Rabbit meat freshness detection method based on electronic nose technology

Abstract

The invention relates to a rabbit meat freshness detection method based on an electronic nose technology. The volatile gas release mode of the rabbit meat under different storage conditions including temperature, humidity and time is analyzed through experiments, and key putrefying gas including dimethylamine, hydrogen sulfide and aldehydes and the concentration change trend are recognized; a metal oxide semiconductor (MOS) or an electrochemical sensor matched with the gas is screened according to the identified main decay gas, and a sensitive material means comprising a tin oxide thin film doped with platinum is optimized; a response matrix of different sensors to target gas is established through experiments, and a multi-sensor signal fusion algorithm is optimized; a temperature and humidity sensor is introduced, and environmental noise correction is performed in combination with odor signals; adopting a convolutional neural network feature extraction method to extract nonlinear features associated with the rabbit meat decay stage; docking the extracted smell features with the dynamic distribution model, and unifying the quantization dimension through a feature standardization algorithm; and correcting the predicted corruption trend in real time in combination with a smell distribution dynamic model.

Description

Technical Field

[0001] The invention relates to a rabbit meat freshness detection method based on electronic nose technology. Background Art

[0002] At present, the rabbit meat freshness detection method based on electronic nose technology, although it can improve the efficiency and accuracy of food freshness monitoring to a certain extent, still has some obvious shortcomings and drawbacks. Taking Chinese invention patent CN2012102915877 as an example, a kind of electronic nose technology for hairtail freshness detection is described. Although certain application effects have been achieved, its method also exposes some problems, which are also applicable to other meat freshness detection systems based on electronic nose technology, especially in rabbit meat freshness detection. First, the existing electronic nose technology usually needs to place the sample in a special gas collection bottle, and process the sample through steps such as preheating and vacuuming. This process has high requirements for sample preparation and operation, and requires a large amount of manpower and time investment. For the detection of rabbit meat, similar processing steps will appear particularly complicated, especially in the production and transportation process, the processing of rabbit meat usually requires efficient and convenient methods. The existing electronic nose technology method is difficult to meet the needs of real-time and high-throughput detection due to complex operating steps, and cannot achieve fast and accurate judgment of rabbit meat freshness. Secondly, the existing methods have strict requirements on the temperature control of the gas collection bottle. Usually, the gas collection bottle needs to be preheated to 40-50°C and is required to stand for 20-25 minutes. Although this process can help stabilize the concentration of the gas sample, for foods that are easy to spoil, such as rabbit meat, this long standing time may cause some of the corrupt gases in the sample to volatilize, thereby affecting the accuracy of freshness detection. In addition, rabbit meat will have different volatile gas release patterns under different storage conditions (such as cold chain transportation, room temperature storage, etc.), and existing technologies cannot be effectively adjusted and optimized for these different environmental conditions.

[0003] Furthermore, existing electronic nose systems usually rely on multiple gas sensor arrays, such as volatile aroma component sensors, volatile nitrogen oxide sensors, ammonia sensors, sulfide sensors, etc. These sensors can identify different types of volatile gases and evaluate freshness through load analysis. However, in applications, the selection and layout of these sensors usually do not fully consider the complexity of odor components and mutual interference. For complex foods such as rabbit meat, odor signals often contain a large number of complex compounds, and a single sensor array may not be able to fully and accurately capture all volatile components that affect freshness. For example, there may be some trace amounts of volatile ammonia compounds in rabbit meat, and the concentration of these compounds gradually increases during the spoilage process, but the existing gas sensor array may not be able to effectively identify these components, thereby affecting the accuracy of freshness assessment. Although the existing electronic nose technology can effectively identify different volatile gases, its lack of real-time and sensitivity also limits its application effect. For foods with large biological volume such as rabbit meat, the change in freshness is often gradual and affected by many factors. When faced with complex food samples, traditional electronic nose systems may not be able to obtain sufficiently accurate odor signals in a short period of time, resulting in reduced accuracy in judging freshness. In addition, the existing system's gas sampling methods and analysis speed are limited, which restricts the real-time monitoring function and cannot meet high-frequency, large-scale detection needs. Summary of the invention

[0004] The purpose of the present invention is to provide a rabbit meat freshness detection method based on electronic nose technology, thereby solving some of the drawbacks and shortcomings pointed out in the background technology.

[0005] The present invention solves the above-mentioned technical problems by adopting the following technical solutions: A rabbit meat freshness detection method based on electronic nose technology, comprising: S1, constructing a dynamic distribution model of volatile gases in the process of rabbit meat corruption:

[0006] S1.1. Analyze the volatile gas release pattern of rabbit meat under different storage conditions including temperature, humidity and time through controlled experiments, and identify the key corruption gases including dimethylamine, hydrogen sulfide and aldehydes and their concentration change trends;

[0007] S1.2. Based on the experimental data, a dynamic model of gas release was constructed using multivariate regression analysis to define the functional relationship between gas concentration and corruption process;

[0008] S2. Multi-dimensional response of sensor array for corruption gas characteristics:

[0009] S2.1. For the main corruption gases identified by the dynamic distribution model, metal oxide semiconductor MOS or electrochemical sensors suitable for the gas are screened, and sensitive material means including platinum-doped tin oxide film are optimized;

[0010] S2.2. Establish the response matrix of different sensors to target gas through experiments and optimize the multi-sensor signal fusion algorithm;

[0011] S2.3, introduce temperature and humidity sensors and combine them with odor signals to correct environmental noise;

[0012] S3. Hierarchical multi-level extraction of odor fingerprint features:

[0013] S3.1. Perform hierarchical analysis on the multi-dimensional odor signal output by the sensor array, including the base layer: low-frequency features; the enhancement layer: dynamic features, respectively extracting static features including the mean value of gas concentration and dynamic features including the concentration change rate;

[0014] S3.2, using a convolutional neural network feature extraction method to extract nonlinear features associated with the spoilage stage of rabbit meat;

[0015] S3.3, connect the extracted odor features with the dynamic distribution model, and unify the quantification dimensions through the feature normalization algorithm;

[0016] S4. Scenario-driven rabbit meat freshness detection classification model:

[0017] S4.1. Collect rabbit meat sample data based on different storage scenarios including cold chain transportation, cold storage, and room temperature storage, and manually annotate odor signals and freshness levels, including the correspondence between fresh, initial corruption, and severe corruption;

[0018] S4.2. Use variational autoencoder to classify and model the extracted odor fingerprint features;

[0019] S5. Dynamic trend prediction and freshness time series evaluation:

[0020] S5.1. Based on the collected odor signals and environmental data including temperature and humidity, a time series prediction model of freshness change is constructed using a long short-term memory network (LSTM);

[0021] S5.2. Combined with the odor distribution dynamic model, the predicted corruption trend is corrected in real time.

[0022] Furthermore, the hierarchical multi-level extraction method of the odor fingerprint feature includes:

[0023] Extract the static characteristics of the gas signal and monitor the stable concentration information of the volatile gas; introduce the static characteristic calculation method for the collected sensor output signal, and integrate the overall stability of the concentration signal over time and the impact of non-target gas interference on the detection; the definition of the static characteristic value is as follows:

[0024]

[0025] In the formula, F s It represents the static characteristic value, which is used to reflect the average concentration level of the gas signal; S(t) is the output signal value of the sensor at time t, which represents the change of gas concentration over time; t 0 and t n are the start time and end time of signal acquisition respectively; e -αt is the time attenuation weight factor, where α is the attenuation coefficient, which is used to reduce the noise interference of the acquisition.

[0026] Furthermore, the hierarchical multi-level extraction method of the odor fingerprint feature includes:

[0027] The dynamic characteristics of the sensor signal are extracted; the dynamic characteristics mainly include the concentration change rate and the frequency domain characteristics of the signal; the concentration change rate is calculated by the first-order derivative formula:

[0028]

[0029] In the formula, F d represents a dynamic characteristic value, which is used to quantify the rate at which gas concentration changes with time; It is the time derivative of gas concentration, indicating the rate of change of concentration; cos(βt) is the dynamic adjustment factor, which controls the influence of the signal per unit time, where β is the frequency adjustment parameter, which is used to highlight the characteristics of periodic odor changes.

[0030] Furthermore, the hierarchical multi-level extraction method of the odor fingerprint feature includes:

[0031] Design a fusion mechanism of static features and dynamic features, take static features as the base layer input, and dynamic features as the enhancement layer supplement; the fused feature vector F c Defined by the following formula:

[0032]

[0033] In the formula, F c Represents the integrated feature value after fusion, which is used to finally input the deep learning model; 1 and λ 2 They are the weight coefficients of static and dynamic features, respectively, reflecting the importance of the two types of features to the final features; F s is a static eigenvalue, indicating the average concentration of the base layer signal; F d is the dynamic eigenvalue, which indicates the dynamic characteristics of the signal changing with time; e -γt It is the time attenuation factor of the dynamic feature, which controls the weight distribution of the dynamic feature in different time ranges, where γ is the attenuation coefficient.

[0034] Furthermore, the dynamic trend prediction and freshness time series evaluation method includes:

[0035] The static features of the mean concentration of the odor signal and the dynamic features of the change rate are extracted, and the temperature and humidity data are collected synchronously, and the two are fused to form a multi-modal input; the static feature F s Calculation:

[0036]

[0037] Among them: F s ′ is a static characteristic value, which is used to reflect the concentration-weighted mean of the odor signal; S(t) is the output signal of the sensor at time t, which indicates the change of gas concentration over time; T(t) is the temperature data, which indicates the ambient temperature at time t; κ is the temperature adjustment factor, which reflects the enhancement effect of temperature on odor concentration; e -αt is the time decay function, which is used to reduce the weight of the data at early time points; α is the time decay coefficient; t 0 and t n Define the integration interval for the start and end time of the acquisition, respectively.

[0038] Furthermore, the dynamic trend prediction and freshness time series evaluation method includes:

[0039] The dynamic change rate is used to analyze the change rate of the odor signal concentration, and the humidity adjustment factor is combined to enhance the temporal correlation of the features. d Perform the calculation:

[0040]

[0041] Among them: F d ′ is the dynamic characteristic value, which quantifies the rate of change of the odor signal concentration; is the rate of change of odor signal concentration over time; H(t) is humidity data, indicating the ambient humidity at time t; sin(β·H(t)) is the humidity modulation factor, which strengthens the role of humidity in dynamic characteristics through nonlinear functions; β is the humidity weight factor, which determines the adjustment amplitude of humidity on dynamic characteristics.

[0042] Furthermore, the dynamic trend prediction and freshness time series evaluation method includes:

[0043] Receive the time series feature matrix X and jointly model the odor signal and environmental data; the input layer divides X into time step features, and represents the signal and environmental data of each time step through an embedding vector of a specific dimension; use the LSTM memory unit to capture the long-term dependency between the odor signal and the environmental data, and filter out irrelevant information through the forget gate; the model dynamically updates the hidden state, synthesizing the input features of the current time step and the hidden state of the previous time step to obtain a new memory state;

[0044] The hidden state C t Update

[0045] C t =φ(C t-1 ,X t )+ω·Δ(X t )

[0046] in:

[0047] C t is the hidden state of the current time step, which is used to store the comprehensive results of long-term and short-term information; C t-1 is the hidden state of the previous time step, representing the memory of historical information; X t is the input feature of the current time step, including odor signal and environmental data; φ(C t-1 ,X t ) is the hidden state update function, which merges the state of the previous time step with the current feature; ω·Δ(X t ) is the feature compensation term, which is used to enhance the influence of the current feature on the hidden state update; ω is the compensation weight coefficient, which controls the degree of correction of the input feature to the hidden state.

[0048] Beneficial effects of the present invention:

[0049] The electronic nose sensor array collects volatile gas data from the surface of rabbit meat in real time, and by analyzing the relationship between the odor fingerprint characteristics and freshness, it can accurately detect the freshness changes of rabbit meat. Especially during storage and transportation, real-time monitoring of odor signals can promptly reflect the corruption of rabbit meat and provide protection for food safety.

[0050] The present invention combines odor signals with environmental data (such as temperature and humidity) to form a multimodal input. This data fusion method enhances the model's ability to predict the changes in the freshness of rabbit meat under different environmental conditions, especially when the temperature fluctuates and the humidity changes greatly, and can more accurately reflect the actual freshness of rabbit meat.

[0051] By using LSTM (Long Short-Term Memory Network) to perform time series analysis on odor signals and environmental data, this method can capture long-term and short-term dependencies and dynamically predict the freshness trend of rabbit meat. Compared with traditional static detection methods, the LSTM model can dynamically update the prediction results based on real-time data, has strong adaptability, and is particularly suitable for storage environments that change rapidly. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 The present invention is a flow chart of the rabbit meat freshness detection method based on the electronic nose technology.

[0053] Figure 2 The figure is a flow chart of the hierarchical multi-level extraction method of odor fingerprint features of the present invention.

[0054] Figure 3 This is a flow chart of the dynamic trend prediction and freshness time series evaluation method of the present invention. DETAILED DESCRIPTION

[0055] The specific implementation modes of the present invention will be described in detail below in conjunction with the accompanying drawings.

[0056] Combined with Figure 1 The process of the rabbit meat freshness detection method based on the electronic nose technology of the present invention first needs to construct a dynamic distribution model of volatile gases during the corruption process of rabbit meat. The key goal of the S1 stage is to conduct an in-depth analysis of the volatile gas release pattern of rabbit meat under different storage conditions (including factors such as temperature, humidity and time), so as to reveal the dynamic changes of gases during the corruption process. S1.1 First, through a controlled experiment, rabbit meat is stored under set environmental conditions (such as different temperatures, humidity and time periods), and the release of volatile gases is continuously monitored. In the experiment, electronic nose technology is used to capture and analyze the gas signals released by rabbit meat in real time, focusing on key volatile gases in the corruption process such as dimethylamine, hydrogen sulfide and aldehydes. The concentration changes of these gases can reveal the corruption process of rabbit meat. Therefore, accurately identifying and monitoring the concentration change trends of these gases becomes the core content of the experiment. By analyzing the time series data of these gases, the experiment can capture the corresponding relationship between gas concentration and corruption stage, thereby providing necessary experimental data for the establishment of subsequent models. S1.2 is based on the experimental data and uses multivariate regression analysis method to construct a dynamic model of gas release. This model can achieve a quantitative description of the spoilage process by defining the functional relationship between gas concentration and spoilage progression. Specifically, multivariate regression analysis combines multiple variables (such as temperature, humidity and time) with gas concentration data to establish a mathematical model of gas concentration changes over time, thereby predicting the dynamic pattern of gas release under different conditions and further evaluating its relationship with rabbit meat spoilage.

[0057] In the second stage, the multi-dimensional response of the sensor array for the characteristics of corrupt gas first needs to screen suitable sensors for the main corrupt gases identified by the dynamic distribution model in the S1 stage. S2.1 focuses on identifying the sensor types that match the corrupt gases (such as dimethylamine, hydrogen sulfide and aldehydes, etc.) based on experimental data. In this process, the main sensor technologies include metal oxide semiconductor (MOS) sensors and electrochemical sensors, which have important applications in gas detection. In particular, MOS sensors produce conductivity changes through the reaction of gas molecules with the surface of metal oxides, thereby realizing gas detection. Electrochemical sensors generate current changes through the reaction of electrodes and gases to detect gas concentrations. In order to improve the performance and sensitivity of the sensor, platinum-doped tin oxide thin film materials are further used as sensitive materials. Platinum-doped tin oxide thin films have high thermal stability and gas selectivity, which can significantly improve the sensitivity and response speed of the sensor to the target gas. By optimizing the selection of materials, the detection accuracy and stability of the sensor can be improved, providing high-quality response data for subsequent odor signal acquisition. In step S2.2, based on experimental data, the response matrix of different sensors to the target gas is established to form a multi-dimensional response pattern of the sensor array. Each sensor responds differently to a specific gas, so a multi-sensor array can obtain more comprehensive and accurate odor data through a comprehensive analysis of the responses of multiple gases. In order to process multi-dimensional response data, the sensor signals also need to be fused and optimized. The optimized multi-sensor signal fusion algorithm can synthesize more accurate odor recognition signals based on the response characteristics of each sensor, thereby improving the robustness and accuracy of the detection system. Finally, S2.3 introduces temperature and humidity sensors and combines odor signals for environmental noise correction. Temperature and humidity have a significant impact on the response of gas sensors, resulting in interference from environmental noise and affecting the accurate measurement of gas concentrations.

[0058] The third stage is the hierarchical multi-level extraction of odor fingerprint features, which aims to extract key features closely related to the corruption stage of rabbit meat from the multidimensional odor signal output by the sensor array. S3.1 First, the odor signal output by the sensor array is hierarchically analyzed and the signal is divided into two parts: the base layer and the enhancement layer. The base layer mainly includes low-frequency features, such as the mean of gas concentration. These features reflect the stable concentration level of the gas over a long time scale and can usually be used to evaluate the freshness of rabbit meat. The extraction of low-frequency features relies on the statistical analysis of the gas concentration time series and the calculation of static features such as its mean. These features directly reflect the basic concentration level of the gas. The enhancement layer focuses on the extraction of dynamic features, mainly the rate of change of gas concentration. These features can capture the rapid changes and nonlinear dynamic changes in the corruption process. By analyzing the rate of change of gas concentration, the enhancement layer can provide detailed information about the corruption process, which is crucial for accurately judging the corruption stage. S3.2 further uses the convolutional neural network (CNN) feature extraction method to extract nonlinear features closely related to the corruption stage of rabbit meat from the multidimensional odor signal. Convolutional neural networks have powerful pattern recognition capabilities and can extract useful features from raw odor signals, which can help identify different stages of corruption. CNN can capture the potential complex patterns and features in odor signals, such as mixed signals of different gases, which are often difficult to capture with traditional methods. The convolution layer convolves the input signal through multiple filters to extract multi-level features, and finally integrates these features through the fully connected layer to obtain accurate identification of the rabbit meat corruption stage. S3.3 connects the extracted odor features with the previously established dynamic distribution model and unifies the quantization dimensions through the feature normalization algorithm. Feature normalization is to ensure the comparability between different features, eliminate the differences caused by different dimensions and numerical ranges, and enable all extracted features to be analyzed under the same standard. Through feature normalization, information from different sensors and different feature levels can be effectively quantified, further improving the accuracy and stability of the model.

[0059] In the fourth stage, the scene-driven rabbit meat freshness detection classification model aims to establish a classification model that can accurately judge the freshness of rabbit meat by combining the data features of different storage scenarios. S4.1 First, it is necessary to collect rabbit meat sample data in multiple storage scenarios, including cold chain transportation, cold storage and room temperature storage. Each storage scenario corresponds to different environmental conditions and temperature and humidity, so the odor characteristics of rabbit meat in these different scenarios will also be different. In this process, electronic nose technology is used to collect odor signals of rabbit meat samples under different storage conditions, and the odor signals are associated with freshness levels through manual annotation. These freshness levels include "fresh", "initial corruption" and "serious corruption". By annotating the odor signals of different freshness levels, the corresponding relationship between odor signals and freshness levels is established. This manual annotation process is to ensure the accuracy of sample data and provide high-quality annotated data for subsequent model training. S4.2 uses variational autoencoder (VAE) to classify and model the extracted odor fingerprint features. Variational autoencoder is a generative model that performs unsupervised learning by learning the latent representation of data. In this process, VAE first maps the input odor signal features to the latent space through the encoder, and then reconstructs the latent representation into the odor signal through the decoder. In this process, VAE can not only learn the compressed representation of odor features, but also model the complex patterns of odor signals through the learning of the latent space. By classifying and modeling the odor signal features, VAE can find the distribution of odor features of different freshness levels in the latent space, and map these distributions to the corresponding corruption stages, and finally achieve efficient classification of the freshness of rabbit meat.

[0060] The fifth stage, dynamic trend prediction and freshness time series evaluation, focuses on accurately predicting and evaluating the freshness change trend of rabbit meat through time series models and real-time correction mechanisms. S5.1 First, based on the collected odor signals and environmental data (including temperature, humidity, etc.), a long short-term memory network (LSTM) is used to build a time series prediction model for freshness change. LSTM is a special recurrent neural network (RNN) designed to capture long-term dependencies in time series data, and is particularly suitable for processing time series data such as odor signals. By taking odor signals and temperature and humidity data as input, the LSTM model can learn the change pattern of odor signals over time, thereby predicting the evolution of rabbit meat freshness over time. In this process, LSTM screens and memorizes the input data through its gating mechanism (input gate, forget gate, output gate), thereby effectively capturing the key dynamic features of the rabbit meat corruption process. The output of LSTM will be the probability distribution or category label of freshness change, indicating the degree of corruption of rabbit meat at the current moment and in the future. S5.2 combines the odor distribution dynamic model to make real-time corrections to the predicted corruption trend. The odor distribution dynamic model provides a dynamic reference framework for changes in odor characteristics based on the changing patterns of odor at different time points and under different environmental conditions. By comparing the odor signal collected in real time with the dynamic model, the prediction results of the LSTM model can be corrected to improve the accuracy of the prediction. The real-time correction mechanism can capture changes in odor signals caused by changes in the external environment (such as temperature and humidity fluctuations) or other unknown factors. Through this correction mechanism, the LSTM model can more flexibly adapt to the temporal changes in rabbit meat corruption in complex environments. Combined with the correction mechanism of the odor distribution dynamic model, the predicted corruption trend will be more stable and accurate, thereby providing more accurate temporal prediction information for the freshness assessment of rabbit meat.

[0061] Embodiment 1:

[0062] Combined with Figure 2 In this embodiment, in the rabbit meat freshness detection method based on electronic nose technology of the present invention, static feature extraction is one of the key steps to monitor the concentration of volatile gases in the process of rabbit meat corruption. By analyzing the stability of the gas signal, the change trend of the odor can be effectively monitored, providing an important reference for freshness assessment.

[0063] Ten rabbit meat samples were stored in a refrigerated environment, and the sensor array of electronic nose technology was used to monitor the release of volatile gases in real time. The sensor array used in the setting included a variety of sensors, including tin oxide (SnO 2 ) type sensor is used to monitor corruption gases such as ammonia, hydrogen sulfide, and aldehydes.

[0064] In this experiment, the signal St output by the sensor is a real-time response to the gas concentration. The data collected by the sensor within 10 hours are as follows:

[0065] S(t)=[0.5,0.6,0.7,0.75,0.8,0.85,0.9,0.95,1.0,1.05] (unit: ppm, indicating the gas concentration value detected by the sensor)

[0066] These data represent the 0 = 0 hours to t n = The sensor's response to gas concentration over a 10-hour period. The following formula is required to calculate the static characteristic value F s :

[0067]

[0068] The meanings of the symbols in the formula are as follows:

[0069] F s It is a static characteristic value, reflecting the stable level of gas concentration;

[0070] S(t) is the gas concentration signal output by the sensor at time t;

[0071] t 0 and t n are the start time and end time of signal acquisition, respectively, which are 0 hours and 10 hours in this example;

[0072] e -αt It is the time attenuation weight factor, which is used to reduce the noise interference of earlier acquisitions. α is the attenuation coefficient, and the setting value range is 0.1 to 0.3, which means that the signal at a more distant time has less influence on the calculation result.

[0073] The attenuation coefficient α is set to 0.2, which means that the sensitivity to the change of gas concentration over time is reduced, especially the noise at the earlier time contributes less to the final result.

[0074] Next, the static eigenvalue F is calculated s For the sake of simplicity, the above formula is solved by numerical integration method, and the calculation results are as follows:

[0075] 1.S(t)=[0.5,0.6,0.7,0.75,0.8,0.85,0.9,0.95,1.0,1.05]

[0076] 2. Calculate the decay factor at each time point:

[0077] e -αt=[1,0.8187,0.6692,0.5476,0.4493,0.3681,0.3012,0.2469,0.2014,0.1645]

[0078] 3. Calculate the weighted gas concentration at each moment:

[0079] S(t)×e -αt =[0.5,0.4902,0.4684,0.4107,0.3594,0.3139,0.2709,0.2350,0.2014,0.1739]

[0080] 4. Sum and divide by the total time interval:

[0081]

[0082] Therefore, the static eigenvalue F s It is 0.3928ppm, which means that the average level of gas concentration during the entire observation period is 0.3928ppm.

[0083] This static eigenvalue F s It can be used to reflect the overall stability of gas concentration, and it can also provide a preliminary judgment on the freshness of rabbit meat. By comparing with the standard freshness template, it is found that the current gas concentration value is far from the standard value of the "fresh" sample, then it can be further judged that the sample has begun to spoil.

[0084] In addition to the extraction of static features, it is also necessary to extract dynamic features to more comprehensively evaluate the changes in gas concentration and further improve the accuracy of freshness detection. The extraction of dynamic features can be achieved through the concentration change rate and the frequency domain characteristics of the signal. The concentration change rate reflects the speed at which the gas concentration changes over time, and is usually achieved by calculating the rate of change of the gas concentration over time. In the present invention, the first-order derivative formula is used to calculate the concentration change rate, and a dynamic adjustment factor is introduced to highlight the characteristics of periodic odor changes.

[0085] In an experimental environment where rabbit meat is stored, we continue to use the previous sensor data set and have a signal sequence of gas concentration changing over time. In the previous experiment, we have obtained the gas concentration value S(t) = [0.5, 0.6, 0.7, 0.75, 0.8, 0.85, 0.9, 0.95, 1.0, 1.05] ppm at each time point within 10 hours. In this experiment, we will focus on the rate of change of gas concentration and introduce a frequency adjustment factor to consider the impact of periodic odor changes.

[0086] The calculation formula of dynamic characteristics is as follows:

[0087]

[0088] Among them, F d It represents the dynamic characteristic value, which measures the rate of change of gas concentration; It is the rate of change of gas concentration over time, that is, the derivative of concentration; cos(βt) is a dynamic adjustment factor used to control the impact of the signal per unit time, where β is a frequency adjustment parameter with a setting range of 0.1 to 0.5, indicating the response intensity of different periodic odor changes.

[0089] Step 1: Calculate the time derivative of the concentration

[0090] First, we need to calculate the rate of change of gas concentration over time That is, the derivative of the concentration. Since there is only discrete time data, it can be approximated by the finite difference method:

[0091]

[0092] For example, at time point t 0 = 0 hours to t 1 = The concentration change rate between 1 hour, concentration S(0) = 0.5pppm and S(1) = 0.6ppm, so the concentration change rate is:

[0093]

[0094] Similarly, the concentration change rate at other times can be calculated to obtain:

[0095]

[0096] Step 2: Calculate the dynamic adjustment factor

[0097] Next, we need to calculate the frequency adjustment factor cos(βt), where β is the frequency adjustment parameter. If β=0.2, we can calculate the adjustment factor at each moment:

[0098] cos(0.2×t)=[1,0.9801,0.9211,0.8141,0.6617,0.4685,0.2409,-0.0043,-0.2563,-0.5114]

[0099] Step 3: Calculate dynamic eigenvalues

[0100] Now, the dynamic eigenvalue F can be calculated at each time point d :

[0101]

[0102] Substituting the results from the first two steps into the calculation, we get:

[0103] F d =[0.1·1,0.1·0.9801,0.05·0.9211,0.05·0.8141,0.05·0.6617,0.05·0.4685,0.05·0.2409,0.05·-0.0043,0.05·-0.2563,0.05·-0.5114]

[0104] Right now:

[0105] F d =[0.1,0.098,0.046,0.041,0.033,0.023,0.012,-0.0002,-0.0128,-0.0256]

[0106] Step 4: Analyze dynamic characteristics

[0107] From the calculation results, it can be seen that the dynamic eigenvalue F d In the early stage, the concentration change rate is high. As time goes by, the change of gas concentration gradually slows down and becomes negative, which means that the change of gas concentration tends to be stable or reversed, which is related to the early stage of the spoilage process. During the storage of fresh rabbit meat, the rate of odor change is high. As the spoilage progresses, the rate of concentration change gradually decreases, which is reflected in the calculation of dynamic characteristics.

[0108] By introducing a fusion mechanism, static features and dynamic features are combined into a unified feature vector F c , providing more comprehensive input information for subsequent deep learning models. This fusion mechanism can not only reflect the stability of gas concentration, but also reflect the characteristics of gas concentration change over time, so as to more accurately evaluate the freshness of rabbit meat. The fused feature vector is calculated by the following formula:

[0109]

[0110] Among them, F c Represents the integrated eigenvalue after fusion, which is used as the feature vector finally input to the deep learning model; 1 and λ 2 They are the static features F s and dynamic characteristics F d The weight coefficient is used to adjust the importance of the two in the fusion process; F s Indicates static characteristics, reflecting the average level of gas concentration; F d Represents dynamic characteristics and quantifies the rate of change of gas concentration; e -γtIt is the time decay factor, which is used to control the weight distribution of dynamic features in different time ranges, where γ is the decay coefficient, which reflects the degree of decay of dynamic features over time.

[0111] In the experimental environment of testing the freshness of rabbit meat, the odor signal collected by electronic nose technology is used. Static and dynamic feature data have been obtained from the sensor. In the previous step, the static feature F s and dynamic characteristics F d , and the corresponding time decay factor.

[0112] Static eigenvalue F s The calculations show that:

[0113]

[0114] Dynamic eigenvalue F d The calculations show that:

[0115] F d =[0.1,0.098,0.046,0.041,0.033,0.023,0.012,-0.0002,-0.0128,-0.0256]

[0116] Based on these feature data, it is necessary to set an appropriate weight coefficient λ 1 and λ 2 , and the attenuation coefficient γ, in order to effectively integrate static features and dynamic features. The coefficients set are: λ 1 =0.6,λ 2 =0.4, and the attenuation coefficient γ is selected to be 0.05, which means that the influence of the dynamic characteristics will gradually decay, but still maintain a high sensitivity to changes in a shorter time range.

[0117] Step 1: Calculate the weighted integral of the dynamic features

[0118] In order to incorporate dynamic features into the fusion calculation, it is necessary to first calculate the weighted integral of the dynamic features within the time range. d Data and decay factor e -γt Combined, the weighted integral value of the dynamic characteristics is obtained. Set the time step to 1 hour, and the calculation is as follows:

[0119]

[0120] For a given F d Numeric sequence:

[0121] [0.1,0.098,0.046,0.041,0.033,0.023,0.012,-0.0002,-0.0128,-0.0256]

[0122] Compute the weighting coefficients for each time step:

[0123]

[0124] And so on, we get the weighting factor for each time step.

[0125] Then F d The value is multiplied by the weighting factor and summed to obtain the weighted integral value of the dynamic feature. The calculation result is set as:

[0126]

[0127] Step 2: Calculate the fused eigenvalues

[0128] Now we have obtained the static eigenvalue F s = 0.75ppm and the weighted integral value of the dynamic characteristics Use the fusion formula to calculate the comprehensive eigenvalue F c :

[0129]

[0130] Substituting the values:

[0131] F c =0.6×0.75+0.4×0.45=0.45+0.18=0.63ppm / h

[0132] Step 3: Analysis and practical application

[0133] The obtained fusion feature value F c = 0.63ppm / h reflects the freshness of rabbit meat. In the application, this feature value will be provided as input to the deep learning model, and through model training and optimization, the degree of corruption or remaining shelf life of rabbit meat can be predicted. For example, when F c When the value is high, it means that the rabbit meat is still relatively fresh; c When the value is low and gradually decreasing, it means that the corruption process is accelerating.

[0134] Embodiment 2:

[0135] Combined with Figure 3In this embodiment, the freshness of a batch of rabbit meat is monitored in real time, and sensor data from the refrigerated transportation process is collected. At this time, the corruption trend of rabbit meat is predicted based on the mean and change rate of the odor signal concentration and the temperature and humidity data, and the static and dynamic features are combined to form a multimodal input to provide accurate freshness evaluation.

[0136] In this process, the static characteristic F′ s The calculation formula is as follows:

[0137]

[0138] in:

[0139] F′ S Represents the static characteristic value, which is used to reflect the concentration-weighted mean of the odor signal;

[0140] S(t) is the output signal of the sensor at time t, indicating the change of gas concentration over time;

[0141] T(t) is temperature data, indicating the ambient temperature at time t;

[0142] κ is the temperature regulation factor, which is used to reflect the enhancing effect of temperature on odor concentration;

[0143] e -αt is a time decay function, which is used to reduce the weight of data at early time points;

[0144] α is the time decay coefficient;

[0145] t 0 and t n They are the start and end time of the acquisition, defining the integration interval.

[0146] An experiment was conducted on a batch of rabbit meat, and data was collected synchronously using electronic nose technology and temperature and humidity sensors. During the experiment, the odor signals and ambient temperature data of the sensors were as follows:

[0147] Sensor output signal S(t) (unit: ppm): [0.08, 0.12, 0.16, 0.18, 0.22, 0.24, 0.28, 0.32, 0.36, 0.4];

[0148] Ambient temperature T(t) (unit: ℃): [4.5, 5.0, 5.2, 5.5, 5.7, 6.0, 6.2, 6.5, 6.8, 7.0].

[0149] Set the selected sampling time period to t 0 = 0 hours, t n =9 hours.

[0150] Step 1: Calculate the effect of the temperature adjustment factor

[0151] First, the value of the temperature adjustment factor κ needs to be defined. Experimental results show that the effect of temperature on odor concentration needs to be adjusted by a factor of κ = 0.05, that is, for every 1°C increase in temperature, the odor concentration will increase by 5%.

[0152] Substitute the temperature data T(t) into the formula κ·log(1+T(t)), where T(t) is the temperature data at each moment. For example, at t=0 hours, the temperature is 4.5°C, and the calculation is:

[0153] κ·log(1+4.5)=0.05·log(5.5)≈0.05·1.7047=0.0852

[0154] Applying this calculation method to all temperature data, we get the values ​​of the temperature adjustment factor: [0.0852, 0.0871, 0.0882, 0.0895, 0.0905, 0.0915, 0.0924, 0.0934, 0.0941, 0.0947].

[0155] Step 2: Calculate the impact of the time decay factor

[0156] It is also necessary to adjust the influence of the data at earlier time points according to the time decay coefficient α = 0.1. Using formula e -αt Calculate the decay factor for each time point. For example, at t = 0 hours, the decay factor is:

[0157] e -0.1·0 =1

[0158] At t = 1 hour, the attenuation factor is:

[0159] e -0.1·1 ≈0.9048

[0160] And so on, calculate the attenuation factor value at each time point:

[0161] [1,0.9048,0.8187,0.7408,0.6703,0.6065,0.5493,0.4983,0.4532,0.4136].

[0162] Step 3: Calculate the integral of the weighted signal

[0163] Now, the static characteristic F′ can be calculated by combining the gas concentration signal S(t) with the temperature adjustment factor and the time decay factor: s For each time point, first multiply S(t)+κ·log(1+T(t)) by the decay factor and then integrate. Specifically:

[0164] At t = 0 hours:

[0165] (S(0)+0.0852)·1=(0.08+0.0852)·1=0.1652

[0166] At t = 1 hour:

[0167] (S(1)+0.0871)·0.9048=(0.12+0.0871)·0.9048≈0.1872

[0168] After continuing to calculate the weighted signals at other time points, we get the weighted signal sequence:

[0169] [0.1652,0.1872,0.2335,0.2664,0.2986,0.3335,0.3731,0.4181,0.4543,0.4851].

[0170] Step 4: Calculate the static eigenvalue F′ s

[0171] Finally, the weighted signal is integrated over time and divided by the length of the time interval t n -t 0 =9 hours:

[0172]

[0173] The calculation results are:

[0174]

[0175] Through the above steps, the static eigenvalue F′ is calculated s = 0.2986ppm / h, which is used to reflect the concentration weighted mean of the odor signal. Combined with other dynamic features, temperature and humidity data, and sensor signals, these feature values ​​can be further used to predict dynamic trends and evaluate the freshness trend of rabbit meat.

[0176] The key to dynamic trend prediction and freshness time series evaluation is how to monitor the changes in odor signals and their relationship with ambient humidity through dynamic feature analysis. Below, a specific calculation example is used to show how to use the dynamic feature F′ d To quantify the rate of change of odor signal concentration, and combine it with the humidity adjustment factor to enhance the temporal correlation of features, thereby achieving freshness assessment.

[0177] Step 1: Dynamic feature calculation

[0178] Dynamic feature F′ d The calculation formula is:

[0179]

[0180] Among them, F′ d It represents the dynamic characteristic value, quantifying the rate of change of the odor signal concentration; is the rate of change of odor signal concentration over time, H(t) is the environmental humidity data; β is the humidity weight factor, which is used to adjust the effect of humidity on dynamic characteristics. Here, the effect of humidity on dynamic characteristics is set to be greater, so the humidity weight factor β = 0.5 is selected.

[0181] The odor signal S(t) and humidity H(t) data monitored in the experiment are set as follows:

[0182] Odor signal concentration S(t) (unit: ppm): [0.08, 0.12, 0.16, 0.18, 0.22, 0.24, 0.28, 0.32, 0.36, 0.4]

[0183] Ambient humidity H(t) (unit: %): [60,62,64,65,66,67,68,69,70,71]

[0184] Step 2: Calculate the rate of change of the odor signal

[0185] Rate of change of odor signal It can be obtained by calculating the difference between adjacent moments. For example, set the calculation of the rate of change of the first time period:

[0186]

[0187] And so on, calculate the rate of change at each moment:

[0188] [0.04,0.04,0.02,0.04,0.02,0.04,0.04,0.04,0.04,0.04]

[0189] Step 3: Calculate the humidity adjustment factor

[0190] According to the humidity data H(t) and the humidity weight factor β=0.5, the humidity adjustment factor can be calculated. The humidity adjustment factor is:

[0191] sin(β·H(t))

[0192] For example, when H(t) = 60%, it is calculated as:

[0193] sin(0.5·60)=sin(30°)=0.5

[0194] And so on, calculate the humidity adjustment factor at each moment:

[0195] [0.5,0.522,0.544,0.557,0.57,0.582,0.595,0.607,0.62,0.632]

[0196] Step 4: Calculate dynamic features

[0197] Combining the rate of change of the odor signal and the humidity adjustment factor, the dynamic feature F′ at each moment can be calculated d . For example, at the first time point (t=0):

[0198] F′ d (0) = 0.04 · 0.5 = 0.02 ppm / h

[0199] Continue to calculate the dynamic characteristics at other time points:

[0200] [0.02,0.021,0.011,0.022,0.012,0.023,0.024,0.025,0.025,0.025]

[0201] Step 5: Dynamic trend prediction and time series evaluation

[0202] The calculated dynamic features can be used to further analyze the freshness trend of rabbit meat. Due to the combination of the odor signal change rate and the humidity adjustment factor, the dynamic features can more accurately reflect the freshness changes of rabbit meat. These dynamic feature values ​​can be used to build a time series model to predict future spoilage trends.

[0203] The dynamic trend prediction and freshness time series evaluation method can monitor the freshness changes of rabbit meat in real time by using LSTM (Long Short-Term Memory Network) to jointly model the odor signal and environmental data. This paper will demonstrate how to use the memory unit of LSTM to combine the odor signal and environmental data to predict the corruption trend of rabbit meat, and show the process of time series data processing through the hidden state update formula.

[0204] Step 1: Construction of input feature matrix

[0205] Suppose that the smell signal of rabbit meat and environmental data (including temperature and humidity) are collected within a certain period of time. For example, data of 10 time steps are collected:

[0206] Odor signal S(t) (unit: ppm): [0.1, 0.12, 0.14, 0.16, 0.18, 0.2, 0.22, 0.24, 0.26, 0.28]

[0207] Ambient temperature T(t) (unit: °C): [10, 11, 12, 12.5, 13, 13.5, 14, 14.5, 15, 15.5]

[0208] Ambient humidity H(t) (unit: %): [60,61,62,63,64,65,66,67,68,69]

[0209] This set of data constitutes the feature matrix X, which contains the odor signal and environmental data. In order to adapt to the LSTM model, the data will be divided into multiple time step features, and the input features of each time step include odor signal, temperature and humidity. For example, the data of the first time step is: [0.1, 10, 60], the data of the second time step is: [0.12, 11, 61], and so on.

[0210] Step 2: Construction of LSTM model

[0211] The LSTM model has three key parts: input gate, forget gate, and output gate. LSTM controls long-term and short-term memory through these gate mechanisms and captures the long-term dependencies of input features through memory units. The input layer divides the feature matrix X into time step features and represents the odor signal and environmental data of each time step through the embedding vector.

[0212] In LSTM, the update formula of the hidden state is as follows:

[0213] C t =φ(C t-1 ,X t )+ω·Δ(X t )

[0214] Among them, C t Represents the hidden state of the current time step, which stores the comprehensive results of long-term and short-term information; C t-1 is the hidden state of the previous time step, containing the memory of historical information; X t is the input feature of the current time step, including odor signal, temperature and humidity; φ(C t-1 ,X t ) is the hidden state update function, which represents the fusion of the state of the previous time step and the current feature; ω·Δ(X t ) is a feature compensation term, which is used to enhance the influence of the current feature on the hidden state update.

[0215] Step 3: Update the hidden state

[0216] Taking the first time step as an example, C 0 = 0 as the initial hidden state, assuming that the current input feature X 1 =[0.1,10,60]. The hidden state C of the first time step is calculated by the hidden state update formula 1 :

[0217] C0 = 0 (Initial hidden state)

[0218] X 1 = [0.1, 10, 60] (Input features at the first time step)

[0219] Set the update function φ(C 0 , X 1 ) = 0.05, which means the fusion of the current feature and the hidden state at the previous time step results in 0.05.

[0220] Set the feature compensation term ω·Δ(X 1 ) = 0.03, which means the compensation effect of the current feature on the hidden state is 0.03.

[0221] Then, the hidden state at the first time step is:

[0222] C 1 = 0.05 + 0.03 = 0.08

[0223] Continue with similar calculations. At the 2nd time step, use C 1 to update C 2 , and so on. Set φ(C 1 , X 2 ) = 0.06 and ω·Δ(X 2 ) = 0.04, then:

[0224] C 2 = 0.06 + 0.04 = 0.10

[0225] Step 4: Training and prediction of the LSTM model

[0226] By training the LSTM model and combining with actual data, the changing pattern of the hidden state C t can be learned, thus capturing the long-term dependence relationship between the odor signal and the environmental data. During the training process, the LSTM will adjust the weight coefficients ω and the state update function φ so that the model can optimally predict the change trend of the freshness of rabbit meat.

[0227] Step 5: Application

[0228] In the application, the change trend of the freshness of rabbit meat is obtained through the LSTM model, and according to the model prediction results, it is found that starting from the 5th time step, the change rate of the hidden state C t significantly increases, indicating that the rabbit meat has started to show signs of spoilage. Based on this information, the cold chain transportation conditions can be adjusted in a timely manner to ensure the freshness of the rabbit meat.

[0229] By combining the memory cells of LSTM and environmental data (such as temperature and humidity), as well as the time series characteristics of odor signals, it is possible to efficiently model and predict the changes in the freshness of rabbit meat. This method can not only capture the long-term trend of rabbit meat spoilage, but also optimize the prediction effect by dynamically updating the hidden state and combining historical information with current features, providing a scientific basis for real-time monitoring and control of the freshness of rabbit meat.

[0230] The above shows and describes the basic principles, main features and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A rabbit meat freshness detection method based on electronic nose technology, characterized in that The following steps are involved: S1. Construction of dynamic distribution model of volatile gases during rabbit meat spoilage: S1.

1. Analyze the volatile gas release pattern of rabbit meat under different storage conditions including temperature, humidity and time through controlled experiments, and identify the key corruption gases including dimethylamine, hydrogen sulfide and aldehydes and their concentration change trends; S1.

2. Based on the experimental data, a dynamic model of gas release was constructed using multivariate regression analysis to define the functional relationship between gas concentration and corruption process; S2. Multi-dimensional response of sensor array for corruption gas characteristics: S2.

1. For the main corruption gases identified by the dynamic distribution model, metal oxide semiconductor MOS or electrochemical sensors suitable for the gas are screened, and sensitive material means including platinum-doped tin oxide film are optimized; S2.

2. Establish the response matrix of different sensors to target gas through experiments and optimize the multi-sensor signal fusion algorithm; S2.3, introduce temperature and humidity sensors and combine them with odor signals to correct environmental noise; S3. Hierarchical multi-level extraction of odor fingerprint features: S3.

1. Perform hierarchical analysis on the multi-dimensional odor signals output by the sensor array, including the basic layer: low-frequency features; Enhanced layer: dynamic features, extracting static features including the mean value of gas concentration and dynamic features including the concentration change rate; S3.2, using a convolutional neural network feature extraction method to extract nonlinear features associated with the spoilage stage of rabbit meat; S3.3, connect the extracted odor features with the dynamic distribution model, and unify the quantification dimensions through the feature normalization algorithm; S4. Scenario-driven rabbit meat freshness detection classification model: S4.

1. Collect rabbit meat sample data based on different storage scenarios including cold chain transportation, cold storage, and room temperature storage, and manually annotate odor signals and freshness levels, including the correspondence between fresh, initial corruption, and severe corruption; S4.

2. Use variational autoencoder to classify and model the extracted odor fingerprint features; S5. Dynamic trend prediction and freshness time series evaluation: S5.

1. Based on the collected odor signals and environmental data including temperature and humidity, a time series prediction model of freshness change is constructed using a long short-term memory network (LSTM); S5.

2. Combined with the odor distribution dynamic model, the predicted corruption trend is corrected in real time.

2. The method for detecting the freshness of rabbit meat based on electronic nose technology according to claim 1, characterized in that The hierarchical multi-level extraction method of the odor fingerprint feature includes: Extract the static characteristics of the gas signal and monitor the stable concentration information of the volatile gas; introduce the static characteristic calculation method for the collected sensor output signal, and integrate the overall stability of the concentration signal over time and the impact of non-target gas interference on the detection; the definition of the static characteristic value is as follows: In the formula, F s represents the static characteristic value, which is used to reflect the average concentration level of the gas signal; S(t) is the output signal value of the sensor at time t, which represents the change of gas concentration over time; t0 and t n are the start time and end time of signal acquisition respectively; e -αt is the time attenuation weight factor, where α is the attenuation coefficient, which is used to reduce the noise interference of the acquisition.

3. The method for detecting the freshness of rabbit meat based on electronic nose technology according to claim 2, characterized in that The hierarchical multi-level extraction method of the odor fingerprint feature includes: extracting the dynamic features of the sensor signal; the dynamic features mainly include the concentration change rate and the frequency domain characteristics of the signal.

4. The method for detecting the freshness of rabbit meat based on electronic nose technology according to claim 3, characterized in that The hierarchical multi-level extraction method of the odor fingerprint feature includes: Design a fusion mechanism of static features and dynamic features, take static features as the base layer input, and dynamic features as the enhancement layer supplement; the fused feature vector F c Defined by the following formula: In the formula, F c represents the integrated feature value after fusion, which is used to input the deep learning model; λ1 and λ2 are the weight coefficients of static and dynamic features, respectively, reflecting the importance of the two types of features to the final features; F s is a static eigenvalue, indicating the average concentration of the base layer signal; F d is the dynamic eigenvalue, which indicates the dynamic characteristics of the signal changing with time; e -γt It is the time attenuation factor of the dynamic feature, which controls the weight distribution of the dynamic feature in different time ranges, where γ is the attenuation coefficient.

5. The method for detecting the freshness of rabbit meat based on electronic nose technology according to claim 1, characterized in that The dynamic trend prediction and freshness time series evaluation method includes: extracting static features of the mean value of odor signal concentration and dynamic features of the change rate, and synchronously collecting temperature and humidity data, and fusing the two to form a multi-modal input.

6. The method for detecting the freshness of rabbit meat based on electronic nose technology according to claim 5, characterized in that The dynamic trend prediction and freshness time series evaluation method comprises: The dynamic change rate is used to analyze the change rate of the odor signal concentration, and the humidity adjustment factor is combined to enhance the temporal correlation of the features. d Perform the calculation: Among them: F d ′ is the dynamic characteristic value, which quantifies the rate of change of the odor signal concentration; is the rate of change of odor signal concentration over time; H(t) is humidity data, indicating the ambient humidity at time t; sin(β·H(t)) is the humidity modulation factor, which strengthens the role of humidity in dynamic characteristics through nonlinear functions; β is the humidity weight factor, which determines the adjustment amplitude of humidity on dynamic characteristics.

7. The method for detecting the freshness of rabbit meat based on electronic nose technology according to claim 6, characterized in that The dynamic trend prediction and freshness time series evaluation method comprises: Receive the time series feature matrix X and jointly model the odor signal and environmental data; the input layer divides X into time step features, and represents the signal and environmental data of each time step through an embedding vector of a specific dimension; use the LSTM memory unit to capture the long-term dependency between the odor signal and the environmental data, and filter out irrelevant information through the forget gate; the model dynamically updates the hidden state, synthesizing the input features of the current time step and the hidden state of the previous time step to obtain a new memory state; The hidden state C t Update C t =φ(C t-1 ,X t )+ω·Δ(X t ) in: C t is the hidden state of the current time step, which is used to store the comprehensive results of long-term and short-term information; C t-1 is the hidden state of the previous time step, representing the memory of historical information; X t is the input feature of the current time step, including odor signal and environmental data; φ(C t-1 ,X t ) is the hidden state update function, which merges the state of the previous time step with the current feature; ω·Δ(X t ) is the feature compensation term, which is used to enhance the influence of the current feature on the hidden state update; ω is the compensation weight coefficient, which controls the degree of correction of the input feature to the hidden state.

Images