MAMÍFEROS

Thermal processing is one of the major operations in the food processing and preservation system. It has generally been viewed as an energy-intensive pres- ervation technique, but persists as the most widely used method of preservation. One major objective of thermal processing is to destroy pathogenic and spoilage microorganisms present in foods being processed so that they can be stored for extended periods and consumed with no safety concerns. Quality factors in foods are also affected by heat treatments; however, they usually show much higher heat resistance than targeted microorganisms. Therefore, an optimal thermal process procedure for a given food product means that it will result in minimal quality destruction while being sufficient to make the product safe for consump- tion. In order to solve an optimization problem, the key step is to develop a suitable model capable of describing the relationships between inputs and out- puts. To date, a variety of conventional methods have been used for both mod- eling and optimization purposes in food thermal processing areas, from which different models have been developed, including numerical, analytical, and experimental. These conventional methods applied for optimizing thermal pro- cessing operations need specific inputs. First, the models need a knowledge and understanding of relationships between the input and output variables. Second, it is necessary that information of physical and thermal properties of food products being modeled be available. Unlike other disciplines, food processing deals with biomaterials, which show much more complicated thermophysical properties and uncertainties during the processing period. This also results in more complicated relationships between input and output variables. Thus, often it is difficult to use a simple partial differential equation or model to accurately describe the phenomena occurring in food processing operations. In addition, the lower calculation speed of most conventional methods under complex situ- ations of process optimization and control limits them to be optimally applied for online application in industrial processes. Neural network models offer an attractive alternative in such instances. Hence, neural networks are being con- tinually extended to the food thermal processing area as a modeling and opti- mization technique.

Since the 1980s, neural networks have received more and more interest in food processing areas. So far, the applications of neural networks in food processing have covered various areas, such as drying,7–12 _{fermentation,}13,14 extrusion,15–17 _freezing,18 _baking,19,20 _postharvest,21,22 _{experimental design,}23 _etc. The application in food thermal processing began in the mid-1990s. Sablani et al.24 _{published one of the early reports on the application of neural networks} in food thermal processing. They developed a four-layer neural network with three inputs and three outputs to predict optimal sterilization temperatures under different processing conditions. Sablani et al.25 _{used artificial neural network} models for the overall heat transfer coefficient and the fluid-to-particle heat transfer associated with liquid particle mixtures, in cans subjected to end-over-end rotation.

The application of the neural computing approach for prediction of the residence

time distribution (RTD) under aseptic processing conditions was reported by

Chen and Ramaswamy.26 _{In this paper, neural networks were explored for} modeling two RTD functions: the time-specific (E-type distribution) and the cumulative particle concentration function (F-type distribution) of carrot cubes in starch solutions in a vertical scraped-surface heat exchanger (SSHE) of a pilot-scale aseptic processing system. Neural networks have been used as an alternative tool to the Ball and Stumbo methods, which are the most often used methods for thermal calculation to predict process time and or process lethality in cans during thermal processing.27 _{A more systematic and in-depth application} of neural networks in food thermal processing areas was carried out by Chen and Ramaswamy.28–33 _{In these studies, separate ANN prediction models were} developed involving main input parameters such as retort temperature profile, thermophysical properties of food products, kinetics of microorganisms, and quality factors and outputs, such as the process time, cumulative lethality value, quality retention, unit energy consumption, and transient temperature at the can center. These ANN models were able to be directly used for the process establishment and validation for a given food product, but also could be com- bined with a search technique to build optimal thermal process conditions in order to meet different optimization objectives. Some details of selected studies in thermal processing based on the ANN approach are given in the following sections.

4.5.1 NEURAL NETWORK MODELING OF HEAT TRANSFER TO LIQUID PARTICLE MIXTURES IN CANS SUBJECTED TO END-OVER-END PROCESSING

The overall heat transfer coefficient (U ) and fluid-to-particle heat transfer

coefficient (hfp) are fundamental data needed to develop prediction models for the transient temperature of canned foods undergoing agitation thermal pro- cessing, which is necessary to establish and optimize the thermal process schedule for the canned liquid–particle food system. Traditionally, the dimen- sionless correlations are used for development of an experimental model of U and hfp involving other influencing parameters by use of multiple regression analysis. However, selection of appropriate dimensionless groups requires prior knowledge of the phenomena under investigation. Sablani et al.25 _devel- oped an artificial neural network (ANN) model for the overall heat transfer coefficient and the fluid-to-particle heat transfer coefficient associated with liquid particle mixtures, in cans subjected to end-over-end rotation. Experi- mental data obtained for U and hfp under various test conditions (shown in Table 4.1) were used for both training and evaluation. Multilayer neural net- works with seven inputs and two output neurons (for a single particle in a can), and six inputs and two output neurons (for multiple particles in a can) were trained. The optimal network was obtained by initial trials as number of hidden layers = 2, number of neurons in each hidden layer = 10, and learning

runs = 50,000. By use of trained NN models with optimal configurations, the prediction performance of all NN models for both U and hfp was found to be higher than 0.98, meaning that the developed NN models could safely be used for prediction of U and hfp under the given experimental conditions. The comparison of NN models and dimensionless regression models using the same experimental data is summarized in Table 4.2. Prediction errors using ANN were less than 3 and 5%, respectively, for U and hfp, which were about 50% better than those associated with dimensionless number models, indicat- ing that the predictive performance of the ANN was far superior than that of dimensionless correlations.

TABLE 4.1

Range of System and Product Parameters Used in the Determination of Heat Transfer Coefficients (U and h_fp)

No. Parameter Experimental Range

1 Retort temperature 110, 120, and 130°C

2 Radius of rotation 0, 0.09, 0.19, and 0.27 m

3 Rotation speed 10, 15, and 20 rpm

4 Can headspace 0.0064 and 0.01

5 Test fluid Water and oil

6 Test particle Polypropylene, nylon

7 Particle concentration Single particle, 20, 30, and 40% (v/v)

8 Particle shape and size

Cube 0.01905 m

Cylinder 0.01905 × 0.01905 m

Sphere 0.01905, 0.02225, and 0.025 m

9 Can dimension 307 × 409 (8.73 × 11.6)

TABLE 4.2

Comparison of Error Parameters for Neural Network (NN) Models and Dimensionless Correlation (DC) Models

Single Particle Multiple Particle

U hfp U hfp Error Parameters DC NN DC NN DC NN DC NN MAE 17.1 5.11 31.3 17.2 25.1 9.85 75.4 48.1 SDE 25.4 4.76 43.3 16.0 32.0 11.0 63.4 40.7 MRE (%) 5.00 2.46 16.9 5.82 5.70 2.57 8.26 4.52 SRE (%) 3.76 2.51 11.9 7.00 4.65 1.96 7.12 3.90 R2 _{0.99 0.99 0.83 0.98 0.98 0.99 0.96 0.98}

4.5.2 A NEURO-COMPUTING APPROACH FOR MODELING OF RESIDENCE TIME DISTRIBUTIONOF CARROT CUBES INA VERTICAL SCRAPED-SURFACE HEAT EXCHANGER

The residence time distribution (RTD) is one of the important parameters for establishing the aseptic processing of particulate liquids. Although a lot of different models have been developed for describing RTD characteristics using conventional mathematical methods, none of them give a fully satisfactory solution for the RTD covering the wide range of processing conditions. A neuro- computing approach was used by Chen and Ramaswamy26 _{for modeling two} residence time distribution (RTD) functions: the time-specific (E-type distribu- tion) and the cumulative particle concentration function (F-type distribution) of carrot cubes in starch solutions in a vertical scraped-surface heat exchanger (SSHE) of a pilot-scale aseptic processing system. In this study, 356 experi- mental data pairs obtained for E(t) and F(t) under various test conditions, including the concentration of particles, flow rate, particle dimension, and test time, were used for both training and evaluation. The optimal configurations of the neural network model were determined by adjusting the number of hidden layers, the number of neurons in each hidden layer and learning runs, and a combination of learning rule and transfer functions. The results showed that the trained ANN model can accurately map experimental results with R2 _{value =} 0.98 and 0.99 for E and F functions, respectively. The prediction performance of the ANN model under several typical processing conditions is shown in Figure 4.5. The ANN models were also compared with conventional models developed based on multiple variable regression techniques. The comparison indicated that average modeling errors associated with the ANN model were 5.7 and 3.0%, respectively, for E and F, while those for the multiple regression models were 15.5 and 12.3%, meaning that the ANN model had higher precision for predicting E and F functions.

4.5.3 MODELING AND OPTIMIZATION OF CONSTANT RETORT