A temperature soft measurement method and system based on a physical information neural network, an electronic device, and a storage product

Abstract

This invention provides a temperature soft measurement method, system, electronic device, and storage product based on a physical information neural network, relating to the field of intelligent monitoring technology for wine aging processes. The invention first acquires the time-series data of the ambient temperature and the outer boundary temperature of a target container during aging. It then constructs a physical information neural network model containing a cascaded spatiotemporal feature extraction network of Transformer and Bi-LSTM modules. Furthermore, it introduces a physical equation residual loss based on unsteady-state heat conduction partial differential equations into the loss function to output the predicted temperature field of the fluid domain inside the target container. This achieves non-invasive soft measurement of the internal temperature, offering advantages such as small reconstruction error, high robustness, and strong generalization ability.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology for the wine aging process, and in particular to a temperature soft measurement method, system, electronic device and storage product based on physical information neural network. Background Technology

[0002] The domestic wine industry is promoting the digital transformation of the entire industry chain, and building an intelligent brewing system has become a new trend. Aging is a crucial link between the initial and subsequent stages of winemaking for most high-quality wines. The aging process is not merely simple physical storage, but a complex, unsteady-state thermo-fluid coupling process. Oak barrels, as the most classic aging container, are a typical thick-walled, porous medium. Their significant thermal damping effect has a substantial low-pass filtering and phase lag effect on ambient temperature fluctuations, leading to severe attenuation of external observation information. Simultaneously, due to natural convection, a weak but stable large-scale thermal stratification structure easily forms inside the container. Subtle instabilities in the spatiotemporal distribution of the temperature field can easily trigger nonlinear biochemical evolution of the wine, leading to oxidation, spoilage, or aroma fragmentation.

[0003] Real-time monitoring of the internal temperature field of aging containers can be achieved using contact thermometry or numerical simulation. However, three major problems exist in practical applications: First, physical probe methods disrupt the micro-oxygen environment. Traditional contact temperature probes need to be immersed in the wine using a barrel stopper, which not only disrupts the original sealed micro-oxygen balance of the oak barrel, increasing the risk of contamination and excessive oxidation, but also makes the probe prone to tartaric acid crystals from long-term immersion, leading to data drift. Furthermore, single-point monitoring data cannot comprehensively reflect the three-dimensional temperature gradient distribution inside a large-inertia container. Second, traditional computational fluid dynamics (CFD) simulations are computationally expensive and disconnected from real-world conditions. CFD computation is extremely costly and cannot meet the needs of real-time online monitoring; moreover, CFD heavily relies on idealized boundary condition assumptions and cannot dynamically correct parameters based on changes in ventilation and humidity or individual thermophysical properties of oak planks in a real cellar environment. Third, purely data-driven deep learning models suffer from the challenges of small sample sizes and poor generalization. Traditional deep learning models require massive amounts of real-world labeled data for training, but in real-world brewing scenarios, it's impossible to obtain accurate internal temperature data for the entire field. Without the constraints of fluid dynamics equations, purely data-driven models are prone to getting trapped in local optima, producing spurious solutions that violate physical laws. Once environmental conditions change, the model's accuracy will suffer a significant drop.

[0004] Currently, under the dual constraints of severely lacking observational information and simplified physical mechanisms, accurately, non-destructively, and adaptively reconstructing the spatiotemporal distribution of the thermal-fluid coupling field inside thick-walled containers based solely on sparse external monitoring data is a critical scientific problem that urgently needs to be solved in the field of intelligent brewing. Therefore, balancing computational efficiency, prediction accuracy, and non-destructive and non-invasive methods for specific variable operating conditions, and adaptively transferring the model, is key to researching and improving the performance of non-invasive soft measurement in the aging process. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide a temperature soft measurement method, system, electronic device, and storage product based on physical information neural networks. This invention addresses the deficiencies of existing technologies, such as invasive measurements damaging the sealing of the wine, inaccurate and time-consuming simulation models, and the lack of generalization ability in purely data-driven models. It aims to utilize the constraint force of physical equation residuals to compensate for the lack of internal spatial data, reduce temperature field reconstruction errors, and improve the robustness of soft measurements.

[0006] To achieve the above objectives, the present invention provides the following solution: A soft temperature measurement method based on a physical information neural network includes: Acquire the time-series data of the ambient temperature and the time-series data of the outer wall boundary temperature of the target container during the aging period; A physical information neural network model is constructed. The physical information neural network model includes a spatiotemporal feature extraction network, which is composed of a Transformer module and a bidirectional long short-term memory network Bi-LSTM module cascaded together. The loss function of the physical information neural network model includes the physical equation residual loss based on the unsteady heat conduction partial differential equation. The ambient temperature time series data and the outer wall boundary temperature time series sequence are input into a trained physical information neural network model. The Transformer module extracts the global long-term impact features of the ambient temperature change on the internal temperature, and the Bi-LSTM module extracts the local time series dynamic features and decodes and outputs the temperature field prediction results of the fluid domain inside the target container, realizing non-invasive soft measurement of the internal temperature.

[0007] Preferably, the ambient temperature time series data and the outer wall boundary temperature time series data are input into a trained physical information neural network model, including: The ambient temperature time series data and the outer wall boundary temperature time series data are used as input tensors to input the Transformer module. The multi-head self-attention mechanism is used to identify key thermal shock moments and extract the global long-term impact features of the abrupt change in ambient temperature on the internal temperature. The global long-term impact features are input into the Bi-LSTM module, and the time-series recursion of heat and energy dissipation are simulated using the gating mechanism. The local time-series dynamic features are extracted and the temperature field prediction results are decoded and output.

[0008] Preferably, the loss function of the physical information neural network model is a composite loss function, which includes not only the physical equation residual loss, but also data fitting loss, boundary matching loss, and interface coupling loss. During model training, the weights of each loss in the composite loss function are dynamically adjusted through an adaptive gradient balancing algorithm.

[0009] Preferably, the process of obtaining the residual loss of the physical equation includes: Randomly sample unlabeled physical configuration points within the spatiotemporal domain of the defined target container; Substitute the temperature field prediction results from the physical information neural network model into the unsteady heat conduction partial differential equation; The partial derivatives of the target coordinate points are calculated using automatic differentiation techniques. The deviation between the transient heat storage term and the heat conduction term in the unsteady heat conduction partial differential equation is then calculated as the residual loss of the physical equation.

[0010] Preferably, the physical information neural network model is trained and validated according to a progressive transfer strategy: Phase 1: Fully supervised pre-training in the simulation domain. The model is pre-trained using a multi-condition simulation spatiotemporal evolution dataset, and the data fitting loss is activated. Phase Two: Real Domain Physical Fine-Tuning. Acquire mixed data with internal temporary probes during the initial deployment phase, activate the data fitting loss, and introduce physical equation residual loss and interface coupling loss to fine-tune the high-level weights of the model. Phase 3, full-cycle non-invasive soft measurement verification: The measured data of the internal temporary probe is shielded in the algorithm logic to simulate the cold start adaptive process under real variable operating conditions. Unsupervised parameter optimization is performed by minimizing the boundary matching loss, the physical equation residual loss, and the interface coupling loss. The data fitting loss is shielded, and the internal temperature field prediction result is output. Subsequently, the temperature field prediction result is compared with the measured temperature of the internal temporary probe in real time to verify and calibrate the reconstruction accuracy of the model under non-invasive conditions.

[0011] Preferably, the real-domain physics fine-tuning in stage two includes: The effective thermal conductivity and air convection heat transfer coefficient of the target container are embedded into the computational graph as trainable variables. Gradient descent is used to dynamically identify real physical parameters in order to correct systematic biases caused by simulation assumptions.

[0012] Preferably, when changing the target container or applying under varying environmental conditions, an unsupervised adaptive migration mechanism is executed; the unsupervised adaptive migration mechanism includes: Extracting network weights from the backbone features of a frozen physical information neural network model; Based solely on short-term ambient temperature time-series data and outer wall boundary temperature time-series data, unsupervised gradient descent is used to optimize the equivalent thermal properties of the target container, aiming to minimize boundary matching loss, physical equation residual loss, and interface coupling loss. The equivalent thermal properties include effective thermal conductivity and air convection heat transfer coefficient.

[0013] A temperature soft measurement system based on a physical information neural network includes: The acquisition module is used to acquire the time-series data of the ambient temperature and the time-series data of the outer wall boundary temperature of the target container during the aging period; The building module is used to construct the physical information neural network model. The physical information neural network model includes a spatiotemporal feature extraction network, which is composed of a Transformer module and a bidirectional long short-term memory network Bi-LSTM module cascaded together. The loss function of the physical information neural network model includes the physical equation residual loss based on the unsteady heat conduction partial differential equation. The soft measurement module is used to input the ambient temperature time series data and the outer wall boundary temperature time series sequence into the trained physical information neural network model. The Transformer module extracts the global long-term impact features of the ambient temperature change on the internal temperature, and the Bi-LSTM module extracts the local time series dynamic features and decodes and outputs the temperature field prediction results of the fluid domain inside the target container, realizing non-invasive soft measurement of the internal temperature.

[0014] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the temperature soft measurement method described above.

[0015] A storage product storing a computer program, which, when executed by a processor, performs the steps of the temperature soft measurement method described above.

[0016] The present invention discloses the following beneficial effects: This invention constructs a physical equation residual loss constraint model based on unsteady heat conduction partial differential equations, which reduces the error caused by reconstructing the internal temperature field under sparse boundaries in a purely data-driven model; and improves the accuracy and generalization ability of the model in identifying physical parameters under varying operating conditions through model adaptive migration and cold start mechanisms. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention 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 the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart of the method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the technical route provided in the embodiments of the present invention; Figure 3 This is a schematic diagram of the Transformer-Bi-LSTM-PINN network architecture and thermal flow field coupling mechanism provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the system provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the physical structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] The purpose of this invention is to provide a temperature soft measurement method, system, electronic device and storage product based on physical information neural network. By combining physical equation residual constraints with spatiotemporal feature extraction network, it is possible to achieve non-invasive soft measurement of the temperature field of the fluid domain inside the target container based only on ambient temperature time series data and outer wall boundary temperature time series, thereby reducing the temperature field reconstruction error under sparse boundary conditions and improving the model's adaptability under varying operating conditions.

[0021] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] Figure 1 The method flowchart provided in the embodiments of the present invention is as follows: Figure 1 As shown, the present invention provides a temperature soft measurement method based on a physical information neural network, comprising: S1: Obtain the time-series data of the ambient temperature and the time-series data of the outer wall boundary temperature of the target container during the aging period; S2: Construct a physical information neural network model; the physical information neural network model includes a spatiotemporal feature extraction network, which is composed of a Transformer module and a bidirectional long short-term memory network Bi-LSTM module cascaded together, and the loss function of the physical information neural network model includes the physical equation residual loss based on the unsteady heat conduction partial differential equation; S3: Input the ambient temperature time series data and the outer wall boundary temperature time series sequence into the trained physical information neural network model. The Transformer module extracts the global long-term impact features of the ambient temperature change on the internal temperature. The Bi-LSTM module extracts the local time series dynamic features and decodes and outputs the temperature field prediction results of the fluid domain inside the target container, realizing non-invasive soft measurement of the internal temperature.

[0023] As an example, the target container used in this invention is a standard 225L thick-walled French oak barrel, and the internal fluid medium is wine. Figure 2 The technical route flowchart provided by the present invention is as follows: Figure 2 As shown, the technical approach of the present invention includes: Step 100: Obtain the ambient temperature time series data and the outer wall boundary temperature time series of the target container during the aging period.

[0024] The data required for this invention comes from real-world sensor data collected in a wine cellar environment. Since the aging container (oak barrel) is a highly inertial porous media system, temperature fluctuations in the external environment will produce significant thermal hysteresis when penetrating the thick-walled medium. Therefore, obtaining a high-precision time series of ambient temperature and outer wall temperature is a prerequisite for reconstructing the system's energy input boundary.

[0025] To accurately detect subtle changes in heat flux on the outer wall of oak barrels and ensure they are unaffected by noise under complex environmental heat radiation conditions, this invention, in a preferred embodiment, proposes a system integration and hardware calibration scheme for a non-invasive sensing terminal: Select Class A PT100 thin-film resistance thermometer (accuracy) As a temperature patch sensor, it is attached to the top, bottom, and widest part of the oak barrel. To eliminate contact thermal resistance, a high thermal conductivity silicone grease is applied between the probe and the outer wall of the oak barrel, and a weather-resistant silicone tape is used for constant pressure bonding to ensure zero-gap contact between the sensor and the curved barrel wall.

[0026] Furthermore, to address potential wind interference in the underground wine cellar environment, an aerogel-aluminum foil double-layer shielding structure was designed. The sensor is covered with a silica aerogel insulation felt of appropriate thickness, and the outermost layer is wrapped with high-reflectivity aluminum foil tape. This double-layer shielding structure completely blocks heat exchange between the back of the sensor and the ambient air, forcing the sensor to rely solely on heat conduction with the outer wall of the oak barrel. This eliminates the spurious superposition of ambient wind speed and sunlight on wall temperature measurement, ensuring the acquisition of pure solid-gas interface thermal response data. .

[0027] Ambient temperature sensors synchronously acquire time-series data of ambient temperature within the wine cellar. The sensing terminal, through a smart gateway integrating a LoRa spread spectrum communication module, aggregates the aforementioned multiple timing signals and sends them to the processing platform to complete the acquisition of raw data for the soft measurement input tensor.

[0028] Step 200: Construct the unsteady-state thermal-fluid coupling multiphysics field control equations and energy coupling interface.

[0029] Furthermore, in another embodiment, in the aging microenvironment temperature field modeling of the present invention, in order to accurately capture the complex thermal hysteresis effect caused by natural convection inside the target container (aging oak barrel) while ensuring computational efficiency, this embodiment adopts an equivalent heat conduction modeling method based on the effective medium theory.

[0030] In practical applications, the aging system can be decoupled into two independent physical fields: the solid domain (barrel wall) and the fluid domain (internal wine), with energy coupling achieved through Robin boundaries and continuous interfaces. The process follows the principles of energy conservation, Newton's law of cooling, and Fourier's law. The specific construction process is as follows: ①Governing equations of the solid domain (oak barrel wall) ): Inside the outer wall of the target container, heat transfer follows the unsteady-state heat conduction law. This embodiment retains the time derivative term (heat storage term) to characterize the low-pass filtering and interception effect of the thick-walled structure's heat capacity on ambient temperature fluctuations. Its governing equation is as follows:

[0031] In the formula, , These are the density and specific heat capacity of oak, respectively. Let be the thermal conductivity of oak. The heat storage term on the left side of the equation represents the power retained by the barrel wall, explaining the physical mechanism by which the inner wall temperature does not respond immediately when the ambient temperature fluctuates drastically. That is, the heat capacity of the thick wall of the solid domain must be filled first before the remaining heat can be transferred to the interior.

[0032] ② Fluid domain governing equations (internal wine liquid) ): Within a fluid, heat transfer is a mixture of conduction and natural convection. To address the complex convection term... This invention simplifies the thermal conductivity into a parameter form that can be recognized by neural networks, and introduces an effective thermal conductivity coefficient. Perform equivalent substitution:

[0033] In the formula, , These are the density and specific heat capacity of the internal fluid, respectively. Only when internal convection is strong... ( (Static thermal conductivity) characterizes the rapid transfer of heat. However, under stable aging conditions, the interior is dominated by weak natural convection, and the fluid domain maintains a stable large-scale thermal stratification, which forms the physical basis for subsequent multi-point biochemical predictions in space.

[0034] ③ Boundary conditions and coupling interfaces: The external boundary, i.e., the gas-solid (outer wall of the container) interface, uses Robin boundary conditions to describe the heat exchange between the air and the outer wall of the target container, determining the threshold for ambient heat to enter the system.

[0035] In the formula, The temperature refers to the temperature at the outer boundary of the container. It refers to the air convection heat transfer coefficient.

[0036] For solid-liquid coupling interfaces This satisfies both temperature continuity and heat flow continuity (energy conservation), meaning that the heat conducted by the solid domain, after deducting its own heat storage, is entirely transferred to the fluid domain.

[0037] Furthermore, in order to demonstrate the effectiveness of thermal conductivity... To ensure the scientific validity and equivalence of replacing the natural convection term, this invention introduces a dimensionless number correlation (Nusser number) based on the principle of energy conservation. Assuming heat passes through the natural convection boundary layer within the fluid, the heat flux is... If the fluid is considered as a quasi-solid with enhanced thermal conductivity, then the characteristic scale is defined. Then heat flux can be written as From this, the Nusselt number can be derived. The expression:

[0038] The physical essence of this is the ratio of the effective thermal conductivity to the thermal conductivity of a stationary fluid. Adjusting this in the algorithm model... The numerical value is mathematically equivalent to simulating a natural convection effect of a specific intensity. This equivalent conversion provides a rigorous theoretical basis for the parameter identification of the subsequent PINN network.

[0039] Step 300: Construct a spatiotemporal feature extraction network, which is composed of a Transformer module and a bidirectional long short-term memory network Bi-LSTM module cascaded together.

[0040] In practical applications of non-invasive temperature field reconstruction, the significant thermal damping effect of the target container causes substantial phase delay and amplitude attenuation in the transmission of external temperature fluctuations to the internal center. Solving the purely physical equations is extremely time-consuming, and a single neural network struggles to handle such long-period temporal dependencies.

[0041] Therefore, considering the significant thermal damping and phase lag characteristics of aging containers, this embodiment innovatively introduces and constructs a spatiotemporal feature extraction backbone architecture that integrates Transformer and Bi-LSTM in the physical information neural network model. For example... Figure 3 As shown, this invention utilizes the multi-head attention mechanism of Transformer specifically for extracting long-range abrupt changes in ambient temperature, and the gating mechanism of Bi-LSTM specifically for extracting local thermal hysteresis features, establishing a high-precision boundary fluctuation -> internal response nonlinear mapping. Specifically, it includes the following sub-steps: Step 310, Input Tensor Construction: Select the past The hourly boundary state sequence is used as the input tensor, where... Preferably, a complete diurnal cycle (e.g., 24 hours) is used to cover the complete asymmetric thermal hysteresis loop:

[0042] in, For time-series input tensor matrices, For discrete time step variables, The real-time temperatures at three key measuring points on the outer wall (upper, middle, and lower) characterize the energy input boundary of the system.

[0043] Step 320, Transformer Global Feature Encoding: A multi-head self-attention mechanism is introduced to capture the long-term effects of potential drastic fluctuations in ambient temperature on internal temperature. The input tensor... Projection into a query (Q), key (K), value (V) matrix:

[0044] Where Q represents the key time point features of the temperature that need to be reconstructed, K contains the thermal shock feature library for each moment in the entire historical window (the past 24 hours), and V is the corresponding specific thermal state value. This is a learnable weight parameter matrix.

[0045] Furthermore, the correlation score between time steps is calculated using the scaled dot product to identify key thermal shock moments:

[0046] Finally, the multi-head attention outputs are concatenated and linearly mapped to obtain the global feature vector. .

[0047] Step 330, Bi-LSTM Local Timing Dynamics and Output Decoding: The global feature vector extracted by Transformer Inputting the data into a Bi-LSTM network, its gating mechanism is used to simulate the temporal recursion of heat and energy dissipation. This design gives the neural network nodes explicit physical meaning. For each time step in the forward propagation, including: Forgotten Gate ( ): Simulates the heat dissipation process to determine how much of the previous enthalpy state is retained.

[0048] Input gate ( ): Simulate the impact of the current boundary heat flux input on the system energy.

[0049] Cell state update ( ): Represents the total thermodynamic energy state of the system at the current moment.

[0050]

[0051] In the above formula, This represents the Sigmoid activation function. and These represent the weight and bias of the corresponding gating unit, respectively.

[0052] Finally, the forward hidden state and the backward hidden state are concatenated, decoded by a fully connected layer (MLP), and the predicted temperature of the target container at the current moment (or at any query point in space) is output. .

[0053] Step 400: Construct a physical information-driven composite loss function library.

[0054] In traditional deep learning models, neural networks rely solely on black-box fitting of real-label data. This black-box approach, lacking sufficient internal experimental data, easily leads to predictions that violate fundamental thermodynamic laws. To ensure that deep learning models strictly adhere to the thermodynamic laws established in step 200 above, this invention introduces Automatic Differentiation (AD) technology, constructing a composite loss function that includes four components: data fitting, physical residuals, interface coupling, and boundary matching. .

[0055] Furthermore, in another embodiment, the overall expression of the composite loss function is:

[0056] The weights are dynamically adjusted using an adaptive gradient balancing algorithm. This addresses the gradient competition problem between loss functions of different magnitudes in multi-task learning. The specific process for obtaining and constructing each sub-loss function is as follows: ① Data fitting loss : Used to anchor real-world observations. During the pre-training phase, this term targets multi-condition simulation datasets; during the real-domain physical fine-tuning phase, it targets data acquired by sparse sensors (external wall patch sensors + internal temporary probes). Its calculation formula is:

[0057] In the formula, The total number of data points represents a large number of simulation nodes in the pre-training stage and a sparse number of real probe sampling points in the fine-tuning stage. Let be the spacetime coordinate point, representing the th Radial position of each sampling point inside the aging container ( axial depth and the time ; and These represent the predicted temperature from the neural network and the actual measured temperature from the sensor, respectively.

[0058] ② Boundary matching loss ( ): During the non-invasive measurement phase, in which the internal probe is completely removed, the outer wall temperature inferred from the model is forced to be strictly consistent with the measured value of the aerogel shielded patch sensor to prevent drift of the system boundary conditions.

[0059] in, To be on the outer wall surface of the target container ( The total number of boundary points sampled. and These represent the predicted value of the model's outer wall and the actual measured value of the patch sensor, respectively.

[0060] ③ Physical equation residual loss ( ): This is the core physical constraint for achieving non-invasive measurement in this invention. It involves using a neural network to predict the temperature field. The residuals are calculated by substituting them into the unsteady partial differential equation (PDE) for heat conduction. This process requires no real label data; it only involves randomly sampling placement points within the domain.

[0061] in, This represents the number of randomly sampled, unlabeled physical configuration points. For transient heat storage, Let represent the radial and axial heat conduction terms in cylindrical coordinates. Partial derivatives are precisely calculated using automatic differentiation techniques, forcing the neural network to strictly adhere to the law of conservation of energy even within a fluid domain devoid of any observational data. If the model produces a thermal illusion that violates physical laws (e.g., heating without a heat source), the residual will significantly increase and correct the network weights through backpropagation.

[0062] ④ Interface coupling loss ( ): At the solid-liquid interface At this point, the continuity of temperature and heat flow is constrained to ensure the physical conservation of energy transfer across the medium:

[0063] Step 500: Execute the Sim-to-Real progressive migration strategy and unsupervised adaptive cold start.

[0064] To address the challenges of scarce internal flow field data in real-world aging scenarios and the difficulty of adapting to individual differences in target containers (such as thermal property drift caused by oak wood texture) using general simulation models alone, this invention proposes an implementation method that integrates a Sim-to-Real progressive migration strategy and an unsupervised cold start approach.

[0065] Furthermore, in another embodiment, the physical information neural network model is trained according to the following three-stage evolution strategy: Phase 1: Fully supervised pre-training in the simulation domain.

[0066] Utilize multiple operating conditions (such as including different) and The orthogonal combination of these methods generates a spatiotemporal evolution dataset through simulation. At this stage, only the data fitting loss function is activated. Strongly supervised training is performed to enable the Transformer-Bi-LSTM backbone network to quickly grasp the general manifold structure of the temperature field as parameters change, thus injecting physical prior initial weights into the model.

[0067] Phase Two: Real-Domain Physical Fine-Tuning and Parameter Identification.

[0068] In the initial deployment phase of the monitoring system, temporary calibration components (such as rigid Teflon probes) are configured, employing a hybrid acquisition scheme combining non-invasive sensors and internal temporary probes. The combined loss function is activated during this phase.

[0069] The predicted values ​​are corrected using internally measured true values; furthermore, the innovation of this embodiment lies in: using the effective thermal conductivity... and air convection heat transfer coefficient Embedded as a trainable variable in the computational graph, the real physical parameters under the current working conditions are dynamically identified using the gradient descent method with real data, thus correcting the systematic deviations caused by simulation assumptions.

[0070] Phase 3: Full-cycle non-invasive soft measurement verification.

[0071] The algorithm logic masks the measured data of the internal temporary probes to simulate the cold start adaptive process under real varying operating conditions. Unsupervised parameter optimization is performed by minimizing the boundary matching loss, the physical equation residual loss, and the interface coupling loss; the data fitting loss is also masked. The system outputs the predicted internal temperature field; then, it compares the predicted temperature field with the measured temperature of the internal temporary probe in real time to verify and calibrate the reconstruction accuracy of the model under non-invasive conditions.

[0072]

[0073] The model only receives ambient temperature and time-series inputs from the outer wall boundary. Under the strong constraint of physical residuals, it automatically back-evolves the unique and reasonable internal three-dimensional temperature field solution based on the law of conservation of energy.

[0074] Furthermore, in order to improve the robustness of the model under different operating conditions, this invention also provides a cold-start unsupervised adaptive mechanism for variable operating conditions.

[0075] In actual production, when faced with the replacement of target containers (such as replacing old oak barrels with new ones) or drastic environmental changes, it is impossible to perform invasive pin calibration again. Therefore, this embodiment designs a thermal inertia-cycled cold start protocol (e.g., 24-48 hours): In the initial stage of deployment under varying operating conditions, the system only relies on the boundary temperature response sequences captured by the ambient temperature sensor and the external wall sensor to freeze the main weights of the Transformer-Bi-LSTM backbone feature extraction network, and only applies the physical parameter layer ( Unsupervised gradient descent optimization is performed. The objective function is constructed as follows:

[0076] By leveraging the characteristics of the external boundary response data and minimizing the physical equation residuals, the optimal equivalent thermophysical parameters under internal physical constraints can be calculated in reverse. This enables rapid cold start under varying operating conditions and non-destructive physical adaptive calibration for each tank.

[0077] The temperature soft measurement system based on physical information neural network provided by the present invention is described below. The temperature soft measurement system described below can be referred to in correspondence with the temperature soft measurement method described above.

[0078] Figure 4 This invention also provides a schematic diagram of a temperature soft measurement system based on a physical information neural network, as shown below. Figure 4 As shown, the present invention also provides a temperature soft measurement system, the system comprising: The acquisition module is used to acquire the ambient temperature time series data and the outer wall boundary temperature time series of the target container during the aging period. This module uses a smart gateway with an integrated LoRa spread spectrum communication module to collect data from the aerogel double-shielded PT100 thin film thermistor and the ambient temperature sensor to complete the acquisition of the raw data of the soft measurement input tensor.

[0079] The module is used to construct a physical information neural network model. The loss function of the physical information neural network model includes the physical equation residual loss constructed based on the unsteady heat conduction partial differential equation. The module further includes a spatiotemporal feature extraction unit for constructing a Transformer and Bi-LSTM cascaded architecture, and a loss calculation unit for constructing a library of four composite loss functions—data fitting, boundary matching, physical equation, and interface coupling—using automatic differentiation technology.

[0080] The soft measurement module is used to input the ambient temperature time series data and the outer wall boundary temperature time series sequence into the trained physical information neural network model. Under the constraints of physical laws, including at least the residual loss of the physical equations, it outputs the temperature field prediction result of the fluid domain inside the target container, realizing non-invasive soft measurement of the internal temperature. This module also embeds a Sim-to-Real progressive migration unit and an unsupervised adaptive cold start unit, used to freeze the backbone network and minimize the physical residuals when changing the target container or operating conditions, and to unsupervisedly back-calculate the equivalent thermal property parameters (effective thermal conductivity). and air convection heat transfer coefficient This enables physical calibration tailored to each container.

[0081] Figure 5 A schematic diagram of the physical structure of an electronic device provided by the present invention, such as... Figure 5 As shown, the electronic device may include: a processor 810, a communication interface 820, a memory 830, and a communication bus 840, wherein the processor 810, the communication interface 820, and the memory 830 communicate with each other via the communication bus 840. The processor 810 can call logical instructions in the memory 830 to execute a temperature soft measurement method based on a physical information neural network, including: Acquire the time-series data of the ambient temperature and the time-series data of the outer wall boundary temperature of the target container during the aging period; A physical information neural network model is constructed, wherein the loss function of the physical information neural network model includes the physical equation residual loss constructed based on the unsteady heat conduction partial differential equation; The ambient temperature time series data and the outer wall boundary temperature time series are input into the trained physical information neural network model. Under the constraints of physical laws that include at least the residual loss of the physical equations, the temperature field prediction result of the fluid domain inside the target container is output, realizing non-invasive soft measurement of the internal temperature.

[0082] Furthermore, the logical instructions in the aforementioned memory 830 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, an edge computing gateway, a cloud server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0083] On the other hand, the present invention also provides a computer program product, the computer program product including a computer program stored on a non-transitory computer-readable storage medium, the computer program including program instructions, and when the program instructions are executed by a computer, the computer is able to execute the temperature soft measurement method based on physical information neural network provided by the above methods.

[0084] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the temperature soft measurement methods provided in the above embodiments.

[0085] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0086] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, 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 the present invention.

Claims

1. A soft temperature measurement method based on a physical information neural network, characterized in that, include: Acquire the time-series data of the ambient temperature and the time-series data of the outer wall boundary temperature of the target container during the aging period; A physical information neural network model is constructed; the physical information neural network model includes a spatiotemporal feature extraction network, which is composed of a Transformer module and a bidirectional long short-term memory network Bi-LSTM module cascaded together, and the loss function of the physical information neural network model includes the physical equation residual loss based on the unsteady heat conduction partial differential equation; The ambient temperature time series data and the outer wall boundary temperature time series sequence are input into the trained physical information neural network model. The Transformer module extracts the global long-term impact features of the ambient temperature change on the internal temperature, and the Bi-LSTM module extracts the local time series dynamic features and decodes and outputs the temperature field prediction results of the fluid domain inside the target container, thereby realizing non-invasive soft measurement of the internal temperature.

2. The temperature soft measurement method according to claim 1, characterized in that, The environmental temperature time series data and the outer wall boundary temperature time series data are input into the trained physical information neural network model, including: The ambient temperature time series data and the outer wall boundary temperature time series are used as input tensors to input the Transformer module. The multi-head self-attention mechanism is used to identify key thermal shock moments and extract the global long-term impact features of the ambient temperature change on the internal temperature. The global long-term impact features are input into the Bi-LSTM module, and the time-series recursion of heat and energy dissipation are simulated using a gating mechanism. Local time-series dynamic features are extracted and decoded to output the temperature field prediction results.

3. The temperature soft measurement method according to claim 1, characterized in that, The loss function of the physical information neural network model is a composite loss function, which includes, in addition to the physical equation residual loss, data fitting loss, boundary matching loss, and interface coupling loss. During model training, the weights of each loss term in the composite loss function are dynamically adjusted through an adaptive gradient balancing algorithm.

4. The temperature soft measurement method according to claim 3, characterized in that, The process of obtaining the residual loss of the physical equation includes: Randomly sample unlabeled physical configuration points within the spatiotemporal domain of the defined target container; Substitute the temperature field prediction result predicted by the physical information neural network model into the unsteady heat conduction partial differential equation; The partial derivatives of the target coordinate points are calculated using automatic differentiation techniques. The deviation between the transient heat storage term and the heat conduction term in the unsteady heat conduction partial differential equation is then calculated as the residual loss of the physical equation.

5. The temperature soft measurement method according to claim 3, characterized in that, The physical information neural network model is trained and validated according to a progressive transfer strategy; The gradual migration strategy includes: Phase 1: Fully supervised pre-training in the simulation domain. The model is pre-trained using a multi-condition simulation spatiotemporal evolution dataset, and the data fitting loss is activated. Phase Two: Real Domain Physical Fine-Tuning. Acquire mixed data with internal temporary probes during the initial deployment phase, activate the data fitting loss, and simultaneously introduce the physical equation residual loss and the interface coupling loss to fine-tune the high-level weights of the model. Phase 3, full-cycle non-invasive soft measurement verification: The measured data of the internal temporary probe is shielded in the algorithm logic to simulate the cold start adaptive process under real variable operating conditions. Unsupervised parameter optimization is performed by minimizing the boundary matching loss, the physical equation residual loss, and the interface coupling loss. The data fitting loss is shielded, and the internal temperature field prediction result is output. Subsequently, the temperature field prediction result is compared with the measured temperature of the internal temporary probe in real time to verify and calibrate the reconstruction accuracy of the model under non-invasive conditions.

6. The temperature soft measurement method according to claim 5, characterized in that, The real-domain physics fine-tuning in Phase Two includes: The effective thermal conductivity and air convection heat transfer coefficient of the target container are embedded into the computation graph as trainable variables. Gradient descent is used to dynamically identify real physical parameters in order to correct systematic biases caused by simulation assumptions.

7. The temperature soft measurement method according to claim 5, characterized in that, When changing the target container or applying under fluctuating environmental conditions, an unsupervised adaptive migration mechanism is executed; the unsupervised adaptive migration mechanism includes: Freeze the backbone feature extraction network weights of the physical information neural network model; Based solely on the short-term ambient temperature time series data and the outer wall boundary temperature time series data, with the goal of minimizing the boundary matching loss, the physical equation residual loss, and the interface coupling loss, unsupervised gradient descent is used to optimize the equivalent thermal property parameters of the target container; the equivalent thermal property parameters include the effective thermal conductivity and the air convection heat transfer coefficient.

8. A temperature soft measurement system based on a physical information neural network, characterized in that, include: The acquisition module is used to acquire the time-series data of the ambient temperature and the time-series data of the outer wall boundary temperature of the target container during the aging period; The building module is used to build a physical information neural network model. The physical information neural network model includes a spatiotemporal feature extraction network, which is composed of a Transformer module and a bidirectional long short-term memory network Bi-LSTM module cascaded together. The loss function of the physical information neural network model includes the physical equation residual loss based on the unsteady heat conduction partial differential equation. The soft measurement module is used to input the ambient temperature time series data and the outer wall boundary temperature time series sequence into the trained physical information neural network model. The Transformer module extracts the global long-term impact features of the ambient temperature change on the internal temperature, and the Bi-LSTM module extracts the local time series dynamic features and decodes and outputs the temperature field prediction results of the fluid domain inside the target container, thereby realizing non-invasive soft measurement of the internal temperature.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the temperature soft measurement method as described in any one of claims 1 to 7.

10. A storage product, characterized in that, The storage product stores a computer program that, when executed by a processor, implements the steps of the temperature soft measurement method as described in any one of claims 1 to 7.

Images