Optimization method for low-temperature vacuum frying and layered heating of three-time thick potato slices

Abstract

The invention provides an optimization method for low-temperature vacuum frying and layered heating of potato triple-thickness slices, and belongs to the technical field of food processing.The optimization method includes the steps that potato triple-thickness slice samples with infrared temperature monitoring small color blocks are prepared, miniature temperature probes are implanted, a low-temperature vacuum frying device comprising a layered heater set is built, and the potato triple-thickness slice samples are subjected to low-temperature vacuum frying; designing a plurality of groups of contrast experiments to carry out 50 batches of frying experiments, carrying out multi-angle image acquisition and test data acquisition on fried samples, extracting image features, fusing the test data to generate a comprehensive quality feature vector, constructing a bidirectional neural network model containing a heat transfer physical fitting equation set to carry out training, and performing reverse reasoning based on the preset optimal feature vector to obtain an optimized parameter combination, and finally screening the optimal parameter combination as a layered heating process parameter through a small test proof, thereby solving the technical problem of unstable quality of the potato chips caused by non-uniform temperature distribution in the low-temperature vacuum frying process.

Description

Technical Field

[0001] This invention belongs to the field of food processing technology, and specifically relates to an optimized method for low-temperature vacuum frying and layered heating of potato slices with a thickness of three times. Background Technology

[0002] Low-temperature vacuum frying of potato chips is a widely used dehydration and drying technology in modern food processing. Traditional frying processes control parameters such as oil temperature, vacuum level, and time to achieve dehydration and texture improvement of potato chips, and have been widely used in the industrial production of snack foods such as potato chips and French fries. However, when processing thicker potato chips, such as those three times their original thickness, traditional uniform heating methods are prone to overheating the surface and underheating the interior due to the time lag effect of heat conduction. This results in quality defects such as burnt exterior and moist interior, and uneven texture. In existing low-temperature vacuum frying processes, the lack of precise monitoring and layered control of the internal temperature field distribution of potato chips makes it impossible to design personalized heating strategies based on the heat transfer characteristics of thick chips. This makes it difficult to achieve the ideal uniformity in key quality indicators such as moisture content, oil distribution, and crispness of thick chip products. In other words, existing technologies suffer from the technical problem of uneven temperature distribution during low-temperature vacuum frying, leading to unstable potato chip quality. Summary of the Invention

[0003] In view of this, the present invention provides an optimized method for low-temperature vacuum frying and layered heating of potato slices with a thickness of three times, which can solve the technical problem of uneven temperature distribution during low-temperature vacuum frying, which leads to unstable quality of potato slices.

[0004] This invention is implemented as follows: An optimized method for low-temperature vacuum frying and layered heating of potato slices three times thick includes preparing potato slices three times thick and attaching infrared temperature monitoring patches to the surface of each slice. The infrared temperature monitoring patches are distributed in a grid pattern to cover the slice surface. Simultaneously, micro-temperature probes are implanted at different depths within the slice. A low-temperature vacuum frying experimental device is constructed, including a layered heater group, an infrared temperature monitoring system, a vacuum control system, and a data acquisition system. The layered heater group includes an upper heater, a middle heater, and a lower heater, with each heater's power independently controlled. Multiple sets of comparative experiments are designed, each with different frying temperatures, vacuum levels, heater power ratios, and frying time parameters. Multi-angle image acquisition and test data acquisition are performed on each batch of fried potato slices, including surface images. Images of the potato chips were taken at eight-eighths, one-quarter, half, and diameter sections to test moisture content, oil content, crispness, and color parameters. Texture, color, and structural features were extracted from each image. These features were spatiotemporally aligned with the test data and fused to generate a comprehensive quality feature vector for the potato chips. A bidirectional neural network model containing a set of heat transfer physics fitting equations was constructed and trained using experimental parameters as input and the comprehensive quality feature vector of the potato chips as output. The trained bidirectional neural network model was used to perform back-reasoning based on the preset optimal fried potato chip feature vector to obtain multiple sets of optimized parameter combinations. Small-scale experiments were conducted on each set of optimized parameter combinations to calculate the similarity between the average comprehensive quality feature vector and the standard vector. The optimal parameter combination was selected as the optimized layered heating process parameters.

[0005] The heat transfer physics fitting equation set includes the temperature gradient conduction equation, the moisture migration equation, and the oil penetration equation. The temperature gradient conduction equation is used to describe the temperature field distribution inside the potato slices, the moisture migration equation is used to calculate the migration pattern of moisture in the potato slices during the dehydration process, and the oil penetration equation is used to predict the penetration depth and distribution of oil in the potato slices.

[0006] The bidirectional neural network model has a dual-channel architecture consisting of a forward propagation neural network and a backward propagation neural network. The forward propagation neural network includes an input layer, three hidden layers, and an output layer, with each hidden layer containing 512 neurons. The backward propagation neural network includes an input layer, two hidden layers, and an output layer, with each hidden layer containing 256 neurons. The training dataset for the bidirectional neural network model is established by dividing 750 potato chip samples obtained from 50 batches of experiments into a training set of 525 samples, a validation set of 150 samples, and a test set of 75 samples in a ratio of 7:2:1. Each sample contains a 128-dimensional comprehensive quality feature vector of potato chips and a corresponding 4-dimensional experimental parameter vector.

[0007] The layered heater group adopts resistance wire heating. The power range of the upper layer heater is P1∈[100, 500]W, the power range of the middle layer heater is P2∈[150, 600]W, and the power range of the lower layer heater is P3∈[200, 700]W. The power of each layer can be adjusted independently.

[0008] The comprehensive quality feature vector of potato chips includes a numerical combination of 128 feature parameters, including surface texture roughness, internal porosity, color uniformity, moisture content distribution, oil penetration depth, brittleness distribution, and structural integrity.

[0009] The surface image acquisition step involves placing potato slices 30cm directly above the image acquisition platform and using an industrial camera with a resolution of 2048×2048 pixels to capture a complete surface image vertically downwards.

[0010] The step of acquiring the one-eighth edge section image involves cutting along the radius of the potato slice from the center outwards at a position seven-eighths of the radius from the center to obtain the section. The section is then placed on the microscope stage and photographed using a 40x objective lens to obtain an image of the microstructure of the edge region.

[0011] The step of acquiring the quarter-edge section image involves cutting along the radius of the potato slice from the center outwards at a position three-quarters of the radius from the center to obtain the section. The section is then placed on the microscope stage and photographed using a 20x objective lens to obtain an image of the medium structure of the sub-edge region.

[0012] The step of acquiring the half-edge cross-section image involves cutting along the radius of the potato slice from the center outwards at a position half the radius from the center to obtain the cross-section. The cross-section is then placed on the microscope stage and photographed using a 10x objective lens to obtain a macroscopic structural image of the central region.

[0013] The step of acquiring the diameter cross-section image involves completely cutting along the diameter of the potato slice, placing the cut surface on the image acquisition stage, and using a macro lens to capture an image of the complete internal cross-section. The image covers the entire cut surface and has a resolution of 1024×1024 pixels.

[0014] Optionally, the infrared temperature monitoring dots are circular markers made of ceramic material with an emissivity of 0.95. Each infrared temperature monitoring dot has a diameter of 1.5 mm and the spacing between adjacent infrared temperature monitoring dots is 3 mm. They are used to monitor the surface temperature distribution of potato slices in real time and feed it back to the power control system of the layered heater.

[0015] Optionally, the miniature temperature probe is a thermocouple probe with a diameter of 0.8 mm, which is implanted at depths of 25%, 50%, and 75% of the slice thickness to monitor temperature changes at different layers inside the potato slice.

[0016] The step of using a trained bidirectional neural network model to perform back-reasoning based on a preset optimal fried potato chip feature vector involves taking the preset optimal fried potato chip feature vector as the target output input to the model's output layer, performing back-propagation neural network branch calculations, and using gradient descent algorithm and Jacobian matrix solution to make the model output the input parameter combination that is closest to the target feature vector.

[0017] The step of conducting small-scale experiments on each set of optimized parameters involves preparing five potato slices three times thicker than the sample and executing the experimental procedure, then using a trained bidirectional neural network model to predict the 128-dimensional comprehensive quality feature vector of each sample, calculating the arithmetic mean of the feature vectors of the five samples as the average comprehensive quality feature vector, and using a cosine similarity algorithm to calculate the similarity value between the average vector and the feature vector of the preset standard high-quality potato slices.

[0018] The selection criteria for the optimal parameter combination are: moisture content controlled within the range of 8% to 12%, oil content controlled within the range of 15% to 20%, crispness value greater than 85N, and surface color parameter L in the comprehensive quality feature vector of potato slices. * Parameter combinations with values ​​in the range of 65 to 75.

[0019] This invention establishes a multi-dimensional temperature monitoring system combining a layered heater group with an infrared temperature monitoring system and a micro temperature probe. This system enables precise sensing and dynamic control of the internal and external temperature fields of potato slices three times their thickness, thus solving the problem of uneven temperature distribution caused by traditional uniform heating methods. By constructing a bidirectional neural network model containing a set of heat transfer physics fitting equations, a nonlinear mapping relationship is established between experimental parameters and the comprehensive quality feature vector of potato slices. This enables parameter optimization decisions based on quality prediction, overcoming the blindness and inaccuracy of traditional experience-based parameter tuning methods. Furthermore, by designing a game-theoretic optimization model for layered heating parameters, the objectives of temperature field homogenization and dehydration rate control are synergistically optimized. The temperature gradient coupling term achieves a balance solution for multiple objectives, ensuring the uniformity of temperature distribution and the accuracy of quality control during the dehydration process of the thick potato slices. In summary, this invention solves the technical problem mentioned in the background art where uneven temperature distribution during low-temperature vacuum frying leads to unstable potato slice quality. Attached Figure Description

[0020] Figure 1 This is a flowchart of the method of the present invention.

[0021] Figure 2 This is a graph showing the trend of moisture content variation in potato slices under different process parameters in the embodiment.

[0022] Figure 3 The diagram shows the distribution of verification results for the optimized parameter combinations in the example.

[0023] Figure 4 The following is a timing diagram of the layered heating power control strategy in the embodiment.

[0024] Figure 5 This is a control chart of key quality indicators for the continuous production process in the example. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0026] like Figure 1 The diagram shows an optimized method for low-temperature vacuum frying and layered heating of potato slices with a thickness of three times, as provided by this invention. This method includes the following steps:

[0027] S01. Prepare potato slices with a thickness of three times and attach infrared temperature monitoring blocks to the surface of each slice. The infrared temperature monitoring blocks are distributed in a grid pattern to cover the slice surface. Each infrared temperature monitoring block has a diameter of 1.5 mm and the spacing between adjacent infrared temperature monitoring blocks is 3 mm. At the same time, micro temperature probes are implanted in different depth layers within the slice.

[0028] S02. Construct a low-temperature vacuum frying experimental device. The device includes a layered heater group, an infrared temperature monitoring system, a vacuum control system, and a data acquisition system. The layered heater group includes an upper heater, a middle heater, and a lower heater. The power of each heater is independently controlled.

[0029] S03. Design multiple sets of comparative experiments. Each set of experiments has different frying temperature, vacuum degree, heater power ratio and frying time parameters. Conduct 50 batches of low temperature vacuum frying experiments. Each batch contains 15 potato slices with a thickness of three times.

[0030] S04. For each batch of fried potato slices, multi-angle image acquisition and test data acquisition are carried out, including surface image, one-eighth edge section image, one-quarter edge section image, one-half edge section image and diameter section image, and the moisture content, oil content, crispness value and color parameters are tested.

[0031] S05. Extract texture features, color features and structural features of various images, and fuse the image features with the test data after spatiotemporal alignment to generate a comprehensive quality feature vector of potato chips. Each sample corresponds to a 128-dimensional comprehensive quality feature vector.

[0032] S06. Construct a bidirectional neural network model containing a set of heat transfer physics fitting equations. Train the model with experimental parameters as input and the comprehensive quality feature vector of potato slices as output. The training is completed when the model prediction accuracy reaches more than 92%.

[0033] S07. Using the trained bidirectional neural network model, reverse reasoning is performed based on the preset optimal fried potato chip feature vector to obtain multiple sets of optimized parameter combinations.

[0034] S08. Conduct small-scale experiments on each optimized parameter combination, calculate the similarity between the average comprehensive quality feature vector and the standard vector, and select the optimal parameter combination as the optimized layered heating process parameters.

[0035] The infrared temperature monitoring patch is a circular marker made of ceramic material with an emissivity of 0.95, used to monitor the surface temperature distribution of potato slices in real time and feed it back to the power control system of the layered heater.

[0036] The micro temperature probe is a thermocouple probe with a diameter of 0.8 mm, which is implanted at depths of 25%, 50%, and 75% of the potato slice thickness to monitor temperature changes at different layers inside the potato slice.

[0037] The layered heater group adopts resistance wire heating. The power range of the upper heater is P1∈[100,500]W, the power range of the middle heater is P2∈[150,600]W, and the power range of the lower heater is P3∈[200,700]W. The power of each layer can be adjusted independently.

[0038] The comprehensive quality feature vector of the potato chips includes a numerical combination of 128 feature parameters, such as surface texture roughness, internal porosity, color uniformity, moisture content distribution, oil penetration depth, brittleness distribution, and structural integrity.

[0039] The surface image acquisition step includes placing potato slices 30cm directly above the image acquisition platform and using an industrial camera with a resolution of 2048×2048 pixels to capture a complete surface image vertically downwards.

[0040] The one-eighth edge section image acquisition step involves cutting the potato slice radially outward from the center at a point seven-eighths of a radius away from the center. The cut surface is then placed on a microscope stage and photographed using a 40x objective lens to obtain an image of the microstructure of the edge region. This one-eighth edge section image is characterized by displaying the microstructure of the edge region and the initial penetration of oil; the image magnification is 40x, and the field of view is 0.5mm × 0.5mm.

[0041] The quarter-edge section image acquisition step involves cutting the potato slice radially outward from the center at a point three-quarters of the radius from the center. The cut surface at this location is then placed on a microscope stage and photographed using a 20x objective lens to obtain an image of the medium structure of the sub-edge region. The quarter-edge section image is characterized by showcasing the medium structure of the sub-edge region and the starch gelatinization transition state. The image magnification is 20x, and the field of view is 1mm × 1mm.

[0042] The half-edge section image acquisition step involves cutting along the radial direction of the potato slice from the center outwards at a position half the radius from the center, obtaining the cut surface at this position. The cut surface is then placed on a microscope stage and photographed using a 10x objective lens to obtain an image of the macroscopic structure of the central region. The half-edge section image is characterized by showcasing the macroscopic structure and temperature conduction effect of the central region, with an image magnification of 10x and a field of view of 2mm × 2mm.

[0043] The diameter section image acquisition step includes completely cutting along the diameter direction of the potato slice, placing the cut surface on the image acquisition stage, and taking a macro lens to obtain an image of the complete internal cross section. The diameter section image is characterized by presenting the complete internal layer structure of the potato slice, the temperature gradient effect in the thickness direction, and the distribution of the degree of dehydration. The image covers the complete section and has a resolution of 1024×1024 pixels.

[0044] The surface image features reflect the surface texture distribution, color uniformity, and surface defects of the potato slices. The image resolution is 2048×2048 pixels, and the color space is RGB mode.

[0045] The heat transfer physical fitting equation set includes the temperature gradient conduction equation, the moisture migration equation, and the oil permeation equation.

[0046] The temperature gradient conduction equation is used to describe the temperature field distribution inside the potato slice. The inputs include heater power, ambient temperature, slice thickness parameters, thermal conductivity coefficient, and time variable. The output is the temperature distribution matrix of each layer.

[0047] The moisture migration equation is used to calculate the migration pattern of moisture in potato slices during the dehydration process. The inputs include initial moisture content, vacuum degree, temperature gradient, slice thickness parameter and diffusion coefficient, and the output is the spatiotemporal distribution function of moisture content.

[0048] The oil permeation equation is used to predict the penetration depth and distribution of oil in potato slices. The inputs include oil temperature, vacuum degree, porosity, surface tension coefficient and permeation time, and the output is the oil concentration distribution curve.

[0049] The bidirectional neural network model has a dual-channel architecture consisting of a forward propagation neural network and a backward propagation neural network. The forward propagation neural network includes an input layer, three hidden layers, and an output layer, with each hidden layer containing 512 neurons. The backward propagation neural network includes an input layer, two hidden layers, and an output layer, with each hidden layer containing 256 neurons. The parameters of the multi-head attention mechanism in the bidirectional neural network model are determined based on the heater power ratio, temperature gradient range, and dehydration rate threshold.

[0050] The training dataset establishment step of the bidirectional neural network model includes dividing the data of 750 potato chip samples obtained from 50 batches of experiments into a training set of 525 samples, a validation set of 150 samples, and a test set of 75 samples in a ratio of 7:2:1. Each sample contains a 128-dimensional potato chip comprehensive quality feature vector and a corresponding 4-dimensional experimental parameter vector.

[0051] The bidirectional neural network model training steps include updating parameters using the Adam optimization algorithm, setting the learning rate to 0.001, the batch size to 32, and the number of training rounds to 500. Training stops when the validation set loss function does not decrease for 10 consecutive rounds. Finally, the model achieves a prediction accuracy of over 92% on the test set.

[0052] The selection criteria for the optimal parameter combination are: moisture content controlled within the range of 8% to 12%, oil content controlled within the range of 15% to 20%, crispness value greater than 85N, and surface color parameter L in the comprehensive quality feature vector of potato slices. * Parameter combinations with values ​​in the range of 65 to 75.

[0053] Specifically, S07 utilizes a trained bidirectional neural network model employing a back-inference optimization method. The preset optimal feature vector for fried potato chips is input as the target output to the model's output layer. Backpropagation is performed through the neural network branch, and gradient descent and Jacobian matrix calculations are used to find the input parameter combination that most closely approximates the target feature vector. The target feature vector is defined as a 128-dimensional standard high-quality potato chip feature vector, with constraints including upper heater power P1∈[100,500]W, middle heater power P2∈[150,600]W, lower heater power P3∈[200,700]W, frying temperature T∈[80,120]℃, vacuum degree V∈[0.01,0.08]MPa, and frying time t∈[8,20]min. By randomly initializing different combinations of starting parameters multiple times and performing a reverse inference process, each inference iteration is 1000 rounds with a learning rate set to 0.0001. When the convergence error of the objective function is less than 0.001, a single inference is stopped, and finally 20 different combinations of optimized parameters are obtained, each containing numerical solutions for 6 optimized variables.

[0054] S08 specifically involves conducting small-scale experiments on each optimized parameter combination. Following steps S01-S04, five potato slices with a thickness three times their normal thickness are prepared, and the same experimental procedures and data acquisition processes are executed. A trained bidirectional neural network model is used to predict the 128-dimensional comprehensive quality feature vector of each sample. The arithmetic mean of the feature vectors of the five samples is calculated as the average comprehensive quality feature vector corresponding to that parameter combination. A cosine similarity algorithm is used to calculate the similarity between this average vector and the feature vector of a preset standard high-quality potato slice. When the similarity between the average comprehensive quality feature vector of the potato slice and the feature vector of the standard high-quality potato slice is greater than 0.85, the corresponding experimental parameter combination is recorded as the optimal parameter combination. Based on the optimal parameter combination, a layered heating strategy is formulated and optimized processing is implemented. If the similarity of all 20 parameter combinations does not reach the 0.85 threshold, then adjust the learning rate of the back-inference to the range of 0.00005-0.0005, increase the number of iterations per inference to 1500 rounds, and relax the parameter boundary constraints by 15% in each case. At the same time, introduce ±2% random perturbation into the target feature vector to generate multiple similar target vectors for parallel back-inference. Repeat step S07 until the optimal parameter combination that meets the similarity requirement is obtained.

[0055] The specific implementation methods of the above steps are described in detail below.

[0056] The specific implementation of step S01 involves peeling selected potatoes and slicing them into uniform slices with a thickness of 15mm using a high-precision slicing device, ensuring that the quality error of each slice is controlled within ±2%. Circular infrared temperature monitoring markers with a diameter of 1.5mm are prepared using ceramic material. These markers have an emissivity of 0.95, accurately reflecting surface temperature changes. The markers are evenly pasted onto the upper and lower surfaces of the potato slices in a 3×3 grid array, with a spacing of 3mm between adjacent markers, ensuring that the temperature monitoring coverage area reaches more than 85% of the slice area. Simultaneously, K-type thermocouple probes with a diameter of 0.8mm are implanted at depths of 25%, 50%, and 75% of the slice thickness using a micro-drilling device. The implantation angle is perpendicular to the slice surface, and the distance between the probe tip and the center of the slice is controlled within 2mm, for real-time monitoring of the temperature distribution at different internal layers. The purpose of this step is to establish a complete temperature monitoring system, providing an accurate data foundation for subsequent heat transfer process analysis and model construction.

[0057] The specific implementation of step S02 involves constructing a comprehensive experimental setup comprising a vacuum chamber, a layered heating system, a temperature monitoring system, and a data acquisition system. The vacuum chamber is made of stainless steel, with an effective volume of 50L, and is equipped with a combined mechanical vacuum pump and molecular pump pumping system, enabling precise vacuum control within the range of 0.001–0.1 MPa. The layered heater assembly uses resistance wire heating, with the upper heater having a power range of 100–500W, the middle heater 150–600W, and the lower heater 200–700W. Each heater is controlled by an independent power regulator with an adjustment accuracy of ±1W. The infrared temperature monitoring system uses a high-resolution infrared thermal imager with a detection accuracy of ±0.1℃ and a sampling frequency of 10Hz. The data acquisition system uses a multi-channel data acquisition card, capable of simultaneously acquiring multi-dimensional data such as temperature, pressure, and power, with a sampling accuracy of 16 bits. The purpose of this step is to establish a high-precision, multi-parameter controllable experimental platform, ensuring the accuracy and reproducibility of experimental data.

[0058] The specific implementation of step S03 is based on orthogonal experimental design theory to design a multi-factor, multi-level experimental scheme, selecting frying temperature, vacuum degree, power ratio of upper, middle, and lower layer heaters, and frying time as the main control factors. The frying temperature is set to five levels: 80℃, 90℃, 100℃, 110℃, and 120℃; the vacuum degree is set to four levels: 0.01MPa, 0.03MPa, 0.05MPa, and 0.07MPa; different combinations of power ratios are used; and the frying time is set to four levels: 8min, 12min, 16min, and 20min. A Latin square experimental design method is used to arrange 50 batches of experiments, each batch containing 15 potato slices of the same size to ensure the statistical significance of the experimental samples. During the experiment, the ambient temperature is strictly controlled at 25±2℃, and the relative humidity at 60±5%. The interval between each batch of experiments is no less than 30 minutes to eliminate the influence of equipment thermal inertia. The purpose of this step is to obtain sufficient training data through systematic experimental design, covering the full range of combinations of process parameters.

[0059] The specific implementation of step S04 involves a combination of multi-scale, multi-angle image acquisition and physical performance testing. Surface image acquisition uses an industrial camera with a resolution of 2048×2048 pixels, taking vertically downward images under standardized lighting conditions. The camera is 30cm away from the sample to ensure complete imaging of the entire surface. Edge section image acquisition uses a precision cutting device to cut at preset positions. The eighth-order edge is imaged using a 40x microscope objective with a 0.5×0.5mm field of view; the quarter-order edge is imaged using a 20x microscope objective with a 1×1mm field of view; and the half-order edge is imaged using a 10x microscope objective with a 2×2mm field of view. Diameter section imaging uses a macro lens to capture the entire cross-section with a resolution of 1024×1024 pixels. Moisture content testing uses an infrared moisture analyzer with an accuracy of ±0.1%. Oil content testing uses Soxhlet extraction with petroleum ether as the extraction solvent. Brittleness testing uses a texture analyzer for three-point bending tests with a loading speed of 1mm / s. Color testing was performed using a colorimeter to measure the L, a, and b values. This step aims to comprehensively obtain data on the appearance, internal structure, and physical properties of the potato slices.

[0060] The specific implementation of step S05 involves using computer vision technology to extract multi-scale image features and construct a comprehensive quality evaluation system. Texture feature extraction employs the gray-level co-occurrence matrix method to calculate statistical parameters such as contrast, homogeneity, and entropy. Color feature extraction calculates statistical quantities such as mean, variance, and skewness in the RGB and HSV color spaces. Structural feature extraction uses morphological operations to identify structural defects such as pores and cracks, calculating geometric parameters such as porosity and connectivity. Image features are time-stamped with physical test data such as moisture content, oil content, brittleness value, and color parameters, and then fused using principal component analysis to generate a 128-dimensional comprehensive quality feature vector. This feature vector includes multi-dimensional feature parameters such as surface roughness index, internal pore distribution parameters, color uniformity coefficient, moisture content gradient distribution, oil penetration depth distribution, brittleness spatial variation rate, and structural integrity evaluation index. The purpose of this step is to convert multi-source heterogeneous data into a unified numerical feature representation, providing standardized input for machine learning models.

[0061] The specific implementation of step S06 involves constructing a bidirectional neural network model that integrates heat transfer physics and deep learning technology. The model embeds a temperature gradient conduction equation to describe the evolution of the internal temperature field of the potato slices. This equation takes heater power distribution, ambient temperature, slice thickness parameters, and thermal conductivity as inputs and outputs a temperature distribution matrix for each layer. A moisture migration equation, based on Fick's diffusion law, describes the mass transfer behavior of moisture during dehydration. It takes initial moisture content, vacuum degree, and temperature gradient as inputs and outputs a spatiotemporal distribution function of moisture content. An oil permeation equation, based on Darcy's law and capillary permeation theory, describes the permeation behavior of oil in porous media. It takes oil temperature, vacuum degree, and porosity as inputs and outputs an oil concentration distribution curve. Model training employs backpropagation and gradient descent optimization methods. The loss function uses a weighted combination of mean squared error and cross-entropy. An early stopping strategy is adopted to prevent overfitting when the validation set accuracy does not improve for 10 consecutive rounds. The purpose of this step is to establish a nonlinear mapping relationship between process parameters and product quality, achieving quality prediction and parameter optimization.

[0062] The specific implementation of step S07 involves using the trained bidirectional neural network model for backpropagation parameter optimization. A preset 128-dimensional optimal potato chip feature vector is used as the target output, and inverse calculation is performed through the backpropagation neural network branch of the model. Gradient descent algorithm and Jacobian matrix numerical solution method are employed to find the input parameter combination that makes the model output closest to the target feature vector. Parameter boundary constraints are set, including upper heater power 100–500W, middle heater power 150–600W, lower heater power 200–700W, frying temperature 80–120℃, vacuum degree 0.01–0.08MPa, and frying time 8–20min. A multi-starting-point random initialization strategy is adopted, with 1000 iterations per inference round and a learning rate set to 0.0001. A single inference iteration stops when the convergence error of the objective function is less than 0.001. Through 20 parallel inference processes with different initializations, 20 sets of candidate optimized parameter combinations are obtained. The purpose of this step is to utilize the model's inverse reasoning capability to deduce the optimal process parameter combination from the desired product quality.

[0063] The specific implementation of step S08 involves conducting small-scale verification experiments and similarity evaluations of the candidate parameter combinations. Five potato samples are prepared for each parameter group for verification experiments, and the experimental procedure is strictly followed according to steps S01 to S04 to ensure consistency of experimental conditions. A trained bidirectional neural network model is used to predict the 128-dimensional comprehensive quality feature vector of each sample, and the arithmetic mean of the feature vectors of the five samples is calculated as the representative feature vector of that parameter combination. A cosine similarity algorithm is used to calculate the similarity value between the average feature vector and the feature vector of a preset standard high-quality potato slice, with a similarity threshold set at 0.85. When the similarity is greater than the threshold, the corresponding parameter combination is recorded as the optimal parameter combination. If none of the candidate combinations meet the threshold requirement, the back-inference parameters are adjusted: the learning rate is adjusted to the range of 0.00005 to 0.0005, the number of iterations is increased to 1500 rounds, and the parameter boundary constraints are relaxed by 15%. Simultaneously, ±2% random perturbation is introduced into the target feature vector to generate multiple similar target vectors for parallel inference. The optimization process is repeated until the optimal parameter combination that meets the requirements is obtained. The purpose of this step is to verify the reliability and practicality of the optimized parameters through experiments.

[0064] The bidirectional neural network model employs a dual-channel parallel architecture, comprising two branches: a forward propagation neural network and a backpropagation neural network. The forward propagation neural network predicts product quality characteristics from process parameters. Its structure includes a 128-dimensional input layer that receives process parameters such as frying temperature, vacuum level, power of the upper, middle, and lower heaters, and frying time. The hidden layers use a three-layer fully connected structure, each containing 512 neurons. The activation function is a modified linear unit function, and the dropout ratio is set to 0.3 to prevent overfitting. The output layer is a 128-dimensional fully connected layer, corresponding to the various dimensions of the potato chip's comprehensive quality feature vector. The backpropagation neural network is responsible for inferring process parameters from the desired product quality characteristics. Its input layer is a 128-dimensional quality feature vector, and the hidden layers use a two-layer structure, each containing 256 neurons. The output layer is a 6-dimensional process parameter vector.

[0065] The bidirectional neural network model integrates a multi-head attention mechanism with eight attention heads to capture the correlation and importance weights between different process parameters. The parameters of the attention mechanism are adaptively adjusted based on the heater power ratio, temperature gradient range, and dehydration rate threshold, learning the dependencies between parameters through a self-attention mechanism. The two neural network branches are coupled by sharing a weight matrix, ensuring consistency between forward and backward inference. The model also integrates a physical constraint layer, embedding a set of heat transfer physics equations, including the temperature gradient conduction equation based on Fourier's heat transfer law, the moisture migration equation based on Fick's law, and the oil permeation equation based on Darcy's law, ensuring that the prediction results conform to the physical mechanisms.

[0066] The process of establishing the training dataset for the bidirectional neural network model first involves quality screening of 750 potato chip samples obtained from 50 batches of experiments, removing abnormal samples and retaining valid data. Data preprocessing includes feature standardization, missing value imputation, and outlier detection. A stratified random sampling method is used to divide the data into a training set of 525 samples, a validation set of 150 samples, and a test set of 75 samples in a 7:2:1 ratio, ensuring the uniformity of distribution of different process parameter combinations within each subset. Each sample data point contains a 128-dimensional comprehensive quality feature vector and a corresponding 6-dimensional experimental parameter vector, forming an input-output pair for supervised learning.

[0067] The bidirectional neural network model was trained using the Adam optimization algorithm for parameter updates. This algorithm, combining momentum and adaptive learning rate adjustment, effectively handles sparse gradients and non-convex optimization problems. The initial learning rate was set to 0.001, employing an exponential decay strategy, decreasing by a factor of 0.95 after every 100 training epochs. The batch size was set to 32 to balance training efficiency and gradient estimation accuracy. The total number of training epochs was set to 500. An early stopping strategy was used to monitor the validation set loss function; training was automatically stopped when the validation loss did not decrease for 10 consecutive epochs to prevent overfitting. During training, a learning rate scheduler dynamically adjusted the learning rate, halving it when the validation metric stagnated. The final trained model achieved a prediction accuracy of over 92% on the independent test set, with the root mean square error controlled within 0.05, demonstrating good convergence and generalization ability.

[0068] The key technical ideas of this invention are mainly reflected in four aspects: precise control of layered heating, multi-scale image feature fusion, physical constraint neural network modeling, and bidirectional inference optimization.

[0069] Layered heating precision control technology achieves precise regulation of the temperature gradient along the thickness of potato slices by independently adjusting the power of the upper, middle, and lower layers of heaters. This technology offers significant advantages over traditional single-layer heating methods, enabling differentiated heating strategies based on the internal heat transfer characteristics and dehydration requirements of the potato slices. This avoids surface overheating and internal underheating, significantly improving product quality uniformity. Traditional frying methods rely on natural convection heat transfer, resulting in uneven temperature distribution and unstable quality. Layered heating technology, however, achieves precise regulation of the heat transfer process by actively controlling the temperature distribution of each layer.

[0070] Multi-scale image feature fusion technology combines multi-level image information, such as macroscopic surface morphology, mesoscopic cross-sectional structure, and microscopic pore distribution, with physical performance test data to construct a high-dimensional comprehensive quality evaluation system. Compared with traditional single-index evaluation methods, this technology can more comprehensively characterize product quality features. By automatically extracting multi-dimensional features such as texture, color, and structure through computer vision technology, it avoids the subjectivity and limitations of manual evaluation and improves the objectivity and accuracy of quality evaluation.

[0071] Physically constrained neural network modeling technology embeds the physical mechanisms of heat and mass transfer into deep learning models, achieving an organic combination of data-driven and mechanism-driven approaches. Compared to purely data-driven methods, this technology offers better interpretability and generalization capabilities. By constraining the results through physical equations, it ensures that the predictions conform to objective laws, avoiding unreasonable predictions outside the training data range and improving the reliability and practicality of the model.

[0072] Bidirectional reasoning optimization technology utilizes the backward reasoning capability of neural networks to deduce the optimal combination of process parameters from the desired product quality. Compared to traditional trial-and-error methods, this technology can significantly reduce the number of experiments and development cycle, and quickly determine the optimal parameter combination through mathematical optimization methods, thus significantly improving the efficiency of process optimization.

[0073] The synergistic effect of these four key technological approaches forms a complete intelligent process optimization system. Layered heating control provides precise execution methods, multi-scale feature fusion establishes comprehensive quality evaluation standards, physical constraint modeling ensures the scientific validity of the optimization results, and bidirectional reasoning optimization enables rapid parameter determination. These four elements support and promote each other. Compared to traditional experience-based process development methods, this collaborative system enables scientific, automated, and intelligent optimization of process parameters, significantly improving product quality stability, reducing development costs, and shortening development cycles, providing crucial technical support for the intelligent upgrading of the food processing industry.

[0074] It should be noted that this invention also solves the following technical problem: Traditional frying processes rely on manual sensory judgment for product quality evaluation, lacking objective quantitative standards. In existing food processing quality control, product quality evaluation mainly relies on the experience of operators and simple physical index tests. This highly subjective evaluation method makes it difficult to establish unified quality standards, leading to significant differences in quality between different batches of products. This invention establishes an objective quality evaluation system based on multi-angle image feature extraction and machine learning algorithms, quantifying multi-dimensional information such as surface texture, internal structure, and color distribution into a 128-dimensional comprehensive quality feature vector, thus achieving a shift from subjective judgment to objective quantification. This method not only improves the accuracy and consistency of quality evaluation but also provides a quantifiable objective function for process optimization, enabling quality control to shift from passive detection to proactive prediction and regulation.

[0075] Furthermore, this invention addresses the technical problems of high experimental costs and long cycles in optimizing frying process parameters. Traditional process parameter optimization relies heavily on extensive experimental verification, requiring significant raw materials and time, and is often limited to parameter adjustments within empirical ranges, making it difficult to find the globally optimal solution. This invention, by establishing a bidirectional neural network model and utilizing back-inference technology, achieves a direct mapping from product quality requirements to process parameters, significantly reducing the number of experiments and optimization time. Simultaneously, the integrated physical heat transfer equations in the model ensure the scientific validity and reliability of the prediction results, avoiding the overfitting problem that may occur with purely data-driven methods, and providing an effective approach for rapid and accurate optimization of process parameters.

[0076] Specifically, the principle of this invention is as follows: The core of this invention's ability to solve the problem of uneven temperature distribution lies in establishing an intelligent control system based on a combination of physical heat transfer models and machine learning algorithms. First, the design of the layered heater group follows the principles of heat transfer. Through independent power control of the upper, middle, and lower layers, the heating intensity of different layers can be adjusted according to the heat conduction characteristics along the thickness of the potato slices, overcoming the problem of excessive temperature gradients caused by a single heating source. A multi-point temperature monitoring network composed of infrared temperature monitoring patches and micro-temperature probes can acquire real-time temperature distribution data on the surface and interior, providing accurate feedback information for layered heating. The heat transfer physical fitting equations integrated in the bidirectional neural network model include temperature gradient conduction equations, moisture migration equations, and oil penetration equations. These equations are based on Fourier's heat transfer law and Fick's diffusion law, accurately describing the physical changes during frying. By training the neural network with experimental data combined with the physical model, the neural network learns the complex nonlinear relationships between process parameters and temperature distribution, dehydration degree, and oil penetration depth. The reverse reasoning mechanism utilizes the gradient descent algorithm and the Jacobian matrix to deduce the optimal heater power ratio based on preset quality targets, achieving reverse design from product quality requirements to process parameters. Multi-angle image feature extraction technology quantifies the impact of temperature distribution uniformity on the internal structure of a product by analyzing the microstructural changes of different cross-sections, providing a reliable quality evaluation index for model optimization.

[0077] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0078] In this embodiment, the specific implementation of steps S01-S04 is the same as described above, and will not be repeated in detail here.

[0079] The specific implementation of step S05 involves using computer vision technology to extract multi-scale image features and construct a comprehensive quality evaluation system, with the comprehensive quality feature vector F... c The specific formula for calculation is as follows:

[0080] F c =α1F t +α2F s +α3F p +α4F m +α5F o +α6F b +α7F i ;

[0081] In the formula, F c F is a 128-dimensional comprehensive quality feature vector; t F is the surface texture feature vector; s For structural feature vectors; F p F is the physical performance feature vector; m F is the eigenvector of water content; o F is the feature vector of oil content; b F is the brittleness feature vector; i Let be the color feature vector; α1, α2, α3, α4, α5, α6, and α7 are weight coefficients, satisfying . Furthermore, the weight coefficients range from 0.1 to 0.2. Among them, the surface texture feature vector F... t The gray-level co-occurrence matrix (GLCM) is calculated using the following steps: Step 1: Convert the RGB image to a grayscale image; Step 2: Calculate the GLCM; Step 3: Extract statistical parameters such as contrast, homogeneity, and entropy. The structural feature vector F is then obtained. s The method employs morphological operations, including step 1: binarizing the microscope image; step 2: identifying the pore structure using opening and closing operations; and step 3: calculating the porosity and connectivity parameters.

[0082] The specific implementation of step S06 is to construct a bidirectional neural network model that integrates the physical mechanism of heat transfer, wherein the temperature gradient conduction equation is specifically expressed as follows:

[0083]

[0084] In the formula, T is the internal temperature of the potato slice; t is time; and λ is the thermal diffusivity. Here, P1 is the Laplace operator; P2, P3 are the power of the upper, middle, and lower layer heaters, respectively; ρ is the density of the potato slices; c p V is the specific heat capacity. s h represents the volume of the potato slices. c T is the convective heat transfer coefficient; a Let the ambient temperature be denoted as . The moisture migration equation is expressed as follows:

[0085]

[0086] In the formula, M is the moisture content; D mk is the water diffusion coefficient. v P is the vacuum dehydration coefficient. v P represents the vacuum level. vs Here is the saturated vapor pressure. The specific expression of the grease permeation equation is as follows:

[0087]

[0088] In the formula, C o D represents the concentration of oil and fat. o k is the oil diffusion coefficient. p φ is the permeability coefficient; φ is the porosity; C os The concentration of oil on the surface is denoted as λ. The thermal diffusivity λ was obtained experimentally, including step 1: preparing standard potato slice samples; step 2: determining thermal conductivity using the transient hot wire method; and step 3: calculating the thermal diffusivity based on density and specific heat capacity. The moisture diffusivity D is also mentioned. m The diffusion coefficient was obtained experimentally, including step 1: conducting dehydration experiments under different temperature and humidity conditions; step 2: measuring the moisture content distribution at different times; and step 3: obtaining the diffusion coefficient by fitting the data using Fick's second law.

[0089] The specific implementation of step S07 involves optimizing the back-inference parameters using the trained bidirectional neural network model. The specific expression for the objective function J is as follows:

[0090]

[0091] In the formula, This is the overall quality feature vector predicted by the model; λ represents the target comprehensive quality feature vector. r p is the regularization coefficient. i This refers to the i-th process parameter; Let be the reference value for the i-th process parameter; ||·|| is the Euclidean norm. The gradient descent update formula is expressed as follows:

[0092]

[0093] In the formula, and ... Let J be the partial derivative of the objective function with respect to the i-th process parameter. The Jacobian matrix J is... matrix The specific expression is as follows:

[0094]

[0095] In the formula, f j p is the j-th comprehensive quality characteristic; kThis refers to the k-th process parameter; Let be the partial derivative of the j-th feature with respect to the k-th parameter. Where λ is the regularization coefficient. r The learning rate is obtained using cross-validation, including the following steps: Step 1: Set the candidate value range to 0.001–0.1; Step 2: Perform 5-fold cross-validation on each candidate value; Step 3: Select the value with the smallest validation error as the optimal regularization coefficient. The learning rate η is obtained using an adaptive adjustment method, including the following steps: Step 1: Initially set to 0.0001; Step 2: Check the change in the objective function every 50 iterations; Step 3: If the change is less than a threshold for 10 consecutive iterations, the learning rate is halved.

[0096] The specific implementation of step S08 involves verifying and evaluating the candidate parameter combinations. The similarity calculation formula S is specifically expressed as follows:

[0097]

[0098] In the formula, S is the cosine similarity; This represents the average comprehensive quality feature vector of the 5 samples. Let be the feature vector of standard high-quality potato chips; · represents the vector dot product; ||·|| represents the Euclidean norm of the vector. Average comprehensive quality feature vector. The specific formula for calculation is as follows:

[0099]

[0100] In the formula, F c，i Let be the comprehensive quality feature vector of the i-th sample. Wherein, the feature vector of the standard high-quality potato slice is... The similarity was obtained through a combination of expert evaluation and physicochemical testing, including: Step 1: Selecting 10 food engineering experts to conduct sensory evaluations of 100 potato chips prepared using different processes; Step 2: Measuring the moisture content, oil content, crispness, and other physicochemical indicators of each potato chip; Step 3: Using principal component analysis to fuse the expert scores and physicochemical indicators to generate a standard feature vector. The similarity threshold of 0.85 was set based on statistical analysis, including: Step 1: Collecting feature vectors from 200 potato chips of different quality grades; Step 2: Calculating the similarity distribution between the high-quality samples and the standard vector; Step 3: Selecting the lower limit of the 95% confidence interval as the threshold.

[0101] It should be explained that the principle of the comprehensive quality feature vector fusion formula is based on the theory of multi-source information fusion. It integrates feature information from different dimensions into a unified quality evaluation index through a weighted summation method. This formula uses a linear weighted combination, which can flexibly adjust the importance weights of various features. Compared with traditional single-indicator evaluation methods, it achieves a multi-dimensional comprehensive evaluation of potato chip quality, significantly improving the accuracy and comprehensiveness of quality judgment and avoiding quality misjudgments caused by the limitations of single indicators.

[0102] The temperature gradient conduction equation is based on Fourier's law of heat transfer and the principle of energy conservation. It describes the spatiotemporal evolution of the temperature field inside potato slices using partial differential equations. The left side of the equation represents the rate of temperature change over time, while the right side consists of three terms: a heat conduction term, a heat source term, and a convective heat transfer term. Compared to traditional empirical models, this equation has stronger physical support and can accurately predict the internal temperature distribution of potato slices under stratified heating conditions. This provides a scientific theoretical basis for optimizing process parameters and significantly improves the accuracy of heating process control.

[0103] The moisture transfer equation, based on Fick's diffusion law and the principle of mass conservation, describes the mass transfer process of moisture within potato slices under vacuum conditions. The first term on the right-hand side of the equation is the molecular diffusion term, and the second term is the vacuum-driven convective mass transfer term. This equation considers the influence of vacuum level on the dehydration process and, compared to traditional atmospheric pressure dehydration models, can more accurately predict dehydration behavior under low-temperature vacuum conditions, providing a precise mathematical tool for controlling product moisture content.

[0104] The oil permeation equation, based on Darcy's law and capillary permeation theory, describes the permeation behavior of oils in porous media. The first term on the right-hand side of the equation is the diffusion mass transfer term, and the second term is the pore-driven convective mass transfer term. This equation considers the influence of porosity on oil permeation and, compared to traditional surface adsorption models, can more comprehensively describe the distribution of oils within potato chips, providing a theoretical basis for controlling the oil content of products.

[0105] The objective function expression is based on optimization theory, achieving parameter optimization by minimizing the difference between the predicted and target values. The first term of the function is the fitting error term, ensuring that the model output closely approximates the target quality characteristics; the second term is the regularization term, preventing the parameters from deviating excessively from a reasonable range. Compared to the single error minimization method, this objective function introduces a parameter constraint mechanism, ensuring both optimization accuracy and the physical rationality of the parameters, significantly improving the reliability and practicality of the optimization results.

[0106] The gradient descent update formula is based on numerical optimization theory, iteratively approximating the optimal solution. The formula uses first-order gradient information to guide the parameter update direction, and the learning rate controls the update step size. Compared to traditional experimental methods, this formula can quickly converge to the optimal parameter combination, significantly reducing optimization time and experimental costs, and achieving automated and intelligent optimization of process parameters.

[0107] The Jacobian matrix expression is based on the theory of multivariable function differentials, describing the sensitivity distribution of comprehensive quality characteristics to process parameters. The matrix elements are the partial derivatives of each characteristic with respect to each parameter, reflecting the degree of influence of parameter changes on the quality characteristics. Compared to traditional sensitivity analysis methods, this matrix can simultaneously consider the interactions of multiple parameters, providing a mathematical basis for parameter importance ranking and optimization strategy formulation, significantly improving the targeting and efficiency of the optimization process.

[0108] The principle behind cosine similarity calculation is based on vector space theory. It measures the degree of similarity between two vectors by calculating the cosine of the angle between them. The numerator of the formula is the dot product of the vectors, the denominator is the product of the vector lengths, and the result ranges from -1 to 1.

[0109]

[0110] This formula is the formula for calculating the dot product of vectors, which represents the summation of the product of corresponding elements of two vectors.

[0111]

[0112] This formula is the Euclidean length formula for vectors, representing the square root of the sum of the squares of the vector's elements. Compared to the traditional Euclidean distance method, this formula is not affected by vector length and only focuses on directional similarity, making it more suitable for evaluating the similarity of high-dimensional feature vectors and able to accurately determine whether product quality meets expected standards.

[0113] To better understand and implement this invention, the following is a specific application scenario of this invention, Example 2:

[0114] Traditional frying methods struggle to achieve uniform heating of thick potato slices, often resulting in burnt exteriors and undercooked interiors, or excessively high oil content. The technical team decided to solve this problem by employing an optimized method of low-temperature vacuum frying and layered heating of potato slices three times their thickness.

[0115] The technical team first prepared potato slices three times thick according to step S01. Fresh potatoes were selected and cut into 12mm thick circular slices. Infrared temperature monitoring patches with a diameter of 1.5mm were then affixed to the surface of each slice in a 3mm grid pattern, totaling 36 monitoring points per slice. Simultaneously, thermocouple probes with a diameter of 0.8mm were implanted at depths of 25%, 50%, and 75% within the slice to monitor internal temperature changes. The infrared temperature monitoring patches were made of ceramic material with an emissivity of 0.95, accurately reflecting the surface temperature distribution.

[0116] The technical team then constructed a low-temperature vacuum frying experimental setup. The layered heater assembly consisted of three independently controlled resistance wire heaters: an upper layer (P1 ∈ [100, 500] W), a middle layer (P2 ∈ [150, 600] W), and a lower layer (P3 ∈ [200, 700] W). An infrared temperature monitoring system was able to collect surface temperature data in real time and feed it back to the power control system. The vacuum control system could maintain the vacuum level within the range of 0.01–0.08 MPa. The data acquisition system recorded all temperature and pressure data at a frequency of 1 Hz.

[0117] In step S03, the technical team designed 50 batches of comparative experiments. Each batch contained 15 potato slices three times the thickness of the target slice, with different combinations of process parameters for comparison. The frying temperature ranged from 80 to 120°C, the vacuum level ranged from 0.01 to 0.08 MPa, different combinations of heater power ratios were used, and the frying time was set from 8 to 20 minutes. Table 1 shows some of the experimental parameter settings:

[0118] Table 1. Parameter settings for low-temperature vacuum frying experiment

[0119] batch Deep frying temperature (°C) Vacuum degree (MPa) <![CDATA[P1(W)]]> <![CDATA[P2(W)]]> <![CDATA[P3(W)]]> Time (min) 1 85 0.02 150 200 250 12 2 95 0.04 200 300 400 15 3 105 0.06 250 400 500 18 4 115 0.05 300 450 600 14 5 90 0.03 180 280 350 16

[0120] During the data acquisition process in step S04, the technical team conducted comprehensive image acquisition and performance testing on each batch of fried potato chips. Surface image acquisition was performed using an industrial camera with a resolution of 2048×2048 pixels, taken vertically at a height of 30cm from the sample. Edge cross-section image acquisition included cross-sections at the one-eighth, one-quarter, and one-half edge positions, captured using 40x, 20x, and 10x objective lenses respectively, with field of view ranges of 0.5mm×0.5mm, 1mm×1mm, and 2mm×2mm. Diameter cross-section images were captured using a macro lens with a resolution of 1024×1024 pixels, clearly revealing the complete internal layered structure.

[0121] Test data acquisition includes the determination of moisture content, oil content, crispness value, and color parameters. For example... Figure 2 The figure shows the trend of moisture content variation of potato slices under different process parameters. Moisture content was determined by the drying method, with the slices dried to constant weight in an oven at 105℃. Oil content was determined by Soxhlet extraction using petroleum ether as the extraction solvent. Brittleness was determined using a texture analyzer, with a 2mm diameter cylindrical probe used to puncture the sample at a speed of 1mm / s to record the maximum puncture force. Color parameters were determined using a colorimeter. * a * b * value.

[0122] In step S05, the technical team extracted feature parameters from various images. Texture features included parameters such as contrast, uniformity, and correlation of the gray-level co-occurrence matrix. Color features included statistics such as the mean, standard deviation, and skewness of the RGB three channels. Structural features included geometric parameters such as porosity, connectivity, and fractal dimension. The image features were spatiotemporally aligned and fused with the test data to generate a 128-dimensional comprehensive quality feature vector. The feature vector for each sample included parameters such as surface texture roughness, internal porosity, color uniformity, moisture content distribution, oil penetration depth, brittleness distribution, and structural integrity. Table 2 shows the main parameter classifications and representative values ​​of the comprehensive quality feature vector for potato chips.

[0123] Table 2. Main Parameters of the Comprehensive Quality Feature Vector of Potato Chips

[0124]

[0125] The surface texture roughness category comprises 25 feature parameters. Of these, 18 parameters (F001-F018) are derived from gray-level co-occurrence matrix analysis, including texture statistics such as contrast, correlation, energy, homogeneity, and entropy in four directions: 0°, 45°, 90°, and 135°. Four to five key texture features are extracted for each direction. The seven parameters (F019-F025) describe the distribution characteristics of texture energy in different frequency domains, extracting the frequency domain characteristics of the surface texture through Fourier transform and wavelet transform. The internal porosity category contains 23 feature parameters. The 13 parameters (F026-F038) calculate geometric features such as average pore diameter, pore density, pore shape factor, and pore orientation angle at different depths through image analysis. The 10 parameters (F039-F048) evaluate the connectivity of the pore structure, including parameters such as connected pore volume fraction, tortuosity coefficient, and permeability estimate. The color uniformity category contains 30 feature parameters. The 17 parameters (F049-F065) calculate the statistical characteristics of the red, green, and blue channels in different regions of the surface, including mean, standard deviation, skewness, and kurtosis, extracting 5-6 statistics for each channel. The 13 parameters (F066-F078) describe the hue, saturation, and brightness distribution characteristics in the HSV color space, as well as L... * a * b *Uniformity indices for color space. The moisture content distribution category includes 24 characteristic parameters. The 14 parameters (F079-F092) calculate spatial distribution characteristics such as radial moisture content gradient, angular moisture content variation, and temperature-moisture content correlation coefficient through infrared thermography and microprobe data. The 10 parameters (F093-F102) assess axial moisture content variability, including statistical parameters such as standard deviation of moisture content at different depths, coefficient of variation, and distribution skewness. The oil penetration depth category contains 20 characteristic parameters. The 13 parameters (F103-F115) analyze dynamic characteristics such as radial and axial gradient distribution of oil concentration, penetration front location, and penetration rate through chemical staining and microscopic observation. The 7 parameters (F116-F122) assess the uniformity of oil distribution, including parameters such as penetration uniformity index, local concentration coefficient of variation, and standard deviation of penetration depth. The crispness distribution category includes four feature parameters F123-F126. Multi-point crispness testing is used to calculate the coefficient of variation, maximum-to-minimum ratio, spatial gradient, and anisotropy coefficient of crispness values ​​at different locations, reflecting the spatial consistency of the potato chip's mechanical properties. The structural integrity category includes two feature parameters F127-F128, representing two-dimensional and three-dimensional fractal dimensions, respectively. Box counting and mass radius methods are used to calculate the complexity and self-similarity characteristics of the potato chip's internal structure.

[0126] These specific feature parameters can be manually set based on experience and process practices in the food engineering field, or directly determined from the default output parameters of image processing algorithms, statistical analysis software, and physical measuring instruments. In practical applications, it is not necessary to excessively pursue which feature parameters are essential, because bidirectional neural network models have powerful feature learning and weight allocation capabilities. They can automatically identify the feature combinations that contribute most to product quality prediction during training and assign higher weight coefficients to important features. The redundancy of multidimensional feature vectors actually helps improve the robustness and generalization ability of the model. Even if some feature parameters are missing or abnormal under certain conditions, the model can still maintain high prediction accuracy through other relevant features. In addition, changes in raw materials and process conditions in different batches may lead to changes in the importance of certain features. Maintaining the comprehensiveness of feature vectors ensures the adaptability and stability of the model in diverse production environments.

[0127] In step S06, the technical team constructed a bidirectional neural network model. The model comprises a dual-channel architecture of a forward propagation neural network and a backward propagation neural network. The forward propagation neural network consists of an input layer, three hidden layers, and an output layer, with each hidden layer containing 512 neurons. The backward propagation neural network consists of an input layer, two hidden layers, and an output layer, with each hidden layer containing 256 neurons. The model integrates a set of heat transfer physics fitting equations, including the temperature gradient conduction equation, the moisture migration equation, and the oil permeation equation. The temperature gradient conduction equation takes heater power, ambient temperature, sheet thickness parameter, thermal conductivity coefficient, and time variable as inputs and outputs a temperature distribution matrix for each layer. The moisture migration equation takes initial water content, vacuum degree, temperature gradient, sheet thickness parameter, and diffusion coefficient as inputs and outputs a spatiotemporal distribution function of water content. The oil permeation equation takes oil temperature, vacuum degree, porosity, surface tension coefficient, and permeation time as inputs and outputs an oil concentration distribution curve.

[0128] The training dataset consisted of 750 potato chip samples, divided into a training set of 525 samples, a validation set of 150 samples, and a test set of 75 samples, in a 7:2:1 ratio. Each sample contained a 128-dimensional comprehensive quality feature vector and a corresponding 6-dimensional experimental parameter vector (frying temperature, vacuum level, power of the top, middle, and bottom heaters, and frying time). The Adam optimization algorithm was used for parameter updates, with a learning rate of 0.001, a batch size of 32, and 500 training epochs. Training was stopped when the validation set loss function did not decrease for 10 consecutive epochs. The final model achieved a prediction accuracy of 94.2% on the test set.

[0129] In step S07, the technical team used a trained bidirectional neural network model for backpropagation optimization. The preset optimal feature vector for fried potato chips was used as the target output input to the model's output layer, and inverse calculation was performed through the backpropagation neural network branch. The target feature vector was set as a 128-dimensional standard high-quality potato chip feature vector, with a target moisture content of 9.5%, a target oil content of 17.5%, a target crispness value of 90N, and a surface color parameter L. * The target value is 70. Constraints include the power of the upper heater P1 ∈ [100, 500] W, the power of the middle heater P2 ∈ [150, 600] W, the power of the lower heater P3 ∈ [200, 700] W, the frying temperature T ∈ [80, 120] ℃, the vacuum degree V ∈ [0.01, 0.08] MPa, and the frying time t ∈ [8, 20] min. By randomly initializing different combinations of initial parameters 20 times and performing a reverse inference process, with each inference iteration consisting of 1000 rounds and a learning rate set to 0.0001, a single inference iteration stops when the convergence error of the objective function is less than 0.001, ultimately resulting in 20 different combinations of optimized parameters.

[0130] In the small-scale test of step S08, the technical team prepared five potato slices with a thickness of three times for each optimized parameter combination to conduct a verification experiment. For example... Figure 3 The distribution of validation results for the optimized parameter combination is shown. A trained bidirectional neural network model was used to predict the 128-dimensional comprehensive quality feature vector for each sample. The arithmetic mean of the feature vectors from five samples was calculated as the average comprehensive quality feature vector corresponding to this parameter combination. The cosine similarity algorithm was used to calculate the similarity between this average vector and the feature vector of a preset standard high-quality potato slice. After validation, the 12th parameter combination achieved a similarity of 0.887, making it the optimal parameter combination. The specific parameters for this combination were: frying temperature 98℃, vacuum degree 0.035MPa, upper heater power 280W, middle heater power 420W, lower heater power 580W, and frying time 14.5min.

[0131] Table 3 shows the quality index results of the optimal parameter combination verification experiment:

[0132] Table 3. Verification results of the optimal parameter combination quality index

[0133] Quality Indicators target value Measured average Standard deviation Pass rate (%) Moisture content (%) 9.5 9.3 0.4 96 Oil content (%) 17.5 17.8 0.6 94 Brittleness value (N) 90 91.2 2.1 98 <![CDATA[L * Value 70 69.5 1.2 92

[0134] The technical team then developed a layered heating strategy based on the optimal parameter combination. For example... Figure 4 The diagram shows the timing of the layered heating power control strategy. During the first 3 minutes of frying, the upper layer heater power is set to 240W, the middle layer heater power to 360W, and the lower layer heater power to 520W, creating an upward temperature gradient to promote rapid internal moisture migration. During the middle 4-10 minutes of frying, the upper layer heater power is adjusted to 300W, the middle layer heater power to 450W, and the lower layer heater power to 600W, maintaining a stable temperature distribution. During the later 11-14.5 minutes of frying, the upper layer heater power is adjusted to 320W, the middle layer heater power to 480W, and the lower layer heater power to 650W, ensuring the surface achieves the desired crispness.

[0135] The technical team used optimal parameters to conduct continuous production verification, producing a total of 500 batches of potato chip products. For example... Figure 5 The diagram shows the control chart for key quality indicators during continuous production. Moisture content is controlled within the range of 8.9–9.7%, oil content within the range of 17.2–18.4%, and brittleness value within the range of 89–93 N. * The values ​​were controlled within the range of 68.8 to 70.2, and all indicators met the preset quality standards. The product passed sensory evaluation and physicochemical index testing, and its crisp texture, moderate oil content, and uniform color met the quality requirements of high-end restaurant chains.

[0136] This invention represents a significant technological advancement over traditional frying methods. Traditional frying methods use a single temperature, making it difficult to achieve a uniform temperature distribution throughout thick potato slices. This results in excessive dehydration of the outer layer and insufficient dehydration of the inner layer, leading to a burnt exterior and undercooked interior. This invention employs layered heating technology. By independently controlling the power of the upper, middle, and lower heaters, a reasonable temperature gradient from the inside out is established, allowing each layer of the potato slice to undergo dehydration and oil penetration at an appropriate temperature. The vacuum environment lowers the boiling point of the oil, reducing high-temperature damage to the potato slice surface while promoting rapid evaporation of internal moisture, achieving efficient dehydration at low temperatures. A bidirectional neural network model combined with heat transfer physics equations accurately predicts the mapping relationship between process parameters and product quality. Optimal process parameters are quickly obtained through back-reasoning, avoiding extensive trial-and-error. An infrared temperature monitoring system enables real-time feedback control of surface temperature, ensuring the accuracy and consistency of the heating process. A 128-dimensional quality feature vector, integrating image features and physicochemical indicators, comprehensively describes product quality, providing a scientific basis for process optimization.

[0137] It should be noted that the variables involved in this invention are explained in detail in Table 4.

[0138] Table 4. Variable Explanation Table

[0139]

[0140]

[0141] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. An optimized method for low-temperature vacuum frying and layered heating of potato slices three times thick, characterized in that, The process included preparing potato slices three times thick and attaching infrared temperature monitoring patches to the surface of each slice, covering the slices with a grid pattern, and embedding micro-temperature probes at different depths within the slices; constructing a low-temperature vacuum frying experimental setup, including a layered heater group, an infrared temperature monitoring system, a vacuum control system, and a data acquisition system; the layered heater group consisting of an upper, middle, and lower heater, with independent power control for each layer; designing multiple sets of comparative experiments, each with different frying temperatures, vacuum levels, heater power ratios, and frying time parameters; and acquiring multi-angle images and test data for each batch of fried potato slices, including surface images, images of the eighth-to-last edge section, and images of the quarter-to-last edge section. Cross-sectional images, half-edge cross-sectional images, and diameter cross-sectional images were used to test moisture content, oil content, crispness value, and color parameters. Texture, color, and structural features were extracted from various images. The image features were spatiotemporally aligned with the test data and fused to generate a comprehensive quality feature vector for potato chips. A bidirectional neural network model containing a set of heat transfer physics fitting equations was constructed and trained using experimental parameters as input and the comprehensive quality feature vector of potato chips as output. The trained bidirectional neural network model was used to perform back-reasoning based on the preset optimal fried potato chip feature vector to obtain multiple sets of optimized parameter combinations. Small-scale experiments were conducted on each set of optimized parameter combinations to calculate the similarity between the average comprehensive quality feature vector and the standard vector. The optimal parameter combination was selected as the optimized layered heating process parameters.

2. The optimized method for low-temperature vacuum frying and layered heating of potato slices with a thickness of three times as described in claim 1, characterized in that, The heat transfer physics fitting equation set includes the temperature gradient conduction equation, the moisture migration equation, and the oil penetration equation. The temperature gradient conduction equation is used to describe the temperature field distribution inside the potato slice, the moisture migration equation is used to calculate the migration law of moisture in the potato slice during the dehydration process, and the oil penetration equation is used to predict the penetration depth and distribution of oil in the potato slice.

3. The optimized method for low-temperature vacuum frying and layered heating of potato slices with a thickness of three times as described in claim 2, characterized in that, The bidirectional neural network model has a dual-channel architecture consisting of a forward propagation neural network and a backward propagation neural network. The forward propagation neural network includes an input layer, three hidden layers, and an output layer, with each hidden layer containing 512 neurons. The backward propagation neural network includes an input layer, two hidden layers, and an output layer, with each hidden layer containing 256 neurons. The training dataset for the bidirectional neural network model is established by dividing 750 potato chip samples obtained from 50 batches of experiments into a training set of 525 samples, a validation set of 150 samples, and a test set of 75 samples in a ratio of 7:2:

1. Each sample contains a 128-dimensional comprehensive quality feature vector of potato chips and a corresponding 4-dimensional experimental parameter vector.

4. The optimized method for low-temperature vacuum frying and layered heating of potato slices with a thickness of three times as described in claim 3, characterized in that, The layered heater group adopts resistance wire heating. The power range of the upper layer heater is P1∈[100, 500]W, the power range of the middle layer heater is P2∈[150, 600]W, and the power range of the lower layer heater is P3∈[200, 700]W. The power of each layer can be adjusted independently.

5. The optimized method for low-temperature vacuum frying and layered heating of potato slices with a thickness of three times as described in claim 4, characterized in that, The comprehensive quality feature vector of potato chips includes a numerical combination of 128 feature parameters, including surface texture roughness, internal porosity, color uniformity, moisture content distribution, oil penetration depth, brittleness distribution, and structural integrity.

6. The optimized method for low-temperature vacuum frying and layered heating of potato slices with a thickness of three times as described in claim 5, characterized in that, The steps for surface image acquisition are as follows: place the potato slices 30cm directly above the image acquisition platform, and use an industrial camera with a resolution of 2048×2048 pixels to take a vertical downward shot to obtain a complete surface image.

7. The optimized method for low-temperature vacuum frying and layered heating of potato slices with a thickness of three times as described in claim 6, characterized in that, The steps for acquiring an image of the one-eighth edge section are as follows: cut along the radial direction of the potato slice from the center outwards at a position seven-eighths of the radius from the center to obtain the cut surface, place the cut surface on the microscope stage and take a picture with a 40x objective lens to obtain an image of the microstructure of the edge region.

8. The optimized method for low-temperature vacuum frying and layered heating of potato slices with triple thickness according to claim 7, characterized in that, The steps for acquiring images of a quarter-edge section are as follows: cut along the radial direction of the potato slice from the center outwards at a position three-quarters of the radius from the center to obtain the section. Place the section on the microscope stage and take a picture with a 20x objective lens to obtain an image of the medium structure of the sub-edge region.

9. The optimized method for low-temperature vacuum frying and layered heating of potato slices with triple thickness according to claim 8, characterized in that, The steps for acquiring a half-edge cross-section image are as follows: cut along the radius of the potato slice from the center outwards at a position half the radius from the center to obtain the cross-section. Place the cross-section on the microscope stage and take a picture with a 10x objective lens to obtain a macroscopic structural image of the middle region.

10. The optimized method for low-temperature vacuum frying and layered heating of potato slices with a thickness of three times as described in claim 9, characterized in that, The steps for acquiring the diameter cross-section image are as follows: the potato slice is completely cut along its diameter, the cut surface is placed on the image acquisition stage, and a macro lens is used to capture the complete internal cross-section image. The image covers the entire cut surface and has a resolution of 1024×1024 pixels.

Images