A DAS flow rate identification method based on one-dimensional timing feature extraction guided by physical prior

Abstract

The application provides a DAS flow velocity identification method based on one-dimensional time sequence feature extraction guided by physical priori, and belongs to the technical field of fluid acoustics, and comprises the following steps: collecting DAS original time sequence acoustic signals related to fluid flow, and obtaining DAS original time sequence data distributed according to spatial positions and time; performing flow velocity sensitive input sample construction on the DAS original time sequence data based on flow related physical priori, so as to obtain flow velocity sensitive one-dimensional time sequence input samples used for flow velocity identification; inputting the flow velocity sensitive one-dimensional time sequence input samples into a time sequence feature identification model, so as to obtain time sequence features related to fluid flow velocity; inputting the time sequence features into a regression output layer, so as to obtain fluid flow velocity identification results; and outputting the flow velocity identification results, so as to represent the actual flow velocity state of the measured pipeline or wellbore fluid. The application further discloses a system and a storage medium for executing the method, and is suitable for a wellbore or pipeline flow velocity identification scene.

Description

Technical Field

[0001] This invention belongs to the field of fluid acoustics technology, specifically relating to a DAS flow velocity identification method and system based on physical prior guidance for one-dimensional temporal feature extraction. It is applicable to experimental or field conditions and outputs fluid flow velocity based on the original temporal acoustic signal of DAS related to fluid flow. Technical Background

[0002] Distributed Acoustic Sensing (DAS) is a distributed measurement technology based on fiber optic sensing. It utilizes optical fiber as a continuous sensing medium to acquire acoustic response signals related to external disturbances along the fiber axis and output raw time-series data that varies with spatial location and time. Compared to traditional point sensors, DAS has advantages such as long deployment distance, continuous axial monitoring, resistance to electromagnetic interference, and ease of long-term online data acquisition. Therefore, it has broad application prospects in the fields of condition monitoring, anomaly diagnosis, and flow process characterization in flow channels such as wells and pipelines. In well flow and surface pipeline transportation processes, fluid velocity is an important parameter reflecting changes in operating conditions and flow state. In existing technologies, Flow characterization is typically achieved by performing frequency domain transformation, energy statistics, or empirical feature extraction on DAS signals to construct energy indices and spectral amplitude indices. However, the original DAS time series often contains local responses related to flow disturbances, cross-channel consistent common-mode components introduced by pump vibration / structural vibration, and background noise. Common-mode components tend to dominate in energy or spectral amplitude indices, making the features insensitive to flow velocity. At the same time, the dominant frequency band related to flow velocity will drift with changes in the scale of disturbance characteristics and flow velocity level. Fixed frequency band or artificial frequency band selection schemes are prone to failure under different pipe diameter / structural scales, different noise levels, and different flow velocity ranges, resulting in insufficient consistency across operating conditions. Therefore, there is an urgent need for a flow velocity identification method that takes the original time series of DAS multi-channel as input, determines the flow velocity-related target frequency band based on the flow-related physical prior, performs common-mode suppression and differential processing of adjacent channels in sequence within the target frequency band, constructs flow velocity-sensitive input samples, and combines the time series feature identification model to realize continuous numerical identification of flow velocity, thereby improving the stability and applicability of the identification results under different noise and operating conditions. Summary of the Invention

[0003] This invention addresses the problems of strong common-mode noise, weak local spatial disturbance, and shift of velocity-related frequency bands with changing operating conditions in DAS multi-channel raw time-series signals in well or pipeline scenarios. It proposes a method for constructing velocity-sensitive input samples based on flow-related physical priors. By limiting the target frequency band, suppressing common-mode noise, and differential processing of spatially adjacent channels, the flow velocity-related information in the raw time series is extracted and coupled with a time-series feature recognition model to achieve robust identification of fluid velocity.

[0004] A DAS flow velocity identification method based on physical prior-guided one-dimensional temporal feature extraction includes the following steps: S100 acquires raw time-series acoustic signals of DAS related to fluid flow, and obtains raw time-series DAS data distributed by spatial location and time. S200, based on flow-related physical priors, flow velocity-sensitive input samples are constructed from the original DAS time-series data to obtain flow velocity-sensitive one-dimensional time-series input samples for flow velocity identification: S300, input the velocity-sensitive one-dimensional time-series input sample into the time-series feature recognition model to obtain time-series features related to fluid velocity; S400, input the time-series features into the regression output layer to obtain the fluid velocity identification result; S500, output the flow velocity identification result, which is used to characterize the actual flow velocity state of the fluid in the measured pipe or well.

[0005] The present invention also provides a DAS flow velocity identification system based on physical prior guidance for one-dimensional time-series feature extraction, used to perform the above method steps. The system 10 includes a data acquisition unit 101, an input sample construction unit 102, a feature extraction unit 103, a flow velocity regression unit 104, and a result output unit 105. It can process the raw time-series DAS data related to fluid flow and output the identification result characterizing the fluid flow velocity state. In addition, the present invention also provides a storage medium storing a computer program that, when run on a processor, causes the processor to execute the various steps of the fluid flow rate state recognition method described above.

[0006] This invention achieves quantitative identification and output of fluid velocity based on raw DAS time-series data, reducing reliance on empirical feature construction and manual parameter setting. It is applicable to the characterization and comparative analysis of flow conditions in wells or pipelines under experimental and field conditions. At the same time, the implementation is flexible and easy to integrate with existing DAS acquisition and processing systems. It can serve as a general method and system framework for conducting velocity identification and result comparison analysis under well flow conditions. Attached Figure Description

[0007] Figure 1 This is a flowchart of a DAS flow velocity identification method based on physical prior guidance for one-dimensional temporal feature extraction; Figure 2 A time series comparison chart of flow velocity identification results and reference flow velocities; Figure 3 This is a schematic diagram of a DAS flow velocity identification system based on physical prior guidance for one-dimensional temporal feature extraction. Detailed Implementation

[0008] To further illustrate the technical solution of the present invention, the following describes the implementation process of a DAS flow velocity identification method based on physical prior guidance for one-dimensional temporal feature extraction, in conjunction with specific embodiments. This embodiment describes the corresponding steps S100 to S500 of the method to demonstrate the feasibility and process consistency of the present invention under experimental or field conditions.

[0009] Example 1: As Figure 1 As shown, this embodiment provides a DAS flow velocity identification method based on physical prior guidance for one-dimensional temporal feature extraction, including the following steps: S100 acquires raw time-series acoustic signals of DAS related to fluid flow, and obtains raw time-series DAS data distributed by spatial location and time. S200, based on flow-related physical priors, flow velocity-sensitive input samples are constructed from the original DAS time-series data to obtain flow velocity-sensitive one-dimensional time-series input samples for flow velocity identification: S300, input the velocity-sensitive one-dimensional time-series input sample into the time-series feature recognition model to obtain time-series features related to fluid velocity; S400, input the time-series features into the regression output layer to obtain the fluid velocity identification result; S500, output the flow velocity identification result, which is used to characterize the actual flow velocity state of the fluid in the measured pipe or well. The method described in this embodiment of the invention can achieve quantitative identification and output of fluid velocity based on raw DAS time-series data related to fluid flow, thereby providing support for the characterization and comparative analysis of flow conditions under wellbore flow conditions.

[0010] Further, in step S100, raw time-series acoustic signals of DAS related to fluid flow are acquired to obtain raw time-series DAS data distributed according to spatial location and time. The following scheme can be adopted: In this embodiment, a circulating pipeline flow experiment device is used as the implementation platform. The total length of the experimental pipe section is 10m, and the inner diameter is 32mm. The experimental liquid is injected into the pipe section through a pumping system to form a stable flow condition. The flow velocity range can be set to 0.5–3.0m / s. In order to obtain raw DAS time-series data under multiple conditions and use it for flow velocity identification training and verification, the experimental flow velocity is set to a segmented constant condition. After each change of target flow velocity, a steady-state waiting time is maintained to allow the flow to stabilize sufficiently before collecting effective data under the corresponding condition. In this embodiment, six target flow velocities are set sequentially: 0.80m / s, 1.5 ... The flow rates were 0 m / s, 2.40 m / s, 1.20 m / s, 2.80 m / s, and 3.00 m / s, with each velocity lasting 300 s, thus obtaining segmented stable flow rate data covering the range of 0.5–3.0 m / s. A single-mode sensing fiber was axially deployed along the inner wall of the pipe segment and connected to a DAS signal acquisition device. The sampling frequency of the DAS signal acquisition device was set to 5 kHz, and the spatial resolution was 0.5 m, thereby obtaining raw DAS time-series data at different spatial locations along the pipe segment's axis. This raw DAS time-series data was organized into a matrix X according to the spatial sampling location and time sampling point, with elements x. i [n], where i=1,2,...,M are spatial channel indices, and n=0,1,...,N-1 are time sampling point indices; in this embodiment, the number of spatial channels M=20, and the time sampling interval Δt=1 / f s =0.0002s, the effective sampling time for a single flow velocity is 300s, corresponding to the number of sampling points N=1500000.

[0011] Further, in step S200, based on flow-related physical priors, flow velocity-sensitive input samples are constructed from the original DAS time-series data to obtain flow velocity-sensitive one-dimensional time-series input samples for flow velocity identification. The following scheme can be adopted: In this embodiment, the original timing x of each channel i obtained in step S100 is... i [n] Calculate the mean: In the formula, μ i Let x be the mean of the i-th channel, N be the number of sampling points, and x be the mean of the i-th channel. i [n] represents the original timing amplitude of the DAS; And obtain the DC removal sequence: In the formula, This is the timing amplitude after DC removal; Calculate the standard deviation for the DC-DC sequence: In the formula, σ iLet be the standard deviation of the DC sequence for the i-th channel; Normalization yields: In the formula, The normalized time series is represented by ε, which is a very small positive number to prevent the denominator from being 0. After obtaining the normalized sequence, to ensure that the main energy of subsequent input samples is concentrated in the dominant frequency band related to fluid flow disturbances, and to reduce the interference of out-of-band components unrelated to flow velocity mechanisms on the recognition process, the normalized sequence is subjected to target frequency band limitation processing based on the prior of flow-related dimensionless characteristic frequencies. The number of dimensionless characteristic frequencies is defined as: In the formula, St is the dimensionless characteristic frequency number, f is the frequency, L is the characteristic scale, and U is the fluid velocity; Under the condition that the disturbance mechanism and scale remain consistent, the dominant frequency f and the flow velocity U can be correlated through St. Therefore, the range constraints of St and U can be transformed into frequency range constraints, so that the frequency band selection can have a verifiable physical basis. The value of the characteristic scale L is determined based on the device geometry and the identifiability of the disturbance scale. Preferably, when a definite disturbance structure exists within the pipe section, L is taken as the characteristic size of the disturbance structure so that the defined frequency band directly corresponds to the target disturbance mechanism. When there is no independent disturbance structure and the dominant disturbance scale is on the same order of magnitude as the pipe diameter, L is taken as the inner diameter D of the pipe. In this embodiment, the inner diameter D of the experimental pipe section is 32 mm, therefore, L is taken as: To cover all possible flow velocity conditions of the experimental setup and ensure that the frequency band boundaries derived from physical priors are valid for all conditions, the preset flow velocity range is as follows: In the formula, U min For the minimum flow velocity boundary, U max The maximum flow velocity boundary; In this embodiment, the stably controllable flow velocity range is 0.5-3.0 m / s, therefore, the value is: The preset dimensionless characteristic frequency range is: In the formula, St min St is the lower bound of the dimensionless characteristic frequency. max This is the upper bound of the dimensionless characteristic frequency; The value used in this embodiment is: The interval is a preset parameter that can be adjusted according to the disturbance scale and noise level. It is used to cover the dimensionless characteristic frequency fluctuation range of the target disturbance under different flow velocity conditions, so that the frequency band derived from the interval can cover the dominant frequency band related to flow velocity, and avoid the introduction of too much out-of-band noise into the input sample due to the excessively wide frequency band. According to the L and U min U max St min St max The boundary frequencies defined by the target frequency band are: In the formula, f min f is the lower limit of the target frequency band. max The upper limit of the target frequency band; Substituting the parameters L=0.032m and U into this embodiment... min =0.5m / s, U max =3.0m / s, St min =0.14、St max =0.24, therefore: Therefore, the target bandwidth is limited to 2.19-22.5Hz, and the normalized sequence is subjected to bandpass filtering for energy focusing to obtain the bandwidth-limited sequence: In the formula, Let be the time sequence of the i-th channel after band limiting, and BPF() be the bandpass filter operator; In this embodiment, f s =5kHz, Nyquist frequency f Nyq =f s / 2, A fourth-order Butterworth bandpass filter is used to define the target frequency band in order to obtain a smooth passband response and effectively attenuate out-of-band components, so that subsequent common-mode suppression and spatial difference can be carried out within the target frequency band; To characterize the concentration of signal energy within the target frequency band after band definition, an in-band energy percentage index is introduced: In the formula, η i P represents the percentage of energy within the band. i (f) represents the power spectral density of channel i; This metric is used to characterize the degree to which the target frequency band is concentrating energy within the target frequency band, and it varies with η. i The increase indicates that out-of-band components are further suppressed, and the proportion of velocity-related frequency band information in the input sample is increased; After completing the target frequency band limiting process and obtaining Subsequently, common-mode suppression processing is performed on the multi-channel timing signals within the target frequency band to reduce cross-channel consistent or highly correlated background common-mode disturbance components. The signal after frequency band limitation is represented as follows: In the formula, s i [n] represents the local flow disturbance component, c[n] represents the common mode component, and v i [n] represents the channel-independent noise components; This embodiment employs a target-band common-mode suppression method based on a reference channel. First, a reference channel r is determined among M spatial channels. To ensure that the reference channel best represents the common-mode disturbance and weakens the local flow response, the in-band energy percentage η of each channel is calculated within the steady-state effective data segment. i Based on its time fluctuation and to determine the reference channel, the energy percentage within the band is defined as follows: In the time dimension, the steady-state effective segment is divided into time windows W. k Segmentation, calculate the in-band energy for each channel within the k-th time window: In the formula, For the i-th channel within the k-th time window, [f min f max Energy within the frequency band Let be the power spectral density of the i-th channel within the k-th time window; Further calculation of the coefficient of variation of the energy within the band: In the formula, CV i Let be the coefficient of variation of the in-band capability sequence of the i-th channel, std() be the standard deviation operator, and mean() be the mean operator. The reference channel is determined based on the criterion of low in-band energy proportion and small fluctuation: while satisfying η i ≤η th Select CV from the channels i The smallest channel is used as the reference channel r; in this embodiment, η is taken. th =0.30; After determining the reference channel r, common-mode suppression is performed on each channel i: In the formula, y i [n] represents the value of the time sequence of the i-th channel after common-mode suppression at sampling point n. The amplitude matching coefficient of the i-th channel within the k-th time window. The values ​​of the time sequence at sampling point n after the frequency band of the reference channel r is defined; W in each time window k Inside, to Using least squares estimation, let n∈W k ,but: In the formula, n is the sampling point index, W k Let k be the set of sampling points contained in the k-th time window; Will Substituting into the above equation, we obtain the common-mode suppression result within the window: To highlight the spatial variation characteristics of local flow disturbances and weaken the slowly varying background components along the axial direction, the signal y after common-mode suppression processing is... i [n] undergoes spatial difference processing, and the difference form between adjacent channels is: In the formula, Δy i [n] represents the difference result of the i-th differential channel at sampling point n, and M is the original number of channels; After completing the target frequency band limitation, common mode suppression, and spatial difference, in order to construct a flow velocity-sensitive one-dimensional time series input sample, the processed sequence is divided into time windows, with the number of time window points and the number of step points being: In the formula, N w T represents the number of points in the time window. w N is the length of the time window. s T represents the step size. s The sliding step size, This is the floor operator; In this embodiment, T w =0.2 s, T s =0.05 s、f s =5 kHz, resulting in: Define the starting sampling point of the k-th time window as n. k =(k−1)N s Then, the velocity-sensitive one-dimensional time-series input sample for any differential channel i is represented as: In the formula, For the flow velocity-sensitive one-dimensional time series input sample corresponding to the i-th differential channel within the k-th time window, n k Let [ ] be the index of the starting sampling point for the k-th time window. T This is a transpose operation; The velocity-sensitive one-dimensional time series input samples corresponding to multiple spatial locations along the axial direction within the same time window k are used as samples. The input samples X are organized in spatial order. (k) This yields a flow velocity-sensitive one-dimensional time-series input sample set. To avoid the transient impact of flow rate switching and ensure that the samples come from the steady-state period, this embodiment takes the subsequent 100 seconds as the effective steady-state data segment within each 300-second operating condition segment for time window segmentation; under the conditions of Tw=0.2 s and Ts=0.05 s, the number of time windows that can be obtained for a single 100-second segment is as follows: In the formula, K is the number of time windows that can be obtained by dividing the time window into sliding time windows within a steady-state effective segment; This completes the preprocessing in step S200 and yields a flow-sensitive one-dimensional time-series input sample.

[0012] Further, in step S300, the velocity-sensitive one-dimensional time-series input sample is input into the time-series feature recognition model to obtain time-series features related to fluid velocity. The following scheme can be adopted: In this embodiment, a pre-trained temporal feature recognition model is used to extract temporal features from the flow velocity-sensitive one-dimensional temporal input sample. The input sample corresponding to the k-th time window is X. (k) Its dimensions are represented as: In the formula, X (k) The input sample corresponding to the k-th time window is formed by stacking multi-channel one-dimensional sequences according to the channel dimension. R is the real number field, and C is the number of spatial difference channels. In this embodiment, the input sample is constructed from a spatial difference sequence, with 19 spatial difference channels C; the time window length T... w The sampling time is 0.2s, and the sampling frequency is f. s 5kHz, N w Since it is 1000, therefore: Input sample X (k) The temporal feature recognition model is input, which includes an input layer, a one-dimensional convolutional feature extraction module, and a feature fusion module. Let H0 = X. (k) The convolution kernel of the l-th layer one-dimensional convolution is W. l The bias is b l The convolution output is H l The convolution operation of the l-th layer is represented as: In the formula, H lW represents the output feature map of the l-th one-dimensional convolution layer, where l is the convolution layer index. l Let b be the weight parameters of the one-dimensional convolution kernel in the l-th layer. l For the bias parameters of the l-th layer, This is a one-dimensional convolution operation; In this embodiment, to obtain flow velocity-related features at different time scales, the one-dimensional convolutional feature extraction module adopts a multi-scale convolutional structure. Specifically, it sets parallel one-dimensional convolutional branches with different kernel lengths within the same layer and concatenates the outputs of each branch to form a fused feature. The kernel lengths of the three parallel branches are p1, p2, and p3, and their corresponding outputs are... , , Multi-scale splicing is represented as: In the formula, In the l-th layer, the kernel length is p j The branch output feature map, the Concat() concatenation operator means concatenating multiple branch outputs along the channel dimension; In this embodiment, p1 is 3, p2 is 5, and p3 is 7. Batch normalization and max pooling are performed after each convolutional layer to stabilize training and compress the time dimension. The pooling kernel length is 2 and the stride is 2. After performing multi-layer convolutional feature extraction and feature fusion, global average pooling is performed on the final feature map to obtain the fixed-length feature vector h corresponding to the k-th input sample. (k) : In the formula, h (k) H is the fixed-length feature vector corresponding to the k-th sample. L This is the feature map output by the last layer (layer L), where L is the network layer index, and GAP() is the global average pooling operator, which means averaging the features of each channel over time to obtain a vector of fixed length. In this embodiment, the time-series feature recognition model is obtained through pre-training. The training data consists of the velocity-sensitive one-dimensional time-series input samples constructed in step S200 and their corresponding synchronous velocity labels. The velocity label of the k-th sample is U. (k) The training sample pairs are (X) (k) U (k) In this embodiment, the six steady-state flow velocity conditions are 0.80 m / s, 1.20 m / s, 1.50 m / s, 2.40 m / s, 2.80 m / s, and 3.00 m / s, respectively. The effective data segment for each steady-state condition is 100 s, and the time window length T is [missing information]. w =0.2 s, step size T s =0.05 s, the number of samples in a single segment is: Therefore, the total number of samples is 11982. These samples are divided into training and validation sets at 80% and 20% respectively. Supervised training is then performed on the model parameters, with the training objective being to minimize the mean squared error loss function between the predicted flow velocity and the flow velocity label. In the formula, Here, B represents the training objective function, and B is the batch size, which is the number of samples used for one parameter update. U is the predicted flow velocity value for the k-th sample. (k) The flow rate label for the k-th sample; In this embodiment, the batch size B is 128, and the initial learning rate is 1×10⁻⁶. -3 The training iterations are 50. Training stops when the validation set loss does not decrease for 5 consecutive iterations, and the model parameters with the minimum validation loss are saved. After training is complete, the saved optimal model parameters are used to extract temporal features from the input samples, and the fused temporal feature h is output. (k) .

[0013] Further, in step S400, the time-series features are input into the regression output layer to obtain the fluid velocity identification result, which can be achieved using the following scheme: The fused temporal feature h (k) The input regression output layer yields continuous numerical fluid velocity identification results corresponding to the velocity-sensitive one-dimensional time-series input samples. Let the parameters of the regression output layer be the weight vector w and the bias b. Then, the velocity identification result corresponding to the k-th time window is expressed as: In the formula, w is the weight vector of the regression output layer, and b is the bias term of the regression output layer; This yields the flow velocity prediction sequence that varies with the time window sequence. .

[0014] Furthermore, in step S500, the flow velocity identification result is output to characterize the actual flow velocity state of the fluid in the measured pipe or well. The following scheme can be adopted: In this embodiment, The flow velocity time series is generated by outputting the data in chronological order. For each steady-state condition segment, a statistical measure is calculated based on the window prediction results as the equivalent flow velocity identification value for that condition. This value is defined as follows: In the formula, The equivalent flow velocity is the identification value for the steady-state operating condition. This refers to the set of time window indices used for statistical analysis within the steady-state operating condition period. Each working condition section With time series The combined output is used to characterize the steady-state level and time-varying process of the fluid velocity in the tested pipe section.

[0015] like Figure 3 As shown, the present invention also provides a DAS flow rate identification system 10 based on physical prior guidance for one-dimensional temporal feature extraction; the system 10 includes a processor, a memory, and a computer program stored in the memory and executable on the processor; when the computer program is executed by the processor, it is used to implement the operations described in steps S100 to S500 in the method embodiment of the present invention. The system 10 includes the following functional modules: Acquisition unit 101 is used to acquire raw time-series acoustic signals of DAS related to fluid flow, and obtain raw time-series DAS data distributed according to spatial location and time. The input sample construction unit 102 is used to construct flow-sensitive input samples on the original DAS time series data based on flow-related physical priors, so as to obtain flow-sensitive one-dimensional time series input samples for flow velocity identification. Feature extraction unit 103 is used to input the velocity-sensitive one-dimensional time-series input sample into the time-series feature recognition model to obtain time-series features related to fluid velocity; The velocity regression unit 104 is used to input the time-series features into the regression output layer to obtain the fluid velocity identification result; The result output unit 105 is used to output the flow velocity identification result, which characterizes the actual flow velocity state of the fluid in the measured pipe or well.

[0016] Furthermore, the system 10 can be deployed in a computer, well site data acquisition terminal, edge computing device, remote processing server or cloud platform to process the raw DAS time series data and output flow rate identification results; Furthermore, the system 10 can transmit the flow velocity identification results to the well site monitoring system or data analysis platform through the communication interface for continuous monitoring, recording and comparative analysis of the flow conditions.

[0017] The present invention also provides a computer-readable storage medium storing a computer program that, when executed on a processor, causes the processor to perform the DAS flow rate identification method based on physical prior guidance for one-dimensional temporal feature extraction, as described in steps S100 to S500 of the method embodiment of the present invention; the computer program includes instructions for implementing the following operations: (1) Instructions for acquiring raw time-series acoustic signals of DAS related to fluid flow and obtaining raw time-series data of DAS distributed by spatial location and time. (2) An instruction to construct flow velocity-sensitive input samples from the original DAS time series data based on flow-related physical priors, so as to obtain flow velocity-sensitive one-dimensional time series input samples for flow velocity identification; (3) An instruction for inputting the velocity-sensitive one-dimensional time-series input sample into the time-series feature recognition model to obtain time-series features related to fluid velocity; (4) An instruction for inputting the time-series features into the regression output layer to obtain the fluid velocity identification result; (5) An instruction for outputting the flow velocity identification result, which is used to characterize the actual flow velocity state of the fluid in the measured pipe or well; The computer-readable storage medium may be a read-only memory (ROM), random access memory (RAM), flash memory, solid-state drive (SSD), disk, memory card, or other medium capable of storing program code.

[0018] The parameters described above are merely an example setting in this embodiment. Those skilled in the art can reasonably adjust the processing parameters, time window length, and step size based on factors such as the size and structure of the flow channel, fluid properties, and DAS configuration, without affecting the implementation and effectiveness of the method of this invention. Through the above processing flow, even under different wellbore conditions, different flow conditions, and different DAS acquisition configurations, the corresponding fluid velocity identification results can be stably output based on the original DAS time series data, improving the consistency and comparability of cross-condition applications. It should be understood that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the present invention; the scope of protection of the present invention shall be determined by the appended claims, and the embodiments in this specification do not constitute a limitation on the claims.

Claims

1. A DAS flow velocity identification method based on physical prior-guided one-dimensional temporal feature extraction, characterized in that, Includes the following steps: S100 acquires raw time-series acoustic signals of DAS related to fluid flow, and obtains raw time-series DAS data distributed by spatial location and time. S200, based on flow-related physical priors, flow velocity-sensitive input samples are constructed from the original DAS time-series data to obtain flow velocity-sensitive one-dimensional time-series input samples for flow velocity identification: S300, input the velocity-sensitive one-dimensional time-series input sample into the time-series feature recognition model to obtain time-series features related to fluid velocity; S400, input the time-series features into the regression output layer to obtain the fluid velocity identification result; S500, output the flow velocity identification result, which is used to characterize the actual flow velocity state of the fluid in the measured pipe or well.

2. The method according to claim 1, wherein, S100 includes: laying distributed optical fibers along the shaft or pipeline axis and acquiring raw DAS time-series acoustic signals related to fluid flow; organizing the raw time-series acoustic signals according to spatial location channels to form raw DAS time-series data that is spatially distributed along the shaft or pipeline axis and continuously sampled over time.

3. The method according to claim 1, wherein, S200 includes: constructing velocity-sensitive input samples from the raw DAS time-series data. The construction of velocity-sensitive input samples includes determining a velocity-related target frequency band based on the perturbation structure characteristic scale, candidate velocity range, and dimensionless characteristic frequency range, and limiting the target frequency band of the raw DAS time-series data; performing common-mode suppression on multi-channel time-series signals within the target frequency band; performing differential processing on spatially adjacent channels after common-mode suppression; and performing time window segmentation and sample organization on the processed time-series data to obtain the velocity-sensitive one-dimensional time-series input samples.

4. The method according to claim 1, wherein, The S300 includes: employing a time-series feature recognition model, which is used to perform feature representation and flow-velocity-related feature extraction on the flow velocity-sensitive one-dimensional time-series input sample, and the time-series feature recognition model is a time-series learning model obtained based on sample training.

5. The method according to claim 1, wherein, The S400 includes: inputting the time-series features extracted by the time-series feature recognition model into the regression output layer, and outputting a continuous numerical fluid velocity recognition result corresponding to the velocity-sensitive one-dimensional time-series input sample, wherein the fluid velocity recognition result is any one of the average velocity, instantaneous velocity, or equivalent velocity corresponding to the time window segmentation.

6. A DAS flow velocity identification system based on physical prior-guided one-dimensional temporal feature extraction, comprising an acquisition unit, an input sample construction unit, a feature extraction unit, a flow velocity regression unit, and a result output unit, wherein the unit is used to execute the method steps of claims 1 to 5.

7. A computer-readable storage medium for DAS flow rate identification based on one-dimensional temporal feature extraction guided by physical priors, wherein, The computer-readable storage medium stores a computer program that, when executed on a processor, causes the processor to perform the method of any one of claims 1 to 5.

Images