Method for establishing, training and detecting distiller's yeast psychological value detection model

Abstract

The invention relates to the field of distiller's yeast physical and chemical value detection, and provides a distiller's yeast psychological value detection model establishment, training and detection method in order to realize nondestructive detection of distiller's yeast psychological values, and the distiller's yeast psychological value detection model uses different branches of a convolutional neural network to process hyperspectral data of each surface of distiller's yeast. And each branch comprises an independent convolutional layer and a pooling layer, so that the model can comprehensively consider data features of different surfaces and distance relationships between the data features and the center of curvature, thereby enhancing the robustness and generalization ability of the model. The coefficient k is obtained by establishing the linear relation model of the koji powder spectral information and the koji powder physical and chemical values, when the koji powder psychological value detection model is trained, the coefficient k serves as the initial weight, and the model training speed is increased. And nondestructive testing of the koji psychological value is realized based on the distiller's yeast koji psychological value detection model.

Description

Technical Field

[0001] The present invention relates to the field of physicochemical value detection of koji, and specifically to a method for establishing, training and detecting a physicochemical value detection model of the koji core. Background Art

[0002] In the brewing of Chinese liquor, the quality of koji is crucial for the physicochemical values of the final product. Traditionally, judging the fermentation quality of koji relies on manual experience. This method is not only highly subjective but also requires destroying the sample to observe the state of the koji core. The method of combining an intelligent model for detecting the physicochemical values of koji can only non-destructively check the characteristics of the surface of the daqu, and it is still impossible to non-destructively detect the physicochemical values inside the daqu. However, the quality of daqu fermentation is more determined by the physicochemical values inside the daqu. Therefore, there is an urgent need for a method to achieve non-destructive detection of the physicochemical values of the koji core. Summary of the Invention

[0003] In order to achieve non-destructive detection of the physicochemical values of the koji core, the present application provides a method for establishing, training and detecting a physicochemical value detection model of the koji core.

[0004] The technical solution adopted by the present invention to solve the above problems is as follows:

[0005] A method for establishing a physicochemical value detection model of the koji core, comprising:

[0006] Creating a physicochemical value detection model of the koji core with multi-modal data as input and the physicochemical values of the koji core as output. The multi-modal data includes at least the hyperspectral information of each surface of the koji. In the physicochemical value detection model of the koji core, a plurality of hyperspectral information processing branches corresponding one-to-one to each surface of the koji are set. By weighted fusion of the feature vectors generated by all hyperspectral information processing branches, and then outputting the prediction result of the physicochemical values of the koji core through a fully connected layer, wherein the weight d i represents the actual distance from the i-th surface to the koji core, and n is the total number of surfaces of the koji.

[0007] Further, the multi-modal data further includes the temperature and humidity information during the koji fermentation process.

[0008] Further, the physicochemical value detection model of the koji core is further provided with a feature extraction branch for generating a feature vector based on the temperature and humidity information during the koji fermentation process; by splicing the feature vector with the weighted fused feature, and then outputting the prediction result of the physicochemical values of the koji core through a fully connected layer.

[0009] Further, the physicochemical values include acidity, moisture content and starch content.

[0010] Further, the hyperspectral information includes the length, width of the hyperspectral image and the number of wavelengths of a single pixel.

[0011] Training method for physicochemical value detection model of distiller's yeast koji, including:

[0012] Step 1: Use a hyperspectral near-infrared instrument to comprehensively scan each surface of the distiller's yeast koji, collect the hyperspectral information x 表 of each surface, and record the geometric dimensions of the distiller's yeast koji to obtain the actual distance from each surface to the koji center;

[0013] Step 2: Obtain the physicochemical value y 内 of the koji center;

[0014] Step 3: Grind the whole distiller's yeast koji into powder and mix it evenly, use a hyperspectral instrument to scan the koji powder to obtain the spectral information x 总 of the koji powder, and obtain the physicochemical value y 总 of the koji powder by chemical methods;

[0015] Step 4: Repeat Steps 1-3 to obtain a data set;

[0016] Step 5: Establish a relationship model between the spectral information x 总 of the koji powder and the physicochemical value y 总 of the koji powder: y 总 = kx 总 ;

[0017] Step 6: Divide the data set into a training set and a test set, set the initial weight of each hyperspectral information processing branch convolutional layer of the distiller's yeast koji physicochemical value detection model to k, use the training set to train the distiller's yeast koji physicochemical value detection model, and use the test set to verify the distiller's yeast koji physicochemical value detection model.

[0018] Physicochemical value detection method for distiller's yeast koji, including:

[0019] Step 1: Obtain the spectral information of each surface of the distiller's yeast koji to be detected;

[0020] Step 2: Input the spectral information of each surface of the distiller's yeast koji to be detected into the trained distiller's yeast koji physicochemical value detection model to obtain the predicted physicochemical value of the koji center.

[0021] Furthermore, it also includes Step 3: Regularly obtain the actual physicochemical value of the koji center. When the proportion of the actual value inconsistent with the predicted value is higher than the threshold, adjust the data set and retrain the distiller's yeast koji physicochemical value detection model.

[0022] The beneficial effects of the present invention compared with the prior art are as follows: The physical and chemical value detection model of the koji core uses different branches of the convolutional neural network to process the hyperspectral data of each surface of the koji. Each branch contains an independent convolutional layer and pooling layer, enabling the model to comprehensively consider the data characteristics of different surfaces and their distance relationship with the koji core, thereby enhancing the robustness and generalization ability of the model. By establishing a linear relationship model between the spectral information of the koji powder and the physical and chemical values of the koji powder to obtain the coefficient k, when training the physical and chemical value detection model of the koji core, the coefficient k is used as the initial weight, improving the model training speed. Based on the physical and chemical value detection model of the koji core, non-destructive detection of the physical and chemical values of the koji core is realized. Brief Description of the Drawings

[0023] Figure 1 It is a schematic structural diagram of the physical and chemical value detection model of the koji core. Detailed Embodiments

[0024] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0025] A method for establishing a physical and chemical value detection model of the koji core includes:

[0026] Create a physical and chemical value detection model of the koji core with multi-modal data as the input and the physical and chemical values of the koji core as the output. The multi-modal data includes at least the hyperspectral information of each surface of the koji. The physical and chemical value detection model of the koji core includes: a plurality of hyperspectral information processing branches corresponding to each surface of the koji one by one. Each hyperspectral information processing branch includes a convolutional layer and a pooling layer. The feature vectors generated by all hyperspectral information processing branches are weighted and fused, and then the physical and chemical value prediction result of the koji core is output through a fully connected layer. Among them, the weight d i represents the actual distance from the i-th surface to the koji core, and n is the total number of surfaces of the koji. Specifically, the hyperspectral information includes the length and width of the hyperspectral image and the number of wavelengths of a single pixel, and the physical and chemical values include acidity, moisture content and starch content.

[0027] Hyperspectral imaging (HSI) is a further extension of near-infrared technology. It is a high-precision imaging technology based on fine spectroscopy. The data obtained by it is a three-dimensional hyperspectral data packet combining image space and spectrum, with rich interpretation information. This technology can be used for pixel-level positioning of targets that are similar in shape and color but different in material in the image, and at the same time, accurately calculate the super-visual information such as the material and component content of the target.

[0028] On this basis, this embodiment uses different branches of the convolutional neural network to process the hyperspectral data of each surface of the koji. Each branch contains an independent convolutional layer and pooling layer, and then determines the weights according to the actual distance from each surface of the koji to the koji center. After weighted summation, the prediction result of the physical and chemical values of the koji center is output through the fully connected layer, enabling the system to comprehensively consider the data characteristics of different surfaces and their distance relationship with the koji center, thereby enhancing the robustness and generalization ability of the model and realizing the non-destructive detection of the physical and chemical values of the koji center.

[0029] Furthermore, the physical and chemical values of the koji center are also affected by the temperature and humidity during the fermentation process of the koji. To improve the accuracy of the prediction result, the detection model of the physical and chemical values of the koji center also includes a feature extraction branch for generating a feature vector based on the temperature and humidity information during the fermentation process of the koji; splicing the feature vector with the weighted and fused features, and then outputting the prediction result of the physical and chemical values of the koji center through the fully connected layer.

[0030] A method for training a detection model of the physical and chemical values of the koji center includes:

[0031] Step 1: Use a hyperspectral near-infrared instrument to comprehensively scan each surface of the koji, collect the hyperspectral information x of each surface 表 , and record the geometric dimensions of the koji to obtain the actual distance from each surface to the koji center;

[0032] Step 2: Obtain the physical and chemical values y of the koji center 内 ;

[0033] Step 3: Grind the whole koji into powder and mix it evenly, use a hyperspectral instrument to scan the koji powder to obtain the spectral information x of the koji powder 总 , and obtain the physical and chemical values y of the koji powder through chemical methods 总 ;

[0034] Step 4: Repeat Steps 1-3 to obtain a data set;

[0035] Step 5: Establish a relationship model between the spectral information x of the koji powder 总 and the physical and chemical values y of the koji powder 总 : y 总 = kx 总 ; In this embodiment, the partial least squares regression (PLSR) method is used to establish the relationship model between the spectral information x of the powder 总 and the physical and chemical values y of the koji powder 总 .

[0036] Step 6: Divide the data set into a training set and a test set, set the initial weight of each branch convolutional layer of the detection model of the physical and chemical values of the koji center to k, use the training set to train the detection model of the physical and chemical values of the koji center, and use the test set to verify the detection model of the physical and chemical values of the koji center.

[0037] Taking the cube-shaped koji as an example, as Figure 1 shown, the koji core physical and chemical value detection model has 6 hyperspectral information processing branches. The input of each branch is (batch_size, height, width, channels), where batch_size corresponds to the number of batches (i.e., all samples are divided into corresponding batches and input into the model for training), height and width correspond to the length and width of the hyperspectral image, and channels correspond to the number of wavelengths of a single pixel. First, through a one-dimensional convolution, the convolution kernel size is set to channels, and the initial weight size uses k determined in the modeling in step 5. Using the coefficients of the PLSR algorithm as the initial weights, which reflects the correlation between the spectrum of the whole koji powder and the physical and chemical values. And for the physical and chemical values of the koji core, there is a certain relationship with the external physical and chemical value distribution. By setting the initial weights of the one-dimensional convolution kernel to k for model training, the model training speed can be effectively improved.

[0038] When the koji core physical and chemical value detection model includes a feature extraction branch, temperature and humidity information during the koji fermentation process is also obtained when acquiring the dataset.

[0039] The method for detecting the physical and chemical values of the koji core includes:

[0040] Step 1: Obtain the spectral information of each surface of the koji to be detected;

[0041] Step 2: Input the spectral information of each surface of the koji to be detected into the trained koji core physical and chemical value detection model to obtain the predicted physical and chemical values of the core; or input the spectral information of each surface of the koji to be detected and the temperature and humidity information during the koji fermentation process into the trained koji core physical and chemical value detection model to obtain the predicted physical and chemical values of the core.

[0042] Furthermore, it also includes Step 3: Regularly obtain the actual physical and chemical values of the core. When the proportion of the inconsistent actual value and predicted value is higher than the threshold, adjust the dataset and retrain the koji core physical and chemical value detection model.

Claims

1. A method for establishing a psychological value detection model for koji, characterized in that: include: Taking multimodal data as input and the physical and chemical values ​​of the core of the koji as output, a koji psychological value detection model is created. The multimodal data at least includes the hyperspectral information of each surface of the koji. The koji psychological value detection model is provided with multiple hyperspectral information processing branches corresponding to each surface of the koji. The feature vectors generated by all hyperspectral information processing branches are weightedly fused, and then the prediction result of the koji psychological value is output through the fully connected layer, where the weight of each branch is d i It represents the actual distance from the ith surface to the center of the koji, and n is the total number of surfaces of the koji.

2. The method for establishing a psychological value detection model for koji according to claim 1, characterized in that: Multimodal data also includes temperature and humidity information during the fermentation process of koji.

3. The method for establishing a psychological value detection model for koji according to claim 2, characterized in that: The koji psychological chemical value detection model is also equipped with a feature extraction branch, which is used to generate a feature vector based on the temperature and humidity information during the koji fermentation process; the feature vector is concatenated with the weighted fused features, and then the koji psychological chemical value prediction result is output through a fully connected layer.

4. The method for establishing a psychological value detection model for koji according to claim 1, characterized in that: Physical and chemical values ​​include acidity, moisture content and starch content.

5. The method for establishing a psychological value detection model for koji according to any one of claims 1 to 4, characterized in that: Hyperspectral information includes the length, width and number of wavelengths of a single pixel of the hyperspectral image.

6. A method for training a model for detecting the psychological value of a koji, characterized in that: include: Step 1: Use a hyperspectral near-infrared instrument to comprehensively scan each surface of the koji and collect hyperspectral information of each surface. 表 , and record the geometric dimensions of the koji to obtain the actual distance from each surface to the center of the koji; Step 2: Get the curve y 内 ; Step 3: Crush the whole koji into powder and mix evenly, and use a hyperspectral instrument to scan the koji powder to obtain the koji powder spectrum information x 总 , and obtain the physicochemical value of koji powder by chemical method 总 ; Step 4: Repeat steps 1-3 to obtain the dataset; Step 5: Establish the spectrum information x of Koji powder based on the data set 总 Physicochemical value of koji powder 总 Relationship model: y 总 =kx 总 ; Step 6: Divide the dataset into a training set and a test set, set the initial weight of the convolutional layer of each hyperspectral information processing branch of the koji psychological value detection model to k, use the training set to train the koji psychological value detection model, and use the test set to verify the koji psychological value detection model.

7. A method for detecting the psychological value of koji, characterized in that: include: Step 1: Obtain the spectral information of each surface of the wine yeast to be tested; Step 2: Input the spectral information of each surface of the koji to be tested into the trained koji psychological value detection model to obtain the predicted koji psychological value.

8. The method for detecting the psychological value of koji according to claim 7, characterized in that: It also includes step 3: regularly obtaining the actual physical and chemical values ​​of the koji core, and when the ratio of inconsistency between the actual value and the predicted value is higher than a threshold, adjusting the data set to retrain the koji koji physical and chemical value detection model.

Images