Deep semi-supervised soft measurement method for carbon emission in beer production

Abstract

The invention discloses a deep semi-supervised soft measurement method for carbon emission in beer production, and the method specifically comprises the following steps: a semi-supervised soft prediction model based on a wavelet threshold noise reduction auto-encoder comprises three modules: a semi-supervised pre-training module, a supervised fine tuning module and a noise reduction module; the method comprises the following steps: recycling carbon dioxide in beer production, removing foams by using a foam remover, feeding gas from which the foams are removed into a washing tower through an air bag, removing water-soluble impurities and particles, removing water by using a precooler, removing residual water and peculiar smell in a deodorizer, filtering gas from the deodorizer by using a dust filter, and removing the residual water and peculiar smell in the deodorizer. Gas is cooled by a condenser and then can be sent to a storage tank to be stored, the total reserve of carbon dioxide is predicted based on SS-SDAE, and a semi-supervised pre-training model is proposed by adopting a soft measurement method based on a wavelet threshold noise reduction auto-encoder, so that the recovery amount of carbon dioxide in the beer production process is accurately predicted.

Description

Technical Field

[0001] The invention relates to the technical field of beer, and in particular to a deep semi-supervised soft measurement method for carbon emissions in beer production. Background Art

[0002] Beer is one of the most popular drinks in the world, and it has a wide range of consumers in various countries. However, the beer industry is a high-investment, high-consumption, and high-pollution industry, and the carbon emissions from its production process have a significant impact on the environment.

[0003] The stacked autoencoder algorithm is an unsupervised learning algorithm based on neural networks. When learning from actual industrial production data, it only uses unlabeled samples, but ignores labeled data. Moreover, the features learned through unsupervised learning cannot be guaranteed to be relevant to regression or classification tasks. The loss of information greatly limits the improvement of soft sensor prediction accuracy. At present, there are also some time series machine learning algorithms, but when applied to carbon emission recovery, the data requirements are high. If the available historical data is limited or poor, its performance will also be limited. At the same time, there are often outliers and noise in time series data, which may have an adverse effect on the performance of the model.

[0004] To this end, the present invention provides a deep semi-supervised soft measurement method for carbon emissions in beer production to solve the above problems. Summary of the invention

[0005] In view of the deficiencies in the prior art, the present invention provides a deep semi-supervised soft measurement method for carbon emissions in beer production, which solves the above problems.

[0006] To achieve the above objectives, the present invention is implemented by the following technical scheme: A deep semi-supervised soft measurement method for carbon emissions in beer production, including, specifically, the following steps:

[0007] Step 1: Semi-supervised soft prediction model based on wavelet threshold denoising autoencoder;

[0008] Step 2: Recovering carbon dioxide from beer production;

[0009] Step 3: Predict the total carbon dioxide storage based on SS-SDAE.

[0010] Preferably, the semi-supervised soft prediction model in step one includes three modules, which are a semi-supervised pre-training module, a supervised fine-tuning module and a denoising module.

[0011] Preferably, the semi-supervised pre-training module uses a semi-supervised stacked denoising autoencoder to learn label-related information through supervised learning with a labeled data set, and then put the learned label-related information into each hidden layer of each unsupervised learning to perform unsupervised learning. Unsupervised learning will use the features learned in supervised learning, that is, semi-supervised learning. The specific algorithm operation is to use labeled data for supervised learning to simultaneously learn the input variable x and the target variable y to obtain the corresponding reconstruction features. Unsupervised training on unlabeled data to learn reconstructed features of input variables Finally, the features learned from supervised and unsupervised training are substituted into the top hidden layer of the SS-SDAE model to predict the final target variable.

[0012] Preferably, the supervised fine-tuning module is a gradient descent algorithm based on back propagation (BP), which optimizes the weight W and the bias term b by iteratively updating parameters, and the objective function is as follows:

[0013]

[0014] Preferably, the denoising module uses a wavelet threshold algorithm to denoise the original data. The process of the wavelet threshold denoising algorithm is as follows:

[0015] Step 1: Select a wavelet with N layers to perform wavelet decomposition on the signal;

[0016] Step 2: After decomposition, select a suitable threshold and use the threshold function to quantize the coefficients of each layer;

[0017] Step 3: Reconstruct the signal using the processed coefficients.

[0018] Preferably, the recovery of carbon dioxide from beer production in step 1 specifically comprises the following steps:

[0019] Step 1: Remove the foam through the defoamer;

[0020] Step 2: The gas after the foam is removed passes through the air bag to the scrubber to remove water-soluble impurities and particles;

[0021] Step 3: Precooler removes moisture;

[0022] Step 4: Remove remaining moisture and odor in the deodorizer;

[0023] Step 5: Filter the gas coming out of the deodorizer through a dust filter;

[0024] Step 6: The gas is cooled through a condenser and then sent to a storage tank for storage.

[0025] Preferably, the prediction steps for predicting the total carbon dioxide reserves based on SS-SDAE in step 1 are as follows:

[0026] Step 1: Unlabeled training set Represents a sample set where no carbon dioxide storage data can be collected; there is a labeled training data set represents the sample set of CO2 storage data collected, where y is the collected CO2 storage data. In fact, the number of samples in the former is much larger than that in the latter;

[0027] Step 2: The weight matrix and bias vector W between the i-th and i-1-th hidden layers i (i=1,...,m) and b i (i=1,...,k); weight matrix and bias vectors Wo and bo of the last hidden layer and output layer; predicted total carbon dioxide storage value

[0028] Step 3: Create SS-DAE objective function model;

[0029] Step 4: Predict the performance evaluation function through the SS-DAE objective function model.

[0030] Preferably, the objective function of the SS-DAE objective function model is as follows:

[0031]

[0032] The model prediction performance evaluation function is as follows:

[0033]

[0034] Beneficial Effects

[0035] The present invention provides a deep semi-supervised soft measurement method for carbon emissions in beer production. Compared with the prior art, it has the following beneficial effects:

[0036] This deep semi-supervised soft measurement method for carbon emissions in beer production uses a soft measurement method based on a wavelet threshold denoising autoencoder and proposes a semi-supervised pre-training model, thereby accurately predicting the amount of carbon dioxide recovered during the beer production process. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a schematic diagram of the process of the present invention;

[0038] Figure 2 4 is a model framework diagram of the semi-supervised denoising stacked autoencoder of the present invention;

[0039] Figure 3It is a carbon dioxide recovery process diagram of the present invention. DETAILED DESCRIPTION

[0040] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0041] See also Figure 1-Figure 3 , a deep semi-supervised soft sensing method for carbon emissions in beer production, specifically comprising the following steps:

[0042] Step 1: Semi-supervised soft prediction model based on wavelet threshold denoising autoencoder;

[0043] Step 2: Recovering carbon dioxide from beer production;

[0044] Step 3: Predict the total carbon dioxide storage based on SS-SDAE.

[0045] Embodiment 1:

[0046] See also Figure 1-Figure 3 The semi-supervised soft prediction model in step 1 includes three modules, namely, the semi-supervised pre-training module, the supervised fine-tuning module and the denoising module. The semi-supervised pre-training module uses a semi-supervised stacked denoising autoencoder to learn label-related information through supervised learning with a labeled data set, and then puts the learned label-related information into each hidden layer of each unsupervised learning to perform unsupervised learning. Unsupervised learning will use the features learned through supervised learning, that is, semi-supervised learning. The specific algorithm operation is to use labeled data for supervised learning to simultaneously learn the input variable x and the target variable y to obtain the corresponding reconstruction features. Unsupervised training on unlabeled data to learn reconstructed features of input variables Finally, the features learned from supervised and unsupervised training are substituted into the top hidden layer of the SS-SDAE model to predict the final target variable;

[0047] The model framework of semi-supervised denoising stacked autoencoder. In this model, x is the data after wavelet threshold denoising. Inputting the denoised data x into the first AE will obtain a hidden layer feature d1. Using d1 as input, input it into the second AE to obtain d2, and so on until a dm is obtained. After each AE, a weight matrix and a bias vector {W, b} will be obtained. For labeled signals, the input variables and output feature variables will be optimized at the same time to obtain (for the first AE) or (for the second to m AEs). For unlabeled data, only the input variables will be optimized to obtain (for the first AE) or (for the second to m AEs). Therefore, the above model can train labeled and unlabeled data at the same time, which is equivalent to performing supervised learning and unsupervised learning at the same time, which is the so-called semi-supervised soft measurement. The purpose of supervised learning is to ensure the extraction of label association information during the learning process, while the purpose of unsupervised learning is to retain the information of unlabeled data. Taking the first AE as an example, the unlabeled data set xw and the labeled data set (xj, yj) are used to calculate respectively and The expression of is as follows: (1) and (2), and the loss function of SS-SDAE is as follows:

[0048]

[0049] Formula (3) is equivalent to formula (4)

[0050]

[0051] in is {W1,b1}, is {W′ 1,x ,b′ 1,x}, δ is {W′ 1,y ,b′ 1,y}, μ is the weighting coefficient. The first term in (3) and (4) is the reconstruction error of unlabeled input, and the second term is the reconstruction error and feature prediction error of labeled input.

[0052] The supervised fine-tuning module is a gradient descent algorithm based on back propagation (BP), which optimizes the weight W and bias term b by iteratively updating parameters. The objective function is as follows;

[0053]

[0054] The purpose of this module is to find the parameter set {W i ,W i′ ,b i ,b i′}, the optimization method used is: after executing forward propagation during training, execute the supervised fine-tuning procedure based on the BP gradient descent algorithm. The basic idea of ​​the supervised fine-tuning module is to use labeled data, that is, supervised data, for learning, and use the learned information in unsupervised training. The denoising module uses the wavelet threshold algorithm to denoise the original data. The process of the wavelet threshold denoising algorithm is as follows:

[0055] Step 1: Select a wavelet with N layers to perform wavelet decomposition on the signal;

[0056] Step 2: After decomposition, select a suitable threshold and use the threshold function to quantize the coefficients of each layer;

[0057] Step 3: Reconstruct the signal using the processed coefficients.

[0058] Embodiment 2:

[0059] See also Figure 1-Figure 3 In step 1, the recovery of carbon dioxide from beer production specifically includes the following steps:

[0060] Step 1: Remove the foam through the defoamer;

[0061] Step 2: The gas after the foam is removed passes through the air bag to the scrubber to remove water-soluble impurities and particles;

[0062] Step 3: Precooler removes moisture;

[0063] Step 4: Remove remaining moisture and odor in the deodorizer;

[0064] Step 5: Filter the gas coming out of the deodorizer through a dust filter;

[0065] Step 6: The gas is cooled through a condenser and then sent to a storage tank for storage;

[0066] The entire carbon dioxide recovery process is actually to collect the CO2 produced by brewing through a filtration system. Specifically, the gas produced by the fermentation tank first passes through a defoamer to remove the foam, and then the gas after the foam is removed passes through an air bag to a scrubber, where water-soluble impurities and particles are removed. The gas coming out of the scrubber is pressurized by a compressor, and then goes to a precooler to remove most of the moisture, and then goes to a deodorizer to remove the remaining moisture and odor. The gas coming out of the deodorizer must pass through a dust filter, and finally the gas is cooled by a condenser and sent to a storage tank for storage. The CO2 in the storage tank only needs to be purified by a simple purifier and can be directly sent to various places in the factory where CO2 is needed. It should be noted that the starting conditions for recovering carbon dioxide are: from the 16th hour of fermentation or when the carbon dioxide concentration is greater than 99.5%.

[0067] Embodiment three:

[0068] See also Figure 1-Figure 3 , the prediction steps for total carbon dioxide reserves based on SS-SDAE in step 1 are as follows:

[0069] Step 1: Unlabeled training set Represents a sample set where no carbon dioxide storage data can be collected; there is a labeled training data set represents the sample set of CO2 storage data collected, where y is the collected CO2 storage data. In fact, the number of samples in the former is much larger than that in the latter;

[0070] Step 2: The weight matrix and bias vector W between the i-th and i-1-th hidden layers i (i=1,...,m) and b i (i=1,...,k); weight matrix and bias vectors Wo and bo of the last hidden layer and output layer; predicted total carbon dioxide storage value

[0071] Step 3: Create SS-DAE objective function model;

[0072] Step 4: Predict the performance evaluation function through the SS-DAE objective function model;

[0073] First layer pre-training SS-DAE: use unlabeled training dataset by minimizing the loss function And a labeled training dataset Pre-train the first layer of SS-DAE to obtain {W1,b1} and {W ′ 1,b ′ 1}, then use the data Calculate the forward propagation from the input layer to the first hidden layer to obtain the hidden features of the first hidden layer The loss function is as follows:

[0074]

[0075] The second layer of pre-training SS-DAE: using unlabeled training dataset by minimizing the loss function And a labeled training dataset Pre-train the second layer SS-DAE to obtain {W2,b2} and {W ′ 2,b ′ 2}, and then use the previously trained Calculate the forward propagation from the first hidden layer to the second hidden layer to obtain the hidden features of the second hidden layer The loss function is as follows:

[0076]

[0077] Repeat the above steps until the mth layer of SS-DAE: After training, {Wm,bm} and {W ′ m ,b ′ m}, and then use the previously trained Calculate the forward propagation from the m-1th layer to the mth hidden layer to obtain the hidden features of the mth layer The loss function is as follows:

[0078]

[0079] After obtaining the training result of the last SS-DAE, the entire SS-SDAE training is completed;

[0080] After the pre-training, the feature output layer learned by supervised learning and unsupervised learning is added to the top hidden layer of SS-SDAE to predict the target variable; the trained {Wm, bm} is used to initialize the soft sensor network of SS-SDAE; the weights and bias terms {Wo, bo} of the output layer are initialized with random values; Continuously iterate forward and backward propagation to fine-tune the relevant parameters of the entire network.

[0081] The model prediction performance evaluation function is as follows:

[0082]

[0083] Meanwhile, the contents not described in detail in this specification belong to the prior art known to those skilled in the art.

[0084] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0085] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A deep semi-supervised soft measurement method for carbon emissions in beer production, characterized by: The specific steps include: Step 1: Semi-supervised soft prediction model based on wavelet threshold denoising autoencoder; Step 2: Recovering carbon dioxide from beer production; Step 3: Predict the total carbon dioxide storage based on SS-SDAE.

2. The deep semi-supervised soft sensing method for carbon emissions in beer production according to claim 1 is characterized in that: The semi-supervised soft prediction model in step 1 includes three modules, which are a semi-supervised pre-training module, a supervised fine-tuning module and a denoising module.

3. The deep semi-supervised soft sensing method for carbon emissions in beer production according to claim 2, characterized in that: The semi-supervised pre-training module uses a semi-supervised stacked denoising autoencoder to learn label-related information through supervised learning with a labeled data set, and then puts the learned label-related information into each hidden layer of each unsupervised learning to perform unsupervised learning. Unsupervised learning uses the features learned through supervised learning, i.e. semi-supervised learning. The specific algorithm operation is to use labeled data for supervised learning to simultaneously learn the input variable x and the target variable y to obtain the corresponding reconstructed features. Unsupervised training on unlabeled data to learn reconstructed features of input variables Finally, the features learned from supervised and unsupervised training are substituted into the top hidden layer of the SS-SDAE model to predict the final target variable.

4. The deep semi-supervised soft sensing method for carbon emissions in beer production according to claim 2, characterized in that: The supervised fine-tuning module is a gradient descent algorithm based on back propagation (BP), which optimizes the weight W and the bias term b by iteratively updating the parameters. The objective function is as follows:

5. The deep semi-supervised soft sensing method for carbon emissions in beer production according to claim 2, characterized in that: The denoising module uses a wavelet threshold algorithm to denoise the original data. The process of the wavelet threshold denoising algorithm is as follows: Step 1: Select a wavelet with N layers to perform wavelet decomposition on the signal; Step 2: After decomposition, select a suitable threshold and use the threshold function to quantize the coefficients of each layer; Step 3: Reconstruct the signal using the processed coefficients.

6. The deep semi-supervised soft sensing method for carbon emissions in beer production according to claim 1, characterized in that: The step 1 of recovering carbon dioxide from beer production specifically includes the following steps: Step 1: Remove the foam through the defoamer; Step 2: The gas after the foam is removed passes through the air bag to the scrubber to remove water-soluble impurities and particles; Step 3: Precooler removes moisture; Step 4: Remove remaining moisture and odor in the deodorizer; Step 5: Filter the gas coming out of the deodorizer through a dust filter; Step 6: The gas is cooled through a condenser and then sent to a storage tank for storage.

7. The deep semi-supervised soft sensing method for carbon emissions in beer production according to claim 1, characterized in that: The prediction steps for predicting the total carbon dioxide reserves based on SS-SDAE in step 1 are as follows: Step 1: Unlabeled training set {X w } w＝1...Nw , represents the sample set for which no carbon dioxide storage data can be collected; there is a labeled training data set {x j ,y j } j＝1...Nj , represents the sample set of collected carbon dioxide storage data, where y is the collected carbon dioxide storage data. In fact, the sample size of the former is much larger than that of the latter; Step 2: The weight matrix and bias vector W between the i-th and i-1-th hidden layers i (i=1,...,m) and b i (i=1,...,k); weight matrix and bias vectors Wo and bo of the last hidden layer and output layer; predicted total carbon dioxide storage value Step 3: Create SS-DAE objective function model; Step 4: Predict the performance evaluation function through the SS-DAE objective function model.

8. The deep semi-supervised soft sensing method for carbon emissions in beer production according to claim 7, characterized in that: The objective function of the SS-DAE objective function model is as follows: The model prediction performance evaluation function is as follows:

Images