Sensor precision improvement method and system based on neural network algorithm fusion

Abstract

This invention discloses a method and system for improving sensor accuracy based on neural network algorithm fusion, relating to the field of sensor temperature compensation technology. The method includes acquiring the original measurement signals of multiple pressure points of a pressure sensor at different temperatures, aligning them with a standard input signal to form an original dataset, and constructing a first temperature drift compensation model using a first neural network. Simultaneously, it collects a time-series temperature and pressure value set and calculates the adjacent differences within a preset time period to obtain a temperature and pressure difference value set. The temperature and pressure value set is substituted into the first model to generate a first accurate pressure value set, and the temperature and pressure value set and the difference value set are substituted into a second neural network to generate a second accurate pressure value set. Finally, the two are dynamically fused to obtain an accurate pressure value set. By capturing the nonlinear coupling relationship between temperature and pressure through dual neural networks, the time-series difference feature is introduced to incorporate the influence of dynamic temperature changes and adjacent data differences. Combined with data alignment, data matching is ensured. The dynamic fusion of the dual models further improves the measurement accuracy of the pressure sensor.

Description

Technical Field

[0001] This invention belongs to the field of sensor temperature compensation technology, specifically relating to a method and system for improving sensor accuracy based on neural network algorithm fusion. Background Technology

[0002] With the rapid development of industries such as intelligent manufacturing, smart wearables, and automotive electronics, pressure sensors have become a core component for precise sensing in various electronic devices. Their measurement accuracy directly determines the reliability of equipment functions and the accuracy of data-driven decisions. However, the challenges they face regarding accuracy interference are becoming increasingly complex and diverse, such as the complex coupling between temperature fluctuations and pressure signals, and the impact of instantaneous changes in time-series data under dynamic operating conditions. These problems not only lead to increased sensor measurement deviations but may also cause risks such as equipment control malfunctions, product quality fluctuations, and increased operating costs. Therefore, building an efficient and accurate sensor temperature drift compensation system to achieve real-time correction and accuracy optimization of pressure measurements has become a key requirement for improving the overall performance of equipment. Traditional sensor temperature drift compensation methods often suffer from difficulties in modeling the nonlinear relationship between temperature and pressure, capturing dynamic time-series characteristics, and the poor adaptability of single models, which hinders the compensation effect and makes it difficult to adapt to complex and ever-changing application scenarios.

[0003] Patent CN103048085A discloses a pressure sensor temperature compensation system and method, including a multi-channel sensor test pressure base, a high and low temperature chamber, a pressure controller, a multi-channel bridge test module, a programmable power supply, a multimeter, and an industrial control computer. This system and method enable temperature compensation for batch pressure sensors. Each sensor is tested at different temperatures, and the resistance value in its bridge is calculated. Based on a mathematical model, the industrial control computer precisely calculates the required compensation resistor network for each sensor. Resistors are selected based on the calculated data and soldered into the bridge to form the compensation resistor network, thus completing precise temperature compensation for each sensor. However, this invention does not effectively overcome the limitations of traditional mathematical models in addressing the nonlinear modeling challenges caused by the complex coupling of temperature fluctuations and pressure signals. The compensation lag caused by instantaneous changes in time-series data under dynamic operating conditions still results in insufficient real-time accuracy improvement, thereby reducing the accuracy improvement effect of the pressure sensor. Summary of the Invention

[0004] The purpose of this invention is to solve the problem of low accuracy improvement effect, and to propose a sensor accuracy improvement method and system based on neural network algorithm fusion. In addition, the algorithm described in this invention is not only applicable to pressure sensors, but can also be widely used in the application of instruments such as differential pressure, temperature and humidity, wind speed and air volume, liquid level, and flow rate, and has high versatility and adaptability.

[0005] In a first aspect of this invention, a method for improving sensor accuracy based on neural network algorithm fusion is first proposed, the method comprising:

[0006] By applying pressure to the pressure sensor at different operating temperatures using a preset pressure value, the corresponding detection pressure values ​​at different operating temperatures can be obtained;

[0007] The original input signal is obtained by combining the detection pressure value corresponding to the target temperature and the target temperature; the target temperature is any one of the different operating temperatures.

[0008] The original dataset is obtained by performing a data alignment operation on the original input signal and the standard input signal; the standard input signal is the standard pressure value obtained by applying pressure to the pressure sensor under ideal conditions.

[0009] The first temperature drift compensation model is obtained by processing the original dataset and learning its features using the first neural network algorithm.

[0010] Temperature and pressure values ​​are collected from the pressure sensor over a time series.

[0011] Calculate the difference between adjacent temperature values ​​in the temperature value set within a preset time period to obtain a temperature difference value set;

[0012] The pressure difference set is obtained by calculating the difference between adjacent pressure values ​​in the pressure value set within the preset time period;

[0013] Substituting the temperature value set and the pressure value set into the first temperature drift compensation model, we obtain the first accurate pressure value set;

[0014] Substituting the temperature value set, the pressure value set, the temperature difference value set, and the pressure difference value set into the second neural network model yields the second precise pressure value set;

[0015] The first fluctuation coefficient and the second fluctuation coefficient are obtained by calculating the fluctuation coefficients of the first precise pressure value set and the second precise pressure value set respectively;

[0016] A precise pressure value set is obtained by dynamically fusing the first precise pressure value set and the second precise pressure value set based on the first fluctuation coefficient and the second fluctuation coefficient.

[0017] Optionally, the original dataset is transformed into a grid to obtain a three-dimensional relation matrix; each data point in the three-dimensional relation matrix corresponds to a temperature value, a standard pressure value, and a detection pressure value corresponding to the standard pressure value at that temperature;

[0018] An error dataset is obtained by calculating the difference between the detected pressure value and the standard pressure value for each data point in the three-dimensional relationship matrix; the error data set in the error dataset includes temperature values, standard pressure values, and corresponding error data.

[0019] The temperature values ​​and standard pressure values ​​of each error data group in the error dataset are arranged in order to obtain the input matrix;

[0020] The output matrix is ​​obtained by arranging the error data corresponding to the temperature value and standard pressure value of each error data group in the error dataset in order.

[0021] The first temperature drift compensation model is obtained by establishing a network architecture based on the input matrix and the output matrix.

[0022] Optionally, the input matrix can be connected to two neurons to obtain an input layer connection structure;

[0023] The input layer connection structure is connected to a fully connected layer with a preset depth adjustable to obtain a reinforced intermediate layer connection structure;

[0024] The output matrix is ​​connected to the dynamic weighted neurons to obtain the enhanced output layer connection structure;

[0025] The first temperature drift compensation model is obtained by connecting the intermediate layer connection structure and the enhanced output layer connection structure.

[0026] Optionally, a multi-condition threshold judgment is performed on the temperature value set, the pressure value set, the temperature difference value set, and the pressure difference value set to determine whether a compensation signal is needed;

[0027] If signal compensation is required, a second accurate pressure value is obtained by performing deep time-series fusion calculation on the temperature value set, the pressure value set, the temperature difference value set, and the pressure difference value set.

[0028] If no compensation signal is required, the second precise pressure value is 0.

[0029] Optionally, the temperature value set, the pressure value set, the temperature difference value set, and the pressure difference value set are down-sampled and arranged according to a time window to obtain an initial dataset;

[0030] The first dataset is obtained by extracting features from the initial dataset through a fully connected layer;

[0031] The second dataset is obtained by performing feature encoding on the first dataset through multiple sliding windows of different preset short sizes; the feature encoding operation includes one-dimensional convolution, followed by batch normalization, ReLU activation and max pooling;

[0032] The third dataset is obtained by performing feature encoding on the first dataset using multiple sliding windows with different preset lengths;

[0033] The fourth feature set is obtained by performing dilated convolution on the first dataset;

[0034] The first dataset, the second dataset, the third dataset, and the fourth dataset are weighted and then concatenated along the channel dimension to obtain a fused dataset.

[0035] An enhanced dataset is obtained by performing temporal dynamic modeling of the fused features using LSTM layers;

[0036] The augmented dataset is connected by two fully connected layers to obtain a second accurate stress value set.

[0037] In a second aspect of this invention, a sensor accuracy improvement system based on neural network algorithm fusion is proposed, comprising:

[0038] The pressure value acquisition module is used to apply pressure to the pressure sensor at different operating temperatures based on a preset pressure value to obtain the corresponding detection pressure value at different operating temperatures.

[0039] The raw input signal generation module is used to combine the detection pressure value corresponding to the target temperature and the target temperature to obtain the raw input signal; the target temperature is any one of the different operating temperatures;

[0040] The raw dataset generation module is used to perform data alignment operations on the raw input signal and the standard input signal to obtain the raw dataset; the standard input signal is the standard pressure value obtained by applying pressure to the pressure sensor under ideal conditions.

[0041] The first compensation model generation module is used to process the original dataset and learn its features through a first neural network algorithm to obtain a first temperature drift compensation model.

[0042] The dataset acquisition module is used to acquire temperature and pressure value sets from the pressure sensor over a time series.

[0043] The temperature difference set calculation module is used to calculate the difference between adjacent temperature values ​​in the temperature value set within a preset time period to obtain a temperature difference set.

[0044] The pressure difference set calculation module is used to calculate the difference between adjacent pressure values ​​in the pressure value set within the preset time period to obtain the pressure difference set.

[0045] The first precise pressure value set calculation module is used to substitute the temperature value set and the pressure value set into the first temperature drift compensation model to obtain the first precise pressure value set.

[0046] The second precise pressure value set calculation module is used to substitute the temperature value set, the pressure value set, the temperature difference value set, and the pressure difference value set into the second neural network model to obtain the second precise pressure value set;

[0047] The fluctuation coefficient generation module is used to calculate the fluctuation coefficients of the first precise pressure value set and the second precise pressure value set to obtain the first fluctuation coefficient and the second fluctuation coefficient.

[0048] The precise pressure value set generation module is used to dynamically fuse the first precise pressure value set and the second precise pressure value set according to the first fluctuation coefficient and the second fluctuation coefficient to obtain a precise pressure value set.

[0049] Optionally, the first compensation model generation module includes:

[0050] A three-dimensional relation matrix generation module is used to perform a gridded transformation on the original dataset to obtain a three-dimensional relation matrix; each data point in the three-dimensional relation matrix corresponds to a temperature value, a standard pressure value, and a detection pressure value corresponding to the standard pressure value at that temperature;

[0051] The error dataset calculation module is used to calculate the difference between the detected pressure value and the standard pressure value of each data point in the three-dimensional relationship matrix to obtain the error dataset; the error data group in the error dataset includes temperature value, standard pressure value and corresponding error data.

[0052] The input matrix generation module is used to arrange the temperature values ​​and standard pressure values ​​of each error data group in the error dataset in order to obtain the input matrix;

[0053] The output matrix generation module is used to arrange the error data corresponding to the temperature value and standard pressure value of each error data group in the error dataset in order to obtain the output matrix.

[0054] The network architecture establishment module is used to establish a network architecture based on the input matrix and the output matrix to obtain the first temperature drift compensation model.

[0055] Optionally, the network architecture establishment module includes:

[0056] An input structure generation module is used to connect the input matrix with two neurons to obtain the input layer connection structure;

[0057] The intermediate structure generation module is used to connect the input layer connection structure with a preset depth adjustable fully connected layer to obtain an enhanced intermediate layer connection structure.

[0058] An enhanced output structure generation module is used to connect the output matrix with dynamic weight neurons to obtain an enhanced output layer connection structure.

[0059] A connection module is used to connect the intermediate layer connection structure and the enhanced output layer connection structure to obtain a first temperature drift compensation model.

[0060] Optionally, the second precise pressure value set calculation module includes:

[0061] The judgment module is used to perform multi-condition threshold judgment on the temperature value set, the pressure value set, the temperature difference value set, and the pressure difference value set to determine whether a compensation signal is needed;

[0062] The first condition calculation module is used to perform deep time-series fusion calculation on the temperature value set, the pressure value set, the temperature difference value set, and the pressure difference value set to obtain a second accurate pressure value if a compensation signal is required.

[0063] The second condition calculation module is used to calculate the second precise pressure value as 0 if no compensation signal is required.

[0064] Optionally, the first condition calculation module includes:

[0065] The initial dataset generation module is used to perform frequency reduction extraction on the temperature value set, the pressure value set, the temperature difference value set, and the pressure difference value set, and arrange them according to a time window to obtain the initial dataset;

[0066] The first dataset generation module is used to extract features from the initial dataset through a fully connected layer to obtain the first dataset;

[0067] The second dataset generation module is used to perform feature encoding on the first dataset through multiple sliding windows of different preset short sizes to obtain the second dataset; the feature encoding operation includes one-dimensional convolution, followed by batch normalization, ReLU activation and max pooling;

[0068] The third dataset generation module is used to perform feature encoding on the first dataset through multiple sliding windows with different preset lengths to obtain the third dataset;

[0069] The fourth dataset generation module is used to perform dilated convolution on the first dataset to obtain the second dataset;

[0070] The fused dataset generation module is used to weight the first dataset, the second dataset, the third dataset, and the fourth dataset and concatenate them according to the channel dimension to obtain the fused dataset.

[0071] The augmented dataset generation module is used to obtain an augmented dataset by performing temporal dynamic modeling of the fused features through LSTM layers;

[0072] The second precise stress value set generation module is used to connect the augmented dataset through two fully connected layers to obtain the second precise stress value set.

[0073] The beneficial effects of this invention are as follows: This invention proposes a sensor accuracy improvement method based on neural network algorithm fusion. It acquires the original measurement signals of multiple pressure points from a pressure sensor at different temperatures, aligns them with a standard input signal to form an original dataset, and constructs a first temperature drift compensation model using a first neural network. Simultaneously, it collects time-series temperature and pressure value sets and calculates adjacent differences within a preset time period to obtain a temperature and pressure difference value set. The temperature and pressure value set is substituted into the first model to generate a first accurate pressure value set, and the temperature and pressure value set and the difference value set are substituted into a second neural network to generate a second accurate pressure value set. Finally, the two are dynamically fused to obtain an accurate pressure value set. By capturing the nonlinear coupling relationship between temperature and pressure through dual neural networks, and incorporating time-series difference features to account for dynamic temperature changes and the influence of adjacent data differences, combined with data alignment to ensure data matching, the dynamic fusion of the dual models further improves the measurement accuracy of the pressure sensor. Attached Figure Description

[0074] The invention will now be further described with reference to the accompanying drawings.

[0075] Figure 1 A flowchart illustrating a sensor accuracy improvement method based on neural network algorithm fusion, provided as an embodiment of the present invention;

[0076] Figure 2 This is a framework diagram of a sensor accuracy improvement system based on neural network algorithm fusion, provided for an embodiment of the present invention. Detailed Implementation

[0077] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0078] Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0079] This invention provides a method for improving sensor accuracy based on neural network algorithm fusion. See also... Figure 1 , Figure 1 A flowchart illustrating a sensor accuracy improvement method based on neural network algorithm fusion, provided as an embodiment of the present invention, is shown. The method includes the following steps:

[0080] S101, by applying pressure to the pressure sensor at different operating temperatures using a preset pressure value, the corresponding detection pressure value at different operating temperatures is obtained;

[0081] S102, combine the detection pressure value corresponding to the target temperature and the target temperature to obtain the original input signal;

[0082] S103, perform data alignment operation on the original input signal and the standard input signal to obtain the original dataset;

[0083] S104, The first temperature drift compensation model is obtained by processing the original dataset and learning its features through the first neural network algorithm.

[0084] S105, acquires temperature and pressure value sets from pressure sensors over time;

[0085] S106, calculate the difference between adjacent temperature values ​​in the temperature value set within a preset time period to obtain a temperature difference value set;

[0086] S107, calculate the difference between adjacent pressure values ​​in the pressure value set within a preset time period to obtain the pressure difference value set;

[0087] S108, Substitute the temperature value set and pressure value set into the first temperature drift compensation model to obtain the first accurate pressure value set;

[0088] S109, Substitute the temperature value set, pressure value set, temperature difference value set, and pressure difference value set into the second neural network model to obtain the second accurate pressure value set;

[0089] S110, calculate the fluctuation coefficients of the first precise pressure value set and the second precise pressure value set respectively to obtain the first fluctuation coefficient and the second fluctuation coefficient;

[0090] S111, a precise pressure value set is obtained by dynamically fusing the first precise pressure value set and the second precise pressure value set according to the first fluctuation coefficient and the second fluctuation coefficient.

[0091] The target temperature is any one of the different operating temperatures;

[0092] The standard input signal is the standard pressure value obtained by applying pressure to the pressure sensor under ideal conditions.

[0093] This invention provides a sensor accuracy improvement method based on neural network algorithm fusion. The method acquires raw measurement signals from multiple pressure points at different temperatures from a pressure sensor, aligns them with a standard input signal to form a raw dataset, and constructs a first temperature drift compensation model using a first neural network. Simultaneously, it collects time-series temperature and pressure value sets and calculates adjacent differences within a preset time period to obtain a temperature and pressure difference set. This set is then substituted into the first model to generate a first accurate pressure value set. The temperature and pressure value set and the difference set are substituted into a second neural network to generate a second accurate pressure value set. Finally, the two are dynamically fused to obtain an accurate pressure value set. By using dual neural networks to capture the nonlinear coupling relationship between temperature and pressure, incorporating time-series difference features to account for dynamic temperature changes and the influence of adjacent data differences, and combining data alignment to ensure data matching, the dynamic fusion of the dual models further improves the measurement accuracy of the pressure sensor.

[0094] In one implementation, the preset pressure value and preset time period are set by technicians.

[0095] In one implementation, the original input signals of all pressure points associated with each temperature value are, for example: (-40℃, 0MPa), (-40℃, 0.1MPa), ..., (-40℃, 1MPa); (-35℃, 0MPa), (-35℃, 0.1MPa), ..., (-35℃, 1MPa).

[0096] In one implementation, the system first aligns the data by temperature value groups to ensure that each temperature value of the original input signal can be matched with a group of the same temperature value in the standard input signal. Then, it aligns the data point by pressure point. Under the same temperature group, the original measured value of each pressure point in the original input signal is bound to the standard pressure value of the same pressure point in the standard input signal, forming a quadruple data set of temperature value-pressure point-original pressure measurement value-standard pressure value.

[0097] In one implementation, if the set of temperature values ​​collected in chronological order within a preset time period is <25.0℃, 25.3℃, 25.2℃, 25.5℃, 25.4℃, 25.7℃, 25.8℃>, when calculating the difference between adjacent temperature values, the calculation is performed sequentially according to the rule of subtracting the previous temperature value from the next temperature value, and the final set of temperature difference values ​​is <0.3℃, -0.1℃, 0.3℃, -0.1℃, 0.3℃, 0.1℃>; the calculation logic for pressure difference is similar.

[0098] In one implementation, the first precise pressure value at each time point is designated as P1 and the second precise pressure value is designated as P2. The final value is calculated as P = P1 × weight 1 + P2 × weight 2, where weight 1 + weight 2 = 1. The final values ​​of all time points are organized into a set of precise pressure values, which is used as the final output of the sensor to ensure high-precision measurement in both static and dynamic temperature change scenarios. Weight 1 and weight 2 are set by technicians.

[0099] In one implementation, firstly, for the first precise pressure value set P1 and the second precise pressure value set P2, continuous data sequences within a set time period are extracted respectively. The first volatility coefficient C1 and the second volatility coefficient C2 are obtained by calculating the ratio of the standard deviation to the mean of each data sequence and multiplying it by 100%, thus quantifying the stability of the two sets of data in the time series dimension. Subsequently, the basic weight allocation logic is determined by comparing the magnitudes of C1 and C2: if C1 < C2, it indicates that P1 has better time series stability, and its basic weight is set to 0.6. A weight correction factor is calculated using the formula 0.1 × [(C2 - C1) / C2], and this correction factor is added to the basic weight to obtain the final weight of P1. The weights of P2 and P2 are then adjusted accordingly. The final weight is 1 minus the final weight of P1; if C2 < C1, it indicates that P2 has a better dynamic tracking effect. In the same way, the basic weight of P2 is set to 0.6 and the correction factor is calculated to obtain its final weight. The final weight of P1 is 1 minus the final weight of P2; finally, the pressure values ​​at the corresponding time in P1 and P2 are multiplied by their respective final weights, and the two products are added together to obtain the accurate pressure value at that time. The accurate pressure values ​​at all times are combined in chronological order to form the final set of accurate pressure values.

[0100] In one embodiment, the original dataset is transformed into a grid to obtain a three-dimensional relation matrix; each data point in the three-dimensional relation matrix corresponds to a temperature value, a standard pressure value, and a detection pressure value corresponding to the standard pressure value at that temperature;

[0101] The error dataset is obtained by calculating the difference between the detected pressure value and the standard pressure value for each data point in the three-dimensional relationship matrix. The error data set contains temperature values, standard pressure values, and corresponding error data.

[0102] The input matrix is ​​obtained by arranging the temperature values ​​and standard pressure values ​​of each error data group in the error dataset in order.

[0103] The output matrix is ​​obtained by arranging the error data corresponding to the temperature value and standard pressure value of each error data group in the error dataset in chronological order.

[0104] The first temperature drift compensation model is obtained by establishing a network architecture based on the input and output matrices.

[0105] In one implementation, the temperature range in the original dataset is used as the vertical axis dimension of the two-dimensional grid, and the standard pressure range is used as the horizontal axis dimension. Each original data point in the original dataset is traversed, and the corresponding coordinate position in the two-dimensional grid is found according to the temperature value and the standard pressure value. The measured pressure value is then filled into the grid point to form a three-dimensional correspondence between temperature, standard pressure value, and measured pressure value.

[0106] In one implementation, the error data is calculated as follows: Error data for a single grid point = Measured pressure value - Standard pressure value for that grid point. Each grid point in the two-dimensional data network is traversed, and its temperature, standard pressure, and measured pressure values ​​are extracted. After calculating the error data according to the above rules, these three values ​​are combined into an error data set. For example, if a grid point has a temperature of 25℃, a standard pressure of 0.5MPa, and a measured pressure of 0.502MPa, then the error data set would be (25℃, 0.5MPa, 0.002MPa).

[0107] In one implementation, each row of the input matrix corresponds to an error data set, and each row contains two feature values: the temperature value and the standard pressure value in the error data set. That is, the input matrix is ​​N×2 dimensional, where N is the total number of error data sets.

[0108] In one implementation, by transforming the original dataset into a grid, the discrete temperature and pressure data can be presented in a structured manner, clearly reflecting the distribution of measured pressure values ​​under different temperature and pressure combinations. The calculation error dataset can accurately quantify the measurement deviation at each temperature and pressure point, providing a clear target for model learning. Organizing the temperature and pressure values ​​and error data into input-output matrices respectively can standardize the model training data format, facilitating efficient learning of the neural network. Finally, the first temperature drift compensation model established based on these data can fully capture the nonlinear coupling relationship between temperature and pressure, significantly improving the compensation accuracy for pressure measurement errors under complex temperature change scenarios, and overcoming the limitations of traditional linear compensation methods.

[0109] In one embodiment, the input matrix is ​​connected to two neurons to obtain the input layer connection structure;

[0110] A reinforced intermediate layer connection structure is obtained by connecting the input layer connection structure with a fully connected layer of adjustable preset depth.

[0111] The output matrix is ​​connected to the dynamic weight neurons to obtain the enhanced output layer connection structure;

[0112] The first temperature drift compensation model is obtained by connecting the intermediate layer connection structure and the enhanced output layer connection structure.

[0113] In one implementation, each row of the input matrix contains two features: temperature and standard pressure. Therefore, the input layer has two neurons, each corresponding to one of these features, ensuring a one-to-one mapping between input features and neurons. Each row of data in the input matrix is ​​connected to the two neurons in the input layer via a fully connected manner: the first neuron receives the temperature feature from the sample, and the second neuron receives the standard pressure feature. During the connection process, each feature value directly serves as the input signal for the corresponding neuron. The two neurons in the input layer are precisely connected to the feature dimensions of the input matrix, forming a unidirectional connection structure from the input matrix to the two neurons. This structure serves as the signal entry point for the entire network, responsible for transmitting the temperature and pressure feature signals to the next layer.

[0114] In one implementation, a pre-defined, adjustable-depth fully connected layer, such as layers 1-5, is used, which can be dynamically adjusted according to data complexity. The number of neurons in each layer is set according to feature mapping requirements, such as 32 in the first layer, 64 in the second layer, and 32 in the third layer. Fully connected layers are used between each other. The two neurons in the input layer are connected to all neurons in the first fully connected layer. The output signal of the input layer neurons is transmitted to each neuron in the first fully connected layer through a weight matrix. Each neuron performs a weighted summation of the received signal, which is then processed by an activation function to generate a nonlinear feature mapping result. The output signal of the first fully connected layer is used as the input of the second fully connected layer, and the signal is transmitted layer by layer according to the above logic. Through multi-layer nonlinear transformation, the complex correlation pattern between temperature and pressure features and errors is gradually extracted. Finally, the output of the last fully connected layer forms a reinforced intermediate layer connection structure. This structure contains high-order features that have been deeply mined, providing accurate feature support for the output layer.

[0115] In one implementation, the output layer is configured with one dynamic weight neuron to correspond to the error data in the output matrix. Its weight parameters can be dynamically adjusted based on the output features of the intermediate layers, enhancing adaptability to different temperature and pressure scenarios. The error data in the output matrix serves as the target label, forming a supervisory relationship with the output of the dynamic weight neuron: the neuron's output value needs to be compared with the actual error data in the output matrix, the deviation is calculated using a loss function, and then backpropagation is used to adjust the neuron's weight parameters. The dynamic weight neuron continuously optimizes its own weights by receiving external supervisory signals, forming a closed-loop adjustment mechanism of neuron output, error feedback, and weight update, ultimately constructing an enhanced output layer connection structure that accurately maps the relationship between features and errors.

[0116] In one implementation, the input matrix is ​​precisely connected to two neurons to construct the input layer, ensuring that the temperature and pressure features are efficiently input into the network without dimensional redundancy. The input layer is combined with a pre-defined, adjustable fully connected layer to form a reinforced intermediate layer, which can be flexibly adjusted to adapt to different data complexities and deeply explore the nonlinear relationship between temperature and pressure and error. The output matrix is ​​connected to a dynamically weighted neuron to create an enhanced output layer, which allows the output layer weights to be dynamically optimized according to data features, improving the accuracy of error prediction. Finally, the first temperature drift compensation model, which integrates the reinforced intermediate layer and the enhanced output layer, has both feature extraction depth and output adaptability, can more accurately capture the temperature and pressure coupling law, and significantly improve the temperature drift compensation effect.

[0117] In one embodiment, a multi-condition threshold judgment is performed on the temperature value set, pressure value set, temperature difference set, and pressure difference set to determine whether a compensation signal is needed.

[0118] If signal compensation is required, a second accurate pressure value is obtained by performing deep time-series fusion calculation on the temperature value set, pressure value set, temperature difference value set, and pressure difference value set.

[0119] If no compensation signal is required, the second precise pressure value is 0.

[0120] In one implementation, the multi-condition threshold judgment is to trigger a compensation signal if the temperature value is greater than the absolute temperature value threshold.

[0121] If the pressure value exceeds the absolute pressure threshold, a compensation signal is triggered.

[0122] If the temperature difference is greater than the temperature difference threshold and the pressure difference is greater than the pressure difference threshold, a compensation signal is triggered.

[0123] In this embodiment, the algorithm of this technical solution is not only applicable to pressure sensors, but can also be widely used in instruments such as differential pressure, temperature and humidity, wind speed and air volume, liquid level, and flow rate, and has high versatility and adaptability.

[0124] In one embodiment, the temperature value set, pressure value set, temperature difference set, and pressure difference set are down-sampled and arranged according to a time window to obtain an initial dataset;

[0125] The first dataset is obtained by extracting features from the initial dataset using a fully connected layer;

[0126] The second dataset is obtained by performing feature encoding on the first dataset using multiple sliding windows of different preset short sizes; the feature encoding operation includes one-dimensional convolution, followed by batch normalization, ReLU activation and max pooling;

[0127] The third dataset is obtained by performing feature encoding on the first dataset using multiple sliding windows with different preset lengths;

[0128] The fourth feature set is obtained by performing dilated convolution on the first dataset;

[0129] The first, second, third, and fourth datasets are weighted and then concatenated along the channel dimension to obtain the fused dataset;

[0130] An enhanced dataset is obtained by performing temporal dynamic modeling of the fused features using LSTM layers;

[0131] The augmented dataset is connected by two fully connected layers to obtain a second accurate set of stress values.

[0132] In one implementation, the input dimension of the fully connected layer is four times the length of the time window, and the output dimension is set according to the feature compression requirements. Each neuron establishes connections with all features in the input layer. The fully connected layer performs a weighted summation operation on each initial data point in the initial dataset, transforming the time-series data across the four feature dimensions into a high-dimensional feature vector, achieving preliminary cross-feature information fusion. After processing by the fully connected layer, all initial data points form a first dataset with a shape of initial data quantity × window length × dimension. This dataset retains the temporal structure while enhancing the expressive power of the features.

[0133] In one implementation, technicians set a dilation rate to allow the convolutional kernel to sample at intervals along the temporal dimension, thereby expanding the receptive field without increasing the parameters. Dilated convolution is applied to the first dataset to capture the correlation between discontinuous but distant temporal points, such as the delayed relationship between sudden temperature changes and subsequent pressure changes. The output feature map is then batch normalized and ReLU activated to form a preliminary result. The features processed by dilated convolution are integrated into a fourth feature set, which focuses on characterizing the dependencies between long-distance temporal features.

[0134] In one implementation, the number of hidden units in the LSTM layer is set to utilize its gating mechanism to process the temporal dependencies in the fused dataset and retain key historical information. The LSTM layer processes each time point of the fused dataset in chronological order, filters irrelevant historical features through a forget gate, updates the current features through the input gate, and generates a feature vector containing both historical and current information through the output gate, capturing the dynamic coupling relationship between temperature and pressure over time. The output of the LSTM layer is mapped through a fully connected layer to form an enhanced dataset, whose features contain temporal dynamic evolution information, which is closer to the dynamic change patterns in actual measurement scenarios.

[0135] In one implementation, the augmented dataset is passed through a first fully connected layer and ReLU activation, and then outputs the final pressure prediction value through a second fully connected layer, forming a second accurate pressure value set corresponding to the input time series. The second accurate pressure value set fully integrates multi-scale time series features and dynamic evolution laws, and can effectively compensate for the impact of dynamic temperature changes on pressure measurement.

[0136] Based on the same inventive concept, this invention also provides a sensor accuracy improvement system based on neural network algorithm fusion. See also Figure 2 , Figure 2 A framework diagram of a sensor accuracy improvement system based on neural network algorithm fusion provided in this embodiment of the invention includes:

[0137] The pressure value acquisition module is used to apply pressure to the pressure sensor at different operating temperatures based on a preset pressure value to obtain the corresponding detection pressure value at different operating temperatures.

[0138] The raw input signal generation module is used to combine the detection pressure value corresponding to the target temperature and the target temperature to obtain the raw input signal; the target temperature can be any temperature among different operating temperatures;

[0139] The raw dataset generation module is used to perform data alignment operations on the raw input signal and the standard input signal to obtain the raw dataset; the standard input signal is the standard pressure value obtained by applying pressure to the pressure sensor under ideal conditions.

[0140] The first compensation model generation module is used to process the original dataset and learn its features through the first neural network algorithm to obtain the first temperature drift compensation model.

[0141] The dataset acquisition module is used to acquire temperature and pressure value sets from the pressure sensor over a time series.

[0142] The temperature difference set calculation module is used to calculate the difference between adjacent temperature values ​​in a temperature value set within a preset time period to obtain a temperature difference set.

[0143] The pressure difference set calculation module is used to calculate the difference between adjacent pressure values ​​in the pressure value set within a preset time period to obtain the pressure difference set.

[0144] The first precise pressure value set calculation module is used to substitute the temperature value set and pressure value set into the first temperature drift compensation model to obtain the first precise pressure value set.

[0145] The second precise pressure value set calculation module is used to substitute the temperature value set, pressure value set, temperature difference value set, and pressure difference value set into the second neural network model to obtain the second precise pressure value set;

[0146] The fluctuation coefficient generation module is used to calculate the fluctuation coefficients of the first precise pressure value set and the second precise pressure value set to obtain the first fluctuation coefficient and the second fluctuation coefficient.

[0147] The precise pressure value set generation module is used to dynamically fuse the first precise pressure value set and the second precise pressure value set according to the first fluctuation coefficient and the second fluctuation coefficient to obtain a precise pressure value set.

[0148] This invention provides a sensor accuracy improvement system based on neural network algorithm fusion. It acquires raw measurement signals from multiple pressure points at different temperatures from a pressure sensor, aligns these signals with a standard input signal to form a raw dataset, and constructs a first temperature drift compensation model using a first neural network. Simultaneously, it collects time-series temperature and pressure value sets and calculates adjacent differences within a preset time period to obtain a temperature and pressure difference set. This set is then substituted into the first model to generate a first accurate pressure value set. The temperature and pressure value set and the difference set are substituted into a second neural network to generate a second accurate pressure value set. Finally, the two are dynamically fused to obtain an accurate pressure value set. By using dual neural networks to capture the nonlinear coupling relationship between temperature and pressure, and by introducing time-series difference features to incorporate the dynamic changes in temperature and the influence of adjacent data differences, combined with data alignment to ensure data matching, the dynamic fusion of the dual models further improves the measurement accuracy of the pressure sensor.

[0149] In one embodiment, a three-dimensional relation matrix generation module is used to perform a gridded transformation on the original dataset to obtain a three-dimensional relation matrix; each data point in the three-dimensional relation matrix corresponds to a temperature value, a standard pressure value, and a detection pressure value corresponding to the standard pressure value at that temperature.

[0150] The error dataset calculation module is used to calculate the difference between the detected pressure value and the standard pressure value of each data point in the three-dimensional relationship matrix to obtain the error dataset. The error data group in the error dataset contains temperature value, standard pressure value and corresponding error data.

[0151] The input matrix generation module is used to arrange the temperature values ​​and standard pressure values ​​of each error data group in the error dataset in order to obtain the input matrix.

[0152] The output matrix generation module is used to arrange the error data corresponding to the temperature value and standard pressure value of each error data group in the error dataset in order to obtain the output matrix.

[0153] The network architecture building module is used to build the network architecture based on the input matrix and the output matrix to obtain the first temperature drift compensation model.

[0154] In one embodiment, the network architecture establishment module includes:

[0155] The input structure generation module is used to connect the input matrix with two neurons to obtain the input layer connection structure;

[0156] The enhanced intermediate structure generation module is used to connect the input layer connection structure with a preset depth adjustable fully connected layer to obtain an enhanced intermediate layer connection structure.

[0157] An enhanced output structure generation module is used to connect the output matrix with dynamic weight neurons to obtain the enhanced output layer connection structure.

[0158] The connection module is used to connect the intermediate layer connection structure and the enhanced output layer connection structure to obtain the first temperature drift compensation model.

[0159] In one embodiment, the second precise pressure value set calculation module includes:

[0160] The judgment module is used to perform multi-condition threshold judgment on the temperature value set, pressure value set, temperature difference set, and pressure difference set to determine whether a compensation signal is needed;

[0161] The first condition calculation module is used to perform deep time-series fusion calculation on the temperature value set, pressure value set, temperature difference value set and pressure difference value set to obtain the second accurate pressure value if compensation signal is required.

[0162] The second condition calculation module is used to calculate the second precise pressure value as 0 if no compensation signal is required.

[0163] In one embodiment, the first condition calculation module includes:

[0164] The initial dataset generation module is used to extract the temperature value set, pressure value set, temperature difference set, and pressure difference set by reducing the frequency and arranging them according to the time window to obtain the initial dataset;

[0165] The first dataset generation module is used to extract features from the initial dataset through a fully connected layer to obtain the first dataset;

[0166] The second dataset generation module is used to encode the features of the first dataset through multiple sliding windows of different preset short sizes to obtain the second dataset; the feature encoding operation includes one-dimensional convolution, followed by batch normalization, ReLU activation and max pooling;

[0167] The third dataset generation module is used to encode the features of the first dataset through multiple sliding windows with different preset lengths to obtain the third dataset;

[0168] The fourth dataset generation module is used to perform dilated convolution on the first dataset to obtain the second dataset;

[0169] The merged dataset generation module is used to weight the first, second, third, and fourth datasets and concatenate them according to the channel dimension to obtain the merged dataset.

[0170] The augmented dataset generation module is used to obtain an augmented dataset by performing temporal dynamic modeling of the fused features through LSTM layers;

[0171] The second precise stress value generation module is used to connect the augmented dataset through two fully connected layers to obtain the second precise stress value set.

[0172] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for improving sensor accuracy based on neural network algorithm fusion, characterized in that, The method includes: By applying pressure to the pressure sensor at different operating temperatures using a preset pressure value, the corresponding detection pressure values ​​at different operating temperatures can be obtained; The original input signal is obtained by combining the detection pressure value corresponding to the target temperature and the target temperature; the target temperature is any one of the different operating temperatures. The original dataset is obtained by performing a data alignment operation on the original input signal and the standard input signal; the standard input signal is the standard pressure value obtained by applying pressure to the pressure sensor under ideal conditions. The first temperature drift compensation model is obtained by processing the original dataset and learning its features using the first neural network algorithm. Temperature and pressure values ​​are collected from the pressure sensor over a time series. Calculate the difference between adjacent temperature values ​​in the temperature value set within a preset time period to obtain a temperature difference value set; The pressure difference set is obtained by calculating the difference between adjacent pressure values ​​in the pressure value set within the preset time period; Substituting the temperature value set and the pressure value set into the first temperature drift compensation model, we obtain the first accurate pressure value set; Substituting the temperature value set, the pressure value set, the temperature difference value set, and the pressure difference value set into the second neural network model yields the second precise pressure value set; The first fluctuation coefficient and the second fluctuation coefficient are obtained by calculating the fluctuation coefficients of the first precise pressure value set and the second precise pressure value set respectively; A precise pressure value set is obtained by dynamically fusing the first precise pressure value set and the second precise pressure value set based on the first fluctuation coefficient and the second fluctuation coefficient.

2. The sensor accuracy improvement method based on neural network algorithm fusion according to claim 1, characterized in that, The first temperature drift compensation model is obtained by processing the original dataset and learning its features using the first neural network algorithm, including: The original dataset is transformed into a grid to obtain a three-dimensional relation matrix; each data point in the three-dimensional relation matrix corresponds to a temperature value, a standard pressure value, and a detection pressure value corresponding to the standard pressure value at that temperature. An error dataset is obtained by calculating the difference between the detected pressure value and the standard pressure value for each data point in the three-dimensional relationship matrix; the error data set in the error dataset includes temperature values, standard pressure values, and corresponding error data. The temperature values ​​and standard pressure values ​​of each error data group in the error dataset are arranged in order to obtain the input matrix; The output matrix is ​​obtained by arranging the error data corresponding to the temperature value and standard pressure value of each error data group in the error dataset in order. The first temperature drift compensation model is obtained by establishing a network architecture based on the input matrix and the output matrix.

3. The sensor accuracy improvement method based on neural network algorithm fusion according to claim 2, characterized in that, The network architecture established based on the input matrix and the output matrix yields the first temperature drift compensation model, which includes: The input matrix is ​​connected to two neurons to obtain the input layer connection structure; The input layer connection structure is connected to a fully connected layer with a preset depth adjustable to obtain a reinforced intermediate layer connection structure; The output matrix is ​​connected to the dynamic weighted neurons to obtain the enhanced output layer connection structure; The first temperature drift compensation model is obtained by connecting the reinforced intermediate layer connection structure and the reinforced output layer connection structure.

4. The sensor accuracy improvement method based on neural network algorithm fusion according to claim 1, characterized in that, Substituting the temperature value set, the pressure value set, the temperature difference value set, and the pressure difference value set into the second neural network model yields the second accurate pressure value set, which includes: A multi-condition threshold judgment is performed on the temperature value set, the pressure value set, the temperature difference value set, and the pressure difference value set to determine whether a compensation signal is needed; If signal compensation is required, a second accurate pressure value is obtained by performing deep time-series fusion calculation on the temperature value set, the pressure value set, the temperature difference value set, and the pressure difference value set. If no compensation signal is required, the second precise pressure value is 0.

5. The sensor accuracy improvement method based on neural network algorithm fusion according to claim 4, characterized in that, The second precise pressure value is obtained by performing deep time-series fusion calculation on the temperature value set, the pressure value set, the temperature difference value set, and the pressure difference value set, including: The initial dataset is obtained by down-sampling the temperature value set, the pressure value set, the temperature difference value set, and the pressure difference value set and arranging them according to a time window. The first dataset is obtained by extracting features from the initial dataset through a fully connected layer; The second dataset is obtained by performing feature encoding on the first dataset through multiple sliding windows of different preset short sizes; the feature encoding operation includes one-dimensional convolution, followed by batch normalization, ReLU activation and max pooling; The third dataset is obtained by performing feature encoding on the first dataset using multiple sliding windows with different preset lengths; The fourth dataset is obtained by performing dilated convolution on the first dataset; The first dataset, the second dataset, the third dataset, and the fourth dataset are weighted and then concatenated along the channel dimension to obtain a fused dataset. An enhanced dataset is obtained by performing temporal dynamic modeling of the fused features using LSTM layers; The augmented dataset is connected by two fully connected layers to obtain a second accurate stress value set.

6. A sensor accuracy improvement system based on neural network algorithm fusion, characterized in that, The system includes: The pressure value acquisition module is used to apply pressure to the pressure sensor at different operating temperatures based on a preset pressure value to obtain the corresponding detection pressure value at different operating temperatures. The raw input signal generation module is used to combine the detection pressure value corresponding to the target temperature and the target temperature to obtain the raw input signal; the target temperature is any one of the different operating temperatures; The raw dataset generation module is used to perform data alignment operations on the raw input signal and the standard input signal to obtain the raw dataset; the standard input signal is the standard pressure value obtained by applying pressure to the pressure sensor under ideal conditions. The first compensation model generation module is used to process the original dataset and learn its features through a first neural network algorithm to obtain a first temperature drift compensation model. The dataset acquisition module is used to acquire temperature and pressure value sets from the pressure sensor over a time series. The temperature difference set calculation module is used to calculate the difference between adjacent temperature values ​​in the temperature value set within a preset time period to obtain a temperature difference set. The pressure difference set calculation module is used to calculate the difference between adjacent pressure values ​​in the pressure value set within the preset time period to obtain the pressure difference set. The first precise pressure value set calculation module is used to substitute the temperature value set and the pressure value set into the first temperature drift compensation model to obtain the first precise pressure value set. The second precise pressure value set calculation module is used to substitute the temperature value set, the pressure value set, the temperature difference value set, and the pressure difference value set into the second neural network model to obtain the second precise pressure value set; The fluctuation coefficient generation module is used to calculate the fluctuation coefficients of the first precise pressure value set and the second precise pressure value set to obtain the first fluctuation coefficient and the second fluctuation coefficient. The precise pressure value set generation module is used to dynamically fuse the first precise pressure value set and the second precise pressure value set according to the first fluctuation coefficient and the second fluctuation coefficient to obtain a precise pressure value set.

7. The sensor accuracy improvement system based on neural network algorithm fusion according to claim 6, characterized in that, The first compensation model generation module includes: A three-dimensional relation matrix generation module is used to perform a gridded transformation on the original dataset to obtain a three-dimensional relation matrix; each data point in the three-dimensional relation matrix corresponds to a temperature value, a standard pressure value, and a detection pressure value corresponding to the standard pressure value at that temperature; The error dataset calculation module is used to calculate the difference between the detected pressure value and the standard pressure value of each data point in the three-dimensional relationship matrix to obtain the error dataset; the error data group in the error dataset includes temperature value, standard pressure value and corresponding error data. The input matrix generation module is used to arrange the temperature values ​​and standard pressure values ​​of each error data group in the error dataset in order to obtain the input matrix; The output matrix generation module is used to arrange the error data corresponding to the temperature value and standard pressure value of each error data group in the error dataset in order to obtain the output matrix. The network architecture establishment module is used to establish a network architecture based on the input matrix and the output matrix to obtain the first temperature drift compensation model.

8. The sensor accuracy improvement system based on neural network algorithm fusion according to claim 7, characterized in that, The network architecture establishment module includes: An input structure generation module is used to connect the input matrix with two neurons to obtain the input layer connection structure; The intermediate structure generation module is used to connect the input layer connection structure with a preset depth adjustable fully connected layer to obtain an enhanced intermediate layer connection structure. An enhanced output structure generation module is used to connect the output matrix with dynamic weight neurons to obtain an enhanced output layer connection structure. A connection module is used to connect the reinforced intermediate layer connection structure and the reinforced output layer connection structure to obtain a first temperature drift compensation model.

9. A sensor accuracy improvement system based on neural network algorithm fusion according to claim 6, characterized in that, The second precise pressure value set calculation module includes: The judgment module is used to perform multi-condition threshold judgment on the temperature value set, the pressure value set, the temperature difference value set, and the pressure difference value set to determine whether a compensation signal is needed; The first condition calculation module is used to perform deep time-series fusion calculation on the temperature value set, the pressure value set, the temperature difference value set, and the pressure difference value set to obtain a second accurate pressure value if a compensation signal is required. The second condition calculation module is used to calculate the second precise pressure value as 0 if no compensation signal is required.

10. A sensor accuracy improvement system based on neural network algorithm fusion according to claim 9, characterized in that, The first condition calculation module includes: The initial dataset generation module is used to perform frequency reduction extraction on the temperature value set, the pressure value set, the temperature difference value set, and the pressure difference value set, and arrange them according to a time window to obtain the initial dataset; The first dataset generation module is used to extract features from the initial dataset through a fully connected layer to obtain the first dataset; The second dataset generation module is used to perform feature encoding on the first dataset through multiple sliding windows of different preset short sizes to obtain the second dataset; the feature encoding operation includes one-dimensional convolution, followed by batch normalization, ReLU activation and max pooling; The third dataset generation module is used to perform feature encoding on the first dataset through multiple sliding windows with different preset lengths to obtain the third dataset; The fourth dataset generation module is used to perform dilated convolution on the first dataset to obtain the fourth dataset; The fused dataset generation module is used to weight the first dataset, the second dataset, the third dataset, and the fourth dataset and concatenate them according to the channel dimension to obtain the fused dataset. The augmented dataset generation module is used to obtain an augmented dataset by performing temporal dynamic modeling of the fused features through LSTM layers; The second precise stress value set generation module is used to connect the augmented dataset through two fully connected layers to obtain the second precise stress value set.

Images