimulation presentation methods and a comprehensive objective and intuitive feedback collection and analysis mechanism, and uses a machine learning model to complete a comprehensive and accurate assessment of human olfactory function.

[0008] The present invention provides an olfactory function evaluation device for human olfactory detection, the device comprising an olfactory function evaluation model that has been finally optimized, and an evaluation process based on the olfactory function evaluation model comprises the following steps:

[0009] S1, based on the selected odor samples and the test environment system, obtaining intuitive feedback data and objective feedback data of the subject, wherein the intuitive feedback data includes odor preference, odor familiarity, odor intensity perception and odor description; the objective feedback data includes the subject's reaction time and accuracy;

[0010] S2, preprocessing the intuitive feedback data and objective feedback data to obtain standard input data for training the olfactory function evaluation model;

[0011] S3, inputting the standard input data into the trained olfactory function assessment model, and outputting the olfactory health status of the subject, including unhealthy olfaction, generally healthy olfaction, and healthy olfaction;

[0012] The olfactory function evaluation model includes a data embedding module, a feature extraction module and a prediction output module;

[0013] The data embedding module uses text segmentation and word embedding to perform feature embedding on the intuitive feedback data, and performs feature calculation and concatenation on the objective feedback data to obtain intuitive embedding feature data and objective embedding feature data respectively;

[0014] The feature extraction module is used to extract key features from the intuitive embedded feature data and the objective embedded feature data to obtain intuitive data features and objective data features;

[0015] The prediction output module performs information interaction between intuitive data features and objective data features based on a cross-attention mechanism to obtain fused features, and predicts and outputs olfactory health status prediction results based on the fused features.

[0016] Preferably, the preprocessing of the intuitive feedback data and the objective feedback data in S2 includes data cleaning; for the intuitive feedback data, outliers are eliminated by removing irrelevant characters and correcting typos, and for the missing objective feedback data, linear interpolation and polynomial interpolation are used for interpolation.

[0017] Preferably, the specific processing process of the data embedding module is:

[0018] For the feature embedding of intuitive feedback data, we use the text segmentation method and word embedding method to embed the intuitive feedback data and obtain the embedded feature FI of the odor preference degree. x , odor familiarity embedding feature FI s, Odor intensity perception embedding feature FI q and the smell description embedding feature FI m , as follows:

[0019] The text segmentation is to divide the natural language text in the intuitive feedback into smaller units, and obtain a more concise vocabulary list through the steps of word segmentation and stop word removal; the word segmentation is to divide the text string into words or phrases; the stop word removal is to remove common words in the text, which appear frequently in the text but do not contain valid information;

[0020] The word embedding is to convert the segmented text into a numerical vector and use the bag-of-words model for word embedding:

[0021]

[0022] Where w is a word, N is the size of the vocabulary, and Count(w i ) is the word w i The number of times it appears in the text, one_hot(w i ) is the word w i One-hot encoded vector of ;

[0023] For objective feedback data feature embedding, the subject reaction time I is calculated respectively. f And the accuracy of the subjects in identifying the odor concentration I z The variance, maximum and minimum of the subject's reaction time is the vector composed of the variance, maximum and minimum of the subject's reaction time embedding feature FI f ; The vector consisting of the variance, maximum value and minimum value of the subject's recognition accuracy of odor concentration is the embedding feature FI of the subject's recognition accuracy of odor concentration z .

[0024] Preferably, the feature extraction module includes two channels, the input data of channel one is the odor preference embedding feature FI x , odor familiarity embedding feature FI s , Odor intensity perception embedding feature FI q and the smell description embedding feature FI m The vector composed of FZ = [FI x ;FI s ;FI q ;FI m], FZ first passes through three convolution layers with kernel sizes of 1*3, 1*4 and 1*5 to obtain multi-scale local features LFZ1, LFZ2, LFZ3, and then passes through a maximum pooling layer to obtain the parameter-lightweight pooling feature PFZ, and then is normalized by the Softmax activation function to obtain the output of channel one, that is, the intuitive feedback data deep feature F z ;

[0025] The input data of channel 2 is the subject's reaction time embedding feature FI f And the accuracy of the subject's identification of odor concentration embedded feature FI z The vector composed of FK = [FI f ;FI z ], FK first passes through 10 residual convolution modules to obtain the residual feature RFK. The residual convolution module consists of two 1*3 convolution layers and one 1*2 convolution layer, which have two residual connections. Then, after passing through an average pooling layer, the parameter-lightweight pooling feature PFK is obtained. Then, after normalization by the Softmax activation function, the output of the second channel is obtained, that is, the objective data feature F k .

[0026] Preferably, the prediction output module includes a cross attention module, a fully connected layer and a Relu activation function;

[0027] The cross-attention module is based on the intuitive data feature F z and objective master data characteristics F k , to interact with information and obtain intuitive data fusion features and objective data fusion features Then the fine-grained fusion feature F is obtained by weighted summation. c ;

[0028] The fine-grained fusion feature F c Input the fully connected layer and the Relu activation function layer in sequence to obtain the local fusion feature F 1 ;

[0029] The local fusion feature F 1 Input the fully connected layer and the Relu activation function layer in sequence to obtain the global fusion feature F 2 ;

[0030] The deep fusion feature F 2 The deep fusion features obtained after inputting into the fully connected layer and the Relu activation function layer in sequence are passed through the fully connected layer to obtain the prediction result of the olfactory health status.

[0031] Preferably, the cross attention module determines the weighted combination of information in the value Value by calculating the correlation between the query Query and the key Key to achieve deep fusion between features. The formula is:

[0032]

[0033] The weight matrix W Q , W K , W V The degree of adaptive learning information interaction, is the dimension of the key vector, which is used to scale the dot product to prevent the value from being too large during gradient descent; the attention scores are normalized by the softmax function to ensure that the sum of all scores is 1:

[0034]

[0035] x i is an element in the attention score matrix.

[0036] Preferably, supervised training is performed on the olfactory function evaluation model, wherein when the data set is annotated, the intuitive feedback data includes four dimensions, namely, the degree of odor preference I x 、Odor familiarity I s 、Odor intensity perception I q and odor description I m ; The degree of smell preference is divided into five levels: very dislike, dislike, average, like and very like; the degree of smell familiarity is divided into three levels: familiar, moderately familiar and unfamiliar; the smell intensity is divided into three levels: strong, medium and weak; the smell description is divided into three categories: floral, fruity and woody; the objective feedback data includes two dimensions, namely the time the subject responds to the appearance of the smell I f and the accuracy of the subjects in identifying the odor concentration I z Experts classified the subjects' olfactory health status y according to the subjects' intuitive feedback data and objective feedback data, dividing them into three categories: unhealthy olfaction, generally healthy olfaction, and healthy olfaction, represented by labels 0, 1, and 2 respectively.

[0037] Preferably, when training the olfactory function evaluation model, a hybrid loss function is constructed, including cross entropy loss and binary cross entropy loss, wherein the cross entropy loss is used to improve the convergence speed of the model, and the binary cross entropy loss is responsible for improving the accuracy of the model's prediction of the subject's olfactory health status; the cross entropy loss function is specifically as follows:

[0038]

[0039] in is the true label distribution, which is a one-hot encoding vector, that is, for the correct olfactory health status category i, For other categories y is the probability distribution predicted by the model, where y i is the probability of the model predicting category i; the binary cross entropy loss is as follows:

[0040]

[0041] in is the true label, with a value of 0 or 1, 0 represents the negative class and 1 represents the positive class; y is the probability of the model predicting the correct olfactory health status category;

[0042] The model is trained using a hybrid training strategy, and the final model optimization loss is as follows:

[0043] L=λL 1 +μL 2

[0044] Where L 1 is the cross entropy loss, L 2 is the binary cross entropy loss, λ is preset to 1, and μ is preset to 2.

[0045] Preferably, a model training strategy based on adaptive optimization is used during model training, and the weights of each part of the loss function are dynamically adjusted according to the performance of the model during the training process, as follows:

[0046] First, define an adaptive loss function L(θ), which consists of multiple sub-loss functions, each of which corresponds to a different evaluation indicator of the model. The function is expressed as:

[0047]

[0048] Among them, L k (θ) is the kth sub-loss function, w k is the corresponding weight, θ is the parameter of the model;

[0049] The adaptive weight adjustment method is as follows:

[0050] A. Initialization weights: At the beginning of training, an initial weight is assigned to each sub-loss function Artificial settings based on prior knowledge;

[0051] B. Calculate loss gradient: In each iteration, calculate the gradient of each sub-loss function

[0052] C. Update weights: Dynamically adjust weights w based on the performance of the model in the current iteration k , the weight update formula is:

[0053]

[0054] in, is the weight of the kth sub-loss function at the tth iteration, α k is the learning rate, which is used to control the speed of weight adjustment. The denominator is the sum of the squares of the gradients of all sub-loss functions, which is used to normalize the weight update;

[0055] D. Apply weights: Use the updated weights Calculate the new total loss L(θ) and update the model parameters θ accordingly;

[0056] Repeat steps A to D until the model converges or reaches a predetermined number of training rounds.

[0057] Compared with the prior art, the present invention has the following beneficial effects:

[0058] (1) Improving the objectivity and accuracy of the evaluation: This method can improve the objectivity and accuracy of the evaluation of human olfactory function by constructing an olfactory function evaluation model based on multi-dimensional feedback collection and comprehensive analysis. The evaluation model uses a combination of attention mechanism and residual module to enhance the model's ability to capture key features and improve the model's ability to capture rich information. This allows for a more accurate judgment of the subject's olfactory function status when the model is actually applied to the evaluation scenario, ultimately further improving the scientificity and accuracy of the entire evaluation process.

[0059] (2) Improve the applicability of the model: When conducting olfactory function assessment, a model training strategy based on adaptive parameter adjustment was constructed. During inference, the model parameters were dynamically adjusted according to different test populations, environments and other factors. This can enhance the model's adaptability to a variety of complex situations and improve the model's applicability. For a variety of test groups of different ages, health conditions, etc., the model can adapt well when they are undergoing olfactory function assessment, which is conducive to expanding the scope of application of this assessment method.

[0060] (3) Improve training efficiency: Hybrid training allows the model to learn multiple tasks simultaneously during a single training process, which can significantly improve training efficiency. In the training of the olfactory function evaluation model, this method can reduce the time and resources required for training and speed up the model development cycle. At the same time, since the model can learn from multiple perspectives, it also helps to discover hidden patterns and associations in the data, further improving the model's predictive ability.

[0061] (4) Improve the comprehensiveness of the evaluation results: Compared with relying solely on a certain type of feedback data for evaluation, the present invention first adds a cross-attention mechanism to the feedback data of different dimensions (such as intuitive description, objective reaction time, etc.) when constructing the feature extraction unit during the feature extraction operation, which can better grasp the intrinsic connection between different feedback data. Through such an operation, when evaluating the human olfactory function, the ability of comprehensive detection can be further improved due to the adoption of data segmentation and comprehensive weighing. The final olfactory function evaluation results will be more comprehensive, and reasonable evaluations can be given to the complex olfactory performances of different individuals. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 It is a flow chart of the overall technical route of the present invention.

[0063] Figure 2 This is a structural diagram of the data embedding module of the present invention.

[0064] Figure 3 This is a structural diagram of the feature extraction module of the present invention.

[0065] Figure 4 This is a structural diagram of the prediction output module of the present invention.

[0066] Figure 5 This is a structural diagram of the cross-attention module of the present invention.

[0067] Figure 6 This is a diagram of the experimental results of the present invention. DETAILED DESCRIPTION

[0068] The present invention provides an olfactory function evaluation method for human olfactory detection, which completes a comprehensive and accurate evaluation of human olfactory function by integrating multiple odor stimulation presentation methods and a comprehensive objective and intuitive feedback collection and analysis mechanism. Figure 1 shown.

[0069] Odor samples and environmental preparation: First, carefully select a variety of representative odor substances covering different categories (such as natural fragrances, daily common odors, etc.), prepare various types and concentrations of odor samples according to strict ratio requirements, and properly place them in a special odor release device. Then, turn on the air purification system to remove possible odors and impurities in the test space. At the same time, use environmental control equipment to accurately adjust the temperature to a suitable range, maintain humidity within a reasonable range, and keep the air pressure stable, to ensure the stability of the test environment in all aspects, and create good and standard conditions for subsequent olfactory function evaluation.

[0070] Olfactory feedback collection: The olfactory feedback collection module is designed based on multi-dimensional feedback channels, including an intuitive feedback section, which collects the subject's intuitive feelings about the smell through questionnaires, verbal inquiries, etc., such as the familiarity and preference of the smell, and the perceived smell intensity. At the same time, there is an objective feedback section, which uses response buttons and electronic recording equipment to collect objective data such as the time the subject responds to the smell and the concentration of the smell. This ensures that the information related to the subject's olfactory response can be stably and comprehensively collected in the face of different individuals and different test situations, thereby enhancing the reliability of the entire evaluation system.

[0071] Data preprocessing: Data preprocessing includes data cleaning and data labeling. Data cleaning can remove errors, anomalies, and redundant information in the data, ensure the consistency and accuracy of the data, thereby improving data quality and providing a reliable foundation for model training; while data labeling provides clear training goals for machine learning models by giving specific labels or classifications to the data, which helps to improve the recognition and prediction capabilities of the model. The role and benefit of both is that they jointly ensure the efficiency and availability of input data, thereby enhancing the training effect and generalization ability of the model.

[0072] Olfactory function evaluation model: Design an olfactory function evaluation model based on a neural network, including a data embedding module, a feature extraction module, and a prediction output module. The data embedding module processes the collected data into a data type that the model can recognize. The feature extraction module deeply mines the information contained in the data and extracts key features. The prediction output module predicts the output results of the extracted features, and further iterates and updates the model through loss optimization.

[0073] Model training strategy with adaptive parameter adjustment: In the actual training process of the model, in order to further improve the evaluation accuracy of the model, the present invention constructs a hybrid training strategy and a model training strategy with adaptive parameter adjustment. Hybrid training allows the model to learn multiple tasks in one iteration, reducing the time and resources required for training and accelerating the model development cycle. The model training strategy with adaptive parameter adjustment will dynamically adjust the algorithm weights in the model according to factors such as the characteristics of different sample data input and the characteristics of different subject groups, and ultimately obtain accurate olfactory function evaluation results. Therefore, when the model is actually applied for evaluation, more accurate calculations can be achieved using the optimized model, which can ensure more accurate evaluation results of human olfactory function and further improve the scientificity and effectiveness of the entire evaluation process.

[0074] Practical application and evaluation: When it is necessary to conduct an actual evaluation of human olfactory function, the subjects are guided into the prepared standard test environment, and the odor samples are released in sequence according to the set test process. The various feedback information made by the subjects on the odor is collected in real time and sent to the optimized olfactory function evaluation model, so as to conduct a comprehensive analysis and accurate judgment of the subjects' olfactory function in real time, providing a reliable basis for subsequent health diagnosis, related research, etc.

[0075] The invention will be further described below in conjunction with specific embodiments.

[0076] 1. Dataset Construction

[0077] 1. Odor sample selection and environment preparation

[0078] Selection and preparation of odor samples: The preparation of odor samples must be carefully designed to ensure the reliability and validity of the evaluation results. First of all, when selecting odor samples, the present invention needs to consider their representativeness, stability and safety. In order to cover different categories of odors, representative odor substances should be selected, such as natural fragrances, common daily odors, and specific odors. At the same time, odor substances with high stability are selected to ensure that they maintain consistency and reliability during the test process to avoid affecting the evaluation results due to changes in odor. In addition, safety is also an important consideration in selecting odor samples to ensure that they will not cause harm to the human body.

[0079] The preparation of odor samples needs to be processed accordingly according to their physical state. For liquid odors, solvents are used for dilution, and the dilution ratio is accurately controlled by flow meters, burettes and other equipment to ensure the accuracy of the odor concentration. For solid odors, the odor components can be extracted by grinding, soaking and other methods, and further purified by distillation, filtration and other technologies. During the preparation process, the operation process needs to be strictly controlled to avoid cross-contamination and odor emission, and to ensure the purity and stability of the odor sample.

[0080] After the odor sample is prepared, strict quality control is required. First, component analysis is performed to confirm whether the odor sample contains the expected components and exclude impurities. Secondly, concentration testing is performed to ensure that the concentration of the odor sample meets the test requirements, and stability testing is performed to simulate the actual test environment and observe the changes of the odor sample under specific conditions to ensure its consistency and reliability during the test process.

[0081] Test environment preparation: According to the type of odor and release requirements, select devices such as sprayers, diffusers, heaters, and heating plates. These devices should have the ability to accurately control the concentration, speed, and time of odor release, and be able to operate stably to avoid affecting the test results due to equipment failure. Air purifiers, activated carbon, and other methods are needed to remove odors and impurities that may exist in the test space to ensure fresh air. At the same time, air conditioners, humidifiers, dehumidifiers, and other equipment can be used to adjust the temperature, humidity, and air pressure of the test space to keep it within a suitable range to avoid the impact of environmental factors on the evaluation results. In addition, factors such as sound insulation and lighting in the test space are also considered to provide a comfortable and quiet environment for the subjects.

[0082] 2. Collection of olfactory data feedback

[0083] Intuitive feedback collection: Intuitive feedback collection is an important part of the olfactory function evaluation system. It provides important reference information for the evaluation model by collecting the subject's feelings and experiences of odors. In order to ensure the effectiveness and reliability of intuitive feedback, we follow the principles of clear goals, clear questions, diverse question types, and a moderate number of questions, and cover odor familiarity, odor preference, odor intensity perception, and odor description information.

[0084] Intuitive feedback data collection includes but is not limited to the following aspects: asking subjects about their familiarity with specific odors, including identification of the odor, previous exposure experience and the time interval between exposures; assessing subjects' preference for odors, involving grading of preference levels and analysis of reasons for preference; recording subjects' perception of odor intensity, involving the division of intensity levels and discussion of changes in intensity perception; and collecting subjects' detailed descriptions of odors, including odor types, attributes, characteristics and comparisons with other odors.

[0085] To ensure the accuracy and reliability of intuitive feedback, the present invention takes the following measures: providing a detailed filling guide to ensure that the subjects have a full understanding of the questionnaire filling process; using standardized language to avoid expressions that may lead to misunderstandings; designing non-leading questions to ensure that the subjects respond based on their true feelings; and implementing strict data confidentiality measures to protect the personal information security of the subjects.

[0086] Objective feedback collection: Objective feedback collection is a key component of the olfactory function evaluation system. Its purpose is to supplement and verify the results of intuitive feedback through quantitative indicators. The present invention proposes an objective feedback collection method, which ensures the accuracy and consistency of the collected data through high-precision instruments and standardized operating procedures.

[0087] Specifically, the objective feedback collection of the present invention includes the following aspects: first, selecting appropriate physiological and behavioral measurement tools, such as electroencephalogram (EEG), skin electrode activity (SCL), heart rate variability (HRV), and reaction time measurement devices, to capture the physiological and behavioral responses of the subjects when exposed to odors. Second, all measurement equipment is accurately calibrated to ensure that it maintains optimal performance during the measurement process.

[0088] During the data collection process, the present invention adopts the following measures: formulate a standardized test process to ensure that each subject is evaluated under the same conditions; control the test environment, including temperature, humidity, light and noise level, to reduce the impact of environmental factors on the test results; and record other relevant behavioral indicators such as the subject's reaction time and accuracy to quantify their olfactory response.

[0089] In summary, the present invention provides high-quality data input for the olfactory function evaluation system through careful questionnaire design, appropriate data collection methods and strict data processing procedures, thereby improving the performance and robustness of the evaluation model and achieving accurate evaluation of olfactory function. The objective feedback collection method of the present invention provides reliable quantitative indicators for olfactory function evaluation through precise instrument measurement and standardized operating procedures, thereby enhancing the overall performance and practicality of the evaluation system.

[0090] 2. Data Preprocessing

[0091] Data preprocessing is an important step in building a deep learning model and is crucial to improving the performance and robustness of the model. Through data cleaning, noise and outliers in the data can be removed to improve data quality. Through data annotation, raw data can be converted into a format that can be understood by model evaluation.

[0092] Data cleaning: Clean the collected intuitive and objective feedback data, remove missing values, outliers, etc., to ensure data quality. For intuitive feedback data, remove irrelevant characters and correct typos to eliminate outliers. For missing objective feedback data, use linear interpolation and polynomial interpolation for interpolation.

[0093] Data annotation: In the olfactory function evaluation model, the main contents of data annotation are as follows. Intuitive feedback data includes four dimensions: odor preference I x 、Odor familiarity I s 、Odor intensity perception I q and odor description I m; The odor preference is divided into five levels: very dislike, dislike, average, like and very like; the odor familiarity is divided into three levels: familiar, moderately familiar and unfamiliar; the odor intensity is divided into three levels: strong, medium and weak. The odor description is divided into three categories: floral, fruity and woody. The objective feedback data includes two dimensions: the time the subject responds to the appearance of the odor. f and the accuracy of the subjects in identifying the odor concentration I z Experts classified the subjects' olfactory health status y according to the subjects' intuitive feedback data and objective feedback data, dividing them into three categories: unhealthy olfaction, generally healthy olfaction, and healthy olfaction, represented by labels 0, 1, and 2 respectively.

[0094] 3. Construction of Olfactory Function Evaluation Model

[0095] The olfactory function evaluation model includes a data embedding module, a feature extraction module, and a prediction output module;

[0096] The data embedding module uses text segmentation and word embedding to perform feature embedding for the intuitive feedback data, and performs feature calculation and splicing for the objective feedback data to obtain intuitive embedded feature data and objective embedded feature data respectively; the feature extraction module is used to extract key features of the intuitive embedded feature data and the objective embedded feature data to obtain intuitive data features and objective data features; the prediction output module performs information interaction between intuitive data features and objective data features based on the cross-attention mechanism to obtain fused features, and predicts and outputs the olfactory health status prediction results based on the fused features.

[0097] 1. Data Embedding Module

[0098] First, a data embedding module is constructed to embed the intuitive feedback data and objective feedback data to obtain the intuitive feedback data embedding features and objective feedback data embedding features. The specific structure is as follows: Figure 2 shown.

[0099] Intuitive feedback data feature embedding, using text segmentation method and word embedding method to embed the intuitive feedback data and obtain the odor preference embedding feature FI x , odor familiarity embedding feature FI s , Odor intensity perception embedding feature FI q and the smell description embedding feature FI m , as follows:

[0100] Text segmentation is the process of breaking down the natural language text in intuitive feedback into smaller units. Through the steps of tokenization and stop word removal, a more concise vocabulary list is obtained. Tokenization: Split the text string into words or phrases. Stop word removal: Remove common words in the text, which appear frequently in the text but do not contain valid information.

[0101] Word embedding is the process of converting segmented text into a numerical vector. The present invention adopts the bag-of-words model for word embedding.

[0102]

[0103] Where w is a word, N is the size of the vocabulary, and Count(w i ) is the word w i The number of times it appears in the text, one_hot(w i ) is the word w i The one-hot encoded vector of .

[0104] Objective feedback data feature embedding, and calculating the subject's reaction time I f And the accuracy of the subjects in identifying the odor concentration I z The variance, maximum and minimum of the subject's reaction time is the vector composed of the variance, maximum and minimum of the subject's reaction time embedding feature FI f ; The vector consisting of the variance, maximum value and minimum value of the subject's recognition accuracy of odor concentration is the embedding feature FI of the subject's recognition accuracy of odor concentration z .

[0105] Through these steps, the original intuition and objective feedback data are converted into embedded feature vectors that the model can understand and learn, providing a solid foundation for subsequent model training.

[0106] 2. Feature Extraction Module

[0107] Construct a feature extraction module to perform deep feature extraction on the intuitive feedback data embedding features and the objective feedback data embedding features to obtain the intuitive feedback data deep features and the objective feedback data deep features; the constructed feature extraction module contains two paths, the specific structure is as follows Figure 3 shown.

[0108] It includes two channels, the input data of channel one is the odor preference embedding feature FI x , odor familiarity embedding feature FI s , Odor intensity perception embedding feature FI q and the smell description embedding feature FI m The vector composed of FZ = [FI x ;FI s ;FIq ;FI m ], FZ first passes through three convolution layers with kernel sizes of 1*3, 1*4 and 1*5 to obtain multi-scale local features LFZ1, LFZ2, LFZ3, and then passes through a maximum pooling layer to obtain the parameter-lightweight pooling feature PFZ, and then is normalized by the Softmax activation function to obtain the output of channel one, that is, the intuitive feedback data deep feature F z .

[0109] The input data of channel 2 is the subject's reaction time embedding feature FI f And the accuracy of the subject's identification of odor concentration embedded feature FI z The vector composed of FK = [FI f ;FI z ], FK first passes through 10 residual convolution modules to obtain the residual feature RFK. The residual convolution module consists of two 1*3 convolution layers and one 1*2 convolution layer, which has two residual connections, which helps the model learn more complex and richer feature maps. Then, after an average pooling layer, the parameter-lightweight pooling feature PFK is obtained, and then the output of the second channel is obtained after normalization by the Softmax activation function, that is, the objective data feature F k .

[0110] 3. Prediction output module

[0111] The structure diagram of the prediction output module is as follows: Figure 4 As shown, it mainly includes a cross attention module, a fully connected layer and a Relu activation function.

[0112] First, F z and F k Input cross attention mechanism to obtain fine-grained fusion feature F c , using the cross-attention mechanism to interact with intuitive features and objective information. Through cross-attention, the model can interact with information at a fine-grained level, that is, for each element of the input sequence, the model can dynamically adjust its attention weight according to the current context, thereby achieving accurate information matching and fusion. The structure of the cross-attention mechanism is as follows Figure 5 shown.

[0113] The cross-attention mechanism promotes information interaction between different data features. The core of this mechanism is to determine how to weight the information in the value by calculating the correlation between the query and the key, thereby achieving deep fusion between features. z and objective master data characteristics F kA cross-attention module is constructed to effectively promote the information interaction between the two data features, enabling the model to better understand and utilize the complex relationships in the data. The formula is as follows:

[0114]

[0115] The weight matrix W Q , W K , W V The degree of information interaction that can be learned adaptively. is the dimension of the key vector, which is used to scale the dot product to prevent the value from being too large during gradient descent. The attention scores are normalized by the softmax function to ensure that the sum of all scores is 1:

[0116]

[0117] Here x i is an element in the attention score matrix. The intuitive data fusion feature can be obtained through the cross attention module and objective data fusion features Then the fine-grained fusion feature F is obtained by weighted summation. c .

[0118] Then the fine-grained fusion feature F c Input the fully connected layer and the Relu activation function layer in sequence to obtain the local fusion feature F 1 ;

[0119] Then the local fusion feature F 1 Input the fully connected layer and the Relu activation function layer in sequence to obtain the global fusion feature F 2 ;

[0120] Finally, the deep fusion feature F 2 The deep fusion features obtained after inputting into the fully connected layer and the Relu activation function layer in sequence are passed through the fully connected layer to obtain the prediction result of the olfactory health status.

[0121] 4. Model Training

[0122] 1. Mixed loss strategy

[0123] In order to constrain model training, the present invention constructs a hybrid loss function, including cross entropy loss and binary cross entropy loss, where cross entropy loss is used to improve the convergence speed of the model, and binary cross entropy loss is responsible for improving the accuracy of the model's prediction of the subject's olfactory health status. The cross entropy loss function is as follows:

[0124]

[0125] in is the true label distribution, usually a one-hot encoded vector, i.e., for the correct olfactory health status category i, For other categories y is the probability distribution predicted by the model, where y i is the probability that the model predicts category i. The binary cross entropy loss is as follows:

[0126]

[0127] in is the true label, which takes a value of 0 or 1 (0 represents the negative class and 1 represents the positive class). y is the probability that the model predicts the correct olfactory health status category. The model can be effectively trained through the cross entropy loss and binary cross entropy loss functions, so that the model has the function of correct olfactory function evaluation. In order to further improve the training efficiency more effectively and reduce the time and resources required for training, the present invention adopts a hybrid training strategy to train the model. The loss of the final model optimization is as follows:

[0128] L=λL 1 +μL 2

[0129] Where L 1 is the cross entropy loss, L 2 is the binary cross entropy loss, λ is preset to 1, and μ is preset to 2.

[0130] 2. Dynamic weight adjustment

[0131] This paper proposes a model training strategy based on adaptive optimization to enhance the performance of the olfactory function evaluation model. The strategy can dynamically adjust the weights of each part of the loss function according to the performance of the model during the training process. The following is the detailed process of the adaptive loss function weight method:

[0132] First, define an adaptive loss function L(θ), which consists of multiple sub-loss functions, each of which corresponds to a different evaluation indicator of the model, such as accuracy, reaction time, etc. The function can be expressed as:

[0133]

[0134] Among them, L k (θ) is the kth sub-loss function, w k is the corresponding weight and θ is the parameter of the model.

[0135] The adaptive weight adjustment method is as follows:

[0136] A. Initialization weights: At the beginning of training, an initial weight is assigned to each sub-loss function These weights are manually set based on prior knowledge.

[0137] B. Calculate loss gradient: In each iteration, calculate the gradient of each sub-loss function

[0138] C. Update weights: Dynamically adjust weights w based on the performance of the model in the current iteration k The weight update formula is:

[0139]

[0140] in, is the weight of the kth sub-loss function at the tth iteration, α k is a learning rate that controls the speed of weight adjustment, and the denominator is the sum of the squares of the gradients of all sub-loss functions, which is used to normalize the weight update.

[0141] D. Apply weights: Use the updated weights Calculate the new total loss L(θ) and update the model parameters θ accordingly.

[0142] Iterative process: Repeat steps A to D until the model converges or reaches a predetermined number of training rounds.

[0143] In this way, the model can automatically adjust the focus on different evaluation indicators, so as to better balance performance indicators such as accuracy and reaction time during training and improve the comprehensive evaluation ability of the model. In addition, this adaptive method can also help the model better adapt to changes in the data and improve its generalization ability under different conditions.

[0144] 5. Model Deployment and Evaluation

[0145] Based on the designed olfactory function evaluation model and combined with the described adaptive optimization model training strategy, an optimized olfactory function evaluation model is finally obtained. The following are the specific implementation steps for deploying this optimized model into practical applications.

[0146] Data collection: First, deploy highly sensitive olfactory sensors in the olfactory function evaluation system to collect real-time data on the subjects' responses to various odors. Collect intuitive feedback from testers through questionnaires or interactive software, including odor preference, odor familiarity, odor intensity perception, and odor description. Use professional instruments to test the subjects' response time and accuracy to odors.

[0147] Data preprocessing: The collected raw data is preprocessed, mainly through data cleaning, to enhance the data quality and improve the accuracy of the model's assessment of olfactory function.

[0148] Olfactory function evaluation model deployment: The optimized olfactory function evaluation model is deployed on professional olfactory detection equipment or mobile devices. The model inputs the preprocessed data into the model to evaluate the olfactory function and obtain the olfactory function status of the subject, including three categories: unhealthy olfactory function, generally healthy olfactory function, and healthy olfactory function, which are represented by labels 0, 1, and 2 output respectively.

[0149] Model function evaluation: The present invention uses the following indicators to quantify the model performance: Odor recognition ability: the percentage of correctly recognized odors in the total number of tests. Odor differentiation ability: the percentage of correctly distinguishing two different odors in the total number of tests. Olfactory function total score: the total score combined with multiple test results, used to comprehensively evaluate the olfactory function. The performance comparison results are shown in Table 1.

[0150] Table 1 Performance comparison results

[0151]

[0152] At the same time, in order to verify the effectiveness of the method of the present invention, the present invention predicts the olfactory health status of 20 subjects. Figure 6 As shown, sample numbers 1-20 represent 20 subjects, blue dots represent the true values ​​of the subjects' olfactory health status, and orange dots represent the predicted values ​​of the subjects' olfactory health status. If the predicted value of a subject's olfactory health status is consistent with the true value of the olfactory health status, the blue dots and orange dots overlap in the figure, indicating that the prediction is correct. It can be seen from the figure that the number of samples predicted incorrectly by the method of the present invention is 3, indicating that the method of the present invention has a high prediction accuracy.

[0153] Through the above steps, the optimization and deployment of the olfactory function assessment model was achieved, and an olfactory health monitoring tool was provided to users, which helps to detect and intervene in olfactory-related health problems at an early stage.

[0154] The above description is only the preferred embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

[0155] Although the above describes the specific implementation methods of the present invention, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. An olfactory function evaluation device for human olfactory detection, characterized in that: The device includes a finally optimized olfactory function evaluation model, and the evaluation process based on the olfactory function evaluation model includes the following steps: S1, based on the selected odor samples and the test environment system, obtaining intuitive feedback data and objective feedback data of the subject, wherein the intuitive feedback data includes odor preference, odor familiarity, odor intensity perception and odor description; the objective feedback data includes the subject's reaction time and accuracy; S2, preprocessing the intuitive feedback data and objective feedback data to obtain standard input data for training the olfactory function evaluation model; S3, inputting the standard input data into the trained olfactory function assessment model, and outputting the olfactory health status of the subject, including unhealthy olfaction, generally healthy olfaction, and healthy olfaction; The olfactory function evaluation model includes a data embedding module, a feature extraction module and a prediction output module; The data embedding module uses text segmentation and word embedding to perform feature embedding on the intuitive feedback data, and performs feature calculation and concatenation on the objective feedback data to obtain intuitive embedding feature data and objective embedding feature data respectively; The feature extraction module is used to extract key features from the intuitive embedded feature data and the objective embedded feature data to obtain intuitive data features and objective data features; The prediction output module performs information interaction between intuitive data features and objective data features based on a cross-attention mechanism to obtain fused features, and predicts and outputs olfactory health status prediction results based on the fused features.

2. The olfactory function evaluation device for human olfactory detection according to claim 1, characterized in that: The preprocessing of the intuitive feedback data and the objective feedback data in S2 includes data cleaning; for the intuitive feedback data, outliers are eliminated by removing irrelevant characters and correcting typos, and for the missing objective feedback data, linear interpolation and polynomial interpolation are used for interpolation.

3. The olfactory function evaluation device for human olfactory detection according to claim 1, characterized in that: The specific processing process of the data embedding module is as follows: For the feature embedding of intuitive feedback data, we use the text segmentation method and word embedding method to embed the intuitive feedback data and obtain the embedded feature FI of the odor preference degree. x , odor familiarity embedding feature FI s , Odor intensity perception embedding feature FI q and the smell description embedding feature FI m , as follows: The text segmentation is to divide the natural language text in the intuitive feedback into smaller units, and obtain a more concise vocabulary list through the steps of word segmentation and stop word removal; the word segmentation is to divide the text string into words or phrases; the stop word removal is to remove common words in the text, which appear frequently in the text but do not contain valid information; The word embedding is to convert the segmented text into a numerical vector and use the bag-of-words model for word embedding: Where w is a word, N is the size of the vocabulary, and Count(w i ) is the word w i The number of times it appears in the text, one_hot(w i ) is the word w i One-hot encoded vector of ; For objective feedback data feature embedding, the subject reaction time I is calculated respectively. f And the accuracy of the subjects in identifying the odor concentration I z The variance, maximum and minimum of the subject's reaction time is the vector composed of the variance, maximum and minimum of the subject's reaction time embedding feature FI f ; The vector consisting of the variance, maximum value and minimum value of the subject's recognition accuracy of odor concentration is the embedding feature FI of the subject's recognition accuracy of odor concentration z .

4. The olfactory function evaluation device for human olfactory detection according to claim 1, characterized in that: The feature extraction module includes two channels. The input data of channel one is the odor preference embedding feature FI x , odor familiarity embedding feature FI s , Odor intensity perception embedding feature FI q and the smell description embedding feature FI m The vector FZ = [FI x ;FI s ;FI q ;FI m ], FZ first passes through three convolution layers with kernel sizes of 1*3, 1*4 and 1*5 to obtain multi-scale local features LFZ1, LFZ2, LFZ3, and then passes through a maximum pooling layer to obtain the parameter-lightweight pooling feature PFZ, and then is normalized by the Softmax activation function to obtain the output of channel one, that is, the intuitive feedback data deep feature F z ; The input data of channel 2 is the subject's reaction time embedding feature FI f And the accuracy of the subject's identification of odor concentration embedded feature FI z The vector composed of FK = [FI f ;FI z ], FK first passes through 10 residual convolution modules to obtain the residual feature RFK. The residual convolution module consists of two 1*3 convolution layers and one 1*2 convolution layer, which have two residual connections. Then, after passing through an average pooling layer, the parameter-lightweight pooling feature PFK is obtained. Then, after normalization by the Softmax activation function, the output of the second channel is obtained, that is, the objective data feature F k .

5. An olfactory function evaluation device for human olfactory detection as claimed in claim 4, characterized in that: The prediction output module includes a cross attention module, a fully connected layer and a Relu activation function; The cross-attention module is based on the intuitive data feature F z and objective master data characteristics F k , to interact with information and obtain intuitive data fusion features and objective data fusion features Then the fine-grained fusion feature F is obtained by weighted summation. c ; The fine-grained fusion feature F c Input the fully connected layer and the Relu activation function layer in sequence to obtain the local fusion feature F1; The local fusion feature F1 is sequentially input into the fully connected layer and the Relu activation function layer to obtain the global fusion feature F2; The deep fusion feature F2 is input into the fully connected layer and the Relu activation function layer in sequence, and then the deep fusion feature obtained passes through the fully connected layer to obtain the olfactory health status prediction result.

6. The olfactory function evaluation device for human olfactory detection according to claim 5, characterized in that: The cross attention module calculates the correlation between the query Query and the key Key to decide the weighted combination of the information in the value Value to achieve deep fusion between features. The formula is: The weight matrix W Q , W K , W V The degree of adaptive learning information interaction, is the dimension of the key vector, which is used to scale the dot product to prevent the value from being too large during gradient descent; the attention scores are normalized by the softmax function to ensure that the sum of all scores is 1: x i is an element in the attention score matrix.

7. The olfactory function evaluation device for human olfactory detection according to claim 1, characterized in that: The olfactory function evaluation model is trained in a supervised manner. When the data set is annotated, the intuitive feedback data includes four dimensions: odor preference, I x 、Odor familiarity I s 、Odor intensity perception I q and odor description I m ; The degree of smell preference is divided into five levels: very dislike, dislike, average, like and very like; the degree of smell familiarity is divided into three levels: familiar, moderately familiar and unfamiliar; the smell intensity is divided into three levels: strong, medium and weak; the smell description is divided into three categories: floral, fruity and woody; the objective feedback data includes two dimensions, namely the time the subject responds to the appearance of the smell I f and the accuracy of the subjects in identifying the odor concentration I z Experts classified the subjects' olfactory health status y according to the subjects' intuitive feedback data and objective feedback data, dividing them into three categories: unhealthy olfaction, generally healthy olfaction, and healthy olfaction, represented by labels 0, 1, and 2 respectively.

8. The olfactory function evaluation device for human olfactory detection according to claim 1, characterized in that: When training the olfactory function evaluation model, a hybrid loss function was constructed, including cross entropy loss and binary cross entropy loss. The cross entropy loss is used to improve the convergence speed of the model, and the binary cross entropy loss is responsible for improving the accuracy of the model's prediction of the subject's olfactory health status. The cross entropy loss function is as follows: in is the true label distribution, which is a one-hot encoding vector, that is, for the correct olfactory health status category i, For other categories y is the probability distribution predicted by the model, where y i is the probability of the model predicting category i; the binary cross entropy loss is as follows: in is the true label, with a value of 0 or 1, 0 represents the negative class and 1 represents the positive class; y is the probability of the model predicting the correct olfactory health status category; The model is trained using a hybrid training strategy, and the final model optimization loss is as follows: L=λL1+μL2 Among them, L1 is the cross entropy loss, L2 is the binary cross entropy loss, λ is preset to 1, and μ is preset to 2.

9. The olfactory function evaluation device for human olfactory detection according to claim 8, characterized in that: When training the model, a model training strategy based on adaptive optimization is used to dynamically adjust the weights of each part of the loss function according to the performance of the model during the training process, as follows: First, define an adaptive loss function L(θ), which consists of multiple sub-loss functions, each of which corresponds to a different evaluation indicator of the model. The function is expressed as: Among them, L k (θ) is the kth sub-loss function, w k is the corresponding weight, θ is the parameter of the model; The adaptive weight adjustment method is as follows: A. Initialization weights: At the beginning of training, an initial weight is assigned to each sub-loss function Artificial settings based on prior knowledge; B. Calculate loss gradient: In each iteration, calculate the gradient of each sub-loss function C. Update weights: Dynamically adjust weights w based on the performance of the model in the current iteration k , the weight update formula is: in, is the weight of the kth sub-loss function at the tth iteration, α k is the learning rate, which is used to control the speed of weight adjustment. The denominator is the sum of the squares of the gradients of all sub-loss functions, which is used to normalize the weight update; D. Apply weights: Use the updated weights Calculate the new total loss L(θ) and update the model parameters θ accordingly; Repeat steps A to D until the model converges or reaches a predetermined number of training rounds.

Images