Self-adaptive ultrasonic measurement method and system based on multichannel collaboration

Abstract

The invention discloses a self-adaptive ultrasonic measurement method and system based on multi-channel cooperation, and relates to the technical field of ultrasonic measurement. The method comprises the following steps: collecting original flight time signals of each channel; meanwhile, environment data are collected in real time; original flight time signals are processed, effective signal arrival pre-flight time is extracted, and a fusion data set is constructed; constructing a two-dimensional sound channel spectrogram based on the multi-sound channel time sequence of the pre-flight time, and performing feature extraction based on a lightweight convolutional neural network to generate an abnormal confidence vector; analyzing the sound channel state based on the abnormal confidence vector; constructing a physical information neural network, and analyzing the corrected sound velocity value of each sound channel and the two-dimensional sound velocity field distribution on the section of the whole pipeline; training an online sequence extreme learning machine model in combination with historical measurement data; and based on the final fusion weight of each sound channel and the corresponding sound channel flow velocity, carrying out weighted fusion to generate a flow velocity optimal estimation value. And the measurement precision and robustness are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of ultrasonic measurement, and particularly relates to an adaptive ultrasonic measurement method and system based on multi-channel cooperation. BACKGROUND

[0002] With the rapid development of automatic control technology, computer technology and sensor technology, ultrasonic nondestructive testing has entered the modernization process first, and is now advancing towards the direction of digitization, automation, intelligence and imaging technology, and its application field is continuously expanding, and its market scale is maintaining a rapid growth trend.

[0003] However, the measurement model of the prior art has poor adaptability and cannot accurately compensate for the influence of a complex fluid environment. The mainstream method usually relies on a few environmental parameters to globally and uniformly estimate the sound velocity through a simple empirical formula. This simplification assumption is seriously inconsistent with the physical reality of the uneven distribution of temperature and composition on the pipe cross section. Especially in the case of asymmetric flow, high turbidity or temperature gradient, this uniform sound velocity assumption will introduce significant model errors. In addition, the system error caused by the slow change of the sensor characteristics drift, pipe fouling and other factors in the long-term operation of the instrument cannot be identified and calibrated online by the traditional method, and must rely on periodic shutdown calibration to maintain accuracy, which not only increases the maintenance cost, but also affects the continuity of the production process.

[0004] Therefore, the present application discloses an adaptive ultrasonic measurement method and system based on multi-channel cooperation to solve the above problems. SUMMARY

[0005] The present application aims to provide an adaptive ultrasonic measurement method and system based on multi-channel cooperation to solve the problems in the prior art.

[0006] To achieve the above-mentioned purpose, the present application provides the following technical scheme: an adaptive ultrasonic measurement method based on multi-channel cooperation, which comprises the following steps:

[0007] S1: transmitting coded ultrasonic pulses in a time division multiplexing manner by an ultrasonic transducer array, and synchronously collecting original flight time signals of each channel; simultaneously collecting environmental data in real time; processing the original flight time signals to extract effective signal arrival pre-flight time, and constructing a fusion data set;

[0008] S2: constructing a two-dimensional channel spectrum based on the multi-channel time series of the pre-flight time, performing feature extraction based on a lightweight convolutional neural network to generate an anomaly confidence vector, and analyzing the channel state based on the anomaly confidence vector;

[0009] S3: Construct a physical information neural network, fuse a preset database, analyze the corrected sound velocity value of each sound channel and the two-dimensional sound velocity field distribution on the entire pipeline cross section;

[0010] S4: Based on the corrected sound velocity value and the sound velocity field distribution, combined with historical measurement data, training an online sequence extreme learning machine model; based on the final fusion weight of each sound channel and the corresponding sound channel flow rate weighted fusion, generating an optimal flow rate estimation value.

[0011] According to the above scheme, in S1, the following contents are included:

[0012] S101: Emit a linear Chirp pulse signal with orthogonal coding from an ultrasonic transducer array, the orthogonal coding adopts at least one coding scheme selected from a Walsh sequence, a Gold sequence, a pseudo-random phase code or an orthogonal frequency division scheme; capture the original time-of-flight signal through synchronous sampling of the receiving transducer channel driven by a unified clock by an FPGA platform; the FPGA platform generates a unified timestamp in each transmission and reception cycle, and synchronously latches data from temperature sensors, piezoresistive pressure sensors and turbidity sensors; sort according to the sound channel number to generate an environment vector;

[0013] S102: After band-pass filtering, wavelet packet denoising is performed on each original time-of-flight signal, the wavelet packet denoising uses a Symlets wavelet basis as a base function and is decomposed into several layers, and a denoising signal is reconstructed; the denoising signal and the Chirp pulse signal are applied to the PHAT normalization algorithm of generalized cross-correlation, and the cross-correlation function is calculated; the peak position of the cross-correlation function is recorded as the pre-time-of-flight, and the pre-time-of-flight is sorted according to the sound channel number to generate a pre-time-of-flight vector;

[0014] S103: Calculate the signal-to-noise ratio of each denoising signal, denoted as a signal quality indicator, sort the signal quality indicator according to the sound channel number to generate a signal quality indicator vector; normalize and splice the pre-time-of-flight vector, the signal quality indicator vector and the environment vector to generate a fusion data set.

[0015] The unified clock driving ensures strict time synchronization of all channel data acquisition, which is the basis for high-precision time-of-flight measurement and subsequent cross-sound channel feature analysis, and eliminates system errors introduced by clock asynchronization; synchronous acquisition of temperature, pressure and turbidity provides crucial physical background information for subsequent sound velocity compensation and flow state analysis, and constitutes the data basis for system adaptive capability;

[0016] Wavelet packet denoising can effectively remove non-stationary noise, and is more suitable for processing complex noise in actual industrial environment than traditional Fourier filtering, and can significantly improve the signal-to-noise ratio. The GCC-PHAT algorithm is not sensitive to amplitude, can sharpen the cross-correlation peak, and effectively overcome the peak broadening problem caused by multiple reflections, reverberation and the like, thereby improving the accuracy and reliability of the time-of-flight preliminary estimation (pre-time-of-flight).

[0017] According to the above scheme, in S2, the following is included:

[0018] S201: arrange the continuous historical observation data of the pre-time-of-flight vector in a sliding time window in chronological order into a two-dimensional data matrix; the total number of elements of the two-dimensional data matrix is equal to the total number of channels multiplied by the preset sliding time window length; perform bilinear interpolation processing on the two-dimensional data matrix to generate a gray image, denoted as a channel spectrum image; wherein the row coordinates of the channel spectrum image correspond to the channel index, the column coordinates correspond to the time sequence, and the pixel gray value represents the pre-time-of-flight of the corresponding channel at the corresponding time;

[0019] S202: input the channel spectrum image into a pre-trained lightweight convolutional neural network to output an abnormal confidence vector with the same dimension as the total number of channels;

[0020] S203: input each element in the abnormal confidence vector and the abnormal judgment threshold into an exponential function respectively; if the element in the abnormal confidence vector is greater than the abnormal judgment threshold, the exponential function outputs 1; if the element in the abnormal confidence vector is less than or equal to the abnormal judgment threshold, the exponential function outputs 0; generate a binary abnormal identification vector and transmit it to the administrator; analyze the difference between the all-1 vector and the abnormal confidence vector, denoted as a weight vector, and the elements in the weight vector represent the reliability weight of each channel.

[0021] The application converts one-dimensional time series data into images, and ingeniously converts the abnormal patterns of channel dimension and time dimension into visual features such as texture, spot and edge in the image, lays a foundation for using mature image processing technology for pattern recognition; the CNN can automatically learn and extract complex abnormal features of channel data in space-time; the output is a confidence degree between 0 and 1, not just a binary judgment of "yes / no", providing more abundant information for subsequent weighted fusion; and the usability and reliability of the system are improved.

[0022] According to the above scheme, in S3, the following is included:

[0023] S301: splice the environment vector and the weight vector to construct an input feature vector of a physical information neural network; a loss function of the physical information neural network is composed of a data fitting loss and a physical equation constraint loss weighted; the data fitting loss is equal to a mean square error between a predicted sound velocity value and a reference sound velocity value corresponding to a training data set of a plurality of samples; the predicted sound velocity value is a corresponding output value of a training sample in the physical information neural network; the reference sound velocity value is equal to a sound velocity true value corresponding to the training sample obtained by querying or interpolating from a pre-calibrated sound velocity and environment parameter relationship database under a known temperature and pressure condition; the physical equation constraint loss is equal to a residual error of a Helmholtz equation or an Eikonal equation of sound wave propagation at discrete points in a calculation domain;

[0024] S302: use the trained physical information neural network as a sound velocity field reconstruction model, input the input feature vector and normalized coordinates of any point in the pipeline cross section, and output a sound velocity prediction value of the point; for each sound channel, the sound velocity field is integrated along an ultrasonic wave propagation path to calculate an average sound velocity on the ultrasonic wave propagation path, denoted as a corrected sound velocity value of each sound channel, and integrated to generate a two-dimensional sound velocity field distribution map on the pipeline cross section.

[0025] The application more accurately describes the actual distribution of sound velocity on the pipeline cross section (rather than the assumed uniform distribution), is especially suitable for cases such as asymmetric flow and existence of temperature gradient, thereby more accurately compensating for the time of flight and realizing high-precision measurement; the physical constraint plays a role of regularization, so that the model can learn good rules even in the case that the labeled data (reference sound velocity true value) is not so sufficient, and the generalization ability of the model is enhanced; the accurate flow rate optimal estimation value is a prerequisite for accurately solving the flow rate, significantly reduces the measurement error caused by inaccurate sound velocity estimation, and improves the accuracy of the system.

[0026] According to the above scheme, in S4, the following contents are included:

[0027] S401: take the environment vector, the corrected sound velocity value vector and the weight vector as an input feature vector of an online sequence extreme learning machine model; calculate a preliminary flow rate value V raw ;

[0028]

[0029] wherein w i represents the reliability weight of the i-th sound channel; L i represents the ultrasonic wave propagation path length of the i-th sound channel; t i pre represents the preliminary flight time of the i-th sound channel, and θ idenotes the included angle between the i-th sound channel axis and the pipeline axis; c i denotes the corrected sound velocity value of the i-th sound channel;

[0030] The deviation of the preliminary flow rate value from the optimal estimated flow rate value of the last calibration cycle is used as the supervised learning target deviation of the online sequence extreme learning machine model; the online sequence extreme learning machine model receives new input feature vectors and target deviations in an online sequence manner, and recursively updates the connection weight matrix from the hidden layer to the output layer at each time point;

[0031] S402: using the trained online sequence extreme learning machine model to predict the flow rate deviation value Δv of the current system; combining the weight vector to calculate the final fusion weight of each sound channel; the final fusion weight of the i-th sound channel is denoted as w i final ;

[0032]

[0033] Based on the corrected sound velocity value and the pre-flight time of each sound channel, a sound propagation model is constructed:

[0034]

[0035] wherein v i denotes the flow rate of the i-th sound channel; Newton iteration method or fixed point iteration method is used for flow rate solving; the final fusion weight of each sound channel and the corresponding sound channel flow rate are weighted and fused to generate a flow rate optimal estimation value, and the flow rate optimal estimation value is transmitted to the administrator.

[0036] The present application uses an iterative method to accurately solve a complex sound propagation equation, avoids errors caused by simplified formulas, and further improves the calculation accuracy; when finally fused, not only the reliability weight is considered, but also the confidence after the output deviation correction of the OS-ELM is combined to generate the final "fusion weight", so that the system gives higher weight to more reliable and more accurate sound channel data, and finally outputs an optimal estimation value, improving the measurement accuracy and robustness.

[0037] Another aspect of the present application provides an adaptive ultrasonic measurement method based on multi-channel cooperation, which is applied to the above-mentioned adaptive ultrasonic measurement system based on multi-channel cooperation, and the system comprises a data acquisition and fusion module, a sound channel anomaly analysis module, a sound velocity field reconstruction and correction module, and an adaptive calibration module.

[0038] The data acquisition fusion module is configured to emit coded ultrasonic pulses in a time-division multiplexing manner through an ultrasonic transducer array, and synchronously acquire original time-of-flight signals of each sound channel; meanwhile, real-time acquisition of environmental data is performed; the original time-of-flight signals are processed, and effective signal arrival pre-time-of-flight is extracted to construct a fusion data set;

[0039] The sound channel anomaly analysis module is configured to construct a two-dimensional sound channel spectrum based on a multi-sound channel time sequence of the pre-time-of-flight, perform feature extraction based on a lightweight convolutional neural network to generate an anomaly confidence vector, and analyze a sound channel state based on the anomaly confidence vector;

[0040] The sound velocity field reconstruction and correction module is configured to construct a physical information neural network, fuse a pre-set database, analyze a corrected sound velocity value of each sound channel and a two-dimensional sound velocity field distribution on an entire pipeline cross section;

[0041] The adaptive calibration module is configured to train an online sequence extreme learning machine model based on the corrected sound velocity value and the sound velocity field distribution in combination with historical measurement data, and generate an optimal flow rate estimation value by weighted fusion of a final fusion weight of each sound channel and a corresponding sound channel flow rate.

[0042] According to the above scheme, the data acquisition fusion module comprises a data acquisition unit and a data fusion unit;

[0043] The data acquisition unit is configured to synchronously sample and capture original time-of-flight signals through FPGA platform unified clock driving of a receiving transducer channel, and synchronously latch data from a temperature sensor, a piezoresistive pressure sensor and a turbidity sensor;

[0044] The data fusion unit is configured to sort data of the temperature sensor, the piezoresistive pressure sensor and the turbidity sensor according to sound channel serial numbers to generate an environmental vector;

[0045] A PHAT normalization algorithm of generalized cross-correlation is applied to the de-noised signal and the Chirp pulse signal to calculate a cross-correlation function; a peak position of the cross-correlation function is recorded as a pre-time-of-flight, and the pre-time-of-flight is sorted according to sound channel serial numbers to generate a pre-time-of-flight vector;

[0046] A signal-to-noise ratio of each de-noised signal is calculated and recorded as a signal quality index, the signal quality index is sorted according to sound channel serial numbers to generate a signal quality index vector; the pre-time-of-flight vector, the signal quality index vector and the environmental vector are normalized and spliced to generate a fusion data set.

[0047] According to the above scheme, the sound channel anomaly analysis module comprises a feature analysis unit and a sound channel state analysis unit;

[0048] The feature analysis unit is configured to arrange continuous historical observation data of the pre-flight time vector in a sliding time window in time sequence into a two-dimensional data matrix; perform bilinear interpolation processing on the two-dimensional data matrix to generate a grayscale image, denoted as a sound channel spectrum image; input the sound channel spectrum image into a pre-trained lightweight convolutional neural network to output an abnormal confidence vector with the same dimension as the total number of sound channels;

[0049] The sound channel state analysis unit is configured to input each element in the abnormal confidence vector and an abnormality determination threshold into an exponential function respectively; generate a binary abnormality identification vector and transmit it to an administrator; and analyze the difference between the all-1 vector and the abnormal confidence vector.

[0050] According to the above scheme, the sound velocity field reconstruction correction module includes a neural network construction unit and a sound velocity correction unit.

[0051] The neural network construction unit is configured to construct a physical information neural network, and a loss function of the physical information neural network is composed of a data fitting loss and a physical equation constraint loss weighted; the data fitting loss is equal to the mean square error between the predicted sound velocity value and the reference sound velocity value corresponding to the training data set of a plurality of samples; and the physical equation constraint loss is equal to the residual error of the Helmholtz equation or the Eikonal equation of sound wave propagation at discrete points in the calculation domain.

[0052] The sound velocity correction unit is configured to use the trained physical information neural network as a sound velocity field reconstruction model, input a feature vector and normalized coordinates of any point in the pipeline cross section as inputs, and output a predicted sound velocity value at the point; for each sound channel, the sound velocity field is integrated along the ultrasonic wave propagation path to calculate the average sound velocity on the ultrasonic wave propagation path, denoted as the corrected sound velocity value of each sound channel, and a two-dimensional sound velocity field distribution map on the pipeline cross section is generated.

[0053] According to the above scheme, the adaptive calibration module includes an online learning unit and a calibration unit.

[0054] The online learning unit takes an environment vector, a corrected sound velocity value vector and a weight vector as an input feature vector of an online sequence extreme learning machine model; calculates a preliminary flow rate value based on the corrected sound velocity value vector and the pre-flight time vector using a weighted average method; takes the deviation of the preliminary flow rate value from the optimal estimated flow rate value in the last calibration period as a supervised learning target deviation of the online sequence extreme learning machine model; and the online sequence extreme learning machine model receives new input feature vectors and target deviations in an online sequence manner at each time instant, and recursively updates the connection weight matrix from the hidden layer to the output layer.

[0055] The calibration unit is used for predicting the flow rate deviation value of the current system by using the trained online sequence extreme learning machine model, combining the weight vector to calculate the final fusion weight of each acoustic channel, constructing the sound propagation model based on the corrected sound speed value and the pre-flight time of each acoustic channel, solving the flow rate by using the Newton iteration method or the fixed point iteration method, and generating the optimal estimation value of the flow rate by weighting and fusing the final fusion weight of each acoustic channel and the corresponding acoustic channel flow rate, and transmitting the optimal estimation value of the flow rate to the administrator.

[0056] Compared with the prior art, the beneficial effects of the present application are: the unified clock drive of the present application ensures strict time synchronization of data acquisition of all channels, eliminates system errors introduced by clock asynchronization, synchronously acquires temperature, pressure and turbidity, provides important physical background information for subsequent sound speed compensation and flow state analysis, and constitutes the data basis of the system adaptive ability; the wavelet packet denoising can effectively remove non-stationary noise, is more suitable for processing complex noise in actual industrial environment than traditional Fourier filtering, and significantly improves the signal-to-noise ratio; the GCC-PHAT algorithm is not sensitive to amplitude, can sharpen the cross-correlation peak value, and effectively overcomes the peak broadening problem caused by multiple reflections, reverberation and the like, thereby improving the accuracy and reliability of the pre-flight time; the present application converts one-dimensional time sequence data into an image, ingeniously converts the abnormal patterns of the acoustic channel dimension and the time dimension into visual features such as texture, spots and edges in the image, and lays a foundation for using mature image processing technology for pattern recognition; the CNN can automatically learn and extract complex abnormal features of the acoustic channel data in space and time; the output is a confidence degree between 0 and 1, not just a binary judgment of "yes / no", and provides more abundant information for subsequent weighted fusion; the usability and reliability of the system are improved. The present application more accurately describes the actual distribution of the sound speed on the pipeline cross section, is especially suitable for asymmetric flow, temperature gradient and the like, thereby more accurately compensating the flight time and realizing high-precision measurement; the physical constraint plays a role of regularization, so that the model can learn good rules even when the labeled data is not very sufficient, and the generalization ability of the model is enhanced; the accurate optimal estimation value of the flow rate is a prerequisite for accurately solving the flow rate, significantly reduces the measurement error caused by inaccurate sound speed estimation, and improves the accuracy of the system. The present application accurately solves the complex sound propagation equation by using the iteration method, avoids the error caused by the simplified formula, and further improves the calculation accuracy; when finally fusing, not only the reliability weight is considered, but also the confidence degree after the deviation correction of the OS-ELM output may be combined to generate the final "fusion weight", so that the system gives higher weight to more reliable and more accurate acoustic channel data, and finally outputs an optimal estimation value, thereby improving the measurement precision and robustness. BRIEF DESCRIPTION OF DRAWINGS

[0057] The accompanying drawings, which are included to provide a further understanding of the application and are incorporated in and constitute a part of this specification, illustrate embodiments of the application and together with the description serve to explain the application. In the drawings:

[0058] Figure 1 A flowchart of an adaptive ultrasonic measurement method based on multi-channel cooperation of the present application;

[0059] Figure 2 A structural schematic diagram of an adaptive ultrasonic measurement system based on multi-channel cooperation of the present application. DETAILED DESCRIPTION

[0060] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0061] Please refer to Figure 1 The present application provides a technical solution: an adaptive ultrasonic measurement method based on multi-channel cooperation, which comprises the following steps:

[0062] S1: transmitting coded ultrasonic pulses in a time-division multiplexing manner by an ultrasonic transducer array, and synchronously collecting original flight time signals of each channel; simultaneously, collecting environmental data in real time; processing the original flight time signals, extracting effective signal arrival pre-flight time, and constructing a fusion data set;

[0063] In S1, the following contents are included:

[0064] S101: transmitting orthogonal coded linear Chirp pulse signals by an ultrasonic transducer array, the orthogonal coding adopting at least one coding scheme selected from a Walsh sequence, a Gold sequence, a pseudo-random phase code or an orthogonal frequency division scheme; synchronously sampling and capturing original flight time signals by the FPGA platform for unified clock driving of the receiving transducer channels; the FPGA platform generates a unified time stamp at each transmission and reception cycle, and synchronously latches data from a temperature sensor, a piezoresistive pressure sensor and a turbidity sensor; sorting according to the channel number to generate an environmental vector;

[0065] S102: adopt band-pass filtering and wavelet packet denoising for each original time-of-flight signal, the wavelet packet denoising is based on a Symlets wavelet base and is decomposed into several layers, and a denoised signal is obtained by reconstruction; a generalized cross-correlation PHAT normalization algorithm is applied to the denoised signal and a Chirp pulse signal to calculate a cross-correlation function; a peak position of the cross-correlation function is recorded as a pre-time-of-flight, and the pre-time-of-flight is sorted according to a sound channel serial number to generate a pre-time-of-flight vector;

[0066] S103: calculate a signal-to-noise ratio of each denoised signal, record the signal-to-noise ratio as a signal quality index, sort the signal quality index according to a sound channel serial number to generate a signal quality index vector, and normalize and splice the pre-time-of-flight vector, the signal quality index vector and an environment vector to generate a fusion data set.

[0067] S2: construct a two-dimensional sound channel spectrum based on a multi-sound channel time sequence of the pre-time-of-flight, extract features based on a lightweight convolutional neural network to generate an anomaly confidence vector, and analyze a sound channel state based on the anomaly confidence vector;

[0068] In S2, the following contents are included:

[0069] S201: arrange continuous historical observation data of the pre-time-of-flight vector in a sliding time window in chronological order into a two-dimensional data matrix; the total number of elements of the two-dimensional data matrix is equal to the total number of sound channels multiplied by a preset sliding time window length; perform bilinear interpolation processing on the two-dimensional data matrix to generate a gray-scale image, which is recorded as a sound channel spectrum; wherein the row coordinates of the sound channel spectrum correspond to the sound channel index, the column coordinates correspond to the time sequence, and the pixel gray-scale value represents the pre-time-of-flight of the corresponding sound channel at the corresponding time;

[0070] S202: input the sound channel spectrum into a pre-trained lightweight convolutional neural network to output an anomaly confidence vector with the same dimension as the total number of sound channels;

[0071] In embodiment 1, the lightweight convolutional neural network includes: a first convolutional layer using 16 3x3 convolutional kernels with ReLU as the activation function to extract local features in the sound channel dimension and the time dimension; a first maximum pooling layer with a pooling window size of 2x2; a second convolutional layer using 8 1x1 convolutional kernels with ReLU as the activation function to realize cross-sound channel feature fusion and dimension reduction; a global average pooling layer; and a fully connected layer with Sigmoid as the activation function;

[0072] S203: respectively input each element in the anomaly confidence vector and the anomaly determination threshold into an exponential function; if the element in the anomaly confidence vector is greater than the anomaly determination threshold, the exponential function outputs 1; if the element in the anomaly confidence vector is less than or equal to the anomaly determination threshold, the exponential function outputs 0; a binary anomaly identification vector is generated and transmitted to the administrator; the difference between the all-1 vector and the anomaly confidence vector is analyzed, denoted as a weight vector, and the elements in the weight vector represent the reliability weights of each channel.

[0073] S3: constructing a physical information neural network, fusing a preset database, and analyzing the corrected sound velocity value of each channel and the two-dimensional sound velocity field distribution on the entire pipeline cross section;

[0074] In S3, the following contents are included:

[0075] S301: splicing the environment vector and the weight vector to construct an input feature vector of the physical information neural network; the loss function of the physical information neural network is composed of a data fitting loss and a physical equation constraint loss; the data fitting loss is equal to the mean square error between the predicted sound velocity value and the reference sound velocity value corresponding to the training data set of a plurality of samples; the predicted sound velocity value is the output value corresponding to the training sample in the physical information neural network; the reference sound velocity value is equal to the sound velocity true value corresponding to the training sample obtained by querying or interpolating from the pre-calibrated sound velocity and environment parameter relationship database under the known temperature and pressure conditions; the physical equation constraint loss is equal to the residual error of the Helmholtz equation or the Eikonal equation of sound wave propagation at the discrete points in the calculation domain;

[0076] S302: taking the trained physical information neural network as a sound velocity field reconstruction model, taking the input feature vector and the normalized coordinates of any point in the pipeline cross section as inputs, and outputting the sound velocity prediction value at the point; for each channel, the sound velocity field is integrated along the ultrasonic wave propagation path to calculate the average sound velocity on the ultrasonic wave propagation path, denoted as the corrected sound velocity value of each channel, and the two-dimensional sound velocity field distribution map on the pipeline cross section is generated.

[0077] S4: training an online sequence extreme learning machine model based on the corrected sound velocity value and the sound velocity field distribution, combining historical measurement data; and generating an optimal flow rate estimation value by weighted fusion based on the final fusion weight of each channel and the corresponding channel flow rate.

[0078] In S4, the following contents are included:

[0079] S401: taking the environment vector, the corrected sound velocity value vector, and the weight vector as the input feature vector of the online sequence extreme learning machine model; calculating the preliminary flow rate value V raw ;

[0080]

[0081] wherein w i represents the reliability weight of the ith acoustic channel; L i represents the ultrasonic propagation path length of the ith acoustic channel; t i pre represents the pre-flight time of the ith acoustic channel, θ i represents the included angle between the ith acoustic channel axis and the pipeline axis; c i represents the corrected sound speed value of the ith acoustic channel;

[0082] the deviation of the preliminary flow rate value from the optimal estimated flow rate value of the last calibration cycle is taken as the supervised learning target deviation of the online sequence extreme learning machine model; the online sequence extreme learning machine model receives new input feature vectors and target deviations in an online sequence manner, and recursively updates the connection weight matrix from the hidden layer to the output layer at each time instant;

[0083] S402: the trained online sequence extreme learning machine model is used to predict the flow rate deviation value Δv of the current system; the final fusion weight of each acoustic channel is calculated in combination with the weight vector; and the final fusion weight of the ith acoustic channel is denoted as w i final ;

[0084]

[0085] Based on the corrected sound speed value and the pre-flight time of each acoustic channel, an acoustic propagation model is constructed:

[0086]

[0087] wherein v i represents the flow rate of the ith acoustic channel; Newton iteration method or fixed point iteration method is used for flow rate solving; the final fusion weight of each acoustic channel and the corresponding acoustic channel flow rate are weighted and fused to generate a flow rate optimal estimation value, which is transmitted to the administrator.

[0088] Referring to Figure 2 , the application provides a technical scheme: an adaptive ultrasonic measurement method based on multi-acoustic channel cooperation, which comprises a data acquisition and fusion module, an acoustic channel anomaly analysis module, a sound speed field reconstruction and correction module, and an adaptive calibration module;

[0089] The data acquisition and fusion module is used for transmitting coded ultrasonic pulses in a time division multiplexing manner through an ultrasonic transducer array, and synchronously acquiring original flight time signals of each acoustic channel; meanwhile, environmental data are acquired in real time; the original flight time signals are processed, effective signal arrival pre-flight times are extracted, and a fusion data set is constructed;

[0090] The channel anomaly analysis module is configured to construct a two-dimensional channel spectrum based on a pre-flight time multi-channel time sequence, extract features based on a lightweight convolutional neural network to generate an anomaly confidence vector, and analyze the channel state based on the anomaly confidence vector.

[0091] The sound velocity field reconstruction correction module is configured to construct a physical information neural network, fuse a preset database, analyze a corrected sound velocity value of each channel and a two-dimensional sound velocity field distribution on the entire pipeline cross section, and analyze the sound velocity field distribution.

[0092] The adaptive calibration module is configured to train an online sequence extreme learning machine model based on the corrected sound velocity value and the sound velocity field distribution in combination with historical measurement data, and generate an optimal flow rate estimation value based on a final fusion weight of each channel and a corresponding channel flow rate weighted fusion.

[0093] The data acquisition and fusion module includes a data acquisition unit and a data fusion unit.

[0094] The data acquisition unit is configured to capture original flight time signals by synchronously sampling and capturing the received transducer channels through FPGA platform unified clock driving, and synchronously latch data from temperature sensors, piezoresistive pressure sensors and turbidity sensors.

[0095] The data fusion unit is configured to sort the data from the temperature sensors, piezoresistive pressure sensors and turbidity sensors according to the channel sequence number to generate an environment vector.

[0096] The PHAT normalization algorithm of generalized cross-correlation is applied to the de-noised signal and the Chirp pulse signal to calculate the cross-correlation function; the peak position of the cross-correlation function is recorded as the pre-flight time, and the pre-flight time is sorted according to the channel sequence number to generate a pre-flight time vector.

[0097] The signal-to-noise ratio of each de-noised signal is calculated and recorded as a signal quality index, and the signal quality index is sorted according to the channel sequence number to generate a signal quality index vector; the pre-flight time vector, the signal quality index vector and the environment vector are normalized and spliced to generate a fusion data set.

[0098] The channel anomaly analysis module includes a feature analysis unit and a channel state analysis unit.

[0099] The feature analysis unit is configured to arrange the continuous historical observation data of the pre-flight time vector in a sliding time window into a two-dimensional data matrix in time sequence; perform bilinear interpolation processing on the two-dimensional data matrix to generate a grayscale image, recorded as a channel spectrum; input the channel spectrum into a pre-trained lightweight convolutional neural network to output an anomaly confidence vector with the same dimension as the total number of channels.

[0100] The sound channel state analysis unit is configured to input each element in the anomaly confidence vector into an exponential function respectively with an anomaly determination threshold; generate a binary anomaly identification vector and transmit the binary anomaly identification vector to the administrator; and analyze the difference between the all-1 vector and the anomaly confidence vector.

[0101] The sound speed field reconstruction correction module comprises a neural network construction unit and a sound speed correction unit.

[0102] The neural network construction unit is configured to construct a physical information neural network, and a loss function of the physical information neural network is composed of a data fitting loss and a physical equation constraint loss weighted together; the data fitting loss is equal to a mean square error between predicted sound speed values and reference sound speed values corresponding to a plurality of sample training data sets; and the physical equation constraint loss is equal to a residual error of a Helmholtz equation or an Eikonal equation of sound wave propagation at discrete points in a calculation domain.

[0103] The sound speed correction unit is configured to use the trained physical information neural network as a sound speed field reconstruction model, input a feature vector and normalized coordinates of any point in a pipeline cross section as inputs, and output a predicted sound speed value of the point; for each sound channel, integrate the sound speed field along an ultrasonic wave propagation path to calculate an average sound speed on the ultrasonic wave propagation path, denoted as a corrected sound speed value of each sound channel, and integrate to generate a two-dimensional sound speed field distribution map on the pipeline cross section.

[0104] The adaptive calibration module comprises an online learning unit and a calibration unit.

[0105] The online learning unit uses an environment vector, a corrected sound speed value vector and a weight vector as input feature vectors of an online sequence extreme learning machine model; calculates a preliminary flow rate value based on the corrected sound speed value vector and a pre-flight time vector using a weighted average method; uses a deviation of the preliminary flow rate value from an optimal estimated flow rate value of a previous calibration cycle as a supervised learning target deviation of the online sequence extreme learning machine model; and the online sequence extreme learning machine model receives new input feature vectors and target deviations in an online sequence manner at each time instant, and recursively updates a connection weight matrix from a hidden layer to an output layer.

[0106] The calibration unit is configured to use the trained online sequence extreme learning machine model to predict a flow rate deviation value of a current system, calculate a final fusion weight of each sound channel in combination with the weight vector; construct a sound propagation model based on the corrected sound speed value of each sound channel and the pre-flight time; solve the flow rate using a Newton iteration method or a fixed point iteration method; and generate an optimal estimated flow rate value based on the final fusion weight of each sound channel and a corresponding sound channel flow rate weighted fusion, and transmit the optimal estimated flow rate value to the administrator.

[0107] It is to be understood that the phrases such as first and second, and the like, can refer to different entities or operations without necessarily implying any actual relationship or order between such entities or operations. Also, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0108] It will be apparent to those skilled in the art that the application is not limited to the details of the foregoing exemplary embodiments and that the present application can be implemented in other specific forms without departing from the spirit or essential characteristics of the present application. The present embodiments are therefore to be considered in all respects as illustrative and not restrictive, the scope of the application being indicated by the appended claims rather than by the foregoing description, and all changes which come within the meaning and range of equivalency of the claims are therefore intended to be embraced therein. No feature of the application is considered critical unless expressly stated in the claims.

Claims

1. An adaptive ultrasonic measurement method based on multi-channel coordination, characterized in that, The method includes the following steps: S1: Encoded ultrasonic pulses are transmitted in a time-division multiplexing manner through an ultrasonic transducer array, and the original time-of-flight signals of each channel are acquired simultaneously; environmental data is acquired in real time at the same time; the original time-of-flight signals are processed to extract the effective signals arriving at the pre-flight time and to construct a fusion dataset; S2: A two-dimensional channel spectrum is constructed based on the multi-channel time series of the pre-flight time, and anomaly confidence vector is generated by feature extraction based on a lightweight convolutional neural network. Vocal tract status is analyzed based on anomaly confidence vector; S3: Construct a physical information neural network, integrate a preset database, and analyze the corrected sound velocity value of each channel and the two-dimensional sound velocity field distribution on the entire pipe cross section; S4: Based on the corrected sound velocity value and sound velocity field distribution, combined with historical measurement data, train an online sequence extreme learning machine model; generate the optimal flow velocity estimate based on the final fusion weight of each channel and the corresponding channel flow velocity weighted fusion.

2. The adaptive ultrasonic measurement method based on multi-channel coordination according to claim 1, characterized in that: S1 includes the following: S101: An orthogonally coded linear Chirp pulse signal is emitted by an ultrasonic transducer array. The orthogonal coding employs at least one coding scheme, selected from Walsh sequences, Gold sequences, pseudo-random phase codes, or orthogonal frequency division schemes. The original time-of-flight signal is captured by synchronous sampling driven by a unified clock on the receiving transducer channels through an FPGA platform. The FPGA platform generates a unified timestamp in each transmission and reception cycle and synchronously latches data from temperature sensors, piezoresistive pressure sensors, and turbidity sensors. An environmental vector is generated by sorting the data according to the channel number. S102: After bandpass filtering, wavelet packet denoising is performed on each original time-of-flight signal. The wavelet packet denoising is decomposed into several layers using Symlets wavelet basis as the basis function, and the denoised signal is reconstructed. The PHAT normalization algorithm with generalized cross-correlation is applied to the denoised signal and the Chirp pulse signal to calculate the cross-correlation function. The peak position of the cross-correlation function is recorded as the pre-flight time, and the pre-flight time is sorted according to the channel number to generate a pre-flight time vector. S103: Calculate the signal-to-noise ratio of each denoised signal, denoted as the signal quality index, sort the signal quality index according to the channel number to generate a signal quality index vector; normalize the pre-flight time vector, signal quality index vector and environment vector and concatenate them to generate a fused dataset.

3. The adaptive ultrasonic measurement method based on multi-channel coordination according to claim 2, characterized in that: S2 includes the following: S201: Arrange the continuous historical observation data of the pre-flight time vector within the sliding time window into a two-dimensional data matrix in chronological order; The total number of elements in the two-dimensional data matrix is ​​equal to the total number of channels multiplied by the preset sliding time window length; bilinear interpolation is performed on the two-dimensional data matrix to generate a grayscale image, denoted as a channel spectrum; where the row coordinates of the channel spectrum correspond to the channel index, the column coordinates correspond to the time series, and the pixel grayscale value represents the preflight time of the corresponding channel at the corresponding time. S202: Input the duct spectrogram into a pre-trained lightweight convolutional neural network and output an anomaly confidence vector with the same dimension as the total number of ducts; S203: Input each element of the anomaly confidence vector and the anomaly judgment threshold into the exponential function respectively; if the element in the anomaly confidence vector is greater than the anomaly judgment threshold, the exponential function outputs 1; if the element in the anomaly confidence vector is less than or equal to the anomaly judgment threshold, the exponential function outputs 0; generate a binary anomaly identifier vector and transmit it to the administrator; analyze the difference between the all-1 vector and the anomaly confidence vector, and denote it as the weight vector. The elements in the weight vector represent the reliability weight of each channel.

4. The adaptive ultrasonic measurement method based on multi-channel coordination according to claim 3, characterized in that: S3 includes the following: S301: Concatenate the environment vector and weight vector to construct the input feature vector of the physical information neural network; the loss function of the physical information neural network is composed of a weighted average of the data fitting loss and the physical equation constraint loss; the data fitting loss is equal to the mean square error between the predicted sound velocity value and the reference sound velocity value corresponding to the training dataset of several samples; the predicted sound velocity value is the output value corresponding to the training sample in the physical information neural network. The reference sound velocity value is equal to the true sound velocity value corresponding to the training sample obtained by querying or interpolating from a pre-calibrated database of sound velocity and environmental parameter relationships under known temperature and pressure conditions; the physical equation constraint loss is equal to the residual of the Helmholtz equation or Eikonal equation for sound wave propagation at discrete points in the computational domain. S302: The trained physical information neural network is used as the sound velocity field reconstruction model. The input feature vector and the normalized coordinates of any point in the pipe cross section are used as inputs, and the predicted sound velocity value at that point is output. For each channel, the sound velocity field is integrated along the ultrasonic wave propagation path to calculate the average sound velocity on the ultrasonic wave propagation path. This is recorded as the corrected sound velocity value for each channel, and the data is integrated to generate a two-dimensional sound velocity field distribution map on the pipe cross section.

5. The adaptive ultrasonic measurement method based on multi-channel coordination according to claim 4, characterized in that: S4 includes the following: S401: The environment vector, the corrected sound velocity vector, and the weight vector are used as the input feature vectors of the online sequential extreme learning machine model; the initial flow velocity value v is calculated using a weighted average method based on the corrected sound velocity vector and the pre-flight time vector. raw ; The deviation between the initial flow velocity value and the optimal estimated flow velocity value of the previous calibration period is used as the supervised learning target deviation of the online sequence extreme learning machine model. The online sequence extreme learning machine model receives new input feature vectors and target deviations in an online sequence manner, and continuously recursively updates the connection weight matrix from the hidden layer to the output layer. S402: Predict the current system's flow velocity deviation value Δv using a trained online sequence extreme learning machine model; calculate the final fusion weights for each channel using the weight vector; Based on the corrected sound velocity value and pre-flight time of each channel, a sound propagation model is constructed; the flow velocity is solved using the Newton iteration method or the fixed-point iteration method; the optimal flow velocity estimate is generated by weighted fusion based on the final fusion weight of each channel and the corresponding channel flow velocity, and the optimal flow velocity estimate is transmitted to the administrator.

6. An adaptive ultrasonic measurement method based on multi-channel collaboration, wherein the system is applied to the adaptive ultrasonic measurement system based on multi-channel collaboration as described in any one of claims 1-5, characterized in that, The system includes a data acquisition and fusion module, a duct anomaly analysis module, a sound velocity field reconstruction and correction module, and an adaptive calibration module; The data acquisition and fusion module is used to transmit coded ultrasonic pulses in a time-division multiplexing manner through an ultrasonic transducer array and simultaneously acquire the original time-of-flight signals of each channel; at the same time, it acquires environmental data in real time; processes the original time-of-flight signals, extracts the effective signals arriving at the pre-flight time, and constructs a fusion dataset. The duct anomaly analysis module is used to construct a two-dimensional duct spectral map based on the multi-channel time series of the pre-flight time, and to extract features based on a lightweight convolutional neural network to generate an anomaly confidence vector. Vocal tract status is analyzed based on anomaly confidence vector; The sound velocity field reconstruction and correction module is used to construct a physical information neural network, integrate a preset database, and analyze the corrected sound velocity value of each channel and the two-dimensional sound velocity field distribution on the entire pipe cross section. The adaptive calibration module is used to train an online sequence extreme learning machine model based on the corrected sound velocity value and sound velocity field distribution, combined with historical measurement data; and to generate the optimal flow velocity estimate based on the final fusion weight of each channel and the corresponding channel flow velocity weighted fusion.

7. The adaptive ultrasonic measurement system based on multi-channel coordination according to claim 6, characterized in that: The data acquisition and fusion module includes a data acquisition unit and a data fusion unit; The data acquisition unit is used to synchronously sample and capture the original time-of-flight signal by uniform clock driving the receiving transducer channel through the FPGA platform, and synchronously latch the data from the temperature sensor, piezoresistive pressure sensor and turbidity sensor. The data fusion unit is used to sort the data from the temperature sensor, piezoresistive pressure sensor and turbidity sensor according to the channel number to generate an environmental vector. The PHAT normalization algorithm with generalized cross-correlation is applied to the denoised signal and the Chirp pulse signal to calculate the cross-correlation function; the peak position of the cross-correlation function is recorded as the pre-flight time, and the pre-flight time is sorted according to the channel number to generate a pre-flight time vector; The signal-to-noise ratio of each denoised signal is calculated and denoted as the signal quality index. The signal quality index is sorted according to the channel number to generate a signal quality index vector. The pre-flight time vector, signal quality index vector and environment vector are normalized and concatenated to generate a fused dataset.

8. The adaptive ultrasonic measurement system based on multi-channel coordination according to claim 6, characterized in that: The vocal tract anomaly analysis module includes a feature analysis unit and a vocal tract state analysis unit; The feature analysis unit is used to arrange the continuous historical observation data of the pre-flight time vector within the sliding time window into a two-dimensional data matrix in chronological order; perform bilinear interpolation on the two-dimensional data matrix to generate a grayscale image, denoted as a channel spectrum map; input the channel spectrum map into a pre-trained lightweight convolutional neural network, and output an anomaly confidence vector with the same dimension as the total number of channels; The vocal tract state analysis unit is used to input each element in the abnormal confidence vector and the abnormal judgment threshold into the exponential function respectively. Generate a binary exception identifier vector and transmit it to the administrator; Analyze the difference between the all-1 vector and the outlier confidence vector.

9. The adaptive ultrasonic measurement system based on multi-channel coordination according to claim 6, characterized in that: The sound velocity field reconstruction and correction module includes a neural network construction unit and a sound velocity correction unit; The neural network building unit is used to construct a physical information neural network. The loss function of the physical information neural network is composed of a weighted average of data fitting loss and physical equation constraint loss. The data fitting loss is equal to the mean square error between the predicted sound velocity value and the reference sound velocity value corresponding to a training dataset of several samples. The physical equation constraint loss is equal to the residual of the Helmholtz equation or Eikonal equation for sound wave propagation at discrete points in the computational domain. The sound velocity correction unit is used as a sound velocity field reconstruction model by taking the trained physical information neural network as input and the normalized coordinates of any point in the pipe cross section as input, and outputting the predicted sound velocity value at that point; for each channel, the sound velocity field is integrated along the ultrasonic wave propagation path to calculate the average sound velocity on the ultrasonic wave propagation path, which is recorded as the corrected sound velocity value for each channel, and integrated to generate a two-dimensional sound velocity field distribution map on the pipe cross section.

10. The adaptive ultrasonic measurement system based on multi-channel collaboration according to claim 6, characterized in that: The adaptive calibration module includes an online learning unit and a calibration unit; The online learning unit uses the environment vector, the corrected sound velocity value vector, and the weight vector as the input feature vectors of the online sequential extreme learning machine model; it calculates the preliminary flow velocity value using a weighted average method based on the corrected sound velocity value vector and the pre-flight time vector; and it uses the deviation between the preliminary flow velocity value and the optimal estimated flow velocity value of the previous calibration cycle as the supervised learning target deviation of the online sequential extreme learning machine model. The online sequence extreme learning machine model receives new input feature vectors and target biases in an online sequence manner, and continuously recursively updates the connection weight matrix from the hidden layer to the output layer. The calibration unit is used to predict the current system's flow velocity deviation value using a trained online sequence extreme learning machine model, and calculate the final fusion weight of each channel by combining the weight vector. Based on the corrected sound velocity value and pre-flight time of each channel, a sound propagation model is constructed; the flow velocity is solved using the Newton iteration method or the fixed-point iteration method; the optimal flow velocity estimate is generated by weighted fusion based on the final fusion weight of each channel and the corresponding channel flow velocity, and the optimal flow velocity estimate is transmitted to the administrator.

Images