A method for real-time compensation and self-calibration of dynamic temperature drift in eddy current sensors

Abstract

This disclosure provides a method for real-time compensation and self-calibration of dynamic temperature drift in eddy current sensors. By constructing an intelligent monitoring system integrating multi-point temperature sensors and employing a DBN-LSTM combined deep learning model to establish a complex nonlinear dynamic mapping relationship between sensor output and temperature and displacement, high-precision real-time compensation for dynamic temperature drift is achieved. Furthermore, a multi-objective optimization model is used to adaptively optimize the compensation results under operating conditions, balancing the contradictions between system accuracy, stability, and response speed. Simultaneously, an innovative online sequential learning mechanism and an intelligent self-calibration process based on uncertainty assessment are introduced, enabling the system to continuously optimize model parameters based on operating data and autonomously manage the calibration cycle. This effectively addresses the slow drift of sensor characteristics, significantly improves measurement accuracy and long-term reliability, greatly extends the calibration interval, and reduces maintenance requirements.

Description

Technical Field

[0001] This invention relates to the field of sensor signal processing technology, and in particular to a method for real-time compensation and self-calibration of dynamic temperature drift of an eddy current sensor. Background Technology

[0002] Eddy current sensors are widely used in the measurement of key parameters in high-end equipment such as aero-engines and steam turbines due to their non-contact nature, high reliability, and anti-contamination capabilities. However, the influence of temperature on eddy current sensors is a core technical challenge. Significant and highly nonlinear temperature drift under high-temperature environments severely restricts their measurement accuracy. Traditional hardware compensation circuits based on fixed parameters or piecewise linear lookup table software compensation methods are difficult to accurately cover the entire operating range, especially during rapid dynamic temperature changes, where the compensation effect is poor, and even compensation lag or overcompensation may occur. In addition, due to material aging, stress release, and long-term high-temperature exposure, the temperature drift characteristics of the sensor change slowly but irreversibly over time, causing the original compensation model to gradually become invalid. This necessitates periodic and cumbersome offline calibration, increasing maintenance costs and downtime. The industry urgently needs a new compensation technology that can adapt to complex dynamic temperature drift and has adaptive capabilities. Summary of the Invention

[0003] The first aspect of this disclosure provides a method for real-time compensation and self-calibration of dynamic temperature drift in an eddy current sensor, comprising the following steps:

[0004] S1: Construct a multi-temperature sensor array integrated with the eddy current sensor probe, wherein the temperature sensor array includes at least three temperature sensors arranged at the front sensing surface, the internal coil and the rear interface of the eddy current sensor probe.

[0005] S2: Collect the output voltage signal of the eddy current sensor, the signals of each temperature sensor, and the displacement reference signal of the target being measured. Perform preprocessing on the collected signals. The preprocessing includes filtering to remove noise, normalization to unify the dimensions, and time alignment to eliminate time deviation between channels.

[0006] S3: Extract input features from the preprocessed data, including temperature change rate, temperature gradient features, and historical sequence features;

[0007] S4: Input the input features into the trained deep learning model to calculate the displacement value after temperature drift compensation;

[0008] S5: Input the displacement value into the constructed multi-objective optimization function, and solve for the target compensation displacement value through adaptive weight adjustment based on working conditions. The multi-objective optimization function is as follows:

[0009] min{α·E+β·S+γ·R},

[0010] Where E is the compensation error, S is the system stability index, R is the response speed index, α, β, and γ are the adaptive weight coefficients for the working conditions, and α+β+γ=1. The constraints of the multi-objective optimization function include that the rate of change of the compensated displacement value does not exceed a preset threshold, and the compensation process is smooth without abrupt changes.

[0011] In conjunction with the first aspect, the deep learning model is a DBN-LSTM combined model, comprising:

[0012] The DBN feature extraction part consists of 3 stacked RBM layers with layer sizes of [128, 64, 32].

[0013] The LSTM sequence modeling part consists of two LSTM layers, each with 64 units.

[0014] The output layer has a structure of 64 input units, which are sequentially connected to a first hidden layer of 32 units, a second hidden layer of 16 units, and finally output to a single unit.

[0015] In conjunction with the first aspect, the training of the deep learning model includes offline training of the deep learning model using a calibration experimental dataset containing different combinations of temperature and displacement, in order to establish a nonlinear mapping relationship between sensor output voltage, temperature and actual displacement.

[0016] In conjunction with the first aspect, the method further includes acquiring true displacement data under working conditions that meet preset conditions, and using an online learning algorithm based on a sequential extreme learning machine combined with a forgetting factor mechanism to incrementally update the upper layer parameters of the LSTM layer and the fully connected parameters of the output layer in the deep learning model.

[0017] In conjunction with the first aspect, the operating conditions of the preset conditions include:

[0018] Stable idling condition: temperature change rate |dT / dt| < 1°C / min and duration > 10min;

[0019] Cold start / stop condition: The known displacement point when the equipment starts or stops.

[0020] In conjunction with the first aspect, the method also includes continuously evaluating the uncertainty of the deep learning model output, and triggering the system's self-calibration process when the uncertainty exceeds a preset threshold.

[0021] In conjunction with the first aspect, the uncertainty of the deep learning model output is continuously evaluated by using the Monte Carlo dropout method, maintaining Dropout activation during the model inference phase, performing multiple forward propagations, and statistically analyzing the variance of the output distribution as a measure of uncertainty.

[0022] In conjunction with the first aspect, the self-calibration process includes:

[0023] Based on the magnitude of the uncertainty and the duration of the anomaly, the system status is determined as observation status, calibration warning status, or forced calibration status.

[0024] When the system is determined to be in an observation state, the model monitoring frequency is increased and more calibration data points are acquired for online learning.

[0025] When a calibration warning state is detected, the measurement accuracy rating is lowered and an uncertainty indicator is added to the output results.

[0026] When the sensor is determined to be in a forced calibration state, it is marked as unreliable and a full-range calibration is performed.

[0027] A second aspect of this disclosure provides an electronic device, characterized in that it comprises:

[0028] One or more processors;

[0029] A storage unit is used to store one or more programs, which, when executed by one or more processors, enable the one or more processors to implement the method for real-time compensation and self-calibration of dynamic temperature drift of an eddy current sensor.

[0030] A third aspect of this disclosure provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, enables the implementation of the aforementioned method for real-time compensation and self-calibration of dynamic temperature drift in an eddy current sensor.

[0031] Beneficial Effects: This disclosure provides a real-time compensation and self-calibration method for dynamic temperature drift of eddy current sensors. By constructing an intelligent monitoring system integrating multi-point temperature sensors and using a DBN-LSTM combined deep learning model to establish a complex nonlinear dynamic mapping relationship between sensor output and temperature and displacement, high-precision real-time compensation for dynamic temperature drift is achieved. Furthermore, the compensation results are adaptively optimized under operating conditions through a multi-objective optimization model, balancing the contradiction between system accuracy, stability, and response speed. At the same time, an online sequential learning mechanism and an intelligent self-calibration process based on uncertainty assessment are innovatively introduced, enabling the system to continuously optimize model parameters and autonomously manage the calibration cycle based on operating data. This effectively addresses the slow drift of sensor characteristics, significantly improves measurement accuracy and long-term reliability, greatly extends the calibration interval, and reduces maintenance requirements. Attached Figure Description

[0032] Figure 1 This is a flowchart illustrating a method for real-time compensation and self-calibration of dynamic temperature drift of an eddy current sensor according to an embodiment of this disclosure.

[0033] Figure 2 An electronic device according to an embodiment of this disclosure. Detailed Implementation

[0034] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with those disclosed herein.

[0035] The terminology used in this disclosure is for the purpose of describing particular embodiments only and is not intended to be limiting of the present disclosure. The singular forms “a,” “the,” and “the” as used in this disclosure and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.

[0036] It should be understood that although the terms first, second, third, etc., may be used to describe various information in embodiments of this disclosure, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first information may also be referred to as second information without departing from the scope of embodiments of this disclosure, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."

[0037] like Figure 1 The diagram shown is a flowchart illustrating a method for real-time compensation and self-calibration of dynamic temperature drift in an eddy current sensor according to an embodiment of this disclosure, including:

[0038] S1: Construct a multi-temperature sensor array integrated with the eddy current sensor probe, wherein the temperature sensor array includes at least three temperature sensors arranged at the front sensing surface, the internal coil and the rear interface of the eddy current sensor probe.

[0039] For example, the temperature sensor is placed in key parts of the eddy current sensor, such as the front sensing surface, inside the coil, the rear interface, the mounting base, and the environmental reference point.

[0040] For high-temperature applications, miniature ceramic-encapsulated thermocouples are used and fixed in a designated position with high-temperature ceramic adhesive; for room-temperature applications, NTC or Pt100 thin-film elements can be used and fixed with thermally conductive epoxy resin.

[0041] For thermocouples, a dedicated temperature transmitter is used to convert the weak thermoelectric potential signal into a standard voltage signal; for resistive temperature sensors, a constant current source excitation and a four-wire connection method are used to eliminate the influence of lead resistance.

[0042] Each temperature sensor is independently calibrated to ensure an accuracy better than ±0.1°C over the operating temperature range (typically -40°C to +500°C).

[0043] S2: Collect the output voltage signal of the eddy current sensor, the signals of each temperature sensor, and the displacement reference signal of the target being measured. Perform preprocessing on the collected signals. The preprocessing includes filtering to remove noise, normalization to unify the dimensions, and time alignment to eliminate time deviation between channels.

[0044] Specifically, the pre-amplifier low-pass filter operates at a typical cutoff frequency of 1.2kHz (approximately 1.2 times the measurement bandwidth), effectively suppressing high-frequency noise. Subsequently, the adaptive median filter dynamically processes the signal, selectively filtering out occasional pulse interference. In addition, the system supports on-demand configuration of bandpass filters, facilitating the accurate extraction of target frequency components in specific application scenarios (such as vibration monitoring).

[0045] The first step applies a medium-range filter (window size 5) to effectively filter out outliers in the signal; the second step uses an exponential smoothing filter (smoothing coefficient α=0.1) to reduce the impact of random fluctuations. For scenarios with high dynamic temperature changes, the system can automatically switch to Kalman filtering mode, and its state transition matrix will be adaptively adjusted according to the historical temperature change rate to balance real-time performance and stability.

[0046] The voltage signal is processed using Min-Max normalization: V norm = (V - V min ) / (V max - V min Temperature signals are processed using Z-score normalization: T norm = (T - T mean ) / T std Normalized parameter (V) min V max T mean T std The model is determined based on calibration data, and a regular update mechanism is established to maintain its applicability.

[0047] S3: Extract input features from the preprocessed data, including temperature change rate, temperature gradient features, and historical sequence features;

[0048] For example, such as:

[0049] Instantaneous temperature value T at each temperature measuring pointi (i=1,2,...,n, where n is the number of temperature sensors);

[0050] Average temperature: T avg = (T 1 + T 2 + ... + T n ) / n ;

[0051] Temperature variance: T var = [(T 1 -T avg )² + (T 2 -T avg )² + ... + (T n -T avg )²] / n .

[0052] Rate of temperature change (first derivative): dT i / dt = (T i,t - T i,t -Δt) / Δt ;

[0053] Second derivative of temperature: d²T i / dt² = (dT i / dt,t - dT i / dt,t-Δt) / Δt ;

[0054] Cumulative temperature change: ∫|dT i / dt|dt (Accumulated within the sliding window W).

[0055] Temperature difference between points: ΔT ij = T i - T j (i,j = 1,2,...,n, i≠j) ;

[0056] Maximum temperature difference: ΔT max = max(ΔT ij ) ;

[0057] Normalized temperature gradient: T norm= ΔT ij / d ij (d) ij (The physical distance between the two temperature measurement points).

[0058] Voltage history sequence : V = [V t-k+1 , V t-k+2 , ..., V t ] (k is the sequence length, typically 10~30);

[0059] Temperature history sequence: T = [T t-k+1 , T t-k+2 , ..., T t ] ;

[0060] Exponentially weighted historical values: assign higher weight to recent values. V w = Σ(e (-λ(t-i)) ·V i ) λ is the attenuation factor. V w The feature values ​​are obtained by exponentially weighting historical voltage sequences, where more recent voltage values ​​are given higher weights, while the weights of more distant historical voltage values ​​decay exponentially.

[0061] The product of temperature and voltage T i ·V The product of the rate of temperature change and voltage. dT i / dt·V ; and interaction items between different temperature measurement points T i ·T j .

[0062] It should be noted that not all of the above features are used simultaneously; rather, feature selection is based on application requirements and model performance. To reduce computational complexity, principal component analysis (PCA) or autoencoders are used to reduce the dimensionality of high-dimensional features, retaining more than 90% of the information.

[0063] S4: Input the input features into the trained deep learning model to calculate the displacement value after temperature drift compensation;

[0064] The deep learning model is a DBN-LSTM combined model, including:

[0065] The DBN feature extraction part consists of 3 stacked RBM layers with layer sizes of [128, 64, 32].

[0066] The LSTM sequence modeling part consists of two LSTM layers, each with 64 units.

[0067] The output layer has a structure of 64 input units, which are sequentially connected to a first hidden layer of 32 units, a second hidden layer of 16 units, and finally output to a single unit.

[0068] The training of the deep learning model includes offline training of the deep learning model using a calibration experimental dataset containing different combinations of temperature and displacement, in order to establish a nonlinear mapping relationship between sensor output voltage, temperature and actual displacement.

[0069] Specifically, the DBN-LSTM combined model used in this disclosure has the following detailed structure:

[0070] Input layer: Receives feature vectors of dimension d, where d is determined by the specific implementation of feature engineering, with a typical value of 20~50.

[0071] The DBN feature extraction part consists of three stacked Restricted Boltzmann Machines (RBMs): the first layer is a Gaussian-Bernoulli RBM with 128 hidden units, specifically for handling continuous input data; the second layer uses a Bernoulli-Bernoulli RBM with 64 hidden units; and the third layer also uses a Bernoulli-Bernoulli RBM with 32 hidden units. All layers use the Sigmoid function as the activation function.

[0072] LSTM sequence modeling: The input sequence consists of 32-dimensional features extracted by a Deep Belief Network (DBN). The sequence length L can be flexibly defined according to application requirements, with a typical value of 20. In the network architecture, the first LSTM layer contains 64 units, processing the input and outputting the complete sequence; a Dropout layer is then added with a dropout rate of 0.3 to enhance the model's generalization ability; the second LSTM layer also has 64 units, but only returns the output of the last time step of the sequence. All LSTM units adopt a standard gating structure, including input gate, forget gate, output gate, and unit state.

[0073] Fully Connected Output Network: This neural network consists of three fully connected layers: the first layer reduces the input dimension from 64 to 32, followed by ReLU activation; the second layer further compresses the input to 16 dimensions and applies ReLU activation. Batch normalization is performed after the first two layers, and Dropout regularization with a ratio of 0.2 is applied. The final output layer maps the 16-dimensional features to a 1-dimensional result and uses a linear activation function to perform regression prediction.

[0074] Auxiliary output branch (for uncertainty estimation): A branch is taken from the output of the second LSTM layer. This branch is first processed through a fully connected layer to reduce the input dimension from 64 to 32 and apply the ReLU activation function. Then, it enters the output layer to further compress the dimension to 1 and applies the Softplus activation function to specifically predict the standard deviation.

[0075] The model training disclosed herein employs a two-stage strategy:

[0076] DBN Pre-training Phase: This model employs an unsupervised, layer-by-layer greedy pre-training method, with the objective function set to maximize the log-likelihood estimate of the data probability. The optimization algorithm uses contrastive divergence (CD-1), with an initial learning rate of 0.01 and a cosine annealing strategy applied. The batch size is fixed at 128, and each layer is trained for 100 epochs during the pre-training phase.

[0077] End-to-end joint optimization phase: This model uses mean squared error (MSE) as the main loss function and introduces negative log-likelihood (NLL) as an auxiliary uncertainty loss. The comprehensive loss function is defined as: L = MSE + α · NLL, where the weight coefficient α for the uncertainty loss is typically 0.2. The optimization process uses the Adam optimizer with an initial learning rate of 0.001, and a dynamic adjustment strategy: the learning rate is halved when the validation loss does not decrease for three consecutive rounds. The training batch size is fixed at 64, the maximum number of training rounds is set to 300, and an early stopping mechanism is implemented—training is terminated early if the validation loss does not improve for 10 consecutive rounds. To improve the model's generalization ability, L2 weight decay (coefficient 0.0005) and Dropout are used for regularization control during training, and data augmentation methods such as adding Gaussian noise, random scaling, and time shifting are employed.

[0078] The offline pre-training steps for the DBN-LSTM ensemble model include:

[0079] The raw data collected during the calibration experiment were aligned to microsecond-level time to ensure that the output voltage signal, multi-point temperature signal, and measured displacement signal of the eddy current sensor were strictly synchronized in time.

[0080] An adaptive bandpass filter (cutoff frequency 1.2kHz) is applied to the eddy current sensor signal to eliminate high-frequency noise while preserving the effective signal bandwidth;

[0081] Medium-range filtering (window size 5) and exponential smoothing filtering (smoothing coefficient α = 0.1) are applied to the temperature signal to eliminate measurement noise;

[0082] All signals are normalized, scaling the data to the [0,1] interval, as shown in the formula. .

[0083] Feature engineering includes:

[0084] Calculate the first-order rate of change (dT / dt) and the second-order rate of change (d²T / dt²) at each temperature point.

[0085] Calculate the spatial gradient characteristics between multiple temperature points, including radial and axial temperature gradients.

[0086] Historical sequence features were constructed, and temperature and voltage data from the past 128 time points were extracted using a sliding time window (window size N=128).

[0087] Through correlation analysis and principal component analysis, 16 features with the highest correlation to temperature drift were selected as model inputs.

[0088] The unsupervised pre-training of the DBN part includes:

[0089] The first RBM layer has a visible layer size of 16 (input feature dimension) and a hidden layer size of 128. It is trained using the contrastive divergence algorithm (CD-1) with a learning rate of 0.01, a batch size of 32, and 100 iterations.

[0090] The second RBM layer uses the output of the hidden layer of the first RBM layer as input. The hidden layer size is 64, and the training parameters are the same as the first layer.

[0091] The third RBM layer uses the output of the hidden layer of the second RBM layer as input, with a hidden layer size of 32 and training parameters the same as the first layer.

[0092] Through unsupervised pre-training, the DBN part learns deep feature representations of the data, effectively extracting the nonlinear features of temperature drift.

[0093] The supervised training of the LSTM component includes:

[0094] The feature representation (32-dimensional) obtained from DBN pre-training is used as the input to LSTM.

[0095] Configure a 2-layer LSTM network, with 64 units per layer, using the tanh activation function, and adding a dropout rate of 0.2 to prevent overfitting.

[0096] Add a fully connected layer (64→32→16→1) after the LSTM layer to output the compensated displacement value.

[0097] Using the negative log-likelihood loss function, the Adam optimizer, with an initial learning rate of 0.001, a batch size of 64, and 200 iterations.

[0098] To prevent overfitting, an early stopping method is used: training is stopped when the validation set loss no longer decreases for 10 consecutive rounds.

[0099] Model fusion and end-to-end fine-tuning include:

[0100] By connecting the DBN and LSTM components, a complete DBN-LSTM network is formed.

[0101] The entire network is fine-tuned end-to-end using a small learning rate (0.0001) for 50 iterations.

[0102] Save the optimal model parameters for subsequent real-time compensation.

[0103] S5: Input the displacement value into the constructed multi-objective optimization function, and solve for the target compensation displacement value through adaptive weight adjustment based on working conditions. The multi-objective optimization function is as follows:

[0104] min{α·E+β·S+γ·R},

[0105] Where E is the compensation error, S is the system stability index, R is the response speed index, α, β, and γ are the adaptive weight coefficients for the working conditions, and α+β+γ=1. The constraints of the multi-objective optimization function include that the rate of change of the compensated displacement value does not exceed a preset threshold, and the compensation process is smooth without abrupt changes.

[0106] First, the input to S5 is the "displacement value compensated for temperature drift" initially calculated by the deep learning model in step S4. Although the deep learning model has already performed compensation, its training objective is usually to minimize the overall error, which may not perfectly accommodate the different performance requirements of the system under all real-time operating conditions. For example, when the temperature changes drastically, the model output may exhibit high-frequency jitter in pursuit of accuracy, affecting stability; or in scenarios requiring fast response, the model output may appear sluggish.

[0107] Therefore, S5 constructs a constrained multi-objective optimization function.

[0108] Its objective function, min{α·E+β·S+γ·R}, aims to simultaneously minimize three key performance indicators: compensation error (E), system stability (S), and response speed (R). The key innovation here lies in the introduction of adaptive weighting coefficients α, β, and γ based on operating conditions.

[0109] The system no longer views the importance of accuracy, stability, and speed in a fixed way, but can dynamically adjust the weight of these three factors based on the real-time identified operating conditions. For example, under stable idling conditions, the system will assign a higher weight to accuracy (α); under transient conditions with rapid temperature changes, it will appropriately increase the weight of stability (β) and response speed (γ) to prevent output oscillation and lag. This dynamic trade-off mechanism ensures that the system can output the "optimal compensation displacement value" that best meets the current actual needs at any time.

[0110] Finally, S5 also sets clear constraints: "the rate of change of the compensated displacement value does not exceed a preset threshold, and the compensation process is smooth without abrupt changes." This is a crucial safety and reliability design. It mathematically guarantees that the optimized displacement output will not experience physically impossible large jumps or drastic fluctuations, ensuring the smoothness of the output curve and thus avoiding interference or even accidents caused by noise or abrupt changes introduced by the compensation algorithm itself to the downstream control system.

[0111] The method further includes acquiring true displacement data under working conditions that meet preset conditions, and using an online learning algorithm based on sequential extreme learning machine combined with a forgetting factor mechanism to incrementally update the upper layer parameters of the LSTM layer and the fully connected parameters of the output layer in the deep learning model.

[0112] The preset conditions include the following operating conditions:

[0113] Stable idling condition: temperature change rate |dT / dt| < 1°C / min and duration > 10min;

[0114] Cold start / stop condition: The known displacement point when the equipment starts or stops.

[0115] Specifically, stable idling condition identification involves identifying potential temperature drift calibration points under conditions where the temperature change rate is less than 1°C / min and lasts for more than 10 minutes. Simultaneously, it verifies that equipment operating parameters (such as speed and load) are within ±5% of their rated range. Based on this condition, it is assumed that the displacement of the measured object remains constant, and all output variations are attributed to the temperature drift effect. Calibration sample points are collected by evaluating data from the most recent 10 minutes every 30 minutes; samples meeting the criteria are recorded as valid samples.

[0116] Cold start / shutdown operation: During the initial stage of equipment starting from a cold state or the stable stage after a complete shutdown, the location of critical components is usually at the known initial design location. Sampling should be performed within the first 5 minutes after equipment startup or during the shutdown stabilization period, where the stabilization period must meet the condition that the temperature change rate drops below 0.5°C / min.

[0117] The method also includes continuously evaluating the uncertainty of the deep learning model output, and triggering the system's self-calibration process when the uncertainty exceeds a preset threshold.

[0118] Continuous evaluation of the uncertainty of the deep learning model output includes using the Monte Carlo dropout method, maintaining Dropout activation during the model inference phase, performing multiple forward propagations, and statistically analyzing the variance of the output distribution as a measure of uncertainty.

[0119] The self-calibration process includes:

[0120] Based on the magnitude of the uncertainty and the duration of the anomaly, the system status is determined as observation status, calibration warning status, or forced calibration status.

[0121] When the system is determined to be in an observation state, the model monitoring frequency is increased and more calibration data points are acquired for online learning.

[0122] When a calibration warning state is detected, the measurement accuracy rating is lowered and an uncertainty indicator is added to the output results.

[0123] When the sensor is determined to be in a forced calibration state, it is marked as unreliable and a full-range calibration is performed.

[0124] Specifically, this method introduces a continuous model health monitoring mechanism, namely, "continuously assessing the uncertainty of the deep learning model output".

[0125] This mechanism employs Monte Carlo dropout as a measure of uncertainty. During model inference (prediction), instead of shutting down the Dropout layer as usual, it remains active and performs multiple forward propagations on the same set of input data. Due to the randomness of Dropout, each forward propagation yields a slightly different output. The system then calculates the variance of these multiple output values. This variance is the quantitative indicator of uncertainty: a larger variance indicates a less confident model regarding the current input data, leading to less reliable output; a smaller variance indicates a more resolute model judgment and higher reliability.

[0126] Next, based on the uncertainty assessment results, the system triggers a hierarchical, progressive intelligent self-calibration process. This is not a binary "yes / no" decision, but a refined state machine that adopts different response strategies according to the severity of the problem:

[0127] When uncertainty begins to rise but does not reach a dangerous level, the system enters a state of alert. The response is non-intrusive: increasing monitoring frequency and more proactively seeking opportunities (such as under stable operating conditions) to acquire calibration data points for online model fine-tuning. The goal is to eliminate potential problems in their infancy through proactive learning.

[0128] When uncertainty increases further, it indicates a noticeable degradation in model performance. At this point, the system takes two measures: first, it lowers the measurement accuracy rating externally, signaling to users or higher-level systems that the data's reliability has decreased; second, it adds uncertainty indicators to the output, providing decision-making support for data users. Simultaneously, the system triggers a simplified calibration process to attempt repair.

[0129] When uncertainty is extremely high or persistently abnormal, it indicates that the model may have seriously failed or the sensor may be malfunctioning. This is the highest level of alert. The system will decisively mark the sensor as "unreliable" and may switch to backup mode, while simultaneously notifying that a full-range calibration is required. This ensures that the system does not continue to operate while providing erroneous data, avoiding serious consequences that could result from sensor failure.

[0130] Ultimately, this achieved a shift from "periodic calibration" to "on-demand calibration," greatly enhancing the system's autonomy and reliability, and significantly reducing maintenance costs throughout its entire lifecycle.

[0131] Electronic device 200 can be a desktop computer, laptop, handheld computer, cloud server, or other electronic device. Electronic device 200 may include, but is not limited to, processor 201 and memory 202. Those skilled in the art will understand that... Figure 2 This is merely an example of electronic device 200 and does not constitute a limitation on electronic device 200. It may include more or fewer components than shown, or combine certain components, or different components. For example, electronic device may also include input / output devices, network access devices, buses, etc.

[0132] The processor 201 can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0133] The memory 202 can be an internal storage unit of the electronic device 200, such as a hard disk or RAM of the electronic device 200. The memory 202 can also be an external storage device of the electronic device 200, such as a plug-in hard disk, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card equipped on the electronic device 200. Furthermore, the memory 202 can include both internal and external storage units of the electronic device 200. The memory 202 is used to store the computer program 203 and other programs and data required by the electronic device. The memory 202 can also be used to temporarily store data that has been output or will be output.

[0134] The above embodiments are only used to illustrate the technical solutions of this disclosure, and are not intended to limit it. Although this disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this disclosure, and should all be included within the protection scope of this disclosure.

Claims

1. A method for real-time compensation and self-calibration of dynamic temperature drift of an eddy current sensor, characterized in that, The method comprises the following steps: S1: constructing a plurality of temperature sensor arrays integrated with an eddy current sensor probe, the temperature sensor arrays comprising at least three temperature sensors arranged at a front end sensing surface, an internal coil and a rear end interface part of the eddy current sensor probe; S2: collecting output voltage signals of the eddy current sensor, signals of each temperature sensor and displacement reference signals of a measured target, and pre-processing the collected signals, the pre-processing comprising filtering to remove noise, normalization to unify dimensions and time alignment to eliminate time deviation between channels; S3: extracting input features from the pre-processed data, the input features comprising a temperature change rate, a temperature gradient feature and a historical sequence feature; S4: inputting the input features into a trained deep learning model to calculate a temperature drift compensated displacement value, wherein displacement true value data is obtained under a preset condition of working condition, an online learning algorithm based on a sequential extreme learning machine and combined with a forgetting factor mechanism is used to incrementally update upper layer parameters of an LSTM layer and full connection parameters of an output layer in the deep learning model, and the preset condition of working condition comprises: a stable idle speed working condition: a temperature change rate |dT / dt|<1°C / min and a duration >10 min, a cold start / shut down working condition: within the first 5 minutes after starting the equipment or a stable shutdown period, wherein the stable shutdown period needs to meet a temperature change rate of 0.5°C / min or less, S5: inputting the displacement value into a constructed multi-objective optimization function, solving a target compensation displacement value through working condition adaptive weight adjustment, and determining a system state as an observation state, a calibration warning state or a forced calibration state according to a numerical size of uncertainty and an abnormal duration, when the system state is determined as the observation state, increasing a model monitoring frequency and obtaining more calibration data points for online learning, when the system state is determined as the calibration warning state, reducing a measurement accuracy rating and increasing an uncertainty prompt in an output result, and when the system state is determined as the forced calibration state, marking the sensor as an unreliable state and performing full range calibration, the multi-objective optimization function being: min{α·E+β·S+γ·R}, wherein E is a compensation error, S is a system stability index, R is a response speed index, α, β and γ are working condition adaptive weight coefficients, and α+β+γ=1, and constraint conditions of the multi-objective optimization function include that a change rate of the compensated displacement value does not exceed a preset threshold, and a compensation process is smooth without mutation.

2. The method of claim 1, wherein, The deep learning model is a DBN-LSTM combined model, comprising: a DBN feature extraction part, which is composed of 3 layers of RBM stacked, with layer sizes of [128, 64, 32]; an LSTM sequence modeling part, which comprises 2 layers of LSTM, each layer having 64 units; an output layer, the structure of the output layer being 64 input units, sequentially connected to a first hidden layer of 32 units, a second hidden layer of 16 units, and finally output to 1 unit.

3. The method of claim 1, wherein, The training of the deep learning model comprises offline training of the deep learning model using a calibration dataset comprising different temperature and displacement combinations to establish a nonlinear mapping relationship between sensor output voltage, temperature and true displacement.

4. The method of claim 1, wherein, The method further comprises continuously evaluating the uncertainty of the deep learning model output, and triggering a self-calibration procedure of the system when the uncertainty exceeds a preset threshold.

5. The method of claim 4, wherein, The continuous evaluation of the uncertainty of the deep learning model output comprises maintaining Dropout activation during model inference stage by Monte Carlo dropout method, performing multiple forward propagations, and taking the variance of the output distribution as a measure of uncertainty.

6. An electronic device, comprising: The method comprises: one or more processors; a storage unit for storing one or more programs, which when executed by the one or more processors, enable the one or more processors to implement the method of claim 1.

7. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program, when executed by a processor, enables the implementation of the method of claim 1.

Images