Nut cooking pretreatment slicing processing method

Abstract

The invention provides a nut cooking pretreatment slicing processing method, and belongs to the technical field of nut cooking pretreatment slicing processing. Badam raw materials are subjected to particle size grading, cooking temperature and cooking time are set in groups, and cooking peeling is completed; detecting the moisture penetration depth of the peeled almonds batch by batch by utilizing online near-infrared moisture gradient detection, returning substandard batches to supplement cooking, baking standard batches in a tunnel oven, and transferring the standard batches into a heat preservation intermediate bin to maintain the surface temperature of the nuts; the single feeding amount and the feeding interval time are calculated based on an optimal slice batch distribution algorithm of thermodynamic afterheat attenuation dynamics, and a conveyor belt control system feeds the materials into a slicing machine in batches in an optimal slice window according to the parameters to complete slicing; the yield of each batch of slices and all process data are synchronously input into an artificial intelligence network training data set for iterative updating, and the technical problem that the yield of the slices is low due to lack of dynamic regulation and control of a kernel temperature window in the nut slicing processing process is solved.

Description

Technical Field

[0001] This invention belongs to the field of nut steaming and cooking pretreatment slicing technology, and more specifically, relates to a method for nut steaming and cooking pretreatment slicing. Background Technology

[0002] Slicing almonds and other nuts is a crucial link in the deep processing industry chain of nuts. In traditional slicing, the pre-treatment of steaming and cooking usually relies on the operator's experience to set a fixed steaming temperature and time. The roasting and heat preservation stages also use static parameter control. The feeding schedule of the slicing machine depends on manual judgment or a fixed rhythm. The entire process lacks the ability to perceive the internal moisture state and surface temperature of the kernels in real time.

[0003] Traditional methods cannot adaptively adjust process parameters based on differences in almond particle size during the steaming stage, resulting in insufficient moisture penetration in some batches and substandard kernel toughness. During the transfer stage from roasting to slicing, the surface temperature of the kernels naturally decreases with the environment. Traditional methods lack the ability to quantitatively predict and actively control the temperature decay process, and the surface temperature of the kernels often deviates from the suitable range during slicing, resulting in increased chipping rate and decreased yield.

[0004] In current nut slicing processing, the multi-step process from steaming and roasting to slicing is lengthy, and the surface temperature decay of the kernels is affected by multiple factors such as particle size distribution, ambient temperature, and batch size. Existing static scheduling schemes cannot respond in real time to the temperature decay differences between batches, making it difficult to ensure that each batch of material is sliced ​​within the optimal temperature window. In other words, existing technologies suffer from a low slicing yield due to the lack of dynamic control over the kernel temperature window during nut slicing. Summary of the Invention

[0005] In view of this, the present invention provides a method for slicing nuts after steaming and cooking, which can solve the technical problem of low slicing yield caused by the lack of dynamic control of kernel temperature window during nut slicing process in the prior art.

[0006] This invention is implemented as follows: This invention provides a method for pre-processing and slicing nuts by steaming and boiling, comprising the following steps:

[0007] The raw material of almonds was subjected to particle size detection. The raw material was divided into several particle size interval groups according to the particle size grading boundary. The corresponding cooking temperature and cooking time were set for each particle size interval group to complete the group cooking and peeling.

[0008] After peeling, the almonds are subjected to online near-infrared moisture gradient detection in batches to obtain the moisture penetration depth detection value. The moisture penetration depth detection value is compared with the moisture penetration depth target value. Batches with the moisture penetration depth detection value reaching the moisture penetration depth target value proceed to the next step, while batches that do not reach the moisture penetration depth target value are returned for supplementary steaming.

[0009] Almonds whose moisture penetration depth detection value reaches the target value are sent into a tunnel oven for baking at the baking temperature. After the baking time is over, the almonds are transferred to an insulated intermediate chamber, where they are kept warm at the temperature of the insulated chamber.

[0010] A multimodal temporal fusion attention network collects real-time temporal data of kernel surface temperature, near-infrared moisture gradient, and tool vibration acceleration, outputs the optimal slicing window start time and optimal slicing window end time, and transmits them to the conveyor belt control system.

[0011] Based on the optimal batch allocation algorithm of thermodynamic residual temperature decay kinetics, the current ambient temperature, the rated processing rate of the slicer and the critical temperature of slicing are used as inputs to calculate the amount of material fed at one time and the feeding interval time. The conveyor belt control system adjusts the feeding rate according to the start time of the optimal slicing window, the end time of the optimal slicing window, the amount of material fed at one time and the feeding interval time, so that the almonds are fed into the slicer in batches from the start time of the optimal slicing window to the end time of the optimal slicing window to complete the slicing.

[0012] After slicing, the slice yield of the current batch is calculated. The slice yield, along with the time-series data of the current batch's cooking temperature, cooking time, baking temperature, baking time, heat preservation chamber temperature, kernel surface temperature, near-infrared moisture gradient, and blade vibration acceleration, are synchronously entered into the multimodal temporal fusion attention network training dataset for iterative updates of the multimodal temporal fusion attention network.

[0013] Specifically, the determination of the particle size classification boundary involves conducting a particle size distribution statistical experiment on almond raw materials with a particle size distribution density of no less than the batch threshold of the particle size classification experiment, and selecting the particle size nodes with the larger distribution density as the particle size classification boundary according to the particle size distribution density.

[0014] Specifically, the determination of cooking temperature and cooking time involves conducting gradient cooking experiments for each particle size range group for no less than the batch threshold of the gradient cooking experiment, establishing a moisture penetration depth prediction model, and iteratively optimizing the cooking temperature range and cooking time range corresponding to each particle size range group with the goal of converging the moisture penetration depth detection value to the moisture penetration depth target value.

[0015] Specifically, the determination of the target value for water penetration depth involves slicing almonds with different water penetration depth detection values ​​in a slicing experiment with a batch number of slices not less than the threshold of the water penetration experiment batch. The minimum water penetration depth detection value corresponding to the first time the slice yield reaches the qualified yield threshold is taken as the target value for water penetration depth.

[0016] Among them, the online near-infrared moisture gradient detection specifically utilizes the absorption difference of near-infrared light of different wavelengths in the surface and subsurface layers of the kernel to invert the moisture content gradient at different depths below the surface of the kernel, thereby indirectly characterizing the moisture penetration depth detection value. After scanning each almond, the moisture penetration depth detection value is output.

[0017] Specifically, the determination of baking temperature and baking time involves conducting slicing experiments with no fewer than the batch threshold under different combinations of baking temperature and baking time. The slicing yield and the time series data of kernel surface temperature are used as evaluation indicators to determine the combination of baking temperature and baking time that ensures the kernel surface temperature remains above the critical slicing temperature after baking.

[0018] The determination of the critical slicing temperature involves conducting slicing experiments at different kernel surface temperatures, recording the breakage rate corresponding to each kernel surface temperature, and using the kernel surface temperature at which the breakage rate first exceeds the upper limit of the acceptable breakage rate as the critical slicing temperature.

[0019] Specifically, determining the temperature of the heat preservation chamber involves measuring the decay rate of the surface temperature of almond kernels under different heat preservation chamber temperatures. The minimum maintenance value of the heat preservation chamber temperature is determined through iterative experiments, with the constraint that the surface temperature of the kernels should not be lower than the critical temperature of the slice from the start time to the end time of the optimal slicing window.

[0020] The input layer of the multimodal temporal fusion attention network receives three heterogeneous sensor data streams: kernel surface temperature time-series data, near-infrared moisture gradient time-series data, and tool vibration acceleration time-series data. These three heterogeneous sensor data streams are fed into three independent one-dimensional temporal convolutional branches for local feature encoding. The encoding results are then dynamically fused into a cross-modal attention module and then fed into a cyclic skip unit. The output layer outputs the optimal slice window start time, the optimal slice window end time, and the expected slice yield range.

[0021] Among them, the network end of the multimodal temporal fusion attention network is embedded with a physical constraint layer. The residual temperature decay differential equation is discretized and embedded into the loss function in the form of a penalty term, so that the predicted values ​​at the start and end times of the optimal slice window do not violate the laws of thermodynamics and physics.

[0022] The training of the multimodal temporal fusion attention network involves collecting production data of no less than the initial training batch threshold to form the initial training dataset. Subsequent batches of production data are continuously added after review to form an online expanded training dataset. An incremental fine-tuning is triggered after each batch of production data of no less than the incremental fine-tuning batch threshold is added. During incremental fine-tuning, the parameters of the three one-dimensional temporal convolutional branches are frozen, and only the parameters of the cross-modal cross attention module and the loop skip unit are updated.

[0023] Among them, the optimal batch allocation algorithm for slicing based on thermodynamic residual temperature decay kinetics is specifically based on Newton's law of cooling to establish a differential equation for the thermal decay of almond particle group. The Monte Carlo sampling method is used to generate a probability distribution sample curve for the decay process of the surface temperature of the kernel of the particle group. The duration for which the average surface temperature of the kernel of the particle group is higher than the critical temperature for slicing is used as the effective slicing window duration. Then, the Lagrange multiplier method is used to solve for the optimal single feeding amount and the optimal feeding interval time under constraints.

[0024] The comprehensive thermal scoring function calculates a comprehensive thermal score based on three data points: the average surface temperature of the kernel, the temperature of the heat preservation chamber, and the current processing rate of the slicer. The comprehensive thermal score is used to adjust the feeding rate parameter of the multimodal temporal fusion attention network. When the comprehensive thermal score is lower than the pause feeding threshold, feeding is paused and the multimodal temporal fusion attention network is triggered to re-predict the start and end times of the optimal slicing window.

[0025] The weighting coefficient of the comprehensive thermal score was determined through a multi-factor slicing experiment with a threshold of no less than the batch threshold of the weighted regression experiment. The slicing yield was used as the response variable, and the normalized values ​​of the mean surface temperature of the kernel, the normalized values ​​of the temperature of the heat preservation chamber, and the normalized values ​​of the current processing rate of the slicer were used as independent variables in a multiple regression analysis.

[0026] The experiment included three batch thresholds for particle size classification: 3 batches, with particle size ranges of 8–10 mm, 10–12 mm, and 12–14 mm; a batch threshold for cooking gradient experiments: 5 batches, with cooking temperatures ranging from 95–100℃ and cooking times ranging from 30–60 min depending on the particle size range; a batch threshold for moisture penetration experiments: 5 batches, with a yield qualification threshold of 70% and a moisture penetration depth target of 0.3 mm; a batch threshold for baking experiments: 5 batches, with a baking temperature of 70℃ and a baking time ranging from 25–35 min; a temperature range for the insulation chamber: 55–65℃; an initial training batch threshold of 200 batches; an incremental fine-tuning batch threshold of 50 batches; and a batch threshold for weighted regression experiments: 10 batches, with a comprehensive thermal scoring weight coefficient. The value range is 0.4 to 0.6. The value range is 0.2 to 0.4. The value range is 0.1 to 0.3.

[0027] This invention solves the technical problem of low slicing yield caused by the lack of dynamic control of kernel temperature window during nut slicing by establishing an adaptive cooking parameter system based on particle size classification, a closed-loop online near-infrared moisture gradient detection system, a temperature maintenance mechanism for the heat preservation intermediate chamber, a multimodal temporal fusion attention network for optimal slicing window prediction, and a dynamic feeding schedule based on the optimal slicing batch allocation algorithm of thermodynamic residual heat decay kinetics.

[0028] This invention integrates multi-source heterogeneous sensing signals with a thermodynamic physical model, enabling the system to perceive three interrelated physical quantities in real time: kernel moisture state, surface temperature dynamics, and equipment vibration. Thus, even when temperature decay is heterogeneous due to differences in particle size distribution between batches, the system can still accurately predict the optimal slicing window and dynamically allocate the feeding cycle, avoiding the slicing debris caused by temperature window mismatch in traditional static scheduling schemes.

[0029] In summary, this invention solves the technical problem mentioned in the background art of low slicing yield caused by the lack of dynamic control of kernel temperature window during nut slicing process. Attached Figure Description

[0030] Figure 1 This is a flowchart of the method of the present invention.

[0031] Figure 2 This is a graph showing the change in water penetration depth of almonds in different particle size ranges with cooking time.

[0032] Figure 3 A comparison chart of the optimal slice window output by the multimodal temporal fusion attention network and the temperature decay curve of the kernel surface.

[0033] Figure 4 This is a graph showing the trend of the overall thermal score as the batch of materials is fed. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0035] like Figure 1 The diagram shown is a flowchart of a nut pretreatment and slicing method provided by the present invention, which includes the following steps:

[0036] S01. Perform particle size detection on almond raw materials, divide the raw materials into several particle size interval groups according to the particle size grading boundary, set the corresponding cooking temperature and cooking time for each particle size interval group, and complete the group cooking and peeling.

[0037] S02. Online near-infrared moisture gradient detection is performed on each batch of peeled almonds to obtain the moisture penetration depth detection value. The moisture penetration depth detection value is compared with the moisture penetration depth target value. Batches with moisture penetration depth detection values ​​that reach the moisture penetration depth target value are entered into S03, and batches that do not reach the moisture penetration depth target value are returned to S01 for supplementary steaming.

[0038] S03. Almonds whose moisture penetration depth detection value reaches the target value are sent into the tunnel oven and baked at the baking temperature. After the baking time is over, the almonds are transferred to the heat preservation intermediate chamber and the heat preservation intermediate chamber is used to keep the almonds warm at the temperature of the heat preservation chamber.

[0039] S04. The multimodal temporal fusion attention network collects real-time temporal data of kernel surface temperature, near-infrared moisture gradient, and tool vibration acceleration, outputs the start and end times of the optimal slicing window for the current batch, and transmits the start and end times of the optimal slicing window to the conveyor belt control system.

[0040] S05. Based on the optimal batch allocation algorithm of thermodynamic residual temperature decay kinetics, the current ambient temperature, the rated processing rate of the slicer and the critical temperature of slicing are used as inputs to calculate the amount of material fed at one time and the feeding interval time. The conveyor belt control system adjusts the feeding rate according to the start time of the optimal slicing window, the end time of the optimal slicing window, the amount of material fed at one time and the feeding interval time, so that the almonds are fed into the slicer in batches from the start time of the optimal slicing window to the end time of the optimal slicing window to complete the slicing.

[0041] S06. After slicing, calculate the slice yield of the current batch. Synchronously input the slice yield along with the time series data of the current batch's cooking temperature, cooking time, baking temperature, baking time, heat preservation chamber temperature, kernel surface temperature, near-infrared moisture gradient, and blade vibration acceleration into the multimodal temporal fusion attention network training dataset for iterative updates of the multimodal temporal fusion attention network.

[0042] The method for obtaining the particle size classification boundary is as follows: conduct particle size distribution statistical experiments on no less than 3 batches of almond raw materials, select the particle size nodes with larger distribution density as the particle size classification boundary according to the particle size distribution density, and the typical particle size ranges are 8-10mm, 10-12mm, and 12-14mm. The particle size classification boundary is determined after being verified by experimental data from no less than 3 batches.

[0043] The cooking temperature and cooking time were obtained as follows: at least five gradient cooking experiments were conducted for each particle size range. The moisture penetration depth measured by an online near-infrared moisture gradient detector was used as the evaluation index. The moisture penetration depth was recorded for each combination of cooking temperature and cooking time, and a moisture penetration depth prediction model was established. The formula for the moisture penetration depth prediction model is as follows: ;in The measured value of water penetration depth ( ), The target value for water infiltration depth ( ), The cooking temperature ( ), The reference cooking temperature ( ), Steaming time ( ), The baseline cooking time ( ), Let be the dimensionless mapping function obtained by fitting experimental data, with With the goal of convergence to 1, the cooking temperature range and cooking time range corresponding to each particle size range are iteratively optimized. The typical cooking temperature range is 95-100℃, and the cooking time is 30-60 min depending on the particle size range.

[0044] The target value for water penetration depth was obtained as follows: in no less than 5 batches of slicing experiments, almonds with water penetration depth detection values ​​of 0.1mm, 0.2mm, 0.3mm, 0.4mm, and 0.5mm were sliced ​​respectively, and the slice yield corresponding to each water penetration depth detection value was recorded. The minimum water penetration depth detection value corresponding to the first slice yield reaching 70% was taken as the target value for water penetration depth, and it was determined to be 0.3mm after verification by no less than 5 batches of experiments.

[0045] Among them, the online near-infrared moisture gradient detection is a detection method based on the principle of near-infrared spectral absorption. It utilizes the absorption difference of near-infrared light of different wavelengths in the surface and subsurface layers of the kernel to invert the moisture content gradient at different depths below the surface of the kernel, thereby indirectly characterizing the moisture penetration depth detection value; the detection scans each almond and outputs the moisture penetration depth detection value.

[0046] The typical baking temperature is 70℃, and the typical baking time range is 25-35 min. The baking temperature and baking time are obtained by conducting no less than 5 batches of slicing experiments under different combinations of baking temperature and baking time. The slicing yield and the time series data of kernel surface temperature are used as evaluation indicators to determine the combination of baking temperature and baking time that makes the kernel surface temperature still higher than the critical temperature of slicing after baking.

[0047] The critical temperature for slicing is obtained by conducting slicing experiments at different kernel surface temperatures and recording the breakage rate corresponding to each kernel surface temperature. The kernel surface temperature at which the breakage rate first exceeds the upper limit of the acceptable breakage rate is taken as the critical temperature for slicing. This is verified through no less than 5 batches of experiments. The upper limit of the acceptable breakage rate is determined by the requirements for the breakage ratio in the product quality standard.

[0048] The method for obtaining the temperature of the heat preservation chamber is as follows: the decay rate of the surface temperature of almond kernels under different heat preservation chamber temperatures is measured. The minimum maintenance value of the heat preservation chamber temperature is determined by iterative experiments, with the constraint that the surface temperature of the kernels should not be lower than the critical temperature of the slice from the start time to the end time of the optimal slicing window. The typical range of the heat preservation chamber temperature is 55-65℃.

[0049] The insulated intermediate chamber is an independent temperature-controlled chamber located between the tunnel oven outlet and the slicer inlet. It is used to shorten the transfer time from the tunnel oven to the slicer and extend the duration for which the surface temperature of the nuts is higher than the critical temperature for slicing.

[0050] The comprehensive thermal scoring function is used to adjust the feed rate parameters of the multimodal temporal fusion attention network. The comprehensive thermal scoring function calculates the comprehensive thermal score based on three data points: the average surface temperature of the kernel, the temperature of the insulation chamber, and the current processing rate of the slicer. The formula for the comprehensive thermal score is as follows: ;in The comprehensive thermal state score is dimensionless. The average surface temperature of the kernels in the current batch ( ), The critical temperature for slicing ( ), For the temperature of the heat preservation chamber ( ), The reference temperature for the heat preservation chamber ( ), The current processing speed of the slicer ( ), The rated processing rate of the slicer ( ), , , The weighting coefficients are and satisfy the following conditions: The weighting coefficients are obtained by conducting multi-factor slice experiments in no fewer than 10 batches, using the slice yield rate as the response variable, and... , , Perform multiple regression analysis on the independent variables to determine , , The typical value ranges are as follows: , , ;when When the feed rate parameter is taken as the slicer's rated processing rate; when At that time, the feed rate parameter is taken as 80% of the slicer's rated processing rate; when At that time, the feeding rate parameter is set to 60% of the slicer's rated processing rate, and a heating command for the insulation intermediate chamber is triggered simultaneously; when When the feeding is paused, the multimodal temporal fusion attention network is triggered to re-predict the start and end times of the optimal slice window.

[0051] The specific structure of the multimodal temporal fusion attention network is as follows: The network input layer receives three types of heterogeneous sensor data streams, namely near-infrared moisture gradient time-series data, nut surface temperature time-series data, and tool vibration acceleration time-series data. These three heterogeneous sensor data streams are respectively fed into three independent one-dimensional temporal convolutional branches for local feature encoding. Each one-dimensional temporal convolutional branch consists of three stacked one-dimensional convolutional layers with a kernel size of 3–7 and a stride of 1. Each one-dimensional convolutional layer is followed by a batch normalization layer and a modified linear unit activation function. The feature sequences output from the three one-dimensional temporal convolutional branches are then fed into the cross-modal cross-attention network. The cross-modal attention module uses the output feature sequences of any two one-dimensional temporal convolutional branches as query vectors and key-value pairs to calculate the cross-modal attention weight matrix. It dynamically weights and fuses the three types of feature sequences, enabling the network to adaptively prioritize the heterogeneous sensor data stream with the strongest discriminative power for the current prediction task under different production conditions. The fused feature sequence is then fed into a recurrent skip unit, a variant of the gated recurrent unit with residual gating. This unit introduces an adaptive time-step skipping mechanism based on the standard gated recurrent unit. The network determines whether to skip the silent segment of the sensor signal based on the skipping probability distribution of the gated output at the time step dimension, directly connecting to the key variable. The network employs a multimodal temporal fusion attention network to reduce redundant time steps from interfering with the hidden state. The output layer is a fully connected layer, outputting three predicted values: the start time of the optimal slice window, the end time of the optimal slice window, and the expected slice yield range. A physical constraint layer is embedded at the network's end, discretizing the residual heat decay differential equation and embedding it as a penalty term into the loss function to ensure that the predicted values ​​for the optimal slice window start time, optimal slice window end time, and expected slice yield range do not violate thermodynamic physical laws. The steps for establishing the training dataset for this multimodal temporal fusion attention network specifically include: during the initial production line setup, manually recording no fewer than 200 batches of steaming temperatures and steam... The initial training dataset consists of time-series data on cooking time, baking temperature, baking time, heat preservation chamber temperature, nut surface temperature, near-infrared moisture gradient, and tool vibration acceleration, along with corresponding slicing yield data. This data undergoes outlier removal, time axis alignment, and normalization. Subsequent batches of production data are continuously added to this initial training dataset after review, forming an online expanded training dataset. The training steps for the multimodal temporal fusion attention network specifically include: using the sum of mean squared error loss and physical constraint penalty as the total loss function, and employing an adaptive moment estimation optimization algorithm to update parameters, with an initial learning rate set to... After every 50 training rounds, the learning rate decays to 0.5 times the previous learning rate. The number of training rounds is no less than 200. The convergence criteria are that the prediction error of the validation set slice output rate does not exceed 5% and the prediction error between the start and end times of the optimal slice window does not exceed 30 seconds. After convergence, the model weights are saved. An incremental fine-tuning is triggered after every 50 batches of new production data. During incremental fine-tuning, the parameters of the three one-dimensional temporal convolution branches are frozen, and only the parameters of the cross-modal attention module and the loop skip unit are updated.

[0052] The multimodal temporal fusion attention network integrates three types of heterogeneous physical sensor signals with physical constraints, enabling the network to simultaneously perceive three interrelated physical quantities—moisture state, temperature dynamics, and equipment vibration—when predicting the start and end times of the optimal slicing window. The cross-modal cross-attention module allows the network to dynamically adjust the dependence weights on each heterogeneous sensor data stream under different production conditions. The cyclic skip unit enhances the network's ability to capture short-lived key kernel surface temperature change nodes. The physical constraint layer embeds thermodynamic laws into the total loss function, preventing the pure data-driven model from outputting predictions that violate thermodynamic laws in sparse training sample regions. This ensures that the network continuously outputs physically consistent and highly accurate slicing window scheduling decisions under varying production environments, improving the overall slicing yield and equipment utilization rate.

[0053] Among them, the optimal slice batch allocation algorithm based on thermodynamic residual temperature decay kinetics is based on Newton's law of cooling, and a differential equation for the ensemble thermal decay of almond particles is established. The formula for the ensemble thermal decay differential equation is as follows: ;in The average surface temperature of the kernels in the grain group ( ), The current ambient temperature ( ), For time ( ), Thermal time constant ( The thermal time constant is obtained as follows: Cooling curve measurement experiments are conducted on particle groups of different particle size ranges, and the measurement data are fitted with an exponential function to extract the thermal time constant corresponding to each particle size range; at least 1000 probability distribution sample curves are generated for the decay process of the surface temperature of the kernels in the particle group using the Monte Carlo sampling method, and the duration for which the average surface temperature of the kernels in the particle group is higher than the critical temperature for slicing is statistically analyzed. The duration is defined as the effective slicing window duration; with the objective function being to maximize the proportion of material sliced ​​by the slicer within the effective slicing window duration, and with the constraints being that the single feeding amount does not exceed the product of the slicer's rated processing rate and the effective slicing window duration, and the feeding interval time is not less than the shortest safe transfer time, the optimal single feeding amount and the optimal feeding interval time are solved using the Lagrange multiplier method, and a dynamic batch scheduling scheme that adapts to the current ambient temperature and the slicer's rated processing rate is output. The dynamic batch scheduling scheme includes the single feeding amount and the feeding interval time.

[0054] Among them, the Monte Carlo sampling method is a numerical statistical method based on random number generation. By repeatedly sampling the input parameters according to their probability distribution, it simulates the system output distribution under a large number of random scenarios. In this scheme, the input parameters are the particle size and corresponding thermal time constant of each particle size interval group, which are used to generate a probability distribution sample curve of the decay of the surface temperature of the kernel of the particle group. The statistical law replaces the single deterministic decay curve, adapts to the heterogeneity of the surface temperature decay of the kernel caused by the difference in particle size distribution between batches, and enables the dynamic batch scheduling scheme to maximize the proportion of material slices within the effective slicing window time when the statistical characteristics of the particle group change.

[0055] Among them, the Lagrange multiplier method is a mathematical method for solving the extremum of the objective function under constraints. By introducing multipliers, the constraints and the objective function are combined into an augmented objective function. The partial derivatives of all variables and multipliers are taken simultaneously and set to zero to obtain the optimal solution that satisfies the constraints. In this scheme, the Lagrange multiplier method is used to solve for the optimal single feeding amount and the optimal feeding interval time under the constraints of single feeding amount and feeding interval time, so as to maximize the proportion of material that the slicer completes slicing within the effective slicing window time. The output results are the single feeding amount and feeding interval time in the dynamic batch scheduling scheme.

[0056] Among them, the optimal batch allocation algorithm for slicing based on thermodynamic residual temperature decay kinetics combines Newton's law of cooling with Monte Carlo sampling method to extend the surface temperature decay process of the kernel group from a single deterministic curve to a probability distribution description, so that the dynamic batch scheduling scheme can adapt to the heterogeneity of kernel surface temperature decay caused by the difference in particle size distribution between batches; the Lagrange multiplier method solves the optimal single feeding amount and the optimal feeding interval time under multiple constraints, so that each batch of material is sliced ​​to the maximum extent within the effective slicing window time, thereby reducing the material cooling loss caused by the mismatch of kernel surface temperature window and improving the effective output and slicing yield of the entire line.

[0057] In a specific embodiment of the present invention concerning the pre-treatment and slicing of almonds, the parameters involved include particle size classification boundary, cooking temperature, cooking time, target value of moisture penetration depth, baking temperature, baking time, heat preservation chamber temperature, critical slicing temperature, effective slicing window duration, single feeding amount, feeding interval time, and comprehensive thermal scoring weighting coefficient. , , The thermal time constant is as follows: the particle size classification boundary is determined through at least three batches of particle size distribution statistical experiments, with typical particle size ranges of 8–10 mm, 10–12 mm, and 12–14 mm; the typical cooking temperature range is 95–100℃, and the cooking time is 30–60 min depending on the particle size range; the target value for moisture penetration depth is determined to be 0.3 mm after verification through at least five batches of slice yield experiments; the typical baking temperature is 70℃, and the typical baking time range is 25–35 min; the typical temperature range for the heat preservation chamber is 55–65℃; the critical temperature for slicing is determined through at least five batches of breakage rate experiments; the effective slicing window duration is 5–10 min; the single feeding amount and feeding interval are output online by an optimal slice batch allocation algorithm based on thermodynamic residual heat decay kinetics; and the comprehensive thermal scoring weight coefficient is also specified. , , Determined by no fewer than 10 batches of multifactor regression experiments, the typical value ranges are as follows: , , The thermal time constant was determined by fitting the exponential function of the cooling curves of each particle size range group.

[0058] The specific implementation of step S01 is as follows: After the almond raw materials arrive at the site, a particle size sorting device is used to perform full-batch particle size testing on the raw materials, and the particle size distribution data is recorded. Through at least three batches of particle size distribution statistical experiments, particle size nodes with higher distribution densities are selected as particle size grading boundaries according to particle size distribution density, dividing the raw materials into three particle size interval groups: 8-10 mm, 10-12 mm, and 12-14 mm. At least five gradient cooking experiments are conducted for each particle size interval group. The moisture penetration depth measured by an online near-infrared moisture gradient detector is used as the evaluation index to establish a moisture penetration depth prediction model. The model form is as follows: ,in This is the measured value of water penetration depth. This represents the target value for water penetration depth. This refers to the cooking temperature. Based on the cooking temperature, This refers to the steaming / cooking time. Based on the cooking time, This is a dimensionless mapping function obtained by fitting experimental data. With the goal of convergence to 1, the cooking temperature and cooking time ranges for each particle size group were iteratively optimized. The typical cooking temperature range was 95–100℃, and the cooking time ranged from 30–60 min for each particle size group, completing the grouped cooking and peeling process. The particle size classification boundary was determined after verification using no fewer than three batches of experimental data to ensure that the grouped cooking parameters matched the particle size characteristics of the raw materials.

[0059] The specific implementation of step S02 is as follows: Peeled almonds are batch-by-batch fed into an online near-infrared moisture gradient detector. Utilizing the absorption differences of near-infrared light at different wavelengths in the surface and subsurface layers of the kernel, the moisture content gradient at different depths below the kernel surface is inverted, thus indirectly characterizing the moisture penetration depth detection value. The moisture penetration depth detection value is output after scanning each almond. The method for determining the target value of the moisture penetration depth is as follows: In at least five batches of slicing experiments, almonds with moisture penetration depth detection values ​​of 0.1mm, 0.2mm, 0.3mm, 0.4mm, and 0.5mm are sliced, and the slice yield corresponding to each moisture penetration depth detection value is recorded. The minimum moisture penetration depth detection value corresponding to the first slice yield reaching 70% is taken as the target value of the moisture penetration depth, which has been verified and determined to be 0.3mm. The moisture penetration depth detection value is compared with 0.3mm. Batches that reach the target value proceed to step S03, while batches that do not reach the target value return to step S01 for supplementary steaming, forming a closed-loop quality control system.

[0060] The specific implementation of step S03 is as follows: Almonds whose moisture penetration depth detection value reaches the target value are sent into a tunnel oven for roasting at the specified roasting temperature. The roasting temperature and roasting time are determined by conducting at least five batches of slicing experiments under different combinations of roasting temperature and roasting time. Using the slice yield and kernel surface temperature time series data as evaluation indicators, the roasting temperature and roasting time combination that ensures the kernel surface temperature remains above the critical slicing temperature after roasting is determined. The typical roasting temperature is 70℃, and the typical roasting time range is 25–35 minutes. The critical slicing temperature is determined by conducting slicing experiments at different kernel surface temperatures and recording the breakage rate corresponding to each kernel surface temperature. The kernel surface temperature at which the breakage rate first exceeds the upper limit of the acceptable breakage rate is taken as the critical slicing temperature, verified through at least five batches of experiments. After the roasting time is completed, the almonds are transferred to an insulated intermediate chamber. This insulated intermediate chamber is located between the tunnel oven outlet and the slicer inlet and is an independently temperature-controlled chamber used to shorten the transfer time and extend the duration for which the kernel surface temperature remains above the critical slicing temperature. The method for determining the temperature of the heat preservation chamber is as follows: the decay rate of the surface temperature of almond kernels under different heat preservation chamber temperatures is measured. The minimum maintenance value of the heat preservation chamber temperature is determined by iterative experiments, with the constraint that the surface temperature of the kernels should not be lower than the critical temperature of the slice from the start time to the end time of the optimal slicing window. The typical value range is 55-65℃.

[0061] The specific implementation of step S04 is as follows: The input layer of the multimodal temporal fusion attention network synchronously receives three heterogeneous sensor data streams: kernel surface temperature time-series data, near-infrared moisture gradient time-series data, and tool vibration acceleration time-series data. These three heterogeneous sensor data streams are fed into three independent one-dimensional temporal convolutional branches. Each branch consists of three stacked one-dimensional convolutional layers with kernel sizes of 3-7 and strides of 1. Each layer is followed by a batch normalization layer and a modified linear unit activation function to complete local feature encoding. The feature sequences output from the three branches are fed into a cross-modal attention module. Using the output feature sequences of any two branches as query vectors and key-value pairs, a cross-modal attention weight matrix is ​​calculated, and the three feature sequences are dynamically weighted and fused. The fused feature sequences are then fed into a cyclic skip unit, which is a variant of a gated cyclic unit with residual gating. An adaptive time-step skipping mechanism is introduced, determining whether to skip the silent segment of the sensor signal based on the skipping probability distribution of the gated output, and directly connecting to key change nodes. The output layer is a fully connected layer, outputting three predicted values: the start time of the optimal slice window, the end time of the optimal slice window, and the expected slice yield range. A physical constraint layer is embedded at the network's end, discretizing the residual temperature decay differential equation and embedding it as a penalty term into the loss function to ensure the predicted values ​​do not violate thermodynamic physical laws. During training, the sum of the mean squared error loss and the physical constraint penalty term is used as the total loss function, and an adaptive moment estimation optimization algorithm is employed for parameter updates, with an initial learning rate set to... After every 50 training rounds, the learning rate decays to 0.5 times that of the previous round, with a minimum of 200 training rounds. The convergence criterion is that the prediction error of the validation set slice output rate does not exceed 5% and the prediction error at the optimal slice window time does not exceed 30 seconds. An incremental fine-tuning is triggered after every 50 batches of new production data. During incremental fine-tuning, the parameters of the three one-dimensional temporal convolution branches are frozen, and only the parameters of the cross-modal attention module and the recurrent skip unit are updated.

[0062] The specific implementation of step S05 is as follows: Based on the optimal slice batch allocation algorithm of thermodynamic residual temperature decay kinetics, and based on Newton's law of cooling, a differential equation for the thermal decay of almond particle group is established. ,in This represents the average surface temperature of the kernels in the grain group. The current ambient temperature. For time, The thermal time constant is derived by measuring the cooling curves of particle groups with different particle size ranges and then fitting the data with an exponential function. At least 1000 probability distribution sample curves were generated using Monte Carlo sampling to study the surface temperature decay process of the kernels. The duration for which the average surface temperature of the kernels exceeded the critical slicing temperature was defined as the effective slicing window duration, typically 5–10 minutes. The objective function was to maximize the proportion of material sliced ​​by the slicer within the effective slicing window duration. Constraints included that the single feed rate did not exceed the product of the slicer's rated processing rate and the effective slicing window duration, and that the feeding interval was not less than the shortest safe transfer time. The optimal single feed rate and optimal feeding interval were solved using the Lagrange multiplier method. The comprehensive thermal scoring function was based on the average surface temperature of the kernels. Temperature of the insulated warehouse and the current processing speed of the slicer Calculate the overall thermal score ,when The feed rate is taken as the rated value, when When the value is taken as 80% of the rated value, At that time, the temperature is taken at 60% of the rated value and the heating command of the insulation intermediate chamber is triggered simultaneously. The feeding is paused and a multimodal temporal fusion attention network is triggered to re-predict the optimal slicing window. The conveyor belt control system adjusts the feeding rate based on the start time and end time of the optimal slicing window, the amount of material fed at one time, and the feeding interval, so that the almonds are fed into the slicer in batches within the optimal slicing window to complete the slicing.

[0063] The specific implementation of step S06 is as follows: After slicing, the slice yield of the current batch is calculated. The slice yield, along with the time-series data of the current batch's cooking temperature, cooking time, baking temperature, baking time, heat preservation chamber temperature, kernel surface temperature, near-infrared moisture gradient, and cutter vibration acceleration, are synchronously entered into the multimodal time-series fusion attention network training dataset. Before entry, outlier removal, time axis alignment, and normalization are performed on the data. After review, the data is added to the online expanded training dataset. An incremental fine-tuning is triggered every 50 newly added batches to continuously improve the network's prediction accuracy for the optimal slicing window.

[0064] It should be noted that the key technologies of this invention include: First, closed-loop cooking control based on online near-infrared moisture gradient detection. By comparing the detected moisture penetration depth with the target value in real time, each batch of almonds entering the slicing stage possesses sufficient internal toughness, fundamentally avoiding slice breakage caused by uneven moisture content. Second, a multimodal temporal fusion attention network combined with a physical constraint layer dynamically fuses three heterogeneous sensor signals—temperature, moisture, and vibration—through a cross-modal cross-attention mechanism. The thermodynamic residual temperature decay law is embedded in the loss function as a penalty term, ensuring that the predicted optimal slicing window is both highly accurate and conforms to physical laws, avoiding the output of abnormal scheduling instructions in sparse sample regions by purely data-driven models. Third, an optimal batch slicing allocation algorithm based on thermodynamic residual temperature decay kinetics. The Monte Carlo sampling method expands the particle group temperature decay from a single deterministic curve to a probability distribution description. The Lagrange multiplier method is used to solve for the optimal feeding cycle under multiple constraints, ensuring that each batch of material is sliced ​​to the maximum extent within the effective slicing window. The mechanism of the synergistic effect of the three technologies is as follows: the water-based closed loop ensures the physical property basis, the attention network accurately locates the slice window, and the batch allocation algorithm maximizes the window utilization rate. The three form a complete closed loop from physical property guarantee to window prediction and then to scheduling execution, which systematically improves the overall slice yield.

[0065] It should be noted that this invention also solves the following technical problem: In multi-process nut slicing, due to the coupling of process parameters in steaming, baking, heat preservation, and slicing, traditional methods lack a cross-process data linkage mechanism, resulting in isolated parameter optimization for each process and an inability to dynamically adjust the execution strategy of subsequent processes based on the actual output of preceding processes. This invention, through step S06, synchronously inputs the full-process data of each batch into the training dataset of a multimodal temporal fusion attention network, forming a cross-process data linkage closed loop. After each incremental fine-tuning, the network can learn the cross-process coupling law between steaming temperature, moisture penetration depth, baking parameters, heat preservation chamber temperature, and slicing yield. This allows the prediction of the optimal slicing window to not only depend on the current sensor data but also reflect the cumulative impact of changes in preceding process parameters on the slicing result. This mechanism enables this invention to continuously output scheduling decisions adapted to the current production state under the combined disturbances of different batches of raw material characteristics, environmental temperature fluctuations, and equipment status changes, solving the technical problem of difficult cross-process collaborative optimization under multi-process parameter coupling conditions.

[0066] Specifically, the principle of this invention is as follows: The fundamental reason why this invention can solve the above-mentioned technical problems is that the closed-loop process of particle size classification cooking and online near-infrared moisture gradient detection ensures that each batch of almonds entering the slicing stage has sufficient moisture penetration depth, laying the physical property foundation for subsequent temperature control. The heat-preserving intermediate chamber establishes an independent temperature control buffer zone between baking and slicing, extending the duration for which the kernel surface temperature is higher than the critical slicing temperature and shortening the risk period of temperature window failure. The multimodal temporal fusion attention network dynamically fuses three types of heterogeneous sensor data streams through a cross-modal cross-attention mechanism and embeds a physical constraint layer to make the prediction results conform to thermodynamic laws, thereby outputting a physically consistent and highly accurate optimal slicing window. The optimal batch slicing algorithm based on thermodynamic residual temperature decay kinetics is based on Newton's law of cooling. It combines the Monte Carlo sampling method to characterize the probability distribution of temperature decay between batches, and then uses the Lagrange multiplier method to solve for the optimal single feeding amount and feeding interval under constraints. This ensures that each batch of material is sliced ​​to the maximum extent within the effective slicing window, thereby eliminating the cooling loss of nuts caused by scheduling lag.

[0067] 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.

[0068] The specific implementation method of step S01 is as follows: Particle size of the almond raw material is measured. Through at least three batches of particle size distribution statistical experiments, particle size grading boundaries are determined based on particle size nodes with higher distribution density. Typical particle size ranges are divided into three grades: 8–10 mm, 10–12 mm, and 12–14 mm. At least five gradient cooking experiments are conducted for each particle size range group. The moisture penetration depth measured by an online near-infrared moisture gradient detector is used as the evaluation index. The moisture penetration depth values ​​under each combination of cooking temperature and cooking time are recorded, and a moisture penetration depth prediction model is established. The formula is as follows:

[0069] ;

[0070] In the formula, This is the measured value of water penetration depth, in mm. The target value for water infiltration depth, in mm, has been verified to be 0.3 mm through at least five batches of experiments. This refers to the cooking temperature, expressed in °C, with a typical range of 95–100 °C. The reference cooking temperature is expressed in °C. The cooking time is in minutes, ranging from 30 to 60 minutes depending on the particle size range. The standard cooking time is expressed in minutes. Let be the dimensionless mapping function obtained by fitting experimental data, with With the goal of convergence to 1, the cooking temperature range and cooking time range corresponding to each particle size range are iteratively optimized to complete the grouped cooking and peeling.

[0071] The specific implementation of step S02 is as follows: Online near-infrared moisture gradient detection is performed on batches of peeled almonds. Utilizing the absorption differences of different wavelengths of near-infrared light at the surface and subsurface layers of the kernel, the moisture content gradient at different depths below the kernel surface is retrieved, and the moisture penetration depth detection value is output for each kernel. and the target value of water infiltration depth Comparison. Detection value reached Batches that meet the requirements are sent to S03, while batches that do not meet the requirements are returned to S01 for additional cooking.

[0072] The specific implementation of step S03 is as follows: Almonds meeting the target moisture penetration depth are placed in a tunnel oven and roasted at a roasting temperature (typically 70℃) for a typical roasting time of 25–35 minutes. At least five batches of slicing experiments are conducted under different combinations of roasting temperature and time. The slicing yield and kernel surface temperature time series data are used as evaluation indicators to determine whether the kernel surface temperature remains above the critical slicing temperature after roasting. The baking temperature and time combination is specified. After baking, the almonds are transferred to an insulated intermediate chamber, where the temperature is maintained. The typical temperature range is 55–65℃. The decay rate of the kernel surface temperature was measured under different insulated warehouse temperatures to ensure that the kernel surface temperature within the optimal slicing window was not lower than [a certain value]. As constraints, iterative experiments were conducted to determine the minimum temperature maintenance value of the insulation chamber. (Critical temperature for slicing) The breakage rate was determined by at least five batches of tests, with the kernel surface temperature at the point when the breakage rate first exceeded the upper limit of acceptable breakage rate being used as the criterion. .

[0073] The specific implementation of step S04 is as follows: A multimodal temporal fusion attention network collects three types of heterogeneous sensor data streams in real time, namely near-infrared moisture gradient temporal data. Time series data of kernel surface temperature Timing data of tool vibration acceleration ,in The time step length is defined. The three data streams are fed into three independent one-dimensional temporal convolutional branches for local feature encoding. Each branch consists of three stacked one-dimensional convolutional layers with kernel sizes of 3–7 and a stride of 1. Each convolutional layer is followed by a batch normalization layer and a modified linear unit activation function. Branch number The output feature sequence of the layer is denoted as ,in For the first Number of channels per layer. The outputs of the three branches are fed into the cross-modal attention module, where the feature sequences from any two branches serve as query vectors and key-value pairs. The cross-modal attention weight matrix... The calculation formula is:

[0074] ;

[0075] In the formula, For the first Query matrix of branches For the first Key matrix of branches , For learnable projection matrix The dimension of the attention key vector For row normalized exponential function matrix transpose symbol , This is a branch index, with values ​​ranging from 1, 2, and 3. Cross-modal attention fusion output The calculation formula is:

[0076] ;

[0077] In the formula, For the first Branches on the first Attention-weighted fusion feature sequences of branches For the first Value matrix of branches The learnable projection matrix. The fused feature sequence. Send to the loop jump unit, To fuse feature dimensions, the cyclic skip unit is a variant of the gated cyclic unit with residual gating. It introduces an adaptive time-step skip mechanism on top of the standard gated cyclic unit. Time step jump probability The calculation formula is:

[0078] ;

[0079] In the formula, The sigmoid activation function has an output range of... Learnable weight vector For the first Hidden state vector at time step For learnable bias scalars; when Above the jump threshold When the empirical value is 0.5, the network skips this time step and directly connects to the next critical change node, reducing the interference of redundant time steps on the hidden state. The output layer is a fully connected layer, outputting the optimal slice window start time. Optimal slice window termination time The network outputs three predicted values, along with the expected slice yield interval. A physical constraint layer is embedded at the network's end, discretizing the residual temperature decay differential equation and embedding it as a penalty term into the total loss function. The total loss function... The formula is expressed as follows:

[0080] ;

[0081] In the formula, Total number of training batches , The first Predicted values ​​for the start and end times of the optimal slice window in a batch, in minutes. , The corresponding label value is in min. This is the time-normalized baseline value, in minutes. For the first Batch expected slice yield prediction, dimensionless The corresponding label value is dimensionless. The normalized benchmark value for yield is dimensionless. This is the weighting coefficient for the physical constraint penalty term, dimensionless, with an empirical value of 0.1–0.5. The physical constraint penalty term is dimensionless and is obtained by discretizing the residual temperature decay differential equation. The formula is as follows:

[0082] ;

[0083] In the formula, For the first Batch No. Predicted kernel surface temperature at time step, in °C The current ambient temperature is expressed in °C. The interval between adjacent time steps, in minutes. The time constant is the thermal constant, expressed in minutes. This penalty term ensures that the predicted temperature sequence conforms to the discrete form of Newton's law of cooling, preventing the predicted values ​​from violating thermodynamic laws. Training employs an adaptive moment estimation optimization algorithm with an initial learning rate of... After every 50 training rounds, the learning rate decays to 0.5 times the previous round's rate, with a minimum of 200 training rounds, to verify that the prediction error of the set slice output rate does not exceed 5%. and The convergence criterion is that the prediction error does not exceed 30 seconds. After convergence, the model weights are saved. An incremental fine-tuning is triggered after every 50 batches of new production data are added. During fine-tuning, the parameters of the three one-dimensional temporal convolution branches are frozen, and only the parameters of the cross-modal attention module and the cyclic skip unit are updated.

[0084] The specific implementation of step S05 is as follows: Based on the optimal slice batch allocation algorithm of thermodynamic residual temperature decay kinetics, and based on Newton's law of cooling, a differential equation for the thermal decay of almond particle groups is established, and the formula is expressed as follows:

[0085] ;

[0086] In the formula, The average surface temperature of the kernels in the grain group is expressed in °C. The current ambient temperature is expressed in °C. Cooling process time, in minutes. The thermal time constant, expressed in minutes, was determined through cooling curve measurements of each particle size group, followed by exponential function fitting of the data. At least 1000 probability distribution sample curves were generated using Monte Carlo sampling to study the surface temperature decay process of the kernels within the particle group. The average surface temperature of the kernels within the particle group was statistically higher than... The duration of the slice is defined as the effective slice window duration. The typical time is 5-10 minutes. The objective function is to maximize the proportion of material sliced ​​by the slicer within the effective slicing window. The optimal single feed rate is solved using the Lagrange multiplier method under constraints. Optimal feeding interval augmented objective function The formula is expressed as follows:

[0087] ;

[0088] In the formula, The mass of material sliced ​​by the slicer within the effective slicing window, expressed in kg. This represents the total amount of material fed into the current batch, in kg. The proportion of material slices completed, dimensionless. , These are Lagrange multipliers, all of which are dimensionless. The rated processing speed of the slicer, in units of The optimal feeding interval time, in minutes. The minimum safe transfer time is given in minutes; the constraints are as follows: and ,in The optimal single feed rate is expressed in kg. The comprehensive hot scoring function is used to adjust the feed rate parameter, and the formula is as follows:

[0089] ;

[0090] In the formula, For comprehensive thermal state scoring, dimensionless This is the average surface temperature of the kernels in the current batch, in °C. The critical temperature for slicing, in °C. Temperature of the insulation chamber, in °C The reference temperature for the insulated warehouse is expressed in °C. The current processing speed of the slicer, in units of The rated processing speed of the slicer, in units of , , The weighting coefficients are and satisfy the following conditions: The values ​​were determined by no fewer than 10 batches of multiple regression experiments, with typical value ranges as follows: , , .when At that time, the feed rate parameter is taken as follows: ;when At that time, the feed rate parameter is taken as follows: ;when At that time, the feed rate parameter is taken as follows: And simultaneously trigger the heating command for the insulation intermediate chamber; when At that time, the feeding is paused and the multimodal temporal fusion attention network is triggered to re-predict. and The conveyor belt control system is based on , , and Adjust the feed rate to feed the almonds into the slicer in batches within the optimal slicing window to complete the slicing.

[0091] The specific implementation of step S06 is as follows: After slicing, the slice yield of the current batch is calculated. The slice yield, along with the time-series data of the current batch's cooking temperature, cooking time, baking temperature, baking time, heat preservation chamber temperature, kernel surface temperature, near-infrared moisture gradient, and blade vibration acceleration, are synchronously entered into the multimodal temporal fusion attention network training dataset. After outlier removal, time axis alignment, and normalization, the data is continuously added to the online expanded training dataset for iterative updates of the multimodal temporal fusion attention network.

[0092] To better understand and implement this invention, the following is a specific application scenario of the invention, Example 2: To illustrate the effect of the invention, technicians set up a test environment and conducted a full-process slicing test on a batch of almond raw materials according to the method described in this invention. The following is a complete description of the implementation process.

[0093] Technicians first conducted particle size testing on the incoming almond raw materials. After scanning by particle size sorting equipment, the particle size distribution of this batch of raw materials is shown in Table 1.

[0094] Table 1. Statistical table of particle size distribution of almond raw materials

[0095]

[0096] As shown in Table 1, all three particle size ranges account for a certain proportion, with 10–12 mm being the main distribution range. Based on the particle size grading boundaries, this batch of raw materials was divided into three groups, with cooking temperatures set at 96℃ and cooking times set at 35 min (8–10 mm group), 97℃ and 42 min (10–12 mm group), and 98℃ and 55 min (12–14 mm group), respectively. The materials were then fed into the cooking equipment to complete the grouped cooking and peeling process.

[0097] After steaming and peeling, the peeled almonds were sent batch by batch to an online near-infrared moisture gradient detector for individual scanning. The test results for this batch are shown in Table 2.

[0098] Table 2. Results of moisture penetration depth test of almonds after peeling

[0099]

[0100] As shown in Table 2, the moisture penetration depth of the 12–14 mm particle size group was 0.27 mm, which did not reach the target value of 0.3 mm. This group returned to step S01 for supplementary cooking. After 8 minutes of supplementary cooking, the moisture penetration depth was measured again, reaching 0.31 mm. This passed the closed-loop verification and proceeded to the next process. The 8–10 mm and 10–12 mm particle size groups directly proceeded to step S03.

[0101] Three batches of qualified almonds were sequentially fed into the tunnel oven and roasted at 70℃ for 30 minutes. After roasting, the almonds were transferred to an insulated intermediate chamber, with the temperature set at 60℃. This insulated intermediate chamber serves as an independent temperature-controlled buffer zone between the tunnel oven outlet and the slicer inlet, maintaining the surface temperature of the kernels above the critical slicing temperature and extending the effective slicing window time.

[0102] like Figure 3 As shown, during the heat preservation stage, the multimodal temporal fusion attention network collects real-time temporal data of kernel surface temperature, near-infrared moisture gradient, and tool vibration acceleration. Through the cross-modal cross-attention module, the three types of heterogeneous sensor data streams are dynamically fused. Combined with the residual temperature decay penalty term of the physical constraint layer, the optimal slicing window start time is 6 minutes after baking ends, the optimal slicing window end time is 14 minutes after baking ends, and the effective slicing window duration is 8 minutes. This prediction result is transmitted to the conveyor belt control system.

[0103] Based on the optimal batch allocation algorithm for thermodynamic residual temperature decay kinetics, with the current ambient temperature of 22℃, the slicer's rated processing rate of 12kg / min, and the critical slicing temperature of 52℃ as inputs, a differential equation for the collective thermal decay of the particle group is established using Newton's law of cooling. The thermal time constant is obtained by fitting the cooling curves of each particle size range through experiments: 18min for the 8–10mm group, 22min for the 10–12mm group, and 26min for the 12–14mm group. 1000 temperature decay probability distribution sample curves are generated using Monte Carlo sampling. The effective slicing window duration is statistically determined to be 8min. The Lagrange multiplier method is used to solve the problem under constraints, resulting in a single feed rate of 18kg, a feed interval of 1.5min, and a dynamic batch scheduling scheme that completes the slicing of all materials in this batch through 5 feeds.

[0104] During the slicing process, the comprehensive thermal scoring function continuously monitored the average surface temperature of the kernels, the temperature of the heat preservation chamber, and the current processing rate of the slicer. The comprehensive thermal score at each feeding time of this batch remained above 0.92, and the feeding rate was executed at 100% of the slicer's rated processing rate. No speed reduction or pause commands were triggered, and the slicing process proceeded smoothly.

[0105] After slicing is completed, the slicing yield of this batch is calculated. The slicing yield and the data of the whole process are synchronously entered into the multimodal temporal fusion attention network training dataset and added to the online expanded training dataset to provide sample support for subsequent incremental fine-tuning.

[0106] Compared to traditional slicing methods, the advancements of this invention are reflected in the following aspects: Traditional methods rely on fixed cooking parameters, lacking the ability to adapt to raw materials of different particle sizes, and making it difficult to guarantee the consistency of moisture penetration depth; this invention, through particle size classification and online near-infrared moisture gradient detection in a closed loop, ensures that each batch of almonds entering the slicing stage possesses sufficient internal toughness, eliminating the root cause of slice breakage from a physical property perspective. Traditional methods lack quantitative description of the surface temperature decay of kernels, and slicing scheduling relies on empirical judgment, resulting in a high probability of temperature window mismatch; this invention, through the combination of a multimodal temporal fusion attention network and a physical constraint layer, embeds thermodynamic laws into the network prediction mechanism, ensuring that the output of the optimal slicing window has physical consistency, avoiding the blindness of purely empirical scheduling. Traditional methods have a fixed feeding cycle, which cannot adapt to the heterogeneity of temperature decay caused by differences in particle size distribution between batches. This invention describes temperature decay as a probability distribution using Monte Carlo sampling, and then uses the Lagrange multiplier method to solve for the optimal feeding cycle, so that each batch of material is sliced ​​to the maximum extent within the effective slicing window time, thereby reducing material cooling loss caused by temperature window mismatch from the scheduling mechanism.

[0107] It should be noted that the variables involved in this invention are explained in detail in Tables 3 and 4.

[0108] Table 3. Variable Explanation Table (Part 1)

[0109]

[0110] Table 4. Variable Explanation Table (Part Two)

[0111]

[0112] 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. A method for slicing nuts after steaming and boiling pretreatment, characterized in that, Includes the following steps: The raw material of almonds was subjected to particle size detection. The raw material was divided into several particle size interval groups according to the particle size grading boundary. The corresponding cooking temperature and cooking time were set for each particle size interval group to complete the group cooking and peeling. After peeling, the almonds are subjected to online near-infrared moisture gradient detection in batches to obtain the moisture penetration depth detection value. The moisture penetration depth detection value is compared with the moisture penetration depth target value. Batches with the moisture penetration depth detection value reaching the moisture penetration depth target value proceed to the next step, while batches that do not reach the moisture penetration depth target value are returned for supplementary steaming. Almonds whose moisture penetration depth detection value reaches the target value are sent into a tunnel oven for baking at the baking temperature. After the baking time is over, the almonds are transferred to an insulated intermediate chamber, where they are kept warm at the temperature of the insulated chamber. A multimodal temporal fusion attention network collects real-time temporal data of kernel surface temperature, near-infrared moisture gradient, and tool vibration acceleration, outputs the optimal slicing window start time and optimal slicing window end time, and transmits them to the conveyor belt control system. Based on the optimal batch allocation algorithm of thermodynamic residual temperature decay kinetics, the current ambient temperature, the rated processing rate of the slicer and the critical temperature of slicing are used as inputs to calculate the amount of material fed at one time and the feeding interval time. The conveyor belt control system adjusts the feeding rate according to the start time of the optimal slicing window, the end time of the optimal slicing window, the amount of material fed at one time and the feeding interval time, so that the almonds are fed into the slicer in batches from the start time of the optimal slicing window to the end time of the optimal slicing window to complete the slicing. After slicing, the slice yield of the current batch is calculated. The slice yield, along with the time-series data of the current batch's cooking temperature, cooking time, baking temperature, baking time, heat preservation chamber temperature, kernel surface temperature, near-infrared moisture gradient, and blade vibration acceleration, are synchronously entered into the multimodal temporal fusion attention network training dataset for iterative updates of the multimodal temporal fusion attention network.

2. The nut pretreatment and slicing method according to claim 1, characterized in that, The determination of the particle size classification boundary involves conducting a particle size distribution statistical experiment on almond raw materials with a particle size distribution density of no less than the batch threshold of the particle size classification experiment, and selecting the particle size nodes with the larger distribution density as the particle size classification boundary according to the particle size distribution density.

3. The nut pretreatment and slicing method according to claim 2, characterized in that, The determination of cooking temperature and cooking time involves conducting gradient cooking experiments for each particle size range group for no less than the batch threshold of the gradient cooking experiment, establishing a moisture penetration depth prediction model, and iteratively optimizing the cooking temperature range and cooking time range corresponding to each particle size range group with the goal of converging the moisture penetration depth detection value to the target value of moisture penetration depth.

4. The nut pretreatment and slicing method according to claim 3, characterized in that, The target value for water penetration depth is determined by slicing almonds with different water penetration depth detection values ​​in a slicing experiment with a batch number of slices not less than the threshold of the water penetration test batch. The minimum water penetration depth detection value corresponding to the first time the slice yield reaches the qualified yield threshold is taken as the target value for water penetration depth.

5. The nut pretreatment and slicing method according to claim 4, characterized in that, Online near-infrared moisture gradient detection specifically utilizes the absorption differences of near-infrared light of different wavelengths in the surface and subsurface layers of the kernel to invert the moisture content gradient at different depths below the kernel surface, thereby indirectly characterizing the moisture penetration depth detection value. After scanning each almond, the moisture penetration depth detection value is output.

6. The nut pretreatment and slicing method according to claim 5, characterized in that, The determination of baking temperature and baking time involves conducting slicing experiments with no fewer than the batch threshold under different combinations of baking temperature and baking time. The slicing yield and kernel surface temperature time series data are used as evaluation indicators to determine the combination of baking temperature and baking time that ensures the kernel surface temperature remains above the critical slicing temperature after baking.

7. The nut pretreatment and slicing method according to claim 6, characterized in that, The critical temperature for slicing is determined by conducting slicing experiments at different kernel surface temperatures, recording the breakage rate corresponding to each kernel surface temperature, and taking the kernel surface temperature at which the breakage rate first exceeds the upper limit of the acceptable breakage rate as the critical temperature for slicing.

8. The nut pretreatment and slicing method according to claim 7, characterized in that, The determination of the temperature of the heat preservation chamber is specifically achieved by measuring the decay rate of the surface temperature of almond kernels under different heat preservation chamber temperatures. The minimum maintenance value of the temperature of the heat preservation chamber is determined through iterative experiments, with the constraint that the surface temperature of the kernels should not be lower than the critical temperature of slicing from the start time to the end time of the optimal slicing window.

9. The nut pretreatment and slicing method according to claim 8, characterized in that, The input layer of the multimodal temporal fusion attention network receives three heterogeneous sensor data streams: kernel surface temperature time-series data, near-infrared moisture gradient time-series data, and tool vibration acceleration time-series data. The three heterogeneous sensor data streams are respectively fed into three independent one-dimensional temporal convolution branches for local feature encoding. The encoding results are then fed into the cross-modal cross-attention module for dynamic fusion, and then into the cyclic skip unit. The output layer outputs the optimal slice window start time, the optimal slice window end time, and the expected slice yield range.

10. The nut pretreatment and slicing method according to claim 9, characterized in that, The multimodal temporal fusion attention network embeds a physical constraint layer at its end. The residual temperature decay differential equation is discretized and embedded into the loss function in the form of a penalty term, so that the predicted values ​​at the start and end times of the optimal slice window do not violate the laws of thermodynamics and physics.

Images