Preservation method for water content of preserved fruit based on multi-sensor data driving

Abstract

The invention provides a preserved fruit water content prediction method based on multi-sensor data driving, and belongs to the technical field of food processing intelligent monitoring based on machine learning. The method comprises the following steps: firstly, constructing a heterogeneous sensing system comprising an environment heat and humidity field, a material spectrum and weight response, and acquiring and processing to obtain a unified moment state vector; secondly, designing a dry dynamic behavior feature extraction mechanism, and constructing a multi-dimensional dynamic feature set comprising transient pulses, dynamic rates and a whole-process cumulative effect by using sliding window fitting and a time domain integration strategy; and further, a double-flow space-time attention fusion network is constructed, and high-precision reasoning of the moisture content is realized through decoupling processing of a spectral flow and a physical flow, self-adaptive gating fusion and physical residual direct connection. According to the method, the physical consistency and precision of long-sequence prediction are remarkably improved, deep complementation of spectral traits and the environment is realized, and the generalization ability of the model to different processes and batches is enhanced.

Description

Technical Field

[0001] This invention belongs to the field of intelligent monitoring technology for food processing based on machine learning, and particularly relates to a method for predicting the moisture content of dried fruit based on multi-sensor data. Background Technology

[0002] Moisture content is a core indicator for measuring the flavor, texture, safety, and shelf life of dried fruit, and its accurate monitoring is directly related to product quality control, energy consumption, and processing line efficiency. However, dried fruit undergoes significant physical and chemical changes during the drying process, such as internal moisture migration, changes in external evaporation rates, and the collapse and compaction of the fruit pulp, making it impossible to obtain the internal moisture content directly through simple surface observation. This makes real-time moisture content monitoring during the drying process a technological bottleneck in the food processing industry.

[0003] Current methods for detecting the moisture content of dried fruit can be mainly divided into three categories: manual sampling, traditional single-sensor estimation methods, and prediction methods based on empirical models.

[0004] Manual sampling method: This method relies on staff periodically removing samples from the drying chamber and determining the moisture content using a drying method or a rapid moisture meter. Although it has high measurement accuracy, it has significant drawbacks: low sampling frequency and poor timeliness, failing to reflect the dynamic changes in the drying process; only a few measurement points can be provided for each batch, which cannot represent the overall production situation; frequent opening of the chamber for sampling will disturb the drying environment, reduce production efficiency, and increase labor costs.

[0005] Traditional single-sensor estimation methods: Temperature, humidity, or weight sensors are widely used in processing equipment as indirect estimates of moisture content. For example, the rate of water evaporation is inferred from changes in humidity within the cavity, or the moisture content range is estimated from weight loss. However, these methods have significant limitations: a single sensor can only reflect one physical quantity and cannot describe the multidimensional changes in moisture migration within the dried fruit.

[0006] Estimation methods based on empirical formulas or static models: Some studies construct empirical formulas based on heat and mass transfer theory, or use simple regression methods to infer moisture content. These methods often rely on static or approximately linear variable relationships, making it difficult to handle highly nonlinear dynamic behaviors during the drying process: for example, different batches of dried fruit vary significantly in initial moisture content, pulp thickness, sugar structure, etc., making it difficult to apply empirical formulas uniformly.

[0007] Although existing methods are widely used in actual processing, they still have the following key drawbacks: (1) Moisture content cannot be directly measured and is difficult to estimate accurately: Traditional methods rely on single variables or indirect inference, and cannot obtain the true moisture distribution inside the dried fruit. (2) The drying dynamic process cannot be effectively characterized: Existing methods generally only use point values ​​at a certain moment, and lack reflection of moisture migration trends, evaporation rates and their changing patterns. (3) Multi-source sensor data lacks an effective synchronization and fusion mechanism: Different sensors have inconsistent sampling frequencies and different acquisition delays, resulting in time-series misalignment of data, making it difficult to use directly for modeling and affecting prediction accuracy. Therefore, there is an urgent need for an intelligent moisture content prediction method that can fuse multi-source sensor data, accurately characterize the drying dynamic behavior, and has real-time prediction capabilities, so as to achieve high-quality control and intelligent upgrading of the dried fruit drying process. Summary of the Invention

[0008] To address the above problems, this invention proposes a method for predicting the moisture content of dried fruit based on multi-sensor data, comprising the following steps: S1 collects environmental constraint data and material response data during the fruit drying process, including temperature time-series data, humidity time-series data, weight time-series data, and spectral time-series data; S2, preprocesses the data obtained from S1 to obtain a multi-source synchronous time series dataset; S3. Calculate the first-order difference of the multi-source synchronous time-series dataset to capture transient pulse changes and obtain the transient difference feature vector. Then, extract the weight decrease slope and spectral evolution slope to construct kinetic rate features, characterize the instantaneous evaporation efficiency, and obtain the kinetic rate feature vector. At the same time, adopt the time-domain integration strategy to calculate the cumulative total water evaporation, heat and moisture load history, and spectral evolution distance to construct the full-process cumulative features and obtain the process cumulative feature vector. S4. The feature data obtained from S2 and S3 are normalized through independent channels to obtain a normalized full feature set. The normalized full feature set is divided into spectral feature clusters, physical feature clusters and global cumulative feature clusters. Using the historical backtracking sliding window mechanism, the three types of feature clusters are encapsulated into independent three-dimensional time series tensors, including spectral time series tensors, physical time series tensors and global cumulative time series tensors. S5 takes the three time-series tensors obtained in S4 and inputs them into the trained moisture content prediction model that can decouple the spectrum and physical characteristics, and outputs the predicted moisture content of the dried fruit.

[0009] Preferably, the water content prediction model is a dual-stream spatiotemporal attention fusion network architecture, comprising two main processing streams: the spectral stream uses a 1D convolutional neural network and channel attention mechanism to extract band evolution features; the physical stream uses a long short-term memory network and temporal attention mechanism to capture the effects of environmental lag; the spectral temporal tensor and the physical temporal tensor are dynamically weighted and fused through an adaptive gating unit; in addition, a physical residual path is introduced, and the global cumulative temporal tensor generated by S4 is directly injected into the output layer to constrain the prediction results with the physical conservation law.

[0010] Preferably, the transient difference feature vector is obtained in step S3, and the specific process is as follows: For synchronized time-series data of temperature, humidity, weight, and spectrum, a synchronized state vector is constructed, and a first-order difference operation is performed to obtain the first-order difference scalar features of temperature, humidity, weight, and the full-band spectrum. Instantaneous change indices of each physical quantity are calculated by subtracting the state vector of the current time step from the state vector of the previous time step. Finally, a feature vectorization integration operation is performed, and the calculation results are time-aligned based on a unified timestamp index. Then, the first-order difference scalar features of temperature, humidity, and weight are sequentially concatenated with the flattened first-order difference vector features of the full-band spectrum, and the cumulative change features are also included in the combination to construct a transient differential feature vector characterizing the drying dynamics at the current moment.

[0011] Preferably, the dynamic velocity feature vector is obtained in step S3, and the specific process is as follows: First, based on the multi-source synchronous time-series dataset within the drying chamber, the first derivatives of the synchronous temperature and humidity time-series data with respect to time are calculated to construct temperature and humidity environmental response gradient features reflecting transient environmental changes. Second, for the synchronous weight time-series data, a linear trend fitting method based on a sliding window is adopted. By performing linear regression analysis within a preset time window of length T and extracting the slope, an instantaneous drying rate index that can overcome minor mechanical vibrations and accurately reflect the rate of moisture vaporization and escape is constructed. Finally, for the synchronous spectral data, the evolution rate of intensity of each band on the time axis is calculated to construct a spectral evolution trend feature vector. Finally, a feature vectorization integration operation is performed. Based on a unified timestamp index, the calculated rates and gradient indices are time-aligned. Then, the scalar-form temperature and humidity environmental response gradient features, the instantaneous drying rate index, and the flattened spectral evolution trend feature vector are sequentially concatenated to construct a dynamic rate feature vector that reflects the transition of the drying stage and the intensity of internal moisture migration.

[0012] Preferably, the process cumulative feature vector is obtained in S3, specifically as follows: First, the first-order difference scalar feature of weight from the start of drying to the current time is integrated to generate a cumulative total moisture evaporation vector. Second, the first-order difference scalar feature of humidity in the drying environment is accumulated over time to generate a cumulative environmental dehumidification potential vector. Finally, the first-order difference vector feature of the full-band spectrum is accumulated in vector space to generate a cumulative spectral evolution vector. Finally, a feature vectorization integration operation is performed, and the cumulative indicators are aligned over time based on a unified timestamp index. Then, the cumulative total moisture evaporation vector, the cumulative environmental dehumidification potential vector, and the flattened cumulative spectral evolution vector are sequentially concatenated to construct a process cumulative feature vector that can completely characterize the overall drying process.

[0013] Preferably, in step S4, the feature data obtained in S2 and S3 are normalized through independent channels to obtain a normalized full feature set. The specific process is as follows: First, based on the spectral synchronization time-series data and its evolution characteristics, the calibration saturation value and dark current noise floor value of the spectral sensor are read. Then, using the maximum-minimum linear transformation formula, the original spectral intensity and spectral difference data are mapped to... First, the relative morphological characteristics between bands are preserved within the specified range. Second, for data with clear physical limits, the real-time acquired physical quantities are strictly compressed to a range based on the sensor's hardware range and the safety boundaries set by the manufacturing process. Within the specified range; for the transient difference feature vector, kinetic rate feature vector, and process cumulative feature vector generated by S3, a statistical boundary mapping strategy is adopted. Based on historical batch big data statistics, the maximum and minimum values ​​are manually calibrated to determine the upper and lower bounds of the statistical distribution of each dynamic feature throughout the drying process. Using the statistical boundaries as a benchmark, the first-order difference features containing positive and negative fluctuations and the monotonically changing cumulative features are uniformly mapped to... For the rate of change feature, 0.5 corresponds to zero change, and 0 and 1 correspond to the maximum negative change and the maximum positive change, respectively. All heterogeneous features are converted into a standardized form, and the normalized full feature set is output.

[0014] Preferably, in step S4, the normalized full feature set is divided into a spectral feature cluster, a physical feature cluster, and a global cumulative feature cluster. The specific process is as follows: First, high-dimensional spectral-related feature vectors are extracted from the normalized full feature set to construct a spectral feature cluster, which includes normalized spectral synchronization time-series data, full-band spectral first-difference vector features, and spectral evolution trend feature vectors. Principal component analysis is then used to perform independent projection transformations on the spectral synchronization time-series data, full-band spectral first-difference vector features, and spectral evolution trend feature vectors. For each type of feature, based on the criterion that the cumulative contribution rate is greater than 95%, the top 10 principal component components after transformation are extracted. Finally, the three sets of independently dimensionality-reduced principal component feature vectors are concatenated along the feature dimension to reconstruct a simplified spectral state vector. Secondly, a physical feature cluster is constructed, which aggregates normalized synchronous time-series data of temperature, humidity, and weight, as well as temperature and humidity environmental response gradient features, first-order differential scalar features of temperature, first-order differential scalar features of humidity, first-order differential scalar features of weight, and instantaneous drying rate index. A vector concatenation operation is performed to concatenate the scalar features in a preset order along the channel dimension to construct a physical state vector that can reflect the current drying dynamic boundary conditions. Finally, for the residual path, a global cumulative feature cluster is constructed, and the normalized process cumulative feature vector is directly selected.

[0015] Preferably, in step S4, the three types of feature clusters are encapsulated into independent three-dimensional temporal tensors using a historical backtracking sliding window mechanism. The specific process is as follows: First, for the spectral feature clusters, at the current time... At the cutoff point, a length of [length] is taken forward. Using continuous time-step data, a spectral temporal tensor is constructed, and the dimension of the tensor is defined as follows: ,in This refers to the batch size during model training or inference. The time window length, First, let's define the dimension of the spectral state vector. Second, for each physical feature cluster, perform the same sliding window truncation operation simultaneously to construct a physical temporal tensor. The dimension of this tensor is defined as... ,in Let be the dimension of the physical state vector; then, for the global cumulative feature cluster, construct a global cumulative temporal tensor, the dimension of which is defined as... ,in Accumulate feature vectors for the process.

[0016] Preferably, the moisture content prediction model specifically includes a spectral feature extraction stream unit, a physical environment extraction stream unit, a multimodal gated fusion and physical residual path unit, and an output unit; The spectral feature extraction unit feeds the spectral temporal tensor into a one-dimensional convolutional neural network layer, where the convolutional kernel slides along the feature dimension to extract morphological features of local bands. Next, the squeeze and activation channel attention module compresses the feature map using global average pooling, automatically learns the importance weights of different spectral bands using a fully connected layer, and recalibrates the convolutional feature channels. Finally, the weighted feature sequence is input into a bidirectional long short-term memory network, capturing the evolution trajectory of spectral traits from both forward and reverse time dimensions, and outputting the spectral hidden layer state vector. The multimodal gating fusion unit inputs the physical temporal tensor into a multi-layer stacked long short-term memory network to extract the temporal dependencies of temperature, humidity, and weight changes over a long time span. Next, a time attention mechanism module is connected after the output layer of the long short-term memory network. By calculating the correlation score between the hidden state at a historical moment and the state at the current moment, a time weight vector is generated. After attention-weighted summation, the physical temporal feature vector is output. The multimodal gated fusion and physical residual path unit is based on a learnable gated fusion unit and calculates dynamic gating coefficients through a fully connected layer. The value ranges from 0 to 1, and the spectral hidden layer state vector and physical temporal feature vector are weighted and fused accordingly; the gating coefficient is then used. Soft weighting of spectral features, utilizing The physical features are complementary and weighted; residual connections are established, and the global cumulative temporal tensor is mapped through an independent fully connected layer and then added to the multimodal fusion feature vector in the form of residual connections to obtain the physical constraint fusion feature vector.

[0017] The output unit outputs the predicted moisture content of the dried fruit at the current moment through a linear regression layer.

[0018] Compared with the prior art, the present invention has the following beneficial effects: (1) Deep network architecture based on dual-stream decoupling and physical residual direct connection: Breaking through the limitations of traditional single long short memory network in processing heterogeneous data, it uses a 1D convolutional neural network combined with channel attention to extract spectral trait evolution, uses a long short memory network combined with time attention to extract environmental lag, and directly injects the cumulative amount of the whole process into the output layer through the physical residual path, effectively preventing non-physical drift of deep models in long sequence prediction; (2) Full-process cumulative feature construction technology for the entire drying life cycle: To address the problem that relying solely on instantaneous rates cannot accurately determine the drying stage, a feature construction strategy based on time-domain integration is proposed. By calculating the total cumulative moisture evaporation, the history of heat and moisture load, and the spectral evolution distance, a macroscopic progress bar for the model is constructed, explicitly solving the problems of long-term dependency loss and stage ambiguity in long-sequence modeling; (3) Multi-source heterogeneous data splitting and encapsulation and multi-channel tensor construction mechanism: Abandoning the simple feature stitching method, a splitting and encapsulation strategy is proposed. Based on physical properties, features are divided into spectral feature clusters, physical feature clusters and global cumulative feature clusters, and independent three-dimensional temporal tensors are constructed using historical backtracking sliding windows. This structured input realizes deep adaptation of the data stream and network dual-stream architecture, maximizing the expressive power of features. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the overall technical route of the present invention.

[0020] Figure 2 This is a schematic diagram of the layout of a multi-source sensor collaborative sensing and drying equipment.

[0021] Figure 3 A flowchart for time-series synchronization and vectorization encapsulation of multi-source heterogeneous data.

[0022] Figure 4 This is a flowchart of dynamic feature extraction based on transient difference, dynamic rate, and process accumulation.

[0023] Figure 5 This is a schematic diagram of multidimensional feature transformation and tensor encapsulation for a two-stream architecture.

[0024] Figure 6 This is a schematic diagram of the network structure of the moisture content prediction model.

[0025] Figure 7 This is a comparison chart of the moisture content prediction curve and the cumulative drift of errors in the example.

[0026] Figure 8 This is a schematic diagram illustrating the dynamic evolution of the adaptive gating fusion weights in the embodiment.

[0027] Figure 9 The example shows the attention heatmap of the spectral channels. Detailed Implementation

[0028] The overall technical flowchart of the multi-source sensor data-driven method for predicting the moisture content of dried fruit proposed in this invention is as follows: Figure 1 As shown, the specific steps are as follows: S1. Deployment of Multi-Source Sensor Array and Heterogeneous Data Acquisition: Constructing a hardware acquisition system for comprehensive sensing of the drying chamber's environmental field and material state. This step aims to capture environmental constraint data and material response data during the drying process through a distributed temperature and humidity sensor array, a dynamic weighing module, and a non-contact spectral acquisition unit. Throughout the drying process, each sensor acquires data in parallel at a preset frequency, and the data acquisition system adds a unified timestamp, forming a raw multi-source heterogeneous time-series dataset covering the thermodynamic and kinetic behaviors of the drying process, providing a physical truth foundation for subsequent modeling.

[0029] S2. Time-series synchronization and vectorization encapsulation of heterogeneous data: This step addresses the issues of sampling frequency mismatch from multiple sensor sources and environmental noise interference, constructing a standardized time-series input. The aim is to eliminate the temporal and spatial heterogeneity of the original multi-source heterogeneous time-series dataset from the dried fruit processing process. First, a multi-level statistical filtering algorithm is used to remove instantaneous pulses and random fluctuations from the original multi-source heterogeneous time-series dataset, and outliers are repaired based on physical process constraints. Then, a globally unified time reference axis is established, and upsampling and interpolation mapping are performed on the low-frequency sampled spectral data to ensure strict alignment with high-frequency physical quantities. Finally, synchronized time-series data for temperature, humidity, weight, and spectrum are obtained and integrated into a multi-source synchronized time-series dataset.

[0030] S3. Dynamic Feature Extraction for Drying Kinetics: Mining physical semantic features reflecting the rate of moisture migration and the cumulative effect throughout the process from multi-source synchronous time-series datasets. This step aims to enhance the model's ability to perceive drying speed and progress. First, the first-order differences of synchronous time-series data on temperature, humidity, weight, and spectrum are calculated to capture transient pulse changes and obtain transient difference feature vectors. Second, a sliding window linear fitting algorithm is used to extract the slope of weight decrease and the slope of spectral evolution to construct kinetic rate features, characterizing instantaneous evaporation efficiency and obtaining a kinetic rate feature vector. Third, a time-domain integration strategy is used to calculate the cumulative total moisture evaporation, the history of heat and moisture load, and the spectral evolution distance to construct cumulative features throughout the process and obtain a process cumulative feature vector. This feature serves as a macroscopic progress bar for the model, resolving the stage ambiguity problem in long-sequence prediction. Finally, the transient difference feature vector, the kinetic rate feature vector, and the process cumulative feature vector are concatenated into a time-series dynamic feature sequence.

[0031] S4. Standardization and Structured Tensor Construction in Multidimensional Feature Space: Heterogeneous features are mapped to a unified dimensional space, and multi-channel input tensors are constructed according to the requirements of the deep network architecture. This step adopts a split-encapsulation strategy. First, independent channel normalization based on physical boundaries is performed on features with different physical properties. Second, the features obtained in S2 and S3 are divided into spectral feature clusters, physical feature clusters, and global cumulative feature clusters. Finally, using a history-backtracking sliding window mechanism, the three types of feature clusters are encapsulated into independent three-dimensional temporal tensors: spectral temporal tensor, physical temporal tensor, and global cumulative temporal tensor, respectively, achieving deep adaptation between data flow and model architecture.

[0032] S5. Moisture Content Prediction Model Construction and Training: A deep prediction model capable of deeply decoupling spectral and physical features is constructed and trained. This step proposes a dual-stream spatiotemporal attention fusion network architecture. This network comprises two main processing streams: the spectral stream utilizes a 1D convolutional neural network and channel attention mechanism to extract band evolution features; the physical stream utilizes a long short-term memory network and temporal attention mechanism to capture the effects of environmental lag. The spectral and physical time-series tensors are dynamically weighted and fused using adaptive gating units. Furthermore, a physical residual path is introduced, directly injecting the globally accumulated time-series tensor generated in S4 into the output layer, constraining the prediction results with physical conservation laws. The network is trained using the backpropagation algorithm to obtain a moisture content prediction model with high accuracy and physical consistency.

[0033] S6. Online Inference and Closed-Loop Feedback Control Based on Real-Time Streaming Data: The trained moisture content prediction model is deployed at the edge to achieve moisture content monitoring and closed-loop control of the drying process. This step aims to achieve intelligent production. The system maintains a first-in-first-out rolling time window in memory and uses an incremental iterative algorithm to update the cumulative features defined in S3 in real time, eliminating the need to repeatedly calculate historical data. Three independent tensors generated in real time are input into the dual-stream spatiotemporal attention fusion network model, outputting the current moisture content prediction value. The system compares the prediction value with the target process parameters and generates dynamic adjustment commands or automatic shutdown commands, thereby achieving unmanned closed-loop management of the drying process.

[0034] The specific implementation process of the present invention will be described in detail below with reference to specific embodiments.

[0035] S1, Multi-source sensor data acquisition The purpose of this step is to utilize a data acquisition system composed of multiple sensors throughout the entire fruit drying process to continuously acquire key physical quantities of the drying environment and fruit samples at a fixed sampling period of 1 second. This will form a raw time-series dataset that reflects multi-dimensional physical behaviors such as temperature changes, humidity changes, mass decay, and surface spectral characteristics. The acquired data will serve as the foundational input for subsequent data preprocessing, dynamic feature construction, and moisture content prediction model training.

[0036] S1-1 Distributed Temperature and Humidity Acquisition: This step acquires time-series temperature and humidity data inside the drying chamber. In practice, four temperature sensors and four humidity sensors are deployed at the hot air inlet area, hot air outlet area, above the fruit tray, and below the fruit tray inside the drying chamber. The temperature sensors use digital integrated temperature probes with a measurement range of 0–150℃ and a measurement accuracy of ±0.2℃; the humidity sensors use capacitive humidity probes with a measurement range of 0–100%RH and a measurement accuracy of ±1.5%RH. Finally, the data from each sensor is organized into structured data according to its location number and timestamp to obtain the time-series temperature and humidity data.

[0037] S1-2 Dynamic Weight Real-Time Monitoring: The weight measurement unit records the mass change of the dried fruit samples in real time during the drying process, obtaining weight time-series data. In actual implementation, a strain gauge load cell is installed at the bottom of the tray holding the dried fruit samples. This load cell has a range of 0–5 kg and a resolution of 0.1 g. The load cell is connected to the tray via a fixed connection, and a four-point shock-absorbing isolation structure is installed below it to reduce the impact of equipment vibration and airflow impact on the stability of the weighing measurement. Different batches of dried fruit are distinguished and stored by tray number or batch ID during drying.

[0038] S1-3 Non-contact Spectral Surface Sensing: The spectral acquisition unit is used to collect spectral characteristics or reflectance energy characteristics of the dried fruit surface to obtain spectral time-series data, characterizing the optical behavior of surface moisture, color changes, and tissue state. The spectral acquisition unit uses a linear spectral sensor with a 450–1000 nm wavelength band and a spectral resolution of 5 nm, installed at a fixed position 15 ± 1 cm from the dried fruit surface. A closed light-shielding structure is installed externally to the spectral sensor to eliminate external light interference and ensure optical path stability. The spectral acquisition unit outputs spectral time-series data with a 2-second sampling period, where a single spectral data set contains 110 spectral bands. All spectral data are timestamped to ensure strict temporal alignment and fusion with temperature, humidity, and weight data in subsequent steps.

[0039] The sensor layout diagram for this step is shown below. Figure 2 As shown, the original multi-source heterogeneous time-series dataset of the entire drying process is finally obtained. This dataset contains temperature time-series data, humidity time-series data, weight time-series data, and spectral time-series data.

[0040] S2. Preprocessing of heterogeneous data The overall process is as follows Figure 3As shown, this step aims to systematically process the original multi-source heterogeneous time-series dataset collected in step S1. This step eliminates the temporal and spatial heterogeneity of the data through signal conditioning, full-time-domain alignment, and vectorization encapsulation mechanisms, constructing a multi-source synchronous time-series dataset with a unified time step, continuous numerical values, and a high signal-to-noise ratio. This provides a strictly aligned data foundation for subsequent dynamic feature extraction.

[0041] S2-1 Time Synchronization: The goal of this step is to provide a unified time reference for the multi-source sensor data acquired in S1-1, S1-2, and S1-3, ensuring cross-dimensional consistency analysis between different physical quantities. This unit performs synchronization timestamp marking on all sensor data based on the unified time reference frequency of 1Hz of the drying equipment control system.

[0042] S2-2 Signal Conditioning and Quality Enhancement: This step performs multi-level signal cleaning and anomaly repair on the raw data, including temperature time-series data, humidity time-series data, weight time-series data, and spectral time-series data, to improve the stability and reliability of the time-series signals. For high-frequency noise in the weight time-series data caused by mechanical vibration or airflow impact, and for optical path fluctuations or electromagnetic interference in the spectral time-series data, the system employs an adaptive statistical filtering algorithm, using a preset length... The time window is used to smooth the original sequence, effectively suppressing instantaneous pulses and random jitters while preserving the trend of physical quantity changes.

[0043] S2-3 Full-Time Domain Alignment of Asynchronous Data: This step aims to address the issue of sampling frequency mismatch among multi-source sensors. For data with sampling frequencies consistent with the reference axis, it is directly retained and aligned to the corresponding time step. For data with sampling frequencies lower than the reference axis, upsampling and interpolation mapping operations are performed. Linear interpolation or spline interpolation algorithms on the time axis are used to fill in the data gaps at non-sampling times, improving its time resolution to be consistent with the global reference. This process yields synchronized time-series data for temperature, humidity, weight, and spectrum. These four types of synchronized time-series data are then integrated into a multi-source synchronized time-series dataset, ensuring that subsequent models can capture the complete system state at every moment.

[0044] S3, Construction of Temporal Dynamic Features The main function of this step is to extract feature sequences that characterize the dynamic changes in the dried fruit drying process from the multi-source synchronous time-series dataset generated in step S2. These feature sequences include first-order difference features, rate of change features, and cumulative change parameters, which describe the dynamic behavior of the time series. Figure 4 As shown.

[0045] S3-1 Differential Feature Calculation: Using the multi-source synchronized time series dataset generated in step S2 as input, for the synchronized time series data of temperature, humidity, weight, and spectrum, a synchronization state vector is constructed and a first-order differential operation is performed on the bit-series time series to obtain the first-order differential scalar features of temperature, humidity, weight, and full-band spectrum.

[0046] S3-2 Drying Kinetic Rate Feature Extraction: First, based on the multi-source synchronous time-series dataset within the drying chamber, the first derivatives of the synchronous temperature and humidity time-series data with respect to time are calculated to construct temperature and humidity environmental response gradient features reflecting transient environmental changes. Second, for the synchronous weight time-series data, a linear trend fitting method based on a sliding window is employed. By performing linear regression analysis within a preset time window of length T and extracting the slope, an instantaneous drying rate index capable of overcoming minor mechanical vibrations and accurately reflecting the rate of moisture vaporization and escape is constructed. Finally, for the synchronous spectral time-series data, the evolution rate of intensity of each band on the time axis is calculated to construct a spectral evolution trend feature vector. Finally, all feature vectors are sequentially concatenated to construct a kinetic rate feature vector that effectively reflects the transition between drying stages and the intensity of internal moisture migration.

[0047] S3-3 Whole-Process Cumulative Effect and Drying Process Characterization Unit: This step aims to construct a set of state features reflecting the overall drying process. Considering the long duration and distinct stages of the drying process, this unit employs a time-domain integration strategy, transforming the instantaneous changes obtained in step S3-1 into a whole-process cumulative effect from the start of drying to the current moment. In practical implementation, using the instantaneous change indicators generated in step S3-1 as input, firstly, the first-order difference scalar feature of weight from the start of drying to the current moment is integrated to generate a cumulative total moisture evaporation vector; secondly, the first-order difference scalar feature of humidity in the drying environment is accumulated over time to generate a cumulative environmental dehumidification potential vector; finally, the first-order difference vector feature of the full-band spectrum is accumulated in vector space to generate a cumulative spectral evolution vector. Finally, all vectors are sequentially concatenated to construct a process cumulative feature vector that can fully characterize the overall drying process.

[0048] Finally, the transient difference feature vector, dynamic rate feature vector, and process cumulative feature vector generated by processing S3-1 to S3-3 are concatenated into a time-series dynamic feature sequence.

[0049] S4. Standardization of Multidimensional Feature Space and Construction of Structure Tensor This step aims to fuse the multi-source synchronous time-series dataset obtained in step S2 and the time-series dynamic feature sequence generated in step S3 to construct a multi-dimensional input vector suitable for a two-stream network architecture model. This method is as follows... Figure 5 As shown, a split-encapsulation strategy is adopted to construct independent temporal tensors adapted to spectral flow, physical flow and global cumulative flow respectively, thereby laying the data foundation for feature decoupling and deep fusion of subsequent models.

[0050] S4-1 Unified Mapping of Heterogeneous Feature Space: First, considering the synchronous time-series data of the spectrum and its evolution characteristics, and taking into account the influence of sensor integration time and optical path reflectivity on light intensity data, the system reads the calibration saturation value and dark current noise floor value of the spectral sensor, and uses the maximum-minimum linear transformation formula to map the original spectral intensity and spectral difference data to... The system preserves the relative morphological characteristics between wavebands. Secondly, for time-series data with clearly defined physical limits, such as temperature, humidity, and weight, the system uses safety boundaries set by the sensor's hardware range and manufacturing process. The temperature range is 0–150 degrees Celsius, humidity is 0–100% relative humidity, and weight is 0–the initial weight of the dried fruit. The real-time acquired physical quantities are strictly compressed to [a specific range]. Within this range, a statistical boundary mapping strategy is adopted for the transient difference feature vector, kinetic rate feature vector, and process cumulative feature vector generated by S3. Based on historical batch big data statistics, the maximum and minimum values ​​are manually calibrated to determine the upper and lower bounds of the statistical distribution of each dynamic feature throughout the drying process. Using this statistical boundary as a benchmark, the first-order difference feature containing positive and negative fluctuations and the monotonically changing cumulative feature are uniformly mapped to... For the rate of change feature, 0.5 typically corresponds to zero change, while 0 and 1 correspond to the maximum negative and maximum positive changes, respectively, thus preserving the sign information of the change direction. Through the above processing, all heterogeneous features are standardized, and this step ultimately outputs a normalized full feature set, containing all the normalized features mentioned above.

[0051] S4-2 Feature Grouping and Recombination for a Two-Stream Architecture: First, high-dimensional spectral-related feature vectors are extracted from the normalized full feature set output by S4-1 to construct a spectral feature cluster. This cluster includes normalized spectral synchronization time-series data, full-band spectral first-order difference vector features, and spectral evolution trend feature vectors. Considering the high dimensionality of the spectral data and the differences in the physical distribution of various features, this unit uses Principal Component Analysis (PCA) to perform independent projection transformations on the aforementioned spectral synchronization time-series data, full-band spectral first-order difference vector features, and spectral evolution trend feature vectors. For each type of feature, based on the criterion that the cumulative contribution rate is greater than 95%, the top 10 principal component components after transformation are extracted. Finally, the three sets of independently dimensionality-reduced principal component feature vectors are concatenated along the feature dimension to reconstruct a simplified spectral state vector, which serves as the input to the subsequent neural network spectral stream. Next, a physical feature cluster is constructed. This cluster aggregates normalized synchronous time-series data on temperature, humidity, and weight, along with temperature and humidity environmental response gradient features, first-order differential scalar features on temperature, humidity, and weight, and instantaneous drying rate indicators. Based on this, a vector concatenation operation is performed, linking these scalar features in a predetermined order along the channel dimension to construct a physical state vector reflecting the current drying kinetic boundary conditions, which serves as the input to the long short-term memory neural network branch. Finally, a global cumulative feature cluster is constructed for the residual path. This cluster directly selects the normalized process cumulative feature vector. This unit does not perform complex nonlinear transformations on it, but rather maintains it as a low-dimensional state vector, aiming to provide the most direct physical constraints and global progress information for the network's residual direct connection path, preventing physical drift in the deep model during long-sequence inference.

[0052] S4-3 Timing Sliding Window and Multichannel Tensor Encapsulation Unit: This unit introduces a history backtracking sliding window mechanism and sets a uniform time step parameter. Regarding the current moment The previous time series data was extracted and encapsulated. First, for the spectral feature clusters, this unit uses the current time... At the cutoff point, a length of [length] is taken forward. Using continuous time-step data, a spectral temporal tensor is constructed, the dimension of which is defined as... ,in This refers to the batch size during model training or inference. The time window length, The dimension of the spectral state vector is defined. This tensor preserves the complete trajectory of the spectral bands evolving over time and is dedicated to the 1D convolutional neural network branch input to the network to extract spatiotemporal texture features. Secondly, for the physical feature clusters, this unit synchronously performs the same sliding window truncation operation to construct a physical temporal tensor, the dimension of which is defined as... ,in The dimension of the physical state vector is defined. This tensor reflects the historical trends of arid environment parameters and evaporation kinetics, and is specifically used in the long short-term memory (LSM) branch of the input network to capture long-term temporal dependencies. Furthermore, for the global cumulative feature cluster, this unit constructs a global cumulative temporal tensor, the dimension of which is defined as... ,in The dimension of the process cumulative feature vector is defined. This tensor is dedicated to the global physical constraint residual path input to the network and is injected into the network end through direct mapping to prevent physical prediction drift. Finally, the spectral time series tensor, physical time series tensor, and global cumulative time series tensor are concatenated along the feature dimension to form a multi-channel structured input set.

[0053] S5. Construction and Training of Moisture Content Prediction Model Traditional single-input networks struggle to effectively extract features from spectral data and physical environment data due to the significant differences in feature space distribution. To address this, this invention designs and constructs a dual-stream spatiotemporal attention fusion (DS-STAF) network. This network employs an architecture that dynamically fuses data features based on data type classification, such as... Figure 6 As shown, it includes a spectral feature extraction stream unit, a physical environment extraction stream unit, a multimodal gated fusion and physical residual path unit, and an output unit.

[0054] S5-1 Model Data Reception: Receives the multi-channel structured input set output from step S4 and routes it to the corresponding independent processing branch of the network. Three logically independent input ports are defined: the spectral input port receives the spectral time-series tensor; the physical input port receives the physical time-series tensor; and the cumulative input port receives the global cumulative time-series tensor.

[0055] S5-2 Spectral Feature Extraction Flow Unit: This branch is configured to capture the correlation between spectral bands and their evolution over time, processing spectral time-series tensors. First, the spectral time-series tensor enters a one-dimensional convolutional neural network layer, where the convolutional kernel slides along the feature dimension to extract morphological features of local bands. Second, the squeeze and excitation channel attention module (SE-Block) compresses the feature map using global average pooling, automatically learns the importance weights of different spectral bands using fully connected layers, and recalibrates the convolutional feature channels. Finally, the weighted feature sequence is input into a bidirectional long short-term memory network, capturing the evolution trajectory of spectral traits from both forward and backward time dimensions, outputting the spectral hidden layer state vector.

[0056] S5-3 Physical Environment Extraction Flow Unit: This branch is configured to capture the hysteretic effects of environmental parameters on moisture content changes, processing physical time-series tensors. First, the physical time-series tensor is input into a multi-layered stacked Long Short-Term Memory (LSTM) network to extract the temporal dependencies of temperature, humidity, and weight changes over a long time span. Second, a temporal attention mechanism module is connected after the LSM output layer. This module generates a temporal weight vector by calculating the correlation score between the hidden states at historical moments and the current state. After attention-weighted summation, the physical time-series feature vector is output.

[0057] S5-4 Multimodal Gated Fusion and Physical Residual Path Unit: This unit is the decision core of the network, responsible for fusing spectral and physical features and injecting physical constraints. Firstly, to address the issue of sensor reliability variations at different drying stages, a learnable adaptive gating unit is designed, calculating dynamic gating coefficients through a fully connected layer. The value ranges from 0 to 1, and is used to determine the state vector of the spectral hidden layer. With physical time series feature vector Weighted fusion is performed. The specific calculation formula is as follows: ; ; This represents the dynamic gating coefficient vector, whose dimension is the same as that of the eigenvector. and The dimensions are consistent. Each element in this vector takes values ​​ranging from... It is used to independently control the retention ratio of spectral features and physical features during fusion on a channel-by-channel basis. This represents the Sigmoid nonlinear activation function, used to map the result of a linear transformation to... Intervals, with gating weights in the form of generation probabilities. This represents the learnable weight matrix of the gated perception layer. The network continuously updates this matrix during training to learn how to decide the optimal fusion strategy based on the current state of the input features. This represents the spectral hidden layer state vector from step S5-2, which carries spectral band evolution information after channel attention weighting. This represents the physical time-series feature vector from step S5-3, which carries environmental lag and dynamic information after being weighted by time attention. This represents the learnable bias vector of the gated perception layer, used to adjust the activation threshold of the gated unit. This represents the multimodal fusion feature vector generated by the gating mechanism, which integrates complementary information from the spectral and physical flows. This represents the Hadamard product, which is the element-wise multiplication of two vectors. This operation utilizes a gating coefficient. Soft weighting of spectral features, utilizing The physical features are weighted complementaryly. Finally, to prevent the deep neural network from losing physical truth information in long-sequence nonlinear transformations, this unit establishes residual connections. The global cumulative temporal tensor is mapped through an independent fully connected layer and then added to the multimodal fusion feature vector in the form of residual connections to obtain the physical constraint fusion feature vector.

[0058] S5-5 Output Unit: This step defines the model's output objective and optimization process. In the output layer, a physical constraint fusion feature vector, which integrates multi-source features and physical residuals, is passed through a linear regression layer to output the predicted moisture content of the dried fruit at the current time. The loss function uses mean squared error as the main loss function to measure the difference between the predicted and actual moisture content; simultaneously, an L2 regularization term is added to constrain the network weights and prevent the model from overfitting on small sample data. This step ultimately outputs the trained and fixed DS-STAF network prediction model as the moisture content prediction model.

[0059] S6. Online inference and closed-loop feedback control based on real-time streaming data This step aims to deploy the trained moisture content prediction model to the edge computing terminal and cloud control system at the dried fruit production site, enabling continuous inference of the dried moisture content.

[0060] S6-1 Real-time Data Acquisition and Feature Incremental Update Unit: To address the low sampling frequency of the spectral sensor, the system employs a zero-order hold strategy, automatically using the values ​​from the previous moment to ensure the completeness of various heterogeneous data at the current moment. For feature calculation, this unit uses an efficient incremental iteration strategy. The system reads the historical states stored in memory and directly adds the instantaneous change at the current moment to the accumulated value from the previous moment, thereby quickly updating dynamic features such as drying rate and total cumulative moisture evaporation.

[0061] S6-2 Rolling Time Window Maintenance and Online Model Inference Unit: The system maintains a first-in-first-out (FIFO) rolling time window queue in memory for constructing input data for the deep learning model. Based on the tensor format defined in step S4, the system encapsulates the data within the rolling window into a multi-channel input structure that meets the network requirements in real time and inputs it into the pre-loaded deep prediction model. The model's internal dual-stream branches process spectral and physical features in parallel, outputting the predicted moisture content of the dried fruit at the current moment in real time.

[0062] S6-3 Drying Endpoint Judgment and Closed-Loop Feedback Control Unit: On the one hand, it performs dynamic parameter adjustment. If the predicted moisture content is still higher than the target value and the current drying rate is lower than the preset threshold, the system automatically issues a command to increase the hot air temperature or increase the speed of the exhaust fan to enhance the drying driving force. On the other hand, it performs endpoint judgment and shutdown. When the predicted moisture content is consistently lower than the target moisture content for a continuous period of time, the system determines that drying is complete, automatically triggers the shutdown procedure, shuts down the heater and fan, and sends a processing completion signal to the operator.

[0063] Experimental verification and analysis: To verify the effectiveness of the method of this invention in predicting the moisture content of dried fruit, this embodiment conducted multiple batches of experimental data collection on actual drying equipment. The experimental system consists of a heating fan, an adjustable airflow channel, a dried fruit sample tray, a multi-source sensor array, and data recording equipment.

[0064] The fruit samples selected for the experiment were two common dried products: mango and pineapple, two typical dried fruit varieties, were chosen as the experimental subjects. The drying process was set to constant temperature. Wind speed Each batch takes 4-8 hours to dry until the moisture content drops to 15%.

[0065] A total of 50 batches of dried fruit drying data were collected in the experiment, of which 30 batches were used as training data, 10 batches as validation data, and 10 batches as independent test data. The batch division ensured the generalization ability under different working conditions. The true moisture content label was measured by the standard drying method, with sampling every 20 minutes, and mapped to a 1-second time series through time-aligned interpolation to form the moisture content label sequence required for supervised learning.

[0066] Strictly follow step S2 to perform time-series synchronization and vectorization encapsulation of heterogeneous data, and generate spectral time-series tensors, physical time-series tensors, and global cumulative time-series tensors according to step S4, with a time sliding window. Set to 60 seconds. The true moisture content label is calibrated and interpolated using the drying method.

[0067] 1. Comparison of model setup and ablation experiment design: To comprehensively evaluate the advancement of the method of this invention, three sets of comparative experiments were designed, covering traditional methods, general deep learning baselines, and ablation variants of the key modules of this invention: (1) Baseline Methods Group M1: Weight-based linear regression: relies solely on real-time weight data to estimate moisture content, representing a traditional single-sensor solution in industry; M2: Standard spliced ​​long short-term memory network: directly splices spectral and physical features at the input layer, inputting into a single-stream multilayer long short-term memory network. Used to compare and verify the superiority of the dual-stream decoupled architecture of this invention in processing heterogeneous data.

[0068] (2) Ablation variant group of the present invention M3: Variant without physical residual path: Employs a two-stream network architecture but removes the global physical residual path from step S5-4. Used to verify the contribution of the cumulative features throughout the S3-3 process and the direct connection of physical residuals to prevent long sequence prediction drift; M4: Variant without attention mechanism: Employs a two-stream architecture but removes the SE-Block channel attention from step S5-2 and the time attention module from S5-3. Used to verify the necessity of key band screening and historical lag capture.

[0069] (3) Complete method of the present invention DS-STAF network: A complete model that includes two-stream extraction, dual attention mechanism and physical residual path.

[0070] 2. Evaluation indicators: Using root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (... The final error is used as an evaluation metric. The final error reflects the control accuracy at the end of drying (low moisture content stage) and is crucial for the closed-loop control in S6.

[0071] Table 1. Prediction performance of different methods on the test set.

[0072] Based on the experimental data in Table 1, an in-depth comparative analysis of the prediction performance of different models was conducted to verify the effectiveness of the DS-STAF network architecture and its key components proposed in this invention.

[0073] (1) Analysis of the adaptive advantages of the dual-stream decoupling architecture for heterogeneous features: Compared with the standard spliced ​​long short-term memory network model, the RMSE index of the method of this invention decreased from 2.68% to 1.25%, a reduction of 53.3%. This significant difference is mainly attributed to the different feature processing methods. The standard spliced ​​long short-term memory network adopts a single-stream splicing strategy, which forces the same set of network weights to process high-dimensional dense spectral vectors and low-dimensional sparse physical scalars simultaneously. Due to the huge difference in the feature space distribution of the two, gradient competition occurs during training, and the model finds it difficult to find a balance between the spectrum and the physical quantity. In contrast, the dual-stream architecture adopted in this invention achieves feature decoupling. Experimental results show that this split processing mechanism effectively eliminates feature conflicts of multi-source heterogeneous data, which is the basis for improving the convergence accuracy of the model.

[0074] (2) The effect of physical residual paths on the suppression of long-sequence prediction drift: By comparing the variant without physical residual paths with the complete model, a key phenomenon can be observed: although the RMSE (2.15%) of the model without residual paths is within an acceptable range, its final error is as high as 2.68%, which is much lower than the 0.85% of the present invention. During the drying process of 4-8 hours, the pure data-driven deep neural network is prone to forgetting, that is, as the sequence iterates, the model's perception of the initial state gradually becomes blurred, resulting in cumulative drift of the predicted value. The model without residual paths, due to the lack of the cumulative features defined in S3-3 and the direct connection path in S5-4, cannot maintain the memory of the physical truth value of the total water loss. The present invention introduces physical residual paths to directly inject the total cumulative water evaporation into the output layer, which is essentially applying a physical hard constraint based on the law of conservation of mass to the neural network. Data shows that this design effectively eliminates the cumulative error of long-sequence prediction and controls the final error to within 1%.

[0075] (3) Contribution of the attention mechanism to the capture of key information: By comparing the variant without the attention mechanism with the complete model, the model RMSE was further optimized from 1.86% to 1.25%, contributing approximately 0.6% to the accuracy gain. The original spectral data contains a large number of background noise bands unrelated to moisture. Without a filtering mechanism, this noise can interfere with the model's judgment. In this invention, the SE-Block channel attention automatically filters out near-infrared feature bands sensitive to moisture, suppressing redundant information; while the time attention mechanism accurately captures the hysteresis effect after changes in environmental parameters. The combination of these two mechanisms significantly enhances the model's sensitivity to changes in spectral properties and environmental disturbances.

[0076] Visual analysis, such as Figure 7 , Figure 8 as well as Figure 9 As shown.

[0077] from Figure 7As can be observed, in the initial stage of drying, all models can track the true curve well; however, as the drying time progresses, the blue dashed line gradually deviates upward from the true value, exhibiting a significant long-sequence cumulative drift phenomenon, and its error curve in the subgraph also increases linearly with time. This indicates that pure deep neural networks are prone to losing memory of the physical truth value of total water loss during long-sequence iterations. In contrast, the red solid line closely follows the true curve throughout the entire drying cycle, and its error in the subgraph remains low. This comparison fully demonstrates that the cumulative features constructed in step S3-3, after being injected into the network through the physical residual path in step S5-4, successfully force the model to follow the law of mass conservation, effectively suppressing non-physical drift and ensuring the accuracy of the drying endpoint determination. Furthermore, the hysteresis phenomenon of the green dotted line in the later stage of drying due to decreased sensitivity reflects the perception advantage of the multi-source fusion architecture of this invention in the low moisture content stage.

[0078] from Figure 8 As can be observed, during the constant-rate drying stage in the early stage, due to the rapid evaporation of moisture, the signal-to-noise ratio of the weight and humidity sensors is extremely high. The model assigns a high weight to the physical flow to quickly track changes in moisture content using dynamic characteristics. However, as drying enters the deceleration stage, moisture mainly exists in the form of bound water, and the rate of weight change slows down, causing the physical signal to gradually become less sensitive. At this time, the browning and shrinkage reactions on the surface of the dried fruit intensify. The model sensitively detects this state switch through the gating mechanism described in S5-4, significantly increasing the weight of the spectral flow and instead using chemical characteristics to dominate the prediction. This dynamic weight evolution curve proves that the model of this invention has cross-stage adaptive sensing capabilities, effectively solving the technical problem of traditional single sensors failing due to decreased sensitivity in the later stages of drying.

[0079] like Figure 9 As shown, the heatmap exhibits vertical deep red bands spanning the entire time frame at 970nm and 760nm. This indicates that the channel attention mechanism introduced in step S5-2 successfully learned the spectral physical features and automatically identified these two most decisive physical feature bands for moisture content prediction. Simultaneously, near wavelength 480nm, the attention color shows a gradual change from blue to red as drying time progresses. This indicates that the model learned the implicit correlation between surface browning and decreased moisture content in the later stages of drying, thus utilizing chemical properties to assist in prediction during periods of weak weight signals. The majority of the remaining areas in the figure are in cool tones, demonstrating that the model effectively suppressed a large amount of background spectral noise and redundant information unrelated to moisture, confirming the excellent feature selection capability and physical interpretability of the algorithm of this invention.

[0080] Conclusion: Ablation experiments fully demonstrate that the network of this invention is not a simple stacking of modules, but an organic whole. The two-stream architecture solves the problem of heterogeneous data fusion, the attention mechanism improves the feature signal-to-noise ratio, and the physical residual path ensures the physical conservation of long sequence prediction. The synergistic effect of these three elements enables accurate prediction of the sub-100% moisture content of dried fruit.

[0081] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0082] While the specific embodiments of the present invention have been described above, they are not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for predicting the moisture content of dried fruit based on multi-sensor data, characterized in that, Includes the following steps: S1 collects environmental constraint data and material response data during the fruit drying process, including temperature time-series data, humidity time-series data, weight time-series data, and spectral time-series data; S2, preprocesses the data obtained from S1 to obtain a multi-source synchronous time series dataset; S3. Calculate the first-order difference of the multi-source synchronous time-series dataset to capture transient pulse changes and obtain the transient difference feature vector. Then, extract the weight decrease slope and spectral evolution slope to construct kinetic rate features, characterize the instantaneous evaporation efficiency, and obtain the kinetic rate feature vector. At the same time, adopt the time-domain integration strategy to calculate the cumulative total water evaporation, heat and moisture load history, and spectral evolution distance to construct the full-process cumulative features and obtain the process cumulative feature vector. S4. Perform independent channel normalization on the feature data obtained from S2 and S3 to obtain the normalized full feature set; divide the normalized full feature set into spectral feature cluster, physical feature cluster and global cumulative feature cluster; Using a historical backtracking sliding window mechanism, the three types of feature clusters are encapsulated into independent three-dimensional temporal tensors, including spectral temporal tensors, physical temporal tensors, and global cumulative temporal tensors. S5 takes the three time-series tensors obtained in S4 and inputs them into the trained moisture content prediction model that can decouple the spectrum and physical characteristics, and outputs the predicted moisture content of the dried fruit.

2. The method for predicting the moisture content of dried fruit based on multi-sensor data as described in claim 1, characterized in that: The moisture content prediction model is a dual-stream spatiotemporal attention fusion network architecture, comprising two main processing streams: the spectral stream uses a 1D convolutional neural network and channel attention mechanism to extract band evolution features; the physical stream uses a long short-term memory network and temporal attention mechanism to capture the effects of environmental lag; the spectral temporal tensor and the physical temporal tensor are dynamically weighted and fused through an adaptive gating unit; in addition, a physical residual path is introduced, and the global cumulative temporal tensor generated by S4 is directly injected into the output layer to constrain the prediction results with the physical conservation law.

3. The method for predicting the moisture content of dried fruit based on multi-sensor data as described in claim 1, characterized in that: The transient difference feature vector is obtained in S3, and the specific process is as follows: For synchronized time-series data of temperature, humidity, weight, and spectrum, a synchronized state vector is constructed, and a first-order difference operation is performed to obtain the first-order difference scalar features of temperature, humidity, weight, and the full-band spectrum. Instantaneous change indices of each physical quantity are calculated by subtracting the state vector of the current time step from the state vector of the previous time step. Finally, a feature vectorization integration operation is performed, and the calculation results are time-aligned based on a unified timestamp index. Then, the first-order difference scalar features of temperature, humidity, and weight are sequentially concatenated with the flattened first-order difference vector features of the full-band spectrum, and the cumulative change features are also included in the combination to construct a transient differential feature vector characterizing the drying dynamics at the current moment.

4. The method for predicting the moisture content of dried fruit based on multi-sensor data as described in claim 1, characterized in that: The dynamic velocity feature vector is obtained in S3, and the specific process is as follows: First, based on the multi-source synchronous time-series dataset within the drying chamber, the first derivatives of the synchronous temperature and humidity time-series data with respect to time are calculated to construct temperature and humidity environmental response gradient features reflecting transient environmental changes. Second, for the synchronous weight time-series data, a linear trend fitting method based on a sliding window is adopted. By performing linear regression analysis within a preset time window of length T and extracting the slope, an instantaneous drying rate index that can overcome minor mechanical vibrations and accurately reflect the rate of moisture vaporization and escape is constructed. Finally, for the synchronous spectral time-series data, the evolution rate of intensity of each band on the time axis is calculated to construct a spectral evolution trend feature vector. Finally, a feature vectorization integration operation is performed. Based on a unified timestamp index, the calculated rates and gradient indices are time-series aligned. Then, the scalar-form temperature and humidity environmental response gradient features, instantaneous drying rate indices, and flattened spectral evolution trend feature vectors are sequentially concatenated to construct a dynamic rate feature vector that can reflect the transition of the drying stage and the intensity of internal moisture migration.

5. The method for predicting the moisture content of dried fruit based on multi-sensor data as described in claim 3, characterized in that: The process of obtaining the cumulative feature vector in S3 is as follows: First, the weight first-order difference scalar feature from the start of drying to the current time is integrated to generate a cumulative total moisture evaporation vector. Second, the humidity first-order difference scalar feature in the drying environment is accumulated over time to generate a cumulative environmental dehumidification potential vector. Finally, the full-band spectral first-order difference vector feature is accumulated in vector space to generate a cumulative spectral evolution vector. Finally, a feature vectorization integration operation is performed. Based on a unified timestamp index, the cumulative indicators are time-series aligned. Then, the cumulative total moisture evaporation vector, the cumulative environmental dehumidification potential vector, and the flattened cumulative spectral evolution vector are sequentially concatenated to construct a process cumulative feature vector that can fully characterize the overall drying process.

6. The method for predicting the moisture content of dried fruit based on multi-sensor data as described in claim 1, characterized in that: In step S4, the feature data obtained from S2 and S3 are normalized through independent channels to obtain a normalized full feature set. The specific process is as follows: First, based on the spectral synchronization time-series data and its evolution characteristics, the calibration saturation value and dark current noise floor value of the spectral sensor are read. Then, using the maximum-minimum linear transformation formula, the original spectral intensity and spectral difference data are mapped to... The interval preserves the relative morphological characteristics between bands; Secondly, for data with clearly defined physical limits, the real-time acquired physical quantities are strictly compressed to a minimum based on the sensor's hardware range and the safety boundaries set by the manufacturing process. Within the specified range; for the transient difference feature vector, kinetic rate feature vector, and process cumulative feature vector generated by S3, a statistical boundary mapping strategy is adopted. Based on historical batch big data statistics, the maximum and minimum values ​​are manually calibrated to determine the upper and lower bounds of the statistical distribution of each dynamic feature throughout the drying process. Using the statistical boundaries as a benchmark, the first-order difference features containing positive and negative fluctuations and the monotonically changing cumulative features are uniformly mapped to... For the rate of change feature, 0.5 corresponds to zero change, and 0 and 1 correspond to the maximum negative change and the maximum positive change, respectively. All heterogeneous features are converted into a standardized form, and the normalized full feature set is output.

7. The method for predicting the moisture content of dried fruit based on multi-sensor data as described in claim 6, characterized in that: In step S4, the normalized full feature set is divided into spectral feature clusters, physical feature clusters, and global cumulative feature clusters. The specific process is as follows: First, high-dimensional spectral-related feature vectors are extracted from the normalized full feature set to construct a spectral feature cluster, which includes normalized spectral synchronization time series data, full-band spectral first-order difference vector features, and spectral evolution trend feature vectors. Principal component analysis was used to perform independent projection transformations on the spectral synchronization time series data, the first-order difference vector features of the full-band spectrum, and the spectral evolution trend feature vector. For each type of feature, the top 10 principal component components after transformation were extracted based on the criterion that the cumulative contribution rate was greater than 95%. Finally, the three sets of principal component feature vectors that had undergone independent dimensionality reduction were concatenated along the feature dimension to reconstruct a simplified spectral state vector. Secondly, a physical feature cluster is constructed, which aggregates normalized synchronous time-series data of temperature, humidity, and weight, as well as temperature and humidity environmental response gradient features, first-order differential scalar features of temperature, first-order differential scalar features of humidity, first-order differential scalar features of weight, and instantaneous drying rate index. A vector concatenation operation is performed to concatenate the scalar features in a preset order along the channel dimension to construct a physical state vector that can reflect the current drying dynamic boundary conditions. Finally, for the residual path, a global cumulative feature cluster is constructed, and the normalized process cumulative feature vector is directly selected.

8. The method for predicting the moisture content of dried fruit based on multi-sensor data as described in claim 7, characterized in that: In step S4, the historical backtracking sliding window mechanism is used to encapsulate the three types of feature clusters into independent three-dimensional temporal tensors. The specific process is as follows: First, for the spectral feature clusters, at the current time... At the cutoff point, a length of [length] is taken forward. Using continuous time-step data, a spectral temporal tensor is constructed, and the dimension of the tensor is defined as follows: ,in This refers to the batch size during model training or inference. The time window length, First, let's define the dimension of the spectral state vector. Second, for each physical feature cluster, perform the same sliding window truncation operation simultaneously to construct a physical temporal tensor. The dimension of this tensor is defined as... ,in Let be the dimension of the physical state vector; then, for the global cumulative feature cluster, construct a global cumulative temporal tensor, the dimension of which is defined as... ,in Accumulate feature vectors for the process.

9. A method for predicting the moisture content of dried fruit based on multi-sensor data as described in claim 1 or 2, characterized in that: The moisture content prediction model specifically includes a spectral feature extraction stream unit, a physical environment extraction stream unit, a multimodal gated fusion and physical residual path unit, and an output unit; The spectral feature extraction unit feeds the spectral temporal tensor into a one-dimensional convolutional neural network layer, where the convolutional kernel slides along the feature dimension to extract morphological features of local bands. Next, the squeeze and activation channel attention module compresses the feature map using global average pooling, automatically learns the importance weights of different spectral bands using a fully connected layer, and recalibrates the convolutional feature channels. Finally, the weighted feature sequence is input into a bidirectional long short-term memory network, capturing the evolution trajectory of spectral traits from both forward and reverse time dimensions, and outputting the spectral hidden layer state vector. The multimodal gating fusion unit inputs the physical time series tensor into a multi-layer stacked long short-term memory network to extract the temporal dependencies of temperature, humidity and weight changes over a long time span. Secondly, a time attention mechanism module is connected after the output layer of the Long Short-Term Memory Network; by calculating the correlation score between the hidden state at the historical moment and the state at the current moment, a time weight vector is generated; after attention weighting and summing, a physical time sequence feature vector is output. The multimodal gated fusion and physical residual path unit is based on a learnable gated fusion unit and calculates dynamic gating coefficients through a fully connected layer. The value ranges from 0 to 1, and the spectral hidden layer state vector and physical temporal feature vector are weighted and fused accordingly; the gating coefficient is then used. Soft weighting of spectral features, utilizing The physical features are complementary and weighted; residual connections are established, and the global cumulative temporal tensor is mapped through an independent fully connected layer and then added to the multimodal fusion feature vector in the form of residual connections to obtain the physical constraint fusion feature vector; The output unit outputs the predicted moisture content of the dried fruit at the current moment through a linear regression layer.

Images