A method for predicting soy sauce yield based on Cantonese soy sauce fermentation process parameters

Abstract

The invention discloses a method for predicting the yield of soy sauce according to the process parameters of Cantonese soy sauce fermentation, comprising the following steps: S1 sets the values ​​of the natural oil evaluation index in the fermentation tank on the 15th, 45th and 90th day of feeding and the temperature index from the 15th to the 180th day after feeding; S2 normalizes the values ​​of the natural oil evaluation index and the temperature index respectively and then converts them into two three-dimensional matrices to obtain two index sequences; S3 inputs the obtained two index sequences into two long short-term memory neural networks respectively, and finally obtains two one-dimensional feature vectors; S4 splices and fuses the quality index of the fermentation tank raw material soybean with the two one-dimensional feature vectors obtained by S3 to form a long feature vector, and then reduces the dimension through a fully connected layer to output the predicted value of the soy sauce yield of each fermentation tank. The invention solves the problem that the existing soy sauce natural oil fermentation process is opaque and the yield is difficult to quantify according to the fermentation process, and has guiding significance for the control of the soy sauce fermentation production process.

Description

Technical Field

[0001] The invention relates to a method for predicting the yield of soy sauce according to the fermentation process parameters of Cantonese-style soy sauce. Background Art

[0002] Soy sauce plays an extremely important role in our daily life and food industry. With the improvement of the overall technical level, management level and people's living standards of our food industry, the output and output value of soy sauce in our country have been growing rapidly in recent years.

[0003] At present, 70% of the soy sauce products in China adopt the Cantonese high-salt dilute fermentation process. The production process of Cantonese soy sauce adopts the natural fermentation process under the sun and at night. The fermentation process is affected by multiple factors such as raw material ratio, strain, climate environment, equipment parameters, etc., and the quality fluctuates greatly.

[0004] The fermentation of Cantonese soy sauce is a multi-parameter, long-cycle, and dynamically changing process. Most of the equipment used is traditional non-standard, which makes it difficult to collect real-time data. Therefore, although the actual parameters of the material during the processing cycle, such as temperature, are changing values, they are always collected manually, and the retained data is only single-point data, and the representativeness of the data is very limited. The key quality indicators in the soy sauce fermentation process fluctuate with seasonal changes. For example, in terms of fermentation, the sauce formula needs to be adjusted according to climate changes, which further reduces the value of the collected single-point data.

[0005] In short, the fermentation process of Cantonese soy sauce has always been recorded manually, which has limited traceability data and is difficult to analyze, restricting the optimization and development of the fermentation process of Cantonese soy sauce. Even providing guidance for the optimization of the sauce formula in the same month of the next year is difficult. This is also the reason why the soy sauce yield in the first and fourth quarters of the year has always been significantly low. Summary of the invention

[0006] The prediction of soy sauce yield plays an important role in discovering deficiencies in the production process, reducing raw material conversion costs, optimizing the entire production process and achieving sustainable high-quality production. The main purpose of the present invention is to explore a method for predicting the soy sauce yield based on the fermentation process parameters of Cantonese soy sauce based on the relevant data currently recorded in the fermentation process of Cantonese soy sauce, so as to help discover the key parameters for improving the product yield and guide the regulation method of the soy sauce fermentation production process.

[0007] The current Cantonese soy sauce fermentation process requires that the fermentation of soy sauce needs to continue in the sun and dew for 180 days, and the fermentation tank is sampled and tested on the 15th, 45th and 90th days after feeding, to obtain the values ​​of natural oil fermentation indicators after the corresponding number of days of fermentation, including total acid (g / 100mL), amino acid nitrogen (g / 100mL), salt (g / 100mL) and pH value, etc., and the temperature of the fermentation tank is recorded every day starting from the day of feeding.

[0008] To achieve the purpose of the present invention, the technical scheme adopted is as follows: a method for predicting the yield of soy sauce according to the fermentation process parameters of Cantonese soy sauce, comprising the following steps:

[0009] S1 sets the values ​​of the natural oil evaluation index in the fermentation tank on the 15th, 45th and 90th days after feeding and the temperature index from the 15th to the 180th day after feeding;

[0010] S2 normalizes the values ​​of the natural oil evaluation index and the temperature index respectively to eliminate the dimension effect, and then converts them into two three-dimensional matrices including the number of fermentation tank samples, time series and characteristic indexes according to the characteristics of the two data, to obtain the natural oil evaluation index sequence and the temperature index sequence for the input of the neural network;

[0011] S3 inputs the two indicator sequences obtained by S2 into two long short-term memory neural networks to extract feature vectors, and finally obtains two one-dimensional feature vectors;

[0012] S4 concatenates and fuses the quality index of soybean, the raw material of the fermentation tank, with the two one-dimensional feature vectors obtained by S3 in the feature dimension to form a long feature vector, and then reduces the dimension through the fully connected layer to output the predicted value of the soy sauce yield of each fermentation tank;

[0013] The weight parameters of the two long short-term memory neural networks in step S3 are determined as follows:

[0014] S3-1 obtains historical values ​​of natural oil evaluation indicators in the fermentation tank on the 15th, 45th and 90th days after feeding and temperature indicators from the 15th to the 180th day after feeding;

[0015] S3-2 normalizes the values ​​of the natural oil evaluation index and the temperature index respectively to eliminate the dimension effect, and then converts them into two three-dimensional matrices including the number of fermentation tank samples, time series and characteristic indexes according to the characteristics of the two data, to obtain the natural oil evaluation index sequence and the temperature index sequence for the input of the neural network;

[0016] S3-3 inputs the two indicator sequences obtained in S3-2 into two long short-term memory neural networks to extract feature vectors, and finally obtains two one-dimensional feature vectors;

[0017] S3-4 combines the quality index of soybean, the raw material of the fermentation tank, with the two one-dimensional feature vectors obtained in S3-3 in the feature dimension to form a long feature vector, and then reduces the dimension through a fully connected layer to output the predicted value of the soy sauce yield of each fermentation tank;

[0018] S3-5 uses the loss function to measure the difference between the predicted value and the true value in S3-4, calculates the neural network loss through forward propagation, and then updates the weight parameters of the S3-3 neural network through the back propagation algorithm, and then repeats S3-3 to S3-5 until the loss function converges.

[0019] After training through S3-5, it is shown that the production process and product yield can be associated through S3-1 to S3-4. The present invention uses the characteristic of the neural network that can automatically extract features to seek the corresponding relationship between fermentation time, temperature and natural oil evaluation indicators, and by splicing the raw material indicator data of the fermentation tank, the initial state information of the raw materials is supplemented, so that the constructed model can more accurately describe the relationship between the process parameters and the soy sauce yield, and solve the problem that the existing soy sauce natural oil fermentation process is opaque and the yield is difficult to quantify according to the fermentation process.

[0020] As a preferred embodiment of the present invention:

[0021] The natural oil evaluation indicators are total acid (g / 100mL), amino acid nitrogen (g / 100mL), salinity (g / 100mL) and pH value, and the temperature indicators are the highest temperature (°C) and the lowest temperature (°C) of each day.

[0022] The loss function used in S3-5 is the SmoothL1 loss function, which can avoid the defects of inflection points and non-smoothness in model training. Its expression is:

[0023]

[0024] x represents the difference between the predicted value and the true value. This loss function is insensitive to outliers and abnormal values, and the gradient change is relatively smaller, which helps to improve the robustness of the model.

[0025] The quality indicators are total nitrogen (g / 100mL) and moisture (%).

[0026] In the above steps, MinMax is used for normalization to compress the data value to between 0 and 1. The specific expression is:

[0027]

[0028] xi represents the input vector. In order to quantify the degree of change of the index with the number of days, the values ​​of max and min are selected from the indicators of all the day sequences of all fermentation tanks.

[0029] The long short-term memory neural network described in the above steps is bidirectional, and can process the sequence obtained by S2 from two opposite directions, learn the dependence of the indicator on time, and enable the extracted deep features to better retain the implicit information of the indicator.

[0030] Beneficial effects:

[0031] The present invention uses the characteristic that a neural network can automatically extract features, and uses the neural network to seek the corresponding relationship between fermentation time, temperature and natural oil evaluation indexes, and by splicing the raw material index data of the fermentation tank, the initial state information of the raw materials is supplemented, so that the constructed model can more accurately describe the relationship between the process parameters and the soy sauce yield, and solve the problem that the existing soy sauce natural oil fermentation process is opaque and the yield is difficult to quantify according to the fermentation process. Through the method of the present invention, the association between the production process and the product yield can be established, so that the product yield can be predicted according to the range of the controllable parameters in the production process, and then the controllable production parameter interval that can obtain a stable yield is reversely deduced, thereby guiding stable production and increasing economic benefits. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is a flowchart of the present invention;

[0033] Figure 2 It is a framework diagram of a neural network in a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0034] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be further described below in conjunction with specific embodiments and drawings. Obviously, the following specific embodiments are only preferred embodiments of the present invention, rather than all optional embodiments. Based on the technical concept of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work should fall within the scope of protection of the present invention as long as they are within the scope of protection required by the claims of the present invention.

[0035] The fermentation of Cantonese soy sauce requires 180 days of continuous exposure to the sun and night dew. Generally, sampling tests will be carried out at the 15th, 45th and 90th day of fermentation after feeding the fermentation tank, respectively, to obtain the values ​​of natural oil fermentation indicators after the corresponding number of days of fermentation, including total acid (g / 100mL), amino acid nitrogen (g / 100mL), salt (g / 100mL) and pH value, to represent the changing trend of the fermentation degree. These four indicators are not independent of each other, and there is interaction between them. At the same time, starting from the day of feeding, the temperature changes of each fermentation tank are recorded every day.

[0036] The specific process of predicting the soy sauce yield based on the relevant data obtained from the above-mentioned Cantonese soy sauce fermentation process is as follows:

[0037] Step 1: Obtain historical values ​​of natural oil evaluation indicators in the fermentation tank on the 15th, 45th and 90th days after feeding and temperature indicators from the 15th to the 180th day after feeding, wherein the natural oil evaluation indicators are total acid (g / 100mL), amino acid nitrogen (g / 100mL), salt (g / 100mL) and pH value. For simplicity, the temperature indicators used are the highest temperature (°C) and the lowest temperature (°C) of each day.

[0038] In this step, only the temperature data after 15 days is obtained because the natural oil evaluation index in the present invention is recorded from the 15th day, and the temperature data from the 15th day of fermentation after feeding affects the index.

[0039] Step 2: The first step: normalize the values ​​of the natural oil evaluation index and the temperature index to eliminate the dimension effect. The normalization method is as follows:

[0040] Perform MinMax normalization, the expression is:

[0041]

[0042] xi represents the input vector. In order to quantify the degree of change of the index with the number of days, the values ​​of max and min are selected from the indicators of all the day sequences of all fermentation tanks.

[0043] Step 2: According to the characteristics of the two types of data, they are converted into two three-dimensional matrices including the number of fermentation tank samples N, time series (the natural oil index has three collection times, namely the 15th, 45th, and 90th days, and the temperature data is collected from the 15th to the 180th day, with 165 collection times) and characteristic indicators (there are 4 natural oil indicators and 2 time indicators, which are the highest and lowest temperatures of the day), and the natural oil evaluation index sequence and temperature index sequence are obtained for the input of the neural network model. Finally, for the natural oil fermentation index, it can be expanded to a sequence of N×3×4; for the temperature index, it can be expanded to a sequence of N×165×2.

[0044] Step 3: Input the two indicator sequences obtained in step 2 into two long short-term memory neural networks respectively to extract feature vectors and obtain two one-dimensional feature vectors.

[0045] The long short-term memory neural network used here is bidirectional, which can process the sequence obtained in step 2 from two opposite directions, learn the dependence of the indicator on time, and enable the extracted deep features to better retain the implicit information of the indicator. Of course, the present invention can also use a unidirectional long short-term memory neural network.

[0046] The long short-term memory neural network used in the present invention increments and discards the input information through the internal gate structure, including the forget gate, update gate, and output gate. Among them, the forget gate is used to control whether the information of the previous neuron needs to be discarded in the current layer. If the hidden state at time t is h t , then the forget gate can be expressed as:

[0047] f t =σ(W f *[h t-1 , x t ]+b f ),

[0048] It first connects the hidden state of the previous moment, and obtains a number of 0-1 through the sigmoid activation function, where 0 means all are discarded and 1 means all are retained. The numbers between them represent the retention ratio; the update gate integrates the past unit state information, the hidden information inside the unit at the old moment, and the new input data, which can be expressed as:

[0049] i t =σ(W i *[h t-1 , x t ]+b i );

[0050] The output gate can obtain the output value of the current unit and the hidden state value passed to the next unit. The hidden state contains relevant information of the previous input. The output value of the current unit can be expressed as:

[0051] o t =σ(W o *[h t-1 , x t ]+b o ),

[0052] The hidden state of the next unit can be expressed as:

[0053] h t =o t *tanh(C t ),

[0054] Among them, C t Indicates the unit status at the current moment.

[0055] The long short-term memory neural network in the present invention is implemented through the Pytorch framework. Pytorch is a manifold development framework in deep learning, which supports dynamic construction of neural networks and has strong scalability.

[0056] Step 4: For the two feature vectors output by the neural network in step 3, concatenate and fuse them in the feature dimension to form a long feature vector. In order to combine the quality information of the raw material soybean and form a coherent prediction of the production line, the quality index of soybean is concatenated on the basis of the concatenated long feature vector. The quality index used in the present invention includes total nitrogen (g / 100mL) and moisture (%), such as Figure 2 The concatenated final vector is reduced in dimension through a fully connected layer, and the predicted value of the soy sauce yield of each fermentation tank is output.

[0057] Step 5: Use the SmoothL1 loss function to measure the difference between the yield prediction value in step 4 and the actual value, calculate the neural network loss through forward propagation, and then update the weight parameters of the neural network in step 3 through the back propagation algorithm, and then repeat steps 3 to 5 until the loss function converges. After that, the soy sauce yield of each fermentation tank can be predicted by setting the relevant parameters in step 1 according to steps 2 to 4.

[0058] The present invention uses the SmoothL1 loss function to train the model because it can avoid the defects of inflection points and non-smoothness in model training. Its expression is:

[0059]

[0060] x represents the difference between the predicted value and the true value. This loss function is insensitive to outliers and abnormal values, and the gradient change is relatively smaller, which helps to improve the robustness of the model.

[0061] After training through step 5, it is shown that the production process and product yield can be associated through steps 1 to 4, so that the product yield can be predicted according to the range of controllable parameters in the production process, and then the controllable production parameter range that can obtain a stable yield can be reversely deduced, thereby guiding stable production and increasing economic benefits.

[0062] The present invention is based on 3600 production data from the Cantonese soy sauce fermentation production line, and divides the training set and the test set of the neural network according to a ratio of 7:3. After the weight parameters of the network are trained with the data of the training set, the accuracy is verified on the test set. The mean absolute error (MAE) of the yield prediction in the test set is finally obtained to be 0.109, while the mean absolute value deviation of the actual yield is 0.176. It can be seen that the yield prediction by the present invention has reference value.

[0063] Yield prediction is a key task in the soy sauce production process. The prediction of soy sauce yield plays an important role in discovering deficiencies in the production process, reducing raw material conversion costs, optimizing the entire production process, and achieving sustainable high-quality production. The present invention utilizes the characteristics of neural networks that can automatically extract features, and uses them to seek the corresponding relationship between fermentation time, temperature, and natural oil evaluation indicators, and by splicing fermentation tank raw material index data, the initial state information of the raw materials is supplemented, so that the constructed model can more accurately describe the relationship between process parameters and soy sauce yield, and solve the problem that the existing soy sauce natural oil fermentation process is opaque and the yield is difficult to quantify according to the fermentation process. Through the method of the present invention, the association between the production process and the product yield can be established, so that the product yield can be predicted according to the range of controllable parameters in the production process, and then the controllable production parameter interval that can obtain a stable yield is reversely deduced, thereby guiding stable production and increasing economic benefits.

[0064] The above are only preferred embodiments of the present invention. However, the protection scope of the patent of the present invention is not limited thereto. Equivalent replacements or changes made by those skilled in the art based on the technical solution of the patent of the present invention and its inventive concept are all within the protection scope of the patent of the present invention.

Claims

1. A method for predicting soy sauce yield based on Cantonese soy sauce fermentation process parameters, characterized in that: The steps include: S1 sets the values ​​of the natural oil evaluation index in the fermentation tank on the 15th, 45th and 90th days after feeding and the temperature index from the 15th to the 180th day after feeding; S2 normalizes the values ​​of the natural oil evaluation index and the temperature index respectively to eliminate the dimension effect, and then converts them into two three-dimensional matrices including the number of fermentation tank samples, time series and characteristic indexes according to the characteristics of the two data, to obtain the natural oil evaluation index sequence and the temperature index sequence for the input of the neural network; S3 inputs the two indicator sequences obtained by S2 into two long short-term memory neural networks to extract feature vectors, and finally obtains two one-dimensional feature vectors; S4 concatenates and fuses the quality index of soybean, the raw material of the fermentation tank, with the two one-dimensional feature vectors obtained by S3 in the feature dimension to form a long feature vector, and then reduces the dimension through the fully connected layer to output the predicted value of the soy sauce yield of each fermentation tank; The weight parameters of the two long short-term memory neural networks in step S3 are determined as follows: S3-1 obtains historical values ​​of natural oil evaluation indicators in the fermentation tank on the 15th, 45th and 90th days after feeding and temperature indicators from the 15th to the 180th day after feeding; S3-2 normalizes the values ​​of the natural oil evaluation index and the temperature index respectively to eliminate the dimension effect, and then converts them into two three-dimensional matrices including the number of fermentation tank samples, time series and characteristic indexes according to the characteristics of the two data, to obtain the natural oil evaluation index sequence and the temperature index sequence for the input of the neural network; S3-3 inputs the two indicator sequences obtained in S3-2 into two long short-term memory neural networks to extract feature vectors, and finally obtains two one-dimensional feature vectors; S3-4 combines the quality index of soybean, the raw material of the fermentation tank, with the two one-dimensional feature vectors obtained in S3-3 in the feature dimension to form a long feature vector, and then reduces the dimension through a fully connected layer to output the predicted value of the soy sauce yield of each fermentation tank; S3-5 uses the loss function to measure the difference between the predicted value and the true value in S3-4, calculates the neural network loss through forward propagation, and then updates the weight parameters of the S3-3 neural network through the back propagation algorithm, and then repeats S3-3 to S3-5 until the loss function converges; The natural oil evaluation index includes total acid, in g / 100mL, amino acid nitrogen, in g / 100mL, salt, in g / 100mL, and pH value. The temperature index is the highest and lowest temperatures of each day, in °C. The loss function used in S3-5 is the SmoothL1 loss function, which is expressed as: x represents the difference between the predicted value and the true value.

2. The method for predicting soy sauce yield according to Cantonese soy sauce fermentation process parameters according to claim 1, characterized in that: The quality indicators are total nitrogen and moisture, where the total nitrogen unit is g / 100mL and the moisture is expressed as a percentage.

3. The method for predicting soy sauce yield according to Cantonese soy sauce fermentation process parameters according to claim 1, characterized in that: It uses MinMax for normalization to compress the data value to between 0 and 1. The specific expression is: Among them, xi represents the input vector, and the values ​​of max and min are selected from the indicators of all day sequences of all fermentation tanks.

4. The method for predicting soy sauce yield according to Cantonese soy sauce fermentation process parameters according to claim 1, characterized in that: The long short-term memory neural network is a bidirectional long short-term memory neural network.

Images