A wireless pressure wave code communication denoising method based on a CNN-LSTM hybrid model

Abstract

The present application relates to the field of artificial intelligence and oilfield separate layer water injection cross technology, specifically to a wireless pressure wave code communication denoising method based on CNN-LSTM hybrid model. The method comprises: constructing a composite simulation signal set containing Gaussian noise, impulse noise, periodic noise and complex frequency band noise; designing a CNN-LSTM hybrid model, the encoder extracts local features of the signal through convolution layer and pooling layer, the bidirectional LSTM layer captures the time sequence dependence, and the decoder reconstructs the signal through upsampling and convolution layer; the model is trained using mean square error loss function and Adam optimizer; when applied, the collected signal is normalized and edge filled, then input into the trained model for denoising, and the output signal is decoded to get the code value. The method can effectively learn the space-time features of the signal, adaptively suppress the complex noise from multiple sources downhole, significantly improve the signal-to-noise ratio and decoding accuracy of the signal, and enhance the reliability of wireless pressure wave communication.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence and oilfield stratified water injection technology, and more specifically to a wireless pressure wavecode communication noise reduction method based on a CNN-LSTM hybrid model. Background Technology

[0002] As oilfields enter the high water-cut development stage, the need for real-time monitoring and precise control of downhole stratified water injection technology is becoming increasingly urgent. Wireless pressure wave code communication technology, with its advantages of not requiring cable laying and being easy to maintain, has become the main solution for realizing information exchange between the well and the surface. This technology uses the fluid inside the wellbore as the transmission medium, transmitting key parameters such as water injection pressure, flow rate, and temperature through pressure waves, providing data support for the refined management of stratified water injection.

[0003] However, the complex noise environment downhole severely restricts the reliability and decoding accuracy of wireless pressure wave code communication. In practical applications, pressure wave signal transmission faces multi-source noise interference, mainly including sensor temperature drift, hydrodynamic noise, and mechanical vibration noise. Specifically, sensor temperature drift originates from the sensitivity drift of the pressure sensor caused by downhole temperature changes, resulting in signal amplitude distortion; hydrodynamic noise includes turbulence generated by high-speed water injection, Gaussian noise introduced by the bursting of bubbles generated by gas-liquid mixing flow, and pulse noise caused by water hammer effect; mechanical vibration noise comes from harmonic noise generated by the periodic vibration of the pump unit, as well as broadband complex noise generated by other equipment. These noises collectively lead to a significant reduction in the signal-to-noise ratio, increasing the decoding error rate at the receiver, directly affecting the accurate adjustment of the water injection scheme, and reducing the efficiency of oilfield development.

[0004] Currently, traditional model-driven denoising methods are mainly used in wireless pressure code communication. However, these methods suffer from three common problems: first, they rely on prior assumptions and have poor adaptability; second, they lack the ability to collaboratively suppress multi-source noise; and third, it is difficult to balance signal fidelity and noise suppression. Although deep learning technology has made progress in acoustics, seismic signal denoising, and other fields, it faces unique challenges in pressure code communication denoising applications. For example, while single deep learning models such as Convolutional Neural Networks (CNNs) are good at extracting local signal features, or Long Short-Term Memory Networks (LSTMs) can capture temporal dependencies, neither can simultaneously take into account both spatial and temporal features, resulting in limited denoising performance in complex multi-source noise environments.

[0005] Therefore, existing technologies cannot effectively solve the noise reduction problem of wireless pressure wavecode signals in multi-source dynamic noise environments downhole, resulting in low signal-to-noise ratio and high decoding error rate, making it difficult to meet the communication reliability requirements of layered water injection technology. Summary of the Invention

[0006] In view of this, the present invention provides a wireless pressure wave code communication denoising method based on a CNN-LSTM hybrid model, which aims to solve the technical problem that existing denoising methods have poor adaptability and limited signal denoising effect in complex multi-source noise environments downhole, resulting in high bit error rate in wireless pressure wave code decoding, thereby improving the signal-to-noise ratio and decoding accuracy, and enhancing the reliability of downhole wireless communication.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A method for denoising wireless pressure wavecode communication based on a CNN-LSTM hybrid model, characterized by the following steps:

[0009] S1: Generate a set of simulated wireless pressure wavecode signals containing multi-source noise, including Gaussian noise, impulse noise, periodic noise, and complex frequency band noise;

[0010] S2: Construct a CNN-LSTM hybrid model, which sequentially includes an encoder module, a temporal modeling module, and a decoder module. The encoder module consists of convolutional layers and max pooling layers to extract local features of the signal. The temporal modeling module consists of bidirectional LSTM layers to capture temporal dependencies. The decoder module consists of convolutional layers and upsampling layers to reconstruct the signal.

[0011] S3: The CNN-LSTM hybrid model is trained using the mean squared error loss function and the adaptive moment estimation optimizer;

[0012] S4: Perform dynamic threshold decoding on the noise-reduced output signal to parse the discrete code values ​​of the wireless pressure wave code.

[0013] For example, the output of the encoder module is flattened and then connected to the input of the bidirectional LSTM layer of the timing modeling module, and the output of the timing modeling module is reshaped and then connected to the input of the convolutional layer of the decoder module; the encoder module achieves feature dimensionality reduction through a max pooling layer, and the decoder module restores the original dimension of the signal through an upsampling layer, ensuring that the timing lengths of the input and output signals are consistent.

[0014] In one specific implementation, in step S1, the mean of the Gaussian noise is 0, and the standard deviation is randomly generated within the range of 0.2 to 0.4.

[0015] In a specific implementation scheme, in step S1, the pulse noise is generated by a probability triggering mechanism, the pulse occurrence probability p is randomly set, and a pulse with a random amplitude is randomly selected at a position in the signal sequence with probability p.

[0016] For example, the random amplitude ranges from 0.5 to 2.0 times the maximum amplitude of the clean wireless pressure wavecode signal, and the pulse polarity is randomly selected to match the pulse interference intensity characteristics caused by downhole hydrodynamic noise.

[0017] For example, in step S1, when generating the impulse noise, the probability p of the impulse occurrence ranges from 0.01 to 0.05.

[0018] In one specific implementation scheme, in step S1, the method for generating the periodic noise is as follows: superimposing a fundamental frequency sine wave, a second harmonic, and a third harmonic, with the amplitude of each harmonic decreasing proportionally.

[0019] For example, the amplitude of each harmonic decreases in a ratio of "the amplitude of the second harmonic is 0.5 times that of the fundamental sine wave, and the amplitude of the third harmonic is 0.25 times that of the fundamental sine wave", with the fundamental frequency ranging from 5 to 50 Hz, matching the frequency range of periodic vibrations generated by the operation of the pump group.

[0020] In a specific implementation scheme, in step S1, the method for generating the complex frequency band noise is as follows: filtering and shaping the Gaussian white noise, and then linearly superimposing the filtered signals.

[0021] In one specific implementation scheme, in step S2, the encoder module has the following specific structure:

[0022] Level 1: Convolutional layers with 64 kernels and a size of 3, and max pooling layers with a pooling window of 2;

[0023] Second level: Convolutional layer with 128 kernels and a size of 3.

[0024] In a specific implementation scheme, in step S2, the specific structure of the timing modeling module is: two bidirectional LSTM layers, each layer having 64 bidirectional LSTM units, followed by a Dropout layer with a 20% deactivation rate.

[0025] In one specific implementation scheme, in step S2, the specific structure of the decoder module is as follows:

[0026] Level 1: Convolutional layers with 128 kernels and a size of 3, and upsampling layers with an upsampling factor of 2;

[0027] Second level: Convolutional layer with 64 kernels and a size of 3.

[0028] In a specific implementation scheme, in step S3, when training the CNN-LSTM hybrid model, the training set and validation set signals of the input model are preprocessed, including: performing min-max normalization, and padding the beginning and end of each noisy signal and its corresponding clean label signal, padding with M sampling points before the start end and after the end end of the signal.

[0029] For example, the value of M is 32, which is specifically determined based on the signal sampling rate and the convolution kernel size. The padding value is the linear extrapolation result of the sampling values ​​at the beginning and end of the signal to avoid signal distortion caused by edge convolution operations.

[0030] In one specific implementation scheme, the dynamic threshold decoding method in step S4 is as follows:

[0031] The average of the first 500 largest sampling points of the signal is used as the high voltage threshold, and the average of the first 500 smallest sampling points is used as the low voltage threshold.

[0032] The arithmetic mean of the high-pressure threshold and the low-pressure threshold is taken as the decision threshold;

[0033] Compare the signal sample values ​​with the decision threshold point by point to classify the high-voltage state or the low-voltage state and count the duration.

[0034] Convert the duration into discrete code values ​​according to the time encoding rules;

[0035] Specifically, when a state flip is detected, if the duration is less than 0.5 times the basic unit time T0, it is determined to be an invalid code segment and the previous state is maintained.

[0036] For example, the basic unit time T0 is the symbol period of the wireless pressure wave code, which is 0.1 seconds, and is specifically set according to the downhole communication rate requirements; the selection basis of the "first 500 maximum / minimum sampling points" is 10% of the total number of sampling points of a single communication signal to ensure that the threshold calculation covers the main amplitude range of the signal.

[0037] Compared with existing technologies, the wireless pressure code communication denoising method based on a CNN-LSTM hybrid model described in this invention is used for wireless pressure code signal processing in downhole layered water injection systems. By constructing a simulated signal set containing multi-source noise and combining the local feature extraction capabilities of CNN with the temporal modeling advantages of LSTM, the method achieves joint learning of signal spatiotemporal features and adaptive noise suppression, effectively improving the signal-to-noise ratio and decoding accuracy of signals in complex noise environments. It has the following beneficial effects:

[0038] This invention overcomes the limitations of a single model in feature extraction by designing a hybrid model architecture of encoder-temporal modeling-decoder, utilizing convolutional neural networks to effectively extract local spatial features of signals, and capturing long-term temporal dependencies in signals through bidirectional long short-term memory networks.

[0039] This invention uses a composite simulation signal set including Gaussian noise, impulse noise, periodic noise, and noise in complex frequency bands for model training, which enhances the model's adaptability and generalization ability to the actual noise environment downhole.

[0040] This invention trains the model using a mean squared error loss function and an adaptive moment estimation optimizer, and combines this with a dynamic threshold decoding strategy, which significantly improves the quality of the denoised signal and the accuracy of the decoding decision.

[0041] This invention can effectively suppress multi-source mixed noise and improve signal waveform fidelity, thereby significantly reducing the communication error rate and providing solid technical support for the reliability of wireless pressure wave code communication in oilfields. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0043] Figure 1 This is an overall flowchart of a wireless pressure wavecode communication denoising method based on a CNN-LSTM hybrid model as described in this invention.

[0044] Figure 2 This is a schematic diagram of a clean signal generated during simulation for model training in an embodiment of a wireless pressure wavecode communication denoising method based on a CNN-LSTM hybrid model as described in this invention.

[0045] Figure 3 This is a schematic diagram of a noisy signal superimposed with noise used for model training in an embodiment of a wireless pressure wavecode communication denoising method based on a CNN-LSTM hybrid model as described in this invention.

[0046] Figure 4 This is a schematic diagram of the structure of the CNN-LSTM hybrid model in an embodiment of the wireless pressure wavecode communication denoising method based on the CNN-LSTM hybrid model described in this invention.

[0047] Figure 5This is a training loss curve of the CNN-LSTM hybrid model in an embodiment of the wireless pressure wavecode communication denoising method based on the CNN-LSTM hybrid model described in this invention.

[0048] Figure 6 This is a schematic diagram of a test signal in an embodiment of a wireless pressure wavecode communication denoising method based on a CNN-LSTM hybrid model as described in this invention.

[0049] Figure 7 This is a schematic diagram of the denoised signal output by the trained model in an embodiment of the wireless pressure wavecode communication denoising method based on a CNN-LSTM hybrid model described in this invention.

[0050] Figure 8 This is a comparison diagram of the original code value and the decoded value of the denoised signal after decoding in an embodiment of a wireless pressure wavecode communication denoising method based on a CNN-LSTM hybrid model according to the present invention. Detailed Implementation

[0051] The technical solutions in the embodiments of the present invention will be clearly and completely described below. 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 are within the scope of protection of the present invention.

[0052] like Figure 1 As shown, during the model training phase, a noisy simulation signal set with a dimension of 1×5000 and its corresponding clean label signal are generated by synthesizing randomly generated clean wireless pressure wavecode signals and multi-source noise. The generated dataset is then processed by edge padding with M sampling points at the beginning and end (e.g., 100 sampling points at the beginning and end in the embodiment) and normalized to obtain a dataset with a dimension of 1×5200, which is input into the model for model training. The model consists of an encoder, a temporal modeling module, and a decoder. During the model application phase, the collected 1×5000 signal to be denoised is processed by edge padding and normalization preprocessing consistent with the training phase to obtain an expanded signal with a dimension of 1×5200, which is input into the trained model. The model encodes the signal, extracts temporal features, and decodes and reconstructs the signal, outputting the denoised expanded signal. The sampling points pruned at the beginning and end are removed to restore the original length of the signal, and the signal is back-calculated from the normalized interval back to the actual pressure value range, finally obtaining a 1×5000 denoised signal. The present invention discloses a wireless pressure wavecode communication denoising method based on a CNN-LSTM hybrid model, which specifically includes the following steps:

[0053] S1: Randomly generate 500 clean pressure wavecode signals, and add four types of randomly generated noise to the clean pressure wavecode signals to construct a multi-source composite noise simulation signal dataset corresponding to each clean signal.

[0054] S2: Construct a CNN-LSTM hybrid model, which includes, in sequence: an encoder module consisting of convolutional layers and max pooling layers; a temporal modeling module consisting of bidirectional LSTM layers; and a decoder module consisting of convolutional layers and upsampling layers.

[0055] S3: The dataset is preprocessed and used for model training. The CNN-LSTM hybrid model is trained using the MSE loss function (mean squared error loss function) and the Adam optimizer (adaptive moment estimator optimizer).

[0056] S4: After preprocessing, the collected signal is input into the trained model to output the denoised pressure wave code signal. The denoised signal can be decoded into code value through dynamic threshold decoding.

[0057] Furthermore, S1 specifically includes:

[0058] Clean signals are combined into code segments according to time-coded rules: the wake-up code has a fixed duration of 5 seconds. High-voltage signal; stratum code randomly generates stratum value Convert to duration Low-voltage signal; instruction code randomly generates instruction value. Convert to duration High voltage signal; data code tens digit value (duration) (low voltage signal), unit code value (duration) (high voltage signal), tenth bit code value (duration) The low-voltage signal is generated independently and randomly; the check code is generated based on the sum and parity of the layer code + instruction code + data code. High voltage (odd number) or 2 High voltage (even); the end code has a fixed duration of 1 second. Low-voltage signal. Time-domain signal generated at a sampling frequency of 1Hz.

[0059] Random generation of four types of noise:

[0060] 1. Gaussian noise: Generates a Gaussian distributed sequence with a mean of 0, whose standard deviation is uniformly and randomly sampled within the interval [0.2, 0.4].

[0061] 2. Impulse noise: Set the probability of pulse occurrence. pRandomly generated within the range [0.01, 0.05], and randomly selected from the signal sequence. p × signal length at each position, superimposed with pulses randomly generated within the range of [0.1, 0.5] MPa;

[0062] 3. Periodic noise: fundamental frequency f Randomly generated and superimposed with second harmonics (2 f ) and third harmonic (3 f This constitutes periodic noise, with harmonic amplitudes decreasing by 0.6 and 0.3 respectively from the fundamental frequency amplitude.

[0063] 4. Complex frequency band noise: Gaussian white noise is spectrally shaped by passing it through three different bandpass filters and then superimposed to generate noise with specific frequency domain characteristics.

[0064] Four types of noise are superimposed on each clean signal to generate 500 independent samples. Each sample contains a clean signal and the corresponding noisy signal. The samples are randomly divided into training set and validation set in an 8:2 ratio.

[0065] Furthermore, in S2, constructing the CNN-LSTM hybrid model specifically includes:

[0066] The input layer receives data in the following shape: n , L The noisy signal of ,1), where n The batch size is fixed at 32. L This is the signal length.

[0067] The encoder module consists of a first-level convolutional layer with 64 convolutional kernels of size 3, stride 1, ReLU activation, and an output dimension of (32, ...). L ,64); Max pooling layer: pooling window size 2, stride 2, output dimension (32, L / 2,64); Second-level convolutional layer: 128 convolutional kernels of size 3, ReLU activation, output dimension (32, L / 2,128).

[0068] The time series modeling module has two bidirectional LSTM layers, each with 64 LSTM units, returning a complete sequence with an output dimension of (32, L / 2,128), each bidirectional LSTM layer is followed by a Dropout layer (random deactivation layer) with a deactivation rate of 20%.

[0069] The decoder module consists of a first-level convolutional layer with 128 convolutional kernels of size 3, ReLU activation, and an output dimension of (32, ...). L / 2,128); Upsampling layer: upsampling factor 2, output dimension (32, 128); L,128); Second-level convolutional layer: 64 convolutional kernels of size 3, ReLU activation, output dimension (32, L The output layer consists of a single convolutional kernel of size 3, with linear activation, and an output dimension of (32, 64). L ,1).

[0070] Furthermore, in S3, the dataset preprocessing specifically includes:

[0071] Clean signals for training and validation sets Perform min-max normalization and save the scaler for the noisy signals on the training and validation sets. The minimum-maximum normalization expression for subsequent data denoising and preprocessing is as follows:

[0072]

[0073] In the formula, min (·) represents the minimum value function. max (·) represents the maximum value function.

[0074] Furthermore, in S3, before model training, in order to avoid the loss of signal boundary features that may be caused by convolution operation and to ensure data consistency between the training and inference stages, before inputting batch data into the model, edge filling processing is required for each noisy signal and its corresponding clean label signal in the training set. Specifically, expansion filling is performed before the signal start end and after the signal end end, with a preferred filling length of 100 sampling points on each side.

[0075] Furthermore, the model training uses MSE as the loss function, and the MSE calculation formula is as follows:

[0076]

[0077] In the formula, Indicates the model output, This indicates a clean signal label.

[0078] The model was trained using the Adam optimizer, with the learning rate set as the parameter. The training parameters are set to batch size 32 and training epochs 50.

[0079] Furthermore, in step S4, the specific steps for denoising the acquired signal include:

[0080] The raw wireless pressure code signal is normalized using a scaler saved during model training. The normalized signal is then input into the trained CNN-LSTM hybrid model, sequentially passing through an encoder convolutional layer to extract local features, a bidirectional LSTM layer to capture temporal dependencies, and a decoder upsampling layer to reconstruct the signal waveform, ultimately outputting a denoised signal. The dimension is the same as the input.

[0081] Furthermore, in step S4, before inputting the normalized signal into the trained CNN-LSTM hybrid model, a signal padding step is also included:

[0082] To prevent the model from experiencing missing receptive fields at the beginning and end of the signal due to convolution operations, the input signal needs to be preprocessed. Specifically, a symmetrical padding pattern is used, padding the normalized signal with 100 sampling points before the start and 100 after the end. The padded extended signal is then input into the model. After the model outputs the denoised extended signal, the first and last 100 padding points are truncated to obtain the final denoised signal with the same length as the original acquired signal, which is then used for subsequent dynamic threshold decoding.

[0083] Furthermore, in step S4, dynamic threshold decoding is implemented through the following sub-steps:

[0084] (4.1) Extraction The mean of the top 500 sampling points is calculated as follows: As a high-voltage threshold, extract The mean of the smallest 500 sample points is calculated as follows: The low-pressure threshold is calculated as the arithmetic mean of the high-pressure threshold and the low-pressure threshold. The threshold is used to determine whether the sampling point is in a high-pressure state or a low-pressure state.

[0085] (4.2) Point-by-point comparison and ,like Greater than Then it is marked as a high-voltage state; if Less than This is then marked as a low-pressure state. The duration of the continuous state is recorded. t Then, the duration of the code segment in this state is converted into the corresponding code value. The conversion expression is as follows:

[0086]

[0087] In the formula, x This represents the converted code value, and round(·) represents the rounding function. It represents the basic unit of time.

[0088] (4.3) When a state transition is detected, if the duration of the new state is less than 0.5 seconds... If the signal is not clear, it is considered noise interference, the current state flip is canceled, and the previous state is maintained; otherwise, it is considered a valid code segment, and the corresponding code value is output. Finally, the signal is analyzed in the order of wake-up code, layer code, instruction code, three-digit data code, check code, and end code.

[0089] This invention, through its innovative hybrid model architecture and dynamic decoding strategy, surpasses existing technologies in terms of complex noise suppression, communication reliability, and engineering practicality, providing key support for the large-scale application of wireless pressure wavecode communication technology in industrial scenarios such as oil fields.

[0090] Example: Figure 1 As shown, it includes the following steps:

[0091] S1: Based on the pressure wave transmission principle of a stratified water injection system, noise-free pressure wave code signals and multi-source noise signals are generated to construct noisy data. Based on the pressure wave transmission model derived from Bernoulli's equation, the original signal is generated by simulating changes in the opening degree of ground valves, such as... Figure 2 The signal shown is a clean signal with the code value "54177121". The horizontal axis represents time in seconds (s), and the vertical axis represents pressure in megapascals (MPa). The signal contains a wake-up code (5). The high voltage code corresponds to code value 5, the layer code corresponds to code value 4, the instruction code corresponds to code value 1, the data code corresponds to three-digit code value 771, the check code corresponds to code value 2, and the end code (1) The low voltage corresponds to the code value 1. Set to 120s, sampling frequency 1Hz, and subsequently supplement with low voltage values ​​to make the length of a single signal 5000 sampling points.

[0092] Gaussian noise is generated to simulate turbulence and gas-liquid two-phase flow interference, with a mean of 0 and a standard deviation of 0.2-0.4 MPa, randomly sampled. Impulse noise is generated to simulate water hammer effect, generated through a probability triggering mechanism, with pulse amplitude of 10%-30% of the maximum signal amplitude and an occurrence probability of 0.01-0.05. Periodic noise is generated to simulate pump vibration, with second and third harmonics superimposed on the fundamental frequency, and harmonic amplitudes decreasing at 0.6 times and 0.3 times the fundamental frequency, respectively. Complex frequency band noise is generated to simulate multi-device coupling interference, with Gaussian white noise superimposed after being filtered by three different bandpass filters.

[0093] like Figure 3 As shown, four types of noise are superimposed onto a clean signal with a code value of "54177121" to generate 500 independent samples. Each sample contains a clean signal and the corresponding noisy signal, and they are randomly divided into a training set and a validation set in an 8:2 ratio.

[0094] S2: Constructing a CNN-LSTM hybrid model: The encoder module contains 64 Conv1D layers (one-dimensional convolutional layers with ReLU activation) with 3×1 convolutional kernels, 2×1 max pooling layers, and 128 Conv1D layers (with ReLU activation) with 3×1 convolutional kernels.

[0095] The temporal modeling module uses a two-layer bidirectional LSTM (64 units per layer), with a Dropout layer following each bidirectional LSTM layer, resulting in a 20% inactivation rate.

[0096] The decoder module contains 128 Conv1D layers (ReLU activation) with 3×1 convolutional kernels, 2×1 upsampling layers, and 64 Conv1D layers (ReLU activation) with 3×1 convolutional kernels; the output layer is a single Conv1D layer (linear activation) with 3×1 convolutional kernels.

[0097] The overall structure of the CNN-LSTM hybrid model of the present invention is as follows: Figure 4 As shown, the encoder module first extracts local signal features through a first-level convolutional layer (64 convolutional kernels of size 3), and then performs feature dimensionality reduction through a max pooling layer (pooling window of 2), changing the output dimension from (32, L, 1) of the input to (32, L / 2, 64), where L is the length of the input signal. In this embodiment, L is set to 5200. Subsequently, a second-level convolutional layer (128 convolutional kernels of size 3) further extracts high-level features, with an output dimension of (32, L / 2, 128). The temporal modeling module consists of two bidirectional LSTM layers, each containing 64 LSTM units, returning a complete sequence with an output dimension maintained at (32, L / 2, 128). Each layer is followed by a Dropout layer with a 20% deactivation rate to enhance generalization ability. The decoder module first performs feature mapping through a convolutional layer (128 convolutional kernels of size 3), and then restores the sequence length to the original size through an upsampling layer (upsampling factor of 2), with an output dimension of (32, L, 128). Subsequently, the features are further refined through a second-level convolutional layer (64 convolutional kernels of size 3), and finally the signal is reconstructed through an output layer (1 convolutional kernel of size 3, linear activation), with the output dimension consistent with the input, which is (32, L, 1).

[0098] S3: Perform min-max normalization on the dataset, and pad each signal's beginning and end edges with 100 sampling points before training the model. Use MSE as the loss function and Adam optimizer, with a learning rate of [missing information]. Batch size 32, training epochs 50. The training loss curve of the CNN-LSTM hybrid model during training is shown below. Figure 5As shown in the figure, the horizontal axis represents the number of training epochs, and the vertical axis represents the mean squared error loss (MSE). As can be seen from the figure, both the training loss and validation loss decrease rapidly with increasing training epochs, eventually dropping below 0.001. During training, the model continuously learns data patterns. While the validation loss fluctuates, it does not continuously increase, indicating no significant overfitting. This demonstrates that the model has good generalization ability.

[0099] S4: Input the test signal into the trained model. The test signal is as follows: Figure 6 As shown, the test signal (SNR=1.98dB, RMSE=0.43MPa) is independent of the training and validation sets. Its noise combination is consistent with the training signal, but the noise parameters (such as Gaussian noise standard deviation, impulse occurrence probability, and periodic noise fundamental frequency) are randomly generated to simulate the dynamic changes of noise in the actual downhole environment. A symmetrical filling method is used to fill 100 sampling points at each of its beginning and end edges. This is input into the trained model, and the output is the denoised extended signal. The filled portions at the beginning and end of this denoised signal are truncated to obtain the final denoised signal (ΔSNR=16.11dB, RMSE=0.07MPa). Figure 7 As shown. With Figure 6 As can be seen from the comparison, the denoised signal waveform is significantly smoother, noise components are effectively suppressed, the boundaries between high-voltage and low-voltage code segments in the signal are clear, amplitude fluctuations are significantly reduced, and the signal waveform is highly consistent with the original clean signal, demonstrating the model's excellent noise suppression capability and signal fidelity in complex multi-source noise environments. The average of the top 500 largest sampling points of the denoised signal is calculated as the high-voltage threshold, and the average of the top 500 smallest sampling points of the denoised signal is calculated as the low-voltage threshold. The average of the high-voltage and low-voltage thresholds is taken as the decision threshold for the signal. The denoised signal is divided into high-voltage and low-voltage segments according to the decision threshold. Invalid code segments with a duration of less than 60 seconds are filtered out to extract the code value. The decoding result is as follows: Figure 8 As shown in the figure, the upper part of the figure displays the original code value of each code segment (5 3 1 8 5 2 1 1), and the lower part displays the decoded value obtained after dynamic threshold decoding of the model output (5 3 1 8 5 2 1 1). The comparison results show that the decoded value of all code segments is completely consistent with the original code value, and the symbol error rate (SER) is 0%, indicating that the method of the present invention achieves extremely high accuracy in both signal denoising and decoding decision stages, effectively improving the reliability of wireless pressure wave code communication.

[0100] Finally, the comparison of the CNN-LSTM hybrid model of the present invention with the single CNN model, LSTM model and traditional wavelet transform denoising method on 100 noisy signal samples independent of the training dataset for the three indicators of average ΔSNR, average RMSE and average SER is shown in Table 1.

[0101] Table 1. Comparison of average signal-to-noise ratio (SNR) improvement, average root mean square error (RMSE), and average symbol error rate (SER) for four noise reduction methods.

[0102]

[0103] The CNN-LSTM hybrid model proposed in this invention exhibits significant advantages in noise suppression performance, signal reconstruction accuracy, and digital signal decoding accuracy. Specifically, the CNN-LSTM hybrid model proposed in this invention achieves an average signal-to-noise ratio improvement of 16.24 dB, an average root mean square error of 0.09 MPa, and an average bit error rate of 0.14%, all of which are the best among all models.

[0104] In summary, the CNN-LSTM hybrid model proposed in this invention outperforms other single neural network models and wavelet transform denoising methods, demonstrating excellent denoising and generalization capabilities.

[0105] The various embodiments described in this specification are presented in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for denoising wireless pressure wavecode communication based on a CNN-LSTM hybrid model, characterized in that, Includes the following steps: S1: Generate a set of simulated wireless pressure wavecode signals containing multi-source noise, including Gaussian noise, impulse noise, periodic noise, and complex frequency band noise; S2: Construct a CNN-LSTM hybrid model, which includes an encoder module, a temporal modeling module, and a decoder module in sequence; the encoder module consists of convolutional layers and max pooling layers, used to extract local features of the signal; The timing modeling module consists of bidirectional LSTM layers to capture timing dependencies; the decoder module consists of convolutional layers and upsampling layers to reconstruct the signal. S3: The CNN-LSTM hybrid model is trained using the mean squared error loss function and the adaptive moment estimation optimizer; S4: Perform dynamic threshold decoding on the noise-reduced output signal to parse the discrete code values ​​of the wireless pressure wave code; the specific method of dynamic threshold decoding is as follows: The average of the first 500 largest sampling points of the signal is used as the high voltage threshold, and the average of the first 500 smallest sampling points is used as the low voltage threshold. The arithmetic mean of the high-pressure threshold and the low-pressure threshold is taken as the decision threshold; Compare the signal sample values ​​with the decision threshold point by point to classify the high-voltage state or the low-voltage state and count the duration. Convert the duration into discrete code values ​​according to the time encoding rules; Specifically, when a state flip is detected, if the duration is less than 0.5 times the basic unit time T0, it is determined to be an invalid code segment and the previous state is maintained.

2. The wireless pressure code communication denoising method based on a CNN-LSTM hybrid model according to claim 1, characterized in that, In step S1, the mean of the Gaussian noise is 0, and the standard deviation is randomly generated within the range of 0.2 to 0.

4.

3. The wireless pressure code communication denoising method based on a CNN-LSTM hybrid model according to claim 1, characterized in that, In step S1, the pulse noise is generated by a probability triggering mechanism. The probability of pulse occurrence is randomly set to p, and a pulse with a random amplitude is randomly selected at a position in the signal sequence with probability p.

4. The wireless pressure code communication denoising method based on a CNN-LSTM hybrid model according to claim 1, characterized in that, In step S1, the method for generating the periodic noise is as follows: superimposing a fundamental frequency sine wave, a second harmonic, and a third harmonic, with the amplitude of each harmonic decreasing proportionally.

5. The wireless pressure code communication denoising method based on a CNN-LSTM hybrid model according to claim 1, characterized in that, In step S1, the method for generating the complex frequency band noise is as follows: Gaussian white noise is filtered and shaped, and the filtered signals are linearly superimposed.

6. The wireless pressure wavecode communication denoising method based on a CNN-LSTM hybrid model according to claim 1, characterized in that, In step S2, the specific structure of the encoder module is as follows: Level 1: Convolutional layers with 64 kernels and a size of 3, and max pooling layers with a pooling window of 2; Second level: Convolutional layer with 128 kernels and a size of 3.

7. The wireless pressure wavecode communication denoising method based on a CNN-LSTM hybrid model according to claim 1, characterized in that, In step S2, the specific structure of the timing modeling module is as follows: two bidirectional LSTM layers, each with 64 bidirectional LSTM units, followed by a Dropout layer with a 20% deactivation rate.

8. A method for denoising wireless pressure wavecode communication based on a CNN-LSTM hybrid model according to claim 1, characterized in that, In step S2, the specific structure of the decoder module is as follows: Level 1: Convolutional layers with 128 kernels and a size of 3, and upsampling layers with an upsampling factor of 2; Second level: Convolutional layer with 64 kernels and a size of 3.

9. A method for denoising wireless pressure wavecode communication based on a CNN-LSTM hybrid model according to claim 1, characterized in that, In step S3, when training the CNN-LSTM hybrid model, the training set and validation set signals of the input model are preprocessed, including: performing min-max normalization, and padding the beginning and end of each noisy signal and its corresponding clean label signal, and padding M sampling points before the start end and after the end end of the signal.

Images