Capacitive pressure transmitter temperature compensation method based on RMEA-DACA-SVR

Abstract

The invention discloses a capacitive pressure transmitter temperature compensation method based on RMEA-DACA-SVR, and the method comprises the steps: 1, collecting corresponding pressure real-time data at different temperatures, and generating a comprehensive data set; 2, preprocessing the acquired data, including noise data removal, data scaling and key feature extraction; 3, decomposing the data by adopting HHT (Hilbert-Huang Transform); 4, a comprehensive temperature compensation model based on the RMEA-DACA-SVR is constructed; rMEA-DACA combination is used for dynamic optimization and real-time compensation, SVR is used for constructing a nonlinear relation between temperature and pressure value output, and MHA and ResNet structures are combined, so that attention to key temperature features is improved, and overfitting and model complexity are reduced; and 5, integrating the compensation results of the RMEA, the DACA and the SVR by adopting a weighted fusion method, and generating high-precision temperature compensation output. According to the invention, the complex compensation requirements of the capacitive pressure transmitter in different temperature environments can be effectively met, and the temperature compensation precision is improved; the method is suitable for the field of industrial automation and high-precision measurement, and has a wide application prospect.

Description

Technical Field

[0001] The present invention relates to the technical field of temperature compensation, and particularly relates to a temperature compensation method for a capacitive pressure transmitter based on RMEA-DACA-SVR. Background Art

[0002] As a commonly used precision measurement device, capacitive pressure transmitters are widely used in high-precision measurement fields such as industrial automation, aerospace, and medical equipment. However, since the change in ambient temperature will directly affect the output signal of the capacitive sensor, resulting in measurement errors, in practical applications, capacitive pressure transmitters often need to use temperature compensation technology to improve their measurement accuracy.

[0003] Currently, temperature compensation technologies mainly include methods based on linear regression, neural networks, time series analysis, etc. These traditional methods perform well in dealing with simple linear or static temperature compensation problems, but for complex non-linear temperature changes or sudden temperature distortions, the performance of traditional methods is relatively limited. Especially in multi-dimensional temperature scenarios, linear compensation models are easily affected by temperature fluctuations, resulting in a decrease in the accuracy of the output results. The temperature compensation method based on linear regression has a simple structure and is difficult to accurately compensate for complex temperature effects in non-linear temperature scenarios; although the compensation method based on neural networks has certain non-linear processing capabilities, due to being prone to overfitting and requiring a large amount of labeled data during the training process, its generalization ability is often insufficient.

[0004] With the development of intelligent algorithms, the combination of optimization algorithms and non-linear models has been increasingly widely used in sensor temperature compensation. Especially the SVR model, with its good non-linear fitting ability and strong generalization performance, has gradually become one of the commonly used methods for non-linear temperature compensation. However, a single SVR model still has deficiencies when facing complex multi-dimensional temperature changes. It depends on the setting of initial parameters and cannot adapt to a rapidly changing temperature environment. SVR relies on fixed parameters optimized offline and cannot adaptively adjust in a dynamic temperature environment, resulting in prediction lag during rapid temperature changes; at the same time, its pure data-driven characteristic makes it difficult to explicitly model the physical deformation of the pressure transmitter caused by temperature, and its ability to extract features of multi-dimensional temperature coupling effects is limited. The high-dimensional mapping of the kernel function is difficult to cover the complex non-linear interactions in extreme temperature regions, ultimately leading to the failure of structural error compensation. In addition, the physical structure of the pressure transmitter may also be distorted due to temperature changes, making it difficult for a single SVR model to comprehensively handle the structural errors caused by temperature. Summary of the Invention

[0005] Aiming at the technical problems of insufficient accuracy and poor adaptability to complex temperature changes in the above-mentioned existing temperature compensation technologies, this technical solution provides a temperature compensation method for capacitive pressure transmitters based on RMEA-DACA-SVR, which deeply integrates three technologies, namely RMEA, DACA, and SVR, to form an efficient comprehensive temperature compensation framework specifically for the problem of poor measurement accuracy of capacitive pressure transmitters in complex temperature environments, effectively solving the above problems.

[0006] The present invention is realized through the following technical solutions:

[0007] A temperature compensation method for capacitive pressure transmitters based on RMEA-DACA-SVR, comprising the steps:

[0008] Step 1: Collect real-time pressure data of capacitive pressure transmitters under different temperature conditions to generate a comprehensive data set;

[0009] Step 2: Preprocess the comprehensive data set, including removing noise data, data scaling, and extracting key features, and decompose the data using the HHT transform;

[0010] Step 3: Build a comprehensive temperature compensation model based on RMEA-DACA-SVR:

[0011] Step 3.1: Introduce a learning and reflection mechanism into the RMEA algorithm. The learning mechanism allows communication and learning of excellent features and parameter settings between different individuals; the reflection mechanism is achieved by comparing the fitness values of individuals, eliminating individuals with lower fitness values, and stimulating the evolutionary vitality of the population;

[0012] Step 3.2: Connect RMEA and DACA to each other through a feedback mechanism to form a closed-loop system; continuously adjust parameter settings, dynamically adjust weights, and re-evaluate to improve the compensation accuracy;

[0013] Step 3.3: Build an improved SVR model, optimize model parameters, select a suitable kernel function, adjust hyperparameters, and combine the multi-head attention mechanism MHA and the residual network ResNet to adapt to temperature compensation in different situations;

[0014] Step 3.4: Combine the global search ability of RMEA with the local correction ability of DACA, and perform weighted fusion with the non-linear temperature-capacitance output relationship established by SVR; compare the comprehensive model with SVR, LSTM, and BP models respectively, and evaluate the comparison results;

[0015] Step 4: Dynamically adjust the weights of REMA, DACA, and SVR according to their performance under different temperature scenarios; in the initial stage, the weights of REMA, DACA, and SVR are equally distributed. As the training iteration progresses, the system monitors the temperature scenario features in real time, combines the error performance on the validation set, evaluates the performance of each model on the validation set, gradually adjusts the weights, and generates the integrated prediction result through the dynamic weighted fusion formula to strengthen the adaptive ability of the model under specific scenarios;

[0016] Step 5: The process of dynamically generating and outputting the temperature compensation result: The system collects the current temperature value T(t) and the original pressure value P(t) in real time, and inputs them in parallel into the RMEA, DACA, and SVR models to generate the global compensation amount ΔP RMEA (t), the local distortion correction amount ΔP DACA (t), and the non-linear mapping compensation amount ΔP SVR (t). Based on the dynamic weights w RMEA , w DACA , w SVR Non-linearly fuse the compensation amounts of the three to generate the final compensation value ΔP final (t) and add it to the original pressure value, and synchronously output the high-precision pressure signal P compensated (t) through 4-20mA analog output and Modbus digital protocol. At the same time, calculate the compensation error e(t) online. When the error exceeds the threshold, trigger the SVR parameter gradient fine-tuning and RMEA population update to form a closed-loop optimization, and use the hardware timer to achieve millisecond-level synchronous output. Combine the limit protection mechanism to ensure the signal stability under extreme working conditions, and finally realize the dynamic generation of the temperature compensation result, multi-protocol output, and adaptive accuracy maintenance.

[0017] Furthermore, the specific method of Step 1 is as follows:

[0018] Step 1.1: Determine the model and technical specifications of the capacitive pressure transmitter for data collection; clarify the key parameters to be monitored, including: pressure value P(t), temperature value T(t), time stamp; configure the data acquisition hardware, select the appropriate acquisition frequency f = 1Hz, and set the acquisition interval Δt:

[0019]

[0020] Step 1.2: Conduct experiments in different temperature environments and record the pressure output under each condition; monitor and record the relationship between P(t) and T(t) to ensure that there is a corresponding P(t) and T(t) at each time point t;

[0021] Step 1.3: Integrate the pressure data under all temperature conditions to form a comprehensive data set D, and check the distribution characteristics of the data; conduct basic statistical analysis on the comprehensive data set and calculate the mean μ P and the standard deviation σ P :

[0022]

[0023] where μ P is the average value of the pressure data; N is the total number of pressure data; P i (t) is the pressure value corresponding to the i-th sample; t is the time variable indicating the specific time point corresponding to each sample; i is an integer starting from 1 and ending at N; σ P is the standard deviation of the pressure data, which is used to measure the degree of fluctuation of the pressure data;

[0024] Step 1.4: Integrate the pressure data under all temperature conditions, store the collected data in the database, and form a comprehensive data set.

[0025] Furthermore, the specific method of step 2 is as follows:

[0026] Step 2.1: Use the Kalman filter to remove the noise data and obtain the denoised estimated value;

[0027] Step 2.2: Adopt the Z-score normalization method for data scaling, convert the data into the form of a standard normal distribution, and ensure that the mean is 0 and the standard deviation is 1; the specific normalization formula is:

[0028]

[0029] where Z i is the normalized data, X i is the original data value, μ is the sample mean, and σ is the sample standard deviation;

[0030] Then use the covariance matrix C for normalization processing:

[0031]

[0032] Perform eigenvalue decomposition on the covariance matrix to construct a standardized data matrix, is the mean of the sample data, and T is the transpose operation;

[0033] Step 2.3: Use PCA for dimensionality reduction and feature extraction of the pressure and temperature data, project the data into the feature space, and calculate the eigenvectors and eigenvalues;

[0034] Step 2.4: Conduct data decomposition, take the mean of each multi-IM component respectively to eliminate the error introduced by noise addition, and obtain the final IMF:

[0035]

[0036] where, IMF j is the average value of the j-th intrinsic mode function component, is the value of the j-th IMF component corresponding to the n-th sample;

[0037] Perform Hilbert transform on each IMF to obtain the instantaneous frequency and envelope amplitude for analyzing the time-frequency characteristics of the data; the Hilbert transform formula is:

[0038]

[0039] where, P.V. represents the principal value integral, H(IMF j )(t1) is the transformation result of the j-th intrinsic mode function at t1, IMF j(τ) is the value of the j-th intrinsic mode function at time τ, τ is the integration variable representing continuous points in time, t1 is the target time point of the Hilbert transform, and dτ is the differential of the integration variable τ; the final decomposition result of HHT, synthesizing the final signal decomposition form:

[0040]

[0041] where, X(t) is the final decomposition result of the original signal, IMF j are the respective IMF components, r is the residual term, and j is an integer starting from 1 and ending at m.

[0042] Furthermore, the specific method of step 3.1 is as follows:

[0043] Step 3.11: Randomly generate the initial population X = {X1, X2, …, X N}, where each individual x i represents a potential parameter setting for the temperature compensation model;

[0044] Step 3.12: Introduce a learning mechanism, and individuals improve their performance by learning from better individuals; assume that individual X i learns from the individual X j with higher fitness, and the update formula is as follows:

[0045] X i (t + 1) = X i (t) + ζ · (X j (t) - X i (t)) + β · η

[0046] Among them, ζ is the learning rate, which determines the step size of learning; β is the noise intensity coefficient, which maintains the population diversity; η is the random noise term, which follows a normal distribution Among them is the mathematical symbol of the normal distribution; X i (t + 1) is the position of the i-th individual at time step t + 1, X i (t) is the position of the i-th individual at the current time step t, X j (t) is the position of the j-th individual at the current time step t;

[0047] Step 3.13: Introduce a reflection mechanism to eliminate individuals with low fitness and introduce new individual configurations; Let the individual with relatively low fitness currently be X k , and the update formula is as follows:

[0048] X k (t + 1) = x best (t) + γ·(X k (t) - X avg (t)) + δ·β

[0049] Among them, X best (t) is the individual with the highest fitness in the current generation, X avg (t) is the average parameter value of the population in the current generation, X k (t + 1) is the optimal individual position at the current time step t, γ is a regulation parameter, which controls the step size of the individual moving according to the difference between the current optimal position and the current individual position, X k (t) is the position of the k-th individual at the current time step t, δ is an adjustment factor, which is used to affect the random change of the individual;

[0050] Step 3.14: In the population update, introduce an enhanced crossover and an adaptive mutation strategy at the same time; During the crossover process, use the fitness value as the crossover weight coefficient;

[0051]

[0052] Among them, X p and X q are two parent individuals, and F(X p ) and F(X q ) are their fitness values respectively, X new is the newly calculated position;

[0053] The newly generated individuals are added to the population, and fitness evaluation is carried out before the next generation starts; Retain the individuals with higher fitness to ensure that the population continuously evolves in a better direction;

[0054] Step 3.15: Repeat the above steps until the termination condition is met; finally, the individual with the highest fitness is selected as the optimal parameter configuration for temperature compensation.

[0055] Further, the specific method of step 3.2 is as follows:

[0056] Step 3.21: During each iteration cycle, DACA receives the initial parameters optimized by RMEA as input, adjusts the compensation parameters in real time, and feeds back the correction results to RMEA; at the beginning of each generation, RMEA calculates the fitness score according to the feedback from DACA, and the feedback value is used to adjust the fitness of individuals to form a preliminary evaluation.

[0057] Step 3.22: During the compensation process, DACA dynamically corrects the pressure value according to the current temperature change; RMEA updates the fitness of individuals in the population by receiving the correction feedback from DACA.

[0058] At the end of each generation, the feedback factor θ is calculated according to the correction effect of DACA and fed back to RMEA to affect its population optimization process:

[0059]

[0060] where E is the total number of feedback cycles, and this factor is used to measure the correction effect of DACA, P observed (t) is the true pressure value observed at time t, P DACA (t) is the pressure value predicted by the DACA algorithm at time t;

[0061] Step 3.23: According to the θ factor fed back by DACA, RMEA adjusts the parameter settings of individuals in the new generation, and the specific formula is:

[0062]

[0063] where, is the learning rate; X best is the individual with the highest fitness in the current generation, X i (t) is the state of the i-th individual at time t;

[0064] Through the adjustment of the feedback factor, RMEA can tend to a higher fitness in the next generation, and at the same time compensate for the temperature influence through the real-time correction of DACA to form a closed-loop optimization.

[0065] Further, the specific method of step 3.3 is as follows:

[0066] Step 3.31: The parameters of the SVR model are the key steps to ensure its best performance in the temperature compensation problem; it mainly involves three important parameters: the penalty parameter L, the kernel function v, and the tolerance error e; the specific optimization process is as follows:

[0067] 3.31.1: Use grid search combined with 5-fold cross-validation for optimization, set multiple candidate values of L, perform cross-validation on the training set, calculate the model performance under different L values, and select the best L value with the minimum validation error;

[0068] 3.31.2: Adopt grid search to select a suitable e, test a series of e values, and select the best e through cross-validation;

[0069] 3.31.3: Select the radial basis kernel function RBF to solve the modeling problem of nonlinear data, which is applicable to most nonlinear data sets:

[0070] K(x,x′) = exp(-v||x - x′|| 2 )

[0071] where K(x,x′) is the kernel function value between the input points x and x′, ν is a positive constant, usually called the width parameter of the kernel function, which controls the decay rate of the function, and ||x - x′|| 2 is the square of the Euclidean distance, representing the difference between these two points in the input space;

[0072] Step 3.3.2: Introduce the multi-head attention mechanism MHA to enable the model to focus on features in different subspaces, linearly transform the input features to obtain the Q, K, and V matrices, select the number of attention heads h according to the data complexity, calculate the attention outputs of each head and splice them, and finally restore the original dimension through linear transformation;

[0073] In the input layer of the SVR model, take the multi-head attention output as the enhanced input feature, further stack the ResNet structure on the enhanced features generated by MHA, and prevent information loss during the feature transmission process through residual connection; the final enhanced feature is used as the input feature and directly input into the input layer of the SVR model for regression prediction.

[0074] Furthermore, the specific method of step 3.4 is as follows:

[0075] Step 3.41: Set the population size to 100 individuals, evolve for 100 generations, the individuals adaptively update by learning high-fitness individuals, and eliminate low-fitness individuals through the reflection mechanism, randomly generate new individuals to maintain diversity, and gradually approach the optimal solution;

[0076] By introducing a feature self-enhancement module, SVR allows the model to learn feature interactions in different dimensions; a multi-layer perceptron (MLP) is introduced at the input layer, and through non-linear transformations of multiple hidden layers, the mapping ability to complex temperature fluctuations and sensor outputs is enhanced;

[0077] Step 3.42: Train the RMEA-DACA-SVR model by combining the outputs of RMEA, DACA, and SVR through a dynamic weighted fusion strategy; at the initial stage of the model, the weights of the three models are set to be equal; as the training process progresses, the model automatically adjusts the weights according to its performance under different temperature conditions; according to the performance of each model, the dynamic adjustment formula is:

[0078] w RMEA +w DACA +w SVR =1

[0079] where w RMEA is the weight of the RMEA model, used to measure the contribution of this model to the final result; w DACA is the weight of the DACA model; w SVR is the weight of the SVR model; each weight is dynamically updated according to its performance in the previous generations of iterations to ensure the optimal output of the combined model;

[0080] Step 3.43: Divide the dataset into multiple scenarios based on the volatility of temperature, and each scenario is divided into a training set, a validation set, and a test set. The training set accounts for 60%, the validation set accounts for 20%, and the test set accounts for 20%. To verify the superiority of the comprehensive model, select the BP model, the LSTM model, and the SVR model as comparison objects; use the root mean square error (MSE), the mean absolute error (MAE), and the coefficient of determination (R 2 ) as evaluation metrics. First, optimize the hyperparameters of each model on the validation set, and then calculate the final performance of each model on the test set; the specific formulas are:

[0081]

[0082] where H is the number of samples, is the model prediction value, and P i is the true value.

[0083] (III) Beneficial Effects

[0084] A temperature compensation method for a capacitive pressure transmitter based on RMEA-DACA-SVR proposed by the present invention has the following beneficial effects compared with the prior art:

[0085] (1) The present invention deeply integrates three technologies, namely RMEA, DACA, and SVR, to form an efficient temperature compensation comprehensive framework specifically for the measurement accuracy problem of capacitive pressure transmitters in complex temperature environments. RMEA is responsible for global search, DACA corrects structural distortions in real time, and SVR accurately captures the relationship between temperature and output through non-linear modeling. The combination of the three forms an efficient and intelligent highly coupled compensation system through multiple rounds of iteration. Specifically, the combination of RMEA and DACA realizes the collaborative work of global optimization and local correction by establishing a feedback closed-loop system. RMEA provides initial compensation parameters through global search, and DACA provides real-time correction when there are structural distortions, and the two work together. On this basis, SVR further constructs an accurate mapping relationship between temperature and capacitance output through non-linear modeling. By modeling the complex non-linear relationship between temperature and output signal, it makes up for the limitation that RMEA and DACA cannot directly handle complex non-linear mappings. The intelligent evolutionary algorithm of RMEA provides dynamic parameter search ability, solves the problem that traditional SVR depends on static initialization, and adapts to drastic temperature fluctuations; DACA directly offsets structural errors through a physical deformation real-time compensation mechanism, making up for the lack of mechanism of pure data-driven models; SVR uses high-dimensional mapping of kernel functions to accurately fit non-linear relationships, and its output and the local correction amount of DACA achieve a leap in compensation accuracy through weighted fusion. The three form a synergistic enhancement effect through closed-loop feedback. After RMEA optimizes global parameters, it triggers the transient response of DACA, and SVR synchronously updates the non-linear model, finally achieving a full-dimensional performance breakthrough in steady-state accuracy, dynamic robustness, and multi-physical field coupling compensation under complex working conditions. This complementary integration enables the comprehensive model to maintain a robust compensation for the global temperature trend while quickly responding to local temperature fluctuations and structural distortions when compensating for the temperature of capacitive pressure transmitters in complex temperature environments. At the same time, SVR accurately models the complex non-linear relationship between temperature and output, significantly improving the compensation accuracy and system adaptability.

[0086] (2) The present invention can not only effectively capture the multi-level compensation requirements of capacitive pressure transmitters in different temperature environments, but also cope with the non-linear errors caused by complex temperature fluctuations and structural distortions. Through global search and optimization, RMEA captures the long-term trends during the temperature change process and provides robust compensation when the temperature changes sharply. DACA responds in a timely manner to the instantaneous changes in the structure through real-time distortion correction. SVR maps the relationship between temperature and output to a high-dimensional space through kernel methods, thereby capturing non-linear characteristics. This process enables the model to perceive minute temperature fluctuations and predict their precise impact on the output value. In this integrated model, RMEA and DACA work together to provide precise global and local adjustment capabilities for the compensation system. The introduction of SVR further enhances the system's ability to capture the complex non-linear relationship between temperature and pressure output, enabling the system to not only cope with long-term trend changes in temperature, but also sensitively perceive local temperature fluctuations and their impact on the output. This method can capture both global and local characteristics in temperature compensation and has the ability to handle complex and multi-dimensional temperature change scenarios. Experimental verification shows that the temperature compensation method based on RMEA-DACA-SVR is significantly superior to traditional methods and single compensation models in terms of the compensation accuracy and system stability of capacitive pressure transmitters, especially showing higher robustness and precision in complex environments. The proposed method provides a new technical means for high-precision compensation of capacitive pressure transmitters in complex temperature environments, with broad application prospects and important practical significance.

[0087] (3) RMEA in the present invention plays a key role in global optimization. By using the learning and reflection mechanisms of the algorithm, it conducts global search for compensation parameters, ensuring that the system can quickly approach the optimal compensation solution and effectively solving the problem of low parameter selection efficiency of traditional methods in multi-dimensional temperature scenarios. At the same time, DACA tracks and dynamically corrects the physical distortions caused by temperature in real time. Through the distortion area compensation mechanism, DACA can meticulously capture the microscopic changes of the structure under different temperature conditions and adjust the output signal in real time, thereby effectively eliminating the influence of temperature changes on the output accuracy. SVR, through non-linear modeling methods, precisely captures the complex non-linear relationship between temperature and sensor output, making up for the deficiency that RMEA and DACA cannot accurately handle complex temperature effects in global and local corrections. The combination of the three realizes the complementary effect of global and local compensation through the complementarity and synergy of features. Experimental results show that the RMEA-DACA-SVR method can effectively reduce the influence of temperature on pressure measurement under various temperature conditions. Whether in extremely low or high temperature environments, this method can maintain high measurement accuracy and effectively eliminate the error fluctuations caused by temperature. This not only improves the application reliability of capacitive pressure transmitters in industrial fields, but also provides technical guarantee for precise temperature compensation, greatly enhancing the stability and robustness of the system. Description of the Drawings

[0088] Figure 1 is the overall process schematic diagram of the present invention.

[0089] Figure 2 is the structural diagram of the non - linear compensation system in the present invention.

[0090] Figure 3 is the structural diagram of RMEA - DACA - SVR in the present invention.

[0091] Figure 4 is the structural diagram of the dynamic optimization of RMEA in the present invention.

[0092] Figure 5 is the characteristic curve graph of the pressure measurement results varying with different temperatures in the present invention.

[0093] Figure 6 is the graph of the pressure values detected after non - linear compensation of RMEA - DACA - SVR in the present invention. Detailed Description of the Preferred Embodiments

[0094] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. The described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Without departing from the design concept of the present invention, various modifications and improvements made by those of ordinary skill in the art to the technical solutions of the present invention shall fall within the protection scope of the present invention.

[0095] Embodiment 1:

[0096] As Figure 1 shown, a temperature compensation method for a capacitive pressure transmitter based on RMEA - DACA - SVR includes the steps:

[0097] Step 1: Collect the real - time pressure data of the capacitive pressure transmitter under different temperature conditions to generate a comprehensive data set;

[0098] Step 1.1: Determine the model and technical specifications of the capacitive pressure transmitter used for data collection; clarify the key parameters to be monitored, including: pressure value P(t), temperature value T(t), time stamp, etc.; configure the data acquisition hardware, select an appropriate acquisition frequency f = 1Hz, and set the acquisition interval Δt:

[0099]

[0100] Step 1.2: Conduct experiments in different temperature environments and record the pressure output under each condition; monitor and record the relationship between P(t) and T(t) to ensure that there are corresponding P(t) and T(t) at each time point t;

[0101] Step 1.3: Integrate the pressure data under all temperature conditions to form a comprehensive data set D, and check the distribution characteristics of the data; conduct basic statistical analysis on the comprehensive data set, and calculate the mean μ P and the standard deviation σ P :

[0102]

[0103] where, μ P is the average value of the pressure data; N is the total number of the pressure data; P i (t) is the pressure value corresponding to the i-th sample; t is the time variable indicating the specific time point corresponding to each sample; i is an integer starting from 1 and ending at N; σ P is the standard deviation of the pressure data, which is used to measure the degree of fluctuation of the pressure data;

[0104] Step 1.4: Integrate the pressure data under all temperature conditions, store the collected data in the database, and form a comprehensive data set.

[0105] Step 2: Preprocess the comprehensive data set, including removing noise data, data scaling, and extracting key features, and decompose the data using the HHT transform; the specific operation steps include:

[0106] Step 2.1: Use the Kalman filter to remove the noise data and obtain the denoised estimated value.

[0107] Step 2.2: Adopt the Z-score normalization method for data scaling, convert the data into the form of a standard normal distribution, and ensure that the mean is 0 and the standard deviation is 1; the specific normalization formula is:

[0108]

[0109] where, Z i is the normalized data, X i is the original data value, μ is the sample mean, and σ is the sample standard deviation.

[0110] Then use the covariance matrix C for normalization processing:

[0111]

[0112] Conduct eigenvalue decomposition on the covariance matrix to construct the normalized data matrix, is the mean of the sample data, and T is the transpose operation.

[0113] Step 2.3: Use PCA to reduce the dimension and extract features for the pressure and temperature data, project the data into the feature space, and calculate the eigenvectors and eigenvalues.

[0114] Step 2.4: Perform data decomposition, take the mean of each multi-IM component respectively to eliminate the error introduced by noise addition, and obtain the final IMF:

[0115]

[0116] where, IMF j is the average value of the j-th intrinsic mode function component, is the value of the j-th IMF component corresponding to the n-th sample.

[0117] Perform Hilbert transform on each IMF to obtain the instantaneous frequency and envelope amplitude for analyzing the time-frequency characteristics of the data; the Hilbert transform formula is:

[0118]

[0119] where, P.V. represents the principal value integral, H(IMF j )(t1) is the transformation result of the j-th intrinsic mode function at t1, IMF j(τ) is the value of the j-th intrinsic mode function at time τ, τ is the integration variable representing continuous points in time, t1 is the target time point of the Hilbert transform, and dτ is the differential of the integration variable τ; the final decomposition result of HHT, synthesize the final signal decomposition form:

[0120]

[0121] where, X(t) is the final decomposition result of the original signal, IMF j are each IMF component, r is the residual term, and j is an integer starting from 1 and ending at m.

[0122] Step 3: Construct a comprehensive temperature compensation model based on RMEA-DACA-SVR, and the structure of the comprehensive temperature compensation model based on RMEA-DACA-SVR is as Figure 3 shown. The specific operation steps for construction include:

[0123] Step 3.1: Introduce a learning and reflection mechanism into the RMEA algorithm. The learning mechanism allows communication and learning of excellent features and parameter settings between different individuals; the reflection mechanism is realized by comparing the fitness values of individuals, eliminating individuals with lower fitness values, and stimulating the evolutionary vitality of the population;

[0124] Step 3.11: Randomly generate an initial population X = {X1, X2, …, X N}, where each individual x i represents a potential parameter setting for the temperature compensation model;

[0125] Step 3.12: Introduce a learning mechanism where individuals improve their performance by learning from better individuals; assume individual X i learns from individual X j with a higher fitness value, and the update formula is as follows:

[0126] X i (t + 1) = X i (t) + ζ · (X j (t) - X i (t)) + β · η

[0127] where ζ is the learning rate, determining the step size of learning; β is the noise intensity coefficient, maintaining population diversity; η is a random noise term, following a normal distribution where is the mathematical symbol of the normal distribution; X i (t + 1) is the position of the i-th individual at time step t + 1, X i (t) is the position of the i-th individual at the current time step t, X j (t) is the position of the j-th individual at the current time step t;

[0128] Step 3.13: Introduce a reflection mechanism to eliminate individuals with low fitness and introduce new individual configurations; assume the individual with relatively low fitness at present is X k , and the update formula is as follows:

[0129] X k (t + 1) = X best (t) + γ · (X k (t) - X avg (t)) + δ · β

[0130] where X best (t) is the individual with the highest fitness in the current generation, X avg (t) is the average parameter value of the population in the current generation, X k (t + 1) is the position of the optimal individual at the current time step t, γ is a regulating parameter, controlling the step size for the individual to move according to the difference between the current optimal position and the current individual position, X k (t) is the position of the k-th individual at the current time step t, δ is an adjustment factor, used to affect the random variation of the individual;

[0131] Step 3.14: In population update, introduce enhanced crossover and adaptive mutation strategies simultaneously; during the crossover process, use the fitness value as the crossover weight coefficient;

[0132]

[0133] where X p and Xq are two parent individuals, and F(X p ) and F(X q ) are their fitness values respectively, and X new is the newly calculated position;

[0134] The newly generated individuals are added to the population, and fitness evaluation is carried out before the next generation; the individuals with higher fitness are retained to ensure that the population evolves continuously in a better direction;

[0135] Step 3.15: Repeat the above steps until the termination condition is met; finally, the individual with the highest fitness is selected as the optimal parameter configuration for temperature compensation; the RMEA dynamic optimization structure is as Figure 4 shown.

[0136] Step 3.2: Connect RMEA and DACA to each other through a feedback mechanism to form a closed-loop system; continuously adjust the parameter settings, dynamically adjust the weights, and re-evaluate to improve the compensation accuracy.

[0137] Step 3.21: In each iteration cycle, DACA receives the initial parameters optimized by RMEA as input, adjusts the compensation parameters in real time, and feeds back the correction result to RMEA; at the beginning of each generation, RMEA calculates the fitness score according to the feedback of DACA, and the feedback value is used to adjust the fitness of individuals to form a preliminary evaluation.

[0138] Step 3.22: During the compensation process, DACA dynamically corrects the pressure value according to the current temperature change; RMEA updates the fitness of individuals in the population by receiving the correction feedback of DACA.

[0139] At the end of each generation, the feedback factor θ is calculated according to the correction effect of DACA and fed back to RMEA to affect its population optimization process:

[0140]

[0141] where E is the total number of feedback cycles, and this factor is used to measure the correction effect of DACA, P observed (t) is the observed true pressure value at time t, and P DACA (t) is the pressure value predicted by the DACA algorithm at time t.

[0142] Step 3.23: According to the θ factor fed back by DACA, RMEA adjusts the parameter settings of individuals in the new generation, and the specific formula is:

[0143]

[0144] where, is the learning rate; X bestis the individual with the highest fitness in the current generation.

[0145] Through the adjustment of the feedback factor, RMEA can tend to a higher fitness in the next generation, and at the same time compensate for the temperature influence through the real-time correction of DACA to form a closed-loop optimization.

[0146] Step 3.3: Construct an improved SVR model, optimize the model parameters, select a suitable kernel function, adjust the hyperparameters, and combine the multi-head attention mechanism MHA and the residual network ResNet to adapt to temperature compensation in different situations.

[0147] Step 3.31: The parameters of the SVR model are the key steps to ensure its best performance in the temperature compensation problem; it mainly involves three important parameters: the penalty parameter L, the kernel function v, and the tolerance error e; the specific optimization process is as follows:

[0148] 3.31.1: Use grid search combined with 5-fold cross-validation for optimization, set multiple candidate values of L, perform cross-validation on the training set, calculate the model performance under different L values, and select the best L value with the minimum validation error.

[0149] 3.31.2: Adopt grid search to select a suitable e, test a series of e values, and select the best e through cross-validation.

[0150] 3.31.3: Select the radial basis kernel function RBF to solve the modeling problem of nonlinear data, which is suitable for most nonlinear data sets:

[0151] K(x,x′) = exp(-v||x - x′|| 2 )

[0152] where K(x,x′) is the kernel function value between the input points x and x′, v is a positive constant, usually called the width parameter of the kernel function, which controls the decay rate of the function, and ||x - x′|| 2 is the square of the Euclidean distance, representing the difference between these two points in the input space.

[0153] Step 3.3.2: Introduce the multi-head attention mechanism MHA to enable the model to focus on features in different subspaces, linearly transform the input features to obtain the Q, K, V matrices, select the number of attention heads h according to the data complexity, calculate the attention outputs of each head and splice them, and finally restore the original dimension through a linear transformation.

[0154] In the input layer of the SVR model, the output of the multi-head attention is used as the enhanced input feature. The ResNet structure is further stacked on the enhanced features generated by the MHA to prevent information loss during the feature transmission through residual connections. The final enhanced features are used as the input features and directly input into the input layer of the SVR model for regression prediction.

[0155] Step 3.4: Combine the global search ability of RMEA and the local correction ability of DACA, and perform weighted fusion with the non-linear temperature-capacitance output relationship established by SVR. Compare the integrated model with the SVR, LSTM, and BP models respectively, and evaluate the comparison results.

[0156] Step 3.41: Set the population size to 100 individuals and evolve for 100 generations. The individuals adaptively update by learning the high-fitness individuals, and the low-fitness individuals are eliminated through the reflection mechanism. New individuals are randomly generated to maintain diversity, gradually approaching the optimal solution.

[0157] By introducing a feature self-enhancement module, SVR allows the model to learn feature interactions in different dimensions. A multi-layer perceptron (MLP) is introduced in the input layer, and through non-linear transformations of multiple hidden layers, the mapping ability to complex temperature fluctuations and sensor outputs is enhanced.

[0158] Step 3.42: Train the RMEA-DACA-SVR model through a dynamic weighted fusion strategy, combining the outputs of RMEA, DACA, and SVR. At the initial stage of the model, the weights of the three models are set to be equal. As the training process progresses, the model automatically adjusts the weights according to its performance under different temperature conditions. According to the performance of each model, the dynamic adjustment formula is:

[0159] w RMEA +w DACA +w SVR =1

[0160] where w RMEA is the weight of the RMEA model, used to measure the contribution of this model to the final result; w DACA is the weight of the DACA model; w SVR is the weight of the SVR model; each weight is dynamically updated according to its performance in the previous generations of iterations to ensure the optimal output of the combined model.

[0161] Step 3.43: Divide the dataset into multiple scenarios based on the volatility of temperature. Each scenario is divided into a training set, a validation set, and a test set respectively. The training set accounts for 60%, the validation set accounts for 20%, and the test set accounts for 20%. To verify the superiority of the integrated model, select the BP model, LSTM model, and SVR model as the comparison objects; use the root mean square error (MSE), mean absolute error (MAE), and coefficient of determination (R2 ) As an evaluation metric, first tune the hyperparameters of each model on the validation set, and then calculate the final performance of each model on the test set; the specific formula is:

[0162]

[0163] where H is the number of samples, is the model prediction value, and P i is the true value.

[0164] Step 3.44: Compare the integrated model with the SVR, LSTM, and BP models respectively;

[0165] Through the comparison experiment, under different temperature scenarios, the MSE, MAE, and R of the RMEA-DACA-SVR model 2 performance will be compared in detail with other comparison models to demonstrate its stability and accuracy under severe temperature fluctuations; the experimental results are shown in Table 1:

[0166] Table 1 Experimental result data

[0167] Model MSE MAE R2 SVR 0.04253 0.16125 0.872 LSTM 0.04020 0.15542 0.885 BP 0.04469 0.17653 0.830 RMEA-DACA-SVR 0.03475 0.14051 0.938

[0168] Step 3.45: Evaluate the comparison results; the MSE of the integrated model is 0.03475, the MAE is 0.14051, and R 2 is 0.938, which is significantly better than other models, Figure 6 showing the pressure measurement effect after non-linear temperature compensation based on RMEA-DACA-SVR; the horizontal axis is the actually applied pressure value, and the vertical axis is the compensated pressure output value. Different symbols and colors represent the compensation results under different temperature conditions from -40°C to 100°C; compared with Figure 5 before non-linear compensation in, the results show that the compensated output values under all temperature conditions coincide highly with the diagonal line, indicating that the compensated pressure value is highly consistent with the actually applied pressure, verifying the non-linear compensation effect of the RMEA-DACA-SVR model; whether at low temperature or high temperature, the compensated pressure curve effectively eliminates the influence of temperature on the measurement, significantly improving the measurement accuracy and showing good adaptability.

[0169] Step 4: Dynamically adjust the weights of REMA, DACA, and SVR according to their performance under different temperature scenarios; in the initial stage, the weights of REMA, DACA, and SVR are equally distributed. As the training iteration progresses, the system monitors the temperature scenario characteristics in real time, combines the error performance on the validation set, evaluates the performance of each model on the validation set, gradually adjusts the weights, and generates the integrated prediction results through the dynamic weighted fusion formula to strengthen the adaptive ability of the model in specific scenarios;

[0170] Step 5: Dynamic generation and output process of temperature compensation results: The system collects the current temperature value T(t) and the original pressure value P(t) in real time, and inputs them in parallel into the RMEA, DACA, and SVR models to generate the global compensation amount ΔP RMEA (t), the local distortion correction amount ΔP DACA (t), and the non-linear mapping compensation amount ΔP SVR (t). Based on the dynamic weights w RMEA , w DACA , w SVR perform non-linear fusion on the compensation amounts of the three to generate the final compensation value ΔP final (t) and superimpose it on the original pressure value. As Figure 2 shown, synchronously output the high-precision pressure signal P compensated (t) through 4-20mA analog output and Modbus digital protocol. At the same time, calculate the compensation error e(t) online. When the error exceeds the threshold, trigger the SVR parameter gradient fine-tuning and the RMEA population update to form a closed-loop optimization, and use a hardware timer to achieve millisecond-level synchronous output. Combine the limit protection mechanism to ensure the signal stability under extreme working conditions, and finally realize the dynamic generation of temperature compensation results, multi-protocol output, and adaptive accuracy maintenance.

[0171] The above embodiments are only for illustrating the technical concept and features of the present invention, and the purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. It should not be used to limit the protection scope of the present invention. Any equivalent transformation or modification made according to the spirit of the present invention should be covered within the protection scope of the present invention.

Claims

1. A temperature compensation method for a capacitive pressure transmitter based on RMEA-DACA-SVR, characterized in that: Including the steps: Step 1: Collect the real-time pressure data of the capacitive pressure transmitter under different temperature conditions to generate a comprehensive data set; Step 2: Preprocess the comprehensive data set, including removing noise data, data scaling, and extracting key features, and decompose the data using the HHT transform; Step 3: Construct a comprehensive temperature compensation model based on RMEA-DACA-SVR: Step 3.1: Introduce a learning and reflection mechanism into the RMEA algorithm. The learning mechanism allows communication and learning of excellent features and parameter settings among different individuals; the reflection mechanism is achieved by comparing the fitness values of individuals, eliminating individuals with lower fitness values, and stimulating the evolutionary vitality of the population; Step 3.2: Connect RMEA and DACA to each other through a feedback mechanism to form a closed-loop system; continuously adjust the parameter settings, dynamically adjust the weights, and re-evaluate to improve the compensation accuracy; Step 3.3: Construct an improved SVR model, optimize the model parameters, select a suitable kernel function, adjust the hyperparameters, and combine the multi-head attention mechanism MHA and the residual network ResNet to adapt to temperature compensation in different situations; Step 3.4: Combine the global search ability of RMEA with the local correction ability of DACA, and perform weighted fusion with the non-linear temperature and capacitance output relationship established by SVR; compare the comprehensive model with the SVR, LSTM, and BP models respectively, and evaluate the comparison results; Step 4: Dynamically adjust the weights of REMA, DACA, and SVR according to their performances in different temperature scenarios; In the initial stage, the weights of REMA, DACA, and SVR are equally distributed. As the training iteration progresses, the system monitors the temperature scenario features in real time, combines the error performance on the validation set, gradually adjusts the weights by evaluating the performances of each model on the validation set, and generates an integrated prediction result through the dynamic weighted fusion formula to strengthen the adaptive ability of the model in specific scenarios; Step 5: Dynamic generation and output process of temperature compensation results: The system collects the current temperature value T(t) and the original pressure value P(t) in real time, and inputs them in parallel into the RMEA, DACA, and SVR models to generate the global compensation amount ΔP RMEA (t), the local distortion correction amount ΔP DACA (t), and the non-linear mapping compensation amount ΔP SVR (t). Based on the dynamic weights w RMEA , w DACA , w SVR Non-linearly fuse the compensation amounts of the three to generate the final compensation value ΔP final (t) and superimpose it on the original pressure value, and synchronously output the high-precision pressure signal P compensated (t) through 4-20mA analog output and Modbus digital protocol. At the same time, calculate the compensation error e(t) online. When the error exceeds the threshold, trigger the SVR parameter gradient fine-tuning and the RMEA population update to form a closed-loop optimization, and use the hardware timer to achieve millisecond-level synchronous output. Combine the amplitude-limiting protection mechanism to ensure the signal stability under extreme working conditions, and finally realize the dynamic generation of temperature compensation results, multi-protocol output, and adaptive accuracy maintenance.

2. A temperature compensation method for a capacitive pressure transmitter based on RMEA-DACA-SVR according to claim 1, characterized in that: The specific method of the said Step 1 is: Step 1.1: Determine the model and its technical specifications of the capacitive pressure transmitter for data collection; clarify the key parameters to be monitored, including: pressure value P(t), temperature value T(t), time stamp; configure the data acquisition hardware, select a suitable acquisition frequency f = 1Hz, and set the acquisition interval Δt: Step 1.2: Conduct experiments in different temperature environments and record the pressure output under each condition; monitor and record the relationship between P(t) and T(t) to ensure that there are corresponding P(t) and T(t) at each time point t; Step 1.3: Integrate the pressure data under all temperature conditions to form a comprehensive data set D, and check the distribution characteristics of the data; perform basic statistical analysis on the comprehensive data set to calculate the mean μ P and the standard deviation σ P : Among them, μ P is the average value of the pressure data; N is the total number of the pressure data; P i (t) is the pressure value corresponding to the i-th sample; t is a time variable indicating the specific time point corresponding to each sample; i is an integer starting from 1 and ending at N; σ P is the standard deviation of the pressure data, which is used to measure the degree of fluctuation of the pressure data; Step 1.4: Integrate the pressure data under all temperature conditions, store the collected data in the database, and form a comprehensive data set.

3. A temperature compensation method for a capacitive pressure transmitter based on RMEA-DACA-SVR according to claim 1, characterized in that: The specific method of the said Step 2 is: Step 2.1: Use the Kalman filter to remove the noise data to obtain the denoised estimated value; Step 2.2: Adopt the Z-score normalization method for data scaling to convert the data into the form of a standard normal distribution, ensuring that the mean is 0 and the standard deviation is 1; the specific normalization formula is: Among them, Z i is the standardized data, X i is the original data value, μ is the sample mean, and σ is the sample standard deviation; Then use the covariance matrix C for normalization processing: Perform eigenvalue decomposition on the covariance matrix to construct a standardized data matrix, is the mean of the sample data, and T is the transpose operation; Step 2.3: Use PCA to reduce the dimension and extract features from the pressure and temperature data, project the data into the feature space, and calculate the eigenvectors and eigenvalues; Step 2.4: Perform data decomposition, take the mean of each multi-IM component respectively to eliminate the error introduced by noise addition, and obtain the final IMF: Among them, IMF j is the average value of the j-th intrinsic mode function component, is the value of the j-th IMF component corresponding to the n-th sample; Perform Hilbert transform on each IMF to obtain the instantaneous frequency and envelope amplitude for analyzing the time-frequency characteristics of the data; The Hilbert transform formula is: where P.V. represents the principal value integral, and H(IMF j )(t1) is the transformation result of the j-th intrinsic mode function at t1, and IMF j(τ) is the value of the j-th intrinsic mode function at time τ, where τ is the integration variable representing continuous points in time, t1 is the target time point of the Hilbert transform, and dτ is the differential of the integration variable τ; the final decomposition result of HHT, synthesizing the final signal decomposition form: Among them, X(t) is the final decomposition result of the original signal, and IMF j is each IMF component, r is the residual term, and j is an integer starting from 1 and ending at m.

4. A temperature compensation method for a capacitive pressure transmitter based on RMEA-DACA-SVR according to claim 1, characterized in that, The specific method of step 3.1 is as follows: Step 3.11: Randomly generate an initial population X = {X1, X2, …, X N}, where each individual x i represents a potential parameter setting for the temperature compensation model; Step 3.12: Introduce a learning mechanism where individuals improve their performance by learning from better individuals; assume individual X i learns from individual X j with a higher fitness, and the update formula is as follows: X i (t + 1) = X i (t) + ζ·(X j (t) - X i (t)) + β·η where ζ is the learning rate, which determines the step size of learning; β is the noise intensity coefficient to maintain population diversity; η is the random noise term, which follows a normal distribution where is the mathematical symbol of the normal distribution; X i (t + 1) is the position of the i-th individual at time step t + 1, X i (t) is the position of the i-th individual at the current time step t, X j (t) is the position of the j-th individual at the current time step t; Step 3.13: Introduce a reflection mechanism to eliminate individuals with low fitness and introduce new individual configurations; assume the individual with relatively low current fitness is X k , and the update formula is as follows: X k (t + 1)= X best (t)+γ·(X k (t)- X avg (t))+δ·β Among them, X best (t) is the individual with the highest fitness in the current generation, X avg (t) is the average parameter value of the population in the current generation, X k (t + 1) is the position of the optimal individual at the current time step t, γ is a regulation parameter that controls the step length for the individual to move according to the difference between the current optimal position and the current individual position, X k (t) is the position of the k-th individual at the current time step t, and δ is an adjustment factor used to affect the random variation of the individual; Step 3.14: In population update, introduce enhanced crossover and adaptive mutation strategies simultaneously; during the crossover process, use the fitness value as the crossover weight coefficient; Among them, X p and X q are two parental individuals, and F(X p ) and F(X q ) are their fitness values respectively, and X new is the newly calculated position; The newly generated individuals are added to the population, and fitness evaluation is performed before the next generation starts; retain the individuals with higher fitness to ensure that the population continuously evolves towards a better direction; Step 3.15: Repeat the above steps until the termination condition is met; finally, the individual with the highest fitness is selected as the optimal parameter configuration for temperature compensation.

5. A temperature compensation method for a capacitive pressure transmitter based on RMEA-DACA-SVR according to claim 1, characterized in that, The specific method of step 3.2 is as follows: Step 3.21: In each iteration cycle, DACA receives the initial parameters optimized by RMEA as input, adjusts the compensation parameters in real time, and feeds back the correction result to RMEA; at the beginning of each generation, RMEA calculates the fitness score according to the feedback of DACA, and the feedback value is used to adjust the fitness of individuals to form a preliminary evaluation; Step 3.22: During the compensation process, DACA dynamically corrects the pressure value according to the current temperature change; RMEA updates the fitness of individuals in the population by receiving the correction feedback of DACA; At the end of each generation, calculate the feedback factor θ according to the correction effect of DACA and feed it back to RMEA to affect its population optimization process: where E is the total number of feedback cycles, which is a factor used to measure the correction effect of DACA, and P observed (t) is the true pressure value observed at time t, and P DACA (t) is the pressure value predicted by the DACA algorithm at time t; Step 3.23: According to the θ factor fed back by DACA, RMEA adjusts the parameter settings of individuals in the new generation, and the specific formula is: Among them, is the learning rate; X best is the individual with the highest fitness in the current generation, X i (t) is the state of the i-th individual at time t; Through the adjustment of the feedback factor, RMEA can tend to a higher fitness in the next generation, and at the same time compensate for the temperature influence through the real-time correction of DACA to form a closed-loop optimization.

6. A temperature compensation method for a capacitive pressure transmitter based on RMEA-DACA-SVR according to claim 1, characterized in that, The specific method of step 3.3 is as follows: The parameters of the SVR model are the key steps to ensure its best performance in the temperature compensation problem; it mainly involves three important parameters: the penalty parameter L, the kernel function v, and the tolerance error e; the specific optimization process is as follows: 3.31.1: Use grid search combined with 5-fold cross-validation for optimization, set multiple candidate values of L, perform cross-validation on the training set, calculate the model performance under different L values, and select the best L value with the smallest validation error; 3.31.2: Adopt grid search to select the appropriate e, a series of e values can be tested, and the best e is selected through cross-validation; 3.31.3: Select the radial basis kernel function RBF to solve the modeling problem of nonlinear data, which is suitable for most nonlinear data sets: K(x,x′) = exp(-v||x - x′|| 2 ) Among them, $K(x, x')$ is the kernel function value between the input points $x$ and $x'$, and $v$ is a positive constant, usually called the width parameter of the kernel function, which controls the decay rate of the function, and $\|x - x'\|$ 2 is the square of the Euclidean distance, representing the difference between these two points in the input space; Step 3.3.2: Introduce the multi-head attention mechanism MHA to enable the model to focus on features in different subspaces, linearly transform the input features to obtain the Q, K, and V matrices, select the number of attention heads h according to the data complexity, calculate the attention outputs of each head and splice them, and finally restore the original dimension through a linear transformation; In the input layer of the SVR model, use the multi-head attention output as the enhanced input features, further stack the ResNet structure on the enhanced features generated by MHA, and prevent information loss during feature transmission through residual connections; the final enhanced features are used as input features and directly input into the input layer of the SVR model for regression prediction.

7. A temperature compensation method for a capacitive pressure transmitter based on RMEA-DACA-SVR according to claim 1, characterized in that The specific method of step 3.4 is as follows: Step 3.41: Set the population size to 100 individuals and evolve for 100 generations. The individuals adaptively update by learning high-fitness individuals, and eliminate low-fitness individuals through a reflection mechanism. Randomly generate new individuals to maintain diversity and gradually approach the optimal solution; SVR allows the model to learn feature interactions in different dimensions by introducing a feature self-enhancement module; Introduce a multi-layer perceptron (MLP) in the input layer to enhance the mapping ability to complex temperature fluctuations and sensor outputs through non-linear transformations of multiple hidden layers; Step 3.42: Through a dynamic weighted fusion strategy, combine the outputs of RMEA, DACA, and SVR to train the RMEA-DACA-SVR model; at the initial stage of the model, the weights of the three models are set to be equal; as the training process progresses, the model will automatically adjust the weights according to its performance under different temperature conditions; according to the performance of each model, the dynamic adjustment formula is: w RMEA +w DACA +w SVR = 1 Among them, w RMEA is the weight of the RMEA model, which is used to measure the contribution of the model to the final result; w DACA is the weight of the DACA model; w SVR is the weight of the SVR model; each weight is dynamically updated according to its performance in the previous generations of iterations to ensure the optimal output of the combined model; Step 3.43: Divide the dataset into multiple scenarios based on the temperature volatility. Each scenario is divided into a training set, a validation set, and a test set. The training set accounts for 60%, the validation set accounts for 20%, and the test set accounts for 20%. To verify the superiority of the comprehensive model, select the BP model, LSTM model, and SVR model as comparison objects; use the root mean square error (MSE), mean absolute error (MAE), and coefficient of determination (R 2 ) as evaluation indicators. First, tune the hyperparameters of each model on the validation set, and then calculate the final performance of each model on the test set; the specific formulas are as follows: where H is the number of samples, is the model prediction value, P i is the true value.

Images