Application of Artificial Neural Network in the Baking Process of Salmon

/e global production of farmed Atlantic salmon amounts to over 2 million tons per year. Consumed all over the world, salmon is not only delicious but also nutritious./is paper deals with the relationship between moisture content, low-field nuclear magnetic resonance (LF-NMR), scanning electron microscope (SEM), and sensory evaluation in the baking process of salmon. An artificial neural network (ANN)model has been established to simulate the change of moisture content and energy consumed in the baking process. /rough the study of LF-NMR, SEM, and sensory evaluation, it was found that the change of sensory indexes was consistent with the results observed by LF-NMR and SEM. With the increase of temperature, muscle fibers contracted, the interstices increased, the rate of water loss increased, and the sensory score decreased. Initial moisture content, baking time, baking temperature, baking humidity, and baking air velocity were employed as the baking control parameters for the ANN. ANN can be used to determine the moisture content and energy consumed of baking salmon./e best network topology occurred with 5 input layer neurons, 17 hidden layer neurons, and 2 output layer neurons, and the MSE was 0.00153, and Rall was 0.99661. According to the experiment, it was demonstrated that the ANN is a reliable software-based method.


Introduction
Salmon is a large-and medium-sized cold water migratory fish, mainly distributed in the northern Pacific Ocean and the boundary between the Atlantic Ocean and the Arctic Ocean [1]. Salmon is not only delicious but also nutritious [2]. e global production of farmed Atlantic salmon (Salmo salar L.) reached 2.36 million tons in 2017 [3]. e salmon market was dominated by raw food and smoked products [4]. Veiseth-Kent et al. [5] studied the effect of sensory assessment and texture assessment on the postmortem process of raw salmon by sensory and instrumental methods. Birkeland et al. [6] researched cold smoking procedures and raw material characteristics on product yield and quality parameters of cold smoked fillets. Other scholars also studied raw salmon and cold smoked salmon preservation technology [7,8], processing technology [9], food safety [10], and other aspects. With the improvement of people's living standard, higher requirements have been put forward for salmon products.
Baking was a common cooking method which induces water loss in the food [11,12]. Evaporation of water was one of the several fundamental complex physical processes during baking [13]. e main conditions affecting baking include baking temperature, air velocity, and humidity.
Artificial neural network (ANN) is one of the black-box modeling approaches, which is a heuristic soft computing method used for the nonlinear and complex systems [14,15]. ANN has been used in the food industry for modeling many processes, such as estimation of antioxidant activity of foods, recognition in the drying of guava pieces in the spouted bed, and prediction of paddy drying a fluidized-bed drier [16][17][18].
Generally speaking, ANN was used to predict product indicators (moisture content, crumb temperature, color change, and relative volume) with the inputs of drying parameters (jet temperature, jet velocity, and baking time) [19]. Similar study was conducted to predict the modified moisture ratio of pepper-tree fruits with two inputs of mass, air temperature, and air velocity [20]. In addition, there were many researchers who used different variables to predict the performance of moisture content [21][22][23]. Current research studies generally had taken a single parameter (such as moisture, active ingredients, and color) as the output layer to establish the ANN model. When studying multiple indicators, multiple models were generally established to study one by one. erefore, if the ANN model can be established by simultaneously studying two parameters as output layers, such as moisture content and energy consumed in this paper, it will be more conducive to the wider application of ANN in food research, which is also the significance of this paper.

Materials and Drying Equipment.
e salmon used in this experiment was Pacific salmon. Chilled salmons were purchased from a local market and were transported to the laboratory using refrigerated transportation. At the beginning of the experiment, salmons were removed the head, scales, skin, and bones and taken the anterior back muscles of the salmon to be the materials. e salmons were cut manually using the cubic device with dimensions of 1 cm × 1 cm with a thickness of 4 cm. e baking experiment was conducted using the universal steam oven (model SCC WE 101, Rational Co., Ltd., Germany).

Experimental Procedure.
e initial moisture content of the experiment materials was controlled at about 59.91 ± 0.27%. e baking temperature, time, humidity, and air velocity are given in Table 1. e levels for the process variables were decided based on trial experiments. e prepared samples were picked in the seasoning solution for 2 hours at 4°C. In the first half minutes, the sample moisture content was measured at an interval of 2 min. After 30 minutes, the sample moisture content was measured at an interval of 5 min.

Measurement of Energy Consumption.
Fifty-four groups of salmons were investigated for energy consumption measurement. An energy meter (model DTS 7738 3 × 220/ 380 V, Shanghai Huali Co., Ltd., China) bridging the connection between a voltage stabilizer and the universal steam oven was installed. Energy consumed per experiment was estimated in kilowatt hours.

Low-Field Nuclear Magnetic Resonance (LF-NMR) Imaging Technology.
e samples were prepared according to the treatment method of the fish in method 2.1. e salmon samples were divided into 54 groups with 3 parallel for each group. Samples were baked at different temperatures (100°C, 110°C, 120°C, 130°C, 140°C, and 150°C), and each temperature was set for different times (2 min, 4 min, 6 min, 8 min, 10 min, 12 min, 14 min, and 16 min). e processed sample was detected by using an LF-NMR imaging analyzer. e resonance frequency of the proton was 22.7 mhz, the temperature of the magnet was 32°C, and the strength of the magnet was 0.47 T. e salmon sample was placed in a cylindrical feeding tube, with imaging parameters settings as follows: TW � 1500 ms, TE � 20 ms, average � 2, slice width � 2.5 mm, and slice � 1 [24].

Scanning Electron Microscope (SEM).
e SEM was used to analyze the microstructure of salmons after baking 16 min. e sample microstructure was observed by JEOL model JSM-7800F, Tokyo, Japan. e specimen fragments for SEM were taken from the center of baked sample and dehydrated by freeze-drying. Small piece of about 4 × 4 × 1 mm was cut from the dried samples and fixed on the SEM stub, which were coated with gold to provide a reflective surface for electron beam. e gold-coated samples were viewed under the microscope, and a 50× magnification was used in all SEM observations [25].

Sensory Evaluation.
e sensory qualities of different salmons were analyzed in terms of taste (1-10 points), odour (1-10 points), color (1-10 points), hard (1-10 points), and springiness (1-10 points). An eight-member panel, all of whom were experienced in the sensory evaluation of salmon foods, scored the five parts. e judges were asked to give their remarks about each of the samples [26].

Artificial Neural Network (ANN) Implementation.
e ANN is a multilayer feedforward neural network trained according to the error back propagation algorithm with a momentum adjustment and an adaptive learning rate [27][28][29].
e ANN implemented a three-layer ANN like the one shown in Figure 1(a). e three kinds of layers in our ANN are known as input, hidden, and output layers. Equation (1) through (3) express the inputs of the input layer: Equation (4) expresses the outputs of the hidden layer: e input signal to the output layer is estimated using

Journal of Food Quality
e final output can be expressed as We used the Neural Network Toolbox and MATLAB R2012a to develop our implementation, employing MAT-LAB's toolbox to write the program, load data files, train and validate the network, and save the model architecture. e model structure is shown in Figure 1(b). e established three-layer ANN had five input variables, including initial moisture content, baking temperature (six levels), baking air velocity (three levels), baking humidity (three levels), and baking time. e final moisture content and energy consumed of the salmon were taken as the two output variables. Before training the network, it is necessary to standardize the input and output data to express the correlation between them accurately. We normalized the general weight value of both input and output data between [0, 1] according to the equation [30,31].
where x i , x i , x min , and x max are the weight values before and after pretreatment of neutral i, and the minimum and maximum weights of each neural network, respectively. We used only a single layer in our implementation because more layers may cause the local minimum problem [32,33]. We proposed the feedforward-backpropagation learning algorithm along with a Levenberg-Marquardt (LM) training function in this model [34]. e feedforward neural network is organized in three or more layers, an input layer, an output layer, and one or more hidden layers. From the input layer to the output layer, the network is one-way connection [35]. In this study, we randomly divided the analysis data for the drying process into three parts: the first part was used to train the network and consisted of approximately 70% of the total data points. e second part was used to validate the network and consisted of approximately 15% of the samples. e remaining 15% were used as experimental inputs [36,37]. We took into account the different numbers of neurons in the hidden layer. e predicted moisture content and energy consumed change computed to evaluate the performance of fitting and predicting by using the least mean square error (MSE) metric and coefficient of determination (R 2 ). An MSE of 0.01 was deemed to indicate convergence. We allowed a maximum of 1000 iterations to ensure that the network completed the training process. e linear (purelin) transfer function provided better correlation coefficients for the processed hidden layer output data (O 1 I ) and was therefore found suitable for the output neuron [38].

Effects of Different Baking Conditions of Salmon.
LF-MRI can obtain the fault visualization information of the sample and obtain the H + proton density and distribution in the sample, so as to reflect the content and distribution of water or oil in the sample. e higher the content of water and oil in the sample, the greater the proton density. H + proton density imaging was performed on salmon pieces at different baking temperatures, as shown in Figure 2. At the beginning of baking, the sample water distribution was relatively uniform, the proton density was larger, and the fat was mainly concentrated in the fat grain of the sample. With the extension of baking time, the color of the NMR image of the sample gradually becomes lighter and the moisture gradually decreases. At the same time, the reduction rate of proton density was accelerated as the sample temperature increased. According to the results of LF-MRI, the water loss rate is different with different heating temperatures. e higher the temperature, the faster the water loss. SEM of baked salmon at different baking temperatures is illustrated in Figure 3. e effect on the microstructure of baked salmon was characterized by thick or thin muscle fibrils. It can be seen that muscle fibrils of samples baked at 100°C were thick and had small interstices. e thickness and interstice of muscle fibrils changed with increase in the  drying temperature. When samples were baked at 150°C, a highly interstices final product was obtained with thinner muscle fibrils. ese interstices not only affect the textural property but also the transport phenomenon, such as diffusivities of gases and liquids in the sample and resulting higher rate of effective moisture diffusivity at a higher drying temperature [39,40]. Tables 1 and 2, the results of sensory evaluation of baking salmon in different temperature ranges can be seen. When the baking condition was 100/8/0 (temperature/ air velocity/humidity) and 150/25/10 (temperature/air velocity/humidity), the color sensory score of baked salmon was the best, reaching the highest score of 7.88, and when the baking condition was 150/16/0 (temperature/air velocity/ humidity), the color sensory score was the lowest.

Effects of Different Baking Conditions on Sensory Evaluation of Salmon. From
According to the evaluation of the baked salmon from the odour, when the condition was 110/8/0 (temperature/air velocity/humidity), the flavor was rich; basically, no fishy taste, the odour sensory score was the highest, and when the condition was 150/16/0, the fishy taste was heavy, the flavor was slightly light, and the odour sensory score was the lowest. Combined with the LF-NMR image, it is easy to see that the higher the temperature, the more fat was reduced. Fat is an important carrier of flavor substances, so the flavor score decreases with the increase of temperature. Fat is an important carrier of flavor substances, so the odour sensory score decreased with the increase of temperature.
From the hardness and springiness analysis of sensory evaluation, when the baking temperature was low, the fish was too soft, which affected the taste. When the baking condition was 100/25/20 (temperature/air velocity/humidity), the hardness score is the lowest, and when the baking condition was 150/25/10, the hardness and springiness score were the highest. Combined with the scanning electron    microscope results, it can be seen that with the increase of temperature, the aggregation of fibers increases the hardness of fish correspondingly. According to the evaluation of the baked salmon from the taste, when the baking condition was 140/25/0 (temperature/air velocity/humidity), the taste was best. When the baked salmon was rated by the total score of the five senses, the baked fish score reached the highest when the baking condition was 110/8/0 (temperature/air velocity/humidity), reaching 38.13. erefore, the control of baking conditions had a great impact on the quality of baked salmon, and the control of baking conditions provided strong technical support for the production of high-quality baked salmon in the factory.

ANN Model Performance.
e behavior of biological products under processing conditions is highly nonlinear in nature [38]. e moisture content and energy consumed in the case of salmons also show a similar kind of trend. It is therefore justified to apply ANN modeling to such complex data. Due to the adaptable nature of ANN, further addition of data can be performed to a pre-existing data set, and the model can be retrained to cover a wider range of levels for the process variables under study. Data sets generated through 54 experiments amounted to 824 points of which 576 data points were taken for training, 124 for testing, and the remaining 124 for validation. e data set for training, testing, and validation was created randomly using diver and function available in MATLAB, based on the overall correlation coefficient. Figure 4 and Table 3 illustrate the network performance for varying numbers of neurons in the hidden layer with the testing data. We determined the number of neurons in the hidden layer by predicting the percentage change in the moisture content and energy consumed. After repeated trials, it was found that a network with 20 hidden neurons produced the best performance during model development. However, according to Table 3, a network with 17 hidden neurons, R training , R validation , and R all were better than a network with 20 hidden neurons, just R test was less.
In addition, Figure 5 depicts the predicted moisture content and energy consumed versus experimental moisture    Table 4 shows the weights and bias estimation model data obtained by the ANN tool MATLAB R2012a. ANN accurately predicted the drying behavior of the sturgeon bone marrow. We chose the BP model suitable for this study not only because its accuracy but also its generality, being able to predict the behavior of the entire experimental range [41]. e model parameters described in this section (Table 1) along with the others defined are almost certainly useful for applying this model to moisture content prediction in other food products [38].