Ultrasonic testing method for defects of external delivery hose based on reciprocity difference and hybrid neural network

Abstract

The application relates to the technical field of ultrasonic detection and noise reduction, and provides an ultrasonic detection method for defects of an external delivery hose based on reciprocity difference and a hybrid neural network, which comprises the following steps: filtering an initial signal to obtain a first signal; decomposing the first signal into an approximate coefficient and a first detail coefficient; performing soft threshold processing on the first detail coefficient to obtain a second detail coefficient; reconstructing the second detail coefficient and the first approximate coefficient to obtain a second signal; performing normalization processing on the second signal to obtain a third signal; inputting the third signal into a convolutional neural network model to obtain a spatial feature vector; inputting the spatial feature vector into a recurrent neural network model to extract a time feature vector; fusing the spatial feature vector and the time feature vector to obtain a noise reduction signal; and integrating the convolutional neural network model and the recurrent neural network model with an ultrasonic detection system after training, so as to output a defect feature signal and improve the precision and detection efficiency of online detection of the external delivery hose.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of ultrasonic detection and noise reduction, and particularly relates to an ultrasonic detection method for defects of an external delivery hose based on a reciprocity difference and a hybrid neural network. BACKGROUND

[0002] Rubber hoses are widely used in the petroleum, chemical and other industries, and their quality and safety directly affect the stability and safety of production. Ultrasonic detection is one of the effective methods for detecting defects in rubber hoses. However, due to the complexity of the material of the rubber hose and the diversity of the use environment, various noises often interfere with the ultrasonic detection process, affecting the accuracy and reliability of the detection. Therefore, how to effectively reduce the noise in ultrasonic detection has become a problem to be solved.

[0003] Existing ultrasonic detection methods often encounter environmental noise, material self-noise, and detection equipment noise when facing rubber hoses. The existence of these noises reduces the signal-to-noise ratio of the detection signal, increases the uncertainty of the detection result, and affects the accurate judgment of defects. In addition, existing noise reduction methods are not effective in dealing with complex noise, and cannot meet the needs of high-precision detection. SUMMARY

[0004] The present application aims to at least solve one of the technical problems in the related art. To this end, the present application provides an ultrasonic detection method for defects of an external delivery hose based on a reciprocity difference and a hybrid neural network, which realizes improving the precision and detection efficiency of online detection of the external delivery hose.

[0005] The present application provides an ultrasonic detection method for defects of an external delivery hose based on a reciprocity difference and a hybrid neural network, and the construction of a hybrid neural network model includes the following steps:

[0006] S1: An ultrasonic detection system detects an initial signal, and filters the initial signal to obtain a first signal;

[0007] S2: The first signal is decomposed into an approximation coefficient and a first detail coefficient;

[0008] S3: The first detail coefficient is subjected to soft threshold processing to obtain a second detail coefficient, and the second detail coefficient and the approximation coefficient are reconstructed to obtain a second signal;

[0009] S4: The second signal is subjected to normalization processing to obtain a third signal;

[0010] S5: The third signal is input into a convolutional neural network model to obtain a spatial feature vector of the third signal;

[0011] S6: input the spatial feature vector into a recurrent neural network model to extract a time feature vector of the third signal;

[0012] S7: fuse the spatial feature vector and the time feature vector to obtain a fusion feature vector, and obtain a denoising signal according to the fusion feature vector;

[0013] S8: obtain a hybrid neural network model by jointly training the convolutional neural network model and the recurrent neural network model, integrate the trained hybrid neural network model into the ultrasonic detection system, and output a defect feature signal according to the denoising signal. According to the ultrasonic detection method for defects of an external delivery hose provided by the application, an initial signal is filtered to obtain a first signal, including the following steps:

[0014] The initial signal is processed by an adaptive filtering algorithm, and the adaptive filtering algorithm includes a least mean square error algorithm (LMS algorithm), and the formula is as follows:

[0015]

[0016] wherein, y ( t ) is a first signal output by a filter at time point n, t ( w ) is a filter coefficient vector of the filter at time point n, t ( t ) is an initial signal vector input by the filter at time point n, x = 0, 1, 2, 3, …; t t t The difference between the expected signal and the output first signal at time point n is calculated according to the following formula:

[0017] t

[0018]

[0019] wherein, d ( t ) is an expected signal of the filter at time point n, t ( e ) is the difference between the expected signal and the output first signal of the filter at time point n; t t

[0020] The filter coefficient of the adaptive filtering algorithm is dynamically adjusted according to the calculated difference between the expected signal and the output first signal, and the formula of the filter coefficient is as follows:

[0021] ​​​​​

[0022] wherein, w t is the filter coefficient vector at time n, t is the filter coefficient vector at time n+1, μ is a learning factor used to control the speed of filter coefficient update.

[0023] According to the method for ultrasonic detection of defects of an export hose based on reciprocity difference and a hybrid neural network provided by the application, decomposing the first signal into approximation coefficients and first detail coefficients comprises the following steps: performing continuous wavelet transform on the first signal, and decomposing the first signal into approximation coefficients and first detail coefficients of different scales through wavelet transform.

[0024] According to the method for ultrasonic detection of defects of an export hose based on reciprocity difference and a hybrid neural network provided by the application, obtaining the second signal comprises the following steps:

[0025] performing soft threshold processing on the first detail coefficients to obtain second detail coefficients, and using the following formula:

[0026]

[0027] wherein, is the first approximation coefficient of the first signal at time n, t is the first approximation coefficient of the first signal at time n+1, is the first approximation coefficient of the first signal at time n+1, is the first approximation coefficient of the first signal at time n+1, is the first approximation coefficient of the first signal at time n, t is the first approximation coefficient of the first signal at time n+1, is the first approximation coefficient of the first signal at time n+1, is a threshold value of the first detail coefficient;

[0028] reconstructing the second detail coefficients and the approximation coefficients, comprising the following steps:

[0029]

[0030] wherein, is the second signal at time n, t is the second signal at time n+1, is the first approximation coefficient of the first signal at time n, is the first approximation coefficient of the first signal at time n+1.

[0031] According to the method for ultrasonic detection of defects of an export hose based on reciprocity difference and a hybrid neural network provided by the application, obtaining the third signal comprises the following steps:

[0032] performing normalization processing on the second signal, and using the following formula:

[0033] ​

[0034] wherein, is t a third signal at the moment, is the minimum value of the second signal, is the maximum value of the second signal.

[0035] According to the application, a method for ultrasonic detection of defects in an outflow hose based on reciprocity difference and a hybrid neural network is provided, and the spatial feature vector of the third signal comprises the following steps:

[0036] The third signal is input into a convolutional neural network model, and a convolutional layer of the convolutional neural network model extracts local features of the third signal through a plurality of groups of convolutional kernels.

[0037] A pooling layer of the convolutional neural network model adopts a maximum pooling or average pooling method to reduce the feature map dimension of the local features.

[0038] A fully connected layer of the convolutional neural network model performs feature fusion and outputs the spatial feature vector of the third signal.

[0039] According to the application, a method for ultrasonic detection of defects in an outflow hose based on reciprocity difference and a hybrid neural network is provided, and the time feature vector of the third signal comprises the following steps:

[0040] The spatial feature vector is input into the recurrent neural network model, a long short-term memory network is adopted to capture the time dependence in the third signal, and the time feature vector of the third signal is obtained.

[0041] According to the application, a method for ultrasonic detection of defects in an outflow hose based on reciprocity difference and a hybrid neural network is provided, and the fusion of the spatial feature vector and the time feature vector comprises the following steps:

[0042] The spatial feature vector and the time feature vector are spliced into a fusion feature vector, and the formula is as follows:

[0043]

[0044] wherein, is the spatial feature vector, is the time feature vector, is the fusion feature vector.

[0045] According to the application, a method for ultrasonic detection of defects in an outflow hose based on reciprocity difference and a hybrid neural network is provided, and the joint training of the convolutional neural network model and the recurrent neural network model to obtain a hybrid neural network model comprises the following steps:

[0046] Collecting the ultrasonic detection signals of the defect-free external conveying hose and the plurality of defective external conveying hoses to form a data set, repeating steps S1-S7 for each set of detection signals in the data set, and jointly training the convolutional neural network model and the recurrent neural network model to obtain a hybrid neural network model.

[0047] According to the external conveying hose defect ultrasonic detection method based on reciprocity difference and hybrid neural network provided by the application, the data set is established by the following steps:

[0048] A plurality of defective external conveying hoses are provided, and a first excitation probe and a first detection probe are arranged on the defect-free external conveying hose, and a second detection probe and a second excitation probe are arranged on the defective external conveying hose, the position of the first excitation probe corresponds to the position of the second detection probe, and the position of the first detection probe corresponds to the position of the second excitation probe, the first excitation probe emits a first excitation signal, the first detection probe receives and processes the first excitation signal to obtain a first ultrasonic detection signal, the second excitation probe emits a second excitation signal, and the second detection probe receives and processes the second excitation signal to obtain a second ultrasonic detection signal, and the first ultrasonic detection signal and the plurality of second ultrasonic detection signals constitute the data set.

[0049] The one or more technical solutions in the above embodiment of the application have at least one of the following technical effects:

[0050] According to the external conveying hose defect ultrasonic detection method based on reciprocity difference and hybrid neural network, the hybrid neural network model trained and optimized is integrated into the ultrasonic detection system, the problem of the influence of various types of noise on the accuracy of detection is solved, real-time online noise reduction detection is realized, and the precision and detection efficiency of online detection of the external conveying hose are improved.

[0051] Additional aspects and advantages of the application will be described in part in the description which follows, and in part will become apparent to those skilled in the art from the description, or by practicing the application. BRIEF DESCRIPTION OF DRAWINGS

[0052] In order to more clearly illustrate the technical solutions of the application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or the prior art description. Obviously, the drawings in the following description are some embodiments of the application, and those skilled in the art can also obtain other drawings according to these drawings without creative labor.

[0053] Figure 1 is a flowchart of the external conveying hose defect ultrasonic detection method based on reciprocity difference and hybrid neural network provided by the application.

[0054] Figure 2 The figure is a structural schematic diagram of a defect-free external conveying hose and a defect external conveying hose of a method for detecting defects of an external conveying hose based on reciprocity difference and a hybrid neural network provided by the application.

[0055] Figure 3 The figure is an output signal of a defect-free external conveying hose and a defect external conveying hose of a method for detecting defects of an external conveying hose based on reciprocity difference and a hybrid neural network provided by the application. DETAILED DESCRIPTION

[0056] In order to make the objects, technical solutions and advantages of the present application clearer, the technical solutions in the present application will be clearly and completely described below. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the protection scope of the present application. The following embodiments are used to illustrate the present application, but cannot be used to limit the scope of the present application.

[0057] In the description of the embodiments of the present application, it should be noted that the orientations or positional relationships indicated by the terms "center", "longitudinal", "lateral", "upper", "lower", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer" and the like are based on the orientations or positional relationships shown in the drawings, and are only for the convenience of describing the embodiments of the present application and simplifying the description, and therefore cannot be understood as indicating or implying that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the embodiments of the present application. In addition, the terms "first", "second", "third" are only for the purpose of description, and cannot be understood as indicating or implying relative importance.

[0058] In the description of the present application, the description of the terms "one embodiment", "some embodiments", "an example", "a specific example", or "some examples" means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the present application, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any appropriate manner in any one or more embodiments or examples. In addition, those skilled in the art can combine and combine the different embodiments or examples described in the present application and the features of the different embodiments or examples without contradiction.

[0059] Figure 1 The figure is a flowchart of a method for detecting defects of an external conveying hose based on reciprocity difference and a hybrid neural network provided by the application. The figure is a flowchart of a method for detecting defects of an external conveying hose based on reciprocity difference and a hybrid neural network provided by the application.

[0060] This invention provides an ultrasonic testing method for defects in export hoses based on reciprocity and a hybrid neural network, such as... Figure 1 As shown, it includes the following steps:

[0061] S1: The ultrasonic testing system detects an initial signal and filters the initial signal to obtain a first signal;

[0062] S2: Decompose the first signal into approximation coefficients and first detail coefficients;

[0063] S3: Perform soft thresholding on the first detail coefficients to obtain the second detail coefficients, and reconstruct the second detail coefficients with the approximation coefficients to obtain the second signal;

[0064] S4: Normalize the second signal to obtain the third signal;

[0065] S5: Input the third signal into the convolutional neural network model to obtain the spatial feature vector of the third signal;

[0066] S6: Input the spatial feature vector into the recurrent neural network model to extract the temporal feature vector of the third signal;

[0067] S7: The spatial feature vector and the temporal feature vector are fused to obtain a fused feature vector, and a noise reduction signal is obtained based on the fused feature vector;

[0068] S8: A hybrid neural network model is obtained by jointly training the convolutional neural network model and the recurrent neural network model. The trained hybrid neural network model is integrated into the ultrasonic detection system, and a defect feature signal is output based on the noise reduction signal.

[0069] In this embodiment, by integrating a trained and optimized hybrid neural network model into the ultrasonic testing system, the problem of various types of noise affecting the accuracy of testing is solved, and real-time online noise reduction testing is achieved, thereby improving the accuracy and efficiency of online testing of export hoses.

[0070] According to the present invention, an ultrasonic testing method for defects in export hoses based on reciprocity and hybrid neural networks includes the following steps: filtering the initial signal to obtain a first signal.

[0071] The initial signal is processed by an adaptive filtering algorithm, which includes the minimum mean square error (LMS) algorithm, and its formula is as follows:

[0072]

[0073] in, y (t ) is the filter in t The first signal output at time 10:00 w ( t ) is a filter t The filter coefficient vector at time t. x ( t ) is the filter in t The initial signal vector of the input at time t. t =0, 1, 2, 3, ...;

[0074] t The formula for calculating the difference between the expected signal and the first output signal at time t is as follows:

[0075]

[0076] in, d ( t ) is the filter in t Expected signal at time, e ( t ) is the filter in t The difference between the expected signal at time t and the first output signal;

[0077] Based on the difference between the calculated desired signal and the first output signal, the filter coefficients of the adaptive filtering algorithm are dynamically adjusted. The formula for the filter coefficients is as follows:

[0078]

[0079] in, w ( t +1) is the filter t The filter coefficient vector at time +1 μ It is a learning factor used to control the speed at which the filter coefficients are updated. Used to adjust filter coefficients.

[0080] In this embodiment, the signal after adaptive filtering has a high signal-to-noise ratio, providing a high-quality data foundation for subsequent defect detection.

[0081] According to some embodiments of the present invention, adaptive filtering processing involves using an adaptive filtering algorithm to monitor the ultrasonic detection signal in real time and dynamically adjust the filter coefficients to suppress environmental noise and equipment noise.

[0082] According to some preferred embodiments of the present application, the initial signal is first filtered by a filter with initial parameters to preliminarily suppress ambient noise and equipment noise and improve the signal-to-noise ratio of the signal. Then, the filtered signal is input into the LMS algorithm to determine the expected signal and calculate the difference between the expected signal and the output first signal. According to the calculated difference between the expected signal and the output first signal, the adaptive filter algorithm dynamically adjusts the filter coefficients to minimize the difference between the expected signal and the output first signal. The adjustment process can use the gradient descent method to realize the iterative update of the filter coefficients. The above filtering process and parameter adjustment process are repeated until the preset filtering effect or the difference between the expected signal and the output first signal is minimized. The signal after adaptive filtering has a high signal-to-noise ratio, providing a high-quality data basis for subsequent wavelet transform and defect detection.

[0083] According to some preferred embodiments of the present application, the formula for calculating the signal-to-noise ratio is as follows:

[0084]

[0085]

[0086]

[0087] wherein, is the initial signal power, is the total number of signal sampling points, is the original signal, is the signal after noise reduction is the noise signal power, is the noise signal, SNR (dB) is the signal-to-noise ratio, and the larger the SNR (dB) is, the better the noise reduction effect is and the higher the signal quality is.

[0088] According to the ultrasonic detection method for defects of an export hose based on reciprocity difference and a hybrid neural network provided by the present application, decomposing the first signal into approximation coefficients and first detail coefficients comprises the following steps: performing continuous wavelet transform (CWT) on the first signal, and decomposing the first signal into approximation coefficients and first detail coefficients of different scales through wavelet transform.

[0089] In this embodiment, the first signal is decomposed into approximation coefficients and first detail coefficients of different scales to analyze the characteristics of the signal in different frequency and time ranges.

[0090] According to some preferred embodiments of the present application, the approximation coefficients represent the characteristics of the first signal in the low-frequency scale, and the first detail coefficients represent the characteristics of the first signal in the high-frequency scale.

[0091] According to the present invention, an ultrasonic detection method for defects in export hoses based on reciprocity and hybrid neural networks is provided, and obtaining the second signal includes the following steps:

[0092] The first detail coefficients are subjected to soft thresholding to obtain the second detail coefficients, using the following formula:

[0093]

[0094] in, for t The first signal at time 1 The first detail coefficient of level 1 is obtained after the soft thresholding process. Level 2 detail coefficient, for t The first signal at time 1 Level 1 detail coefficient, The threshold value for the first detail coefficient;

[0095] Reconstructing the second detail coefficients and the approximation coefficients includes the following steps:

[0096]

[0097] in, for t The second signal at time, The first signal represents the first signal. Approximation coefficients of the order.

[0098] In this embodiment, by performing soft thresholding on the obtained first detail coefficients, the first detail coefficients larger than the threshold are smoothly reduced to suppress noise and preserve important features in the signal, ensuring a natural transition of coefficients when they are near the threshold close to zero.

[0099] According to some embodiments of the present invention, coefficient spectra are obtained after CWT transformation, the time-frequency characteristics of the signal are analyzed, feature quantities related to defects, such as energy, frequency distribution, instantaneous phase, etc., are extracted, and a defect identification model is constructed to determine the location and nature of defects in the signal.

[0100] According to the present invention, an ultrasonic detection method for defects in export hoses based on reciprocity and hybrid neural networks is provided, and obtaining the third signal includes the following steps:

[0101] The second signal is normalized using the following formula:

[0102]

[0103] in, for ta third signal at the time instant, a minimum value of the second signal, a maximum value of the second signal.

[0104] In the embodiment, the signal after the wavelet transform processing is converted into a data format suitable for neural network processing through normalization processing, the ultrasonic detection signal is normalized to scale the amplitude range to [0, 1], so as to improve the training effect of the neural network.

[0105] According to the ultrasonic detection method for defects of an outflow hose based on a reciprocity difference and a hybrid neural network provided by the application, obtaining the spatial feature vector of the third signal comprises the following steps:

[0106] The third signal is input into a convolutional neural network model (CNN), and a convolutional layer of the convolutional neural network model extracts local features of the third signal through a plurality of groups of convolutional kernels.

[0107] A pooling layer of the convolutional neural network model adopts a maximum pooling or average pooling method to reduce the feature map dimension of the local features.

[0108] A full connection layer of the convolutional neural network model performs feature fusion and outputs the spatial feature vector of the third signal.

[0109] In the embodiment, the spatial feature vector of the third signal is extracted for subsequent identification of the position of the defect.

[0110] According to the ultrasonic detection method for defects of an outflow hose based on a reciprocity difference and a hybrid neural network provided by the application, obtaining the time feature vector of the third signal comprises the following steps:

[0111] The spatial feature vector is input into the recurrent neural network model (RNN), a long short-term memory network is adopted to capture the time dependence in the third signal, and the time feature vector of the third signal is obtained.

[0112] In the embodiment, the time feature vector of the third signal is extracted, the suppression effect on noise is improved, and the nature and severity of the defect are facilitated to be detected.

[0113] According to some embodiments of the application, the long short-term memory network solves the gradient disappearance problem, thereby improving the learning ability, the RNN processes dynamic changes through an internal memory unit, extracts time sequence features, and improves the suppression effect on noise.

[0114] According to the ultrasonic detection method for defects of an outflow hose based on a reciprocity difference and a hybrid neural network provided by the application, fusing the spatial feature vector and the time feature vector comprises the following steps:

[0115] The spatial feature vector and the time feature vector are spliced into a fusion feature vector, and the formula is as follows:

[0116]

[0117] wherein, is a spatial feature vector, is a time feature vector, is a fusion feature vector.

[0118] In the embodiment, by splicing the spatial feature vector and the time feature vector into a fusion feature vector, it is ensured that the final feature can effectively represent the characteristics of the signal.

[0119] According to some embodiments of the application, the output features of the CNN and the RNN are fused, they are combined by using a fully connected layer, the spliced high-dimensional feature vector is further mapped and fused, the final denoising signal is obtained, and the denoising effect is evaluated by calculating the signal-to-noise ratio (SNR) and the like, so as to ensure the quality improvement of the final denoising signal.

[0120] According to some embodiments of the application, obtaining the final denoising signal comprises the following steps:

[0121] The fused feature vector is transmitted into one or more fully connected layers to realize further feature fusion and dimension compression. The function of the fully connected layer is to linearly combine all the input features, and then realize nonlinear mapping through an activation function, and finally obtain the final denoising signal. The mathematical form of the fully connected layer can be expressed as:

[0122]

[0123] wherein, is a weight matrix of the fully connected layer, is a bias vector, is an activation function, is a final denoising signal, that is, an output feature after processing by the fully connected layer.

[0124] According to the method for ultrasonic detection of defects of an export hose provided by the application, the convolutional neural network model and the recurrent neural network model are jointly trained to obtain a hybrid neural network model, and the method comprises the following steps:

[0125] The ultrasonic detection signals of a defect-free export hose and a plurality of defective export hoses are collected to form a data set, steps S1-S7 are repeated for each group of detection signals in the data set, and the convolutional neural network model and the recurrent neural network model are jointly trained to obtain a hybrid neural network model.

[0126] In this embodiment, the hybrid neural network model is obtained by jointly training the convolutional neural network model and the recurrent neural network model, thereby improving the performance and applicability of the hybrid neural network model in the ultrasonic denoising detection of rubber hoses.

[0127] According to some preferred embodiments of the present application, during the training of the hybrid neural network model, the data set is divided into a training data set and a validation data set using a cross-validation method, the model parameters are optimized using a backpropagation algorithm, and the mean square error is used to improve the performance of the hybrid neural network model. In order to prevent overfitting, a regularization strategy and an early stopping mechanism are implemented to ensure that the performance of the hybrid neural network model on the validation data set is consistent with that on the training data set.

[0128] According to some preferred embodiments of the present application, the optimization of the hybrid neural network model is a continuous process. After preliminary verification, the hyperparameters of the hybrid neural network model, including learning rate, batch size, network structure, etc., are continuously adjusted according to the verification results. Through strategies such as grid search and random search, the best combination of hyperparameters is found to improve the denoising effect of the hybrid neural network model.

[0129] According to some preferred embodiments of the present application, the trained hybrid neural network model is integrated into the ultrasonic detection system to realize real-time online denoising detection. The hybrid neural network model can continue to learn in actual detection, combined with user feedback, and continuously optimized and adjusted to adapt to the complex and variable rubber hose detection environment. Through this dynamic learning mechanism, the system can maintain efficient denoising performance and ensure the accuracy and reliability of the ultrasonic detection results.

[0130] Figure 2 The present application provides a structure diagram of a defect-free export hose and a defect export hose based on the mutual difference and the hybrid neural network method for ultrasonic detection of defects in export hoses.

[0131] According to the present application, a data set is established based on the mutual difference and the hybrid neural network method for ultrasonic detection of defects in export hoses.

[0132] A defect-free export hose and a plurality of defect export hoses are set up, such as Figure 2As shown, the first excitation probe and the first detection probe are arranged on the defect-free external delivery hose, the second detection probe and the second excitation probe are arranged on the defective external delivery hose, the position of the first excitation probe corresponds to the position of the second detection probe, the position of the first detection probe corresponds to the position of the second excitation probe, the first excitation probe emits a first excitation signal, the first detection probe receives and processes the first excitation signal to obtain a first ultrasonic detection signal, the second excitation probe emits a second excitation signal, and the second detection probe receives and processes the second excitation signal to obtain a second ultrasonic detection signal, and the first ultrasonic detection signal and a plurality of second ultrasonic detection signals constitute the data set.

[0133] In this embodiment, by arranging the first excitation probe and the first detection probe and the second detection probe and the second excitation probe at opposite positions, the specific position of the defect can be analyzed and determined by using the acoustic reciprocity difference principle.

[0134] According to some preferred embodiments of the present application, for the defect-free external delivery hose, the first excitation probe emits a preset light signal, and the first detection probe receives the returned signal. In this process, the detection system needs to be calibrated initially to ensure the stability and accuracy of the signal transmission path. The calibration process includes adjusting the position and angle of the first excitation probe and the first detection probe to obtain the optimal signal receiving effect, and recording the intensity and phase characteristics of the reference signal.

[0135] According to some embodiments of the present application, for the defective external delivery hose, the second excitation probe emits a preset light signal, and the second detection probe receives the returned signal. At this time, it needs to be calibrated again to ensure the consistency of the signal transmission path under different excitation and detection conditions. The signal collection needs to be repeated multiple times at the same position to ensure the accuracy and stability of the data. After each signal collection, the detection system will automatically record the intensity, phase and other related characteristics of the signal, and perform preliminary comparative analysis to identify possible abnormal signal characteristics.

[0136] Figure 3 is the output signal of the defect-free external delivery hose and the defective external delivery hose of the external delivery hose defect ultrasonic detection method based on reciprocity difference and hybrid neural network provided by the present application.

[0137] According to some embodiments of the present application, as shown in Figure 3 The trained hybrid neural network model is integrated into the ultrasonic detection system, the first ultrasonic detection signal and the second ultrasonic detection signal are subjected to noise reduction processing by the hybrid neural network, and the defect position of the defective external delivery hose can be obtained by comparison.

[0138] According to some embodiments of the present application, a large amount of outflow hose ultrasonic detection data is collected and labeled to construct a training data set and a validation data set. Then, the hybrid neural network model is trained using the training data set, and the third signal is input into the hybrid neural network for intelligent noise reduction processing. First, CNN is used to extract the spatial feature vector of the signal, and the signal is feature-extracted through the convolution kernel. In this process, the ReLU activation function is used to speed up the training convergence. Then, the spatial feature vector extracted by the CNN is input into the RNN, and the RNN is used to process the time sequence features of the signal, aiming to capture the dependence of the signal on time.

[0139] According to some embodiments of the present application, in the training process of the hybrid neural network model, the training data set and the validation data set are divided by the cross-validation method, the model parameters are optimized by the backpropagation algorithm, and the performance of the hybrid neural network model is improved by the mean square error. In order to prevent overfitting, the L2 regularization strategy and the early stopping mechanism are implemented, and the performance of the model on the validation data set is ensured to be consistent with the training data set by combining the two methods.

[0140] According to some embodiments of the present application, the early stopping mechanism is a strategy for determining when to stop training by dynamically monitoring the performance of the validation data set during the training process of the model. When the model performs well on the training data set but the performance on the validation data set starts to deteriorate (i.e., the loss of the validation data set starts to rise), it indicates that the model has started to overfit. During the training process, the loss function on the validation data set is monitored, and when the loss of the validation data set no longer decreases after several iterations (i.e., exceeds the set tolerance number), the training is stopped. In this way, the model can be stopped training when it reaches the best generalization performance, avoiding the risk of overfitting caused by continued training.

[0141] According to some embodiments of the present application, the model is strictly controlled at each step of the training. L2 regularization restricts the complexity of the model by adding a penalty term to the loss function, making the model more robust when facing new data. And the early stopping mechanism detects the trend of the loss of the validation data set to dynamically determine when to stop training, to ensure that the model does not perform worse on the validation data set. Through this combination, the performance of the model on the training data set and the validation data set is more consistent, significantly improving the generalization ability and reliability in practical applications.

[0142] According to some embodiments of the present application, the optimization of the model is a continuous process. After preliminary verification, the hyperparameters of the model are adjusted according to the verification results, including learning rate, batch size, network structure, etc. The size of the learning rate determines the amplitude of the weight update in each iteration. If the learning rate is too high, the training will be unstable, and if it is too low, the convergence will be slow. The learning rate is generally set in [10 −2 ,10 −1If the loss of the validation dataset is higher than the loss of the training dataset, it means that the model is overfitting, and the number of layers or the number of neurons in each layer can be reduced. If the loss of both the recruitment dataset and the training dataset is high, it means that the complexity of the model may not be enough, and the performance of the model can be improved by increasing the number of convolutional layers or fully connected layers and the number of neurons.

[0143] According to some embodiments of the present application, the set of multiple possible configurations formed by the values of different hyperparameters during model training is called a hyperparameter combination. By selecting and optimizing these hyperparameter combinations, the noise reduction effect of the model can be improved, and the model can have good performance on both the validation dataset and the training dataset, thereby ensuring that the model has good generalization ability and application value. To improve the noise reduction effect of the model. After training and optimizing the model, the system has the function of high-precision defect detection in ultrasonic signals and can identify the specific location, type, and severity of defects while reducing noise.

[0144] According to some embodiments of the present application, in the implementation of defect detection, the model first identifies the specific location of the abnormal signal according to the spatial feature vector of the third signal. For the identification of the defect location, the system uses the spatial feature vector extracted by the CNN to accurately locate the area in the signal that shows abnormalities. In addition, the model further identifies the nature and evolution trend of the defect through the time-dependent features captured by the RNN. With this feature fusion method, the system can comprehensively analyze the characteristics of the signal and effectively distinguish different defect types, such as cracks and wear, and judge the severity of the defect according to the intensity and frequency changes of the defect signal.

[0145] According to some embodiments of the present application, finally, the entire system completes signal acquisition under the reciprocity difference principle through the connection of the probe and the data acquisition device, and transmits the signal into the data preprocessing module for noise reduction processing. The trained hybrid neural network model further extracts deep features and reduces noise of the signal to realize the real-time online defect recognition function of ultrasonic detection. After the model is integrated into the ultrasonic detection system, it can identify the specific location, type and severity of defects during the detection process, and present these defect information to the operator to support real-time decision and operation. The system not only displays the location and nature of the defects, but also provides detailed detection data, including detection time, defect location, type, severity and signal-to-noise ratio (SNR), etc. to support subsequent analysis and management of data. At this time, the model can continue to learn in actual detection, combined with user feedback, and constantly optimize and adjust to adapt to the complex and variable rubber hose detection environment. Through this dynamic learning mechanism, the system can maintain efficient noise reduction performance to ensure the accuracy and reliability of ultrasonic detection results.

[0146] The ultrasonic detection method for defects of the external delivery hose based on the reciprocity difference and the hybrid neural network has the following technical effects compared with the prior art:

[0147] The ultrasonic detection method for defects of the external delivery hose based on the reciprocity difference and the hybrid neural network proposed in the present application realizes accurate comparison of ultrasonic signals of samples with defects and reference samples without defects by introducing the "reciprocity difference" principle, laying a solid data foundation for the detection process. By combining the advantages of CNN and RNN, not only the spatial features of ultrasonic signals are analyzed in depth, but also their time sequence features are captured in detail, enriching the dimensions of defect recognition. In the complex and variable ultrasonic scattering wave field, the system can efficiently recognize and accurately extract defect feature signals, improving the accuracy and reliability of external delivery hose defect detection, providing strong technical support for safety production and pipeline maintenance, and reducing potential safety risks and economic losses.

[0148] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent substitutions for some technical features; and these modifications or substitutions do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for ultrasonic testing of defects in an outflow hose based on reciprocity differences and a hybrid neural network, characterized in that The method comprises the following steps: S1: an ultrasonic detection system detects an initial signal, and filters the initial signal to obtain a first signal; S2: the first signal is decomposed into an approximation coefficient and a first detail coefficient; S3: the first detail coefficient is subjected to soft threshold processing to obtain a second detail coefficient, and the second detail coefficient is reconstructed with the approximation coefficient to obtain a second signal; S4: the second signal is subjected to normalization processing to obtain a third signal; S5: the third signal is input into a convolutional neural network model to obtain a spatial feature vector of the third signal; S6: the spatial feature vector is input into a recurrent neural network model to extract a time feature vector of the third signal; S7: the spatial feature vector and the time feature vector are fused to obtain a fusion feature vector, and a denoised signal is obtained according to the fusion feature vector; S8: a hybrid neural network model is obtained by jointly training the convolutional neural network model and the recurrent neural network model, the trained hybrid neural network model is integrated into the ultrasonic detection system, and a defect feature signal is output according to the denoised signal; The convolutional neural network model and the recurrent neural network model are jointly trained to obtain a hybrid neural network model, which comprises the following steps: An ultrasonic detection signal of a defect-free external conveying hose and a plurality of defect external conveying hoses is collected to form a data set, steps S1-S7 are repeated for each set of detection signals in the data set, and the convolutional neural network model and the recurrent neural network model are jointly trained to obtain a hybrid neural network model; The data set is established by the following steps: A defect-free external conveying hose and a plurality of defect external conveying hoses are set, a first excitation probe and a first detection probe are arranged on the defect-free external conveying hose, a second detection probe and a second excitation probe are arranged on the defect external conveying hose, the position of the first excitation probe corresponds to the position of the second detection probe, and the position of the first detection probe corresponds to the position of the second excitation probe, the first excitation probe emits a first excitation signal, the first detection probe receives and processes the first excitation signal to obtain a first ultrasonic detection signal, the second excitation probe emits a second excitation signal, and the second detection probe receives and processes the second excitation signal to obtain a second ultrasonic detection signal, and the first ultrasonic detection signal and a plurality of second ultrasonic detection signals form the data set.

2. The reciprocity difference and hybrid neural network based ultrasonic inspection method of an export hose for defects according to claim 1, characterized in that, The initial signal is filtered to obtain a first signal, which comprises the following steps: The initial signal is processed by an adaptive filtering algorithm, and the adaptive filtering algorithm comprises a least mean square error algorithm, and the formula is as follows: wherein y t is a first signal being an output of the filter at a time instant t w t is a filter coefficient vector of the filter at the time instant t x t is an initial signal vector being an input of the filter at the time instant t t = 0, 1, 2, 3,...​​​​​​ t The difference between the desired signal at time t and the output first signal is calculated as follows: in, d ( t ) is the filter in t Expected signal at time, e ( t ) is the filter in t The difference between the expected signal at time t and the first output signal; The filter coefficient of the adaptive filtering algorithm is dynamically adjusted according to the difference between the calculated expected signal and the output first signal, and the formula of the filter coefficient is as follows: wherein w t +1) is a filter t the filter coefficient vector at time instant μ is a learning factor used to control the speed of filter coefficient update.​ 3. The reciprocity difference and hybrid neural network based ultrasonic inspection method of an export hose for defects according to claim 2, characterized in that, The first signal is decomposed into an approximation coefficient and a first detail coefficient, which comprises the following steps: the first signal is subjected to continuous wavelet transform, and the first signal is decomposed into approximation coefficients and first detail coefficients of different scales through wavelet transform.

4. The reciprocity difference and hybrid neural network based ultrasonic inspection method of an export hose for defects according to claim 3, characterized in that, The second signal is obtained by the following steps: The first detail coefficient is soft threshold processed to obtain a second detail coefficient, and the following formula is used: wherein is t a first signal at a time instant a first detail coefficient of the first signal at the time instant a second detail coefficient of the first signal at the time instant is t a first detail coefficient of the first signal at the time instant a first detail coefficient of the first signal at the time instant is a threshold value for the first detail coefficient, j = 0, 1, 2, 3,... The second detail coefficient is reconstructed with the approximation coefficient, including the following steps: wherein is t a second signal at the time instant, denotes a first signal of the order approximation coefficient.

5. The reciprocity difference and hybrid neural network based ultrasonic inspection method of an export hose for defects according to claim 4, characterized in that, The third signal is obtained, including the following steps: The second signal is normalized, and the following formula is used: wherein is t the third signal at the moment, is the minimum value of the second signal, is the maximum value of the second signal.

6. The reciprocity difference and hybrid neural network based ultrasonic inspection method of an export hose for defects according to claim 5, characterized in that, The spatial feature vector of the third signal is obtained, including the following steps: The third signal is input into a convolutional neural network model, and a convolutional layer of the convolutional neural network model extracts local features of the third signal through a plurality of groups of convolution kernels; A pooling layer of the convolutional neural network model uses a maximum pooling or average pooling method to reduce a feature map dimension of the local features; A fully connected layer of the convolutional neural network model performs feature fusion and outputs a spatial feature vector of the third signal.

7. The reciprocity difference and hybrid neural network based ultrasonic inspection method of an export hose for defects according to claim 6, characterized in that, The time feature vector of the third signal is obtained, including the following steps: The spatial feature vector is input into the recurrent neural network model, and a long short-term memory network is used to capture a time dependence relationship in the third signal to obtain a time feature vector of the third signal.

8. The method of claim 7, wherein, The spatial feature vector and the time feature vector are fused including the following steps: The spatial feature vector and the time feature vector are spliced into a fusion feature vector, and the following formula is used: wherein, is a spatial feature vector, is a temporal feature vector, is a fused feature vector.

Images