A deep learning-based semi-infinite long pipe pulsating pressure signal reconstruction method and device, storage medium and electronic equipment
By constructing a hybrid reconstruction model using deep learning methods, the problem of pressure signal distortion caused by semi-infinitely long pressure guide pipes in heavy-duty gas turbines was solved, achieving high-precision signal reconstruction and online monitoring while reducing computational costs.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHINA UNITED GAS TURBINE TECH CO LTD
- Filing Date
- 2026-03-31
- Publication Date
- 2026-07-07
AI Technical Summary
In heavy-duty gas turbines, pressure signal distortion caused by semi-infinite pressure guide tubes makes it difficult for existing technologies to achieve high-precision signal reconstruction, especially under high-temperature and confined space conditions, making it impossible to accurately measure the pulsating pressure inside the combustion chamber.
A deep learning-based method for reconstructing pulsating pressure signals in a semi-infinite tube is adopted. By constructing a hybrid reconstruction model, signal feature extraction and reconstruction are performed using CNN and Transformer modules, and training is carried out in combination with a physical constraint loss function to achieve end-to-end signal reconstruction.
It achieves high-precision signal reconstruction, overcomes pressure signal amplitude attenuation and phase delay, meets online monitoring requirements, and reduces computing costs.
Smart Images
Figure CN121935569B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of signal reconstruction technology, and in particular to a method, apparatus, storage medium, and electronic device for reconstructing a semi-infinite tube pulsating pressure signal based on deep learning. Background Technology
[0002] In the research, development, and operational monitoring of high-temperature power equipment such as heavy-duty gas turbines, accurate measurement of pulsating pressure inside the combustion chamber is crucial for assessing combustion stability, suppressing thermoacoustic oscillations, and ensuring structural safety. However, due to the harsh operating conditions of high temperatures and confined spaces in some areas of heavy-duty gas turbines, pulsating pressure sensors cannot be directly installed on the combustion chamber wall. In engineering practice, a thin pressure-conducting tube can be used to guide the pressure signal from the combustion chamber to a location far from the high-temperature zone, facilitating sensor installation. By selecting appropriate tube length and diameter to minimize internal reflection effects, the measured signal can be made as close as possible to the original pulsating pressure (e.g., ...). Figure 1 (As shown).
[0003] However, during the propagation of the pulsating pressure along the pressure tap, the combined effects of sound wave reflection, air viscosity, pipe wall heat loss, and the mounting effect cause significant distortion in both amplitude and phase between the oscillating pressure measured by the pressure sensor and the actual oscillating pressure in the combustion chamber, resulting in severely distorted measurement signals. Therefore, it is necessary to reconstruct the distorted signal to obtain accurate source-end pressure information.
[0004] In existing technologies, the measured signal is Fourier transformed to the frequency domain using frequency domain correction, multiplied by the inverse transfer function, and then reconstructed using inverse Fourier transform. However, this method has limitations due to excessive idealization assumptions, neglect of practical factors, and inability to handle nonlinear distortion problems. Alternatively, a three-dimensional geometric model of the pipeline can be established through simulation, and the fluid flow and sound wave propagation can be numerically solved in a computer using the Navier-Stokes equations, continuity equations, or energy equations. However, the computational cost of these methods is high, and if parameters such as the precise dimensions of the pipeline and the fluid viscosity are inaccurate, the reconstruction results will also be inaccurate, failing to meet the needs of online monitoring. Another approach is to treat the pipeline system as a distributed parameter system using modal analysis, where the complex pressure fluctuation field inside can be decomposed into a linear superposition of a series of orthogonal characteristic modes. However, this method is sensitive to boundary conditions, and the shape and frequency of the modes are determined by the system's boundary conditions, leading to inaccurate high-frequency signal amplitudes. Summary of the Invention
[0005] The present invention aims to at least partially solve one of the technical problems in the related art.
[0006] To address this, this invention proposes a deep learning-based method for reconstructing pulsating pressure signals in semi-infinite pipes. This method uses a target hybrid reconstruction model to perform forward inference on the pressure distortion signal at the pipe outlet, and outputs the reconstructed pressure signal at the pipe source end in real time. This overcomes the pressure signal amplitude attenuation and phase delay caused by the semi-infinite pipe, achieving high-precision signal reconstruction. Simultaneously, the target hybrid reconstruction model automatically performs end-to-end signal reconstruction in real time, eliminating the need for manual intervention, meeting the requirements of online monitoring, and reducing computational costs.
[0007] Another objective of this invention is to propose a deep learning-based device for reconstructing pulsating pressure signals in a semi-infinite tube.
[0008] To achieve the above objectives, this invention proposes a method for reconstructing pulsating pressure signals in a semi-infinite tube based on deep learning, the method comprising:
[0009] In the semi-infinite tube experimental system, the source signal is acquired by a high-precision reference sensor installed at the source end of the tube, and the distortion signal is measured at the outlet of the tube.
[0010] The source signal and the distorted signal are time-aligned and frequency domain features are extracted to obtain the first dataset;
[0011] Construct an initial hybrid reconstruction model, wherein the initial hybrid reconstruction model includes a CNN module, a Transformer module, and a decoding and reconstruction module;
[0012] The initial hybrid reconstruction model is trained based on the first dataset and the physical constraint loss function to obtain the target hybrid reconstruction model;
[0013] The pressure distortion signal at the pipeline outlet is acquired and input into the target hybrid reconstruction model for forward inference, and the reconstructed pipeline source end pressure signal is output in real time.
[0014] The deep learning-based method for reconstructing a semi-infinite tube pulsating pressure signal according to an embodiment of the present invention may also have the following additional technical features:
[0015] In one embodiment of the present invention, training the initial hybrid reconstruction model based on the first dataset and the physical constraint loss function to obtain the target hybrid reconstruction model includes:
[0016] The first dataset is preprocessed to obtain the second dataset;
[0017] Based on the second dataset, a training dataset and a test dataset are obtained for training the initial hybrid reconstruction model;
[0018] The initial hybrid reconstruction model is trained using the training dataset and the physical constraint loss function to obtain the trained hybrid reconstruction model;
[0019] The trained hybrid reconstruction model is tested and optimized using the test dataset to obtain the target hybrid reconstruction model.
[0020] In one embodiment of the present invention, training the initial hybrid reconstruction model using the training dataset and the physical constraint loss function to obtain the trained hybrid reconstruction model includes:
[0021] The distorted signal in the training dataset is input into the initial hybrid reconstruction model to obtain the reconstruction pressure signal;
[0022] The reconstructed pressure signal and the source signal are input into the physical constraint loss function to obtain the corresponding loss value;
[0023] The network parameters in the initial hybrid reconstruction model are updated using the loss value until the initial hybrid reconstruction model converges, or the number of iterations of the network parameters reaches a preset number, to obtain the trained hybrid reconstruction model.
[0024] In one embodiment of the present invention, the step of inputting the distortion signal in the training dataset into the initial hybrid reconstruction model to obtain the reconstruction pressure signal includes:
[0025] The CNN module extracts local temporal features of the distorted signals in the training dataset and outputs a high-dimensional feature sequence.
[0026] The Transformer module captures the global long-term dependencies of the high-dimensional feature sequence and outputs a reconstructed feature vector sequence.
[0027] The feature vector sequence is mapped back to a time-domain signal of the same length as the source signal by the decoding and reconstruction module to obtain the reconstructed pressure signal.
[0028] In one embodiment of the present invention, the CNN module comprises multiple sequentially connected feature extraction modules, and each feature extraction module includes a one-dimensional convolutional layer, an activation function layer, and a pooling layer; the step of extracting local temporal features of the distorted signal in the training dataset through the CNN module and outputting a high-dimensional feature sequence includes:
[0029] By using multiple learnable convolutional kernels in the one-dimensional convolutional layer, the input temporal signal is convolved using a sliding window method to obtain the local temporal features of the signal.
[0030] The local time-domain features of the signal are nonlinearly activated by the activation function layer to obtain the activated local time-domain features;
[0031] The activated local temporal features are downsampled by the pooling layer to obtain the downsampled local temporal features.
[0032] By stacking multiple feature extraction modules, the input distorted signal is subjected to layer-by-layer dimensionality reduction and feature abstraction, and the downsampled local temporal features obtained by the final feature extraction module are determined as a high-dimensional feature sequence.
[0033] In one embodiment of the present invention, the Transformer module includes a Transformer encoder consisting of N sequentially stacked encoding layers, wherein the encoding layers include a multi-head self-attention module and a feedforward neural network module; the step of capturing the global long-term dependencies of the high-dimensional feature sequence through the Transformer module and outputting a reconstructed feature vector sequence includes:
[0034] Add positional encoding to each feature vector in the high-dimensional feature sequence to obtain a new sequence containing temporal positional information;
[0035] The new sequence is input into the Transformer encoder, and the multi-head self-attention module performs self-attention calculation on the new sequence to obtain the feature vector at each position.
[0036] The feedforward neural network module performs nonlinear transformation and mapping on the feature vector at each position to obtain a temporal reconstruction vector sequence.
[0037] A Transformer encoder consisting of N sequentially stacked coding layers is used, and the temporal reconstruction vector sequence output by the Transformer encoder is determined as the reconstructed feature vector sequence.
[0038] In one embodiment of the present invention, the physical constraint loss function includes:
[0039]
[0040] in, The absolute error between the reconstructed pressure signal and the source signal. For transfer function constraint loss, The loss is constrained by the wave equation. and The weights are for each constraint.
[0041] To achieve the above objectives, another aspect of the present invention proposes a deep learning-based device for reconstructing a semi-infinite tube pulsating pressure signal, the device comprising:
[0042] The acquisition module is used to acquire the source signal in the semi-infinite tube experimental system by means of a high-precision reference sensor installed at the source end of the tube, and to acquire the measured distortion signal at the outlet of the tube.
[0043] The data processing module is used to perform time-series alignment and frequency domain feature extraction on the source signal and the distorted signal to obtain the first dataset;
[0044] A building module is used to build an initial hybrid reconstruction model, wherein the initial hybrid reconstruction model includes a CNN module, a Transformer module, and a decoding and reconstruction module;
[0045] The training module is used to train the initial hybrid reconstruction model based on the first dataset and the physical constraint loss function to obtain the target hybrid reconstruction model;
[0046] The reconstruction module is used to acquire the pressure distortion signal at the pipeline outlet, input the pressure distortion signal into the target hybrid reconstruction model for forward inference, and output the reconstructed pipeline source end pressure signal in real time.
[0047] This invention discloses a method, apparatus, storage medium, and electronic device for reconstructing pulsating pressure signals in a semi-infinite pipe based on deep learning. The method includes: acquiring a source signal in a semi-infinite pipe experimental system using a high-precision reference sensor installed at the source end of the pipe, and acquiring a measured distorted signal at the outlet of the pipe; performing time-series alignment and frequency domain feature extraction on the source signal and the distorted signal to obtain a first dataset; constructing an initial hybrid reconstruction model, wherein the initial hybrid reconstruction model includes a CNN module, a Transformer module, and a decoding and reconstruction module; training the initial hybrid reconstruction model based on the first dataset and a physical constraint loss function to obtain a target hybrid reconstruction model; acquiring the pressure distortion signal at the pipe outlet, inputting the pressure distortion signal into the target hybrid reconstruction model for forward inference, and outputting the reconstructed pressure signal at the source end of the pipe in real time. Therefore, this invention can use a target hybrid reconstruction model to perform forward inference on the pressure distortion signal at the pipeline outlet and output the reconstructed pipeline source pressure signal in real time, thereby overcoming the pressure signal amplitude attenuation and phase delay caused by semi-infinite pipes and achieving high-precision signal reconstruction. At the same time, the target hybrid reconstruction model can automatically perform end-to-end signal reconstruction in real time without relying on manual intervention, meeting the needs of online monitoring and reducing computational costs.
[0048] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description
[0049] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:
[0050] Figure 1 This is a schematic diagram of a semi-infinite tube according to an embodiment of the present invention;
[0051] Figure 2 This is a flowchart of a method for reconstructing a semi-infinite tube pulsating pressure signal based on deep learning, according to an embodiment of the present invention.
[0052] Figure 3 This is a structural diagram of a deep learning-based semi-infinite tube pulsating pressure signal reconstruction device according to an embodiment of the present invention. Detailed Implementation
[0053] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.
[0054] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.
[0055] The following description, with reference to the accompanying drawings, describes a method and apparatus for reconstructing a semi-infinite tube pulsating pressure signal based on deep learning, according to an embodiment of the present invention.
[0056] Figure 2 This is a flowchart of a method for reconstructing a semi-infinite tube pulsating pressure signal based on deep learning, according to an embodiment of the present invention.
[0057] like Figure 2 As shown, the method includes:
[0058] S1, in the semi-infinite tube experimental system, the source signal is obtained by a high-precision reference sensor installed at the source end of the tube, and the distortion signal is obtained by measurement at the outlet of the tube.
[0059] In one embodiment of the present invention, the above-described semi-infinite tube experimental system can generate a known and controllable pulsating pressure signal p(t) at the source end of the tube by creating a known excitation using devices such as a shock tube, an acoustic generator, or a pulse valve. Furthermore, a high-precision reference sensor can be installed at the source end of the tube, and the source signal p(t) can be acquired through the high-precision reference sensor.
[0060] Furthermore, in one embodiment of the present invention, the corresponding distortion signal y(t) is simultaneously measured at the outlet of the long tube.
[0061] S2 performs time-series alignment and frequency domain feature extraction on the source signal and the distorted signal to obtain the first dataset.
[0062] In one embodiment of the present invention, after obtaining the source signal and the distorted signal through the above steps, the source signal and the distorted signal can be time-aligned and frequency domain features extracted to obtain the first dataset.
[0063] Specifically, in one embodiment of the present invention, the source signal p(t) and the distorted signal y(t) obtained after time-series alignment and frequency domain feature extraction of the source signal and the distorted signal can be combined to form a corresponding [y(t), p(t)] data pair, which is determined as the first dataset.
[0064] S3, construct the initial hybrid reconstruction model, which includes a CNN module, a Transformer module, and a decoding and reconstruction module.
[0065] In one embodiment of the present invention, the CNN module may be composed of multiple sequentially connected feature extraction modules, and each feature extraction module includes a one-dimensional convolutional layer, an activation function layer and a pooling layer; the Transformer module may include a Transformer encoder composed of N sequentially stacked encoding layers, wherein the encoding layers include a multi-head self-attention module and a feedforward neural network module.
[0066] S4. The initial hybrid reconstruction model is trained based on the first dataset and the physical constraint loss function to obtain the target hybrid reconstruction model.
[0067] In one embodiment of the present invention, after obtaining the first dataset through the above steps, the initial hybrid reconstruction model can be trained based on the first dataset and the physical constraint loss function to obtain the target hybrid reconstruction model.
[0068] In one embodiment of the present invention, the method for training an initial hybrid reconstruction model based on a first dataset to obtain a target hybrid reconstruction model may include the following steps:
[0069] S41, preprocess the first dataset to obtain the second dataset.
[0070] In one embodiment of the present invention, the data pairs [y(t), p(t)] in the first dataset are normalized or standardized, and a suitable time window length h is selected according to the sampling frequency to construct discrete time series data of length h, thereby obtaining the second dataset.
[0071] S42, based on the second dataset, obtain the training dataset and test dataset for training the initial hybrid reconstruction model.
[0072] In one embodiment of the present invention, after obtaining the second dataset through the above steps, the data pairs in the second dataset can be divided into a training dataset and a test dataset for training the initial hybrid reconstruction model according to a preset ratio. For example, assuming the preset ratio is 8:2, that is, 80% of the data pairs in the second dataset are randomly divided into the training dataset and 20% of the data pairs are divided into the test dataset.
[0073] S43, the initial hybrid reconstruction model is trained using the training dataset and the physical constraint loss function to obtain the trained hybrid reconstruction model.
[0074] In one embodiment of the present invention, after obtaining the training dataset through the above steps, the initial hybrid reconstruction model can be trained using the training dataset and the physical constraint loss function to obtain the trained hybrid reconstruction model.
[0075] In one embodiment of the present invention, the method for training an initial hybrid reconstruction model using a training dataset and a physical constraint loss function to obtain a trained hybrid reconstruction model may include the following steps:
[0076] S431, input the distorted signal from the training dataset into the initial hybrid reconstruction model to obtain the reconstruction pressure signal.
[0077] In one embodiment of the present invention, after obtaining the training dataset through the above steps, the distortion signal in the training dataset can be input into the initial hybrid reconstruction model to obtain the reconstruction pressure signal.
[0078] In one embodiment of the present invention, the above-mentioned initial hybrid reconstruction model may include a CNN module, a Transformer module, and a decoding and reconstruction module.
[0079] Furthermore, in one embodiment of the present invention, the method of inputting the distorted signal in the training dataset into the initial hybrid reconstruction model to obtain the reconstructed pressure signal may include the following steps:
[0080] S4311 extracts local temporal features of distorted signals in the training dataset through a CNN module and outputs a high-dimensional feature sequence.
[0081] In one embodiment of the present invention, the CNN module described above may be composed of multiple feature extraction modules connected in sequence, and each feature extraction module includes a one-dimensional convolutional layer, an activation function layer and a pooling layer.
[0082] In one embodiment of the present invention, the method described above for extracting local temporal features of distorted signals in the training dataset using a CNN module and outputting a high-dimensional feature sequence may include the following steps:
[0083] S43111 uses multiple learnable convolutional kernels in a one-dimensional convolutional layer and a sliding window method to perform convolution operations on the input time-series signal to obtain the local time-domain features of the signal.
[0084] In one embodiment of the present invention, the above-mentioned one-dimensional convolutional layer can adopt a multi-scale parallel structure, using multiple convolutional kernels with different time window sizes to simultaneously perform convolution operations on the input time-series signal, so as to extract local features at different time scales in parallel and obtain the local time-domain features of the signal.
[0085] S43112 performs nonlinear activation on the local time-domain features of the signal through an activation function layer to obtain the activated local time-domain features.
[0086] In one embodiment of the present invention, the activation function used in the above-mentioned activation function layer may be the ReLU function or a variant thereof.
[0087] S43113 uses a pooling layer to downsample the activated local temporal features to obtain the downsampled local temporal features.
[0088] In one embodiment of the present invention, the activated local temporal features can be compressed into a feature sequence through a pooling layer to preserve the main pulsating structure.
[0089] In one embodiment of the present invention, the pooling layer may be a one-dimensional max pooling layer or a one-dimensional average pooling layer.
[0090] S43114 performs layer-by-layer dimensionality reduction and feature abstraction on the input distorted signal by stacking multiple feature extraction modules, and determines the downsampled local temporal features obtained by the final feature extraction module as a high-dimensional feature sequence.
[0091] S4312 captures the global long-term dependencies of high-dimensional feature sequences through the Transformer module and outputs a reconstructed feature vector sequence.
[0092] In one embodiment of the present invention, the Transformer module includes a Transformer encoder consisting of N sequentially stacked encoding layers. The encoding layers may include a multi-head self-attention module and a feedforward neural network module.
[0093] In one embodiment of the present invention, the method described above for capturing the global long-term dependencies of the high-dimensional feature sequence through the Transformer module and outputting a reconstructed feature vector sequence may include the following steps:
[0094] S43121, add position encoding to each feature vector in the high-dimensional feature sequence to obtain a new sequence containing temporal position information.
[0095] In one embodiment of the present invention, the above-mentioned position encoding may be a fixed encoding generated based on sine and cosine functions, or a position embedding vector that can be trained together with the network to maintain the time order.
[0096] S43122, the new sequence is input into the Transformer encoder, and the new sequence is self-attention calculated by the multi-head self-attention module to obtain the feature vector at each position.
[0097] S43123 uses a feedforward neural network module to perform nonlinear transformation and mapping on the feature vector at each position to obtain a temporal reconstruction vector sequence.
[0098] In one embodiment of the present invention, each of the above-described coding layers may further include residual connections and layer normalization operations. Specifically, in one embodiment of the present invention, a residual connection is made between the output and input of the multi-head self-attention module, and the summed result is subjected to layer normalization; a residual connection is made between the output and input of the feedforward neural network module, and the summed result is subjected to layer normalization.
[0099] S43124 uses a Transformer encoder consisting of N sequentially stacked coding layers, and determines the temporal reconstruction vector sequence output by the Transformer encoder as the reconstructed feature vector sequence.
[0100] S4313 uses a decoding and reconstruction module to map the feature vector sequence back to a time-domain signal of the same length as the source signal, thus obtaining a reconstructed pressure signal.
[0101] In one embodiment of the present invention, the feature vector sequence can be processed by one or more one-dimensional convolutional layers to fuse global context information within the sequence; and the fused feature sequence output by the one-dimensional convolutional layer is input into a fully connected layer to map the feature dimension to a single channel, thereby generating a reconstructed pressure signal with the same length as the original target signal.
[0102] S432 inputs the reconstructed pressure signal and the source signal into the physical constraint loss function to obtain the corresponding loss value.
[0103] In one embodiment of the present invention, the physical information of pipeline dynamics can be incorporated into the training process. By embedding transfer function constraints and wave equation constraints in the loss function, the physical generalization ability of the reconstruction model can be improved, the model overfitting can be avoided, and the reconstruction results obtained by the model can still satisfy physical laws and have good generalization ability even when experimental data is sparse, thereby improving the generalization ability of the model.
[0104] In one embodiment of the present invention, the above-mentioned physical constraint loss function can be:
[0105]
[0106] in, To reconstruct the absolute value error between the pressure signal and the source signal, For the transfer function constraint loss, L wave The loss is constrained by the wave equation. and The weights are for each constraint.
[0107] In one embodiment of the present invention, it is assumed that the pressure wave propagates in the pipe as a plane wave, the pressure fluctuation satisfies the linear acoustic theory, the fluid is a homogeneous medium, and the linearized derivation is made using the energy conservation equation, the flow continuity equation, and the Navier-Stokes equations. When the pressure wave propagates from x=0 to x=L, its amplitude attenuates by e. -αL α is the attenuation coefficient; the phase delay is e -iβL β is the phase coefficient, and i is the imaginary unit; they satisfy the relationship k = β - iα for the complex wavenumber k. Referring to the classical expression for the value of k proposed by Kinsler, we can obtain... ), where d is the pipe diameter, Kinematic viscosity, For signal frequency, Pr is the specific heat ratio, and Pr is the Prandtl number.
[0108] Furthermore, the transfer function is derived as follows: , where c is the speed of sound and P is pressure. For density.
[0109] Furthermore, the transfer function constraint loss is:
[0110]
[0111] The FFT function converts the time-domain acquired signal into a frequency-domain signal.
[0112] Furthermore, in one embodiment of the present invention, for sound wave propagation in a semi-infinite tube, the most fundamental physical law is the sound wave equation. Discretizing the wave equation in time and space yields the wave equation constraint loss as follows:
[0113]
[0114] S433, update the network parameters in the initial hybrid reconstruction model using the loss value until the initial hybrid reconstruction model converges, or when the number of iterations of the network parameters reaches a preset number, the trained hybrid reconstruction model is obtained.
[0115] In one embodiment of the present invention, the above-mentioned preset number of times can be set as needed, such as 100.
[0116] S44. The trained hybrid reconstruction model is tested and optimized using the test dataset to obtain the target hybrid reconstruction model.
[0117] In one embodiment of the present invention, after obtaining the trained hybrid reconstruction model through the above steps, the trained hybrid reconstruction model is tested and optimized using a test dataset to obtain the accuracy performance of the trained hybrid reconstruction model on new samples and evaluate the model's generalization ability. Then, based on the evaluation results of the test dataset, the trained hybrid reconstruction model is adjusted and optimized, for example, by changing the network structure or hyperparameter settings, to improve the model's accuracy, robustness, and performance.
[0118] Furthermore, in one embodiment of the present invention, after obtaining the target hybrid reconstruction model through the above steps, the target hybrid reconstruction model can be exported into a format suitable for the production environment, and a script can be set to load the solidified model. Additionally, the script can be run on a data acquisition computer or a dedicated edge computing device to perform reconstruction using the aforementioned target hybrid reconstruction model.
[0119] S5 acquires the pressure distortion signal at the pipeline outlet and inputs the pressure distortion signal into the target hybrid reconstruction model for forward inference, and outputs the reconstructed pipeline source end pressure signal in real time.
[0120] In one embodiment of the present invention, after obtaining the target hybrid reconstruction model through the above steps, the pressure distortion signal at the pipeline outlet can be obtained, and the pressure distortion signal can be input into the target hybrid reconstruction model for forward inference, and the reconstructed pipeline source end pressure signal can be output in real time.
[0121] According to an embodiment of the present invention, a deep learning-based method for reconstructing pulsating pressure signals in a semi-infinite pipe includes: acquiring a source signal in a semi-infinite pipe experimental system using a high-precision reference sensor installed at the source end of the pipe, and acquiring a measured distorted signal at the outlet of the pipe; performing time-series alignment and frequency domain feature extraction on the source signal and the distorted signal to obtain a first dataset; constructing an initial hybrid reconstruction model, wherein the initial hybrid reconstruction model includes a CNN module, a Transformer module, and a decoding and reconstruction module; training the initial hybrid reconstruction model based on the first dataset and a physical constraint loss function to obtain a target hybrid reconstruction model; acquiring the pressure distortion signal at the outlet of the pipe, and inputting the pressure distortion signal into the target hybrid reconstruction model for forward inference, and outputting the reconstructed pressure signal at the source end of the pipe in real time. Therefore, the present invention can use the target hybrid reconstruction model to perform forward inference on the acquired pressure distortion signal at the outlet of the pipe and output the reconstructed pressure signal at the source end of the pipe in real time, thereby overcoming the pressure signal amplitude attenuation and phase delay caused by the semi-infinite pipe and achieving high-precision signal reconstruction; at the same time, the target hybrid reconstruction model automatically performs end-to-end signal reconstruction in real time, eliminating the need for manual intervention, meeting the requirements of online monitoring, and reducing computational costs.
[0122] To achieve the above embodiments, such as Figure 3 As shown, this embodiment also provides a semi-infinite tube pulsating pressure signal reconstruction device 10 based on deep learning. The device includes an acquisition module 301, a data processing module 302, a construction module 303, a training module 304, and a reconstruction module 305.
[0123] The acquisition module 301 is used to acquire the source signal through a high-precision reference sensor installed at the source end of the pipe in the semi-infinite pipe experimental system, and to acquire the measured distortion signal at the outlet of the long pipe.
[0124] Data processing module 302 is used to perform time-series alignment and frequency domain feature extraction on the source signal and the distorted signal to obtain a first dataset;
[0125] Module 303 is used to construct an initial hybrid reconstruction model, wherein the initial hybrid reconstruction model includes a CNN module, a Transformer module, and a decoding and reconstruction module;
[0126] Training module 304 is used to train the initial hybrid reconstruction model based on the first dataset and the physical constraint loss function to obtain the target hybrid reconstruction model;
[0127] The reconstruction module 305 is used to acquire the pressure distortion signal at the pipeline outlet, input the pressure distortion signal into the target hybrid reconstruction model for forward inference, and output the reconstructed pipeline source end pressure signal in real time.
[0128] In one embodiment of the present invention, the training module 304 is specifically used for:
[0129] Preprocess the first dataset to obtain the second dataset;
[0130] Based on the second dataset, we obtain the training dataset and test dataset for training the initial hybrid reconstruction model;
[0131] The initial hybrid reconstruction model is trained using the training dataset and the physical constraint loss function to obtain the trained hybrid reconstruction model;
[0132] The trained hybrid reconstruction model was tested and optimized using the test dataset to obtain the target hybrid reconstruction model.
[0133] Furthermore, the aforementioned training module 304 is also used for:
[0134] The distorted signal in the training dataset is input into the initial hybrid reconstruction model to obtain the reconstruction pressure signal;
[0135] The reconstructed pressure signal and the source signal are input into the physical constraint loss function to obtain the corresponding loss value;
[0136] The network parameters in the initial hybrid reconstruction model are updated using the loss value until the initial hybrid reconstruction model converges, or until the number of iterations of the network parameters reaches a preset number, to obtain the trained hybrid reconstruction model.
[0137] Furthermore, the aforementioned initial hybrid reconstruction model includes a CNN module, a Transformer module, and a decoding and reconstruction module; the aforementioned training module 304 is also used for:
[0138] The local temporal features of the distorted signals in the training dataset are extracted using a CNN module, and a high-dimensional feature sequence is output.
[0139] The Transformer module captures the global long-term dependencies of high-dimensional feature sequences and outputs a reconstructed feature vector sequence.
[0140] The feature vector sequence is mapped back to a time-domain signal of the same length as the source signal by the decoding and reconstruction module, thus obtaining the reconstructed pressure signal.
[0141] Furthermore, the aforementioned CNN module consists of multiple sequentially connected feature extraction modules, and each feature extraction module includes a one-dimensional convolutional layer, an activation function layer, and a pooling layer; the aforementioned training module 304 is also used for:
[0142] By using multiple learnable convolutional kernels in a one-dimensional convolutional layer and employing a sliding window approach, the input temporal signal is convolved to obtain the local temporal features of the signal.
[0143] The local time-domain features of the signal are nonlinearly activated by an activation function layer to obtain the activated local time-domain features.
[0144] The activated local temporal features are downsampled using a pooling layer to obtain the downsampled local temporal features.
[0145] By stacking multiple feature extraction modules, the input distorted signal is subjected to layer-by-layer dimensionality reduction and feature abstraction, and the downsampled local temporal features obtained by the final feature extraction module are determined as a high-dimensional feature sequence.
[0146] Furthermore, the aforementioned Transformer module includes a Transformer encoder consisting of N sequentially stacked encoding layers, wherein the encoding layers include a multi-head self-attention module and a feedforward neural network module; the aforementioned training module 304 is also used for:
[0147] By adding positional encoding to each feature vector in the high-dimensional feature sequence, a new sequence containing temporal positional information is obtained;
[0148] The new sequence is input into the Transformer encoder, and the multi-head self-attention module performs self-attention calculation on the new sequence to obtain the feature vector at each position.
[0149] The feedforward neural network module performs nonlinear transformation and mapping on the feature vector at each position to obtain the temporal reconstruction vector sequence.
[0150] A Transformer encoder consisting of N sequentially stacked coding layers is used, and the temporal reconstruction vector sequence output by the Transformer encoder is determined as the reconstructed feature vector sequence.
[0151] Furthermore, the aforementioned physical constraint loss function includes:
[0152]
[0153] in, To reconstruct the absolute value error between the pressure signal and the source signal, For the transfer function constraint loss, L wave The loss is constrained by the wave equation. and The weights are for each constraint.
[0154] According to an embodiment of the present invention, a deep learning-based device for reconstructing pulsating pressure signals in a semi-infinite pipe can acquire source signals and measured distorted signals at the pipe outlet in a semi-infinite pipe experimental system. The device performs time-series alignment and frequency domain feature extraction on the source and distorted signals to obtain a first dataset. An initial hybrid reconstruction model is constructed, comprising a CNN module, a Transformer module, and a decoding and reconstruction module. The initial hybrid reconstruction model is trained based on the first dataset and a physical constraint loss function to obtain a target hybrid reconstruction model. The pressure distorted signal at the pipe outlet is acquired and input into the target hybrid reconstruction model for forward inference, outputting the reconstructed pipe source-end pressure signal in real time. Therefore, the present invention can use the target hybrid reconstruction model to perform forward inference on the acquired pressure distorted signal at the pipe outlet and output the reconstructed pipe source-end pressure signal in real time, thereby overcoming the pressure signal amplitude attenuation and phase delay caused by the semi-infinite pipe and achieving high-precision signal reconstruction. Simultaneously, the target hybrid reconstruction model automatically performs end-to-end signal reconstruction in real time, eliminating the need for manual intervention, meeting the requirements of online monitoring, and reducing computational costs.
[0155] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.
[0156] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.
Claims
1. A method for reconstructing pulsating pressure signals in a semi-infinite tube based on deep learning, characterized in that, include: In the semi-infinite tube experimental system, the source signal is acquired by a high-precision reference sensor installed at the source end of the tube, and the distortion signal is measured at the outlet of the tube. The source signal and the distorted signal are time-aligned and frequency domain features are extracted to obtain the first dataset. Construct an initial hybrid reconstruction model, wherein the initial hybrid reconstruction model includes a CNN module, a Transformer module, and a decoding and reconstruction module; The initial hybrid reconstruction model is trained based on the first dataset and the physical constraint loss function to obtain the target hybrid reconstruction model. The physical constraint loss function includes the transfer function constraint loss and the wave equation constraint loss. The pressure distortion signal at the pipeline outlet is acquired, and the pressure distortion signal is input into the target hybrid reconstruction model for forward inference, and the reconstructed pipeline source end pressure signal is output in real time. The step of training the initial hybrid reconstruction model based on the first dataset and the physical constraint loss function to obtain the target hybrid reconstruction model includes: The first dataset is preprocessed to obtain the second dataset; Based on the second dataset, a training dataset and a test dataset are obtained for training the initial hybrid reconstruction model; The distorted signal in the training dataset is input into the initial hybrid reconstruction model to obtain the reconstruction pressure signal; The reconstructed pressure signal and the source signal are input into the physical constraint loss function to obtain the corresponding loss value; The network parameters in the initial hybrid reconstruction model are updated using the loss value until the initial hybrid reconstruction model converges. The trained hybrid reconstruction model is tested and optimized using the test dataset to obtain the target hybrid reconstruction model.
2. The method according to claim 1, characterized in that, The step of inputting the distortion signal from the training dataset into the initial hybrid reconstruction model to obtain the reconstruction pressure signal includes: The CNN module extracts local temporal features of the distorted signals in the training dataset and outputs a high-dimensional feature sequence. The Transformer module captures the global long-term dependencies of the high-dimensional feature sequence and outputs a reconstructed feature vector sequence. The feature vector sequence is mapped back to a time-domain signal of the same length as the source signal by the decoding and reconstruction module to obtain the reconstructed pressure signal.
3. The method according to claim 2, characterized in that, The CNN module consists of multiple sequentially connected feature extraction modules, and each feature extraction module includes a one-dimensional convolutional layer, an activation function layer, and a pooling layer. The step of extracting local temporal features of the distorted signal in the training dataset through the CNN module and outputting a high-dimensional feature sequence includes: By using multiple learnable convolutional kernels in the one-dimensional convolutional layer, the input temporal signal is convolved using a sliding window method to obtain the local temporal features of the signal. The local time-domain features of the signal are nonlinearly activated by the activation function layer to obtain the activated local time-domain features; The activated local temporal features are downsampled by the pooling layer to obtain the downsampled local temporal features. By stacking multiple feature extraction modules, the input distorted signal is subjected to layer-by-layer dimensionality reduction and feature abstraction, and the downsampled local temporal features obtained by the final feature extraction module are determined as a high-dimensional feature sequence.
4. The method according to claim 2, characterized in that, The Transformer module includes a Transformer encoder consisting of N sequentially stacked encoding layers, wherein each encoding layer includes a multi-head self-attention module and a feedforward neural network module; the step of capturing the global long-term dependencies of the high-dimensional feature sequence through the Transformer module and outputting a reconstructed feature vector sequence includes: Add positional encoding to each feature vector in the high-dimensional feature sequence to obtain a new sequence containing temporal positional information; The new sequence is input into the Transformer encoder, and the multi-head self-attention module performs self-attention calculation on the new sequence to obtain the feature vector at each position. The feedforward neural network module performs nonlinear transformation and mapping on the feature vector at each position to obtain a temporal reconstruction vector sequence. A Transformer encoder consisting of N sequentially stacked coding layers is used, and the temporal reconstruction vector sequence output by the Transformer encoder is determined as the reconstructed feature vector sequence.
5. The method according to claim 3, characterized in that, The physical constraint loss function includes: in, The absolute error between the reconstructed pressure signal and the source signal. For the transfer function constraint loss, L wave The loss is constrained by the wave equation. and The weights for each constraint.
6. A device for reconstructing pulsating pressure signals in a semi-infinite tube based on deep learning, characterized in that, include: The acquisition module is used to acquire the source signal in the semi-infinite tube experimental system by means of a high-precision reference sensor installed at the source end of the tube, and to acquire the measured distortion signal at the outlet of the tube. The data processing module is used to perform time-series alignment and frequency domain feature extraction on the source signal and the distorted signal to obtain the first dataset; A building module is used to build an initial hybrid reconstruction model, wherein the initial hybrid reconstruction model includes a CNN module, a Transformer module, and a decoding and reconstruction module; The training module is used to train the initial hybrid reconstruction model based on the first dataset and the physical constraint loss function to obtain the target hybrid reconstruction model. The physical constraint loss function includes the transfer function constraint loss and the wave equation constraint loss. The reconstruction module is used to acquire the pressure distortion signal at the pipeline outlet, input the pressure distortion signal into the target hybrid reconstruction model for forward inference, and output the reconstructed pipeline source end pressure signal in real time. The training module is specifically used for: The first dataset is preprocessed to obtain the second dataset; Based on the second dataset, a training dataset and a test dataset are obtained for training the initial hybrid reconstruction model; The distorted signal in the training dataset is input into the initial hybrid reconstruction model to obtain the reconstruction pressure signal; The reconstructed pressure signal and the source signal are input into the physical constraint loss function to obtain the corresponding loss value; The network parameters in the initial hybrid reconstruction model are updated using the loss value until the initial hybrid reconstruction model converges. The trained hybrid reconstruction model is tested and optimized using the test dataset to obtain the target hybrid reconstruction model.
7. The apparatus according to claim 6, characterized in that, The training module is also used for: The CNN module extracts local temporal features of the distorted signals in the training dataset and outputs a high-dimensional feature sequence. The Transformer module captures the global long-term dependencies of the high-dimensional feature sequence and outputs a reconstructed feature vector sequence. The feature vector sequence is mapped back to a time-domain signal of the same length as the source signal by the decoding and reconstruction module to obtain the reconstructed pressure signal.
8. The apparatus according to claim 7, characterized in that, The CNN module consists of multiple sequentially connected feature extraction modules, and each feature extraction module includes a one-dimensional convolutional layer, an activation function layer, and a pooling layer; the training module is further used for: By using multiple learnable convolutional kernels in the one-dimensional convolutional layer, the input temporal signal is convolved using a sliding window method to obtain the local temporal features of the signal. The local time-domain features of the signal are nonlinearly activated by the activation function layer to obtain the activated local time-domain features; The activated local temporal features are downsampled by the pooling layer to obtain the downsampled local temporal features. By stacking multiple feature extraction modules, the input distorted signal is subjected to layer-by-layer dimensionality reduction and feature abstraction, and the downsampled local temporal features obtained by the final feature extraction module are determined as a high-dimensional feature sequence.
9. The apparatus according to claim 7, characterized in that, The Transformer module includes a Transformer encoder consisting of N sequentially stacked encoding layers, wherein the encoding layers include a multi-head self-attention module and a feedforward neural network module; the training module is further used for: Add positional encoding to each feature vector in the high-dimensional feature sequence to obtain a new sequence containing temporal positional information; The new sequence is input into the Transformer encoder, and the multi-head self-attention module performs self-attention calculation on the new sequence to obtain the feature vector at each position. The feedforward neural network module performs nonlinear transformation and mapping on the feature vector at each position to obtain a temporal reconstruction vector sequence. A Transformer encoder consisting of N sequentially stacked coding layers is used, and the temporal reconstruction vector sequence output by the Transformer encoder is determined as the reconstructed feature vector sequence.
10. An electronic device, characterized in that, include: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-5.
11. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-5.