Remote heat meter data calibration method based on deep learning

Abstract

The invention relates to the technical field of deep learning, and discloses a remote heat meter data calibration method based on deep learning, which comprises the following steps: constructing a resource awareness optimization framework with maximized information retention, and optimizing model performance under resource constraint by utilizing a mutual information principle; constructing a self-adaptive wavelet preprocessing engine, dynamically selecting a wavelet decomposition level according to hardware capability, and realizing effective dimension reduction of input data; constructing a neural architecture search framework capable of differentiating resource constraint, and automatically generating an optimal network architecture meeting equipment resource limitation; a variable precision calculation mechanism is realized, and differential precision processing strategies are implemented on different layers in the network; implementing a task decomposition strategy, and decomposing the complex calibration task into parallel simple sub-tasks; a progressive model distillation process is realized, and complex model knowledge on high-performance equipment is migrated to a lightweight model on low-performance equipment. The model deployment and updating process is simplified, and the system maintenance cost and the technical threshold are reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep learning technology, more specifically, it relates to a remote heat meter data calibration method based on deep learning. BACKGROUND

[0002] With the rapid development of smart cities and intelligent heating systems, remote heat meters, as the core metering equipment in the heating pipe network, are widely deployed. These metering equipment needs to measure the water flow and the inlet and outlet water temperature in real time, calculate the heat consumption and remotely transmit data to realize accurate monitoring and management of the heating system. However, in actual application, heat meter data often has various errors and noises, which need to be calibrated effectively to ensure the accuracy of metering.

[0003] Currently, heat meter data calibration mainly adopts the following technologies: first, the calibration method based on physical model, which corrects the measurement error by establishing a thermodynamic model, but this method is difficult to adapt to complex changing conditions; second, traditional machine learning calibration methods such as support vector machine, random forest, etc., which have limited performance in dealing with nonlinear relationships; third, deep learning calibration methods developed in recent years, although the accuracy has improved, but the demand for computing resources is large, and it is difficult to deploy on resource-limited devices.

[0004] The existing technology faces several key problems in actual application: first, the heat meter deployment environment varies from high-performance edge devices to low-power embedded systems, with a huge difference in computing power, and traditional calibration methods are difficult to adapt to such a wide range of hardware differences, usually requiring separate development and optimization of models for each type of device, resulting in high development and maintenance costs; second, existing deep learning calibration models often require high computing resources to achieve ideal accuracy, when deployed on resource-limited devices, either the accuracy is significantly reduced due to excessive model pruning, or the device cannot bear or responds slowly due to the large model; in addition, although general models have a wide range of applications, their accuracy is often insufficient, while specialized models have high accuracy but narrow application range, making it difficult to achieve both; moreover, heat meter data often contains noise and outliers, and existing calibration methods are not effective in dealing with these interference factors, especially the difference in signal processing capability on different hardware platforms; finally, existing calibration systems often lack adaptive optimization mechanisms for different hardware platforms, making the model deployment and updating process cumbersome and difficult to achieve efficient iteration of the model.

[0005] Therefore, there is an urgent need for a remote heat meter data calibration method that can efficiently run in a heterogeneous hardware environment, meeting the deployment needs of different computing resources while ensuring calibration accuracy and real-time performance, thereby reducing system maintenance costs and improving the overall operation efficiency of the heating system. SUMMARY

[0006] To address the aforementioned technical problems, this invention provides a remote heat meter data calibration method based on deep learning, comprising the following steps: We construct a resource-aware optimization framework that maximizes information retention, quantifies the degree of information retention through mutual information estimation technology, and achieves joint optimization of network architecture parameters and model weight parameters. An adaptive wavelet preprocessing engine is constructed to dynamically select the optimal wavelet decomposition level based on hardware capabilities, thereby achieving data dimensionality reduction and key feature extraction. Construct a neural architecture search framework with differentiable resource constraints to automatically generate the optimal network architecture that meets device resource limitations; Implement a variable precision calculation mechanism by adopting a differentiated precision allocation strategy based on the precision sensitivity of each layer of the neural network; implement a task decomposition strategy to break down the complex calibration task into multiple simple subtasks that can be parallelized. Implement an incremental model distillation process, using knowledge distillation technology to transfer complex model knowledge from high-performance devices to lightweight models on low-performance devices; The optimization objective of the resource-aware optimization framework that maximizes information retention is: , This represents the mutual information between the input data and the calibration model output. Represents network architecture parameters, Represents the model weight parameters. Represents a hardware characteristic description vector. Indicates in hardware Running architecture Required computing resources Hardware Maximum computing resource limitations.

[0007] Preferably, the adaptive wavelet preprocessing engine includes: The multi-scale wavelet decomposition algorithm is used to apply discrete wavelet transform to the raw heat measurement data to generate coefficient matrices of different frequency components. A hardware adaptive decomposition level selector is used to dynamically determine the optimal wavelet decomposition level based on the hardware characteristic description vector; The frequency selectivity coefficient retention algorithm is used to score the importance of wavelet decomposition coefficients, retain key frequency components, and filter out non-key information. An adaptive reconstruction mechanism is used to reconstruct signal features based on the retained key coefficients and output a dimension-reduced feature representation.

[0008] Preferably, the neural architecture search framework with differentiable resource constraints includes: Architecture search space, including various network layer types, connection patterns, and hyperparameter options; Hardware-aware search strategy, integrating resource consumption models with architecture search processes to enable architecture search under resource constraints; Gradient-based architecture optimization, converting discrete architecture choices into continuous weight parameters and optimizing architecture parameters through gradient descent methods; Architecture adaptability verification module, testing architecture performance on different hardware configurations to ensure resource constraints and optimal performance.

[0009] Preferably, the variable precision computer mechanism includes: Hierarchical sensitivity analysis algorithm, evaluating the sensitivity of each layer of the neural network to quantization and precision reduction; Adaptive quantization scheme, assigning different calculation precisions to different network layers based on sensitivity analysis results; Dynamic precision adjustment algorithm, adjusting the calculation precision of each layer according to real-time hardware load conditions; Mixed precision training framework, considering the impact of different precisions during the training phase to ensure stable operation of the model under different precision configurations.

[0010] Preferably, the task decomposition strategy includes: Task dependency analysis algorithm, constructing a task dependency graph and identifying groups of tasks that can be executed in parallel; Dynamic task allocation algorithm, determining the granularity and allocation method of task decomposition based on device computing resources and current load status; Subtask-specific model, constructing a specialized lightweight model for different subtasks; Subtask result fusion mechanism, integrating the processing results of each subtask into the final calibration result.

[0011] Preferably, the progressive model distillation process includes: Building a teacher model, training a complex and high-precision calibration model in an environment with sufficient computing resources; Building a resource-aware student model family, constructing lightweight models that meet resource constraints for different hardware platforms; Implementing feature-level distillation, distilling knowledge from both the model output layer and intermediate feature layers; Building a progressive distillation pipeline, gradually transferring knowledge from complex models to simple models through multi-level distillation.

[0012] Preferably, the multi-scale wavelet decomposition algorithm uses db4 wavelet from the Daubechies wavelet family as the mother wavelet, which has time-frequency localization characteristics and orthogonality, and is suitable for processing non-stationary signal features in heat metering data.

[0013] Preferably: the hardware-aware search strategy adopts a resource-efficiency-aware reward function: , wherein is an adaptive weight function in the form of a piecewise function, which dynamically adjusts the reward weight according to the resource utilization.

[0014] Preferably: the sub-task result fusion mechanism adopts an adaptive Bayesian weighting method, which comprehensively considers the confidence and historical accuracy of each sub-task model, and realizes dynamic weight distribution through a small meta-learning network.

[0015] The beneficial effects of the present application are: The present application realizes a unified calibration solution in a heterogeneous hardware environment, and the same calibration algorithm can adapt to a wide range of deployment environments from high-performance edge devices to low-power embedded systems, realizing the goal of "one training, multiple deployment", and significantly reducing development and maintenance costs; Through wavelet decomposition technology, the dimension of input data is significantly reduced, and test results show that the data dimension is reduced by 92%, while the information essential to calibration is retained, greatly improving the processing efficiency of subsequent deep learning models; Through hardware-aware neural architecture search technology, the optimal network architecture is automatically generated for devices with different computing capabilities, and test results show that the floating-point operation amount is reduced by 78%, while the calibration accuracy is maintained; Through the variable precision computing mechanism and task decomposition strategy, efficient use of computing resources is realized, and test results show that memory occupancy is reduced by 65%, and computing delay is significantly reduced; The progressive model distillation technology effectively transfers the knowledge of complex models to simple models, so that even on the most resource-constrained devices, the calibration accuracy can still reach more than 95% of the original complex model; The entire system exhibits resource-aware self-optimization capabilities, and can dynamically adjust the computational complexity according to device load, providing stable and reliable calibration performance in both resource-constrained and resource-abundant situations; The model deployment and update process is significantly simplified, and there is no need to develop and maintain models for each type of device, reducing system maintenance costs and technical barriers. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 is a flowchart of the method of the present application; Figure 2 is a bar chart comparing the effect of wavelet decomposition on reducing data dimension on different devices; Figure 3 is a bar chart comparing the effect of neural architecture search on reducing floating-point operation amount; Figure 4is a column chart of the memory occupation optimization effect of the variable precision computer mechanism; Figure 5 is a column chart of the calculation delay optimization effect of the task decomposition strategy. DETAILED DESCRIPTION

[0017] The subject matter described herein will now be discussed in connection with example embodiments. It should be understood that the discussion of these embodiments is merely meant to provide a better understanding of the subject matter described herein and can include changes, modifications, or additions of elements to the functions and arrangements of the elements discussed without departing from the scope of the present description. Various examples can omit, substitute, or add various procedures or components as appropriate, and the methods described can not necessarily be performed in the order described. Also, some of the described features can be utilized independently of others.

[0018] Before the detailed description of the present application is described, in order to help the understanding of the technical scheme of the present application, the following first explains the terms to be used in the present application: Remote heat meter: refers to a metering device installed in the heat supply pipe network, which can measure the water flow and inlet and outlet water temperature in real time, calculate the heat consumption and remotely transmit data; Wavelet transform: a signal processing technique that can provide local characteristics of a signal in both time and frequency domains, suitable for analyzing non-stationary signals; Neural architecture search (NAS): a technique for automatically designing neural network architecture, which searches for the optimal network structure in a predefined architecture space through a search algorithm; Knowledge distillation: a model compression method that transfers knowledge from a complex model to a simple model by having a small model (student model) imitate the output of a large model (teacher model); Hardware-aware optimization: a technical method for adjusting algorithm implementation according to the specific computing power, memory size and energy consumption characteristics of the deployment hardware to achieve optimal performance; Resource constraints: refer to the limitations of hardware devices in terms of computing power, storage capacity, energy consumption, etc. Mutual information: a measure in information theory that measures the degree of mutual dependence between two random variables, used to evaluate the degree of information retention.

[0019] Application scenario description The deep learning-based remote heat meter data calibration method provided by the present application is applicable to the following application scenarios: Large-scale smart heat supply network: in a heat supply network covering multiple communities or an entire city, different types and different computing power of remote heat meters are installed, and the collected data needs to be calibrated to a unified standard to ensure the accuracy and consistency of the metering data; Heterogeneous edge computing environment: heat meter data processing can occur on a variety of computing devices, from powerful edge servers to limited computing power embedded devices, requiring calibration solutions that adapt to different hardware platforms; Resource-constrained device deployment: many remote heat meters have limited computing resources, and need to implement high-precision data calibration functions without upgrading hardware; Multi-level data processing architecture: heat meter data may need to be processed in multiple levels of systems such as local devices, relay nodes and cloud servers, and each level of device has different computing power, requiring consistent calibration results; Real-time monitoring and analysis system: the heating system management center needs to receive, calibrate and analyze a large amount of heat meter data in real time, and the efficiency and accuracy of the calibration algorithm are required to be high, and the resource fluctuations of the data processing device need to be adapted.

[0020] Embodiments See Figure 1 The present embodiment proposes a remote heat meter data calibration method based on deep learning, which mainly includes the following steps: Step 1: Construct a resource-aware optimization framework that maximizes information retention This step uses information theory to construct an optimization framework to solve the core contradiction between model complexity and resource constraints. According to an embodiment of the present application, the framework is expressed as: Wherein: represents the input data and the mutual information between the output after processing by the calibrated model , quantifying the information retention; represents the network architecture parameters; represents the model weight parameters; represents the hardware characteristic description vector; represents the computing resources required to run the architecture on the hardware ; represents the maximum computing resource limit of the hardware .

[0021] It should be noted that the function in it represents the neural network model determined by the architecture parameters and the weight parameters for the input data The function's specific implementation is a forward propagation calculation of a multi-layer neural network, which sequentially passes input data through various network layers, including convolutional layers, pooling layers, fully connected layers, etc., and finally generates calibrated output results. For the heat meter data calibration task, the function particularly focuses on the time series characteristics and nonlinear relationships of temperature and flow data, capturing these complex patterns through multiple nonlinear transformations.

[0022] In addition, the function The specific implementation of the function is a polynomial model that takes into account the number of network layers, the number of neurons in each layer, the connection method, and the impact of hardware characteristics (such as processor frequency, memory size, cache structure, etc.) on computational resource consumption. The specific form is: where is the number of network layers, is the number of neurons in the th layer, and are coefficients related to the specific operation type, is a hardware performance factor used to adjust the resource consumption differences of different hardware platforms.

[0023] This step specifically includes the following sub-steps: 1.1 Construct a hardware characteristic description vector: Collect the computing power, memory size, energy consumption characteristics, etc. of the deployment device to form a hardware characteristic description vector , which contains processor type, core number, clock frequency, memory capacity, bus bandwidth, etc.

[0024] It should be understood that in some embodiments of the present application, device battery capacity, remaining battery percentage, wireless communication bandwidth, etc. can also be added to the hardware characteristic description vector to more comprehensively describe the characteristics of resource-constrained devices. For remote heat meters and other devices that usually rely on battery power, energy consumption characteristics are particularly important, so the present application gives priority to energy-related parameters.

[0025] 1.2 Establish a resource consumption model: analyze the resource consumption patterns of different network architectures and operations on various hardware, and construct the function , which can accurately predict the computational resource requirements based on network architecture parameters and hardware characteristics.

[0026] According to embodiments of this application, the resource consumption model can be implemented in various forms. Optionally, an analytical model based on computational and storage requirements can be used to directly calculate the floating-point operands (FLOPs) and parameters of the network architecture; alternatively, a regression model based on historical data can be constructed, and the prediction model can be trained using actual measurement data; or, a hardware simulator and performance profiling tools can be combined to accurately simulate the target platform. For the heat meter scenario, the resource consumption model provided in this application pays particular attention to the impact of integer operations, floating-point operations, and memory access patterns on energy consumption. It should be noted that before constructing the resource consumption model, various hardware indicators are standardized, converting indicators with different dimensions (such as CPU frequency, memory size, bus bandwidth, etc.) into uniform standard scores to ensure that different hardware characteristics have reasonable weights in the model.

[0027] 1.3 Definition of Mutual Information Estimation Method: Variational mutual information estimation technique is adopted, and the mutual information is approximated by neural network parameterization. This ensures that information is retained to the maximum extent.

[0028] It should be noted that, since directly calculating mutual information is difficult in practice, this application uses a variational lower bound for approximation. Specifically, an auxiliary network is introduced. , so that: in This represents the output of the calibration model. Parameters are optimized. This method maximizes the variational lower bound, thereby approximately maximizing mutual information. In remote heat meter data calibration tasks, this method effectively preserves key temporal patterns in temperature and flow data, ensuring that the calibration model can capture the essential characteristics of the data.

[0029] 1.4 Constructing a Differentiable Constrained Optimizer: Resource constraints are transformed into differentiable penalty terms. The Lagrange multiplier method is used to transform the constrained optimization problem into an unconstrained optimization problem, enabling optimization of architectural parameters. and model parameters Joint optimization.

[0030] In addition, this application also provides other implementation methods. For example, the augmented Lagrange method can be used to handle the constraints, and the optimization objective can be rewritten as: in It is a Lagrange multiplier. It is the penalty coefficient. Furthermore, heuristic search methods can be combined with gradient descent to form a hybrid optimization strategy, achieving a better balance between search efficiency and optimization quality.

[0031] Step 2: Build an adaptive wavelet preprocessing engine This step uses wavelet transform technology to perform multi-resolution analysis on the input data, dynamically selecting the most suitable decomposition level based on hardware capabilities to achieve signal dimensionality reduction and key feature extraction. Specifically, it includes the following sub-steps: 2.1 Implementation of Multi-Scale Wavelet Decomposition Algorithm: Apply discrete wavelet transform to the original calorific data to generate coefficient matrices for different frequency components, represented as: in: It is the first Wavelet coefficients of level decomposition; These are the raw heat measurement data; The scale is Translation The wavelet basis functions.

[0032] It should be noted that wavelet basis functions The specific implementation is as follows: ,in The mother wavelet function is selected in this application. The db4 wavelet from the Daubechies wavelet family is used as the mother wavelet. This wavelet has good time-frequency localization and orthogonality, and is suitable for processing non-stationary signal features in the heat meter data. The scale parameter controls the "width" of the wavelet; a larger value indicates a lower frequency component in the analysis. These represent the translation parameters, controlling the wavelet's position in the time domain. By adjusting these two parameters, signal characteristics can be analyzed at different time and frequency scales.

[0033] In its implementation, the Daubechies wavelet basis is chosen as the basis function, and efficient computation is achieved through a fast wavelet transform algorithm. To address the time-series characteristics of the calorie meter data, a bidirectional filter bank is employed to achieve accurate time-domain and frequency-domain feature extraction. In practical applications, such as temperature and flow data processing, this algorithm can effectively filter out 50Hz power frequency interference and random noise while preserving key frequency features.

[0034] 2.2 Constructing a Hardware Adaptive Decomposition Level Selector: Based on the hardware characteristic description vector obtained in the first step. Dynamically determine the optimal wavelet decomposition level This ensures that the processing complexity matches the hardware's computing power.

[0035] This selector uses a decision tree model to determine the optimal decomposition level, based on hardware characteristic vectors. The appropriate wavelet decomposition level is determined by the processor type, memory size, and available computing resources. For example, for low-end heat meter devices equipped with an ARM Cortex-M4 processor and less than 256KB of RAM, a decomposition level of 2-3 is selected; for mid-range gateway devices equipped with a quad-core ARM Cortex-A53 or higher processor, a decomposition level of 4-5 is selected; and for edge servers or cloud environments, a decomposition level of 6 or higher can be selected to obtain a more refined feature representation.

[0036] 2.3 Implement a frequency-selective coefficient retention algorithm: score the importance of wavelet decomposition coefficients, retain the frequency components most critical to the calibration task, filter out non-critical information, and generate a simplified representation.

[0037] This algorithm employs a sliding threshold method, combined with information entropy, to evaluate the importance of each frequency band coefficient to the calibration task. For calorimetric data, the low-to-mid frequency bands (e.g., 0.01-0.5Hz) typically contain the main information on flow rate and temperature changes, while high-frequency components (>10Hz) are primarily noise. The algorithm dynamically calculates the information entropy for each frequency band. The algorithm retains frequency band coefficients with information entropy exceeding a set threshold and filters out the remaining coefficients. In practical implementation, for battery-powered meter scenarios, this algorithm can compress the original data volume to 8% of its original size, significantly reducing the computational burden of subsequent processing.

[0038] 2.4 Construct an adaptive reconstruction mechanism: Based on the retained key coefficients, reconstruct the signal features and output a dimension-reduced feature representation that retains the information in the original signal that is crucial for calibration.

[0039] The reconstruction mechanism employs a wavelet packet reconstruction algorithm, adaptively selecting reconstruction methods of varying complexity based on the computational capabilities of different hardware platforms. For resource-constrained devices, an approximate reconstruction algorithm is used, sacrificing a small amount of accuracy for higher efficiency; for high-performance devices, an exact reconstruction algorithm is employed to ensure complete recovery of signal details. In practical applications, such as remote heat network monitoring systems, this mechanism can automatically adjust the reconstruction algorithm based on fluctuations in network bandwidth and terminal device computing power, ensuring stable output of feature representations that meet calibration requirements under various conditions.

[0040] Figure 2 This paper presents a comparison of the dimensionality reduction effects of wavelet decomposition on different devices. The figure illustrates the effectiveness of wavelet decomposition technology in reducing the dimensionality of raw calorie meter data on different types of devices. It also shows the percentage reduction in data dimensionality achieved on different hardware platforms, verifying the technical effect of "wavelet decomposition reducing input data dimensionality by 92%" described in step 2 of this patent.

[0041] Step 3: Construct a neural architecture search framework with differentiable resource constraints This step utilizes neural architecture search technology to automatically generate an optimal network architecture that meets device resource constraints, achieving adaptive optimization of the calibration network. According to one embodiment of this application, it specifically includes the following sub-steps: 3.1 Define the architecture search space: Construct an architecture search space that includes various network layer types, connection modes, and hyperparameter options. This space covers components such as convolutional layers, fully connected layers, and attention mechanisms, as well as different activation functions and regularization methods.

[0042] The search space specifically includes: (1) basic operational units: 1D convolution, depthwise separable convolution, dilated convolution, self-attention unit, and fully connected layer; (2) activation functions: ReLU, LeakyReLU, and GELU; (3) regularization methods: Dropout, BatchNorm, and LayerNorm; (4) connection patterns: residual connection, skip connection, and dense connection. For the task of calibrating calorimeter data, the search space is specifically optimized with temporal feature extraction units, including combinations of different convolution kernel sizes (3, 5, 7, 9) and dilation rates (1, 2, 4, 8), to capture pattern changes at different time scales.

[0043] It should be noted that in other embodiments of this application, recurrent units (such as LSTM and GRU) can also be included in the search space to enhance the model's ability to model time-series data. Furthermore, the scope of the search space can be expanded or limited according to the needs of the specific deployment environment. For example, for extremely constrained microcontroller environments, only a combination of linear layers and basic convolutional layers can be considered; while for high-performance edge devices, more complex attention modules and graph neural network components can be introduced.

[0044] 3.2 Constructing a Hardware-Aware Search Strategy: Integrating the resource consumption model from Step 1 with the architecture search process to achieve architecture search under resource constraints. The search objective is formalized as follows: in: It's the architecture. In the validation dataset Performance metrics; constraints ensure that the selected architecture meets hardware resource limitations.

[0045] It should be noted that the function The specific implementation is an architecture performance evaluation function, which evaluates the architecture... In the validation dataset The architecture performance is measured by a comprehensive evaluation of calibration accuracy, inference speed, and generalization ability. Specifically, it takes the following form: , in Indicates calibration accuracy. This represents the root mean square error. This represents the inference latency on the target hardware. , and It is the weighting coefficient that balances these three indicators.

[0046] For calorimeter calibration tasks, accuracy is typically expressed as the average relative error of temperature and flow measurements, and is required to be controlled within ±0.5%. It should be noted that before calculating this performance evaluation function, the three metrics—accuracy, RMSE, and delay—are standardized and mapped to the [0,1] interval to ensure comparability of the metrics when weighted. Standardization uses the Min-Max method, i.e. ,in and These are the minimum and maximum values ​​of the metric in the candidate architecture set, respectively.

[0047] The hardware-aware search strategy employs a surrogate model-assisted reinforcement learning method, using a resource efficiency-aware reward function: ,in It is a balancing factor. In practical applications, three hardware resource consumption prediction models are established for three types of target hardware: MCUs, MPUs, and high-performance processors, and each candidate architecture is evaluated. For example, for the ARM Cortex-M4 processor commonly used in heat meters, this strategy can automatically generate a network architecture that maintains high accuracy and resource efficiency under the conditions of a memory limit of 128KB and a computation limit of 50MOPS.

[0048] It should be pointed out that the reward function inner function This is a resource efficiency term that evaluates the architecture's efficiency in resource utilization. (Function) It is an adaptive weighting function that dynamically adjusts the reward weight based on the architecture's resource utilization. Specifically, it is implemented as a piecewise function: when resource utilization... Below the threshold hour, Lower values ​​are preferred to encourage fuller use of resources; when utilization is at a certain level... When the interval is, Taking the maximum value indicates that resource utilization is in the optimal range; when the utilization rate is higher than the maximum value, the resource utilization rate is considered optimal. hour, The value decreases as utilization increases to avoid approaching resource limits too closely, which could lead to deployment risks. This setting is typically used in heat meter calibration systems. , This balances performance with resource sufficiency.

[0049] Optionally, this application also provides a search strategy based on evolutionary algorithms to optimize network architecture by simulating the natural selection process. This method maintains a population of architectures and performs selection, crossover, and mutation operations based on a fitness function (combining performance metrics and resource constraints), evolving generation by generation to produce better architectures. Compared to reinforcement learning methods, evolutionary algorithms are less sensitive to initial conditions and can more easily handle discrete architecture selection.

[0050] 3.3 Implementing gradient-based architecture optimization: A differentiable architecture search method is adopted to transform discrete architecture selection into continuous weight parameters. The architecture parameters are then optimized through gradient descent to improve search efficiency.

[0051] Differentiable architecture search employs a hypernetwork approach, constructing a large network encompassing all possible architectures, with each architecture selection represented as the weight parameters for the corresponding operation. Differentiable optimization of the architecture parameters is achieved through soft weight sharing and the Gumbel-Softmax reparameterization technique. For the calorimeter calibration task, a two-stage training process is designed: the first stage optimizes the architecture parameters while keeping the network weights fixed; the second stage fixes the architecture parameters and optimizes the network weights. When deployed on low-power edge devices, this method reduces search time by 65% ​​while ensuring that the found architecture meets the performance requirements of real-time calibration.

[0052] In another embodiment of this application, to further improve search efficiency, a progressive search strategy can be adopted, first quickly finding a coarse architecture in a small search space, and then refining and optimizing it based on this. Furthermore, transfer learning techniques can be utilized to start from existing high-performance architectures and, through knowledge transfer and adaptive modifications, quickly obtain new architectures that meet resource constraints, significantly reducing search time and computational resource consumption.

[0053] 3.4 Construct an architecture adaptability verification module: Through simulation tests on different hardware configurations, verify the actual performance of the searched architecture on the target device, and ensure that optimal performance is achieved while meeting resource constraints.

[0054] The verification module uses an LLVM-based neural network compiler front-end to convert the searched architecture into executable code for the target hardware, and measures the actual execution time, memory usage, and energy consumption using a hardware simulator. A dedicated test set was established for the calibration of heat meter data, including calibration cases under different temperature conditions, flow ranges, and interference modes, ensuring stable operation of the model under various real-world conditions. For example, in a smart heating project, this module successfully verified that the automatically generated model can run in real time at 10Hz on the STM32F4 microcontroller of the heat meter, while meeting the calibration accuracy requirement of ±0.5%.

[0055] It should be noted that this application can also integrate an automatic correction mechanism into the verification module. When it is found that the architecture does not meet the requirements on specific hardware, it can automatically make local adjustments or re-search, forming a closed-loop optimization system. In addition, the verification module can also collect operational data and feed it back into the search strategy to continuously improve the accuracy of the resource consumption prediction model and enhance the efficiency and robustness of the entire architecture search system.

[0056] Figure 3 This paper demonstrates a comparison of the effectiveness of neural architecture search in reducing floating-point operations. The figure illustrates the effect of the differentiable resource-constrained neural architecture search framework in step 3 of this patent on reducing floating-point operations on different devices, verifying the technical effect of "architecture search reducing floating-point operations by 78%". It also compares the differences in computational resource consumption on different computing platforms between using traditional architectures and architectures automatically generated using neural architecture search.

[0057] Step 4: Implement variable precision calculation mechanism This step optimizes computational resource utilization by precisely analyzing the precision sensitivity of different layers in the network and implementing a differentiated processing strategy that maintains high precision for critical layers and reduces computational precision for non-critical layers. Specifically, it includes the following sub-steps: 4.1 Construct a hierarchical sensitivity analysis algorithm: Evaluate the sensitivity of each layer of the neural network to quantization and precision reduction, and quantize the sensitivity index.

[0058] The algorithm first obtains the precision change of each layer in the neural network, which represents the difference in numerical representation when reducing from high-precision computation (e.g., 32-bit floating-point) to the target precision (e.g., 8-bit integer). Then, for each sample in the test dataset, the absolute value of the change in model output error due to the reduction in the precision of that layer is calculated. Next, the ratio of the absolute value of the error change to the precision change for each test sample is calculated; this ratio reflects the degree of error change caused by a unit change in precision. To ensure the comparability of sensitivity metrics between different layers, the original sensitivity metrics of each layer are Z-score standardized, i.e., subtracting the mean and dividing by the standard deviation, converting the sensitivity of each layer into a standard score under a standard normal distribution. Finally, the standardized ratios over all test samples are averaged to obtain the sensitivity metric value for that layer. The higher the metric value, the more sensitive the layer is to the reduction in precision; therefore, high-precision computation should be prioritized during precision optimization. In this way, the key layers in the neural network that are most sensitive to precision can be identified, providing a basis for subsequent differentiated precision allocation.

[0059] In the calibration network of a calorimeter, analysis using this algorithm revealed that the input and output layers are typically the most sensitive to accuracy because they directly process the raw data and produce the final results. The intermediate hidden layers, especially those primarily performing feature transformations, have relatively lower sensitivity. In practice, the sensitivity index of each layer is compared to a preset threshold. Based on the comparison results, the network layers are categorized into high-sensitivity, medium-sensitivity, and low-sensitivity classes, each with different accuracy configurations.

[0060] 4.2 Implementing an adaptive quantization scheme: Based on the sensitivity analysis results, different computational precisions are assigned to different layers in the network. High-precision computation is maintained for critical layers with high sensitivity, while low-precision computation is used for non-critical layers with low sensitivity.

[0061] The quantization scheme employs a mixed-integer linear programming algorithm, with the optimization objective being to minimize overall memory usage. The constraint is to maintain calibration accuracy within a preset threshold (typically 0.5%). For heat meter calibration tasks, the scheme typically uses 32-bit floating-point calculations for the input and output layers, 16-bit floating-point calculations for the feature extraction layer, and 8-bit fixed-point calculations for the intermediate processing layer. This scheme is specialized for different hardware platforms. For example, for ARM processors supporting FP16, half-precision floating-point operations are prioritized; for low-end MCUs that only support integer operations, a fixed-point method simulating floating-point calculations is used. In a city heating project, this scheme reduced the memory usage of the meter calibration model from 4.2MB to 1.5MB while maintaining calibration accuracy.

[0062] 4.3 Construct a dynamic precision adjustment algorithm: Based on the real-time hardware load, dynamically adjust the calculation precision of each layer to maximize resource utilization efficiency while ensuring calibration accuracy.

[0063] This algorithm employs a dynamic scheduling strategy based on load prediction, monitoring metrics such as device CPU utilization, memory usage, and battery level. It dynamically switches between different precision configurations according to a predefined strategy table. For example, when the device battery level is detected to be below 20%, it automatically switches to a low-precision mode to save energy; when a high-priority task is detected running, the precision of the calibration calculation is temporarily reduced to free up resources. Considering the periodic operation of heat meters, the algorithm implements time-window-based resource planning, reserving sufficient resources during peak data collection periods and performing more complex calibration calculations during off-peak periods.

[0064] 4.4 Implementing a mixed-precision training framework: The impact of different precisions is considered during the model training phase, and mixed-precision training technology is adopted to ensure that the model can work stably under different precision configurations after deployment.

[0065] The hybrid precision training framework combines quantization-aware training and precision adaptive regularization techniques to simulate model behavior under different precision conditions during training. Specific implementations include: (1) using quantization operations to simulate low-precision calculations during forward propagation; (2) using a straight-through estimator to handle non-differentiable quantization operations during backpropagation; and (3) adding a precision adaptive loss term to penalize the model's sensitivity to precision changes. In the training of a calorimeter calibration model, this framework can generate robust models that simultaneously adapt to 32-bit, 16-bit, and 8-bit precision environments, maintaining consistent calibration results on devices with different computing capabilities, greatly simplifying the cross-platform deployment process.

[0066] Figure 4 The figure demonstrates the memory usage optimization effect of the variable precision calculation mechanism. It illustrates the optimization effect of the variable precision calculation mechanism implemented in step 4 of this patent on model memory usage, verifying the technical effect of "65% reduction in memory usage through precision optimization." The figure also compares the memory usage differences between using uniform precision across all network layers and using a layer-sensitivity adaptive precision allocation strategy.

[0067] Step 5: Implement the task decomposition strategy This step decomposes the complex task of calibrating calorific meter data into multiple simple, parallelizable subtasks, reducing the computational complexity of individual tasks and improving overall processing efficiency. According to one embodiment of this application, the specific sub-steps include: 5.1 Construct a task dependency analysis algorithm: Analyze the dependencies between subtasks in the process of calorie meter data calibration, construct a task dependency graph, and identify task groups that can be executed in parallel.

[0068] This algorithm uses a directed acyclic graph (DAG) model to represent the subtasks and their dependencies in the heat meter calibration process. Through static analysis and historical execution data mining, it automatically identifies critical paths and parallelizable subtasks. For heat meter data calibration, a typical decomposition includes five main subtasks: data preprocessing, anomaly detection, temperature compensation, flow correction, and result fusion. Analysis revealed that temperature compensation and flow correction have no direct dependency and can be executed in parallel. Applying this algorithm to a regional heating network monitoring system reduced calibration calculation latency by 42% while achieving more balanced resource utilization.

[0069] It should be noted that in other embodiments of this application, task dependency analysis can also be dynamically adjusted based on data characteristics. For example, it can adaptively decide whether to execute certain subtasks based on data quality and signal strength. When high noise is detected in the temperature sensor data, an additional filtering subtask can be added; while when the flow data quality is good, the flow correction steps can be simplified, further optimizing the calculation process.

[0070] 5.2 Implement a dynamic task allocation algorithm: Based on the device's computing resources and current load status, dynamically determine the granularity and allocation method of task decomposition. For high-performance devices, adopt finer-grained parallel processing, and for low-performance devices, adopt coarser-grained serial processing.

[0071] Dynamic task allocation employs a work-stealing scheduling strategy combined with a resource-aware task merging mechanism. The system maintains a task priority queue, describing the priority vectors based on hardware characteristics. The system dynamically adjusts task granularity and parallelism. For example, for a quad-core ARM processor, the calibration calculation is broken down into 12-16 fine-grained tasks, executed in parallel via a thread pool; while for a single-core MCU, these are merged into 3-4 coarse-grained tasks, reducing task switching overhead. When deployed in a real-world heat meter network, this algorithm adapts to various hardware platforms, from single-core STM32 microcontrollers to 8-core gateway servers, achieving optimal utilization of computing resources.

[0072] Optionally, this application also provides a task allocation strategy based on energy consumption optimization. For battery-powered heat meter devices, this strategy dynamically adjusts task allocation according to the battery's state of charge. When the battery is fully charged, parallel processing is used to improve response speed; when the battery is insufficient, serial processing is used to reduce power consumption. Furthermore, hierarchical scheduling can be implemented based on task importance, prioritizing critical sub-tasks that have the greatest impact on calibration accuracy to ensure core functionality under resource constraints.

[0073] 5.3 Constructing Subtask-Specific Models: Constructing a dedicated lightweight model for each subtask. Compared to the general model, these dedicated models have higher computational efficiency and accuracy for their respective tasks.

[0074] For different sub-tasks of heat meter calibration, dedicated neural network models were constructed: (1) a lightweight LSTM autoencoder model was used for the anomaly detection task, focusing on identifying abnormal patterns in time-series data; (2) a multilayer perceptron model was used for the temperature compensation task, correcting the temperature effect through nonlinear mapping; and (3) a convolutional neural network model was used for the flow correction task, extracting the time-frequency features of the flow data. Each dedicated model was optimized through the neural architecture search framework with differentiable resource constraints in step 3, and its structure was customized for specific task characteristics. In a smart heating network project, this dedicated model method reduced computational resource consumption by 65% ​​compared to a single general model, while improving calibration accuracy.

[0075] In another embodiment of this application, a model-sharing mechanism can be introduced, allowing models for different sub-tasks to share the underlying feature extraction network, using task-specific parameters only in the output layer. This method can further reduce the overall number of parameters and computational cost, making it particularly suitable for environments with extremely limited resources. Furthermore, sub-models can be dynamically combined or split based on the correlation between sub-tasks in historical data, forming a more flexible model structure that improves resource utilization efficiency while maintaining accuracy.

[0076] 5.4 Implement a subtask result fusion mechanism: Construct a weighted fusion algorithm to integrate the processing results of each subtask into the final calibration result. The weight coefficients are dynamically adjusted according to the reliability and importance of each subtask model.

[0077] The result fusion mechanism employs an adaptive Bayesian weighted method, comprehensively considering the confidence level and historical accuracy of each sub-task model. The fusion process is achieved by weighted summation of the outputs of each sub-task model. Specifically, first, the processing result of each sub-task model on the current input data is obtained; then, the corresponding weight coefficients are calculated based on the reliability and adaptability of each model to the current input data; finally, the outputs of each model are multiplied by their respective weight coefficients and summed to obtain the final calibration result. This weighted fusion method fully utilizes the strengths of different sub-task models while reducing the potential impact of errors from individual models, thus improving the accuracy and stability of the overall calibration results.

[0078] The model weight calculation considers model confidence, historical accuracy, and current input characteristics, and achieves dynamic weight allocation through a small meta-learning network. In actual heat metering systems, this fusion mechanism can effectively handle the performance fluctuations of each sub-model under different operating conditions and maintain stable calibration accuracy. For example, it automatically increases the weight of the temperature compensation model in low-temperature environments and increases the weight of the flow correction model when flow fluctuations are large, thus adapting to different operating conditions.

[0079] It should be noted that the weighting function is specifically implemented as an adaptive mapping function based on input features and model confidence. This function first extracts key feature vectors from the input samples using a lightweight neural network, and then calculates the final weights by combining the historical performance metrics of each sub-task model with the current prediction confidence estimate. The specific calculation process is as follows: First, feature representations are obtained from the input samples through a feature extraction network. Then, the feature is multiplied by the learnable parameter vector of each sub-task to obtain a feature relevance score. Next, this score is combined with the historical performance metrics and the current confidence estimate through a weighted summation, where the weight coefficients balance the importance of historical performance, input feature relevance, and current confidence, respectively. Finally, the softmax function is used to convert the combined score into a weight distribution with a sum of 1. This is represented as: in, , and These are weighting coefficients that balance historical performance, input feature relevance, and current confidence. Before calculating the weights, all input parameters undergo data preprocessing: historical performance metrics... Normalized to the [0,1] interval using Min-Max; feature correlations are normalized using L2 norm after vector dot product; confidence estimation. The entropy or probability distribution variance of each sub-task model output is calculated and mapped to the [0,1] interval using the sigmoid function, reflecting the model's degree of certainty in predicting the current situation. This standardized preprocessing ensures that parameters with different properties and dimensions can be effectively integrated, preventing any single factor from excessively dominating the weight calculation due to its large numerical range. For calorific meter data calibration, this dynamic weight allocation mechanism can effectively address the differences in sub-task model performance under different operating conditions (such as high / low flow, high / low temperature, etc.), improving overall calibration stability.

[0080] It should be understood that this application also provides other result fusion methods. For example, the Dempster-Shafer fusion method based on evidence theory can be used to better handle uncertainty; or a voting mechanism can be used to adopt a majority decision strategy for outlier detection results. In addition, the fusion process can also consider temporal continuity, combining previous calibration results for smoothing, reducing the impact of single calibration errors on the final result, and improving overall calibration stability.

[0081] Figure 5 The graph illustrates the optimization effect of the task decomposition strategy on calibration computation latency. It compares the computation latency of using a single holistic model with that of using the task decomposition strategy under tasks of varying complexity, verifying the technical effectiveness of "reducing calibration computation latency by 42%".

[0082] Step 6: Implement the progressive model distillation process This step utilizes knowledge distillation technology to transfer complex model knowledge from high-performance devices to lightweight models on low-performance devices, achieving the goal of "train once, deploy everywhere." According to the embodiments provided in this application, it specifically includes the following sub-steps: 6.1 Constructing a Teacher Model: In an environment with sufficient computing resources, a complex and high-precision calibration model is trained as a teacher model. This model does not consider resource limitations and aims to achieve the highest calibration accuracy.

[0083] The teacher model employs a multi-stage deep learning architecture, including a Transformer-based temporal feature extractor, a deep convolutional network with residual connections, and a multi-head attention mechanism. It integrates the predictions from multiple base models through ensemble learning. Trained on server-grade GPUs, the model is approximately 500MB in size, contains about 50 million parameters, and achieves a calibration accuracy of ±0.2%. Specifically designed for the characteristics of heat meter data, the teacher model is particularly robust to temperature variations, flow fluctuations, and system noise, and its generalization ability is improved through extensive data augmentation and adversarial training techniques. In a real heating system, this teacher model is deployed on a central server to process calibration tasks from across the network and serves as the source model for knowledge distillation.

[0084] Furthermore, this application provides an implementation method for integrating teacher models. In this method, instead of training a single large and complex model, multiple models with different architectures or training strategies are trained, and their predictions are then integrated into the final output. This approach captures different perspectives and features of the data, further improving calibration accuracy and robustness, and providing a more comprehensive knowledge source for subsequent knowledge distillation.

[0085] 6.2 Constructing a family of resource-aware student models: Based on the results of the neural architecture search framework with differentiable resource constraints in step 3, a series of student models that meet resource constraints are constructed for different hardware platforms. These models have different computational complexity and memory requirements.

[0086] The resource-aware student model family adopts a hierarchical design, divided into three levels according to the computing power of the target hardware: (1) Edge server level model: model size approximately 50MB, number of parameters approximately 5 million, suitable for regional concentrator devices; (2) Gateway level model: model size approximately 5MB, number of parameters approximately 500,000, suitable for relay gateway devices; (3) Terminal level model: model size approximately 500KB, number of parameters approximately 50,000, suitable for the built-in MCU of the heat meter. Each level of model generates its basic structure through the neural architecture search framework of differentiable resource constraints in step 3, and optimizes it for the instruction set and accelerator characteristics of specific hardware platforms. For example, the terminal level model is specifically optimized for integer arithmetic operations to adapt to MCUs without floating-point units; the gateway level model makes full use of ARM NEON SIMD instructions to accelerate computation; and the edge server level model optimizes multi-core parallel processing capabilities.

[0087] It should be noted that in some embodiments of this application, the student model can be constructed using a modular design approach. Each model consists of multiple functional modules, such as a feature extraction module, a temporal modeling module, and a decision module. Devices with different computing capabilities can selectively activate or replace these modules to achieve a more refined resource-accuracy balance. Furthermore, the student model can be specifically optimized according to actual application scenarios, such as constructing model variants of varying complexity for different operating modes during heating and non-heating seasons.

[0088] 6.3 Implement feature-level distillation: not only distill the knowledge of the model output layer, but also the representation knowledge of the intermediate feature layers. By minimizing the distance between the intermediate features of the teacher model and the student model, the knowledge transfer effect is enhanced.

[0089] Feature-level distillation employs a multi-objective optimization approach, simultaneously considering soft label matching at the output layer and feature representation matching at intermediate layers. The distillation loss function is constructed through a weighted combination of three main components: first, the KL divergence loss between the output layers of the teacher and student models is calculated to measure the difference in their output distributions; next, the distance loss between the teacher and student models at selected feature layers is calculated to measure the similarity of intermediate representations; finally, a task-specific loss based on the ground truth labels is added to ensure that the student model learns the teacher model's knowledge while maintaining its original task performance. These three losses are combined with different weighting coefficients to form the final optimization objective. During training, by minimizing this comprehensive loss function, the student model learns both the teacher model's judgments at the output layer and the structured knowledge of its internal feature representations, thus maintaining calibration performance as much as possible with significantly reduced computational resources.

[0090] For calorimeter calibration tasks, a feature alignment mechanism based on temperature and flow conditions was specifically developed to ensure that the student model can accurately mimic the behavior of the teacher model under various operating conditions. In practical applications, this technology enables the terminal-level model to maintain over 95% calibration accuracy even with only 1 / 1000th the computational resources of the teacher model.

[0091] It should be noted that the KL divergence loss in the distillation loss function is implemented as follows: First, the output distributions of the teacher and student models are obtained. Then, for each output element, the product of the teacher model's output probability and the ratio of its logarithm to the student model's output probability is calculated. Finally, the results of all elements are summed to obtain the KL divergence value. To enhance knowledge transfer, a temperature coefficient is typically applied to soften the teacher model's output. Specifically, the original output is divided by a temperature parameter before being processed using a softmax function. This operation smooths the output distribution, making it easier for the student model to learn the subtle pattern differences captured by the teacher model. For calorimeter calibration tasks, the temperature parameter is usually set to 3-5, which effectively smooths the output distribution and makes it easier for the student model to learn the subtle pattern differences captured by the teacher model.

[0092] The specific implementation of the feature layer distance loss is as follows: First, corresponding intermediate layer features are selected from the teacher model and the student model. Since the architecture and feature dimensions of the two models may differ, they are transformed to the same representation space using an adaptive layer mapping function. Then, the squared Euclidean distance between the transformed features is calculated as a measure of the difference in feature representation. In the calorimeter calibration model, special attention is paid to the matching of key layer features in temperature and flow data processing to ensure that the student model can inherit the teacher model's ability to process these key parameters. The overall distillation loss is expressed as: in, It is a set of feature layers. , and These are coefficients that control the weights of each loss term. To ensure the comparability of the numerical ranges of different loss terms, each loss term is first standardized before combining them: for KL divergence loss... After adjusting the temperature parameter, normalization is performed by dividing by the output dimension; for feature layer distance loss... Normalize by dividing the Euclidean distance by the square root of the feature dimension; apply a task-specific loss. Based on the specific type of loss function, an appropriate normalization method is used (e.g., dividing the MSE loss by the variance of the target variable, or dividing the cross-entropy loss by the logarithmic number of categories). This normalization preprocessing ensures that different types of loss terms are comparable in magnitude, and that the weighting coefficients accurately reflect the relative importance of each loss term. These coefficients are typically determined through cross-validation, and in calorimeter calibration tasks, they are generally set to... Between 0.4 and 0.6 Between 0.2 and 0.4 The value should be between 0.1 and 0.3 to strike a balance between knowledge transfer and the performance of the original task.

[0093] Optionally, this application also provides a distillation method based on attention transfer. In addition to directly matching feature representations, it extracts and transfers the attention map of the teacher model to guide the student model to focus on key regions in the input data. This method is particularly suitable for processing time-series data in calorimeters, helping the student model efficiently capture feature changes at key time points, such as sudden changes in flow or temperature jumps. Furthermore, in environments with extremely limited computing resources, a method based on quantized knowledge distillation can be used to directly vectorize the compressed low-precision feature representations, further reducing the resource requirements of the distillation process and inference stage.

[0094] 6.4 Constructing a progressive distillation pipeline: From high-performance equipment models to medium-performance equipment models, and then to low-performance equipment models, construct a multi-stage distillation pipeline to enable knowledge to be transferred step by step, reduce the difficulty of distillation, and improve the efficiency of transfer.

[0095] The progressive distillation pipeline employs a cascaded knowledge transfer mechanism, rather than directly distilling from the most complex model to the simplest. The distillation process is as follows: teacher model → edge server-level model → gateway-level model → terminal-level model, with intermediate training objectives and validation criteria at each stage. Furthermore, a course-based learning strategy is introduced, starting with simple samples and gradually increasing the proportion of complex samples to help student models master complex knowledge. In a large-scale smart heating project, this distillation pipeline successfully transferred knowledge from the server-level model to thousands of terminal heat meters, achieving a unified calibration standard and significantly reducing model deployment and update costs. Compared to direct distillation, the progressive method improved terminal model accuracy by approximately 3 percentage points and reduced distillation training time by approximately 40%.

[0096] Furthermore, it should be noted that this application also provides a bidirectional distillation mechanism. Under this mechanism, not only is knowledge transferred from the teacher model to the student model, but the advantages of the student model in specific scenarios can also be passed back to the teacher model, forming a knowledge loop. For example, the model on the terminal device may be more adaptable to specific local operating conditions, and this experience can be transmitted back to the upper-level model through the backdistillation mechanism, improving the overall system performance. In addition, the progressive distillation pipeline can also be combined with online learning to support the continued optimization of the model after deployment, adapting to the specific needs and changes of its respective environment.

[0097] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for calibrating remote heat meter data based on deep learning, characterized in that, Includes the following steps: We construct a resource-aware optimization framework that maximizes information retention, quantifies the degree of information retention through mutual information estimation technology, and achieves joint optimization of network architecture parameters and model weight parameters. An adaptive wavelet preprocessing engine is constructed to dynamically select the optimal wavelet decomposition level based on hardware capabilities, thereby achieving data dimensionality reduction and key feature extraction. Construct a neural architecture search framework with differentiable resource constraints to automatically generate the optimal network architecture that meets device resource limitations; A variable precision calculation mechanism is implemented, employing a differentiated precision allocation strategy based on the precision sensitivity of each layer in the neural network; Implement a task decomposition strategy to break down complex calibration tasks into multiple simple subtasks that can be parallelized; Implement an incremental model distillation process, using knowledge distillation technology to transfer complex model knowledge from high-performance devices to lightweight models on low-performance devices; The optimization objective of the resource-aware optimization framework that maximizes information retention is: ， This represents the mutual information between the input data and the calibration model output. Represents network architecture parameters, Represents the model weight parameters. Represents a hardware characteristic description vector. Indicates in hardware Running architecture Required computing resources Hardware Maximum computing resource limitations.

2. The remote calorific meter data calibration method based on deep learning according to claim 1, characterized in that, The adaptive wavelet preprocessing engine includes: The multi-scale wavelet decomposition algorithm is used to apply discrete wavelet transform to the raw heat measurement data to generate coefficient matrices of different frequency components. A hardware adaptive decomposition level selector is used to dynamically determine the optimal wavelet decomposition level based on the hardware characteristic description vector; The frequency selectivity coefficient retention algorithm is used to score the importance of wavelet decomposition coefficients, retain key frequency components, and filter out non-key information. An adaptive reconstruction mechanism is used to reconstruct signal features based on the retained key coefficients and output a dimension-reduced feature representation.

3. The remote calorific meter data calibration method based on deep learning according to claim 1, characterized in that, The neural architecture search framework with differentiable resource constraints includes: The architecture search space includes various network layer types, connection modes, and hyperparameter options; Hardware-aware search strategy is used to integrate resource consumption models with the architecture search process to achieve architecture search under resource constraints; Gradient-based architecture optimization is used to transform discrete architecture selection into continuous weight parameters, and then optimize the architecture parameters through gradient descent. The architecture adaptability verification module is used to test the architecture performance on different hardware configurations to ensure that resource constraints are met and optimal performance is achieved.

4. The remote calorific meter data calibration method based on deep learning according to claim 1, characterized in that, The variable precision calculation mechanism includes: A layer-by-layer sensitivity analysis algorithm is used to evaluate the sensitivity of each layer of a neural network to quantization and precision reduction. An adaptive quantization scheme is used to assign different computational precisions to different network layers based on the results of sensitivity analysis. A dynamic precision adjustment algorithm is used to adjust the calculation precision of each layer according to the real-time hardware load. A mixed-precision training framework is used to consider the impact of different precisions during the training phase, ensuring that the model works stably under different precision configurations.

5. The remote calorific meter data calibration method based on deep learning according to claim 1, characterized in that, The task decomposition strategy includes: Task dependency analysis algorithms are used to construct task dependency graphs and identify groups of tasks that can be executed in parallel. A dynamic task allocation algorithm is used to determine the granularity and allocation method of task decomposition based on the device's computing resources and current load status. Subtask-specific models are used to build dedicated lightweight models for different subtasks. The subtask result fusion mechanism is used to integrate the processing results of each subtask into the final calibration result.

6. The remote calorific meter data calibration method based on deep learning according to claim 1, characterized in that, The progressive model distillation process includes: Build a teacher model and train a complex and high-precision calibration model in an environment with sufficient computing resources; Construct a family of resource-aware student models to build lightweight models that meet resource constraints for different hardware platforms; Implement feature-level distillation, simultaneously distilling knowledge from the model's output layer and intermediate feature layers; Construct an incremental distillation pipeline to transfer knowledge from complex models to simpler models through multi-stage distillation.

7. The remote calorific meter data calibration method based on deep learning according to claim 2, characterized in that, The multi-scale wavelet decomposition algorithm uses the db4 wavelet from the Daubechies wavelet family as the mother wavelet. It has time-frequency localization and orthogonality, making it suitable for processing non-stationary signal features in calorie meter data.

8. The remote calorific meter data calibration method based on deep learning according to claim 3, characterized in that, The hardware-aware search strategy employs a resource-efficiency-aware reward function: ， in It is an adaptive weight function in the form of a piecewise function, which dynamically adjusts the reward weight based on resource utilization.

9. The remote calorific meter data calibration method based on deep learning according to claim 5, characterized in that, The subtask result fusion mechanism adopts an adaptive Bayesian weighted method, which comprehensively considers the confidence and historical accuracy of each subtask model, and achieves dynamic weight allocation through a small meta-learning network.

Images