New fruit juice processing technology intelligent migration method based on migration learning

Abstract

The invention provides an intelligent migration method for a new fruit juice processing technology based on migration learning, and the method comprises the steps: carrying out the structural library building of a technology log, a technology knowledge encoder based on a Transform architecture, a cross-domain feature space mapping network, multi-task loss joint optimization, and dynamic constraint control of a technology window, common operation experience extraction, process parameter intelligent migration and adaptability automatic adjustment and optimization of different fruit categories are realized, the applicability of a new variety and the compliance of a processing process are improved, and the intelligence and the high efficiency of normal juice production are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent food processing technology, and in particular to a novel intelligent transfer method for processing fruit juice based on transfer learning. Background Technology

[0002] Currently, intelligent transfer modeling of fruit juice processing technology is gradually becoming a core technology direction for intelligent food processing and rapid cross-category technology introduction. In this field, transfer learning is widely used to realize the transfer of process knowledge between different fruit categories, aiming to reduce the trial and error costs of process when launching new varieties and accelerate the pre-tuning process of process parameters. Mainstream technical solutions mostly focus on unsupervised or weakly supervised feature space alignment, such as maximum mean difference (MMD) and adversarial training, which improve the generalization ability of feature transfer by minimizing the statistical distance between the feature distributions of the source and target domains. Existing research generally characterizes the processing features of fruits in the source and target domains as multidimensional statistical vectors and uses deep neural networks for automatic alignment. However, these typical solutions are mostly limited to distribution similarity constraints and neglect the impact of process operation boundaries, physical constraints, and equipment characteristics on parameter feasibility in actual juice processing. For example, existing methods rely solely on the mathematical alignment of features, ignoring the rationality verification of key process parameters such as processing temperature, pressing pressure, and pH adjustment. This leads to the model achieving high distribution consistency in the transfer task, but potentially generating risky parameter combinations in the actual operation, resulting in engineering risks such as enzyme inactivation, equipment overload, and quality degradation. In the field of intelligent food processing, although some research involves the embedding of experiential knowledge and prior constraint modeling, customized process constraint integration and data-knowledge jointly driven feature space alignment systems for cross-category migration of fruit juice are still immature. Representative technologies, such as some frameworks based on general domain adaptive alignment, are only applicable to categories with small process differences and highly overlapping boundary conditions. When dealing with fruits with high acidity, viscosity, or special physicochemical properties, it is difficult to guarantee the process feasibility and product quality consistency of migration prediction results. Therefore, the current technology is mainly applicable to categories with relatively similar process conditions, lacking effective adaptability to abrupt changes in process boundaries and extreme parameter requirements. Summary of the Invention

[0003] In order to solve the above-mentioned technical problems, this invention provides a novel intelligent transfer method for processing fruit juice based on transfer learning.

[0004] The technical solution of this invention is implemented as follows: a new fruit juice processing technology intelligent transfer method based on transfer learning, comprising: S1: Collect key process parameter trajectories and quality output indicators of multiple types of fruits during the juice processing. The process parameter trajectories include temperature-time curves, pressure-speed sequences, and pH change trends. The quality output indicators include clarity, vitamin retention rate, and turbidity stability, forming a structured process log library. S2: Perform standardized preprocessing on the data in the structured process log library, including missing value imputation, outlier removal and time series alignment, to improve the temporal consistency and comparability of the process logs; S3: Based on the preprocessed process log data, a lightweight process knowledge encoder is constructed. The Transformer architecture is used to encode the process log into a low-dimensional process semantic vector to capture the common operation patterns and differential boundary conditions among different types of fruits. S4: Construct a feature space mapping network, input the original features of the source domain and the target domain fruit into the mapping network, generate the mapped feature space representation, and introduce a differentiable process consistency loss term to quantify the deviation between the virtual process response triggered by the mapped features and the target domain process window. S5: Based on the process consistency loss term and feature alignment loss term, a multi-task learning framework is used to jointly optimize the mapping function. The feature alignment loss term includes a domain classification accuracy minimization term, and the process consistency loss term adjusts the mapping function parameters through backpropagation to ensure that the mapped features satisfy both cross-domain distribution consistency and process operation specifications. S6: During the transfer modeling process, dynamically determine whether the current mapped feature falls within the process window allowed by the target domain. If it is determined to fall within the window, execute the transfer modeling result output; if it is determined not to fall within the window, trigger the adaptive adjustment mechanism of the mapping function parameters to improve the process compliance of the feature space mapping. S7: Generate a set of pre-tuned parameters for the processing technology of new fruit juice based on the migration modeling results, including temperature setpoint, pressing pressure threshold, and pH control range, so as to quickly import the process parameter initialization in the scenario. S8: Collect process response data of target domain fruits during actual processing, compare and analyze the data with the prediction results of transfer modeling, generate process consistency scores, and perform incremental training and model updates on the process knowledge encoder and mapping network based on the score results.

[0005] The intelligent transfer learning-based method for processing new fruit juices provided by this invention has the following beneficial effects: (1) This invention starts with the acquisition of multi-source real-time process parameters, the establishment of a structured process log library, and standardized cleaning. It associates key processes in fruit juice processing (such as temperature-time, pressure-speed, pH changes, etc.) with a quality indicator system, which not only ensures the temporal consistency and traceability of the data, but also lays the foundation for high-quality input of subsequent models. Through unified normalization, anomaly removal, and dynamic time regularization, the comparability of multi-category process data is significantly improved, and the problems of data heterogeneity and difficulty in summarization within the industry are solved. (2) The lightweight process knowledge encoder constructed in this invention fully utilizes multi-head attention and hierarchical abstraction mechanism to automatically translate complex process states and their category boundary conditions into low-dimensional, learnable semantic embeddings. This achieves effective extraction of commonalities in multi-category operation modes and high-precision expression of isolated differences. In turn, it provides differentiable process constraint priors for the model in the feature mapping stage, effectively preventing the feature space from "drifting" to non-engineering feasible intervals. (3) This invention achieves dynamic avoidance of potential infeasible process combinations by jointly optimizing the multi-task loss through a dual objective of "feature alignment + process consistency". This not only requires the feature distribution to be domain-indiscriminate, but also forces the consideration that the "virtual process response" excited by the mapped features must fall within the allowed window of the target domain. This new path makes it possible to directly implement the combination of migration parameters. (4) This invention can automatically identify and correct the process window out-of-bounds problem in potential migration output through the dynamic discrimination mechanism of process compliance and parameter adaptive optimization, realize intelligent projection and constraint closure of parameter space, ensure that each round of migration closely follows the real physical constraints and quality and safety boundaries, effectively avoid the embarrassment of traditional migration models being "effective but unusable", and improve the model's adaptability and first-round debugging success rate in the scenario of rapid new product introduction. Attached Figure Description

[0006] Figure 1 This is a flowchart of the intelligent transfer learning method for processing new fruit juice based on transfer learning, as described in this invention. Figure 2 This is a sub-flowchart of the intelligent transfer learning method for processing new fruit juice based on transfer learning, as presented in this invention. Figure 3 This is another sub-flowchart of the intelligent transfer method for processing new fruit juice based on transfer learning according to the present invention. Detailed Implementation

[0007] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0008] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0009] like Figure 1 As shown, this invention provides a novel intelligent transfer method for processing fruit juice based on transfer learning, specifically including: S1: Collect key process parameter trajectories and quality output indicators of multiple types of fruits during the juice processing. The process parameter trajectories include temperature-time curves, pressure-speed sequences, and pH change trends. The quality output indicators include clarity, vitamin retention rate, and turbidity stability, forming a structured process log library. S2: Perform standardized preprocessing on the data in the structured process log library, including missing value imputation, outlier removal and time series alignment, to improve the temporal consistency and comparability of the process logs; S3: Based on the preprocessed process log data, a lightweight process knowledge encoder is constructed. The Transformer architecture is used to encode the process log into a low-dimensional process semantic vector to capture the common operation patterns and differential boundary conditions among different types of fruits. S4: Construct a feature space mapping network, input the original features of the source domain and the target domain fruit into the mapping network, generate the mapped feature space representation, and introduce a differentiable process consistency loss term to quantify the deviation between the virtual process response triggered by the mapped features and the target domain process window. S5: Based on the process consistency loss term and feature alignment loss term, a multi-task learning framework is used to jointly optimize the mapping function. The feature alignment loss term includes a domain classification accuracy minimization term, and the process consistency loss term adjusts the mapping function parameters through backpropagation to ensure that the mapped features satisfy both cross-domain distribution consistency and process operation specifications. S6: During the transfer modeling process, dynamically determine whether the current mapped feature falls within the process window allowed by the target domain. If it is determined to fall within the window, execute the transfer modeling result output; if it is determined not to fall within the window, trigger the adaptive adjustment mechanism of the mapping function parameters to improve the process compliance of the feature space mapping. S7: Generate a set of pre-tuned parameters for the processing technology of new fruit juice based on the migration modeling results, including temperature setpoint, pressing pressure threshold, and pH control range, so as to quickly import the process parameter initialization in the scenario. S8: Collect process response data of target domain fruits during actual processing, compare and analyze the data with the prediction results of transfer modeling, generate process consistency scores, and perform incremental training and model updates on the process knowledge encoder and mapping network based on the score results.

[0010] Step S1: Collect key process parameter trajectories and quality output indicators for multiple fruit varieties during the juice processing. The process parameter trajectories include temperature-time curves, pressure-speed sequences, and pH trends. The quality output indicators include clarity, vitamin retention rate, and turbidity stability, forming a structured process log library. Specifically, this includes: S1.1: Multi-source sensing is used to collect data on the juice processing of multiple types of fruits to obtain process parameter trajectory data such as temperature-time curves, pressure-speed sequences, and pH change trends, in order to construct an original process parameter database; For the juice processing of multiple types of fruits, a multi-source sensor array consisting of temperature, pressure, and pH sensors is used (parameters: temperature measurement range 0-150℃, accuracy ±0.1℃; pressure measurement range 0-2MPa, accuracy ±0.01MPa; pH measurement range 0-14, accuracy ±0.01), to achieve continuous acquisition of temperature-time curves, pressure-speed sequences, and pH change trends. A synchronous acquisition controller (parameters: sampling frequency 100Hz, time synchronization accuracy 1ms) is used to synchronously trigger sampling of the output signal of the multi-source sensor array and generate a raw sensor data stream with precise timestamps to ensure strict alignment of the trajectory of each process parameter in the time dimension. Furthermore, the analog signal from the sensor is converted into a digital signal format by an analog-to-digital conversion module (parameters: 16-bit resolution, sampling accuracy ±0.05%FS), and then quantized and encoded to obtain a digital sequence of process parameters suitable for storage and subsequent processing. Furthermore, through the sensor data buffer and batch transmission protocol (parameters: buffer capacity 1GB, maximum transmission rate 200Mbps), the real-time collected digital sequence is transmitted to the central data receiving server, and is initially classified and stored according to fruit type and processing batch to form a multi-category original process parameter data pool. Furthermore, through a data verification algorithm (parameter: verification threshold set to ±5% of sensor range), integrity verification and range verification are performed on the received digital sequence, invalid records caused by acquisition failure are eliminated, and a database of original process parameters that have passed verification is generated. By using multi-source sensor collaborative acquisition and digital encoding processing, the signals of the multi-category fruit processing process in the previous step are transformed into a database of raw process parameters with time synchronization and high precision, thus achieving the basic data acquisition technology effect for the construction of the process log library. For example, temperature sensors (PT100 platinum resistance thermometers, measuring range 0-150℃, accuracy ±0.1℃), pressure sensors (strain gauges, measuring range 0-2MPa, accuracy ±0.01MPa), and pH sensors (glass electrodes, measuring range 0-14, accuracy ±0.01) are respectively deployed in the juice processing of oranges, pineapples, and kiwis. The signals from each sensor are connected to a synchronous acquisition controller (sampling frequency 100Hz, time synchronization accuracy 1ms). Throughout the juicer's operation, the controller triggers batch-by-batch acquisition, obtaining analog signals of temperature-time curves, pressure-speed sequences, and pH change trends for the three types of fruits. These signals are then encoded into digital sequences by a 16-bit analog-to-digital converter. The acquired digital sequences are then transmitted in batches via a buffer module (512MB per batch) using the UDP protocol to a data receiving server and stored in the original process parameter data pool. The data was validated using an algorithm to remove outliers such as temperature signals exceeding 150℃, pressure signals exceeding 2MPa, or pH signals exceeding 14. The validated data retained consisted of 1.5 million records of oranges, 1.3 million records of pineapples, and 1.1 million records of kiwis. This resulted in a database of original process parameters containing precise timestamps, providing high-quality input data for subsequent timestamp alignment and standardization. S1.2: The collected process parameter trajectory data is timestamped and the sampling frequency is standardized. A linear interpolation algorithm is used to fill in the missing data points to obtain a standardized process parameter sequence with enhanced temporal consistency. S1.3: Based on the quality evaluation standard for fruit juice processing, collect quality output index data for each type of fruit after processing, including clarity, vitamin retention rate, and turbidity stability, to form a quality output index dataset; Based on the quality evaluation standard for fruit juice processing, the standardized process parameter sequence obtained in step S1.2 is used as a time anchor point to structure and record the quality data collected by the quality detection sensors and chemical analysis instruments after processing each type of fruit, so as to ensure the consistency between quality evaluation and process. An optical turbidity measurement method (parameters: wavelength 860nm, bandwidth 10nm, integration time 5s) was used to achieve a quantitative evaluation of the clarity index and obtain raw clarity data expressed as a percentage of transmittance. Furthermore, the vitamin retention rate was quantitatively determined by high performance liquid chromatography (HPLC) (parameters: C18 column, mobile phase methanol:water = 3:7, flow rate 1 mL / min), and the vitamin mass ratio of the processed sample to the raw material sample was calculated to obtain vitamin retention rate data. Furthermore, a turbidity stability test method (parameters: centrifugation speed 3000 rpm, centrifugation time 30 min) was used to determine the rate of turbidity change of the sample under mechanical disturbance conditions, and turbidity stability data expressed as the amount of turbidity change per unit time was obtained. Vitamin retention rate is calculated using the following formula:

[0011] in, This represents the vitamin concentration in the processed fruit juice sample. The vitamin concentration of the raw material before processing; The turbidity stability index is calculated using the following formula:

[0012] in, This is the turbidity value. Time, measured in NTU / min, is used to characterize the change in turbidity per unit time; Through the above-mentioned sensor data collection and analysis, the three types of quality data—clarity, vitamin retention rate, and turbidity stability—are used to generate a quality output index dataset that is bound to timestamps and category tags, thereby achieving a consistent match between the processing process and quality evaluation. For example, in the processing of "red-fleshed kiwifruit" batch number 202405, optical clarity was tested immediately after juicing, and the transmittance was measured to be 92.5%, which was judged to be of high clarity grade according to processing standards; HPLC analysis showed that the vitamin C content decreased from 520 mg / L before processing to 498 mg / L after processing, using the formula... The vitamin retention rate was calculated to be 0.958, or 95.8%. In the centrifugation test, the turbidity increased from an initial 3.0 NTU to 3.5 NTU within 30 minutes, and the rate of turbidity change was calculated according to the formula... The result was 0.0167 NTU / min. The quality output data of this batch was timestamped with the temperature-time curve, pressure-speed sequence and pH trend data generated in S1.2 to form a structured kiwifruit processing quality data record, which supports the training input of the subsequent process knowledge encoder. S1.4: The standardized process parameter sequence is structurally associated with the corresponding quality output index data, and a multi-category fruit process log library is established using a relational database format to support the training input of the subsequent process knowledge encoder. S1.5: Label each type of fruit with its category identifier, processing batch number, and equipment model information in the structured process log library to form a traceable process log metadata system, providing a data foundation for the source and target domain division in migration modeling.

[0013] Step S2: Standardize and preprocess the data in the structured process log library, including missing value imputation, outlier removal, and time series alignment, to improve the temporal consistency and comparability of the process logs. Specifically, this includes: S2.1: Perform missing value detection on the temperature-time curves, pressure-speed sequences, and pH trend data collected from the structured process log library, and use linear interpolation to estimate the missing time points to obtain a complete and continuous process parameter sequence. S2.2: Based on the complete process parameter sequence output, outliers are identified and removed using a statistical method based on Z-score. ±3σ is set as the threshold, and data points outside this range are replaced with the moving average of nearby time points to improve data quality and stability. S2.3: Perform time axis standardization and alignment on the processed process parameter sequence, and use the Dynamic Time Warping (DTW) algorithm to map the processing of different types of fruits to a unified time scale to eliminate the time sequence offset caused by the difference in processing time. S2.4: Based on the output time-aligned data, construct a unified multidimensional feature matrix, and organize parameters such as temperature, pressure, and pH into tensor structures according to time steps to support the input requirements of subsequent process knowledge encoders; Based on the time-aligned process parameter data at a unified time scale output by S2.3, a multi-dimensional feature construction algorithm (parameters: time step Δt=1s, parameter dimension=3) is used to realize the function of slicing and recombining core process parameters such as temperature, pressure, and pH by time step. Furthermore, by using the tensor mapping method (parameters: tensor order = 3, mode dimension = [time steps, number of parameter types, number of sample batches]), the sliced ​​multi-channel time series data is mapped into a third-order tensor structure, and a feature tensor dataset with consistent time index and parameter channel index is obtained. Furthermore, by using the matrix index encoding method (parameter: parameter type index = {temperature: 0, pressure: 1, pH: 2}), fixed-position encoding of different parameter channels within the feature tensor is achieved, and a mapping matrix between parameter type and tensor index position is generated for subsequent input data structure positioning of the process knowledge encoder; Furthermore, by using a sparse padding and mask matrix generation algorithm (parameters: mask value = 0, sparsity threshold = 5%), zero-value padding is performed on the low-density sampling positions that still exist in the feature tensor, and a mask matrix is ​​formed to indicate the valid and invalid data states at each time position, thereby supporting the encoder's attention weight calculation. Through the tensor construction process described above, the time-aligned data from the previous step is transformed into structured multidimensional feature matrix data, enabling a unified tensor representation of key process parameters such as temperature, pressure, and pH. This provides a highly consistent input data format for subsequent process semantic extraction and mapping network training. For example, in a scenario involving the processing of multi-category fruit processing data with a batch size of 50 samples, the temperature-time curves, pressure-speed sequences, and pH trend data in S2.3 have been aligned to a uniform time scale of 120 seconds using a dynamic time warping algorithm, with a time step Δt set to 1 second, resulting in 120 time points. A multi-dimensional feature construction algorithm is used to arrange the three parameters sequentially according to the time step, forming a third-order tensor structure of size [120, 3, 50]. The first dimension is the time index from 0 to 119, the second dimension is the parameter type index encoded in the order of temperature: 0, pressure: 1, pH: 2, and the third dimension is the sample batch index. For approximately 4% of the time points in the original sampling that lack valid data, a sparse padding algorithm is used to fill in the missing data with zero values, and a mask matrix of the same size is generated to mark invalid positions as 0 and valid positions as 1. This mask matrix is ​​used to mask invalid temporal inputs during the encoder's self-attention mechanism calculation. The tensor data processed in this step is input into the subsequent process knowledge encoder, which can directly use the structure of the third-order tensor to perform multi-head self-attention modeling at the input layer. S2.5: Perform standardization and normalization on the generated multidimensional feature matrix, and use the Min-Max scaling method to map the parameters of each dimension to the [0,1] interval to eliminate the influence of dimensional differences on the training process of the process knowledge encoder and improve the consistency of the expression of process semantic vectors.

[0014] Step S3: Based on the preprocessed process log data, a lightweight process knowledge encoder is constructed. The Transformer architecture is used to encode the process logs into low-dimensional process semantic vectors to capture common operational patterns and differential boundary conditions among different fruit varieties. For example... Figure 2 As shown, it specifically includes: S3.1: Perform windowed slicing on the time series parameter trajectories in the preprocessed process log data to construct input sample fragments with temporal local correlation, thereby providing a structured input representation for the Transformer architecture; S3.2: Based on the windowed time series input sample segment, a multi-head self-attention mechanism is used to perform feature interaction modeling on the temperature-time curve, pressure-speed sequence, and pH change trend in the process parameter trajectory to obtain a multi-scale process state characterization; S3.3: Perform position encoding and category embedding fusion operations on the multi-scale process state representation to introduce time sequence information and fruit category identification, thereby enhancing the ability to distinguish the operation modes between different fruit categories; S3.4: Based on the fused process state representation, a hierarchical Transformer encoding layer is used for deep semantic abstraction. Low-dimensional process semantic vectors are extracted through multi-layer nonlinear transformation to capture the common operational logic and boundary constraints in the cross-category fruit processing process. Based on the fused process state representation, a hierarchical Transformer encoding layer (parameters: 4 encoding layers, 128 hidden dimensions, and 8 attention heads) is adopted to achieve hierarchical abstract representation of multi-scale process state features. Furthermore, the correlation enhancement model of the joint pattern of temperature-time, pressure-rotation speed and pH change trends across time steps is performed by the multi-head self-attention mechanism of the first two coding blocks (parameter: scaled dot product attention, key / value dimension is 64), and preliminary semantic association results across parameter domains are obtained. Furthermore, a feedforward fully connected sublayer (parameters: GELU activation function, 256-dimensional hidden layer expansion) is introduced through the intermediate layer coding block to achieve nonlinear feature transformation of the preliminary semantic association results and generate process constraint feature vectors in the intermediate abstract space; Furthermore, by using multi-layer residual connections and layer normalization processing of the end coding block (parameter: ε is set to 1e-5), the semantic features of the upper and middle layers are fused while maintaining gradient stability, generating a low-dimensional process semantic vector after dimensional compression. The low-dimensional embedding generation module performs mean pooling on the final encoded output (the window size is equal to the time series length), transforming the result of the previous step into a structurally balanced semantic representation, thereby achieving feature compression that preserves the common operational logic and boundary constraints across product categories. For example, in a multi-category fruit juice processing scenario, the input is a process state representation tensor that integrates positional encoding and category embedding, with a time step of 50 and a feature dimension of 64. A four-layer Transformer encoding layer is used, with eight attention heads, each with a key / value dimension of 64. The first two encoding blocks calculate the feature associations across time steps using a scaled dot product attention mechanism. The attention weight formula is as follows:

[0015] in, For self-attention output, For querying the matrix, The key matrix, For value matrices, The scaling factor is set to a key dimension of 64. The feedforward network of the middle two coding layers expands the feature dimension to 256, applies the GELU activation function for nonlinear transformation, and obtains a feature vector containing process physical constraints. The final coding block uses residual connections and LayerNorm (ε=1e-5) to stabilize the training process, and compresses the time series features into a low-dimensional process semantic vector (dimension 128) through mean pooling. S3.5: Perform normalization and learnable projection operations on the low-dimensional process semantic vector to generate a standardized process knowledge embedding representation, and output it to the feature space mapping network for the construction and joint optimization of the subsequent process consistency loss term.

[0016] Step S4: Construct a feature space mapping network, inputting the original features of the source and target domain fruits into the mapping network to generate a mapped feature space representation, and introducing a differentiable process consistency loss term to quantify the deviation between the virtual process response elicited by the mapped features and the target domain process window. Figure 3 As shown, it specifically includes: S4.1: Normalize the original feature data of fruits in the source and target domains to eliminate the dimensional differences in physical properties, chemical composition and processing parameters of different types of fruits, and obtain the input feature tensor under a uniform scale as the input of the feature space mapping network. S4.2: Based on the input feature tensor, a feature space mapping network with a multilayer perceptron structure is constructed. The input features are embedded in a high dimension through a learnable nonlinear transformation function to generate a latent feature space representation shared by the source and target domains, thereby enhancing the feature distribution alignment capability. For the normalized source and target domain input feature tensors, a feature space mapping network with a multilayer perceptron (MLP) structure is used (parameter: number of input layer nodes). Number of hidden layer nodes h, number of output layer nodes The ReLU activation function is used to perform a non-linear embedding transformation of the original features, generating a high-dimensional latent feature representation. The first hidden layer is constructed based on the weighted summation operation of the input layer, using a weight matrix. With bias vector Perform a linear transformation, and through The activation function achieves non-linear feature enhancement, thereby obtaining the first layer of embedding vectors; Furthermore, by introducing a batch normalization operation (parameters: a sliding estimation window for the mean μ and variance σ²) in each hidden layer, the data distribution across batches is stabilized, and the convergence speed of the gradient backpropagation stage is improved. Furthermore, multiple nonlinear mappings are constructed for the second and subsequent hidden layers, employing... The random deactivation mechanism (parameter: deactivation probability p=0.3) reduces the risk of overfitting the model to specific feature combinations; Furthermore, through the output layer weight matrix The final linear transformation is performed with the corresponding bias to generate latent feature vectors of the source and target domains in a unified high-dimensional space, thereby enhancing the alignment capability of feature distributions. Through the layer-by-layer nonlinear mapping of the above multilayer perceptron network, the normalized input feature tensor is transformed into a potential feature space representation with process distribution adaptability and cross-domain generalization ability, so as to achieve the expected technical effect of preliminary alignment of cross-category fruits in the feature space. For example, the tensors (dimensions) of the physicochemical properties of the source domain mango and the target domain kiwi fruit after normalization are... During feature space mapping (h1=128), the multilayer perceptron is configured with a two-layer hidden layer structure. The first hidden layer has 256 nodes (h1=256) and the second hidden layer has 128 nodes (h2=128), both using the ReLU activation function. The first hidden layer weight matrix... With bias vector The first layer embedding vector is obtained by calculating a linear combination of matrix multiplication and vector addition, followed by ReLU nonlinear activation. The batch normalization module sets the moving average window size to 32 samples to ensure a stable distribution of activation values ​​across different batches during training. The second hidden layer introduces a Dropout mechanism with a deactivation probability of 0.3, improving the model's generalization ability by randomly deactivating some neurons during training. (The number of output layer nodes is not specified in the original text.) =64, with weight matrix The final linear mapping is completed, generating a 64-dimensional latent feature space vector. This vector is then fused with the process semantic vector in the subsequent S4.3 sub-step, ultimately ensuring that mangoes and kiwis exhibit consistent distribution in the high-dimensional space of the transfer model. S4.3: The low-dimensional process semantic vector output by the process knowledge encoder in S3 is used as prior constraint information. Through a differentiable fusion mechanism, it is cross-modeled with the mapped features to generate process-aware feature representations, so as to guide the feature space mapping to conform to the actual processing operation boundary. S4.4: Based on the fusion representation of the mapped features and the process semantic vector, a virtual process response simulator is constructed. The regression model is used to predict the output of the corresponding virtual process parameters under the feature combination, such as the temperature setpoint, the pressing pressure threshold and the pH control range, to form a virtual process response vector. The input conditions include the process-aware feature representation obtained after processing by S4.3, and the standardized process knowledge embedding vector generated by S3.5. Both have scale consistency in a unified feature space. A multimodal feature fusion method (parameter: the fusion weight matrix is ​​initialized to a uniform distribution) is adopted to realize the numerical combination of process perception features and process semantic vectors to form a high-dimensional fusion representation vector; Furthermore, by using principal component regression (parameter: the number of principal components is set to the number of components that account for 95% of the cumulative variance contribution rate), the dimensionality reduction of the high-dimensional fusion representation is achieved, and a feature subspace with reduced redundant dimensions is obtained. Furthermore, a nonlinear regression model based on gradient boosting trees (parameters: tree depth of 6, learning rate of 0.05) is adopted to train the mapping function from the fusion feature subspace to the virtual process parameter output and generate a set of initial parameter prediction values. Furthermore, by using a constrained regression model (parameters: constraints are derived from the physical boundaries of the target domain process window), hard boundary restrictions are introduced during the prediction process to perform feasibility correction on the regression results of parameters such as temperature, pressure, and pH, and to generate virtual process response vectors that satisfy the constraints. Through the above model structure, the fusion features are transformed into process parameters such as virtual temperature setpoint, pressing pressure threshold, and pH control range, so as to realize the mapping effect of the prediction results in the physical feasible domain. For example, in the process transfer scenario from apples to newly introduced blueberries, the input process-aware feature representation is 512-dimensional, the process knowledge embedding vector is 64-dimensional, and the fusion weight matrix is ​​initialized with a uniform distribution of 0.5. The fused high-dimensional vector is 576-dimensional. The top 50 principal components are selected through principal component regression, with a variance contribution rate of 96%. The nonlinear regression model uses a gradient boosting tree with a tree depth of 6, a subsampling rate of 0.8, 4 leaves per tree, and 2000 training samples paired with the fused representation and the actual process response. In the constraint regression stage, the feasible temperature range [20℃, 45℃], the feasible pressure range [0.5MPa, 1.5MPa], and the pH control range [3.0, 4.5] of the target domain blueberries are introduced as boundary conditions to truncate predicted values ​​outside these ranges. Specifically, the temperature prediction formula is:

[0017] in Set the temperature value. For feature representation, Let be the regression mapping function obtained through training. If If the value is less than 20, set it to 20; if it is greater than 45, set it to 45. The constraint corrections for pressure and pH are similar, involving interval projection after formula calculation. The mean square error of the final output virtual process response vector on the test set is:

[0018] in These are the actual process response parameters. For the parameters predicted by the regression model, The sample size is 0.015. In this example, the mean squared error is 0.015, indicating that the virtual process response simulator has high precision and meets the physical constraints of blueberry processing, thus ensuring the precise feasibility of the process. S4.5: Based on the difference between the virtual process response vector and the process window allowed by the target domain, calculate the process consistency loss term. This loss term adopts the joint measurement of L2 norm and KL divergence to quantify the degree of deviation of the mapped features from the actual process operation, and serves as the gradient source for the optimization of the mapping network parameters.

[0019] Step S5: Based on the process consistency loss term and the feature alignment loss term, a multi-task learning framework is used to jointly optimize the mapping function. The feature alignment loss term includes a term that minimizes the domain classification accuracy, and the process consistency loss term adjusts the mapping function parameters through backpropagation to ensure that the mapped features satisfy both cross-domain distribution consistency and conform to process operation specifications. Specifically, this includes: S5.1: Train a domain classifier on the source and target domain feature representations output by the mapping network, and use a gradient reversal layer to suppress the domain discrimination capability in order to achieve domain indistinguishability of the feature space, thereby minimizing the feature alignment loss term and obtaining a domain-independent shared feature space. S5.2: Based on the target domain process semantic vector output by the pre-trained process knowledge encoder, process response simulation is performed on the mapped target domain features. The corresponding virtual process response trajectory is predicted through a differentiable physical constraint model to generate the input signal of the process consistency loss term. S5.3: Calculate the deviation between the virtual process response triggered by the mapped features and the target domain process window, and use the mean square error function to quantify the process consistency loss term to form a constraint term that can participate in gradient optimization. S5.4: The feature alignment loss term and the process consistency loss term are weighted and fused to construct a multi-task joint optimization objective function, and the parameters are updated based on the Adam optimizer to simultaneously optimize feature distribution consistency and process operation compliance. The input parameter matrices of the source and target domain feature alignment loss terms and process consistency loss terms are initialized with weighted coefficients. A weight setting method based on the historical convergence characteristics of the loss terms (parameters: initial weights α1∈[0.5,0.7], α2∈[0.3,0.5]) is adopted to realize the initial influence allocation of different loss terms in the joint optimization process. For weighting coefficients , A fusion operation is performed with the corresponding loss term values ​​using a linear weighted summation algorithm (parameter: batch loss record window length L=32) to achieve the initial construction of the joint optimization objective function. The calculation formula is as follows:

[0020] in, The feature alignment loss value, This represents the loss value due to process consistency. These are the weighting coefficients for each loss term; Furthermore, the Adam optimization algorithm (parameters: learning rate η=0.001, β1=0.9, β2=0.999, ε=1e-8) is used to calculate the gradient of the joint optimization objective function. Based on adaptive first-order moment estimation and second-order moment correction, the parameters of each mapping network are updated in a refined manner to improve the cross-domain feature mapping accuracy under process constraints. Batch normalization is performed on the parameter update vector output by the Adam optimizer. A vector norm constraint method (parameter: upper norm threshold γ=5.0) is used to achieve safe control of the update step size within the feasible solution space, preventing overfitting or out-of-bounds adjustments during the optimization process. Through the above joint optimization process, the optimization results of each loss term are dynamically balanced in the weight space, thereby achieving a simultaneous improvement in the consistency of feature distribution and the compliance of process operation. For example, in a cross-domain migration modeling scenario for multiple fruit varieties, the source domain feature alignment loss value... = Target domain process consistency loss value = Initial weighting coefficients = , = Substituting into the joint optimization objective function calculation formula, we get:

[0021] The calculation result is The parameters are used as input to the Adam optimizer for gradient descent iterations (η=0.001, β1=0.9, β2=0.999). The norm of the changes in the mapped network parameters after a single iteration is... The value is below the constraint threshold γ=5.0, and the optimization process converges stably. S5.5: In each optimization iteration, the loss weight allocation strategy is dynamically adjusted, and a gradient magnitude-based balancing mechanism is adopted to ensure that feature alignment and process consistency maintain the same level of technical abstraction during the optimization process, thereby improving the robustness and generalization ability of the mapping function.

[0022] Step S6: During the transfer modeling process, dynamically determine whether the currently mapped feature falls within the process window allowed by the target domain. If it does, output the transfer modeling result; if it does not, trigger the adaptive adjustment mechanism of the mapping function parameters to improve the process compliance of the feature space mapping. Specifically, this includes: S6.1: Perform process window boundary judgment on the target domain feature representation output by the mapping network, and construct a process feasible domain model based on the predefined target domain process operation boundary conditions. The process feasible domain model includes physical constraint parameters such as upper temperature limit, upper pressure limit, pH change rate limit, and enzyme activity maintenance range, so as to obtain the process feasibility status identifier corresponding to the current feature. S6.2: Based on the process feasibility status identifier, execute the process compliance judgment logic. If the current feature falls within the process window range defined by the process feasibility domain model, generate a migration modeling result output instruction signal to trigger the subsequent process parameter generation module to execute parameter set output; if it does not fall within the range, generate a parameter adjustment trigger signal to drive the adaptive optimization of the mapping function parameters. Based on the process feasibility status identification signal, a logical judgment algorithm (parameter: feasible domain model threshold set) is used to perform process window matching calculation on the target domain feature vector to realize the interval inclusion test between the mapped features and the constraint parameters such as temperature upper limit, pressure upper limit, pH change rate limit, and enzyme activity maintenance interval in the process feasible domain model. Furthermore, based on the interval inclusion test results, a binary decision function (output: {0,1}) is used to implement the compliance labeling of the feature state, and a Boolean flag indicating whether the current mapping result satisfies the process constraints of the target domain is obtained. Furthermore, a signal encoding operation (encoding method: bit weight encoding) is performed on the Boolean identifier to realize the mapping conversion from compliance label to control command signal, and the compliance is true (1) is mapped to the migration modeling output trigger signal, and the compliance is false (0) is mapped to the parameter adjustment trigger signal; Furthermore, through the state machine driving mechanism (parameter: control instruction priority table), the migration modeling output trigger signal is sent to the process parameter generation module start channel to realize the generation of subsequent parameter sets such as temperature setpoint, pressing pressure threshold, and pH control range; the parameter adjustment trigger signal is sent to the mapping function adaptive optimization module start channel to realize the local parameter reconstruction of the feature mapping function. Through the above-mentioned judgment and signal conversion link, the process feasibility judgment result is transformed into a control signal, realizing real-time process compliance closed-loop control in the migration modeling process; For example, in the scenario of juice processing of high-acidity, crisp fruits, the upper limit of the feasible process domain model is set to 85℃, the upper limit of pressure is set to 2.5MPa, the pH change rate limit is set to 0.05 units / min, and the enzyme activity maintenance range is set to pH 3.0 to 4.0. The virtual process response vector of the input mapped features is temperature 83℃, pressure 2.0MPa, pH change rate 0.06 units / min, and pH 3.8. Using the interval inclusion test method, the temperature 83℃ is compared with the upper limit of 85℃ and is determined to be compliant; the pressure 2.0MPa is compared with the upper limit of 2.5MPa and is determined to be compliant; the pH change rate of 0.06 units / min is compared with the limit of 0.05 units / min and is determined to be non-compliant; the pH 3.8 is compared with the interval [3.0, 4.0] and is determined to be compliant. Since any parameter is non-compliant, the binary decision function outputs 0, which is converted into a parameter adjustment trigger signal through position weight encoding, driving the entry of the adaptive optimization module of the mapping function. In the signal triggering chain, the state machine directly connects the trigger signal to the parameter optimization task queue according to the priority table, realizing the iterative optimization of non-compliant features. Ultimately, through the above process, it is ensured that in the high-acidity brittle fruit processing scenario, the output features of the mapping function fall within the process-allowed window range after adjustment, guaranteeing the process executability and safety of the transfer modeling results. S6.3: In response to the parameter adjustment trigger signal, start the gradient-guided adaptive adjustment mechanism for the mapping function parameters, use the first-order Taylor expansion method to approximate the local change trend of the feature mapping function at the current parameter point, and combine the gradient direction of the process consistency loss to calculate the parameter update step size and direction to generate the optimized mapping function parameter set. S6.4: Perform process boundary constraint projection processing on the optimized mapping function parameter set, and use the Lagrange multiplier method to project the parameter update direction to the feasible solution space that satisfies the physical constraints of juice processing, so as to obtain a mapping function update model that conforms to the process specifications and prevent feature mapping results that are not feasible in the process of parameter update. S6.5: Based on the mapping function, update the model to perform secondary mapping on the target domain features, generate an updated feature space representation, and re-execute the process window boundary judgment process until the mapped features fall within the process window range allowed by the target domain, so as to achieve closed-loop process compliance control of the migration modeling process.

[0023] Step S7: Generate a pre-optimization parameter set for the new fruit juice processing technology based on the migration modeling results, including temperature setpoints, pressing pressure thresholds, and pH control ranges, to achieve rapid initialization of process parameters in the new scenario. Specifically, this includes: S7.1: Based on the mapped feature space representation output by the transfer modeling module, extract key feature vectors related to the target domain fruit categories to serve as input for initializing process parameters; S7.2: Execute a rule-based parameter mapping function on the extracted key feature vectors. The rule mapping function is constructed from the mapping relationship extracted from the historical process experience database to generate a preliminary set of candidate parameters such as temperature setpoint, pressing pressure threshold, and pH control range. For the key feature vectors extracted by S7.1, a rule-based parameter mapping algorithm (the parameter mapping rules are derived from the mapping relationship table in the historical process experience database) is used to transform the feature vectors of the target domain category into preliminary estimates of process parameters such as temperature, pressure and pH control. Furthermore, by using the feature normalization matching method (parameter: normalization scale matching threshold is set to 0.85), the quantitative correlation between the source domain historical parameter boundary and the target domain feature components is realized, and a coarse-grained temperature setting candidate interval is obtained. Furthermore, by using pressure mapping rules based on fuzzy set reasoning (parameters: the membership function adopts a triangular function, and the pressure is divided into three levels based on 60%, 80%, and 95% of the equipment's rated pressure), the processing pressure candidate values ​​are generated, and a membership matrix is ​​generated as supporting data for the pressure candidate set. Furthermore, by using the acid-base balance model matching method (parameter: acidity compensation coefficient k is 0.12, and linear regression fitting is performed based on the historical pH regulation parameter sequence), the mapping from the acidity-related components of the feature vector to the pH regulation range is realized, and a preliminary set of pH regulation range candidates is generated. A process parameter set synthesis algorithm is used to combine the aforementioned temperature candidate range, pressure candidate set and pH candidate set into a Cartesian combination to generate a complete set of candidate process parameters, and to record the source rule identifier and mapping confidence index for each set of candidate parameters. Through the above algorithm, the key feature vectors of the previous step are transformed into a structured set of candidate process parameters, realizing the rule-based process parameter initialization function, laying the parameter foundation for subsequent physical constraint verification and optimization. For example, in a processing scenario where the target domain is a new type of mango with high sugar and low acidity, the key feature vector input to the mapping algorithm has a dimension of 12, including a normalized peak temperature component (0.78), a mean pressure component (0.65), and an acidity component (0.22). In the temperature mapping rule table, the temperature matching threshold range for this type of feature is 80-90℃. In the pressure mapping rule table, the pressure membership matrix has a membership of 0.92 for 80% of the rated pressure and a membership of 0.41 for 95% of the rated pressure. The pH mapping model is based on the acidity component. The calculated pH control range is 4.5264–4.5764. The process parameter synthesis algorithm performs Cartesian combinations on the temperature range (80, 90°C), candidate pressure values ​​(0.8, 0.95), and pH range (4.5264, 4.5764), generating four sets of candidate parameter combinations. The mapping confidence scores for each combination are recorded as being higher than 0.85. These candidate parameters will undergo feasibility boundary judgment in the next step of physical constraint verification, ultimately retaining the compliant combinations that meet the actual equipment and process safety requirements. S7.3: Based on the candidate parameter set and the physicochemical properties of the target fruit, execute the physical constraint verification module. The physical constraint verification module performs feasibility boundary judgment on the candidate parameter set according to the thermodynamic, fluid dynamics and chemical reaction kinetics models in the juice processing process, so as to obtain a compliant parameter subset that meets the process safety requirements. S7.4: Perform a multi-objective optimization algorithm on a subset of compliant parameters. The multi-objective optimization algorithm is based on two optimization objectives: maximizing juice yield and minimizing energy consumption. It generates a set of non-dominated parameter combinations through a Pareto front search strategy to serve as an optimization candidate set for the process pre-tuning parameter set. S7.5: Based on the optimized candidate set and the operational constraint parameters of the fruit processing equipment in the target domain, execute the parameter adaptability evaluation module. The parameter adaptability evaluation module uses the fuzzy comprehensive evaluation method to score the candidate parameter combinations and finally outputs the temperature set value, pressing pressure threshold, and pH control range with the highest scores as the final process pre-tuning parameter set.

[0024] Step S8: During actual processing, process response data of the target domain fruit is collected, compared and analyzed with the transfer modeling prediction results, a process consistency score is generated, and incremental training and model updates are performed on the process knowledge encoder and mapping network based on the score results. Specifically, this includes: S8.1: Collect process response data generated by the target domain fruit during the actual juice processing. The process response data includes temperature-time curves, pressure-speed sequences, pH change trends and quality output indicators such as clarity, vitamin retention rate, and turbidity stability, in order to obtain real process behavior data samples. S8.2: Perform standardized preprocessing on the collected target domain process response data, including timestamp alignment, sampling frequency unification, and abnormal fluctuation detection, to generate a set of actual feature vectors of the target domain that are consistent with the input format of transfer modeling; S8.3: Based on the mapping function output from the transfer modeling stage, the original features of the target domain fruit are input into the feature space mapping network to generate the mapped feature representation. The process knowledge encoder is then used to decode and generate the predicted process response trajectory to obtain the process behavior sequence predicted by the model. For the original feature vector of the target domain after standardization and preprocessing, a feature space mapping network based on a multilayer perceptron structure (parameters: 128 hidden layer nodes, activation function ReLU) is used to map the input features to a latent feature space representation shared with the source domain. Furthermore, through the learnable nonlinear transformation function inside the mapping network, feature interaction combination is performed on multidimensional processing parameters such as temperature, pressure, and pH to obtain a high-dimensional embedding vector containing cross-domain semantics. Furthermore, the pre-trained process knowledge encoder (architectural parameters: 4 Transformer encoding layers, 8 multi-head attention layers) is invoked, and the above high-dimensional embedding vector is used as the input to the decoding end to realize process semantic inversion and generate a semantic representation matrix of the processing process. Furthermore, based on this semantic representation matrix, a temporal reconstruction module (algorithm: inverse positional encoding transformation) is executed to restore the process response trajectory tensor according to the processing time axis of the target domain, so as to include the predicted trajectory containing temperature-time curve, pressure-speed sequence, and pH change trend; Furthermore, continuous value smoothing is performed on the predicted process trajectory tensor (method: Savitzky-Golay filtering, window length 7, order 3) to generate a smoothed model prediction process behavior sequence, ensuring that the trajectory curve meets the processing constraints in terms of physical interpretability and signal stability. By jointly calling the feature space mapping network and the process knowledge encoder, the original features of the target domain in the previous step are transformed into a predicted process behavior sequence with consistency between temporal sequence and process constraints, so as to realize the visualization and verifiable output of the processing process under the target domain conditions. For example, in the juice processing test of a batch of tropical fruits (category identifier: TF-A1), the original feature vector of the target domain includes a temperature trajectory of 600 steps, a pressure trajectory of 600 steps, and a pH trajectory of 600 steps. The mapping network is set with 128 hidden layer nodes and the activation function ReLU. The weight matrix obtained through training maps the input feature tensor to a latent space of dimension 64. The process knowledge encoder uses a 4-layer Transformer with 8 multi-head attention per layer to decode the latent space vector into a semantic matrix containing 3 trajectories, each trajectory being 600 steps long. The temporal reconstruction module performs an inverse positional encoding transformation on the semantic matrix to generate a predicted time series containing temperature, pressure, and pH. The smoothing process uses Savitzky-Golay filtering (window length 7, order 3), which reduces the instantaneous fluctuation amplitude of the temperature prediction curve from ±2℃ to ±0.5℃, the noise peak of the pressure curve from 15% to 3%, and effectively eliminates the signal drift trend of the pH curve. The final output of the predicted process behavior sequence achieved a trajectory similarity of 0.94 in the comparison analysis with the actual processing trajectory, which meets the engineering verification standard for process consistency scoring. S8.4: The process behavior sequence predicted by the model is compared and analyzed step by step with the actual process response data collected in the target domain. The trajectory similarity is calculated based on the dynamic time warping (DTW) algorithm, and the process consistency score is calculated in combination with the deviation of the quality output index to quantify the engineering feasibility of the transfer modeling results. S8.5: Construct an incremental training sample set based on the process consistency score results, perform online incremental learning on the process knowledge encoder and mapping network, and use the mini-batch gradient descent algorithm to update the network parameters to improve the model's transfer accuracy and process compliance in the target domain fruit category; S8.6: After the model update is completed, the updated mapping function and process knowledge encoder are written into the model version library, and the corresponding process consistency score, update timestamp and target domain fruit category identifier are recorded to build a traceable model evolution log.

[0025] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0026] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and rules of the present invention should be included within the scope of protection of the present invention.

Claims

1. A novel intelligent transfer learning method for processing fruit juice based on transfer learning, characterized in that, Includes the following steps: S1: Collect key process parameter trajectories and quality output indicators of multiple types of fruits during the juice processing process to form a structured process log library; S2: Standardize and preprocess the data in the structured process log library to generate preprocessed process log data; S3: Based on the preprocessed process log data, a lightweight process knowledge encoder is constructed, and the Transformer architecture is used to encode the process log into a low-dimensional process semantic vector. S4: Construct a feature space mapping network, input the original features of the source domain and the target domain fruits into the mapping network, generate the mapped feature space representation, and introduce a differentiable process consistency loss term; S5: Based on the process consistency loss term and the feature alignment loss term, a multi-task learning framework is used to jointly optimize the mapping function. The feature alignment loss term includes a domain classification accuracy minimization term, and the process consistency loss term adjusts the mapping function parameters through backpropagation. S6: During the transfer modeling process, dynamically determine whether the currently mapped feature falls within the process window allowed by the target domain. If it is determined to fall within the window, execute the transfer modeling result output; if it is determined not to fall within the window, trigger the adaptive adjustment mechanism of the mapping function parameters. S7: Based on the migration modeling results, generate a set of pre-optimization parameters for the processing technology of new fruit juice, including temperature setpoint, pressing pressure threshold, and pH control range.

2. The intelligent transfer learning method for processing new fruit juices based on transfer learning as described in claim 1, characterized in that, Following step S7, the following is also included: S8: Collect process response data of target domain fruits during actual processing, compare and analyze the data with the prediction results of transfer modeling, generate process consistency scores, and perform incremental training and model updates on the process knowledge encoder and mapping network based on the scores.

3. The intelligent transfer learning method for processing new fruit juices based on transfer learning as described in claim 1, characterized in that, Step S1 specifically includes: Multi-source sensing was used to collect data on the juice processing of multiple types of fruits, including temperature-time curves, pressure-speed sequences, and pH change trends, to construct a database of original process parameters. The process parameter trajectory data is timestamped and the sampling frequency is standardized. A linear interpolation algorithm is used to fill in the missing data points to obtain a standardized process parameter sequence with enhanced temporal consistency. Based on the quality evaluation standard for fruit juice processing, we collect quality output index data for each type of fruit after processing, including clarity, vitamin retention rate and turbidity stability, to form a quality output index dataset. The standardized process parameter sequence is structurally associated with the corresponding quality output index data, and a structured process log library for multiple fruit categories is established using a relational database format. In the structured process log library, each type of fruit is labeled with its category identifier, processing batch number, and equipment model information, forming a process log metadata system.

4. The intelligent transfer learning method for processing new fruit juices based on transfer learning as described in claim 3, characterized in that, The multi-source sensing acquisition includes a multi-source acquisition array composed of temperature sensors, pressure sensors, and pH sensors. The temperature measurement range is 0-150℃ with an accuracy of ±0.1℃; the pressure measurement range is 0-2MPa with an accuracy of ±0.01MPa; and the pH measurement range is 0-14 with an accuracy of ±0.

01. It continuously acquires the temperature-time curves, pressure-speed sequences, and pH change trends during the juice processing of multiple types of fruits.

5. The intelligent transfer learning method for processing new fruit juices based on transfer learning according to claim 1, characterized in that, Step S2 specifically includes: Missing value detection was performed on the temperature-time curves, pressure-speed sequences, and pH trend data collected from the structured process log library. Linear interpolation was used to estimate the missing time points to obtain a complete and continuous sequence of process parameters. Based on the complete and continuous process parameter sequence, outliers are identified and removed using a statistical method based on Z-score. ±3σ is set as the threshold, and data points that exceed this range are replaced with the moving average of nearby time points. The processed process parameter sequence is standardized and aligned on the time axis, and the dynamic time warping algorithm is used to map the processing of different types of fruits to a unified time scale. Based on the output time-aligned data, a unified multidimensional feature matrix is ​​constructed, and the temperature, pressure, and pH parameters are organized into a tensor structure according to the time step. The generated multidimensional feature matrix is ​​standardized and normalized, and the Min-Max scaling method is used to map the parameters of each dimension to a unified range.

6. The intelligent transfer learning method for processing new fruit juices based on transfer learning according to claim 5, characterized in that, The multidimensional feature matrix is ​​divided according to the time step Δt=1s, and the parameter type index is sequentially encoded as temperature: 0, pressure: 1, pH:

2. It is organized using a third-order tensor structure, and the low-density sampling positions are filled with zero values ​​to generate a corresponding mask matrix, which is used for the attention weight calculation of the process knowledge encoder.

7. The intelligent transfer learning method for processing new fruit juices based on transfer learning according to claim 1, characterized in that, Step S3 specifically includes: Windowing and slicing are performed on the time series parameter trajectories in the preprocessed process log data to construct input sample segments; Based on the input sample fragment, a multi-head self-attention mechanism is used to perform feature interaction modeling on the temperature-time curve, pressure-speed sequence, and pH change trend in the process parameter trajectory, so as to obtain a multi-scale process state characterization. The multi-scale process state representation is subjected to position encoding and category embedding fusion operations to obtain the fused process state representation; Based on the fused process state representation, a hierarchical Transformer encoding layer is used for deep semantic abstraction, and a low-dimensional process semantic vector is extracted through multi-layer nonlinear transformation. Normalization and learnable projection operations are performed on the low-dimensional process semantic vector to generate a standardized process knowledge embedding representation, which is then output to the feature space mapping network.

8. The intelligent transfer learning method for processing new fruit juices based on transfer learning according to claim 1, characterized in that, Step S4 specifically includes: The original feature data of fruits in the source and target domains are normalized to obtain input feature tensors at a uniform scale. Based on the input feature tensor, a feature space mapping network with a multilayer perceptron structure is constructed. The input features are embedded in a high dimension through a learnable nonlinear transformation function to generate a latent feature space representation shared by the source and target domains. The low-dimensional process semantic vector output by the process knowledge encoder is used as prior constraint information. It is then cross-modeled with the mapped features through a differentiable fusion mechanism to generate process-aware feature representations. Based on the fusion representation of mapped features and process semantic vectors, a virtual process response simulator is constructed. A regression model is used to predict the corresponding virtual process parameter output under this feature combination, forming a virtual process response vector. The process consistency loss term is calculated based on the difference between the virtual process response vector and the process window allowed by the target domain.

9. The intelligent transfer learning method for processing new fruit juices based on transfer learning according to claim 8, characterized in that, The virtual process response simulator uses principal component regression and gradient boosting tree regression models to map high-dimensional process-aware features to virtual process parameters, and uses the physical boundaries of the target domain process window to implement hard range constraints on temperature, pressure and pH parameters during the regression process.

10. The intelligent transfer learning method for processing new fruit juices based on transfer learning according to claim 1, characterized in that, Step S5 specifically includes: A domain classifier is trained on the source and target domain feature representations output by the mapping network. A gradient inversion layer is used to suppress the domain discrimination capability, thereby obtaining a domain-independent shared feature space. Based on the target domain process semantic vector output by the pre-trained process knowledge encoder, process response simulation is performed on the mapped target domain features. The corresponding virtual process response trajectory is predicted through a differentiable physical constraint model, and a process consistency loss term is generated. The deviation between the virtual process response triggered by the mapped features and the target domain process window is calculated, and the mean square error function is used to quantify the process consistency loss term to form a constraint term that can participate in gradient optimization. The feature alignment loss term and the process consistency loss term are weighted and fused to construct a multi-task joint optimization objective function, and parameter updates are performed based on the Adam optimizer. In each optimization iteration, the loss weight allocation strategy is dynamically adjusted.

Images