An IRIME-CNN-DD-based temperature compensation method and system for a silicon piezoresistive pressure sensor

Abstract

This invention discloses a temperature compensation method and system for a silicon piezoresistive pressure sensor based on IRIME-CNN-DD, belonging to the field of pressure detection technology, to solve the problem of low temperature compensation accuracy in existing software compensation methods. The invention first acquires pressure data at different temperatures as a training dataset; then normalizes the training dataset; optimizes the model parameters of the temperature compensation model using an improved frost-freezing algorithm to obtain the optimal model parameters; inputs the normalized training dataset into the temperature compensation model and trains the model using the optimal model parameters; finally, inputs the pressure data to be measured into the trained temperature compensation model for temperature compensation. This invention optimizes the neural network parameters using an improved frost-freezing algorithm, increasing the model's robustness, and further improves the sensitivity of the pressure sensor by extracting local features from the temperature compensation dataset through a convolutional neural network.

Description

Technical Field

[0001] This invention relates to the field of pressure detection technology, specifically to a temperature compensation method and system for a silicon piezoresistive pressure sensor based on IRIME-CNN-DD. Background Technology

[0002] Silicon piezoresistive pressure sensors are widely used in environmental monitoring, shipbuilding and marine engineering, energy industries, and other fields due to their low cost, small size, and high sensitivity, becoming a key component driving technological progress and development in these sectors. However, because silicon-based semiconductors have thermosensitive properties, silicon piezoresistive pressure sensors are highly susceptible to factors such as temperature, which can introduce errors into the measurement results.

[0003] Currently, there are two main types of temperature compensation methods. The first is hardware compensation, which directly offsets the effects of temperature changes by adding physical components to the system. The second is software compensation, which indirectly offsets the effects of temperature changes using software algorithms. Compared with hardware compensation strategies, software compensation strategies are easier to adjust and upgrade, and do not increase the physical complexity of the system, thus becoming a research hotspot. Currently widely used software compensation methods mainly include neural network algorithms, least squares methods, and interpolation compensation. Because neural networks have stronger multidimensional non-mapping capabilities compared to other software compensation algorithms, this capability is widely used in sensor temperature compensation systems.

[0004] However, existing neural network temperature compensation methods generally suffer from problems such as slow convergence speed and susceptibility to getting trapped in local minima of the error function. Especially when dealing with large search spaces, multi-peaked functions, and non-differentiable functions, traditional methods often struggle to find the global minimum effectively, and their generalization ability is also limited. These shortcomings significantly restrict the temperature compensation performance of traditional neural network algorithms in silicon pressure sensors. Summary of the Invention

[0005] In view of the above problems, this invention proposes a temperature compensation method and system for silicon piezoresistive pressure sensors based on IRIME-CNN-DD, in order to solve the problems existing in the current software compensation method.

[0006] According to one aspect of the present invention, a temperature compensation method for a silicon piezoresistive pressure sensor based on IRIME-CNN-DD is proposed, the method comprising:

[0007] Obtain pressure data at different temperatures as a training dataset;

[0008] The training dataset is normalized.

[0009] An improved frost algorithm is used to optimize the model parameters of a temperature compensation model to obtain the optimal model parameters; the model parameters include the initial learning rate and the maximum number of iterations; the temperature compensation model is a temperature compensation model based on a convolutional neural network.

[0010] The normalized training dataset is input into the temperature compensation model, and the temperature compensation model is trained using the optimal model parameters to obtain the trained temperature compensation model.

[0011] The pressure data to be measured is input into the trained temperature compensation model for temperature compensation, and the corrected pressure data is obtained.

[0012] Furthermore, the step of optimizing the model parameters of the temperature compensation model using the improved rime ice algorithm to obtain the optimal model parameters includes:

[0013] Initialize the population, fitness function, global optimal solution of the particle, historical individual optimal solution of the particle, maximum number of iterations T, and current number of iterations t;

[0014] Based on the fitness function of each particle in the population, the population is divided into a deep mining population and a global search population.

[0015] The following process is executed iteratively:

[0016] Update the positions of particles in the deep mining population according to the position update formula of the deep mining learning strategy;

[0017] Update the positions of particles in the global search population according to the global exploration learning strategy position update formula;

[0018] A fixed percentage of particles are selected from both the deep mining population and the global search population, and a cross population is generated based on the cross population position update formula generated by the adaptive Lap crossover operator.

[0019] The merged population is obtained by merging the deep mining, global exploration, and crossover populations.

[0020] Use aggressive greedy updates to merge the global optimal solution and the historical individual optimal solution in the population;

[0021] Adjust the particle ratio between the deep mining population and the global search population according to preset rules;

[0022] The iteration stops when the number of iterations t is greater than or equal to the maximum number of iterations T, and the global optimal solution, i.e., the optimal model parameters, is output.

[0023] Furthermore, the position update formula for the deep learning strategy is as follows:

[0024]

[0025] In the formula, Indicates the particle positions in the updated depth-mined population; R best,j Represents the optimal particle; r1 is a random number ranging from (-1, 1); t is the current iteration number; T is the maximum iteration number; h represents the adhesion degree; Ub ij and Lb ij These are the upper and lower bounds of the escape space, respectively; r2 is a random number in the range (0,1); trnd(T+t) is a random number following the Student t-distribution; and r4 is a random number in the range (0,1).

[0026] Furthermore, the position update formula for the global exploration learning strategy is:

[0027]

[0028] In the formula, This indicates the position of particles in the global exploration population after the update.

[0029] Furthermore, the crossover population position update formula generated by the adaptive Lapp crossover operator is as follows:

[0030]

[0031] In the formula, is the individual's historical optimal solution; c is a random number with a value in the range [-1, 1]; r5 is a random number with a value in the range [0, 1]. r is a random number with a value in the range [0,1], a is a position parameter, and u is a random number with a value in the range [0,1] that follows a uniform distribution.

[0032] Furthermore, the fitness function is the mean absolute error.

[0033] Furthermore, the temperature compensation model is a temperature compensation model based on convolutional neural networks and dendritic networks.

[0034] According to another aspect of the present invention, a temperature compensation system for a silicon piezoresistive pressure sensor based on IRIME-CNN-DD is proposed, the system comprising:

[0035] The data acquisition module is configured to acquire pressure data at different temperatures as a training dataset.

[0036] A preprocessing module configured to normalize the training dataset;

[0037] The parameter optimization module is configured to optimize the model parameters of the temperature compensation model using an improved frost algorithm to obtain the optimal model parameters; the model parameters include the initial learning rate and the maximum number of iterations; the temperature compensation model is a temperature compensation model based on a convolutional neural network.

[0038] The model training module is configured to input the normalized training dataset into the temperature compensation model, and train the temperature compensation model using the optimal model parameters to obtain the trained temperature compensation model.

[0039] The temperature compensation module is configured to input the pressure data to be measured into a trained temperature compensation model for temperature compensation, and obtain the corrected pressure data.

[0040] Furthermore, the parameter optimization module utilizes an improved frost algorithm to optimize the model parameters of the temperature compensation model, obtaining the optimal model parameters including:

[0041] Initialize the population, fitness function, global optimal solution of the particle, historical individual optimal solution of the particle, maximum number of iterations T, and current number of iterations t;

[0042] Based on the fitness function of each particle in the population, the population is divided into a deep mining population and a global search population.

[0043] The following process is executed iteratively:

[0044] Update the positions of particles in the deep mining population according to the position update formula of the deep mining learning strategy;

[0045] Update the positions of particles in the global search population according to the global exploration learning strategy position update formula;

[0046] A fixed percentage of particles are selected from both the deep mining population and the global search population, and a cross population is generated based on the cross population position update formula generated by the adaptive Lap crossover operator.

[0047] The merged population is obtained by merging the deep mining, global exploration, and crossover populations.

[0048] Use aggressive greedy updates to merge the global optimal solution and the historical individual optimal solution in the population;

[0049] Adjust the particle ratio between the deep mining population and the global search population according to preset rules;

[0050] The iteration stops when the number of iterations t is greater than or equal to the maximum number of iterations T, and the global optimal solution, i.e., the optimal model parameters, is output.

[0051] Furthermore, the temperature compensation model in the parameter optimization module is a temperature compensation model based on convolutional neural networks and dendritic networks.

[0052] The beneficial technical effects of this invention are:

[0053] This invention proposes a temperature compensation method and system for silicon piezoresistive pressure sensors based on IRIME-CNN-DD. The invention first acquires pressure data at different temperatures as a training dataset; then normalizes the training dataset; optimizes the model parameters of the temperature compensation model using an improved frost-freezing algorithm to obtain optimal model parameters; the model parameters include the initial learning rate and maximum number of iterations of the neural network; the temperature compensation model is a temperature compensation model based on convolutional neural networks and dendritic networks; the normalized training dataset is input into the temperature compensation model, and the optimal model parameters are used to train the temperature compensation model to obtain a trained temperature compensation model; the pressure data to be measured is input into the trained temperature compensation model for temperature compensation to obtain corrected pressure data. This invention optimizes the neural network parameters using an improved frost-freezing algorithm to increase the model's robustness, then extracts local features from the temperature compensation dataset through a convolutional neural network, and further constructs features through a dendritic network. Attached Figure Description

[0054] The above and other objects, features, and advantages of exemplary embodiments of the present invention will become readily apparent from the following detailed description taken in conjunction with the accompanying drawings. Several embodiments of the invention are illustrated in the drawings by way of example and not limitation, wherein:

[0055] Figure 1 This is a flowchart of a temperature compensation method for a silicon piezoresistive pressure sensor based on IRIME-CNN-DD, as described in an embodiment of the present invention.

[0056] Figure 2 This is an example diagram of the data acquisition system structure in an embodiment of the present invention.

[0057] Figure 3 This is an example diagram of pressure data collected at different temperatures in an embodiment of the present invention.

[0058] Figure 4 This is a schematic diagram of the CNN network and DD network structures in an embodiment of the present invention.

[0059] Figure 5 This is a density function curve of the Laplace distribution in an embodiment of the present invention.

[0060] Figure 6 This is a flowchart illustrating the process of optimizing the model parameters of the temperature compensation model using an improved frost algorithm in an embodiment of the present invention.

[0061] Figure 7This is a diagram showing the absolute error distribution of the pressure sensor in an embodiment of the present invention. Detailed Implementation

[0062] The principles and spirit of the invention will now be described with reference to several exemplary embodiments. It should be understood that these embodiments are given merely to enable those skilled in the art to better understand and implement the invention, and are not intended to limit the scope of the invention in any way. Rather, these embodiments are provided to make this disclosure more thorough and complete, and to fully convey the scope of this disclosure to those skilled in the art.

[0063] Those skilled in the art will recognize that embodiments of the present invention can be implemented as a system, apparatus, device, method, or computer program product. Therefore, this disclosure can be specifically implemented in the following forms: entirely hardware, entirely software (including firmware, resident software, microcode, etc.), or a combination of hardware and software. It should be understood herein that any number of elements in the accompanying drawings is for illustrative purposes only and not as a limitation, and any naming is for distinction only and has no limiting meaning.

[0064] This invention provides a temperature compensation method and system for a silicon piezoresistive pressure sensor based on IRIME-CNN-DD. First, the initial parameters of the model are optimized using an improved frost algorithm. Then, local features in the dataset are extracted through a convolutional neural network (CNN). Finally, the extracted features are fed into a dendritic network (DD) for feature construction, which further improves the sensitivity of the sensor.

[0065] This invention proposes a temperature compensation method for silicon piezoresistive pressure sensors based on IRIME-CNN-DD, such as... Figure 1 As shown, the method includes:

[0066] S1. Obtain pressure data at different temperatures as a training dataset;

[0067] S2. Normalize the training dataset;

[0068] S3. Optimize the model parameters of the temperature compensation model using the improved frost algorithm to obtain the optimal model parameters; the model parameters include the initial learning rate and the maximum number of iterations; the temperature compensation model is a temperature compensation model based on a convolutional neural network.

[0069] S4. Input the normalized training dataset into the temperature compensation model, and train the temperature compensation model using the optimal model parameters to obtain the trained temperature compensation model.

[0070] S5. Input the pressure data to be measured into the trained temperature compensation model for temperature compensation to obtain the corrected pressure data.

[0071] The method begins with S1. In S1, pressure data at different temperatures are acquired as the training dataset.

[0072] According to embodiments of the present invention, a data acquisition system is designed using a silicon piezoresistive pressure sensor, an AD conversion module, a host computer, a DS18B20 temperature sensor, and an STM32F407 microcontroller, such as... Figure 2 As shown, data acquisition was performed. The designed data acquisition system was placed in a high and low temperature chamber for specific pressure tests. Eleven temperature points with a step size of 10℃ were set within the range of -20℃ to 80℃. Each temperature point was stored for 180 minutes before the pressure test began. Eleven pressure test points with a step size of 4KPa were set during the test. Each pressure was increased for 15 minutes, from 0KPa to 40KPa, and then depressurized back to 0KPa. The data collected from the upward and downward pressures were then averaged and combined. The results are shown below. Figure 3 As shown. From Figure 3 It can be concluded that the pressure sensor exhibits a significant temperature drift effect when the ambient temperature changes. Pressure data measured at -20℃ to 10℃ and 50℃ to 80℃ were used as the training dataset for the model; pressure data measured at 20℃ to 40℃ were used as the test data for testing.

[0073] Then, S2 is executed to normalize the training dataset.

[0074] According to embodiments of the present invention, when a dataset contains multiple features, if these features have different units (e.g., length expressed in meters and centimeters), normalization can help eliminate such differences in units, making different features comparable. Therefore, stress data is normalized to eliminate the influence of units between data features. This helps ensure that each feature contributes relatively evenly to the model, thereby improving the stability and accuracy of training.

[0075] Normalization can also accelerate model convergence. By adjusting the scale of features to a similar range, optimization algorithms such as gradient descent take a more direct path in finding the optimal solution, thus speeding up convergence and shortening training time.

[0076] Normalization can also reduce scale sensitivity and improve the accuracy of the algorithm.

[0077] Normalization also helps prevent gradient explosion, which occurs when the gradient becomes too large due to the excessively large values ​​of certain features during gradient descent, thus affecting the stability and convergence of the model.

[0078] Normalization can also ensure that each feature has a relatively equal impact on the model, avoiding the overemphasis on certain features in the model due to their large scale.

[0079] Then, S3 is executed to optimize the model parameters of the temperature compensation model using the improved frost algorithm, and obtain the optimal model parameters; the model parameters include the initial learning rate and the maximum number of iterations.

[0080] According to embodiments of the present invention, the temperature compensation model employs a temperature compensation model based on a convolutional neural network (CNN), or a temperature compensation model based on a CNN and a dendritic network (DD). A CNN is a feedforward neural network with the ability to extract local features, a capability widely used in various fields. A dendritic network (DD) is a novel machine learning algorithm similar to a support vector machine (SVM) or a multilayer perceptron, offering higher accuracy and model generalization ability compared to a backpropagation (BP) neural network.

[0081] like Figure 4 As shown, the CNN network mainly consists of two convolutional layers and two pooling layers. The kernel size in both the first and second convolutional layers is set to 1×1, containing 32 and 64 kernels respectively, and the ReLU activation function is used in both. Both pooling layers use max pooling. The DD network mainly consists of one input layer and two DD layers. The DD layer has three layers.

[0082] Establishing a sound temperature compensation model for a silicon piezoresistive pressure sensor hinges on determining the maximum number of iterations and the initial learning rate of the CNN-DD network. This invention combines an improved haze algorithm to search for the optimal network parameters of the CNN-DD network, minimizing the model's mean absolute error on the test set and thus improving model accuracy.

[0083] The rime optimization algorithm (RIME) is a swarm intelligence optimization algorithm based on the physical phenomenon of rime-ice. This algorithm simulates the growth process of soft and hard rime, constructing a soft rime search strategy, a hard rime puncture mechanism, and an aggressive greedy selection mechanism. The position update formulas for the soft rime search strategy and the hard rime puncture mechanism are shown in equations (1) and (2).

[0084]

[0085] In the formula, Update the position of the particle; R best,j The optimal particle is determined by r1, which is a random number ranging from -1 to 1. w is a constant with a value of 5. Ub ij and Lb ijThese are the upper and lower bounds of the escape space, respectively. h is the adhesion, ranging from (0,1). t is the current iteration number. r2 is a random number ranging from (0,1). T is the maximum number of iterations.

[0086]

[0087] In the formula, r3 is a random number taking values ​​in the range (-1, 1). F normr (S i ) is the normalized value of the current particle fitness value.

[0088] For the RIME algorithm, its optimization capability mainly depends on how to effectively balance global exploration and depth mining capabilities, and how to ensure population diversity. To address these issues, this invention proposes an improved RIME algorithm (IRIME) based on classification discussion and an adaptive Laplace crossover operator strategy. The improved algorithm redefines the position update formula by utilizing classification discussion and an adaptive Laplace crossover operator strategy, balancing the depth mining and global search capabilities of the RIME algorithm while increasing population diversity, thereby enhancing the algorithm's optimization ability.

[0089] To effectively balance global exploration and deep mining, inspired by the concept of classification and discussion, and considering the Euclidean distance between the fitness of the top and best particles in the population, the original population is divided into two categories at a 1:9 ratio: a deep mining population and a global exploration population. As the number of iterations increases, the number of particles in the deep mining population gradually increases, while the number of particles in the global exploration population gradually decreases. Subsequently, learning strategies focused on mining are designed for the deep mining population, and learning strategies favoring global exploration are designed for the global exploration population.

[0090] The position update formula for the deep learning strategy is shown in equation (3):

[0091]

[0092] In the formula, trnd(T+t) is a random number that follows the Student's t-distribution. r4 is a random number that takes values ​​in the range (0,1). This indicates the particle positions in the population after the update and deep mining.

[0093] The position update formula for the global exploration learning strategy is shown in equation (4):

[0094]

[0095] In the formula, This indicates the position of particles in the global exploration population after the update. To avoid the risk of reduced population diversity due to the grouping strategy, a reasonable population information interaction strategy is particularly important. Experiments have shown that during the convergence process of the IRIME algorithm, some particles will get stuck in a local optimum problem as the number of iterations increases. To address this characteristic, this invention introduces an adaptive Laplace crossover operator to help the algorithm increase population diversity.

[0096] The density function of the Laplace distribution is shown in equation (5):

[0097]

[0098] Where x is the value of the random variable; a is the location parameter; and b is the scale parameter. Figure 5 The density function curve of the Laplace distribution is shown.

[0099] If u is a random number that takes values ​​in the range [0,1] and follows a uniform distribution, then the generated random number l is given by equation (6):

[0100]

[0101] Depend on Figure 5 When b = 1, since f(x) has a wide distribution range, the particles can explore the optimal solution with a larger step size in the early stage of deep mining, thus avoiding the algorithm from getting trapped in local optima, as shown in equation (7):

[0102]

[0103] When b = 0.5, since the middle part of f(x) is relatively high, the particle can achieve deep mining in a smaller step in the later stage, and can converge to the global optimal solution better, as shown in equation (8):

[0104]

[0105] In the formula, r is a random number with a value in the range [0,1].

[0106] The final formula for updating the crossover population position generated by the adaptive Lapp crossover operator is shown in equation (9):

[0107]

[0108] In the formula For each individual's historical optimal solution, r5 is a random number ranging from [0,1]. c is a random number ranging from [-1,1].

[0109] The process of optimizing the model parameters of the temperature compensation model using the improved frost algorithm and obtaining the optimal model parameters is as follows: Figure 6 As shown, it includes the following steps:

[0110] S31. Initialize the population; randomly generate multiple sets (e.g., 30 sets) of initial hyperparameter combinations (i.e., individuals in the population), where the hyperparameters are the initial learning rate and the maximum number of iterations of the CNN network or CNN-DD network;

[0111] S32. Initialize the fitness function according to formula (10):

[0112]

[0113] In the formula, It is the true value of the test sample, y n This is the output of the model, where N is the number of test samples;

[0114] S33. Initialize the global optimal solution of the particle, the historical individual optimal solution, the maximum number of iterations T (this maximum number of iterations is the number of iterations set in the rime algorithm) and the current number of iterations t = 0;

[0115] S34. Divide the population into a deep mining population and a global search population according to the fitness function of each particle in the population (Equation (10)). For example, if there are 10 particles, the fitness values ​​of the 10 particles will be obtained according to Equation (10). Then, after arranging the fitness values ​​from smallest to largest, divide the population according to the ratio. The population can be divided into a deep mining population and a global search population at a ratio of 1:9.

[0116] S35. Update the position of particles in the deep mining population according to formula (3);

[0117] S36. Update the positions of particles in the global search population according to formula (4);

[0118] S37. Select a fixed percentage (e.g., 5%) of particles from the deep mining population and the global search population to generate a crossover population using formula (9);

[0119] S38. Merge the deep mining, global exploration population and the crossover population to obtain the merged population;

[0120] S39. Use aggressive greedy updates to merge the global optimal solution and the historical individual optimal solution in the population;

[0121] S40. Adjust the particle ratio between the deep mining population and the global search population according to preset rules; the preset rules are, for example:

[0122]

[0123]

[0124] In the formula, It is the proportion of particles in the deep exploration population. It represents the proportion of particles in the global search population.

[0125] S41. The current iteration number t = t + 1;

[0126] S42. If the algorithm has not yet reached the stopping criterion (i.e., t>T), then return to S34 and continue iterating; once the stopping condition is reached, proceed to S43.

[0127] S43. Obtain the global optimal solution, that is, the optimal initial learning rate and maximum number of iterations for the CNN-DD network.

[0128] Then, execute S4, input the normalized training dataset into the temperature compensation model, and train the temperature compensation model using the optimal model parameters to obtain the trained temperature compensation model.

[0129] Finally, in S5, the pressure data to be measured is input into the trained temperature compensation model for temperature compensation to obtain the corrected pressure data.

[0130] According to an embodiment of the present invention, the pressure data to be measured needs to be normalized first, and then input into the trained temperature compensation model. The model output data needs to be denormalized.

[0131] Furthermore, to verify the superiority of the IRIME-CNN-DD model and the necessity of CNN network local feature extraction and IRIME network optimization, comparative experiments are conducted to illustrate the technical effects of the present invention.

[0132] Temperature compensation for silicon pressure sensors was performed using four models proposed in this invention: IRIME-CNN-DD, RIME-CNN-DD, CNN-DD, and DD. Considering the real-time nature of the algorithms, the maximum number of iterations for the IRIME and RIME algorithms in the optimization module was set to 20, and the population size was set to 30. The results are shown in Table 1.

[0133] Table 1. Model Comparison Results

[0134]

[0135] From Table 1, the following conclusions can be drawn: (1) Compared with DD, CNN-DD reduces the root mean square error by 0.20792 kPa and the mean absolute error by 0.04572 kPa. This indicates that local feature extraction in CNN networks can improve the prediction accuracy of the model. (2) Compared with CNN-DD, RIME-CNN-DD reduces the root mean square error by 0.06213 kPa and the mean absolute error by 0.01984 kPa. This indicates that the RIME algorithm can improve the robustness of the model. (3) Compared with RIME-CNN-DD, IRIME-CNN-DD reduces the root mean square error by 0.03626 kPa and the mean absolute error by 0.02881 kPa. This indicates that the IRIME algorithm can effectively avoid the local optimum problem of the RIME algorithm.

[0136] To further analyze and evaluate the temperature compensation performance of the model, four classic neural network temperature compensation models were selected. The optimal network temperature compensation model in 100 independent runs was used as the test model for comparison. The results are shown in Table 2.

[0137] Table 2. Model Comparison Results

[0138]

[0139] As shown in Table 2, the IRIME-CNN-DD temperature compensation model exhibits the best compensation performance. Compared to the PSO-CNN-DD model, the root mean square error (RMSE) of the model proposed in this invention is reduced by 30.46%, and the mean absolute error (MAE) is reduced by 32.89%. Compared to the DE-SVM model, the RMSE of the model proposed in this invention is reduced by 37.54%, and the MAE is reduced by 35.28%. Compared to the GA-BP model, the RMSE of the model proposed in this invention is reduced by 41.38%, and the MAE is reduced by 38.69%. Therefore, the model proposed in this invention outperforms traditional silicon piezoresistive pressure sensor temperature compensation models such as PSO-CNN-DD, DE-SVM, and GA-BP in both performance metrics.

[0140] After saving the trained model, the measured pressure data was input to obtain the temperature compensation results, and the absolute error was derived from these results. The calibration pressures were 12 kPa, 24 kPa, 36 kPa, and 40 kPa, and the calibration temperatures were -20 to 80℃. The temperature compensation results are shown in Table 3, and the absolute error distribution is shown in [Table 3]. Figure 7 .

[0141] Table 3 Temperature compensation results

[0142]

[0143] From Table 3 and Figure 7It can be seen that the model's compensation result has a small deviation from the actual result, indicating that the temperature compensation effect of the silicon piezoresistive pressure sensor designed in this invention is good, and the output accuracy and reliability of the model meet the requirements of temperature compensation for silicon piezoresistive pressure sensors.

[0144] Another embodiment of the present invention proposes a temperature compensation system for a silicon piezoresistive pressure sensor based on IRIME-CNN-DD, the system comprising:

[0145] The data acquisition module is configured to acquire pressure data at different temperatures as a training dataset.

[0146] A preprocessing module configured to normalize the training dataset;

[0147] The parameter optimization module is configured to optimize the model parameters of the temperature compensation model using an improved frost algorithm to obtain the optimal model parameters; the model parameters include the initial learning rate and the maximum number of iterations; the temperature compensation model is a temperature compensation model based on a convolutional neural network.

[0148] The model training module is configured to input the normalized training dataset into the temperature compensation model, and train the temperature compensation model using the optimal model parameters to obtain the trained temperature compensation model.

[0149] The temperature compensation module is configured to input the pressure data to be measured into a trained temperature compensation model for temperature compensation, and obtain the corrected pressure data.

[0150] In this embodiment, optionally, the parameter optimization module uses an improved frost algorithm to optimize the model parameters of the temperature compensation model, and obtains the optimal model parameters by:

[0151] Initialize the population and divide it into a deep mining population and a global search population according to the proportions.

[0152] Initialize the population, fitness function, global optimal solution of the particle, historical individual optimal solution of the particle, maximum number of iterations T, and current number of iterations t;

[0153] Based on the fitness function of each particle in the population, the population is divided into a deep mining population and a global search population.

[0154] The following process is executed iteratively:

[0155] Update the positions of particles in the deep mining population according to the position update formula of the deep mining learning strategy;

[0156] Update the positions of particles in the global search population according to the global exploration learning strategy position update formula;

[0157] A fixed percentage of particles are selected from both the deep mining population and the global search population, and a cross population is generated based on the cross population position update formula generated by the adaptive Lap crossover operator.

[0158] The merged population is obtained by merging the deep mining, global exploration, and crossover populations.

[0159] Use aggressive greedy updates to merge the global optimal solution and the historical individual optimal solution in the population;

[0160] Adjust the particle ratio between the deep mining population and the global search population according to preset rules;

[0161] The iteration stops when the number of iterations t is greater than or equal to the maximum number of iterations T, and the global optimal solution is output, which is the optimal initial learning rate and maximum number of iterations for the CNN-DD network.

[0162] In this embodiment, optionally, the temperature compensation model in the parameter optimization module is a temperature compensation model based on convolutional neural networks and dendritic networks.

[0163] The function of the temperature compensation system for a silicon piezoresistive pressure sensor based on IRIME-CNN-DD in this embodiment of the invention can be described by the aforementioned temperature compensation method for a silicon piezoresistive pressure sensor based on IRIME-CNN-DD. Therefore, for the parts not described in detail in the system embodiment, please refer to the above method embodiment, and they will not be repeated here.

[0164] It should be noted that although several units, modules, or sub-modules are mentioned in the detailed description above, this division is merely exemplary and not mandatory. In fact, according to embodiments of the present invention, the features and functions of two or more modules described above can be embodied in one module. Conversely, the features and functions of one module described above can be further divided and embodied by multiple modules.

[0165] Furthermore, although the operations of the method of the present invention are described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in that specific order, or that all the operations shown must be performed to achieve the desired result. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

[0166] While the spirit and principles of the invention have been described with reference to several specific embodiments, it should be understood that the invention is not limited to the disclosed specific embodiments, and the division of aspects does not imply that features in these aspects cannot be combined for benefit; such division is merely for ease of description. The invention is intended to cover various modifications and equivalent arrangements included within the spirit and scope of the appended claims.

Claims

1. A temperature compensation method for a silicon piezoresistive pressure sensor based on IRIME-CNN-DD, characterized in that, include: Obtain pressure data at different temperatures as a training dataset; The training dataset is normalized. An improved frost-reducing algorithm is used to optimize the model parameters of a temperature compensation model, including the initial learning rate and the maximum number of iterations, to obtain the optimal model parameters. The temperature compensation model is a convolutional neural network-based model. The optimization of the model parameters using the improved frost-reducing algorithm to obtain the optimal model parameters includes: Initialize the population, fitness function, global optimal solution of the particle, historical individual optimal solution of the particle, maximum number of iterations T, and current number of iterations t; Based on the fitness function of each particle in the population, the population is divided into a deep mining population and a global search population. The following process is executed iteratively: Update the positions of particles in the deep mining population according to the position update formula of the deep mining learning strategy; Update the positions of particles in the global search population according to the global exploration learning strategy position update formula; A fixed percentage of particles are selected from both the deep mining population and the global search population, and a cross population is generated based on the cross population position update formula generated by the adaptive Lap crossover operator. The merged population is obtained by merging the deep mining, global exploration, and crossover populations. Use aggressive greedy updates to merge the global optimal solution and the historical individual optimal solution in the population; Adjust the particle ratio between the deep mining population and the global search population according to preset rules; The iteration stops when the number of iterations t is greater than or equal to the maximum number of iterations T, and the global optimal solution, i.e. the optimal model parameters, is output. The normalized training dataset is input into the temperature compensation model based on the convolutional neural network, and the temperature compensation model based on the convolutional neural network is trained using the optimal model parameters to obtain the trained temperature compensation model. The pressure data to be measured is input into the trained temperature compensation model for temperature compensation, and the corrected pressure data is obtained.

2. The temperature compensation method for a silicon piezoresistive pressure sensor based on IRIME-CNN-DD according to claim 1, characterized in that, The position update formula for the deep learning strategy is: ； In the formula, Indicates the particle positions in the updated depth-mined population; R best,j Represents the optimal particle; It is a random number with a value in the range (-1, 1); h represents the adhesion degree; and These are the upper and lower bounds of the escape space, respectively. It is a random number with a value range of (0,1); trnd(T+t) is a random number that follows a Student t-distribution; It is a random number with a value range of (0,1).

3. The temperature compensation method for a silicon piezoresistive pressure sensor based on IRIME-CNN-DD according to claim 2, characterized in that, The position update formula for the global exploration learning strategy is: ； In the formula, This indicates the position of particles in the global exploration population after the update.

4. The temperature compensation method for a silicon piezoresistive pressure sensor based on IRIME-CNN-DD according to claim 3, characterized in that, The formula for updating the crossover population position generated by the adaptive Lapp crossover operator is as follows: ； In the formula, Update the position of the particles; is the individual's historical optimal solution; c is a random number taking values ​​in the range [-1, 1]. It is a random number with a value in the range [0,1]. r is a random number in the range [0,1], a is a position parameter, and u is a random number in the range [0,1] that follows a uniform distribution.

5. The temperature compensation method for a silicon piezoresistive pressure sensor based on IRIME-CNN-DD according to claim 1, characterized in that, The fitness function is the mean absolute error.

6. The temperature compensation method for a silicon piezoresistive pressure sensor based on IRIME-CNN-DD according to claim 1, characterized in that, The temperature compensation model based on convolutional neural networks is a temperature compensation model based on convolutional neural networks and dendritic networks.

7. A temperature compensation system for a silicon piezoresistive pressure sensor based on IRIME-CNN-DD, characterized in that, include: The data acquisition module is configured to acquire pressure data at different temperatures as a training dataset. A preprocessing module configured to normalize the training dataset; The parameter optimization module is configured to optimize the model parameters of the temperature compensation model using an improved frost algorithm. The model parameters include the initial learning rate and the maximum number of iterations, to obtain the optimal model parameters. The temperature compensation model is a convolutional neural network-based temperature compensation model. The optimization of the model parameters using the improved frost algorithm to obtain the optimal model parameters includes: Initialize the population, fitness function, global optimal solution of the particle, historical individual optimal solution of the particle, maximum number of iterations T, and current number of iterations t; Based on the fitness function of each particle in the population, the population is divided into a deep mining population and a global search population. The following process is executed iteratively: Update the positions of particles in the deep mining population according to the position update formula of the deep mining learning strategy; Update the positions of particles in the global search population according to the global exploration learning strategy position update formula; A fixed percentage of particles are selected from both the deep mining population and the global search population, and a cross population is generated based on the cross population position update formula generated by the adaptive Lap crossover operator. The merged population is obtained by merging the deep mining, global exploration, and crossover populations. Use aggressive greedy updates to merge the global optimal solution and the historical individual optimal solution in the population; Adjust the particle ratio between the deep mining population and the global search population according to preset rules; The iteration stops when the number of iterations t is greater than or equal to the maximum number of iterations T, and the global optimal solution, i.e. the optimal model parameters, is output. The model training module is configured to input the normalized training dataset into the temperature compensation model based on the convolutional neural network, and train the temperature compensation model based on the convolutional neural network using the optimal model parameters to obtain the trained temperature compensation model. The temperature compensation module is configured to input the pressure data to be measured into a trained temperature compensation model for temperature compensation, and obtain the corrected pressure data.

8. The temperature compensation system for a silicon piezoresistive pressure sensor based on IRIME-CNN-DD according to claim 7, characterized in that, The temperature compensation model based on convolutional neural networks in the parameter optimization module is a temperature compensation model based on convolutional neural networks and dendritic networks.

Images