A Deep Learning-Based Nonlinear Correction Optimization Method and System for Miniature Semiconductor Temperature Sensors

Abstract

This invention relates to the field of temperature sensor control technology, and in particular to a deep learning-based method for nonlinear correction and optimization of a micro-semiconductor temperature sensor. The method includes acquiring the original temperature sequence of the micro-semiconductor temperature sensor within a preset temperature range, and simultaneously acquiring the error-related environmental parameter sequence and sensor aging coefficient. The original temperature sequence is then processed by empirical mode decomposition (EMD) to extract the main temperature fluctuation mode components carrying nonlinear errors. This invention uses EMD to perform layered processing of the original temperature signal of the micro-semiconductor temperature sensor, enabling precise separation and extraction of the main temperature fluctuation mode components carrying nonlinear errors. Combined with dual-branch neural network data processing and the use of zero-drift compensation factors and sensitivity compensation factors, the method specifically corrects baseline drift and sensitivity decay caused by long-term use, ensuring that the micro-semiconductor temperature sensor maintains high accuracy even after long-term use, meeting the requirements of high-precision applications.

Description

Technical Field

[0001] This invention relates to the field of temperature sensor control technology, and in particular to a method and system for nonlinear correction and optimization of a micro semiconductor temperature sensor based on deep learning. Background Technology

[0002] Miniature semiconductor temperature sensors are miniaturized devices that measure temperature based on the electrical properties of semiconductor materials. Due to their small size, fast response speed, and low cost, they are widely used in smart wearable devices, new energy vehicle battery packs, medical equipment, and other applications. Their core principle is to convert temperature changes into measurable electrical signals through sensitive elements (such as semiconductor PN junctions or thermistors), which are then processed by signal processing circuits to convert into temperature values, meeting the real-time and miniaturized requirements of temperature monitoring in various scenarios.

[0003] However, existing miniature semiconductor temperature sensors still have technical limitations in practical applications: First, nonlinear errors can occur. Due to the characteristics of semiconductor materials, the electrical characteristics of the sensor in the low-temperature (e.g., -40℃) and high-temperature (e.g., 150℃) ranges deviate from the ideal linear relationship with temperature, leading to measurement deviations. Second, there is the issue of time drift and aging. After long-term use, material fatigue and thermal stress can cause baseline drift of the sensitive element. Third, there are environmental interference issues. Humidity changes can affect measurement stability through leakage current, and voltage fluctuations can introduce additional errors. These problems limit the application of miniature semiconductor temperature sensors in high-precision scenarios. Summary of the Invention

[0004] The main objective of this invention is to provide a method and system for nonlinear correction and optimization of micro semiconductor temperature sensors based on deep learning, aiming to solve the technical problems mentioned in the background art.

[0005] This invention proposes a deep learning-based nonlinear correction optimization method for micro-semiconductor temperature sensors, comprising:

[0006] The original temperature sequence of the miniature semiconductor temperature sensor within a preset temperature range is obtained, and the error-related environmental parameter sequence and sensor aging coefficient are obtained simultaneously. The error-related environmental parameter sequence includes ambient humidity data and power supply voltage fluctuation value.

[0007] The original temperature sequence is processed by empirical mode decomposition to extract the main temperature fluctuation mode components, including nonlinear errors.

[0008] Preliminary correction parameters are generated by fusing the main temperature fluctuation mode component and the error-related environmental parameter sequence through a dual-branch neural network. The first branch neural network processes the main temperature fluctuation mode component, and the second branch neural network processes the error-related environmental parameter sequence.

[0009] The cumulative working time of the sensor is obtained, and a zero-drift compensation factor and a sensitivity compensation factor are generated based on the cumulative working time of the sensor and the aging coefficient of the sensor.

[0010] The preliminary correction parameters are weighted and superimposed with the zero drift compensation factor and the sensitivity compensation factor to generate the final correction temperature value.

[0011] Based on the final calibrated temperature value, the model parameters of the miniature semiconductor temperature sensor are calibrated and optimized through an incremental learning mechanism.

[0012] Preferably, the step of processing the original temperature sequence through empirical mode decomposition to extract the main temperature fluctuation mode components including nonlinear errors includes:

[0013] Based on the sensor aging coefficient, the original temperature sequence is segmented by a sliding window to generate multiple time segment sequences;

[0014] Temperature segment sequence sets are extracted from multiple time segment sequences, and error-related environmental parameter sequence sets are acquired simultaneously;

[0015] Each temperature segment sequence is processed by empirical mode decomposition to obtain multiple temperature fluctuation mode components;

[0016] The error characteristic index of each temperature fluctuation mode component is obtained based on the correlation between the temperature fluctuation mode component and the error-related environmental parameter sequence set.

[0017] Calculate the energy percentage of each of the temperature fluctuation mode components;

[0018] The error index of each temperature fluctuation mode component is calculated by wavelet transform;

[0019] The nonlinear error score is calculated by combining the error characteristic index, the error degree index, and the energy ratio. The temperature fluctuation mode components whose nonlinear error scores are greater than a preset value are selected as the main temperature fluctuation mode components.

[0020] Preferably, the step of fusing the main temperature fluctuation mode components and the error-related environmental parameter sequence through a dual-branch neural network to generate preliminary correction parameters includes:

[0021] Based on harmonic component fitting, intrinsic features of temperature information are extracted from the main temperature fluctuation mode components. These intrinsic features include temperature fluctuation intensity features, dominant temperature fluctuation feature frequencies, and nonlinear error features.

[0022] The intrinsic features of the temperature information are input into the first branch network, which is a CNN. The first branch network extracts local dynamic features in the time-frequency domain through convolutional layers and pooling layers, and outputs a temperature feature vector.

[0023] The error-related environmental parameter sequence is input into the second branch network, which is an LSTM. The second branch network performs nonlinear mapping through a fully connected layer and outputs the feature vector of interference factors.

[0024] Based on a gating network, weight coefficients are dynamically assigned to the temperature feature vector and the interference factor feature vector, and the fused feature is obtained by weighted summation.

[0025] The fusion features are processed based on residual blocks, wherein the main branch extracts deep error features through convolution and batch normalization, the skip branch retains the original fusion features, and the residual features are output after merging.

[0026] By stacking multiple layers of the gated network and the residual block, the residual characteristics are gradually optimized;

[0027] The optimized residual features are mapped to preliminary correction parameters through a fully connected layer.

[0028] Preferably, the step of obtaining the cumulative operating time of the sensor and generating a zero-drift compensation factor and a sensitivity compensation factor based on the cumulative operating time of the sensor and the sensor aging coefficient includes:

[0029] The average temperature of the actual working environment is obtained, and the average temperature of the actual working environment is processed by the Arrhenius equation to obtain the aging acceleration factor.

[0030] Obtain the cumulative working time of the sensor, and obtain the equivalent aging time based on the cumulative working time of the sensor and the aging acceleration factor;

[0031] Obtain the actual working environment temperature deviation and temperature deviation aging coefficient, and correct the sensor aging coefficient based on the actual working environment temperature deviation and temperature deviation aging coefficient;

[0032] The zero-drift compensation factor and sensitivity compensation factor are obtained based on the equivalent aging time and the corrected sensor aging coefficient.

[0033] Preferably, the step of generating the final correction temperature value by superimposing the preliminary correction parameters with the zero drift compensation factor and the sensitivity compensation factor includes:

[0034] The temperature range is divided into multiple temperature intervals, and a polynomial model of corresponding order is configured for each temperature interval based on the preliminary correction parameters.

[0035] Calculate the probability weight of each data point in the original temperature sequence belonging to each of the temperature intervals, and perform a weighted summation of the output of the polynomial model for each of the temperature intervals based on the probability weights to obtain the basic corrected temperature sequence.

[0036] The uncertainty of the base corrected temperature sequence is calculated, a confidence coefficient is generated, the fusion weight of the zero drift compensation factor is obtained based on the confidence coefficient, and the original corrected temperature sequence is combined with the zero drift compensation factor according to the fusion weight to obtain a preliminary corrected temperature sequence.

[0037] The initial corrected temperature sequence is combined with the sensitivity compensation factor to obtain the final corrected temperature value with a confidence interval.

[0038] Preferably, the step of calibrating and optimizing the miniature semiconductor temperature sensor based on the final calibrated temperature value using an incremental learning mechanism includes:

[0039] Calculate the error between the final corrected temperature value and the original temperature sequence. When the error exceeds the preset update value, the incremental learning mechanism is automatically triggered.

[0040] Several sets of the latest data are stored in a circular buffer;

[0041] Using the error value as the loss, the gradient of this loss with respect to the weight matrix and nonlinear gain coefficients of the dual-branch neural network is calculated through backpropagation gradient.

[0042] The parameters of the dual-branch neural network are updated based on the results of the backpropagation gradient, wherein the update amount of the weight matrix and nonlinear gain coefficient is proportional to the gradient value, and the update amplitude is controlled by the learning rate, which decays exponentially with the cumulative working time of the sensor.

[0043] The update amount of the zero drift compensation factor is calculated by combining the exponential moving average of the error value, and the update amount of the sensitivity compensation factor is calculated based on the slope of the error value as a function of temperature. The update amounts of the zero drift compensation factor and the sensitivity compensation factor are then subject to synergistic constraints.

[0044] Generate a parameter update package containing updates to the weight matrix, nonlinear gain coefficients, zero drift compensation factor, and sensitivity compensation factor to complete the correction and optimization.

[0045] This invention also discloses a deep learning-based nonlinear correction and optimization system for a miniature semiconductor temperature sensor, characterized by comprising:

[0046] The parameter synchronization acquisition module is used to acquire the original temperature sequence of the miniature semiconductor temperature sensor within a preset temperature range, and simultaneously acquire the error-related environmental parameter sequence and the sensor aging coefficient. The error-related environmental parameter sequence includes ambient humidity data and power supply voltage fluctuation value.

[0047] The empirical mode decomposition module is used to process the original temperature sequence through empirical mode decomposition to extract the main temperature fluctuation mode components, including nonlinear errors.

[0048] A dual-branch neural network fusion module is used to process the main temperature fluctuation mode component and the error-related environmental parameter sequence through dual-branch neural network fusion to generate preliminary correction parameters, wherein the first branch neural network processes the main temperature fluctuation mode component and the second branch neural network processes the error-related environmental parameter sequence.

[0049] An aging compensation factor generation module is used to obtain the cumulative working time of the sensor and generate a zero drift compensation factor and a sensitivity compensation factor based on the cumulative working time of the sensor and the aging coefficient of the sensor.

[0050] The calibration fusion module is used to generate the final calibration temperature value by weighted superposition of the preliminary calibration parameters with the zero drift compensation factor and the sensitivity compensation factor.

[0051] The incremental learning optimization module is used to correct and optimize the model parameters of the miniature semiconductor temperature sensor based on the final corrected temperature value through an incremental learning mechanism.

[0052] Preferably, the empirical mode decomposition module includes:

[0053] A segmentation unit is used to perform sliding window segmentation on the original temperature sequence based on the sensor aging coefficient to generate multiple time segment sequences;

[0054] The segment synchronous acquisition unit is used to extract a temperature segment sequence set from multiple time segment sequences and synchronously acquire an error-related environmental parameter sequence set;

[0055] The segment empirical mode decomposition unit is used to process each temperature segment sequence based on empirical mode decomposition to obtain multiple temperature fluctuation mode components;

[0056] The error feature index extraction unit is used to obtain the error feature index of each temperature fluctuation mode component based on the correlation between the temperature fluctuation mode component and the error-related environmental parameter sequence set.

[0057] An energy percentage calculation unit is used to calculate the energy percentage of each of the temperature fluctuation mode components.

[0058] The error severity index calculation unit calculates the error severity index for each of the temperature fluctuation mode components using wavelet transform;

[0059] The nonlinear error assessment and screening unit is used to calculate the nonlinear error score by combining the error characteristic index, the error degree index and the energy ratio, and to screen the temperature fluctuation mode components whose nonlinear error scores are greater than a preset value as the main temperature fluctuation mode components.

[0060] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described method for nonlinear correction optimization of a micro semiconductor temperature sensor based on deep learning.

[0061] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described method for nonlinear correction optimization of a micro semiconductor temperature sensor based on deep learning.

[0062] The beneficial effects of this invention are as follows: This invention performs layered processing of the raw temperature signal from a miniature semiconductor temperature sensor through empirical mode decomposition, accurately separating and extracting the main temperature fluctuation mode components carrying nonlinear errors. Combined with a dual-branch neural network for dynamic learning of nonlinear laws, this reduces errors across the entire temperature range. Furthermore, it introduces the Arrhenius equation to obtain an aging acceleration factor and calculates the equivalent aging time based on accumulated working time and ambient temperature, generating zero-drift compensation and sensitivity compensation factors. This specifically corrects baseline drift and sensitivity decay caused by long-term use, ensuring the miniature semiconductor temperature sensor maintains high accuracy even after extended use. Additionally, it simultaneously collects environmental parameters such as humidity and voltage, focusing on key interference periods through the LSTM branch of the dual-branch neural network, dynamically allocating environmental interference weights during the feature fusion stage. Combined with an incremental learning mechanism, it can quickly adapt to new interference modes, reducing errors caused by humidity and voltage fluctuations, meeting the requirements of high-precision applications. Attached Figure Description

[0063] Figure 1 This is a schematic diagram of a method flow according to an embodiment of the present invention.

[0064] Figure 2 This is a schematic diagram of the system structure according to an embodiment of the present invention.

[0065] Figure 3 This is a schematic diagram of the internal structure of a computer device according to an embodiment of this application.

[0066] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0067] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0068] like Figure 1 As shown, this application provides a deep learning-based nonlinear correction optimization method for micro semiconductor temperature sensors, including:

[0069] S1, acquire the original temperature sequence of the miniature semiconductor temperature sensor within a preset temperature range, and simultaneously acquire the error-related environmental parameter sequence and the sensor aging coefficient. The error-related environmental parameter sequence includes ambient humidity data and power supply voltage fluctuation value.

[0070] S2, The original temperature sequence is processed by empirical mode decomposition to extract the main temperature fluctuation mode components, including nonlinear errors;

[0071] S3, the main temperature fluctuation mode component and the error-related environmental parameter sequence are fused and processed by a dual-branch neural network to generate preliminary correction parameters, wherein the first branch neural network processes the main temperature fluctuation mode component and the second branch neural network processes the error-related environmental parameter sequence.

[0072] S4, obtain the cumulative working time of the sensor, and generate a zero drift compensation factor and a sensitivity compensation factor based on the cumulative working time of the sensor and the aging coefficient of the sensor.

[0073] S5, the preliminary correction parameters are weighted and superimposed with the zero drift compensation factor and the sensitivity compensation factor to generate the final correction temperature value;

[0074] S6, based on the final calibrated temperature value, uses an incremental learning mechanism to calibrate and optimize the model parameters of the miniature semiconductor temperature sensor.

[0075] As described in steps S1-S6 above, this invention achieves nonlinear error correction of miniature semiconductor temperature sensors over a wide temperature range through a closed-loop correction mechanism involving multi-source data fusion, deep learning modeling, time drift compensation, and incremental learning. This ultimately aims to improve sensor measurement accuracy and extend effective working life. This method addresses the nonlinear error problem caused by material properties, environmental interference, and long-term aging in practical applications of miniature semiconductor temperature sensors, constructing a solution from data acquisition to dynamic optimization.

[0076] Because the electrical characteristics of the core sensitive components (such as semiconductor PN junctions and thermistors) in miniature semiconductor temperature sensors do not have an ideally linear relationship with temperature—at low temperatures (e.g., -40°C), the sensitivity may be low due to a sharp drop in carrier mobility, while at high temperatures (e.g., 150°C), additional resistance shift may occur due to material thermal expansion. This nonlinear relationship directly leads to measurement errors. Simultaneously, ambient humidity increases component leakage current through the "electrolyte film effect," fluctuations in supply voltage affect the stability of the signal amplification circuit, and aging after long-term use causes a slow baseline drift (i.e., "time drift"). These factors collectively lead to a decrease in sensor accuracy. Traditional calibration methods often use fixed formulas (such as polynomial fitting) or single environmental parameter compensation, which cannot cope with the coupled interference of "temperature-humidity-voltage-aging." Therefore, a more comprehensive calibration scheme is needed.

[0077] The core solution of this invention is as follows: Nonlinear error features in temperature signals are extracted through empirical mode decomposition; a dual-branch neural network is used to learn the fluctuation patterns of the temperature itself and the influence of environmental interference; then, time-zero drift compensation factors and sensitivity compensation factors are combined to correct long-term aging errors; finally, dynamic optimization is achieved through incremental learning. The following details each step:

[0078] During the data acquisition phase, the raw temperature sequence of the miniature semiconductor temperature sensor within a preset temperature range is obtained, along with the error-related environmental parameter sequence and the sensor aging coefficient. Specifically, the preset temperature range is set to -40℃ to 150℃ (covering mainstream application scenarios such as industrial and medical applications). The raw temperature sequence is acquired through the sensor's A / D conversion module, with the sampling frequency selected based on the application scenario. For example, for industrial processes requiring precise temperature control, a sampling frequency of no less than 1kHz can be used to ensure signal integrity. Among the error-related environmental parameters, ambient humidity data is synchronously acquired by an integrated humidity sensor (such as SHT30), power supply voltage fluctuations are monitored in real time by a voltage monitoring circuit (such as a voltage divider resistor + operational amplifier), and the sensor aging coefficient is pre-stored or dynamically calculated based on previous aging experiment data (such as the resistance change rate in accelerated aging tests). This synchronous acquisition of data overcomes the limitations of traditional methods that rely solely on temperature data and ignore environmental interference. For example, when the humidity exceeds 60% RH, subsequent steps can specifically enhance humidity-related error correction.

[0079] In the feature extraction stage, the original temperature sequence is processed by empirical mode decomposition (EMD) to extract the main temperature fluctuation mode components, including nonlinear errors, used to separate the "error modes" in the temperature signal. This step decomposes the original temperature sequence into multiple intrinsic mode functions (IMFs) through EMD, each IMF corresponding to a temperature fluctuation mode (such as slow temperature change, high-frequency noise, or nonlinear error fluctuation). Its advantage is adaptive decomposition, requiring no preset frequency range and automatically matching the actual fluctuation characteristics of the temperature signal. For example, nonlinear errors in the low-temperature region may manifest as a specific temperature fluctuation mode component with a frequency concentrated around 0.1 Hz (corresponding to distortion in slow temperature changes). By calculating the nonlinear error score of each temperature fluctuation mode component, the temperature fluctuation mode component carrying the main nonlinear error can be selected, achieving accurate separation of nonlinear errors and providing "focusing error" input for subsequent neural networks.

[0080] In the dual-branch neural network correction stage, the main temperature fluctuation pattern components and the error-related environmental parameter sequence are fused using a dual-branch neural network to generate preliminary correction parameters. This step achieves collaborative learning of "temperature features" and "environmental features" through the dual-branch neural network. The first branch neural network (CNN) processes the main temperature fluctuation pattern components extracted in the previous step: the CNN's convolutional layers (e.g., 3-layer deep separable convolutions, kernel size 7×7→5×5→3×3) can capture the local time-frequency features of temperature fluctuations (e.g., abrupt changes in nonlinear errors at a certain temperature point), and the activation function uses adaptive Leaky ReLU (slope adjusted with aging coefficient) to enhance sensitivity to weak errors. The second branch neural network (LSTM) processes the environmental parameter sequence: the LSTM's gating mechanism can memorize the temporal changes in humidity and voltage (e.g., the continuous impact of a sudden increase in humidity on errors during rainy days), and the attention mechanism focuses on key interference periods (e.g., within 100ms of a voltage drop). The outputs of the two branches are fused to finally generate the preliminary correction parameters. The advantage of this design is that CNN is good at processing spatial / frequency domain features, and LSTM is good at processing temporal features. The combination of the two can cover complex scenarios of "temperature nonlinearity + environmental time-varying interference". For example, in the battery pack of new energy vehicles, it can simultaneously correct the nonlinear error of battery heating and the interference of vehicle voltage fluctuation.

[0081] The time-drift compensation factor is generated by acquiring the sensor's cumulative operating time and generating zero-drift and sensitivity compensation factors based on this cumulative operating time and the sensor's aging coefficient to correct long-term aging errors. This step addresses the long-term aging problem of the sensor by generating compensation factors based on the cumulative operating time and aging coefficient. The zero-drift compensation factor is used to correct baseline offset (e.g., overall measurement values ​​are 0.5℃ higher after one year of use), and the sensitivity compensation factor is used to correct response amplitude decay (e.g., sensitivity decreases by 20% in high-temperature areas). Specifically, the cumulative operating time is recorded by the sensor's built-in RTC clock, and the aging coefficient is dynamically updated using the Arrhenius equation (considering the accelerating effect of temperature on aging). For example, the aging effect of working for 1 hour at 60℃ is equivalent to working for 10 hours at 25℃. This method of combining a physical model with actual operating data solves the problem that traditional fixed-period calibration cannot correct aging errors in real time.

[0082] In the calibration parameter fusion stage, the preliminary calibration parameters are weighted and superimposed with the zero drift compensation factor and the sensitivity compensation factor to generate the final calibration temperature value. This step combines the preliminary calibration parameters with the zero drift compensation factor and the sensitivity compensation factor through weighted superposition and fusion, taking into account both "dynamic nonlinear error" and "static aging error". For example, in smart wearable devices, it can correct the instantaneous temperature fluctuation error during exercise and compensate for the baseline offset after the device has been used for six months.

[0083] In the incremental learning optimization phase, based on the final calibrated temperature value, the model parameters of the miniature semiconductor temperature sensor are calibrated and optimized through an incremental learning mechanism. This step continuously optimizes the model parameters through incremental learning to address the problem of sensor characteristics slowly changing over time. Specifically, when the deviation of the final calibrated temperature value exceeds 0.3℃, incremental learning is triggered: new error features are extracted from the latest acquired data (such as a sudden increase in error in a certain temperature range), the weights of the dual-branch neural network are fine-tuned through backpropagation (such as adjusting the convolution kernel parameters of the CNN), and the calculation models for the zero-drift compensation factor and sensitivity compensation factor are simultaneously updated. Incremental learning employs a strategy of retaining historical knowledge while learning new knowledge to ensure the stability of the optimization process.

[0084] In summary, this method addresses the nonlinear errors, environmental interference, and time drift issues of miniature semiconductor temperature sensors through a process design that includes comprehensive data acquisition, error feature separation, intelligent model learning, physical compensation correction, and dynamic optimization iteration. In practical applications, it can reduce the full-temperature range error compared to traditional methods and maintain high accuracy even after long-term use.

[0085] In one embodiment of the present invention, the step of extracting the main temperature fluctuation mode components, including nonlinear errors, by processing the original temperature sequence through empirical mode decomposition includes:

[0086] S21, Based on the sensor aging coefficient, the original temperature sequence is divided into multiple time segment sequences by a sliding window.

[0087] S22, extract a temperature segment sequence set from the multiple time segment sequences, and simultaneously obtain a set of error-related environmental parameter sequences;

[0088] S23, based on empirical mode decomposition, each temperature segment sequence is processed to obtain multiple temperature fluctuation mode components;

[0089] S24, Obtain the error characteristic index of each temperature fluctuation mode component based on the correlation between the temperature fluctuation mode component and the error-related environmental parameter sequence set;

[0090] S25, calculate the energy percentage of each temperature fluctuation mode component;

[0091] S26, calculate the error index of each temperature fluctuation mode component by wavelet transform;

[0092] S27, calculate the nonlinear error score by combining the error characteristic index, the error degree index and the energy ratio, and select the temperature fluctuation mode component whose nonlinear error score is greater than the preset value as the main temperature fluctuation mode component.

[0093] As described in steps S21-S27 above, this invention, by combining sensor aging characteristics with multi-source parameter correlation analysis, accurately extracts the main temperature fluctuation mode components carrying nonlinear errors from the original temperature sequence, providing high-quality feature input for subsequent dual-branch neural networks. This addresses the problem of insufficient adaptation of traditional empirical mode decomposition to sensor physical characteristics by dynamically adjusting the decomposition strategy and quantizing the error index, thereby improving the extraction accuracy of nonlinear error features.

[0094] Because the nonlinear error of a miniature semiconductor temperature sensor is not uniformly distributed throughout the temperature series, aged sensors may exhibit response delays at low temperatures, resulting in longer "effective fluctuation segments" of the temperature signal. Furthermore, in high-humidity environments, leakage current on the sensor surface introduces high-frequency noise, causing a strong correlation between error characteristics and humidity data. Traditional EMD methods, using fixed window segmentation and single energy thresholds for component selection, are prone to misclassifying slow-varying errors caused by aging as normal temperature fluctuations, or missing critical error components due to noise interference. Therefore, it is necessary to optimize the decomposition process and component selection by considering both the sensor's own condition and environmental parameters.

[0095] The specific implementation and effects of each step:

[0096] First, the "sliding window segmentation of the original temperature sequence based on the sensor aging coefficient" is to adapt to the aging characteristics of the sensor. A higher sensor aging coefficient (such as the resistance change rate obtained from previous aging experiments) indicates a more significant component response delay. In this case, the window size needs to be increased accordingly (for example, 30 sampling points when the aging coefficient is <0.3, and 50 sampling points when it is >0.5) to fully capture the fluctuation cycle of slowly varying errors. This dynamic adjustment avoids the fragmentation of the aging sensor error characteristics caused by a fixed window.

[0097] Secondly, "extracting the temperature segment sequence set and simultaneously acquiring the error-related environmental parameter sequence set" is to establish the correlation between temperature fluctuations and environmental disturbances. The temperature segment sequence set comes from the effective temperature data after segmentation in the previous step, while the error-related environmental parameter sequence set simultaneously includes humidity and voltage data for the corresponding time period (such as the humidity change curve and voltage fluctuation value within 10 seconds corresponding to a certain temperature segment). These data are aligned with timestamps to ensure that the correlation between temperature fluctuations and environmental parameters can be accurately calculated in subsequent analyses. For example, if the fluctuation pattern of a certain temperature segment highly overlaps with the period of rising humidity, it can be preliminarily determined that the fluctuation includes humidity-induced nonlinear errors.

[0098] Next, the iterative selection process of "processing each temperature segment sequence based on empirical mode decomposition to obtain multiple temperature fluctuation mode components" is as follows: For each temperature segment sequence, local maxima and minima are first identified, upper and lower envelopes are constructed using cubic spline interpolation, the envelope mean is calculated, and the trend term is extracted from the original segment. This process is repeated until temperature fluctuation mode components that meet the conditions are obtained (e.g., the number of extreme points and zero crossings is close, and the envelope mean approaches 0). Finally, each temperature segment will be decomposed into multiple temperature fluctuation mode components, corresponding to fluctuations of different frequencies (e.g., low-frequency trend components, mid-frequency nonlinear error components, and high-frequency noise components). Through the adaptive characteristics of empirical mode decomposition, potential error modes can be separated without preset frequencies. For example, the nonlinear error in the high-temperature region may manifest as a specific mid-frequency temperature fluctuation mode component, whose fluctuation pattern is consistent with the thermal hysteresis characteristics of the sensor material.

[0099] The subsequent "error characteristic index" is used to quantify the correlation between temperature fluctuations and environmental disturbances. The error characteristic index is obtained by calculating the Pearson correlation coefficient between each temperature fluctuation model component and the set of error-related environmental parameter sequences (e.g., a correlation coefficient > 0.6 between a component and humidity data). A higher index indicates that the fluctuation of that component is more likely caused by environmental disturbances. For example, in a humid environment, if the correlation coefficient between a temperature fluctuation model component and humidity data reaches 0.75, its error characteristic index will be correspondingly high, suggesting that the temperature fluctuation model component contains humidity-induced nonlinear errors.

[0100] The "energy percentage" is calculated by dividing the energy of each temperature fluctuation mode component by the sum of the energies of all temperature fluctuation mode components. The energy of the temperature fluctuation mode component is obtained by calculating the square of the amplitude of the temperature fluctuation mode component. It is used to determine the intensity percentage of the temperature fluctuation mode component in the temperature signal. Temperature fluctuation mode components with a low energy percentage (e.g., <5%) are usually noise and can be initially excluded; while temperature fluctuation mode components with a high energy percentage (e.g., >40%) may contain major temperature fluctuations or significant errors.

[0101] The calculation of the error severity index for each temperature fluctuation mode component using wavelet transform focuses on the degree of nonlinear distortion of the components. Wavelet transform can capture local abrupt changes in the signal (such as error jumps during a sudden temperature rise). Specifically, the temperature fluctuation mode component is decomposed into three levels using the db4 wavelet basis, and the singular values ​​of the detail coefficients at each level are extracted. The degree of nonlinear distortion is quantified by the singular value entropy (reflecting signal complexity). The higher the singular value entropy, the more nonlinear abrupt changes are contained in the temperature fluctuation mode component, and the larger the error severity index. For example, when the temperature fluctuation mode component exhibits a significant nonlinear jump during a sudden temperature rise, its singular value entropy is 0.8 (significantly higher than the 0.3 of other components), and the corresponding error severity index is also higher.

[0102] Finally, the "nonlinear error score" is obtained by integrating the error characteristic index, error severity index, and energy proportion. The formula can be simplified to: Nonlinear error score equals weight 1 multiplied by the error characteristic index plus weight 2 multiplied by the error severity index plus weight 3 multiplied by the energy proportion. The preset value can be set to 0.15 (based on experimental data statistics). Temperature fluctuation mode components with scores greater than this preset value are selected as the main temperature fluctuation mode components. For example, when the error characteristic index of a temperature fluctuation mode component is 0.6 (strongly correlated with humidity), the error severity index is 0.5 (significant nonlinearity), and the energy proportion is 0.3 (moderate intensity), its comprehensive score is 0.47, which is much higher than the preset value. Therefore, it is identified as the main component carrying nonlinear error.

[0103] Through the above steps, this method can accurately locate and extract nonlinear error features from the original temperature sequence. Compared with traditional methods, it achieves higher accuracy in identifying the main temperature fluctuation pattern components and reduces the false negative rate of error components in complex scenarios such as high humidity and aging. Furthermore, it provides input features focused on nonlinear errors for subsequent dual-branch neural networks, improving the accuracy of generating initial correction parameters.

[0104] In one embodiment of the present invention, the step of generating preliminary correction parameters by fusing the main temperature fluctuation mode components and the error-related environmental parameter sequence through a dual-branch neural network includes:

[0105] S31, based on harmonic component fitting, the intrinsic features of temperature information are extracted from the main temperature fluctuation mode components. These intrinsic features include temperature fluctuation intensity features, dominant temperature fluctuation frequency features, and nonlinear error features. The extraction formula is as follows:

[0106]

[0107] In the formula, I k (t) represents the k-th dominant temperature fluctuation mode component, C m This represents the intensity characteristic of the m-th temperature fluctuation. f represents the characteristic function of the temperature fluctuation period. m Let denot be the m-th dominant temperature fluctuation characteristic frequency, δ(t) represent the nonlinear error characteristic, t represent the time point, and j represent the imaginary unit;

[0108] S32, the intrinsic features of the temperature information are input into the first branch network, which includes a CNN, and extracts local dynamic features in the time-frequency domain through convolutional layers and pooling layers, and outputs a temperature feature vector;

[0109] S33, the error-related environmental parameter sequence is input into the second branch network, which includes an LSTM, and performs nonlinear mapping through a fully connected layer to output the feature vector of interference factors.

[0110] S34, dynamically assign weight coefficients to the temperature feature vector and the interference factor feature vector based on the gating network, and obtain the fused feature by weighted summation;

[0111] S35, The fusion features are processed based on residual blocks, wherein the main branch extracts deep error features through convolution and batch normalization, the skip branch retains the original fusion features, and the residual features are output after merging.

[0112] S36, stack multiple layers of the gated network and the residual block to progressively optimize the residual features;

[0113] S37, the optimized residual features are mapped to preliminary correction parameters through a fully connected layer, the preliminary correction parameters being polynomial coefficients.

[0114] As described in steps S31-S37 above, this invention uses a dual-branch neural network to learn the nonlinear error characteristics of the temperature signal itself and the temporal influence of environmental interference, respectively. Then, through dynamic fusion and deep optimization, it generates preliminary correction parameters that more accurately match the sensor's error characteristics. The core value of this invention lies in utilizing the fitting ability of deep learning to complex features, transforming the coupling relationship between "temperature fluctuation patterns" and "environmental interference" into quantifiable parameters, providing core algorithmic support for the final temperature correction.

[0115] The nonlinear error of a miniature semiconductor temperature sensor is the result of both internal characteristics and external interference. Internal characteristics manifest as nonlinear fluctuations in the temperature signal itself (such as abrupt changes in sensitivity in high-temperature regions), exhibiting significant local characteristics in the time-frequency domain. External interference (such as humidity changes and voltage fluctuations) displays temporal correlation (such as the continuous impact of slowly increasing humidity on the error), requiring the capture of temporal variations. Traditional calibration methods using fixed polynomial fitting cannot simultaneously accommodate both types of characteristics; they either ignore the temporal characteristics of environmental interference or fail to capture the local nonlinearity of the temperature itself, leading to insufficient accuracy of calibration parameters in complex scenarios. Therefore, it is necessary to design dedicated network branches to address the differences between these two types of characteristics and achieve collaborative calibration through a fusion mechanism.

[0116] The shortcomings of traditional methods are mainly reflected in three aspects: First, if a single model is used to process all features, it cannot simultaneously meet the learning needs of local features in the time-frequency domain and temporal features; second, when using fixed-weight feature fusion, it is impossible to dynamically adjust the importance of the two types of features according to environmental or temperature changes (e.g., environmental features should have higher weights in high-humidity environments); third, deep networks are prone to feature degradation, causing effective features from shallow layers to be lost during propagation in deeper layers. This step systematically solves these problems through a combined design of "dual-branch separation learning - gated dynamic fusion - residual feature enhancement".

[0117] Specifically, the implementation and effects of each step are as follows:

[0118] Extracting intrinsic features of temperature information from the main temperature fluctuation pattern components based on harmonic component fitting is a further analysis of these components. Three core features are separated from the main temperature fluctuation pattern components: temperature fluctuation intensity features (reflecting the amplitude of a certain temperature fluctuation pattern, such as a fluctuation intensity of ±2℃), dominant temperature fluctuation characteristic frequency (reflecting the periodicity of the fluctuation, such as a regular fluctuation of 0.5Hz every 2 seconds), and nonlinear error features (reflecting the distortion part that deviates from the linear relationship, such as the extra offset in the high-temperature region). Each temperature fluctuation pattern component can be represented as a superposition of periodic fluctuations of different intensities and frequencies and nonlinear errors. For example, a certain main temperature fluctuation pattern component exhibits significant nonlinearity around 35℃. Its characteristic frequency (0.3Hz), fluctuation intensity (1.5℃), and nonlinear error features (an extra offset of 0.8℃) can be separated using formulas, providing structured input for subsequent CNN processing.

[0119] Specifically, the three types of characteristics are: temperature fluctuation intensity characteristics, dominant temperature fluctuation frequency characteristics, and nonlinear error characteristics. These three types of characteristics correspond to the three physical causes of nonlinear errors in miniature semiconductor temperature sensors. Among them, the temperature fluctuation intensity characteristics reflect the difference in response amplitude of the sensitive element (such as a semiconductor PN junction) at different temperature ranges (such as the decrease in sensitivity in high-temperature regions leading to a weakening of intensity); the dominant temperature fluctuation frequency characteristics correspond to the signal periodicity caused by the sensor's thermal time constant, circuit filtering characteristics, etc. (such as the thermal hysteresis of the thermistor forming a low-frequency fluctuation of 0.1Hz); and the nonlinear error characteristics focus on the deviations caused by factors such as material nonlinearity (such as abrupt changes in carrier mobility) and environmental coupling (such as leakage current distortion caused by humidity).

[0120] The steps for extracting intrinsic temperature information features from the main temperature fluctuation pattern components based on harmonic component fitting are as follows: First, for the selected main temperature fluctuation pattern components, Fourier transform (FFT) is applied to process them, converting the time-domain temperature signal to the frequency domain to generate a frequency spectrum. From this frequency spectrum, the dominant temperature fluctuation characteristic frequencies (e.g., the number of frequencies accounting for approximately 80% of the total energy) are identified based on their energy proportion, ensuring that the most important dominant temperature fluctuation characteristic frequencies in the temperature signal are identified, while ignoring secondary high-frequency or low-frequency noise that has little impact on the overall characteristics. Next, for each identified dominant temperature fluctuation characteristic frequency, the least squares method is used for harmonic component fitting, matching the fitted harmonic component with the original temperature signal under the principle of minimizing the sum of squared errors. In this process, the temperature fluctuation intensity characteristics corresponding to each dominant temperature fluctuation characteristic frequency can be obtained through fitting calculations, intuitively reflecting the energy strength of the temperature signal at that specific dominant temperature fluctuation characteristic frequency; the dominant temperature fluctuation characteristic frequency corresponds to the dominant fluctuation period of the temperature signal, and factors such as the sensor thermal time constant and the time series period of environmental interference are reflected in these characteristic frequencies. After completing the harmonic component fitting, the nonlinear error characteristics are calculated by subtracting the value obtained through harmonic component fitting from the main temperature fluctuation mode component. The principle is that harmonic component fitting approximates temperature fluctuations based on a linear model, assuming the temperature signal is a superposition of several sine waves. However, miniature semiconductor sensors exhibit material nonlinearity and environmental coupling distortion, and the errors caused by these factors cannot be explained by a linear model. Therefore, the residual δ(t) obtained by subtracting the two can accurately focus on these nonlinear error factors, clarifying the correction target for subsequent calibration work.

[0121] The intrinsic features of temperature information are input into the first branch network (CNN) to capture local dynamic features in the time-frequency domain. The CNN branch employs a three-layer depthwise separable convolutional structure (kernel sizes of 7×7, 5×5, and 3×3, respectively). This design progressively focuses on local details—the 7×7 convolutional kernel extracts large-scale frequency distribution features (such as the overall fluctuation period), the 5×5 convolutional kernel captures medium-scale intensity changes (such as the gradient of fluctuation amplitude), and the 3×3 convolutional kernel identifies small-scale nonlinear abrupt changes (such as error jumps at a specific temperature point). The pooling layer uses max pooling (2×2 window) to compress the feature dimension while retaining key features (such as the maximum fluctuation intensity and the most significant feature frequency). The activation function uses adaptive Leaky ReLU (the leakage slope is dynamically adjusted according to the temperature fluctuation intensity) to enhance the response to areas of strong fluctuation (such as high-temperature segments with significant errors). The final output temperature feature vector (e.g., 128 dimensions) condenses the nonlinear error pattern of temperature itself. For example, in smart wearable devices, CNNs can extract nonlinear error features during movement from body temperature fluctuations, providing a targeted basis for subsequent correction.

[0122] Inputting the sequence of error-related environmental parameters into a second branch network (LSTM) allows for the learning of the temporal patterns of environmental disturbances. The LSTM branch comprises two LSTM layers (64 neurons per layer) and one fully connected layer. Its gating mechanism (input gate, forget gate, output gate) effectively memorizes long-term temporal dependencies. For example, as humidity slowly increases from 40% RH to 80% RH, the LSTM can remember the cumulative impact of previous humidity changes on the current error, avoiding the "vanishing gradient" problem of traditional RNNs. The fully connected layer uses the ReLU activation function for nonlinear mapping, transforming temporal features into a fixed-dimensional feature vector of disturbance factors (e.g., 64-dimensional). For example, in new energy vehicle battery packs, the LSTM can learn the continuous impact of power supply voltage fluctuations (e.g., from 12V to 11V) on temperature measurement, and the output feature vector of disturbance factors accurately reflects this correlation.

[0123] The temperature feature vector and the interference factor feature vector are dynamically weighted using a gating network. A weighted summation is then used to obtain the fused feature, achieving adaptive fusion of the two types of features. The gating network takes the temperature feature vector and the interference factor feature vector as input, and generates weight coefficients (α, ranging from 0 to 1) through two fully connected layers (32 hidden neurons) and a sigmoid activation function. Here, α is the weight of the temperature feature vector, and 1-α is the weight of the interference factor feature vector. The weights are dynamically adjusted according to the scene. For example, in a low-humidity environment (RH < 30%), environmental interference is weak, and α = 0.8 (temperature feature dominant); in a high-humidity environment (RH > 70%), humidity has a significant impact, and α = 0.3 (environmental feature dominant). The fused feature obtained after weighted summation (fused feature = α × temperature feature vector + (1-α) × interference factor feature vector) takes into account the main error sources in the current scene, and compared to fixed-weight fusion, it has higher correction adaptability.

[0124] The fused features are processed using residual blocks to enhance deep feature learning. The main branch of the residual block contains one 3×3 convolutional layer (extracting deep error features) and batch normalization (stabilizing training). The skip branch directly passes the original fused features, and the two are element-wise added to output the residual features. This design solves the "feature degradation" problem in deep networks. For example, humidity-related error features in the fused features may be weakened after three convolutional layers. The skip branch can preserve the original features, ensuring that key information is not lost. Compared with ordinary convolutional output, residual features have higher sensitivity to the recognition of weak errors (such as ±0.1℃ offset).

[0125] Stacking multiple layers of the gated network and residual blocks progressively optimizes the residual features, enhancing feature representation capabilities through incremental optimization. Each layer of the gated network dynamically adjusts feature weights, and each layer of the residual block strengthens deeper features. After 3-4 layers of stacking, the residual features can fully capture the coupling relationship between "temperature nonlinearity and environmental temporal interference." For example, in high-temperature and high-humidity scenarios, the first layer captures basic fluctuations, the second layer focuses on the interaction between temperature and humidity, and the third layer strengthens nonlinear distortion features, resulting in higher accuracy in error pattern recognition.

[0126] By mapping the optimized residual features to preliminary correction parameters through a fully connected layer, abstract features can be transformed into directly applicable numerical values. The fully connected layer consists of two layers (64 hidden neurons and 3 output neurons), outputting polynomial coefficients (such as linear coefficients, quadratic coefficients, and a constant term) corresponding to different temperature ranges. Substituting these coefficients into the polynomial formula (corrected temperature = linear coefficient × original temperature + quadratic coefficient × original temperature² + constant term) achieves preliminary correction.

[0127] Through the above steps, this method fully leverages the adaptability of the dual-branch network to different features. The CNN accurately captures the nonlinear fluctuations of temperature itself, the LSTM effectively learns the temporal patterns of environmental interference, and the gating and residual design ensure the effectiveness and deep expressive power of the fused features. The preliminary correction parameters generated in this step can reduce the nonlinear error of the sensor over a wide temperature range and have a more significant effect on suppressing environmental interference such as humidity and voltage, laying a core foundation for the generation of the final corrected temperature value.

[0128] In one embodiment of the present invention, the step of obtaining the cumulative operating time of the sensor and generating a zero-drift compensation factor and a sensitivity compensation factor based on the cumulative operating time of the sensor and the sensor aging coefficient includes:

[0129] S41, obtain the average temperature of the actual working environment, and process the average temperature of the actual working environment through the Arrhenius equation to obtain the aging acceleration factor.

[0130] S42, obtain the cumulative working time of the sensor, and obtain the equivalent aging time based on the cumulative working time of the sensor and the aging acceleration factor, using the following formula:

[0131]

[0132] In the formula, T eq Indicates the equivalent aging time. The calculation results represent the aging acceleration factor, E a The activation energy is represented by kB, the Boltzmann constant is represented by T1, the nominal temperature of the accelerated aging test is represented by T2, and the average temperature of the actual working environment is represented by T. r Indicates the cumulative operating time of the sensor;

[0133] S43, obtain the actual working environment temperature deviation and temperature deviation aging coefficient (representing the degree of influence of temperature deviation on aging; the greater the temperature deviation from the reference value, the more significant the increase in the aging coefficient, and the stronger the effect of high temperature accelerating aging). Based on the actual working environment temperature deviation and temperature deviation aging coefficient, correct the sensor aging coefficient using the following formula:

[0134] α1 = α0 + p*(ΔT) 2 ;

[0135] In the formula, α1 represents the corrected sensor aging coefficient, α0 represents the sensor aging coefficient, p represents the temperature deviation aging coefficient, and ΔT represents the actual working environment temperature deviation.

[0136] S44, obtain the zero drift compensation factor and sensitivity compensation factor based on the equivalent aging time and the corrected sensor aging coefficient.

[0137] As described in steps S41-S44 above, this invention, by combining the temperature characteristics of the sensor's actual working environment with its cumulative operating status, can more accurately quantify the impact of aging on the sensor, generating a zero-drift compensation factor and a sensitivity compensation factor that can dynamically compensate for zero drift caused by long-term use, providing a basis for correcting time-drift errors in the final temperature calibration. Compared to the limitations of traditional periodic calibration, the above steps transform the "physical decay law" of sensor aging into a mathematically calculable compensation model, allowing calibration to cover the entire life cycle and adapt to complex and ever-changing actual working conditions.

[0138] Because the core sensitive components of miniature semiconductor temperature sensors (such as silicon-based PN junctions and thermistors) age due to long-term thermal stress and material fatigue—high-temperature environments accelerate the accumulation of lattice defects inside the components, causing the sensor baseline to drift slowly (zero drift) and sensitivity to gradually decrease (e.g., sensitivity decreases by 2% every 1000 hours of operation). Traditional calibration methods rely solely on factory parameters or periodic offline calibration, failing to track the dynamic correlation between "actual ambient temperature - cumulative operating time - aging degree" in real time, leading to a significant increase in error after long-term use.

[0139] The specific technical implementation is as follows:

[0140] First, the average temperature of the actual working environment is obtained. This average temperature is then processed using the Arrhenius equation to obtain the aging acceleration factor. The average temperature of the actual working environment is continuously monitored by the temperature acquisition module built into the sensor (from the same source as the measurement circuit to ensure consistent timing), and the average value is calculated using a sliding window (e.g., a 1-hour window). This average value is then substituted into the Arrhenius equation to calculate the aging acceleration factor, as shown in the formula: (where E) a T1 is the activation energy, determined by the sensor material properties, such as 1.1 eV for silicon-based PN junctions; kB is the Boltzmann constant; T1 is the nominal temperature for accelerated aging testing, typically 60℃. The higher the ambient temperature, the larger T2 becomes, and the aging acceleration factor increases exponentially, reflecting the accelerating effect of high temperatures on aging.

[0141] Then, the cumulative operating time of the sensors is obtained. Based on the cumulative operating time and the aging acceleration factor, the equivalent aging time is calculated, thus transforming the aging effect of the dynamic environment into a unified metric. The cumulative operating time of the sensors is accumulated in real time by a built-in timer (synchronized with the system clock), and combined with the aging acceleration factor, the equivalent aging time is calculated. This step is a timescale conversion, mapping the actual operating time under varying conditions to the timescale system of the nominal accelerated aging test, allowing the degree of aging to be quantified and compared.

[0142] Next, the actual operating environment temperature deviation and temperature deviation aging coefficient are obtained. Correcting the sensor aging coefficient based on these two parameters addresses the non-linear effects of temperature deviation. The actual operating environment temperature deviation is the difference between the current average ambient temperature and the reference temperature (e.g., the sensor's nominal operating temperature of 25°C). The temperature deviation aging coefficient is an experimentally calibrated non-linear coefficient (fitted through multi-temperature-point aging tests; for example, the coefficient increases by 0.02 for every 1°C deviation), reflecting the physical characteristic that the greater the temperature deviation, the faster the aging process. The initial sensor aging coefficient α0 in the correction formula is calibrated by factory aging tests or prior operational data. This step more fully considers the intensifying effect of high-temperature deviation on aging.

[0143] Finally, the zero-drift compensation factor and sensitivity compensation factor are obtained based on the equivalent aging time and the corrected sensor aging coefficient. Specifically, the zero-drift compensation factor is obtained by multiplying the equivalent aging time and the corrected sensor aging coefficient, and the sensitivity compensation factor is obtained by subtracting the zero-drift compensation factor from 1. The zero-drift compensation factor is a coefficient that compensates for baseline drift. For example, if aging causes the overall measured value to be 0.3℃ higher, or if the sensor's sensitive element is subjected to long-term temperature cycling stress, causing the temperature coefficient of the material's resistance to change slowly, the measurement baseline will gradually shift. The sensitivity compensation factor is a coefficient that restores the response attenuation capability. For example, if long-term thermal stress causes a decrease in the carrier mobility inside the sensitive element (e.g., after 1000 hours of high-temperature operation, the sensitivity decreases by 10%), the measurement amplitude of temperature changes will be compressed (the actual temperature change is 5℃, but the sensor output is only 4.25℃). The sensitivity compensation factor can compensate for the compressed measurement amplitude.

[0144] Through the above steps, physical modeling from ambient temperature acquisition to compensation factor output is realized. The Arrhenius equation quantifies thermally accelerated aging, the equivalent aging time is standardized with a unified time scale, and the temperature deviation aging coefficient corrects nonlinear effects. Finally, the zero-drift compensation factor and sensitivity compensation factor output accurately match the real-time aging state of the sensor. In practical applications, this can make the zero-drift correction accuracy of long-term operating sensors higher and the sensitivity decay compensation error lower, thereby extending the reliable working life of the miniature semiconductor temperature sensor.

[0145] In one embodiment of the present invention, the step of generating a final correction temperature value by superimposing the preliminary correction parameters with a zero-drift compensation factor and a sensitivity compensation factor includes:

[0146] S51, the temperature range is divided into multiple temperature intervals, and a polynomial model of corresponding order is configured for each temperature interval based on the preliminary correction parameters;

[0147] S52, calculate the probability weight of each data point in the original temperature sequence belonging to each of the temperature intervals, and perform a weighted summation of the output of the polynomial model for each of the temperature intervals based on the probability weights to obtain the basic corrected temperature sequence.

[0148] S53, calculate the uncertainty of the basic corrected temperature sequence, generate a confidence coefficient, obtain the fusion weight of the zero drift compensation factor based on the confidence coefficient, and combine the corrected original temperature sequence with the zero drift compensation factor according to the fusion weight to obtain a preliminary corrected temperature sequence.

[0149] S54, combine the preliminary corrected temperature sequence with the sensitivity compensation factor to obtain the final corrected temperature value with confidence interval.

[0150] As described in steps S51-S54 above, this invention utilizes a collaborative mechanism of "dynamic temperature range adaptation - probability-weighted fusion - uncertainty quantification compensation" to deeply integrate preliminary calibration parameters, zero-drift compensation factors, and sensitivity compensation factors, generating a final calibrated temperature value that combines accuracy and confidence. Compared to the limitations of traditional single-model calibration, this invention adapts to the nonlinear differences of temperature sensors across a wide temperature range and quantifies the uncertainty of the calibration process, ensuring that the calibration results are both accurate and reliable, meeting the stringent temperature measurement requirements of industrial and medical applications.

[0151] Because the nonlinear error of a miniature semiconductor temperature sensor is not uniformly distributed globally—the sensor's response characteristics differ significantly across different temperature ranges (e.g., low sensitivity at low temperatures, severe nonlinear distortion at high temperatures). Traditional calibration methods employ a globally uniform model (e.g., fitting a high-order polynomial across the entire temperature range), which cannot accurately adapt to the nonlinear characteristics of each range, leading to larger errors in local temperature ranges. Furthermore, the uncertainty of the calibration results is not quantified, making it impossible to determine the reliability of the calibration values ​​in reliability-sensitive scenarios such as medical cold chain and automotive battery management. Therefore, this invention constructs a mechanism of "range-specific calibration + probabilistic fusion + uncertainty compensation" to achieve a balance between higher accuracy and higher reliability.

[0152] Specifically:

[0153] First, based on the sensor's physical characteristics (such as material phase transition points and sensitivity abrupt change temperatures), the preset temperature range (e.g., -40℃ to 150℃) is divided into multiple intervals (e.g., [-40, -20), [-20, 0), [0, 20)...[130, 150], totaling 8-10 intervals). This division is based on the sensor's factory calibration data (e.g., the segmented characteristics of the thermistor's B value) or extensive prior testing data (e.g., collecting a set of nonlinear characteristics every 10℃ in the high-temperature range). Next, based on the preliminary calibration parameters, a polynomial model of corresponding order is configured for each interval—a 1-2 order polynomial is used for the low-temperature range (e.g., -40℃ to 0℃) due to weak nonlinearity; a 3-4 order polynomial is used for the high-temperature range (e.g., 100℃ to 150℃) due to severe distortion. The calibration model for each interval can be made to accurately match the nonlinear characteristics of that temperature range. For example, if a sensor experiences enhanced nonlinearity in the 80℃~100℃ range due to material thermal expansion, a third-order polynomial can capture its complex "temperature-resistance" curve. Compared to the global first-order model, the calibration error is reduced from ±0.8℃ to ±0.2℃.

[0154] Then, the probability weight of each data point in the original temperature sequence belonging to each interval is calculated to quantify the strength of the temperature point's affiliation within the interval. This is achieved using a Gaussian membership function: the distance between the temperature point and the center of the interval is calculated, and the probability weight is calculated accordingly. Where d represents the distance between the temperature point and the center of the interval, and σ is 1 / 4 of the interval width, controlling the membership decay rate.

[0155] Based on probability weights, the multinomial model outputs for each interval are weighted and summed to obtain the base corrected temperature sequence. This approach can solve the problem of "nonlinear abrupt changes" at interval boundaries. Traditional hard boundary divisions (such as strictly assigning 80℃ to the [80,100] interval) lead to a sharp increase in boundary point errors. However, probability weighting achieves a soft transition, allowing the correction of boundary points to incorporate the advantages of models from adjacent intervals. For example, for a temperature point of 79.5℃, a weight of 0.4 is applied to the [70,80] interval and a weight of 0.6 is applied to the [80,90] interval. The resulting base corrected temperature output has a lower error than that obtained with hard boundary divisions.

[0156] The uncertainty of the base corrected temperature sequence is calculated, and a confidence coefficient is generated. Based on the confidence coefficient, the fusion weight of the zero-drift compensation factor is obtained. Then, the corrected original temperature sequence is combined with the zero-drift compensation factor according to the fusion weight to obtain a preliminary corrected temperature sequence. Specifically, based on the temporal smoothness characteristics of the original temperature sequence, the first-order difference of the temperature change rate at adjacent times is calculated. The fluctuation degree of the first-order difference sequence is statistically analyzed using a sliding window to obtain local fluctuation values, which represent uncertainty. The local fluctuation values ​​are mapped to a confidence coefficient a = exp(-γ*S) using an exponential function. t), where γ is the adjustment coefficient, a parameter used to control the sensitivity of the dynamic confidence coefficient to local fluctuations, and can take values ​​from 0.8 to 1.2, S t This represents the local fluctuation value. When the temperature series fluctuates drastically, the local fluctuation value is large, and the confidence coefficient automatically decreases. Then, the difference between the base corrected temperature series and the preliminary corrected temperature series within the first three sliding windows at time point t is used to calculate the corrected residual, and the root mean square of the residual is calculated for a second calibration of the confidence coefficient: a1=a*exp(-RMS t ), where a1 represents the confidence coefficient after secondary calibration, and RMS t This represents the root mean square (RMS) of the residuals of the base-corrected temperature series and the initial-corrected temperature series at time point t. The above method quantifies the reliability of the correction process. If the polynomial model has a large fitting error in the high-temperature range for a certain interval, the RMS of the residuals will increase significantly, and the confidence coefficient will decrease. The fusion weight of the zero-drift compensation factor is obtained based on the confidence coefficient: when the confidence coefficient is high, the zero-drift compensation factor is assigned a low weight because the base correction is already sufficiently accurate; when the confidence coefficient is low, the weight of the zero-drift compensation factor is increased to compensate for model deficiencies. Through this dynamic adjustment compensation strategy, zero-drift compensation avoids excessive interference in the accurate interval while promptly filling in intervals with large errors, thus reducing the standard deviation of the compensated error.

[0157] The preliminary corrected temperature series is combined with a sensitivity compensation factor to generate the final corrected temperature value with a confidence interval. The combination method involves multiplying the preliminary corrected temperature series by the sensitivity compensation factor, and simultaneously calculating the final confidence interval (e.g., T) based on the uncertainty of the base corrected temperature series (correction residuals) and the model error of the sensitivity compensation factor (e.g., calibration error of the sensitivity compensation factor). z -q,T z +q), where T z The final corrected temperature value is represented by q, which represents the total error standard deviation of the model error of the correction residual and sensitivity compensation factor. q is used to quantify the degree of random error fluctuation in the temperature correction process, reflecting the degree of dispersion relative to the true temperature. The smaller the value, the more stable and reliable the correction is.

[0158] In summary, this invention improves correction accuracy by dividing temperature ranges to adapt to nonlinearity, using probabilistic fusion to eliminate boundary abrupt changes, and employing a confidence-driven compensation strategy. It also adds a confidence range to the correction results, enabling the system to make decisions based on error boundaries and thus improving system reliability.

[0159] In one embodiment of the present invention, the step of calibrating and optimizing the miniature semiconductor temperature sensor based on the final calibrated temperature value using an incremental learning mechanism includes:

[0160] S61, calculate the error between the final corrected temperature value and the original temperature sequence. When the error exceeds the preset update value, the incremental learning mechanism is automatically triggered.

[0161] S62 stores several sets of the latest data through a circular buffer;

[0162] S63, using the error value as the loss, the gradient of the loss with respect to the weight matrix and nonlinear gain coefficient of the dual-branch neural network is calculated through backpropagation gradient calculation, and the gradient calculation is implemented through hardware acceleration of systolic array.

[0163] S64, the parameters of the dual-branch neural network are updated based on the backpropagation gradient, wherein the update amount of the weight matrix and nonlinear gain coefficient is proportional to the gradient value, and the update amplitude is controlled by the learning rate, wherein the learning rate decays exponentially with the cumulative working time of the sensor.

[0164] S65, calculate the update amount of the zero drift compensation factor by combining the exponential moving average of the error value, and calculate the update amount of the sensitivity compensation factor based on the slope of the error value changing with temperature, and perform collaborative constraints on the update amount of the zero drift compensation factor and the update amount of the sensitivity compensation factor.

[0165] S66 generates a parameter update package containing updates to the weight matrix, nonlinear gain coefficients, zero drift compensation factor, and sensitivity compensation factor, thus completing the correction optimization.

[0166] As described in steps S61-S66 above, this invention tracks the dynamic changes in sensor performance in real time (such as accelerated aging and changes in environmental interference patterns) through an incremental learning mechanism. It transforms the error between the final corrected temperature value and the original temperature sequence into the driving force for updating model parameters, achieving a closed-loop correction of error monitoring, data caching, parameter optimization, and compensation coordination. This ensures that the miniature semiconductor temperature sensor maintains high accuracy throughout its lifespan. Compared to the limitations of periodic offline calibration, this solution allows model parameters, zero-drift compensation factors, and sensitivity compensation factors to dynamically evolve with the actual state of the sensor, adapting to the new patterns of nonlinear characteristic drift and environmental interference during long-term use.

[0167] Because the characteristics of miniature semiconductor temperature sensors are not static, the material properties of the sensing element (such as carrier mobility) will slowly change with aging over long-term use, and environmental interference patterns (such as humidity fluctuation frequency and voltage fluctuation amplitude) may also change due to changes in the application scenario (such as from indoors to outdoors). These changes will cause the parameters of the previously trained model to gradually become invalid, increasing the calibration error. Traditional calibration methods, through periodic compensation, cannot adapt to these dynamic changes in real time and will also accumulate errors due to calibration lag. Therefore, this invention constructs a mechanism that can automatically monitor errors, incrementally update parameters, and collaboratively optimize compensation, enabling the sensor to have self-updating capabilities.

[0168] Specific implementation steps:

[0169] First, the error between the final corrected temperature value and the original temperature sequence is calculated. When this error exceeds a preset update value, an incremental learning mechanism is automatically triggered. The error value is used to quantify the difference between the system model and the actual physical state. The preset update value is set according to the accuracy requirements of the application scenario (e.g., 0.3℃ for medical-grade scenarios and 0.5℃ for industrial-grade scenarios). When the preset update value exceeds this value, the incremental learning mechanism is automatically triggered. This enables dynamic perception of the risk of correction failure. For example, after six months of use, the sensor in a smart wearable device may experience sensitivity drift due to long-term corrosion from sweat, causing the error at a certain body temperature measurement point to increase from 0.2℃ to 0.6℃ (exceeding the 0.3℃ threshold). At this point, incremental learning is triggered to prevent the error from continuing to expand.

[0170] A circular buffer stores several sets of the latest data. These data sets include: the final calibrated temperature value, the original temperature sequence, the error-related environmental parameter sequence (humidity, voltage), the sensor aging coefficient, and the cumulative operating time. The advantage of the circular buffer is that it dynamically overwrites old data, always retaining the latest data and avoiding storage redundancy. This data serves as new samples for incremental learning, recording the complete scenario where the error occurred. For example, among 100 sets of data where the error exceeded the limit, 80% were concentrated in scenarios with humidity > 80%, indicating a new "high humidity-error" correlation pattern and providing targeted basis for subsequent parameter updates. Compared to randomly sampled data, this invention, through error-driven cached data, allows the model to locate problems more quickly, improving parameter update efficiency.

[0171] Using the error value as the loss, the gradient of this loss with respect to the weight matrix and nonlinear gain coefficients of the dual-branch neural network is calculated through backpropagation gradient calculation. This gradient calculation is implemented using hardware acceleration via a pulsating array. The gradient reflects the degree of influence of small parameter changes on the error; the larger the absolute value of the gradient, the greater the contribution of that parameter to the current error, requiring significant adjustment. To meet real-time requirements, gradient calculation can be implemented using hardware acceleration via a pulsating array. The pulsating array decomposes matrix operations into parallel local operations, avoiding redundant data transfer and reducing gradient calculation latency, ensuring that the calculation is completed within the sensor sampling period without affecting the normal temperature measurement process.

[0172] The parameters of the dual-branch neural network are updated based on the results of the backpropagation gradient calculation. The update amounts of the weight matrix and nonlinear gain coefficients are proportional to the gradient values, and the update magnitude is controlled by the learning rate, which decays exponentially with the cumulative working time of the sensor. The weight matrix... ω0 represents the weight matrix before the update, ω n Represents the weight matrix. This represents the result of backpropagation gradient calculation on the weight matrix, and the nonlinear gain coefficient. θ n Let θ0 represent the nonlinear gain coefficient, and η represent the learning rate, which controls the update amplitude and decreases exponentially with the cumulative working time of the sensor. η0 represents the initial learning rate, η represents the learning rate, and λ represents the decay constant, which controls the rate at which the learning rate decays over time. r This indicates the cumulative operating time of the sensor. The more severe the sensor aging, the smoother the parameter updates, preventing parameter oscillations caused by sudden errors due to decreased stability in an aging sensor. Simultaneously, parameter updates can be constrained, for example, by setting the second norm of the weight matrix ||ω. n ||2<10, and the nonlinear gain coefficient γ∈[0.1,0.3] prevents parameter updates from deviating excessively from the physically reasonable range. Through the above scheme, this invention can achieve precise fine-tuning of parameter updates with lower time consumption.

[0173] The update amount of the zero drift compensation factor is calculated by combining the exponential moving average of the error value, and the update amount of the sensitivity compensation factor is calculated based on the slope of the error value as a function of temperature. The update amounts of the zero drift compensation factor and the sensitivity compensation factor are then subject to synergistic constraints.

[0174] The zero-drift compensation factor update is calculated based on the exponential moving average of the error value, smoothing out the impact of sudden errors (such as error spikes caused by transient electromagnetic interference) and ensuring more stable zero-drift compensation. The sensitivity compensation factor update is calculated based on the slope of the error change with temperature; the larger the slope, the more significant the sensitivity deviation, and the larger the update amount. To avoid excessive superposition of zero drift and sensitivity compensation, a synergistic constraint is set. For example, when the zero-drift compensation factor update is >0.2℃, the sensitivity compensation factor update is automatically multiplied by 0.7 to prevent double compensation from causing new deviations and to avoid excessively high temperatures after correction.

[0175] Finally, a parameter update package is generated, which includes updates to the weight matrix, nonlinear gain coefficients, zero drift compensation factor, and sensitivity compensation factor. This completes the calibration and optimization process. After encryption verification (such as CRC check), the package is written to the sensor storage module, thus completing the calibration and optimization.

[0176] In summary, the present invention enables the miniature semiconductor temperature sensor to accurately capture new error patterns and stably retain historical correction knowledge in long-term complex scenarios, thereby achieving high-precision nonlinear correction optimization.

[0177] like Figure 2 As shown, this invention also discloses a deep learning-based nonlinear correction and optimization system for a micro semiconductor temperature sensor, comprising:

[0178] The parameter synchronization acquisition module is used to acquire the original temperature sequence of the miniature semiconductor temperature sensor within a preset temperature range, and simultaneously acquire the error-related environmental parameter sequence and the sensor aging coefficient. The error-related environmental parameter sequence includes ambient humidity data and power supply voltage fluctuation value.

[0179] The empirical mode decomposition module is used to process the original temperature sequence through empirical mode decomposition to extract the main temperature fluctuation mode components, including nonlinear errors.

[0180] A dual-branch neural network fusion module is used to process the main temperature fluctuation mode component and the error-related environmental parameter sequence through dual-branch neural network fusion to generate preliminary correction parameters, wherein the first branch neural network processes the main temperature fluctuation mode component and the second branch neural network processes the error-related environmental parameter sequence.

[0181] An aging compensation factor generation module is used to obtain the cumulative working time of the sensor and generate a zero drift compensation factor and a sensitivity compensation factor based on the cumulative working time of the sensor and the aging coefficient of the sensor.

[0182] The calibration fusion module is used to generate the final calibration temperature value by weighted superposition of the preliminary calibration parameters with the zero drift compensation factor and the sensitivity compensation factor.

[0183] The incremental learning optimization module is used to correct and optimize the model parameters of the miniature semiconductor temperature sensor based on the final corrected temperature value through an incremental learning mechanism.

[0184] Preferably, the empirical mode decomposition module includes:

[0185] A segmentation unit is used to perform sliding window segmentation on the original temperature sequence based on the sensor aging coefficient to generate multiple time segment sequences;

[0186] The segment synchronous acquisition unit is used to extract a temperature segment sequence set from multiple time segment sequences and synchronously acquire an error-related environmental parameter sequence set;

[0187] The segment empirical mode decomposition unit is used to process each temperature segment sequence based on empirical mode decomposition to obtain multiple temperature fluctuation mode components;

[0188] The error feature index extraction unit is used to obtain the error feature index of each temperature fluctuation mode component based on the correlation between the temperature fluctuation mode component and the error-related environmental parameter sequence set.

[0189] An energy percentage calculation unit is used to calculate the energy percentage of each of the temperature fluctuation mode components.

[0190] The error severity index calculation unit calculates the error severity index for each of the temperature fluctuation mode components using wavelet transform;

[0191] The nonlinear error assessment and screening unit is used to calculate the nonlinear error score by combining the error characteristic index, the error degree index and the energy ratio, and to screen the temperature fluctuation mode components whose nonlinear error scores are greater than a preset value as the main temperature fluctuation mode components.

[0192] like Figure 3 As shown, the present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described method for nonlinear correction optimization of a micro semiconductor temperature sensor based on deep learning.

[0193] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described method for nonlinear correction optimization of a micro semiconductor temperature sensor based on deep learning.

[0194] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0195] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A nonlinear correction and optimization method for a micro-semiconductor temperature sensor based on deep learning, characterized in that, include: The original temperature sequence of the miniature semiconductor temperature sensor within a preset temperature range is obtained, and the error-related environmental parameter sequence and sensor aging coefficient are obtained simultaneously. The error-related environmental parameter sequence includes ambient humidity data and power supply voltage fluctuation value. Based on the sensor aging coefficient, the original temperature sequence is segmented by a sliding window to generate multiple time segment sequences; Temperature segment sequence sets are extracted from multiple time segment sequences, and error-related environmental parameter sequence sets are acquired simultaneously; Each temperature segment sequence is processed by empirical mode decomposition to obtain multiple temperature fluctuation mode components; The error characteristic index of each temperature fluctuation mode component is obtained based on the correlation between the temperature fluctuation mode component and the error-related environmental parameter sequence set. Calculate the energy percentage of each of the temperature fluctuation mode components; The error index of each temperature fluctuation mode component is calculated by wavelet transform; The nonlinear error score is calculated by combining the error characteristic index, the error degree index, and the energy ratio. The temperature fluctuation mode components whose nonlinear error scores are greater than a preset value are selected as the main temperature fluctuation mode components. Preliminary correction parameters are generated by fusing the main temperature fluctuation mode component and the error-related environmental parameter sequence through a dual-branch neural network. The first branch neural network processes the main temperature fluctuation mode component, and the second branch neural network processes the error-related environmental parameter sequence. The cumulative working time of the sensor is obtained, and a zero-drift compensation factor and a sensitivity compensation factor are generated based on the cumulative working time of the sensor and the aging coefficient of the sensor. The preliminary correction parameters are weighted and superimposed with the zero drift compensation factor and the sensitivity compensation factor to generate the final correction temperature value. Based on the final calibrated temperature value, the model parameters of the miniature semiconductor temperature sensor are calibrated and optimized through an incremental learning mechanism.

2. The nonlinear correction and optimization method for a micro-semiconductor temperature sensor based on deep learning according to claim 1, characterized in that, The step of fusing the main temperature fluctuation mode components and the error-related environmental parameter sequence through a dual-branch neural network to generate preliminary correction parameters includes: Based on harmonic component fitting, intrinsic features of temperature information are extracted from the main temperature fluctuation mode components. These intrinsic features include temperature fluctuation intensity features, dominant temperature fluctuation feature frequencies, and nonlinear error features. The intrinsic features of the temperature information are input into the first branch network, which is a CNN. The first branch network extracts local dynamic features in the time-frequency domain through convolutional layers and pooling layers, and outputs a temperature feature vector. The error-related environmental parameter sequence is input into the second branch network, which is an LSTM. The second branch network performs nonlinear mapping through a fully connected layer and outputs the feature vector of interference factors. Based on a gating network, weight coefficients are dynamically assigned to the temperature feature vector and the interference factor feature vector, and the fused feature is obtained by weighted summation. The fusion features are processed based on residual blocks, wherein the main branch extracts deep error features through convolution and batch normalization, the skip branch retains the original fusion features, and the residual features are output after merging. By stacking multiple layers of the gated network and the residual block, the residual characteristics are gradually optimized; The optimized residual features are mapped to preliminary correction parameters through a fully connected layer.

3. The nonlinear correction and optimization method for a micro-semiconductor temperature sensor based on deep learning according to claim 1, characterized in that, The step of obtaining the cumulative operating time of the sensor and generating a zero-drift compensation factor and a sensitivity compensation factor based on the cumulative operating time of the sensor and the sensor aging coefficient includes: The average temperature of the actual working environment is obtained, and the average temperature of the actual working environment is processed by the Arrhenius equation to obtain the aging acceleration factor. Obtain the cumulative working time of the sensor, and obtain the equivalent aging time based on the cumulative working time of the sensor and the aging acceleration factor; Obtain the actual working environment temperature deviation and temperature deviation aging coefficient, and correct the sensor aging coefficient based on the actual working environment temperature deviation and temperature deviation aging coefficient; The zero-drift compensation factor and sensitivity compensation factor are obtained based on the equivalent aging time and the corrected sensor aging coefficient.

4. The nonlinear correction and optimization method for a micro semiconductor temperature sensor based on deep learning according to claim 1, characterized in that, The step of generating the final correction temperature value by weighted superposition of the preliminary correction parameters, the zero drift compensation factor, and the sensitivity compensation factor includes: The temperature range is divided into multiple temperature intervals, and a polynomial model of corresponding order is configured for each temperature interval based on the preliminary correction parameters. Calculate the probability weight of each data point in the original temperature sequence belonging to each of the temperature intervals, and perform a weighted summation of the output of the polynomial model for each of the temperature intervals based on the probability weights to obtain the basic corrected temperature sequence. The uncertainty of the base corrected temperature sequence is calculated, a confidence coefficient is generated, the fusion weight of the zero drift compensation factor is obtained based on the confidence coefficient, and the original corrected temperature sequence is combined with the zero drift compensation factor according to the fusion weight to obtain a preliminary corrected temperature sequence. The initial corrected temperature sequence is combined with the sensitivity compensation factor to obtain the final corrected temperature value with a confidence interval.

5. The nonlinear correction and optimization method for a micro-semiconductor temperature sensor based on deep learning according to claim 1 further includes the step of optimizing the micro-semiconductor temperature sensor through an incremental learning mechanism based on the final corrected temperature value: Calculate the error between the final corrected temperature value and the original temperature sequence. When the error exceeds the preset update value, the incremental learning mechanism is automatically triggered. Several sets of the latest data are stored in a circular buffer; Using the error value as the loss, the gradient of this loss with respect to the weight matrix and nonlinear gain coefficients of the dual-branch neural network is calculated through backpropagation gradient. The parameters of the dual-branch neural network are updated based on the results of the backpropagation gradient, wherein the update amount of the weight matrix and nonlinear gain coefficient is proportional to the gradient value, and the update amplitude is controlled by the learning rate, which decays exponentially with the cumulative working time of the sensor. The update amount of the zero drift compensation factor is calculated by combining the exponential moving average of the error value, and the update amount of the sensitivity compensation factor is calculated based on the slope of the error value as a function of temperature. The update amounts of the zero drift compensation factor and the sensitivity compensation factor are then subject to synergistic constraints. Generate a parameter update package containing updates to the weight matrix, nonlinear gain coefficients, zero drift compensation factor, and sensitivity compensation factor to complete the correction and optimization.

6. A nonlinear correction and optimization system for a micro semiconductor temperature sensor based on deep learning, characterized in that, include: The parameter synchronization acquisition module is used to acquire the original temperature sequence of the miniature semiconductor temperature sensor within a preset temperature range, and simultaneously acquire the error-related environmental parameter sequence and the sensor aging coefficient. The error-related environmental parameter sequence includes ambient humidity data and power supply voltage fluctuation value. The empirical mode decomposition module is used to perform sliding window segmentation on the original temperature sequence based on the sensor aging coefficient to generate multiple time segment sequences; extract a temperature segment sequence set from the multiple time segment sequences, and simultaneously acquire an error-related environmental parameter sequence set; Each temperature segment sequence is processed by empirical mode decomposition to obtain multiple temperature fluctuation mode components; The error characteristic index of each temperature fluctuation mode component is obtained based on the correlation between the temperature fluctuation mode component and the error-related environmental parameter sequence set. Calculate the energy percentage of each of the temperature fluctuation mode components; The error index of each temperature fluctuation mode component is calculated by wavelet transform; The nonlinear error score is calculated by combining the error characteristic index, the error degree index, and the energy ratio. The temperature fluctuation mode components whose nonlinear error scores are greater than a preset value are selected as the main temperature fluctuation mode components. A dual-branch neural network fusion module is used to process the main temperature fluctuation mode component and the error-related environmental parameter sequence through dual-branch neural network fusion to generate preliminary correction parameters, wherein the first branch neural network processes the main temperature fluctuation mode component and the second branch neural network processes the error-related environmental parameter sequence. An aging compensation factor generation module is used to obtain the cumulative working time of the sensor and generate a zero drift compensation factor and a sensitivity compensation factor based on the cumulative working time of the sensor and the aging coefficient of the sensor. The calibration fusion module is used to generate the final calibration temperature value by weighted superposition of the preliminary calibration parameters with the zero drift compensation factor and the sensitivity compensation factor. The incremental learning optimization module is used to correct and optimize the model parameters of the miniature semiconductor temperature sensor based on the final corrected temperature value through an incremental learning mechanism.

7. The nonlinear correction and optimization system for a micro semiconductor temperature sensor based on deep learning according to claim 6, characterized in that, The empirical mode decomposition module includes: A segmentation unit is used to perform sliding window segmentation on the original temperature sequence based on the sensor aging coefficient to generate multiple time segment sequences; The segment synchronous acquisition unit is used to extract a temperature segment sequence set from multiple time segment sequences and synchronously acquire an error-related environmental parameter sequence set; The segment empirical mode decomposition unit is used to process each temperature segment sequence based on empirical mode decomposition to obtain multiple temperature fluctuation mode components; The error feature index extraction unit is used to obtain the error feature index of each temperature fluctuation mode component based on the correlation between the temperature fluctuation mode component and the error-related environmental parameter sequence set. An energy percentage calculation unit is used to calculate the energy percentage of each of the temperature fluctuation mode components. The error severity index calculation unit calculates the error severity index for each of the temperature fluctuation mode components using wavelet transform; The nonlinear error assessment and screening unit is used to calculate the nonlinear error score by combining the error characteristic index, the error degree index and the energy ratio, and to screen the temperature fluctuation mode components whose nonlinear error scores are greater than a preset value as the main temperature fluctuation mode components.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 5.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 5.

Images