Processing residual time prediction method based on time sequence multi-dimensional feature fusion
By constructing a time-series prediction model based on multi-dimensional time-series data, the problem of electric ovens being unable to provide real-time perception of food doneness was solved, enabling accurate prediction of the remaining processing time of food and improving the intelligent user experience of the oven.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ZHEJIANG UNIV
- Filing Date
- 2026-05-08
- Publication Date
- 2026-07-31
AI Technical Summary
Commercial and household electric ovens cannot provide users with real-time feedback on changes in the doneness of ingredients, resulting in a lack of intelligent sensing capabilities. This is especially problematic for users with limited cooking experience, who may find it difficult to accurately predict the steaming and baking time and the end time of the food.
By collecting multi-dimensional time-series data during the oven's steaming and baking process, a time-series prediction model is constructed, including a gated loop unit network layer, a multi-task shared parameter layer, a parallel load identification branch, and a remaining time prediction branch, to achieve adaptive and accurate prediction of food load and remaining processing time.
It achieves adaptive and precise perception of the cooking process under different ingredient quantities, providing users with intuitive and real-time feedback on steaming and baking progress, thus optimizing the intelligent user experience of the oven.
Smart Images
Figure CN122490423A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of processing time prediction technology, and in particular to a method for predicting remaining processing time based on temporal multidimensional feature fusion. Background Technology
[0002] Commercial and household electric ovens require users to set the heating temperature and operating time based on the amount of food being steamed or baked and their own cooking experience. The oven simply executes the program according to the user's settings and cannot provide real-time feedback on the changes in the food's doneness. For users lacking cooking experience, it is difficult to accurately estimate the steaming or baking time and the end time of the process. The oven's own PID control system also cannot provide accurate estimates of the remaining processing time, resulting in a lack of intelligent sensing in the user experience.
[0003] By collecting multi-dimensional time-series data during the steaming and baking process of ingredients to train a neural network model, it is possible to achieve intelligent perception of changes in the load and maturity of ingredients. However, the edge-side MCUs used in home appliances are limited by cost, and their storage space and computing power are difficult to meet the computational needs of most neural network models, making it impossible to achieve localized deployment and real-time inference of maturity prediction models. Summary of the Invention
[0004] The purpose of this application is to provide a method for predicting the remaining processing time based on temporal multidimensional feature fusion, which can achieve adaptive and accurate prediction of the remaining processing time of food ingredients under different loads.
[0005] To achieve the above objectives, this application provides the following solution: A method for predicting remaining processing time based on temporal multidimensional feature fusion includes the following steps: The system acquires multidimensional raw time-series data during the oven's steaming and baking process, and performs time-scale alignment on the multidimensional raw time-series data based on a unified time reference to obtain aligned time-series data.
[0006] A sliding window with a preset step size and window length is used to serialize and segment the aligned time-series data, thereby constructing a multidimensional feature vector sequence.
[0007] The multidimensional feature vector sequence is input into the trained time-series prediction model, which outputs the classification result of the load interval corresponding to the current steamed / roasted food and the predicted value of the remaining processing time. The time-series prediction model includes a gated recurrent unit network layer, a multi-task shared parameter layer, and a parallel load identification branch and a remaining time prediction branch connected to the output of the multi-task shared parameter layer. The gated recurrent unit network layer is used to extract high-dimensional time-series feature codes reflecting the heat and mass exchange state of the food from the input multidimensional feature vector sequence. The multi-task shared parameter layer is used to generate a shared feature vector that contains both food load information and remaining processing time information based on the high-dimensional time-series feature code. The load identification branch is used to generate the load interval classification result based on the shared feature vector. The remaining time prediction branch is used to generate the predicted value of the remaining processing time based on the shared feature vector.
[0008] Optionally, the multidimensional raw time-series data includes at least the temperature sequence, humidity sequence, valve opening and closing logic signal, and cooking time data corresponding to the steaming and baking process inside the oven cavity. Acquiring the multidimensional raw time-series data during the oven's steaming and baking process specifically involves: acquiring the multidimensional raw time-series data in real time during the oven's steaming and baking process at a preset sampling frequency; wherein, the temperature sequence is synchronously acquired by two temperature sensors respectively installed on the rear and side of the oven cavity, with the measurement scale values of the two temperature sensors consistent with the sampling frequency; the humidity sequence is acquired by a humidity sensor built into the oven cavity; the valve opening and closing logic signal is directly acquired by the oven controller; and the cooking time data is the relative time data of the current steaming and baking process.
[0009] Optionally, a sliding window with a preset step size and window length is used to serialize and segment the aligned time-series data to construct a multidimensional feature vector sequence, specifically including the following steps: Within each sliding window with a preset step size and window length, temperature change rate features and humidity change rate features are calculated and added to expand the feature dimensions. Specifically, the two change rates are 0 at the first moment within the sliding window, and the change rates at the remaining moments are the differences between the feature values at the corresponding moments and the feature values at the first moment within the sliding window.
[0010] The data from consecutive time periods segmented by the sliding window are combined into a multidimensional feature vector sequence with temporal correlation.
[0011] Optionally, the gated loop unit network layer models the long-term and short-term dependencies between temperature fluctuations, humidity fluctuations, valve state switching and cooking time data through built-in reset gates and update gates, extracts high-dimensional time-series feature codes that reflect the heat and mass exchange state of ingredients, and outputs the high-dimensional time-series feature codes to the multi-task shared parameter layer.
[0012] Optionally, the multi-task shared parameter layer receives the high-dimensional temporal feature encoding output by the gated recurrent unit network layer, learns the thermodynamic performance characteristics of different steamed and baked ingredients under different loads, generates a shared feature vector that simultaneously contains information on the ingredient load and the remaining processing time, and outputs the shared feature vector to the load identification branch and the remaining time prediction branch, respectively.
[0013] Optionally, the load identification branch includes a fully connected layer and a Softmax classifier connected in sequence, receives the shared feature vector output by the multi-task shared parameter layer, outputs the probability distribution of the load level corresponding to the current steamed / baked food, and generates a load interval classification result.
[0014] Optionally, the remaining time prediction branch includes a regression processing layer that receives the shared feature vector output by the multi-task shared parameter layer, calculates and outputs the predicted value of the remaining processing time for the current steamed / baked ingredients.
[0015] Optionally, before inputting the multi-dimensional feature vector sequence into the trained time-series prediction model and outputting the classification result of the load interval corresponding to the current steamed / baked ingredients and the predicted value of the remaining processing time, the method further includes: the training process of the time-series prediction model, specifically including the following steps: Training data for multiple steaming and baking processes of various typical ingredients were collected, and the ingredient load and remaining processing time labels for each steaming and baking process were labeled. The remaining processing time label was calculated from the stop time of the oven's corresponding ingredient steaming and baking program and the timestamp of the steaming and baking process.
[0016] The multidimensional time-series data used for training is constructed into a multidimensional feature vector sequence, which serves as the input to the time-series prediction model. A joint loss function is defined as a linear combination of classification cross-entropy loss and regression mean squared error loss. The classification cross-entropy loss is the result of cross-entropy calculation between the output of the load identification branch of the time-series prediction model and the food load label. The regression mean squared error loss is the result of mean squared error calculation between the output of the remaining time prediction branch of the time-series prediction model and the remaining processing time label.
[0017] The joint loss function value is calculated based on the output of the time-series prediction model, the corresponding food load label, and the remaining processing time label. The network parameters of the time-series prediction model are updated through backpropagation to obtain the trained time-series prediction model.
[0018] Optionally, after obtaining the trained time series prediction model, the process also includes: a lightweight processing procedure for the time series prediction model, specifically including the following steps: A structured pruning strategy is adopted to evaluate the importance of the weight channels of neurons and gated recurrent units in each layer of the trained temporal prediction model. Neurons and structural blocks with contributions below a preset threshold are pruned while preserving the density of the weight matrix, resulting in a lightweight model after pruning.
[0019] The trained time-series prediction model is used as the teacher network, and the pruned lightweight model is used as the student network. The student network is retrained using knowledge distillation based on the soft labels output by the teacher network and the feature information of the intermediate layer, so as to obtain the time-series prediction model with lightweight processing.
[0020] According to the specific embodiments provided in this application, the following technical effects are disclosed: This application provides a method for predicting remaining processing time based on temporal multidimensional feature fusion. This method acquires multidimensional raw temporal data during the oven's steaming and baking process, and aligns the raw data with a unified time reference to obtain aligned temporal data. This eliminates time discrepancies from different data sources and comprehensively covers the oven's internal temperature and humidity environment, actuator actions, and cooking progress. It overcomes the limitations of traditional methods that treat the oven as a closed system, enabling the model to perceive system state changes and external interventions during steaming and baking, thus improving prediction robustness. A sliding window with preset step size and window length is used to serialize and segment the aligned temporal data to construct a multidimensional feature vector sequence. This transforms discrete state data at a single moment into a sequence feature with strong temporal correlation, fully characterizing the dynamic evolution trend of oven thermodynamic environment fluctuations and food heat and mass exchange. This allows the model to accurately capture the current steaming and baking state and deeply explore historical change patterns, providing a high-quality input foundation for subsequent deep temporal feature extraction. The multidimensional feature vector sequence is input into the trained temporal prediction model, which outputs the classification result of the current steaming and baking food's load interval and the predicted value of remaining processing time. The temporal prediction model uses a... The architecture consists of a gated loop unit network layer, a multi-task shared parameter layer, and parallel load identification and remaining time prediction branches connected to the output of the multi-task shared parameter layer. The gated loop unit network layer extracts high-dimensional temporal feature codes reflecting the heat and mass transfer state of the ingredients from the input multi-dimensional feature vector sequence. This effectively models the long-term and short-term dependencies between temperature and humidity fluctuations, valve status switching, and cooking progress, accurately capturing the implicit nonlinear laws of ingredient maturity evolution and solving the problem that traditional threshold judgment methods cannot adapt to complex heat and mass transfer processes. The multi-task shared parameter layer generates a shared feature vector containing both ingredient load information and remaining processing time information based on the high-dimensional temporal feature code, enabling joint learning of load and maturity features and providing a unified high-value feature foundation for dual-task decoupled prediction. Finally, the parallel load identification and remaining time prediction branches generate load interval classification results and remaining processing time prediction values based on the shared feature vectors, respectively. Using ingredient load features to guide remaining time prediction, this achieves adaptive and accurate perception of the maturation process under different ingredient quantities, providing users with intuitive and real-time feedback on steaming and baking progress and effectively optimizing the intelligent user experience of the oven. Attached Figure Description
[0021] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0022] Figure 1 This is a flowchart illustrating a method for predicting remaining processing time based on temporal multidimensional feature fusion, provided as an embodiment of this application.
[0023] Figure 2 This is a flowchart of step A2 in a processing remaining time prediction method based on temporal multidimensional feature fusion provided in an embodiment of this application.
[0024] Figure 3 This is a flowchart illustrating the training process of a temporal prediction model in a processing remaining time prediction method based on temporal multidimensional feature fusion, provided as an embodiment of this application.
[0025] Figure 4 This is a flowchart illustrating the lightweight processing of the temporal prediction model in a processing remaining time prediction method based on temporal multidimensional feature fusion, provided as an embodiment of this application. Detailed Implementation
[0026] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0027] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0028] This application provides a method for predicting remaining processing time based on temporal multidimensional feature fusion. In an exemplary embodiment, such as... Figure 1 As shown, it includes the following steps: A1. Obtain multidimensional raw time-series data during the oven's steaming and baking process, and perform time-scale alignment on the multidimensional raw time-series data based on a unified time reference to obtain aligned time-series data.
[0029] Specifically, the multidimensional raw time-series data includes at least the temperature sequence, humidity sequence, valve opening and closing logic signal, and cooking time data corresponding to the steaming and baking process within the oven cavity. Acquiring the multidimensional raw time-series data during the oven's steaming and baking process involves: real-time acquisition of the multidimensional raw time-series data during the oven's steaming and baking process at a preset sampling frequency; wherein, the temperature sequence is synchronously acquired by two temperature sensors respectively installed on the rear and side of the oven cavity, with the measurement scale values of the two temperature sensors consistent with the sampling frequency; the humidity sequence is acquired by a humidity sensor built into the oven cavity; the valve opening and closing logic signal is directly acquired by the oven controller; and the cooking time data is the relative time data of the current steaming and baking process.
[0030] In this embodiment, the multidimensional raw time-series data includes the temperature sequence, humidity sequence, valve opening and closing logic signal, and cooking time data corresponding to the steaming and baking process within the oven cavity. In this embodiment, real-time data is collected for the steaming and baking process of six typical ingredients: cookies, egg tarts, chicken wings, meat skewers, sweet potatoes, and lemon slices, using a preset sampling frequency of 1Hz. The temperature sequence is synchronously collected by two temperature sensors located on the rear and side of the oven cavity, respectively. The two temperature sensors have identical measurement scale values, sampling frequencies, and other parameters, enabling synchronous temperature acquisition at different locations within the oven cavity. The humidity sequence is collected by a built-in humidity sensor within the oven cavity, reflecting the dynamic changes in moisture evaporation during the steaming and baking process. The valve opening and closing logic signal is directly acquired by the oven controller. The valve opening and closing controls the airflow between the oven cavity and the outside, allowing the release of evaporating water vapor during the steaming and baking process. This signal reflects the energy exchange status between the oven system and the outside environment. The cooking time data is the relative timing data after the start of the current steaming and baking process, used to characterize the time progress of the steaming and baking process.
[0031] The physical parameters and logic state signals from the different sources mentioned above are time-aligned based on a unified time reference to ensure that the feature points at each moment contain data from all the above dimensions, thereby eliminating time deviations from different data sources and ensuring the time consistency of time series data.
[0032] As an optional implementation, the multidimensional raw time-series data collected in step A1 can also incorporate the oven's back fan speed, heating plate temperature, and oven heating power as the original feature dimensions of the feature vector, further enhancing the model's ability to perceive the oven's working status.
[0033] A2. Using a sliding window with a preset step size and window length, the aligned time-series data is serialized and segmented to construct a multi-dimensional feature vector sequence. For example... Figure 2 As shown, step A2 specifically includes the following steps: A21. Within each sliding window with a preset step size and window length, calculate and add temperature change rate features and humidity change rate features to expand the feature dimensions. Specifically, the two change rates are 0 at the first moment within the sliding window, and the change rates at other moments are the differences between the corresponding feature values and the feature values at the first moment within the sliding window. The change rate features visually represent the dynamic trends of temperature and humidity changes within the oven cavity.
[0034] A22. Combine the data from consecutive time periods after the sliding window is segmented into a multi-dimensional feature vector sequence with temporal correlation. By using a sliding window, the complete long-term time-series data collected during a single steaming and baking process is divided into multiple short window sequences of fixed length. The data within these consecutive time periods are combined into a feature vector sequence with temporal correlation, serving as the feature input characterizing the fluctuations in the oven's thermodynamic environment and the evolution of food ripening trends. This allows the model to not only acquire the current state data but also capture the changing trends over historical time periods.
[0035] As an optional implementation, the expanded feature dimension in step A2 can also use the second-order difference of temperature sensor data and the second-order difference of humidity sensor data to further enhance the ability to characterize the trend of temperature and humidity changes.
[0036] A3. Input the multidimensional feature vector sequence into the trained time-series prediction model, and output the classification result of the load interval corresponding to the current steamed and baked ingredients and the predicted value of the remaining processing time; wherein, the time-series prediction model includes a gated recurrent unit (GRU) network layer, a multi-task shared parameter layer, and a parallel load identification branch and a remaining time prediction branch connected to the output of the multi-task shared parameter layer in sequence.
[0037] The gated loop unit network layer, through built-in reset and update gates, models the long-term and short-term dependencies between temperature fluctuations, humidity fluctuations, valve status switching, and cooking time data. It extracts high-dimensional temporal feature codes reflecting the heat and mass exchange state of the ingredients and outputs these codes to the multi-task shared parameter layer. The steaming and baking process of ingredients is a non-linear heat and mass transfer process. The gated loop unit can effectively process long-term time-series data, avoid the gradient vanishing problem, and accurately capture the implicit patterns of ripeness evolution over time.
[0038] The multi-task shared parameter layer receives high-dimensional temporal feature codes output from the gated recurrent unit network layer, learns the thermodynamic characteristics of different steamed and baked ingredients under different loads, generates a shared feature vector that simultaneously contains information on ingredient load and remaining processing time, and outputs the shared feature vector to the load recognition branch and the remaining time prediction branch, respectively. The multi-task shared parameter layer provides a unified feature foundation for the decoupled prediction of the subsequent two tasks.
[0039] The load identification branch includes a fully connected layer and a Softmax classifier connected in sequence. It receives the shared feature vector output by the multi-task shared parameter layer, outputs the probability distribution of the load level corresponding to the current steamed / baked food, and generates a load interval classification result. In this embodiment, the load of cookies, egg tarts, meat skewers, and chicken wings is divided into intervals by quantity (pieces, pieces, skewers), while the load of sweet potatoes and lemon slices is divided into intervals by the number of plates.
[0040] The remaining time prediction branch includes a regression processing layer that receives the shared feature vector output from the multi-task shared parameter layer, calculates and outputs the predicted remaining processing time (RUL) for the current steamed / roasted ingredients. Since the shared feature vector already contains the ingredient load information, the remaining time prediction can adaptively adjust based on the load characteristics, achieving accurate perception of maturity at different ingredient load levels.
[0041] As an alternative implementation, the gated recurrent unit network layer in the time series prediction model can be replaced with a long short-term memory network (LSTM) or a recurrent neural network (RNN). At the same time, a time attention mechanism can be introduced after the gated recurrent unit network layer to assign different weights to features at different times, thereby enhancing the model's ability to capture features during the key steaming and baking stages.
[0042] The trained time-series prediction model is deployed in the edge-side MCU of the oven. The edge-side MCU collects real-time data, constructs features, and performs localized real-time inference during the steaming and baking process. As the steaming and baking process progresses, it outputs the current food load level and remaining processing time in real time, providing users with visual feedback on changes in ripeness and optimizing the user experience. It should be noted that the model's prediction results are not used for the automatic control of the oven's steaming and baking program; they are only displayed to the user as real-time feedback information during the steaming and baking process.
[0043] In one exemplary embodiment, the method further includes a training process for a time-series prediction model, such as... Figure 3 As shown, the specific steps include: B1. Collect multidimensional time-series training data for multiple steaming and baking processes of various typical ingredients, and label the ingredient load and remaining processing time for each steaming and baking process. The remaining processing time label is calculated from the stop time of the oven's corresponding ingredient steaming and baking program and the timestamp of the steaming and baking process.
[0044] For six typical ingredients—cookies, egg tarts, chicken wings, meat skewers, sweet potatoes, and lemon slices—multi-dimensional time-series data were collected from multiple steaming and baking processes. During the data collection, the load of each ingredient was recorded as a load label. Simultaneously, the difference between the oven's built-in stop time for the corresponding ingredient's steaming and baking program and the timestamps recorded during the steaming and baking process was calculated to obtain the remaining processing time label for each timestamp. The stop time for the oven's steaming and baking program is determined by the oven based on humidity sensor data and preset program logic, serving as the termination time of the steaming and baking process.
[0045] B2. Construct a multidimensional feature vector sequence from the multidimensional time-series data used for training, and use it as the input to the time-series prediction model. Define a joint loss function that is a linear combination of classification cross-entropy loss and regression mean squared error loss. Classification cross-entropy loss is the result of cross-entropy calculation between the output of the load identification branch of the time-series prediction model and the food load label. Regression mean squared error loss is the result of mean squared error calculation between the output of the remaining time prediction branch of the time-series prediction model and the remaining processing time label.
[0046] B3. Calculate the joint loss function value based on the output of the time series prediction model, the corresponding food load label, and the remaining processing time label. Update the network parameters of the time series prediction model through backpropagation to obtain the trained time series prediction model.
[0047] To reduce the computational complexity of the edge-side MCU, after obtaining the trained temporal prediction model in step B3, a lightweight processing procedure for the temporal prediction model is also included, such as... Figure 4 As shown, the specific steps include: C1 employs a structured pruning strategy to evaluate the importance of the weight channels of neurons and gated recurrent units in each layer of the trained temporal prediction model. Neurons and structural blocks with contributions below a preset threshold are pruned while preserving the density of the weight matrix, resulting in a lightweight pruned model. This approach adapts to the computing architecture of edge-side MCUs, avoids the sparsity problem of the weight matrix caused by unstructured pruning, and significantly reduces the model's storage requirements and computational load.
[0048] As an optional implementation, quantization technology can also be introduced during the model lightweighting process to quantize the model weights from Float32 to INT8 or INT4, further compressing the model size and improving the speed of edge-side inference.
[0049] C2. Using the trained time-series prediction model as the teacher network and the pruned lightweight model as the student network, the student network is retrained using knowledge distillation based on the soft labels output by the teacher network and the intermediate layer feature information to obtain the lightweight time-series prediction model.
[0050] To compensate for the accuracy loss caused by pruning, a "teacher-student" knowledge distillation technique is employed in step C2. The high-precision time-series prediction model trained in step B3 is used as the teacher network, and the lightweight model pruned in step C1 is used as the student network. The student network is retrained using the soft labels output by the teacher network and the intermediate layer feature information. The soft labels are the probability distributions of different load ranges of ingredients output by the teacher network, used to guide the student network's load classification task training. The intermediate layer feature information is the shared feature vector output by the teacher network, used to guide the student network's remaining processing time regression task training. This allows the student network to maintain a small size while maximizing the inheritance of the teacher model's generalization ability and prediction accuracy, resolving the contradiction between model size and prediction accuracy.
[0051] The processing remaining time prediction method based on temporal multidimensional feature fusion provided in the above embodiments of this application overcomes the limitation of traditional methods that treat the oven as a closed system by simultaneously collecting temperature and humidity data, valve opening and closing logic signals, and cooking time data within the oven cavity. This allows the model to perceive external environmental interventions during the oven's operation, improving the environmental robustness of the prediction model. By constructing a multidimensional feature vector sequence through a sliding window, it can fully characterize the dynamic evolution trend of oven thermodynamic environment fluctuations and food ripeness. Combined with a gated recurrent unit network layer, it can effectively capture the long-term and short-term dependencies between temperature and humidity, valve status, and time progress during the steaming and baking process. By uncovering the implicit features of maturity evolution, the prediction accuracy of remaining processing time is significantly improved compared to traditional threshold judgment methods. A multi-task learning architecture is adopted, simultaneously learning food load characteristics and maturity evolution characteristics through a shared parameter layer. Load characteristics guide the prediction of remaining time, enabling adaptive perception of the maturation process under different food quantities. Through a lightweight strategy combining structured pruning and knowledge distillation, the model size and computational load are significantly reduced while maintaining prediction accuracy, adapting to the computing power and storage limitations of edge-side MCUs. This enables localized deployment and real-time inference of the prediction model, effectively improving the oven's intelligent perception capabilities and user experience.
[0052] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Moreover, the collection, use and processing of the relevant data are carried out in compliance with the relevant data protection laws and policies of the country where the location is located, and with the authorization granted by the owner of the corresponding device.
[0053] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0054] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.
Claims
1. A machining residual time prediction method based on time sequence multi-dimensional feature fusion, characterized in that, include: Acquire multidimensional raw time-series data during the oven's steaming and baking process, and perform time-scale alignment on the multidimensional raw time-series data based on a unified time reference to obtain aligned time-series data; A sliding window with a preset step size and window length is used to serialize and segment the aligned time-series data to construct a multidimensional feature vector sequence. The multidimensional feature vector sequence is input into the trained time-series prediction model, which outputs the load interval classification result and the remaining processing time prediction value corresponding to the current steamed / roasted ingredients. The time-series prediction model includes a gated recurrent unit network layer, a multi-task shared parameter layer, and parallel load identification and remaining time prediction branches connected to the output of the multi-task shared parameter layer. The gated recurrent unit network layer extracts high-dimensional time-series feature codes reflecting the heat and mass exchange state of the ingredients from the input multidimensional feature vector sequence. The multi-task shared parameter layer generates a shared feature vector containing both ingredient load information and remaining processing time information based on the high-dimensional time-series feature codes. The load identification branch generates a load interval classification result based on the shared feature vector. The remaining time prediction branch generates a remaining processing time prediction value based on the shared feature vector.
2. The method for predicting remaining processing time based on temporal multidimensional feature fusion according to claim 1, characterized in that, The multidimensional raw time-series data includes at least the temperature sequence, humidity sequence, valve opening and closing logic signal, and cooking time data corresponding to the steaming and baking process within the oven cavity. Acquiring the multidimensional raw time-series data during the oven's steaming and baking process specifically involves: real-time acquisition of the multidimensional raw time-series data during the oven's steaming and baking process at a preset sampling frequency; wherein the temperature sequence is synchronously acquired by two temperature sensors respectively installed on the rear and side of the oven cavity, with the measurement scale values of the two temperature sensors consistent with the sampling frequency; the humidity sequence is acquired by a humidity sensor built into the oven cavity; the valve opening and closing logic signal is directly acquired by the oven controller; and the cooking time data is the relative time data of the current steaming and baking process.
3. The method for predicting remaining processing time based on temporal multidimensional feature fusion according to claim 1, characterized in that, Using a sliding window with a preset step size and window length, the aligned time-series data is serialized and segmented to construct a multidimensional feature vector sequence, specifically including: Within each sliding window with a preset step size and window length, temperature change rate features and humidity change rate features are calculated and added to expand the feature dimensions. Among them, the two change rates are 0 at the first moment in the sliding window, and the change rates at the remaining moments are the differences between the feature values at the corresponding moments and the feature values at the first moment in the sliding window. The data from consecutive time periods segmented by the sliding window are combined into a multidimensional feature vector sequence with temporal correlation.
4. The method for predicting remaining processing time based on temporal multidimensional feature fusion according to claim 1, characterized in that, The gated loop unit network layer models the long-term and short-term dependencies between temperature fluctuations, humidity fluctuations, valve state switching and cooking time data through built-in reset gates and update gates, extracts high-dimensional time-series feature codes that reflect the heat and mass exchange state of ingredients, and outputs the high-dimensional time-series feature codes to the multi-task shared parameter layer.
5. The method for predicting remaining processing time based on temporal multidimensional feature fusion according to claim 1, characterized in that, The multi-task shared parameter layer receives the high-dimensional temporal feature encoding output by the gated loop unit network layer, learns the thermodynamic performance characteristics of different steamed and roasted ingredients under different loads, generates a shared feature vector that simultaneously contains ingredient load information and remaining processing time information, and outputs the shared feature vector to the load identification branch and the remaining time prediction branch respectively.
6. The method for predicting remaining processing time based on temporal multidimensional feature fusion according to claim 1, characterized in that, The load identification branch includes a fully connected layer and a Softmax classifier connected in sequence. It receives the shared feature vector output by the multi-task shared parameter layer, outputs the probability distribution of the load level corresponding to the current steamed and baked ingredients, and generates a load interval classification result.
7. The method for predicting remaining processing time based on temporal multidimensional feature fusion according to claim 1, characterized in that, The remaining time prediction branch includes a regression processing layer that receives the shared feature vector output by the multi-task shared parameter layer, calculates and outputs the predicted remaining processing time value corresponding to the current steamed / baked ingredients.
8. The method for predicting remaining processing time based on temporal multidimensional feature fusion according to claim 1, characterized in that, Before inputting the multidimensional feature vector sequence into the trained time-series prediction model and outputting the classification result of the load interval corresponding to the current steamed / baked ingredients and the predicted value of the remaining processing time, the method further includes: the training process of the time-series prediction model, specifically including: Training multidimensional time-series data of multiple steaming and baking processes of various typical ingredients were collected, and the ingredient load label and remaining processing time label were labeled for each steaming and baking process; wherein, the remaining processing time label was calculated from the stop time of the oven's corresponding ingredient steaming and baking program and the steaming and baking process timestamp; The multidimensional time-series data used for training is constructed into a multidimensional feature vector sequence, which serves as the input to the time-series prediction model. A joint loss function is defined as a linear combination of classification cross-entropy loss and regression mean squared error loss. The classification cross-entropy loss is the result of cross-entropy calculation between the output of the load identification branch of the time-series prediction model and the food load label. The regression mean squared error loss is the result of mean squared error calculation between the output of the remaining time prediction branch of the time-series prediction model and the remaining processing time label. The joint loss function value is calculated based on the output of the time-series prediction model, the corresponding food load label, and the remaining processing time label. The network parameters of the time-series prediction model are updated through backpropagation to obtain the trained time-series prediction model.
9. The method for predicting remaining processing time based on temporal multidimensional feature fusion according to claim 8, characterized in that, After obtaining the trained time series prediction model, the process also includes: a lightweight processing procedure for the time series prediction model, specifically including: A structured pruning strategy is adopted to evaluate the importance of the weight channels of neurons and gated recurrent units in each layer of the trained temporal prediction model. Neurons and structural blocks with contributions below a preset threshold are pruned while preserving the density of the weight matrix, resulting in a lightweight model after pruning. The trained time-series prediction model is used as the teacher network, and the pruned lightweight model is used as the student network. The student network is retrained using knowledge distillation based on the soft labels output by the teacher network and the feature information of the intermediate layer, so as to obtain the time-series prediction model with lightweight processing.