Oil flavor production line state monitoring method and device based on deep learning

Abstract

The invention provides an oil flavor production line state monitoring method and device based on deep learning, and the method comprises the steps: creating different process environments through the design of a production line flow, the adjustment of the thickness of a material, the adjustment of the temperature of an outlet, and the like, and obtaining sufficient process data and flavor data; an h5 model capable of describing flavor dimensions and corresponding process data is obtained through regression modeling, and the model is transplanted into a self-developed edge end controller to participate in related flavor control; inputting the empirically converted flavor vector data as an input instruction into a flavor large model to obtain a processing technology instruction related to the water content of the material; and the water content of the material and other related process parameters are controlled through a self-research fuzzy double closed-loop algorithm, so that the functions of adjusting the flavor of the oil material and monitoring the process state according to manual selection are realized. According to the method, big data linkage flavor data and process data are fully utilized, the production line is optimized, the production efficiency is improved, and a direction is provided for application of new products and new processes.

Description

Technical Field

[0001] The present invention belongs to the technical field of oil flavor conditioning, and in particular relates to a method and device for monitoring the state of an oil flavor production line based on deep learning. Background Art

[0002] Microwave conditioning technology plays an important role in flavor conditioning and moisture content control in herbaceous or woody oils such as rapeseed, peanuts, and soybeans. By evenly arranging magnetrons on the microwave resonant cavity to establish a controllable electromagnetic field, the water molecules in the material in the resonant cavity are quickly rubbed and evenly heated. Compared with traditional resistive heating methods, microwave heating is more uniform, and by adjusting the residence time of the material in the resonant cavity, it is also more convenient to achieve precise temperature control of the material. In addition to regulating the moisture content of the material, the microwave conditioner also controls the flavor of the material by adjusting the rising curve of the moisture content of the material. Related technologies require the use of electronic noses, self-developed edge intelligent controllers, Dell deep learning workstations, process large models and other tools to achieve the linkage of flavor data and process data and flavor customization for specific rapeseed.

[0003] Existing processes make insufficient use of big data, the relevant flavor data lacks linkage with process data, and there is a lack of digital reconstruction of the aroma-generating mechanisms of relevant processes. Technical bottlenecks restrict the creation of new products and new processes. Summary of the invention

[0004] The technical problem to be solved by the present invention is: to provide a method and device for monitoring the status of an oil flavor production line based on deep learning, which is used to adjust the oil flavor and monitor the process status according to manual selection during the microwave conditioning process of the oil.

[0005] The technical solution adopted by the present invention to solve the above technical problems is: a method for monitoring the state of an oil flavor production line based on deep learning, comprising the following steps:

[0006] S1: Collect process data and flavor data in the oil production environment in real time online through sensors and upload them to the cloud database;

[0007] S2: Cut and process the process data in the cloud database based on the time axis, perform multivariate nonlinear regression modeling on the process data and flavor data of the oil based on deep learning, and obtain a flavor big data model;

[0008] S3: Build an edge controller and transplant the flavor big data model to the edge controller to form an IP core for flavor control;

[0009] S4: Monitor the production line status and adjust the flavor of oil products through the edge controller that has transplanted the flavor big data model.

[0010] According to the above scheme, before step S1, the following steps are also included:

[0011] Use ASCII code communication protocol and modular communication method to handshake the local and cloud communication protocols, and configure the cloud protocol with one click;

[0012] The oil production line is modified to collect crude oil samples under different processing environment: including installing scent holes on the microwave tempering tank, and arranging sampling valves after the crude oil pump and before the solenoid valve in the refining pipeline.

[0013] Furthermore, in the step S1, the process data include different oil thicknesses, press current, press temperature and pressure, microwave outlet temperature, moisture content of the material entering the press, microwave speed and press speed.

[0014] Furthermore, in step S1, the specific steps of collecting flavor data are:

[0015] An air collection pipeline is connected to the aroma collection hole and connected to a buffer bottle for air pressure balance; the other end of the buffer bottle is connected to the collection tube of the electronic nose to measure the flavor data;

[0016] Crude oil samples were collected at the sampling valve, and the flavor components of the crude oil were qualitatively and semi-quantitatively analyzed by combining electronic nose and GC-MS.

[0017] According to the above scheme, in step S2, the specific steps are:

[0018] S21: Use the Spearman matrix to mine the correlation between flavor data and process variables and find the two process variables with the highest correlation with flavor data;

[0019] S22: standardizing the flavor data and the process data so that the processed data is scaled to a distribution with a mean of 0 and a standard deviation of 1;

[0020] S23: A GRU model associated with a multivariate time series is used to generate a nonlinear regression model for flavor characteristics and process characteristics, and the flavor data and process data are fitted according to the time axis.

[0021] Furthermore, in step S23, the specific steps are:

[0022] Build and train the GRU model to generate the h5 neural network regression model for describing the oil flavor model;

[0023] Draw a MAE curve graph;

[0024] The model is used to predict the process data under different flavor requirements, the prediction results are de-standardized, and a curve chart of the actual value and predicted value of the required process requirements is drawn.

[0025] According to the above scheme, in step S3, the specific steps are:

[0026] S31: Build the hardware platform of the edge controller, import the hardware settings into the h5 neural network regression model, and complete the underlying driver configuration, including configuring the communication protocol between the production line and the cloud;

[0027] S32: Modify the application layer according to actual needs, and design a human-computer interaction interface to determine the cause of the fault according to the status of the oil production line, including determining the fault status of the microwave magnetron and the refining electric valve;

[0028] S33: Testing the h5 neural network regression model on the edge controller;

[0029] Denormalize the predicted value data output by the h5 neural network regression model to obtain the predicted current value;

[0030] The following curves of the predicted current value and the actual current value are drawn and compared to obtain the evaluation results of the h5 neural network regression model.

[0031] According to the above scheme, in step S4, the specific steps are:

[0032] S41: The operator selects the type of oil to be produced and the corresponding flavor;

[0033] S42: adjusting the flavor data dimension through the edge controller according to experience, and inputting it into the flavor big model;

[0034] S43: The flavor model predicts process parameters, uses the water loop and temperature loop for dual-loop control, and outputs control instructions;

[0035] S44: The edge controller outputs the control instruction to the actuator of the oil production line to achieve the regulation of the oil flavor.

[0036] Furthermore, in step S43, the specific steps of the dual-loop control method are:

[0037] The dual-loop control method uses the water loop as the outer loop and the temperature loop as the inner loop; the input of the water loop is assumed to be the water content command M re'f , the moisture content of the material entering the press is fed back to M fdb , the output is the material temperature instruction T out ,but:

[0038] T out =fuzzy(M ref ,M fdb );

[0039] The input of the temperature loop is the temperature command T output by the water loop. out , the temperature feedback T of the microwave devicefdb , the output is the microwave belt speed R ref ,but

[0040] R ref =fuzzy(T out ,T fdb ).

[0041] A deep learning-based oil flavor production line status monitoring device includes a sensor, an edge controller, an actuator and a cloud database; it also includes a processor and a memory, the memory stores computer instructions, and the processor is used to execute the computer instructions stored in the memory. When the computer instructions are executed by the processor, the system implements the steps of a deep learning-based oil flavor production line status monitoring method.

[0042] The beneficial effects of the present invention are:

[0043] 1. The present invention discloses a method and device for monitoring the state of an oil flavor production line based on deep learning. By designing the production line process, different process environments are created by adjusting the material thickness and the outlet temperature, and sufficient process data and flavor data are obtained. The h5 model that can describe the flavor dimension and the corresponding process data is obtained through regression modeling, and the model is transplanted into a self-developed edge controller to participate in related flavor control. The flavor vector data converted from experience is input as an input instruction into the flavor big model to obtain processing instructions including material moisture content. The material moisture content and other related process parameters are controlled through a self-developed fuzzy double closed-loop algorithm, thereby realizing the functions of adjusting the oil flavor and monitoring the process state according to manual selection.

[0044] 2. Based on the flavor big model and edge-end intelligent controller, the present invention can customize the flavor of processed oil according to the oil processing and flavor big data, produce oil with different flavors according to the different types of processed materials, and make full use of big data to link flavor data and process data, realizing the upgrade from experience judgment to digital modeling and then to software IP. It optimizes the production line, improves production efficiency, and provides direction for the application of new products and new processes. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 2 is a diagram of the communication configuration of the production line before and after optimization of an embodiment of the present invention.

[0046] Figure 2 1. It is a flavor collection position diagram of a tempering tank according to an embodiment of the present invention.

[0047] Figure 3 It is a crude oil flavor collection position diagram of an embodiment of the present invention.

[0048] Figure 4It is a comparison chart between the actual value and the true value of the predicted data in an embodiment of the present invention.

[0049] Figure 5 This is a diagram of the flavor control interface of the edge controller of an embodiment of the present invention.

[0050] Figure 6 This is a diagram of the communication configuration interface of the edge controller production line of an embodiment of the present invention.

[0051] Figure 7 1 is a circuit diagram of an edge controller according to an embodiment of the present invention.

[0052] Figure 8 It is a diagram of automatic generation of STM32CbuMX code according to an embodiment of the present invention.

[0053] Fig. 9 This is a prediction value graph of a deep learning workstation according to an embodiment of the present invention.

[0054] Fig.10 It is a prediction curve diagram of edge devices according to an embodiment of the present invention.

[0055] Fig.11 This is an oil flavor control topology diagram based on deep learning in an embodiment of the present invention.

[0056] Fig.12 is a flow chart of an embodiment of the present invention. DETAILED DESCRIPTION

[0057] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0058] Example 1

[0059] See also Fig.12 , an embodiment of the present invention comprises the following steps:

[0060] 1. Collect process data and flavor data online

[0061] The oil moisture sensor, temperature sensor, pressure sensor and other equipment are used to collect process data in the oil production environment in real time online, and the relevant process data and communication protocol are directly sent to the cloud database through the edge controller. At the same time, the refining pipeline is modified to facilitate the collection of crude oil samples under different processing technology environments. At the same time, the electronic nose is used to collect the flavor of the materials in the microwave tempering tank in real time online to obtain relevant flavor data.

[0062] 1.1 Collecting oil processing data

[0063] The oil processing data is uploaded through the edge controller. Before the normal data transmission, the communication protocol handshake configuration needs to be executed. Since the original protocol transmission method is relatively complex, the relevant communication protocols are simplified here. The original production line communication configuration is often determined by the on-site PLC programmers according to the actual situation of the production line. The on-site debugging personnel write the program according to the actual installation and wiring conditions, and then write the communication protocol document. The written communication protocol is then passed to the cloud communication configuration personnel in the head office, who will configure it and conduct a handshake test with the on-site personnel. The debugging process is relatively cumbersome. If multiple production lines are debugged at the same time, there will often be a shortage of relevant manpower. The intelligent controller adopts ASCII code communication protocol and modular communication method to achieve extremely simplified communication configuration. Related communication methods such as Figure 1 shown.

[0064] The relevant communication protocol realizes one-click configuration of the cloud protocol, and then the relevant process data can be uploaded to the cloud database normally.

[0065] In order to collect oil flavor data, the oil production line was remodeled at two locations:

[0066] 1) Install the aroma collection hole on the microwave tempering tank. See the relevant installation position for details. Figure 2 :

[0067] 2) The second sampling point is arranged on the refining pipeline. The sampling valve is located after the crude oil pump and before the solenoid valve. The relevant installation position is shown in Figure 3 As shown:

[0068] In order to realize oil collection under different process environments, several different process environments are designed here. The relevant experimental arrangements are shown in Table 1:

[0069] Table 1 Process sampling table

[0070]

[0071]

[0072] 1.2 Obtaining flavor data

[0073] 1.2.1 Determine the flavor of the first sampling point

[0074] For the flavor data of the first sampling point on the line, the electronic nose real-time sampling and detection method is adopted. The gas collection pipeline is connected to the fragrance collection hole, and then connected to the buffer bottle for air pressure balance. The other end of the buffer bottle is connected to the electronic nose collection tube for flavor measurement.

[0075] 1.2.2 Processing samples from the second sampling point

[0076] The crude oil samples were collected from the second sampling point, and the flavor components of the crude oil were subsequently qualitatively and semi-quantitatively analyzed by GC-MS in combination with the electronic nose. The detailed process of flavor determination will be introduced below.

[0077] 2. Multivariate regression modeling based on deep learning

[0078] Based on the time axis, the relevant data in the processing technology database are cut, and the phyton database is used to perform multivariate nonlinear regression modeling on the oil process data and oil flavor data, and the relevant h5 model is generated.

[0079] 2.1 Mining data related to process and flavor variables

[0080] Since the flavor data and process data collected by the electronic nose need to rely on the time axis for fitting, it is proposed to use the GRU model related to the multivariate time series to generate a nonlinear regression model for flavor characteristics and process characteristics.

[0081] Since the relevant data consists of a ten-dimensional flavor vector and a variety of process data, it is necessary to find out which process data the flavor dimension is related to before modeling. Here, the Spearman matrix is ​​used to mine the correlation between process and flavor variables. The relevant code is as follows:

[0082] import pandas as pd

[0083] from scipy.stats import spearmanr

[0084] # Load data

[0085] data = pd.read_csv('5mm_watery_flavor_7.csv')

[0086] # Calculate the Spearman rank correlation coefficient between all variables

[0087] #Note: It is assumed here that the last column of the data is the target variable and the rest are feature variables

[0088] correlation_matrix,p_value_matrix=spearmanr(data)

[0089] #Convert the correlation coefficient matrix and P value matrix to DataFrame for better viewing and analysis

[0090] correlation_df=pd.DataFrame(correlation_matrix,index=data.columns,columns=data.columns)

[0091] p_value_df=pd.DataFrame(p_value_matrix,index=data.columns,columns=data.columns)

[0092] #Print Spearman rank correlation coefficient matrix

[0093] print("Spearman Rank Correlation Coefficient Matrix:")

[0094] print(correlation_df)

[0095] #Print P value matrix

[0096] print("\nP-value Matrix:")

[0097] print(p_value_df)

[0098] After exploration, we found the two process variables that had the highest correlation with the flavor data.

[0099] 2.2 Multivariate nonlinear regression modeling of relevant data

[0100] Since process data involves different dimensions and the numerical distribution ranges vary significantly, it is necessary to convert the data into a distribution with the same scale, a mean μ of 0, and a standard deviation σ of 1, i.e., standardization. The StandardScaler algorithm is a preprocessing method for data standardization, and the relevant algorithm has relevant database support in Python.

[0101] The specific steps for standardization using the StandardScaler algorithm are as follows:

[0102] First, calculate the mean and standard deviation of the data: find the mean of the sample data (such as formula (1)), find the sample variance (such as formula (2)), take the square root of the variance, and get the standard deviation (such as formula (3)).

[0103]

[0104] Where: X i is the sample value, μ is the sample mean, S 2 is the sample variance, and σ is the standard deviation. Data standardization: For each data point, subtract the mean and divide it by the standard deviation. The data processed above is scaled to a distribution with a mean of 0 and a standard deviation of 1 (as shown in formula (4)).

[0105] y i =(x i -μ) / σ (4)

[0106] After the test data is standardized, it is further used for regression modeling. Here is a python code, the input feature is 10-dimensional flavor data, the output feature is 2-dimensional process feature (including water and current), and the related h5 process model is output.

[0107] import pandas as pd

[0108] import numpy as np

[0109] from sklearn.model_selection import train_test_split

[0110] from sklearn.preprocessing import MinMaxScaler

[0111] import matplotlib.pyplot as plt

[0112] from tensorflow.keras.models import Sequential

[0113] from tensorflow.keras.layers import GRU,Dense

[0114] from tensorflow.keras.optimizers import RMSprop

[0115] #Read training and testing data

[0116] train_data=pd.read_csv('5mm_2_train.csv')

[0117] # Make sure the file name and path are correct

[0118] test_data=pd.read_csv('5mm_2_test.csv')

[0119] # Split features and target variables

[0120] X_train=train_data.iloc[:,:10].values

[0121] #Assume that the first 10 columns are features

[0122] y_train=train_data.iloc[:,10:12].values

[0123] #Assume that the next 2 columns are target variables

[0124] X_test=test_data.iloc[:,:10].values

[0125] y_test=test_data.iloc[:,10:12].values

[0126] #Data normalization

[0127] scaler_X = MinMaxScaler()

[0128] scaler_y = MinMaxScaler()

[0129] X_train_scaled=scaler_X.fit_transform(X_train)

[0130] X_test_scaled=scaler_X.transform(X_test)

[0131] y_train_scaled=scaler_y.fit_transform(y_train)

[0132] y_test_scaled=scaler_y.transform(y_test)

[0133] #Reshape the data into the format required by the GRU model

[0134] X_train_scaled=X_train_scaled.reshape((X_train_scaled.shape[0],1,X_train_scaled.shape[1]))

[0135] X_test_scaled=X_test_scaled.reshape((X_test_scaled.shape[0],1,X_test_scaled.shape[1]))

[0136] ###Step 3: Build and train the GRU model

[0137] model=Sequential([

[0138] GRU(64,input_shape=(X_train_scaled.shape[1],X_train_scaled.shape[2])),Dense(2)

[0139] #Two outputs correspond to two dependent variables])

[0140] model.compile(optimizer=RMSprop(learning_rate=0.12),loss='mean_squared_error',metrics=['mae'])

[0141] history=model.fit(X_train_scaled,y_train_scaled,epochs=100,batch_size=32,validation_split=0.2,verbose=1)

[0142] ###Step 4: Draw the MAE curve

[0143] plt.figure(figsize=(12,4))

[0144] plt.subplot(1,2,1)

[0145] plt.plot(history.history['mae'],label='Train MAE')

[0146] plt.plot(history.history['val_mae'],label='Validation MAE')

[0147] plt.title('Model MAE')

[0148] plt.xlabel('Epoch')

[0149] plt.ylabel('MAE')

[0150] plt.legend()

[0151] ###Step 5: Model evaluation and prediction graph

[0152] #Use the model to make predictions

[0153] y_pred_scaled=model.predict(X_test_scaled)

[0154] #Denormalize the prediction results

[0155] y_pred=scaler_y.inverse_transform(y_pred_scaled)

[0156] # Draw a curve graph of the true value and the predicted value

[0157] plt.subplot(1,2,2)

[0158] for iin range(y_test.shape[1]):

[0159] plt.plot(y_test[:,i],label=f'True Value{i+1}')

[0160] plt.plot(y_pred[:,i],label=f'Predicted Value{i+1}',linestyle='--')

[0161] plt.title('True vs Predicted Values')

[0162] plt.xlabel('Sample Index')

[0163] plt.ylabel('Value')

[0164] plt.legend()

[0165] plt.tight_layout()

[0166] plt.show()

[0167] The relevant code implements the feature extraction, standardization, model generation, prediction, and evaluation of the mosture.csv file. The predicted values ​​of the relevant prediction variables can be found in Figure 4 , the above code also generates the corresponding h5 model for the next step.

[0168] 3. Edge controller design and large model migration based on embedded technology

[0169] We developed our own embedded development board and UI interface system based on ARM architecture, and implemented IP core design based on flavor control by transplanting the h5 model to the edge control system.

[0170] 3.1 Building an edge intelligent controller

[0171] In order to ensure that the h5 model generated in the previous section that can describe the oil flavor model can be correctly used to control the oil production line and to ensure the correctness of the relevant control results, an edge intelligent controller is designed by combining a self-developed edge controller board with a serial port screen.

[0172] Based on the research and development of PCB boards, the human-computer interaction interface is designed and developed according to specific functions. The functions related to the edge controller include: 1) Customized processing of oil flavors based on oil processing and flavor big data. The production line can produce different flavors of edible oils according to different types of processed materials, such as strong-flavor rapeseed oil, light-flavor rapeseed oil, etc. The relevant control interface is as follows: Figure 5 As shown. It has realized the improvement from experience judgment → digital modeling → software IP. 2) Judge the fault of the production line according to the status of the oil production line. It mainly realizes the fault status judgment of microwave magnetrons and refining electric valves. Simple fault judgment often relies on technical engineers to solve it remotely via video, but complex problems often require engineers to make judgments on site, which greatly increases the cost. Here, a combination of UI human-machine interface + expert experience method is adopted to realize autonomous fault judgment of some vulnerable electrical equipment. 3) One-click configuration of production line communication protocol, the control interface is as follows Figure 6 shown.

[0173] Traditional machine learning algorithms are usually trained and tested on cloud servers or local deep learning hosts. To solve the problems of engineering and miniaturization applications, STMicroelectronics has developed a The core STM32F7 series chip realizes the engineering application of smaller neural network prediction on miniaturized and scalable embedded devices.

[0174] Here, the STM32F767IGT6 chip is used for the development of related edge embedded devices (see the controller PCB diagram Figure 7 ), internal FLASH of the chip: 1MB, internal RAM: 384KB. In order to expand the storage space of related chips and ensure that the regression model generated by the workstation has enough space to run on the self-developed edge device, the device adds external RAM and NOR-FLASH chips. The STM32F767IGT6 uses an FMC variable storage controller to connect to the external storage chip and exchange data.

[0175] The NOR-FLASH model uses S29GL01GS11TFI010, and the FMC manageable memory space allocated through the malloc.c memory management file is 20MB. The external RAM model uses W9825G6KH, and the memory space is 32MB.

[0176] The external clock frequency used by the circuit here is 25Mhz. After multiplication, the system clock frequency can reach 216Mhz. After completing the hardware design and debugging of the edge intelligent controller, the design of related supporting software begins.

[0177] STMicroelectronics STM32tubeMx integrated development environment can automatically generate relevant programs by manually selecting the chip and configuring its corresponding hardware interface, clock frequency and other parameters. Figure 8 As shown, to simplify the design, we only set the system clock interface and frequency, UART1 and USART4 serial interfaces and communication rates required for the machine learning prediction function, and import the CNN.h5 neural network regression model generated in the previous article through the X-CUBE-AI component for analysis. After the relevant settings are completed, select the integrated development environment and click GENERATE CODE to automatically generate code.

[0178] The generated code basically completes the underlying driver configuration, but the application layer needs to be modified according to actual needs.

[0179] 3.2 Testing the Model on the Edge

[0180] After using the software code generated by the STM32tubeMx software to build the underlying driver, write the application layer according to actual needs. It is necessary to perform de-normalization on the output data in the program in advance. The specific formula is shown in (5):

[0181] x i =y i *σ+μ (5)

[0182] Where: y i is the output variable predicted by the neural network, σ is the standard deviation, μ is the mean, and x i is the predicted value after denormalization.

[0183] STM32 uses Google's Protocol Buffers protocol to transmit neural network data. We obtain the predicted current value by denormalizing the output value of the mathematical model in the protocol. The corresponding results interact with the host computer through the USART hardware interface. The convolutional network step size selected here is 5. The process value parameters are predicted using the workstation and edge device respectively. The predicted data obtained from the two are put into an Excel table together with the actual current data, and the follow-up curves of the two are drawn. The relevant curves are shown as follows: Fig. 9 and 10 As shown:

[0184] The above prediction data results are analyzed and compared with the actual values ​​to obtain the following regression model evaluation results. The results prove that the system has high prediction accuracy. The relevant results are shown in Table 2.

[0185] Table 2 Comparison of prediction results between edge devices and workstations

[0186]

[0187] 4. Implementation of relevant controls on the production line

[0188] The edge control system that has transplanted the flavor big data model sends process instructions to the PLC actuator through the MODBUS protocol to achieve flavor control and production line status monitoring for oil products.

[0189] 4.1 Implementation of Oil Flavor Control Based on Deep Learning

[0190] Through deep learning tools, nonlinear multivariate regression modeling is performed on various parameters including process and flavor, including material moisture content, and the relevant regression model is used to achieve process state control of the instruction target parameters. The relevant control topology diagram is as follows: Fig.11 shown.

[0191] The relevant control methods are described as follows:

[0192] The user selects the type of edible oil to be produced and the corresponding flavor, and then the UI interface outputs the corresponding serial port command to the edge controller SOC where the flavor control model has been transplanted. The relevant controller adjusts the flavor data in the flavor array based on the flavor experience. For example, compared with the light-flavored rapeseed oil, the strong-flavored rapeseed oil needs to adjust the W1C benzene aromatic substances and W1W inorganic sulfides, and other substances remain basically unchanged. The ten-dimensional flavor data is input into the flavor model running on the SOC. The flavor model predicts the moisture content instruction and other process parameters required at this time, such as moisture content instruction. The moisture content instruction is input into the self-developed dual-loop control algorithm. The relevant control algorithm has been described in other patents, and it is only briefly described here.

[0193] The moisture content dual-loop control algorithm takes the material moisture content command as input and adopts a dual-loop control method of moisture loop and temperature loop. The moisture loop is the outer loop and the temperature loop is the inner loop of the system. The moisture loop takes the moisture content command as input and the material temperature command as output. The temperature loop takes the temperature command as input and outputs the microwave conveyor belt speed. The relevant formula is shown in formula (6):

[0194] T out =fuzzy(M ref ,M fdb ) (6)

[0195] M re'f To input the moisture content command, the moisture content feedback of the moisture sensor (into the squeezed material) is M fdb , the output is the microwave material temperature instruction value T out .

[0196] R ref =fuzzy(T out ,T fdb )

[0197] T out is the temperature command output by the water loop, T fdb is the temperature feedback of the temperature sensor (microwave device), R ref is the output microwave conveyor belt speed.

[0198] The edge controller outputs conveyor belt instructions to the PLC actuator to control the flavor of the cooking oil.

[0199] Example 2

[0200] like Fig.12 FIG. 1 is a flow chart of a method for controlling oil flavor according to an embodiment of the present invention. Figure 5 The method shown includes the following steps:

[0201] S1: The edge controller is used to quickly arrange the communication protocol, and the production line process data is uploaded through the DTU. The relevant process data is stored on Alibaba Cloud through the SQL database. The electronic nose odor sensor is further used to continuously collect the flavor of the rapeseed kernels after modulation and the crude oil after pressing. The test results characterize the flavor substances in the rapeseed kernels or the crude oil after pressing from the ten dimensions in Table 3:

[0202] Table 3 Electronic nose sampling content

[0203] Array number Sensor Name Measurement object 1 W1C Benzene aromatic components 2 W5S Broad spectrum sensitivity, very sensitive to nitrogen oxides 3 W3C Ammonia, sensitive to aromatic components 4 W6S Selective for hydrogen radicals 5 W5C Sensitive to short-chain hydrocarbon aromatic components 6 W1S Sensitive to alkanes, such as non-methane total hydrocarbons 7 W1W Sensitive to inorganic sulfides 8 W2S Sensitive to alcohol ethers 9 W2W Aromatic ingredients, sensitive to organic sulfides 10 W3S Sensitive to long chain hydrocarbons

[0204] S2: Design different processing techniques to ensure that sufficient process data can be obtained and different process data present a certain gradient. Targeted experimental processes need to be designed before the experiment. Since the microwave power is fixed, in order to create different process environments, different process environments are created by adjusting the material thickness and adjusting the outlet temperature. Here, the rapeseed grain thickness is adjusted to 5mm, 8mm, and 11mm respectively. At each thickness, the outlet temperature is controlled to 125, 130, 135, 140, and 145 degrees respectively. Each process state is maintained for about 90 minutes to ensure the stability of the relevant sample state.

[0205] The relevant test methods for the samples include:

[0206] Pretreatment of the collected rapeseed oil samples:

[0207] Each portion of rapeseed oil was centrifuged at 5000 r / min for 10 minutes, and the centrifuged sample was filtered and then sealed and stored at 4°C for later use.

[0208] Electronic nose flavor measurement of rapeseed oil samples:

[0209] Take 5g of sample from the middle layer of the sample bottle into a 20ml headspace injection bottle, seal it with a cap and place it in a 40℃ water bath for 30 minutes. Then insert two sampling tubes into the headspace bottle cap, one of which is connected to the electronic nose sampler and the other is connected to the atmosphere to balance the air pressure, and then conduct the electronic nose flavor measurement.

[0210] Semi-quantitative analysis of flavor substances in rapeseed oil samples:

[0211] HS-SPME was used to extract the volatile flavor components of rapeseed oil. 5 g of rapeseed oil was weighed and placed in a 20 mL headspace bottle. 1 μL of 0.816 g / mL internal standard 2-methyl-3-heptanone was added. The mixture was balanced in a 50 °C water bath for 20 min. The DVB / CAR / PDMS extraction material was adsorbed for 30 min and desorbed at the GC injection port for 5 min. Each sample was repeated 3 times. Chromatographic conditions 1: DB-WAX capillary column (30 m × 0.25 mm, 0.25 μm); temperature program: 40 °C for 2 min, then increased to 200 °C at 4 °C / min, then increased to 230 °C at 5 °C / min; helium flow rate was 1.5 mL / min, injection port temperature was 250 °C, and injection mode was splitless. Mass spectrometry conditions: electron energy 70eV; ion source temperature 230℃; transfer line temperature 250℃; quadrupole temperature 150℃; mass scanning range m / z 40~350. Qualitative analysis: GC-MS raw data was matched with NIST17 spectral library for qualitative analysis, and compared with standard products and retention time index. Quantitative analysis: The internal standard method was used for quantitative analysis, and the calculation formula was:

[0212]

[0213] Where: Ci is the concentration of the unknown compound; Si is the concentration of the internal standard in rapeseed oil; CA is the peak area of ​​the unknown compound; SA is the peak area of ​​the internal standard. The relative correction factor in this experiment defaults to 1

[0214] GC-O-MS analysis

[0215] The GC-MS conditions were the same as above. The airflow distribution ratio between the mass spectrometer and the olfactometer was 1:1, and the port temperature of the olfactometer was 260°C. The sensory evaluators perceived and recorded the odor of the odor compounds separated from the chromatographic column by the detector, including the time, odor characteristics and odor intensity of the odor. At the same time, the flavor data measured by the electronic nose were compared.

[0216] The timestamp of the process database is further used, and the corresponding process data is intercepted according to the sampling time of the electronic nose, and the relevant data is made into a corresponding data table.

[0217] S3: Classify the relevant process data into 5mm, 8mm, and 11mm, and each set of data includes 560 sets of data. Since the test set is small, the ratio of training set, validation set, and test set is 6:2:2. The training set, validation set, and test set are 336, 112, and 112 respectively. The relevant data are judged by the Spearman matrix correlation and the nonlinear multivariate regression modeling based on the GRU model to obtain the h5 model that can describe the flavor dimension and the corresponding flavor data.

[0218] S4: Transplant the h5 model of flavor into the self-developed edge controller, and then use the edge controller to participate in related flavor control. The user issues relevant flavor instructions through the controller UI interface, which are converted into corresponding specific flavor vector data through expert experience. Further relevant flavor vector data is input into the flavor model as model input instructions to obtain processing instructions including material moisture content. And through the self-developed fuzzy double closed-loop algorithm, the precise moisture content control of the material and other related process parameters can be achieved, and the production line status monitoring of oil flavor can be realized.

[0219] The specific implementation flow chart is as follows Fig.12 shown.

[0220] It should be understood that the size of the serial numbers of the steps in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0221] The above embodiments are only used to illustrate the design ideas and features of the present invention, and their purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. The protection scope of the present invention is not limited to the above embodiments. Therefore, any equivalent changes or modifications made based on the principles and design ideas disclosed by the present invention are within the protection scope of the present invention.

Claims

1. A method for monitoring the state of an oil flavor production line based on deep learning, characterized in that: The following steps are involved: S1: Collect process data and flavor data in the oil production environment in real time online through sensors and upload them to the cloud database; S2: Cut and process the process data in the cloud database based on the time axis, perform multivariate nonlinear regression modeling on the process data and flavor data of the oil based on deep learning, and obtain a flavor big data model; S3: Build an edge controller and transplant the flavor big data model to the edge controller to form an IP core for flavor control; S4: Monitor the production line status and adjust the flavor of oil products through the edge controller that has transplanted the flavor big data model.

2. The method for monitoring the state of an oil flavor production line based on deep learning according to claim 1, characterized in that: Before step S1, the method further includes the following steps: Use ASCII code communication protocol and modular communication method to handshake the local and cloud communication protocols, and configure the cloud protocol with one click; The oil production line is modified to collect crude oil samples under different processing environment: including installing scent holes on the microwave tempering tank, and arranging sampling valves after the crude oil pump and before the solenoid valve in the refining pipeline.

3. The method for monitoring the state of an oil flavor production line based on deep learning according to claim 2, characterized in that: In the step S1, the process data include different oil thicknesses, press current, press temperature and pressure, microwave outlet temperature, moisture content of the material entering the press, microwave speed and press speed.

4. The method for monitoring the state of an oil flavor production line based on deep learning according to claim 2, characterized in that: In step S1, the specific steps of collecting flavor data are: An air collection pipeline is connected to the aroma collection hole and connected to a buffer bottle for air pressure balance; the other end of the buffer bottle is connected to the collection tube of the electronic nose to measure the flavor data; Crude oil samples were collected at the sampling valve, and the flavor components of the crude oil were qualitatively and semi-quantitatively analyzed by combining electronic nose and GC-MS.

5. The method for monitoring the state of an oil flavor production line based on deep learning according to claim 1, characterized in that: In the step S2, the specific steps are: S21: Use the Spearman matrix to mine the correlation between flavor data and process variables and find the two process variables with the highest correlation with flavor data; S22: standardizing the flavor data and the process data so that the processed data is scaled to a distribution with a mean of 0 and a standard deviation of 1; S23: A GRU model associated with a multivariate time series is used to generate a nonlinear regression model for flavor characteristics and process characteristics, and the flavor data and process data are fitted according to the time axis.

6. The method for monitoring the state of an oil flavor production line based on deep learning according to claim 5, characterized in that: In the step S23, the specific steps are: Build and train the GRU model to generate the h5 neural network regression model for describing the oil flavor model; draw the MAE curve graph; The model is used to predict the process data under different flavor requirements, the prediction results are de-standardized, and a curve chart of the actual value and predicted value of the required process requirements is drawn.

7. The method for monitoring the state of an oil flavor production line based on deep learning according to claim 1, characterized in that: In the step S3, the specific steps are: S31: Build the hardware platform of the edge controller, import the hardware settings into the h5 neural network regression model, and complete the underlying driver configuration, including configuring the communication protocol between the production line and the cloud; S32: Modify the application layer according to actual needs, and design a human-computer interaction interface to determine the cause of the fault according to the status of the oil production line, including determining the fault status of the microwave magnetron and the refining electric valve; S33: Testing the h5 neural network regression model on the edge controller; The predicted value data output by the h5 neural network regression model is denormalized to obtain the predicted current value; the following curves of the predicted current value and the actual current value are drawn to compare and obtain the evaluation result of the h5 neural network regression model.

8. The method for monitoring the state of an oil flavor production line based on deep learning according to claim 1, characterized in that: In the step S4, the specific steps are: S41: The operator selects the type of oil to be produced and the corresponding flavor; S42: adjusting the flavor data dimension through the edge controller according to experience, and inputting it into the flavor big model; S43: The flavor model predicts process parameters, uses the water loop and temperature loop for dual-loop control, and outputs control instructions; S44: The edge controller outputs the control instruction to the actuator of the oil production line to achieve the regulation of the oil flavor.

9. The method for monitoring the state of an oil flavor production line based on deep learning according to claim 8, characterized in that: In step S43, the specific steps of the dual-loop control method are: The dual-loop control method uses the water loop as the outer loop and the temperature loop as the inner loop; the input of the water loop is assumed to be the water content command M re'f , the moisture content of the material entering the press is fed back to M fdb , the output is the material temperature instruction T out ,but: T out =fuzzy(M ref ,M fdb ); The input of the temperature loop is the temperature command T output by the water loop. out , the temperature feedback T of the microwave device fdb , the output is the microwave belt speed R ref ,but R ref =fuzzy(T out ,T fdb )。 10. A deep learning-based oil flavor production line status monitoring device, characterized in that: Includes sensors, edge controllers, actuators, and cloud databases; It also includes a processor and a memory, wherein the memory stores computer instructions, and the processor is used to execute the computer instructions stored in the memory. When the computer instructions are executed by the processor, the system implements the steps of a method for monitoring the status of an oil flavor production line based on deep learning as described in any one of claims 1 to 9.

Images