Pressure gauge calibration process management method and system

Abstract

The invention relates to a pressure gauge verification process management method and system in the technical field of data processing and artificial intelligence. The method comprises the following steps: firstly, acquiring an original pressure measurement signal, a real pressure value and a real space gradient value; thirdly, performing noise separation on the original pressure measurement signal, and extracting time-frequency domain mixed features based on a noise separation result; then, a pressure measurement signal verification and correction model is constructed, and the model comprises a double-flow feature extraction network, a feature interaction and fusion module, a physical constraint output module and a multi-scale residual error correction module; training a pressure measurement signal verification correction model based on the time-frequency domain mixed features, the real pressure value and the real space gradient value; and finally, carrying out noise separation and feature extraction on a to-be-corrected pressure measurement signal, and inputting the extracted features into the trained pressure measurement signal verification correction model to obtain a corrected pressure predicted value. And the accuracy and effectiveness of pressure gauge calibration are improved.

Description

Technical Field

[0001] This application relates to the fields of data processing and artificial intelligence technology, and in particular to methods and systems for managing pressure gauge calibration processes. Background Technology

[0002] In fields such as industrial automation, energy transmission and distribution, water conservancy systems, and aerospace, pressure gauges are critical measuring instruments whose accuracy directly affects system safety and operational efficiency. Therefore, they need to be calibrated regularly according to national standards to ensure equipment reliability and compliance.

[0003] There are two main types of traditional calibration methods: the manual comparison method requires connecting the gauge to be calibrated and the standard gauge to the same pressure source, manually adjusting the pressure, and reading the deviation at multiple calibration points; the simple linear correction method is based on constructing a linear model based on the deviation data to correct the measurement results.

[0004] However, traditional methods have obvious limitations: First, they cannot separate systematic errors from environmental noise in the measurement signal. Systematic errors originate from defects in the instrument itself, while environmental noise comes from electromagnetic interference, mechanical vibration, etc., resulting in insufficient calibration accuracy. Second, pressure signals in industrial environments exhibit nonlinear characteristics and abrupt behavior. Traditional feature extraction methods cannot reflect dynamic trends, and conventional algorithms are difficult to fit nonlinear and abrupt signals, failing to capture multi-scale dynamic changes and affecting the effectiveness of calibration.

[0005] Therefore, there is an urgent need in related technologies for a way to improve the accuracy and effectiveness of pressure gauge calibration. Summary of the Invention

[0006] Therefore, it is necessary to provide a pressure gauge calibration process management method and system that can improve the accuracy and effectiveness of pressure gauge calibration, addressing the aforementioned technical problems.

[0007] Firstly, this application provides a method for managing the pressure gauge calibration process. The method includes:

[0008] Acquire the raw pressure measurement signal, the actual pressure value, and the actual spatial gradient value;

[0009] The original pressure measurement signal is subjected to noise separation, and time-frequency domain hybrid features are extracted based on the noise separation results;

[0010] A pressure measurement signal verification and correction model is constructed, which includes a dual-stream feature extraction network, a feature interaction fusion module, a physical constraint output module, and a multi-scale residual correction module.

[0011] The pressure measurement signal verification and correction model is trained based on the time-frequency domain hybrid features, the true pressure value, and the true spatial gradient value.

[0012] Noise separation and feature extraction are performed on the pressure measurement signal to be calibrated. The extracted features are then input into the trained pressure measurement signal calibration model to obtain the calibrated pressure prediction value.

[0013] Optionally, in one embodiment of this application, the noise separation of the original pressure measurement signal includes:

[0014] The original pressure measurement signal is processed by median filtering to obtain the median filtered baseline signal;

[0015] Based on the absolute difference between the original pressure measurement signal and the median filtered baseline signal, the system error scalar value is adaptively calculated using the Sigmoid function;

[0016] The environmental noise components are determined based on the original pressure measurement signal, the median filtered baseline signal, and the system error scalar value.

[0017] Optionally, in one embodiment of this application, the extraction of time-frequency domain hybrid features based on the noise separation result includes:

[0018] The second derivative of the median filtered baseline signal is calculated, and convolution smoothing is performed using a Hanning window to obtain dynamic differential spectrum eigenvalues.

[0019] The number of frequency bands for wavelet packet decomposition is determined based on the maximum, minimum, and standard deviation of the median filtered baseline signal.

[0020] The environmental noise components are decomposed using wavelet packets, and the energy of each frequency band is extracted as a time-frequency feature.

[0021] Optionally, in one embodiment of this application, the dual-stream feature extraction network is a parallel dual-branch network, including a global trend branch and a local sensitive branch. The global trend branch is composed of multiple layers of dilated causal convolution stacked together, with the dilation rate changing cyclically according to the number of layers to exponentially expand the receptive field, and the output of each layer is subjected to layer normalization. The local sensitive branch adopts a dynamic gated convolution structure, generates a gate signal through the Sigmoid function, and then multiplies it element-wise with the convolution output processed by the hyperbolic tangent activation function.

[0022] Optionally, in one embodiment of this application, the feature interaction fusion module includes:

[0023] A query vector is generated by linear transformation based on the final output features of the global trend branch, and a key vector and a value vector are generated by linear transformation based on the final output features of the local sensitive branch.

[0024] The similarity between the query vector and the transpose of the key vector is calculated, and after being adjusted by a scaling factor, a lower triangular mask matrix is ​​superimposed. The attention weight matrix is ​​then obtained by normalizing the matrix using the Softmax function.

[0025] The value vector is weighted and summed based on the attention weight matrix to obtain local features modulated by global information. These features are then residually connected to the final output features of the original global trend branch. Finally, attention fusion features are output through layer normalization.

[0026] Optionally, in one embodiment of this application, the physical constraint output module includes:

[0027] Based on the attention fusion features, an initial pressure prediction sequence is obtained by linear transformation through a fully connected layer.

[0028] Optionally, in one embodiment of this application, the multi-scale residual correction module includes:

[0029] Based on the attention fusion features, causal convolution is performed, and the shallow features output by the local sensitive branch are upsampled. The two are then added together to generate a multi-scale residual signal.

[0030] The multi-scale residual signal is transformed using a linear projection weight matrix and added to the initial pressure prediction sequence to obtain the corrected pressure prediction value.

[0031] Optionally, in one embodiment of this application, the total loss function for model training includes a main loss function and a physical consistency loss function, wherein the main loss function is determined based on the dynamic importance weight coefficient and the sampling point prediction error, the dynamic importance weight coefficient is determined based on the absolute value of the second derivative of the true pressure value, and the physical consistency loss function is determined based on the temporal gradient and spatial gradient values ​​of the initial pressure prediction sequence.

[0032] Secondly, this application also provides a pressure gauge calibration process management system. The system includes:

[0033] The measurement data acquisition module is used to acquire the raw pressure measurement signal, the actual pressure value, and the actual spatial gradient value;

[0034] A pressure measurement signal noise separation and time-frequency domain hybrid feature construction module is used to perform noise separation on the original pressure measurement signal and extract time-frequency domain hybrid features based on the noise separation result.

[0035] The pressure measurement signal verification and correction model construction module is used to construct a pressure measurement signal verification and correction model. The model includes a dual-stream feature extraction network, a feature interaction fusion module, a physical constraint output module, and a multi-scale residual correction module.

[0036] The pressure measurement signal verification and correction model training module is used to train the pressure measurement signal verification and correction model based on the time-frequency domain hybrid features, the real pressure value, and the real spatial gradient value.

[0037] The pressure measurement signal verification and correction module is used to perform noise separation and feature extraction on the pressure measurement signal to be corrected. The extracted features are then input into the trained pressure measurement signal verification and correction model to obtain the corrected pressure prediction value.

[0038] The aforementioned pressure gauge calibration process management method and system first acquires the original pressure measurement signal, the actual pressure value, and the actual spatial gradient value. Then, noise separation is performed on the original pressure measurement signal, and time-frequency domain hybrid features are extracted based on the noise separation results. Next, a pressure measurement signal calibration and correction model is constructed, including a dual-stream feature extraction network, a feature interaction fusion module, a physical constraint output module, and a multi-scale residual correction module. Then, the pressure measurement signal calibration and correction model is trained based on the time-frequency domain hybrid features, the actual pressure value, and the actual spatial gradient value. Finally, noise separation and feature extraction are performed on the pressure measurement signal to be calibrated, and the extracted features are input into the trained pressure measurement signal calibration and correction model to obtain the calibrated pressure prediction value. In other words, median filtering combined with Sigmoid adaptive weights separates system errors from environmental noise, avoiding the damage to abrupt signals caused by conventional filtering and significantly improving signal fidelity. Dynamic differential spectrum is used to highlight the rate of change at abrupt points, and adaptive wavelet packets are divided into frequency bands according to signal complexity, achieving complementary fusion of time-domain trends and frequency-domain perturbations, thus improving feature representation capabilities. Two branches, dilated causal convolution and dynamic gated convolution, are designed in parallel to extract global trend and local abrupt features, respectively. Fine fusion is achieved through a cross-domain attention mechanism, significantly enhancing the model's dynamic response capability. Simplified momentum conservation constraints are adopted in the output layer, and the loss in abrupt regions is dynamically emphasized through a second-order derivative weighting mechanism, balancing physical rationality and prediction accuracy, and improving generalization ability and engineering usability. Attached Figure Description

[0039] Figure 1 This is a flowchart illustrating a pressure gauge calibration process management method in one embodiment;

[0040] Figure 2 This is a schematic diagram comparing the ability of different methods to handle system errors and environmental noise in one embodiment;

[0041] Figure 3 This is a schematic diagram illustrating the noise separation effect in one embodiment;

[0042] Figure 4 This is a schematic diagram comparing the prediction errors of different feature extraction methods in one embodiment;

[0043] Figure 5This is a schematic diagram illustrating the impact of residual correction on high-frequency detail recovery in one embodiment;

[0044] Figure 6 This is a schematic diagram comparing the pressure prediction accuracy in one embodiment;

[0045] Figure 7 This is a structural block diagram of a pressure gauge calibration process management system in one embodiment;

[0046] Figure 8 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0048] In one embodiment, such as Figure 1 As shown, a method for managing the calibration process of a pressure gauge is provided, including the following steps:

[0049] S101: Acquire the raw pressure measurement signal, the actual pressure value, and the actual spatial gradient value.

[0050] In this embodiment, firstly, the raw pressure measurement signal is acquired from the pressure sensor to be calibrated. Specifically, this is achieved by installing the pressure sensor on the fluid system being measured and connecting it to a data acquisition device. The acquisition device continuously samples at a fixed sampling rate to ensure the complete dynamic characteristics of pressure changes are captured. The sampling rate can be selected from several times per second to hundreds of times per second, such as 10 times per second. The acquired raw signal contains 50 consecutive sampling points, forming the raw pressure measurement signal X. raw Each sampling point corresponds to a pressure reading.

[0051] Similarly, a high-precision calibration pressure sensor is used to acquire the actual pressure value. This calibration pressure sensor offers higher accuracy and stability and serves as the benchmark for pressure measurement. During calibration, the calibration sensor and the pressure gauge to be calibrated are installed at the same pressure point to ensure they are in the same pressure environment, thus guaranteeing consistency of measurement conditions. The acquired actual pressure values ​​are recorded at the same sampling rate, forming a time series synchronized with the original pressure measurement signal. The actual pressure value at the t-th sampling point is defined as... As a label for supervised learning.

[0052] Similarly, multiple high-precision spatial reference sensors are deployed to measure the pressure spatial gradient. These sensors are installed at different spatial locations in the fluid system under test, such as upstream and downstream of a pipe at known intervals. At each sampling moment, all sensors synchronously acquire pressure readings, calculate the pressure difference between adjacent sensors, and divide it by the precise distance between the two points to obtain the spatial gradient value at that moment. The spatial gradient of the true pressure value at the t-th sampling point is defined as... It reflects the rate of change of pressure in the spatial direction.

[0053] S102: Perform noise separation on the original pressure measurement signal, and extract time-frequency domain hybrid features based on the noise separation results.

[0054] In this embodiment, the original pressure gauge reading includes systematic errors and environmental noise. Systematic errors manifest as inherent nonlinear deviations of the pressure gauge, while environmental noise represents transient disturbances. Conventional wavelet denoising methods struggle to effectively distinguish between the two because they overlap in both frequency and amplitude. Single filtering methods are prone to loss of useful information or residual noise, leading to decreased accuracy in subsequent feature extraction and modeling. Furthermore, pressure fluctuation signals simultaneously contain temporal trends and frequency characteristics. Conventional Fast Fourier Transform methods lose temporal correlation information, making it difficult to effectively capture transient features and dynamic changes, thus affecting the accuracy of subsequent classification models. Therefore, a dual-threshold filtering mechanism separates systematic errors and environmental noise, including extracting the baseline signal, adaptively estimating systematic errors, and restoring environmental noise components. Simultaneously, by fusing dynamic differential spectrum and adaptive wavelet packet analysis, complementary features in the time and frequency domains are extracted, achieving the construction of hybrid time-frequency features.

[0055] Specifically, in one embodiment of this application, the noise separation of the original pressure measurement signal includes:

[0056] S201: Perform median filtering on the original pressure measurement signal to obtain the median filtered baseline signal.

[0057] S202: Based on the absolute difference between the original pressure measurement signal and the median filtered baseline signal, the system error scalar value is adaptively calculated using the Sigmoid function.

[0058] S203: Determine the environmental noise components based on the original pressure measurement signal, the median filtered baseline signal, and the system error scalar value.

[0059] In one embodiment of this application, firstly, median filtering is performed on the original pressure measurement signal to eliminate pulse-type environmental noise, resulting in a smooth median filtered baseline signal used to approximate the true pressure trend. The pulse noise is filtered out, as shown below:

[0060] Xbase =MedFilt(X raw (k=5)

[0061] In the formula, X base The base signal is represented by the median filter, used to approximate the true pressure trend and filter out impulse noise; MedFilt(·) represents the median filter function, used to suppress impulse noise and has good edge-preserving characteristics; X raw This represents the original pressure measurement signal, with a dimension of 50, meaning it contains 50 sampling points; k represents the window size for the median filter, in MedFilt(X raw In the term (k=5), we take k=5, which means using a 5-point sliding window for filtering, achieving a balance between noise reduction and maintaining signal response speed.

[0062] Subsequently, based on the absolute difference between the original pressure measurement value and the corresponding median filtered baseline value, the system error scalar value is adaptively calculated using the Sigmoid function, thus reflecting the inherent nonlinear deviation of the pressure gauge, expressed as:

[0063]

[0064] In the formula, ∈ sys t This represents the scalar value of the systematic error at the t-th sampling point, reflecting the inherent, non-random, nonlinear deviation of the pressure gauge; This represents the original pressure measurement value at the t-th sampling point; t is the time index, representing the sequence number of the sampling point, t = 1, 2, ..., 50; γ represents the median filtered baseline value at the t-th sampling point; γ represents the steepness control parameter of the Sigmoid function, used to adjust the sensitivity of the weighting function to the deviation of the baseline signal from the mean. Here, the empirical value is taken as γ = 0.5. exp(·) represents the arithmetic mean of all measurements of the original pressure measurement signal, used as a reference benchmark for evaluating the signal level; exp(·) represents the natural exponential function.

[0065] It should be noted that the estimated system error scalar value ∈ sys The calculation uses the Sigmoid function, with the Sigmoid weighting term used to adaptively assign weights, giving greater weight to baseline signal values ​​that are far from the mean, thus emphasizing systematic errors at significant deviations.

[0066] It should also be noted that this invention transforms the Sigmoid function from its conventional nonlinear activation purpose into a weighter, enabling system error estimation to not only depend on instantaneous deviations but also incorporate the global context, avoiding the shortcomings of fixed threshold methods. The Sigmoid function can dynamically adjust the weights according to the degree of deviation between the baseline signal and the global average. When the baseline signal is far from the average, the Sigmoid function output is close to 1, assigning greater weights and emphasizing system errors at significant deviations. When the baseline signal is close to the average, the weights decrease, suppressing error estimation in stable regions. Therefore, system error estimation is more accurate at abrupt changes in pressure, which are often significant areas of nonlinear deviation. Conventional methods tend to underestimate at abrupt changes or overestimate at stable regions, leading to incomplete noise separation. Sigmoid weighting allows the estimated value to adaptively focus on key regions.

[0067] Finally, element-wise subtraction is used to subtract the median-filtered baseline signal and the system error vector from the original pressure measurement signal to obtain the environmental noise component, which reflects instantaneous, random external disturbances, and is expressed as:

[0068] δ env =X raw -X base -∈ sys

[0069] In the formula, δ env Represents environmental noise components, with a dimension of 50, reflecting instantaneous, random external disturbances; ∈ sys This is the system error vector, with a dimension of 50, the same as the original pressure measurement signal X. raw The same applies, where the t-th element is the estimated system error scalar value of the t-th sampling point ∈ sys t .

[0070] It should be noted that the median filtered baseline signal X base Median filtering removes impulse noise, approximating the true pressure trend, but residual systematic errors remain. The systematic error vector ∈ sys The inherent bias of the pressure gauge is adaptively quantified, and the systematic error is subtracted as a vector instead of relying solely on baseline filtering. Impulse noise and systematic error are processed sequentially through a dual threshold mechanism to achieve a more refined separation.

[0071] In one embodiment, the ability of different methods to separate systematic errors and environmental noise is compared, and the results are as follows: Figure 2As shown in the figure, systematic error is the inherent nonlinear deviation of the pressure gauge, and environmental noise is an external random disturbance. Effective separation can improve calibration accuracy. This experiment uses three indicators to evaluate: residual rate of systematic error, residual rate of environmental noise, and retention rate of useful signal. The lower the residual rate, the better the separation effect; the higher the retention rate, the better the signal integrity. The horizontal axis in the figure represents different processing methods, including conventional methods and the method of this invention; the vertical axis represents the ratio, which is a dimensionless value ranging from 0 to 1. The lower the ratio, the better for the residual rate; the higher the ratio, the better for the retention rate. As can be seen from the figure, conventional methods have high residual rates of both systematic error and environmental noise, indicating that they fail to effectively separate errors and noise. At the same time, the retention rate of useful signal is relatively low, indicating that some useful information is lost during processing. The wavelet denoising combined with linear correction method performs poorly in terms of residual systematic error, the median filtering combined with polynomial fitting has high residual environmental noise, and the moving average combined with neural network is insufficient in terms of retaining useful signal. In comparison, the method of this invention has the lowest residual rate of system error and residual rate of environmental noise, while the highest residual rate of useful signal, proving the effectiveness of the dual threshold filtering mechanism and feature extraction. It can finely separate errors and noise, maximize the retention of real pressure signals, and improve the accuracy of pressure gauge calibration.

[0072] In one embodiment, noise separation effect analysis is performed to verify the effectiveness of the method of the present invention in noise separation and the effectiveness of the dual threshold filtering mechanism. The results are as follows: Figure 3 As shown in the figure, there are two subplots: the upper subplot displays the original measurement signal, baseline signal, and true pressure (in kPa); the lower subplot displays the separated systematic error and environmental noise (also in kPa). In the upper subplot, the baseline signal, extracted through median filtering, roughly follows the trend of the true pressure, but residual error still exists. In the lower subplot, the systematic error shows a slowly changing trend, reflecting the inherent nonlinear deviation of the pressure gauge; the environmental noise exhibits random fluctuations. The method of this invention successfully separates these two noise components, providing clean feature input for subsequent correction, demonstrating the advantages of the dual-threshold filtering mechanism. By adaptively estimating the systematic error and restoring the environmental noise, more refined noise separation is achieved, thereby improving the accuracy of signal processing.

[0073] Specifically, in one embodiment of this application, the extraction of time-frequency domain hybrid features based on the noise separation result includes:

[0074] S301: Calculate the second derivative of the median filtered baseline signal and perform convolution smoothing using a Hanning window to obtain dynamic differential spectrum eigenvalues.

[0075] S302: Determine the number of frequency bands for wavelet packet decomposition based on the maximum value, minimum value, and standard deviation of the median filtered baseline signal.

[0076] S303: Perform wavelet packet decomposition on the environmental noise components and extract the energy of each frequency band as time-frequency features.

[0077] In one embodiment of this application, firstly, the second derivative of the median-filtered baseline signal is calculated to enhance abrupt change characteristics, and then convolutional smoothing is performed using a Hanning window to suppress noise interference, resulting in time-domain enhanced dynamic differential spectrum eigenvalues, expressed as:

[0078]

[0079] In the formula, The dynamic differential spectrum eigenvalue at time t is used to highlight the rate of change of the signal at that time. Represents the median filtered baseline signal X base The second derivative with respect to time is discretized using the central difference method to characterize the acceleration or curvature features of the signal at that moment, and is sensitive to abrupt changes; ReLU(·) represents the activation function of the linear rectifier unit, which is used here to filter out the negative part of the second derivative and retain only the positive acceleration or abrupt change information; This represents the discrete convolution operation; Hann(L=10) represents the Hanning window function with a length of L=10, which is used to smooth and filter the differentiated signal and suppress high-frequency fluctuations that may be amplified by noise; L is the length of the Hanning window, in the number of sampling points, and the preferred value is L=10.

[0080] It should be noted that rising events in pressure signals usually have more important characteristics. The term uses the second derivative form, which is sensitive to the acceleration of the signal, highlights areas of drastic change, such as spikes or steep slopes, amplifies high-frequency noise, and combines the ReLU function to filter out negative values, retaining only positive changes, such as pressure rise edges, to avoid negative interference.

[0081] Then, based on the maximum, minimum, and standard deviation of the median-filtered baseline signal, the number of frequency bands to be used for wavelet packet decomposition is adaptively calculated to adapt the decomposition granularity to the signal complexity, as expressed in:

[0082]

[0083] In the formula, K represents the number of frequency bands to be used in wavelet packet decomposition, aiming to adapt the decomposition granularity to the signal complexity; max(X base ) represents the median filtered baseline signal X base The maximum value among all eigenvalues; min(X) base ) represents the median filtered baseline signal X base The minimum value among all eigenvalues; std(X) base) represents the median filtered baseline signal X base The standard deviation of all eigenvalues ​​measures the degree of dispersion of the signal; This represents the floor function, ensuring that K is an integer.

[0084] It should be noted that 0.5·std(X) base In the ) term, 0.5 is an empirical scaling factor used to adjust the sensitivity of the number of bandwidths to the ratio of the signal's dynamic range to its dispersion.

[0085] It should also be noted that, The ratio of terms reflects the degree of drastic change in the signal. When the ratio is large, the signal is complex and requires more bandwidth. When the ratio is small, the signal is stable and requires less bandwidth. This makes the number of bandwidths match the signal characteristics, providing high-resolution analysis on complex signals and saving computation on simple signals, thus improving the effectiveness of time-frequency features.

[0086] Finally, wavelet packet decomposition is performed on the environmental noise components to extract the energy of each frequency band as a time-frequency feature, characterizing the energy distribution of environmental noise in different frequency ranges, as shown below:

[0087]

[0088] F wp =[E1,E2,…,E k ,…,E K ]

[0089] In the formula, E k E represents the energy of the k-th wavelet packet decomposition band, E1 represents the energy of the 1st wavelet packet decomposition band, E2 represents the energy of the 2nd wavelet packet decomposition band, and E... K WPT represents the energy of the frequency band decomposed from the Kth wavelet packet. k (δ env ,t) represents the environmental noise component δ env The coefficients of the k-th frequency band at time t after wavelet packet decomposition; k is the frequency band index, k = 1, 2, ..., K, representing the k-th frequency band of the wavelet packet decomposition; F wp Let K represent the time-frequency feature vector of the wavelet packet, with dimension K.

[0090] It should be noted that wavelet packet decomposition uses basis functions that match the signal characteristics, such as dB4 wavelets. The number of decomposition layers is determined by the adaptively determined number of frequency bands K, thereby achieving a refined characterization of the frequency domain characteristics of environmental noise.

[0091] It should also be noted that by performing wavelet packet analysis on environmental noise rather than the original signal, the interference of systematic errors and trend terms on frequency domain features can be avoided, and the frequency domain patterns of random disturbances can be extracted more purely.

[0092] In one embodiment, a performance comparison analysis of different feature extraction methods is performed to evaluate the superiority of the time-frequency domain hybrid feature (dynamic differential spectrum + adaptive wavelet packet) of the present invention compared with single-domain features, and its effect on improving prediction accuracy. The results are as follows: Figure 4 As shown in the figure. The horizontal axis of the bar chart represents the four feature extraction methods, and the vertical axis represents the root mean square error of prediction (unit: kPa). From left to right: time-domain features only, frequency-domain features only, conventional time-frequency features, and the hybrid features of this invention. The time-domain features only (blue bars) have the highest error because they ignore frequency-domain perturbation patterns; the frequency-domain features only (orange bars) have the second highest error because they lose information about abrupt changes in the time domain; the conventional time-frequency features (green bars), although combining both, are not adaptable to complex signal segments due to fixed frequency band division. The method of this invention (red bars) has the lowest error, indicating that the dynamic differential spectrum can enhance the curvature characteristics of the pressure rise / fall edges, and the adaptive wavelet packet can dynamically divide the frequency band according to the signal complexity, reflecting the synergistic effect of time-domain acceleration and frequency band energy, and proving the holographic representation capability of the hybrid features for pressure fluctuation patterns.

[0093] S103: Construct a pressure measurement signal verification and correction model, which includes a dual-stream feature extraction network, a feature interaction fusion module, a physical constraint output module, and a multi-scale residual correction module.

[0094] In this embodiment, after preprocessing the data, a pressure measurement signal verification and correction model for pressure gauge calibration is constructed. The model includes a two-stream feature extraction network, a feature interaction fusion module, a physical constraint output module, and a multi-scale residual correction module. Conventional convolutional neural networks struggle to simultaneously capture long-range periodic dependencies and local abrupt changes when processing pressure time-series data. While single-structure networks can expand their receptive field to capture global trends, they lose sensitivity to instantaneous pressure changes, potentially leading to insufficient response to key transient signals. Therefore, a two-stream feature extraction network is used to achieve two-stream feature extraction. This involves constructing a parallel dual-branch network, using dilated causal convolution to capture global trends, and dynamic gated convolution to focus on local abrupt changes. Since the global trend features and local sensitive features extracted by the dual-branch network have different modalities and importance, simple feature concatenation or addition can lead to information overload, failing to establish a guiding relationship between the global context and local abrupt changes, thus reducing the model's ability to identify key transient features in the pressure signal. Therefore, a feature interaction fusion module is used to achieve feature interaction fusion, establishing a guiding weighting of global features on local features through a cross-domain attention mechanism. Conventional fully connected output layers rely solely on data-driven predictions, which may violate fundamental physical laws in fluid mechanics, such as mass and momentum conservation. This can lead to unrealistic pressure abrupt changes or discontinuous gradient variations in the output pressure prediction curve, reducing the reliability and generalization ability of the verification results. Therefore, a physical constraint output module is used to implement physical constraint output. By embedding simplified Navier-Stokes equations and constructing a physical consistency loss function, the prediction is forced to conform to the pressure propagation law. Deep neural networks are prone to hierarchical cumulative errors during feature transmission. Deep networks tend to capture low-frequency macroscopic features, while high-frequency details and phase information are easily lost. This can easily lead to phase lag or high-frequency distortion in the final predicted pressure curve, affecting the verification accuracy. Therefore, a multi-scale residual correction module is used to implement multi-scale residual correction, compensating for lost high-frequency details and phase information by fusing deep and shallow features.

[0095] Specifically, in one embodiment of this application, the dual-stream feature extraction network is a parallel dual-branch network, including a global trend branch and a local sensitive branch. The global trend branch is composed of multiple layers of dilated causal convolution stacked together, with the dilation rate changing cyclically according to the number of layers to exponentially expand the receptive field, and the output of each layer is subjected to layer normalization. The local sensitive branch adopts a dynamic gated convolution structure, generates a gate signal through the Sigmoid function, and then multiplies it element-wise with the convolution output processed by the hyperbolic tangent activation function.

[0096] In one embodiment of this application, the global trend branch feature extraction specifically involves: expanding the wavelet packet time-frequency features into a two-dimensional feature map compatible with time series through a fully connected layer, and then inputting it into the global trend branch. The global trend branch is composed of multiple layers of dilated causal convolutions stacked together, with the dilation rate cyclically changing according to the number of layers to exponentially expand the receptive field. The output of each layer is subjected to layer normalization processing, thereby obtaining the global trend branch output feature tensor, represented as:

[0097]

[0098] In the formula, represents the output feature tensor of the first layer of the global trend branch; LayerNorm(·) represents the layer normalization operation, which standardizes the output features of each layer to stabilize the training process of the deep network and alleviate the internal covariate bias; DCC(·) represents the dilated causal convolution operation, which expands the receptive field and maintains causality on the basis of standard convolution by using dilation rate. Let represent the output feature tensor of the (l-1)th layer of the global trend branch, which is also the input feature tensor of the first layer of the global trend branch. It is the extended wavelet packet time-frequency feature, specifically obtained by extending the wavelet packet time-frequency feature vector F. wp The wavelet packet time-frequency feature vector F of dimension K is expanded into a two-dimensional feature map compatible with time series by a pre-defined fully connected layer. wp Expanded to a dimension of 50×64, where 50 corresponds to the time step, and is related to the original pressure measurement signal X. raw The number of sampling points is consistent. 64 is the preferred value for the output feature dimension of the fully connected layer. It is a hyperparameter that can be adjusted according to actual conditions. The resulting expanded wavelet packet time-frequency feature matrix is ​​used as the input feature matrix of the first layer of the global trend branch. The format is a second-order tensor. d represents the dilation rate of the dilated causal convolution, which controls the interval between values ​​in the convolution kernel. mod is the modulo operator, which returns the remainder of the division. 1 is the network layer index, l = 1, 2, ..., L; L is the total number of network layers, such as L = 6.

[0099] It should be noted that 2 lmod3 This indicates the specific strategy for determining the expansion rate d, which is to take the result of modulo 3 on the current layer number and then iteratively apply 2. 0 =1,2 1 =2,2 2 The three expansion rate values ​​of 4 are used to achieve cyclical exponential growth of the receptive field, so as to effectively capture the dependencies at different time scales.

[0100] It should also be noted that conventional networks may ignore multi-scale dependencies when dealing with slow-changing trends and periodic components in pressure signals. The cyclic expansion rate strategy makes the receptive field growth more flexible, covering multi-scale time patterns, enabling the network to capture both short-term and long-term trends and improve global feature extraction capabilities.

[0101] The local sensitive branch feature extraction process is as follows: Dynamic differential spectrum features are input into the local sensitive branch, which employs a dynamically gated convolution structure. A gate signal is generated using the Sigmoid function, and then multiplied element-wise with the convolution output processed by the hyperbolic tangent activation function to obtain the local sensitive branch output feature tensor, represented as:

[0102]

[0103] In the formula, G gate The gating signal, generated by the gated convolutional layer, has a value range of [0,1] and is used to adaptively control the information flow and filter important features; Sig(·) represents the Sigmoid activation function, which compresses the input to between 0 and 1, generating gating weights; W g This represents the weight parameter matrix used for the convolution operation to generate the gated signal; these are trainable parameters; * indicates the standard convolution operation; b g The bias vector represents the convolution operation used to generate the gated signal, and is a trainable parameter; This represents the output feature tensor of the first layer of the localized sensitive branch; Let represent the output feature tensor of the (l-1)th layer of the local sensitive branch, which is also the input feature tensor of the first layer of the local sensitive branch, and represent the input feature vector of the first layer of the local sensitive branch. The format is a first-order tensor; F dds For dynamic differential spectrum features, the dimension is the same as the original pressure measurement signal X. raw The same, i.e., 50-dimensional, represents all the eigenvalues ​​of the dynamic differential spectrum at time t. The concatenated feature vector; ⊙ represents the Hadamard product, i.e., element-wise multiplication; tanh represents the hyperbolic tangent activation function, which transforms the features to the range [-1, 1]; W c This represents the weight parameter matrix used to perform convolution transformation on the input features; these are trainable parameters.

[0104] It should be noted that, The project uses the Sigmoid function to generate a gating signal to filter important features. The method uses the hyperbolic tangent activation function to transform the feature values ​​and applies the gating mechanism to local feature extraction, emphasizing abrupt change regions. This approach highlights local abrupt change features and avoids smoothing, especially when the network is more sensitive to instantaneous changes in pressure.

[0105] The two branches independently perform multi-level feature extraction, ultimately outputting the final layer's output feature tensor of the global trend branch and the Lth layer's output feature tensor of the local sensitivity branch, respectively. Let L be the output feature tensor of the global trend branch at layer L. Let be the output feature tensor of the Lth layer of the local sensitive branch.

[0106] Specifically, in one embodiment of this application, the feature interaction fusion module includes:

[0107] S401: Based on the final output features of the global trend branch, a linear transformation is used to generate a query vector, and based on the final output features of the local sensitive branch, a linear transformation is used to generate a key vector and a value vector.

[0108] S402: Calculate the similarity between the query vector and the transpose of the key vector, adjust by a scaling factor, superimpose the lower triangular mask matrix, and normalize by the Softmax function to obtain the attention weight matrix.

[0109] S403: The value vector is weighted and summed based on the attention weight matrix to obtain local features modulated by global information. These features are then residually connected with the final output features of the original global trend branch. Finally, attention fusion features are output through layer normalization.

[0110] In one embodiment of this application, firstly, the final output features of the global trend branch are transformed linearly to generate a query vector, and the final output features of the local sensitive branch are transformed linearly to generate a key vector and a value vector, respectively, as follows:

[0111]

[0112] In the formula, Q s This represents a query vector generated from global trend features, whose function is to inquire about local information patterns that require attention from a global perspective; W q K represents the weight matrix used to linearly project global features into the query space; these are trainable parameters. s This represents a key vector generated from locally sensitive features, whose function is to match it with the query vector, measuring the relevance between local information and the global query; W k V represents the weight matrix used to linearly project local features onto the key space, and is a trainable parameter; s W represents a value vector generated from locally sensitive features, whose function is to provide local information content that needs to be aggregated. v This represents the weight matrix used to linearly project local features onto the value space, and consists of trainable parameters.

[0113] It should be noted that the query is generated from global features, while the key and value are generated from local features. This enables global information to query local information, achieving a cross-domain effect. Global features guide the focus of local features in the query, and the global-local interaction establishes a contextual relationship, allowing local features to be modulated by global trends, highlighting local mutations that are related to the global context, and improving the targeting of feature fusion.

[0114] Next, the similarity between the query vector and the transpose of the key vector is calculated. After scaling, a lower triangular mask matrix is ​​superimposed to ensure causality. Then, the attention weight matrix is ​​obtained by normalization using the Softmax function, as follows:

[0115]

[0116] In the formula, A represents the attention weight matrix, and the element A in the i-th row and j-th column of the attention weight matrix is... ij This indicates the correlation strength between the i-th global query position and the j-th local key position; a larger value indicates a stronger correlation. For K s The transpose of ; soft(·) denotes the Softmax function, which operates on the last dimension, normalizing the input score vector into a probability distribution; d k Represents the key vector The dimension of M is used as a scaling factor to prevent the inner product from becoming too large and causing the softmax gradient to vanish; mask This represents a lower triangular mask matrix, where elements on and above the main diagonal are 0, and elements below the main diagonal are negative infinity (-∞). This is used to ensure causality and prevent any location from focusing on future information.

[0117] Finally, the value vector is weighted and summed using the attention weight matrix to obtain the local features modulated by global information. These features are then residually concatenated with the final output features of the original global trend branch. Finally, attention fusion features are output through layer normalization, as follows:

[0118]

[0119] In the formula, H fuse This represents the attention fusion feature, which is the output feature after fusion through a cross-domain attention mechanism.

[0120] It should be noted that residual connections preserve global information while injecting weighted local details to ensure that the global trend is not covered. The fused features not only maintain the long-term trend but also enhance local mutations, thereby improving feature integrity.

[0121] In one embodiment, the impact of residual correction on high-frequency details is analyzed, focusing on the pressure abrupt change region (time axis range: approximately 2-4 seconds, pressure unit: megapascals), to verify the ability of the multi-scale residual correction mechanism to recover high-frequency details. The results are as follows: Figure 5 As shown, the prediction without residual correction (blue dashed square) exhibits significant phase lag at the pressure rise edge, and the rise slope is lower than the true value, indicating that the deep network loses key dynamic response features. The prediction with residual correction (red solid diamond) accurately aligns with the pressure rise start point, reproduces the rise slope, and maintains subtle fluctuations in the pressure plateau region. This demonstrates that the residual correction mechanism, which integrates shallow features (preserving high-frequency details) and deep features (providing semantic information), effectively compensates for information loss during network transmission and solves the problems of phase lag and high-frequency distortion.

[0122] In one embodiment of this application, the physical constraint output module includes:

[0123] Specifically, the initial stress prediction sequence is obtained by performing a linear transformation through a fully connected layer based on the attention fusion features.

[0124] In one embodiment of this application, the attention fusion features are linearly transformed through a fully connected layer to generate an initial stress prediction sequence, as shown below:

[0125]

[0126] In the formula, This represents the initial pressure prediction sequence, calculated through a fully connected layer, with dimensions equal to the original pressure measurement signal X. raw Same; W o is the weight matrix of the fully connected layer, which are trainable parameters used to map high-dimensional attention fusion features to a one-dimensional stress prediction space; b o is the bias vector of the fully connected layer, and is the trainable parameter.

[0127] Specifically, in one embodiment of this application, the multi-scale residual correction module includes:

[0128] S501: Perform causal convolution based on the attention fusion features, upsample the shallow features output by the local sensitive branch, and add the two together to generate a multi-scale residual signal.

[0129] S502: The multi-scale residual signal is transformed by a linear projection weight matrix and added to the initial pressure prediction sequence to obtain the corrected pressure prediction value.

[0130] In one embodiment of this application, firstly, causal convolution is performed on the attention fusion features to preserve their semantic information, while the shallow features output by the locally sensitive branches are upsampled to restore spatial resolution. Then, the two are added together to generate a multi-scale residual signal, represented as follows:

[0131]

[0132] In the formula, R res This is a multi-scale residual signal, representing the residual signal generated after multi-scale feature fusion. Its purpose is to compensate for the high-frequency details and accurate phase information lost during the transmission process of the deep network. It has a dimension of 50, the same as the original pressure measurement signal X. raw Same; Upsample(·) represents the upsampling function. The term specifically refers to using transposed convolution to increase the temporal resolution of feature maps, thereby achieving synchronization with multi-scale residual signals R. res The dimensions remain consistent.

[0133] It should be noted that DCC(H) fuse In the term ,d=1), the dilation rate used for dilated causal convolution is d=1, which means performing standard causal convolution without expanding the receptive field. This allows for further processing of deep features while maintaining their semantic content and ensuring causality.

[0134] It should also be noted that, This represents the output feature of the second layer of the localized sensitive branch. This shallow feature is used to preserve more high-frequency details, edge information, and precise temporal location information of the original input.

[0135] It should also be noted that DCC(H) fuse The term d=1) processes deep features. The method restores the resolution of shallow features, preserves high-frequency information using shallow features, and fuses them with deep features, enabling the residual signal to accurately recover phase and high-frequency details, reducing lag and distortion in the prediction curve.

[0136] Then, the multi-scale residual signal is transformed using a linear projection weight matrix and added to the initial pressure prediction sequence to obtain the corrected pressure prediction value. The corrected prediction value is closer to the true pressure, especially improving the fitting in abrupt change regions, as shown below:

[0137]

[0138] In the formula, This represents the corrected pressure prediction value, processed by a multi-scale residual correction mechanism, which is expected to have more accurate high-frequency details and phase; W resThe linear projection weight matrix maps the number of channels and scale of the multi-scale residual signal to a space that matches the predicted value, so that the residual signal can effectively correct the initial prediction. It is a trainable parameter.

[0139] S104: Train the pressure measurement signal verification and correction model based on the time-frequency domain hybrid features, the real pressure value, and the real spatial gradient value.

[0140] In this embodiment, the time-frequency domain hybrid features obtained based on the original pressure measurement signal, along with the true pressure value and the true spatial gradient value, are used as input training data. The true pressure value is the supervision label, and the true spatial gradient value is used for physical constraints. Predicted values ​​are calculated through forward propagation, and then the total loss function (Loss) is calculated. total The model optimizes all trainable parameters in the network using the backpropagation algorithm, with the Adam optimizer employed during training. Training is repeated multiple times until the loss converges or a preset number of iterations is reached. After training, the model parameters are saved for subsequent inference, ensuring good generalization ability.

[0141] Specifically, in one embodiment of this application, the total loss function for model training includes a main loss function and a physical consistency loss function. The main loss function is determined based on the dynamic importance weight coefficient and the sampling point prediction error. The dynamic importance weight coefficient is determined based on the absolute value of the second derivative of the true pressure value. The physical consistency loss function is determined based on the temporal gradient and spatial gradient values ​​of the initial pressure prediction sequence.

[0142] In one embodiment of this application, the conventional mean squared error loss function treats all sampling points in the pressure time series equally, failing to highlight the importance of regions with drastic pressure changes, such as rapid pressurization and depressurization. This can easily lead to the model optimization process favoring smooth outputs, while underfitting abrupt regions with high requirements for dynamic response accuracy. Therefore, a time-adaptive weighting mechanism is used to dynamically allocate weights to strengthen the penalty for abrupt regions, and a total loss function is constructed by combining physical constraint loss to achieve a dynamically weighted hybrid loss.

[0143] Based on a simplified law of conservation of momentum, a physical consistency loss is calculated by comparing the temporal gradient of the initial pressure prediction sequence with the spatial gradient of the actual pressure value. This loss penalizes instances where the prediction results violate physical laws, and is expressed as follows:

[0144]

[0145] In the formula, Loss phy λ represents the physical consistency loss, used to penalize parts of the neural network predictions that violate the simplified law of conservation of momentum; LpThe adjustment coefficient representing the strength of physical constraints controls the relative importance of the physical loss term in the total loss function; its preferred value is λ. Lp =0.3; The sign for partial derivatives; The pressure value at the t-th sampling point in the initial pressure prediction sequence; This represents the network predicted stress value at time t. The first-order partial derivative with respect to time, i.e., the time gradient, is used in practical calculations for discretization approximation using the central difference method, and characterizes the instantaneous rate of change of the predicted pressure; ρ Lp Density represents the density of the pressure-conducting medium. It is a known physical constant, and its value depends on the specific fluid type. For example, oil ≈ 800-900 kg / m³. 3 Water ≈ 1000 kg / m³ 3 Standard air density ≈ 1.2-1.5 kg / m³ 3 ; The true pressure value at the t-th sampling point is directly measured by the calibration pressure sensor and serves as a label for supervised learning. The spatial gradient representing the true pressure value at the t-th sampling point is approximated by the finite difference method in actual calculations, and characterizes the rate of change of pressure in the spatial direction.

[0146] It should be noted that the cumulative item The reason why the summation is not performed starting from index 1 for the 2nd to 49th sampling points in the time series is to avoid the boundary points at the ends of the series where the central difference cannot be calculated.

[0147] It should also be noted that spatial gradient The acquisition of pressure gradients depends on additional hardware setup or prior knowledge of the system. For example, multiple calibrated high-precision reference sensors can be installed at different spatial locations in the fluid system being measured. In a pipeline, in addition to the measurement point, another reference sensor needs to be installed at a known distance upstream or downstream. At the same time, the pressure readings at these two locations are read simultaneously, and the difference between them is divided by the distance between the two points. The result is the spatial gradient at the current moment, which reflects the degree of non-uniformity or drastic change in the spatial distribution of pressure at that moment.

[0148] Based on the absolute value of the second derivative of the actual pressure value, a dynamic importance weighting coefficient for each sampling point is dynamically calculated to amplify the loss contribution in areas of drastic pressure change, expressed as:

[0149]

[0150] In the formula, w tα represents the dynamic importance weight coefficient calculated at the t-th sampling point, used to amplify the contribution of the loss term in the abrupt change region; w This is the enhancement coefficient, which controls the sensitivity of the weight to the magnitude of the acceleration due to pressure changes. The preferred value is α. w =2.0; express The second derivative with respect to time is calculated using a discrete approximation method based on the central difference. The magnitude of the term represents the degree of drastic change in pressure or the magnitude of acceleration at that moment. The larger the absolute value, the more likely it is a point of abrupt change, such as a peak or a steep edge.

[0151] It should be noted that in the process of calculating the dynamic importance weight coefficient, the addend 1 is the basic weight, ensuring that there is still a basic loss constraint even in the flat region of the signal where the second derivative is zero.

[0152] It should also be noted that the dynamic weighting mechanism assigns weights based on the absolute value of the second derivative of the real signal, which can amplify the loss contribution in abrupt change regions and strengthen the penalty for drastic changes, making the model optimization pay more attention to regions with rapid pressure changes and improving the accuracy of dynamic response.

[0153] Using dynamic importance weighting coefficients, the difference between the corrected pressure prediction value and the actual pressure value is weighted, and the weighted mean square error of all sampling points is calculated as the main loss, expressed as:

[0154]

[0155] In the formula, Loss main Principal loss is the main metric for data fitting accuracy; ||·|| represents the squared L2 norm. The term represents the prediction error at the t-th sampling point; The value represents the true pressure at the t-th sampling point, which is directly measured by the calibration pressure sensor and serves as a label for supervised learning.

[0156] Finally, the weighted principal loss and the physical consistency loss are linearly added together to obtain the total loss function guiding network training, expressed as:

[0157] Loss total =Loss main +Loss phy

[0158] In the formula, Loss total Let be the total loss function, representing the total loss ultimately used to guide the optimization of neural network training, which is the overall goal of model training.

[0159] It should be noted that conventional loss functions are only data-driven and may produce non-physical predictions. This invention incorporates physical constraints into the loss function, combining data-driven approaches with physical laws, making the pressure curve of the predicted output smoother and more reasonable, avoiding unrealistic abrupt changes, and improving generalization ability.

[0160] S105: Perform noise separation and feature extraction on the pressure measurement signal to be calibrated, and input the extracted features into the trained pressure measurement signal calibration model to obtain the calibrated pressure prediction value.

[0161] In this embodiment, the new original pressure measurement signal to be calibrated is input and subjected to the same preprocessing steps, including median filtering, system error estimation, and environmental noise extraction. Then, dynamic differential spectrum features and wavelet packet time-frequency features are extracted. These features are input into a dual-stream feature extraction network for global trend and local sensitive feature extraction. Attention fusion features are then obtained through feature interaction fusion. Finally, a calibrated pressure prediction value is generated through physical constraint output and multi-scale residual correction. The inference process does not rely on the actual pressure value or spatial gradient, but only on the input signal and model parameters, achieving efficient real-time or offline verification. The output prediction value is used for pressure gauge performance evaluation.

[0162] In one embodiment, a comparison of pressure prediction accuracy is performed to verify the accuracy of the present invention's technology in pressure value prediction. The results are as follows: Figure 6 As shown in the figure, the top subplot displays the changes in actual pressure values ​​(solid black line), predicted values ​​using conventional methods (dashed blue line), and predicted values ​​using the present invention (solid red line) over time (unit: seconds). The vertical axis represents pressure values ​​in kPa. The bottom subplot shows the absolute errors of the two methods (unit: kPa). Experimental data shows that during periods of stable pressure change, the two methods have similar prediction performance; however, in areas of abrupt pressure changes (such as peaks and troughs), the conventional method exhibits significant lag and bias, while the present invention closely tracks the actual value. The error curves show that the average error of the present invention is significantly lower than that of the conventional method, especially in abrupt regions where dynamic response requirements are high. This indicates that the dual-stream feature extraction network and dynamic weighted loss mechanism of the present invention can simultaneously capture global trends and local abrupt features, overcoming the shortcomings of a single network in dynamic response.

[0163] In practical applications, for the pressure gauge to be calibrated, its original measurement signal is collected and input into a trained model. The model infers and obtains a corrected pressure prediction value. Then, the prediction value is compared with the standard pressure value, and error indicators such as root mean square error or maximum permissible error are calculated to evaluate the accuracy of the pressure gauge. If the error exceeds the permissible range, the pressure gauge needs to be adjusted, calibrated, or repaired. The calibration process can also be automated and integrated into a management system, providing detailed calibration reports and decision support to ensure the reliability and compliance of pressure gauge measurements in industrial settings.

[0164] In the aforementioned pressure gauge calibration process management method, firstly, the original pressure measurement signal, the actual pressure value, and the actual spatial gradient value are acquired; then, noise separation is performed on the original pressure measurement signal, and time-frequency domain hybrid features are extracted based on the noise separation results; next, a pressure measurement signal calibration and correction model is constructed, which includes a dual-stream feature extraction network, a feature interaction fusion module, a physical constraint output module, and a multi-scale residual correction module; then, the pressure measurement signal calibration and correction model is trained based on the time-frequency domain hybrid features, the actual pressure value, and the actual spatial gradient value; finally, noise separation and feature extraction are performed on the pressure measurement signal to be calibrated, and the extracted features are input into the trained pressure measurement signal calibration and correction model to obtain the calibrated pressure prediction value. In other words, median filtering combined with Sigmoid adaptive weights separates system errors from environmental noise, avoiding the damage to abrupt signals caused by conventional filtering and significantly improving signal fidelity. Dynamic differential spectrum is used to highlight the rate of change at abrupt points, and adaptive wavelet packets are divided into frequency bands according to signal complexity, achieving complementary fusion of time-domain trends and frequency-domain perturbations, thus improving feature representation capabilities. Two branches, dilated causal convolution and dynamic gated convolution, are designed in parallel to extract global trend and local abrupt features, respectively. Fine fusion is achieved through a cross-domain attention mechanism, significantly enhancing the model's dynamic response capability. Simplified momentum conservation constraints are adopted in the output layer, and the loss in abrupt regions is dynamically emphasized through a second-order derivative weighting mechanism, balancing physical rationality and prediction accuracy, and improving generalization ability and engineering usability.

[0165] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0166] Based on the same inventive concept, this application also provides a pressure gauge calibration process management system for implementing the pressure gauge calibration process management method described above. The solution provided by this system is similar to the implementation scheme described in the above method; therefore, the specific limitations of one or more embodiments of the pressure gauge calibration process management system provided below can be found in the limitations of the pressure gauge calibration process management method described above, and will not be repeated here.

[0167] In one embodiment, such as Figure 7 As shown, a pressure gauge calibration process management system 700 is provided, including: a measurement data acquisition module 701, a pressure measurement signal noise separation and time-frequency domain hybrid feature construction module 703, a pressure measurement signal calibration and correction model construction module 705, a pressure measurement signal calibration and correction model training module 707, and a pressure measurement signal calibration and correction module 709, wherein:

[0168] The measurement data acquisition module 701 is used to acquire the raw pressure measurement signal, the actual pressure value, and the actual spatial gradient value.

[0169] The pressure measurement signal noise separation and time-frequency domain hybrid feature construction module 703 is used to perform noise separation on the original pressure measurement signal and extract time-frequency domain hybrid features based on the noise separation result.

[0170] The pressure measurement signal verification and correction model construction module 705 is used to construct a pressure measurement signal verification and correction model. The model includes a dual-stream feature extraction network, a feature interaction fusion module, a physical constraint output module, and a multi-scale residual correction module.

[0171] The pressure measurement signal verification and correction model training module 707 is used to train the pressure measurement signal verification and correction model based on the time-frequency domain hybrid features, the real pressure value, and the real spatial gradient value.

[0172] The pressure measurement signal verification and correction module 709 is used to perform noise separation and feature extraction on the pressure measurement signal to be corrected, and input the extracted features into the trained pressure measurement signal verification and correction model to obtain the corrected pressure prediction value.

[0173] In one embodiment of this application, the pressure measurement signal noise separation and time-frequency domain hybrid feature construction module is further configured to:

[0174] The original pressure measurement signal is processed by median filtering to obtain the median filtered baseline signal;

[0175] Based on the absolute difference between the original pressure measurement signal and the median filtered baseline signal, the system error scalar value is adaptively calculated using the Sigmoid function;

[0176] The environmental noise components are determined based on the original pressure measurement signal, the median filtered baseline signal, and the system error scalar value.

[0177] In one embodiment of this application, the pressure measurement signal noise separation and time-frequency domain hybrid feature construction module is further configured to:

[0178] The second derivative of the median filtered baseline signal is calculated, and convolution smoothing is performed using a Hanning window to obtain dynamic differential spectrum eigenvalues.

[0179] The number of frequency bands for wavelet packet decomposition is determined based on the maximum, minimum, and standard deviation of the median filtered baseline signal.

[0180] The environmental noise components are decomposed using wavelet packets, and the energy of each frequency band is extracted as a time-frequency feature.

[0181] In one embodiment of this application, the dual-stream feature extraction network is a parallel dual-branch network, including a global trend branch and a local sensitive branch. The global trend branch is composed of multiple layers of dilated causal convolution stacked together, with the dilation rate changing cyclically according to the number of layers to exponentially expand the receptive field, and the output of each layer is subjected to layer normalization. The local sensitive branch adopts a dynamic gated convolution structure, generates a gate signal through the Sigmoid function, and then multiplies it element-wise with the convolution output processed by the hyperbolic tangent activation function.

[0182] In one embodiment of this application, the feature interaction fusion module includes:

[0183] A query vector is generated by linear transformation based on the final output features of the global trend branch, and a key vector and a value vector are generated by linear transformation based on the final output features of the local sensitive branch.

[0184] The similarity between the query vector and the transpose of the key vector is calculated, and after being adjusted by a scaling factor, a lower triangular mask matrix is ​​superimposed. The attention weight matrix is ​​then obtained by normalizing the matrix using the Softmax function.

[0185] The value vector is weighted and summed based on the attention weight matrix to obtain local features modulated by global information. These features are then residually connected to the final output features of the original global trend branch. Finally, attention fusion features are output through layer normalization.

[0186] In one embodiment of this application, the physical constraint output module includes:

[0187] Based on the attention fusion features, an initial pressure prediction sequence is obtained by linear transformation through a fully connected layer.

[0188] In one embodiment of this application, the multi-scale residual correction module includes:

[0189] Based on the attention fusion features, causal convolution is performed, and the shallow features output by the local sensitive branch are upsampled. The two are then added together to generate a multi-scale residual signal.

[0190] The multi-scale residual signal is transformed using a linear projection weight matrix and added to the initial pressure prediction sequence to obtain the corrected pressure prediction value.

[0191] In one embodiment of this application, the total loss function for model training includes a main loss function and a physical consistency loss function. The main loss function is determined based on the dynamic importance weight coefficient and the sampling point prediction error. The dynamic importance weight coefficient is determined based on the absolute value of the second derivative of the true pressure value. The physical consistency loss function is determined based on the temporal gradient and spatial gradient values ​​of the initial pressure prediction sequence.

[0192] The various modules in the aforementioned pressure gauge calibration process management system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the processor of a computer device in hardware form or independent of it, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0193] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 8 As shown, the computer device includes a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, NFC (Near Field Communication), or other technologies. When executed by the processor, the computer program implements a pressure gauge calibration process management method. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0194] Those skilled in the art will understand that Figure 8 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0195] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0196] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0197] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0198] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0199] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0200] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0201] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A pressure gauge calibration process management method, characterized in that, The method includes: Acquire the raw pressure measurement signal, the actual pressure value, and the actual spatial gradient value; The original pressure measurement signal is subjected to noise separation, and time-frequency domain hybrid features are extracted based on the noise separation results; A pressure measurement signal verification and correction model is constructed, which includes a dual-stream feature extraction network, a feature interaction fusion module, a physical constraint output module, and a multi-scale residual correction module. The pressure measurement signal verification and correction model is trained based on the time-frequency domain hybrid features, the true pressure value, and the true spatial gradient value. Noise separation and feature extraction are performed on the pressure measurement signal to be calibrated. The extracted features are then input into the trained pressure measurement signal calibration model to obtain the calibrated pressure prediction value.

2. The pressure gauge calibration process management method according to claim 1, characterized in that, The noise separation of the original pressure measurement signal includes: The original pressure measurement signal is processed by median filtering to obtain the median filtered baseline signal; Based on the absolute difference between the original pressure measurement signal and the median filtered baseline signal, the system error scalar value is adaptively calculated using the Sigmoid function; The environmental noise components are determined based on the original pressure measurement signal, the median filtered baseline signal, and the system error scalar value.

3. The pressure gauge calibration process management method according to claim 2, characterized in that, The extraction of time-frequency domain hybrid features based on noise separation results includes: The second derivative of the median filtered baseline signal is calculated, and convolution smoothing is performed using a Hanning window to obtain dynamic differential spectrum eigenvalues. The number of frequency bands for wavelet packet decomposition is determined based on the maximum, minimum, and standard deviation of the median filtered baseline signal. The environmental noise components are decomposed using wavelet packets, and the energy of each frequency band is extracted as a time-frequency feature.

4. The pressure gauge calibration process management method according to claim 1, characterized in that, The dual-stream feature extraction network is a parallel dual-branch network, including a global trend branch and a local sensitive branch. The global trend branch is composed of multiple layers of dilated causal convolution stacked together, with the dilation rate changing cyclically according to the number of layers to exponentially expand the receptive field. The output of each layer is processed by layer normalization. The local sensitive branch adopts a dynamic gated convolution structure, which generates a gate signal through the Sigmoid function and then multiplies it element-wise with the convolution output processed by the hyperbolic tangent activation function.

5. The pressure gauge calibration process management method according to claim 4, characterized in that, The feature interaction fusion module includes: A query vector is generated by linear transformation based on the final output features of the global trend branch, and a key vector and a value vector are generated by linear transformation based on the final output features of the local sensitive branch. The similarity between the query vector and the transpose of the key vector is calculated, and after being adjusted by a scaling factor, a lower triangular mask matrix is ​​superimposed. The attention weight matrix is ​​then obtained by normalizing the matrix using the Softmax function. The value vector is weighted and summed based on the attention weight matrix to obtain local features modulated by global information. These features are then residually connected to the final output features of the original global trend branch. Finally, attention fusion features are output through layer normalization.

6. The pressure gauge calibration process management method according to claim 5, characterized in that, The physical constraint output module includes: Based on the attention fusion features, an initial pressure prediction sequence is obtained by linear transformation through a fully connected layer.

7. The pressure gauge calibration process management method according to claim 6, characterized in that, The multi-scale residual correction module includes: Based on the attention fusion features, causal convolution is performed, and the shallow features output by the local sensitive branch are upsampled. The two are then added together to generate a multi-scale residual signal. The multi-scale residual signal is transformed using a linear projection weight matrix and added to the initial pressure prediction sequence to obtain the corrected pressure prediction value.

8. The pressure gauge calibration process management method according to claim 1, characterized in that, The total loss function for model training includes a main loss function and a physical consistency loss function. The main loss function is determined based on the dynamic importance weight coefficient and the sampling point prediction error. The dynamic importance weight coefficient is determined based on the absolute value of the second derivative of the true pressure value. The physical consistency loss function is determined based on the temporal gradient and spatial gradient values ​​of the initial pressure prediction sequence.

9. A pressure gauge calibration process management system, characterized in that, The system includes: The measurement data acquisition module is used to acquire the raw pressure measurement signal, the actual pressure value, and the actual spatial gradient value; A pressure measurement signal noise separation and time-frequency domain hybrid feature construction module is used to perform noise separation on the original pressure measurement signal and extract time-frequency domain hybrid features based on the noise separation result. The pressure measurement signal verification and correction model construction module is used to construct a pressure measurement signal verification and correction model. The model includes a dual-stream feature extraction network, a feature interaction fusion module, a physical constraint output module, and a multi-scale residual correction module. The pressure measurement signal verification and correction model training module is used to train the pressure measurement signal verification and correction model based on the time-frequency domain hybrid features, the real pressure value, and the real spatial gradient value. The pressure measurement signal verification and correction module is used to perform noise separation and feature extraction on the pressure measurement signal to be corrected. The extracted features are then input into the trained pressure measurement signal verification and correction model to obtain the corrected pressure prediction value.

Images