Intelligent recognition method for incomplete hollow plate membrane based on system signal compensation

Abstract

The application discloses a non-perfect hollow plate type flat plate membrane intelligent identification method based on system signal compensation, and comprises the following steps: 1) collecting defect signals of the hollow plate type flat plate membrane through a flaw detector; 2) collecting signals of a rectangular frame intercepting a water hole; 3) obtaining the intercepted defect signals; 4) establishing a one-dimensional convolutional neural network model fusing attention; and 5) training the model and outputting a detection result. Based on the idea of system signal compensation, the application proposes a method of a rectangular frame and fusing an attention mechanism. The one-dimensional convolutional neural network fusing the attention mechanism avoids signal interference generated by a water hole, an inherent structure of the flat plate membrane, and improves the robustness of the flat plate membrane defect identification method.

Description

Technical Field

[0001] This invention relates to the field of ultrasonic defect signal identification, and particularly to an intelligent identification method for incomplete hollow plate-type flat membranes based on system signal compensation. Background Technology

[0002] Currently, hollow sheet membranes, with their advantages of recyclability, easy decomposition, and long service life, are widely used in wastewater treatment, seawater desalination, and water purification. In water treatment, large impurities can collide with the hollow sheet membrane, causing internal damage. When detecting internal defects, the interference from the flow holes increases the computational load and reduces the recognition rate of the detection signal. Therefore, reducing interference from the flow hole signal and quickly and accurately identifying internal defect signals in the hollow sheet membrane are particularly important.

[0003] In their paper "3D reconstruction and visualization from 2D X-ray CT images in the study of ceramic membrane microstructure," Thomasová Barbora et al. used X-ray computed tomography (CT) to visualize the three-dimensional (3D) microstructure of ceramic membrane scaffolds, reconstructing a spatial 3D geometric model to assess total porosity, discontinuities, failures and defects, the spatial distribution of specific components in the material, and the possibility of simulating liquid flow. While X-ray methods are simple and convenient for inspection, a drawback is the radioactivity of X-rays, which poses a significant hazard to personnel engaged in long-term inspections. In 2020, Y Yan et al., in their paper "A Deep Learning-Based Ultrasonic Pattern Recognition Method for Inspecting Girth Weld Cracking of Gas Pipeline," utilized a method combining deep convolutional neural networks and a pre-trained support vector machine classifier to classify A-scan measurements of circumferential weld crack signals in gas pipelines into defect or non-defect groups. Converting a one-dimensional signal into a two-dimensional time-frequency graph not only increases the computational cost of a one-dimensional convolutional neural network (CNN), but also introduces subjective factors into the signal conversion and extraction process, resulting in a loss of the original signal characteristics. In 2020, Bin Wu et al. proposed a one-dimensional CNN with an attention mechanism in their paper "Radar Emitter Signal Recognition Based on One-Dimensional Convolutional Neural Network with Attention Mechanism." This allows for the direct extraction of features from the original radar signal sequence in the time domain, focusing on extracting key information from the signal. This avoids the need to convert the one-dimensional signal into a two-dimensional time-frequency graph, reducing the computational cost of deep learning and increasing recognition accuracy.

[0004] In summary, while existing methods can detect defects inside hollow plate-shaped sheet membranes, X-ray inspection poses health risks due to radiation exposure, and traditional ultrasonic signal recognition requires multiple signal conversions, resulting in the loss of the original ultrasonic signal information. Therefore, this invention proposes an incomplete intelligent recognition method for hollow plate-shaped sheet membranes based on the concept of system signal compensation. This method achieves high recognition accuracy for ultrasonic signals using a one-dimensional convolutional neural network with an attention fusion mechanism, avoiding interference from the inherent structure of the hollow plate-shaped sheet membrane—the water flow holes—effectively reducing the impact of the neural network on defect signal recognition and significantly improving the robustness of the attention-fused one-dimensional convolutional neural network. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide an intelligent identification method for incomplete hollow plate flat sheet membranes based on system signal compensation, which can effectively improve the identification of defect signals of hollow plate flat sheet membranes.

[0006] The objective of this invention is achieved as follows: A method for intelligent identification of incomplete hollow plate-type flat film based on system signal compensation includes the following steps:

[0007] Step 1): Collect defect signals of the hollow plate-type flat membrane using a flaw detector;

[0008] Step 2): Acquire the signal from the water flow hole by capturing the "rectangular frame";

[0009] Step 3): Obtain the extracted defect signal;

[0010] Step 4): Establish a one-dimensional convolutional neural network model that incorporates the attention mechanism;

[0011] Step 5): Train the model and output the detection results.

[0012] As a further limitation of the present invention, step 1) specifically includes detecting defects in the hollow plate flat film by having an inspector hold the probe by hand. When collecting defect signals of the hollow plate flat film, the probe moves in an "S" shape, and the collected signals and data are stored in a timely manner.

[0013] As a further limitation of the present invention, step 2) specifically includes:

[0014] To obtain the complete defect signal, data processing based on the "rectangular frame" method is performed. By analyzing the ultrasonic defect signal acquired in step 1), and based on the characteristics of the uniform distribution of water flow holes inside the hollow plate membrane and the principle of pulse reflection, the signal segment where the water flow hole waveform is located is accurately determined. The signal segment where the water flow hole is located is extracted using the "rectangular frame". The extracted signals are then spliced ​​together to form a one-dimensional ultrasonic detection signal that lacks the water flow hole waveform signal.

[0015] As a further limitation of the present invention, step 3) specifically includes: performing noise reduction processing on the intercepted defect signal to remove the ultrasonic wave shape of abrupt points, and the ultrasonic signal defect echo model can be constructed as follows:

[0016]

[0017] Where: β is the amplitude of the ultrasonic defect signal; α is the bandwidth factor of the defect wave; τ is the return time of the defect wave; ω c φ is the echo center frequency; φ is the phase of the ultrasonic defect signal;

[0018] The noise in an ultrasound signal is modeled as follows:

[0019] y(t i )=f(t i )+n(t i ) i = , 1, ..., N (2)

[0020] In the formula: f(t) i ) represents the original signal; n(t) i The expected value is 0 and the variance is σ. 2 Independent and identically distributed Gaussian white noise;

[0021] The method of the present invention for removing noise and re-acquiring the information of the original signal is as follows: In order to obtain the original signal, it is necessary to remove noise, which means eliminating n(t) i )Restore f(t) i ); where the autocorrelation function of n(t) is:

[0022] Rn(u,v)=E[n(u)n(v)]=σ 2 δ(uv) (3)

[0023] |Wn(s,x)| 2 =∫∫ z n(u)n(v)ψ s (xu)ψ s (xv)dudv (4)

[0024]

[0025] As the scale s increases, |Wn(s,x)| 2 and E(|Wn(s,x)| 2 The magnitude of the wavelet transform modulus of the original signal decreases; while the maximum value of the wavelet transform modulus of the original signal increases or remains unchanged with the increase of the scale s. Therefore, denoising is performed based on the difference between the two.

[0026] As a further limitation of the present invention, step 4) specifically includes: the one-dimensional convolutional neural network model includes a network structure of five parts, each of which includes a convolutional layer, an activation function, an attention mechanism, a pooling layer, and a normalization layer; the network structure of the first four parts is convolved with small convolutional kernels and then converted into a complete set of feature maps by the ReLU activation function, which are then sent to the average pooling layer for downsampling; the network structure of the fifth part is convolved with large convolutional kernels, which enables the input feature map of the previous layer to automatically learn local features for diagnosis; all feature maps of the last pooling layer are flattened to form a fully connected layer, which is then processed by the overfitting suppression technique Dropout and passed to the final Softmax classification layer;

[0027]

[0028] Where R represents the convolution kernel; f() represents the activation function; j represents the number of kernels; and M represents the input. The channel number; 'b' represents the deviation from the kernel; This represents the convolution operator.

[0029] As a further limitation of the present invention, step 5) specifically includes: after passing through the fully connected layer, to the final Softmax classification layer, the detection result can be output; using the cross-entropy loss function, the probability of Softmax is obtained and -ln is taken to obtain the function L as follows:

[0030]

[0031] Where ||·|| denotes the L2 norm operator; M is the number of categories; y c The value is either 0 or 1; 1 is the value if the predicted class matches the sample label, and 0 is the value otherwise. c w represents the probability that a sample belongs to class c. i ,b i θ represents the weights and biases of the fully connected layer corresponding to class i. i Indicates w i The angle between x and x.

[0032] Based on the Softmax loss function that is based on positive and negative cosine similarity, three factors—scale factor, weight update factor, and margin factor—are introduced to transform the positive and negative cosine similarity respectively, resulting in new positive and negative similarity metrics S. ip ,S jn As shown in equation (8):

[0033] S ip =f(λ,α) p ,s i ,Δp),S jn =g(λ,α) n,s j ,Δn) (8)

[0034] Where λ is the scale factor, Δn and Δp are the margin factors, and α n α p f() and g() are the weight update factors, and f() and g() are similarity transformation functions.

[0035] Compared with existing technologies, the present invention employs the above technical solutions, and its beneficial effects are as follows: Because the signal acquired by the flaw detector is truncated, the overall computational load is reduced, and the accuracy of defect identification is improved. The proposed method of using a "rectangular frame" and incorporating an attention mechanism breaks through the traditional ultrasonic signal identification method, achieving a high accuracy rate for ultrasonic signal identification using a convolutional neural network, and effectively reducing the influence of the water flow hole signal on the identification signal. By incorporating an attention mechanism, the location of the water flow hole signal is improved. After accurately identifying the location of the water flow hole, the position is truncated using a "rectangular frame," reducing the length of the acquired signal and the computational load of the subsequent one-dimensional convolutional neural network. Because the "rectangular frame" is used to remove the signal from the water flow hole—a unique mechanism of hollow plate-type flat membranes—in advance, the "systematic error" acquired by the flaw detector is eliminated, and the computational complexity is effectively reduced, effectively improving the identification speed while ensuring identification accuracy. Attached Figure Description

[0036] Figure 1 Flowchart of the present invention;

[0037] Figure 2 A schematic diagram of the hollow plate-type flat membrane in this invention.

[0038] One of them is a water flow hole. Detailed Implementation

[0039] like Figure 1 The intelligent identification method for incomplete hollow plate flat film based on system signal compensation shown includes the following steps:

[0040] Step 1): Collect defect signals of the hollow plate-type flat membrane using a flaw detector;

[0041] During testing, the hollow sheet membrane is inspected by hand using a probe. When collecting defect signals, the probe moves in an "S" shape to detect defects in the hollow sheet membrane, and the collected signals and data are stored in a timely manner.

[0042] Step 2): Acquire the signal from the water flow hole by capturing the "rectangular frame";

[0043] To obtain the complete defect signal, data processing based on the "rectangular frame" method is performed. By analyzing the ultrasonic defect signal acquired in step 1), and based on the characteristics of uniform distribution of water flow holes inside the hollow plate membrane and the principle of pulse reflection, the signal segment where the water flow hole waveform is located is accurately determined. The signal segment where the water flow hole is located is extracted by using the "rectangular frame". The extracted signals are then spliced ​​together to form a one-dimensional ultrasonic detection signal that lacks the water flow hole waveform signal.

[0044] like Figure 2 As shown, the water flow holes in a hollow plate-type flat membrane have fixed heights and widths and are uniformly distributed within the membrane. During ultrasonic testing, the waveforms of these holes exhibit very similar heights and lengths. This "systematic error" occurs during testing of the hollow plate-type flat membrane, as the signal peaks of the water flow holes affect the signal recognition of the one-dimensional convolutional neural network and increase computational complexity. By using a "rectangular frame" to extract the water flow hole position signal, not only is the signal length of the hollow plate-type flat membrane effectively reduced, but the accuracy of the one-dimensional convolutional neural network in recognizing the ultrasonic signal of the hollow plate-type flat membrane is also improved. Using a "rectangular frame" to extract the complete ultrasonic signal not only eliminates the "systematic error" of the water flow holes in the hollow plate-type flat membrane but also effectively reduces computational complexity; accurate identification of the water flow hole signal is a key factor in improving the recognition rate and accuracy.

[0045] Step 3): Obtain the extracted defect signal;

[0046] After denoising the captured defect signal and removing abrupt ultrasonic waveforms, the ultrasonic signal defect echo model can be constructed as follows:

[0047]

[0048] Where: β is the amplitude, α is the bandwidth factor of the defect wave, τ is the return time of the defect wave, and ω is the return time of the defect wave. c Where φ is the center frequency of the echo, and φ is the phase.

[0049] A signal containing noise is modeled as follows:

[0050] y(t i )=f(t i )+n(t i (2) i = 1, ..., N

[0051] In the formula: f(t) i ) represents the original signal; n(t) i The expected value is 0 and the variance is σ. 2 Independent and identically distributed Gaussian white noise;

[0052] The method of the present invention for removing noise and re-acquiring the information of the original signal is as follows: In order to obtain the original signal, it is necessary to remove noise, which means eliminating n(t) i )Restore f(t) i The autocorrelation function of n(t) is:

[0053] Rn(u,v)=E[n(u)n(v)]=σ 2 δ(uv) (3)

[0054] |Wn(s,x)| 2 =∫∫ z n(u)n(v)ψ s (xu)ψ s (xv)dudv (4)

[0055]

[0056] As the scale s increases, |Wn(s,x)| 2 and E(|Wn(s,x)| 2 The magnitude of the wavelet transform modulus of the original signal decreases; while the maximum value of the wavelet transform modulus of the original signal increases or remains unchanged with the increase of the scale s. Therefore, denoising is performed based on the difference between the two.

[0057] Step 4): Establish a one-dimensional convolutional neural network model that incorporates the attention mechanism;

[0058] The one-dimensional convolutional neural network model consists of a five-part network structure, each including a convolutional layer, activation function, attention mechanism, pooling layer, and normalization layer. The first four parts of the network structure are convolved with small convolutional kernels and then transformed into a complete set of feature maps through the ReLU activation function, which are then fed into the average pooling layer for downsampling. The fifth part of the network structure uses large convolutional kernels to enable the input feature map of the previous layer to automatically learn local features for diagnosis. All feature maps of the last pooling layer are flattened to form a fully connected layer, which is then processed by the overfitting suppression technique Dropout and passed to the final Sigmoid classification layer.

[0059]

[0060] Where R represents the convolution kernel; f() represents the activation function; j represents the number of kernels; and M represents the input. The channel number; 'b' represents the deviation from the kernel; This represents the convolution operator.

[0061] Step 5): Train the model and output the detection results;

[0062] The result of step 4) is obtained, passed through the fully connected layer, and then to the final Softmax classification layer to output the detection result;

[0063] The traditional Softmax classifier exhibits excellent performance for multi-class classification tasks. It normalizes various features based on the number of classifications, making positive features more prominent. Now, using the cross-entropy loss function, the probability of the Softmax classifier is calculated and then -ln is applied to obtain the following function:

[0064]

[0065] Where ||·|| denotes the L2 norm operator; M is the number of categories; y c The value is either 0 or 1; 1 is the value if the predicted class matches the sample label, and 0 is the value otherwise. c w represents the probability that a sample belongs to class c. i ,b i θ represents the weights and biases of the fully connected layer corresponding to class i. i Indicates w i The angle between x and x;

[0066] If the output of a node in the fully connected layer is too large, the Softmax loss function value will approach 0, leading to premature training termination. Introducing a scaling factor into the Softmax loss function expands the range of logistic values, enhancing their separability. Introducing a weight update factor allows for separate optimization of positive and negative cosine similarity. Introducing a margin factor separates the decision boundaries between categories, improving the discriminative performance of the embedding representation. By simultaneously introducing scaling, weight update, and margin factors into the Softmax loss function based on positive and negative cosine similarity, and transforming the positive and negative cosine similarity accordingly, new positive and negative similarity metrics S are obtained. ip ,S jn As shown in equation (8):

[0067] S ip =f(λ,α) p ,s i ,Δp),S jn =g(λ,α) n ,s j ,Δn) (8)

[0068] Where λ is the scale factor, Δn and Δp are the margin factors, and α n α p f() and g() are the weight update factors, and f() and g() are similarity transformation functions.

[0069] This invention provides an intelligent identification method for incomplete hollow plate-type flat sheet membranes based on system signal compensation. By truncating the signal acquired by the flaw detector, the overall computational load is reduced, and the accuracy of defect identification is improved. A method using a "rectangular frame" and incorporating an attention mechanism breaks through the limitations of traditional ultrasonic signal identification methods, achieving high accuracy in ultrasonic signal identification using convolutional neural networks and effectively reducing the influence of the flow hole signal on the identification signal. By using a "rectangular frame" to pre-remove the signal from the flow hole—a unique feature of hollow plate-type flat sheet membranes—the "systematic error" acquired by the flaw detector is eliminated, effectively improving the identification speed while ensuring identification accuracy.

[0070] This invention is not limited to the above embodiments. Based on the technical solutions disclosed in this invention, those skilled in the art can make some substitutions and modifications to some of the technical features without creative effort, and all such substitutions and modifications are within the protection scope of this invention.

Claims

1. A method for intelligent identification of incomplete hollow plate-type flat sheet membranes based on system signal compensation, characterized in that, Includes the following steps: Step 1): Collect defect signals of the hollow plate-type flat membrane using a flaw detector; Step 2): Acquire the signal from the water flow hole by capturing the "rectangular frame"; Step 2) specifically includes: To obtain the complete defect signal, data processing based on the "rectangular frame" method is performed. By analyzing the ultrasonic defect signal acquired in step 1), and based on the characteristics of uniform distribution of water flow holes inside the hollow plate membrane and the principle of pulse reflection, the signal segment where the water flow hole waveform is located is accurately determined. The signal segment where the water flow hole is located is extracted by using the "rectangular frame". The extracted signals are then spliced ​​together to form a one-dimensional ultrasonic detection signal that lacks the water flow hole waveform signal. Step 3): Obtain the extracted defect signal; The method for removing noise and retrieving the original signal information is as follows: In order to retrieve the original signal, it is necessary to remove noise; noise removal is the process of eliminating noise. recover ;in, The autocorrelation function is: （3）； （4）； （5）； With scale The increase, and The magnitude decreases; while the maxima of the wavelet transform modulus of the original signal decreases with scale. The increase is either increasing or remaining constant, therefore, denoising is performed based on the difference between the two; Step 4): Establish a one-dimensional convolutional neural network model that incorporates the attention mechanism; Step 4) specifically includes: the one-dimensional convolutional neural network model comprises a network structure of five parts, each of which includes a convolutional layer, activation function, attention mechanism, pooling layer, and normalization layer; the first four parts of the network structure are convolved with small convolutional kernels and then converted into a complete set of feature maps through the ReLU activation function, which are then sent to the average pooling layer for downsampling; the fifth part of the network structure is a convolutional layer with large convolutional kernels, which enables the input feature map of the previous layer to automatically learn local features for diagnosis; all feature maps of the last pooling layer are flattened to form a fully connected layer, which is then processed by the overfitting suppression technique Dropout and passed to the final Softmax classification layer; （6）； in, Represents the convolution kernel, This represents the activation function. Represents the number of kernels. Representative input Channel number, This represents a deviation from the kernel. Represents the convolution operator; Step 5): Train the model and output the detection results; Step 5) specifically includes: after passing through the fully connected layer, the final Softmax classification layer outputs the detection result; using the cross-entropy loss function, the probability of Softmax is calculated and -ln is taken to obtain the function L as follows: （7）； in, Represents the norm 2 operator; It is the number of categories; It is either 0 or 1; if the predicted class and the sample label are the same, it is 1; otherwise, it is 0. The sample belongs to the category The probability of; express The weights and biases of the fully connected layer corresponding to the class; express and The included angle; Based on the Softmax loss function that is based on positive and negative cosine similarity, three factors—scale factor, weight update factor, and margin factor—are introduced to transform the positive and negative cosine similarity respectively, resulting in new positive and negative similarity measures. , As shown in equation (8): （8）； in, As a scale factor, , Margin factor , For weight update factors, and It is a similarity conversion function.

2. The intelligent identification method for incomplete hollow plate-type flat film based on system signal compensation according to claim 1, characterized in that, Step 1) specifically includes detecting defects in the hollow sheet membrane by having an inspector hold the probe by hand. When collecting defect signals from the hollow sheet membrane, the probe moves in an "S" shape, and the collected signals and data are stored in a timely manner.

3. The intelligent identification method for incomplete hollow plate-type flat film based on system signal compensation according to claim 1, characterized in that, Step 3) specifically includes: performing noise reduction processing on the intercepted defect signal, removing the ultrasonic wave shape of abrupt points, and constructing the ultrasonic signal defect echo model as follows: （1）； in: The amplitude of the ultrasonic defect signal. The bandwidth factor of the defect wave. For the defect wave return time, The center frequency of the echo. The phase of the ultrasonic defect signal; The noise in an ultrasound signal is modeled as follows: （2）； In the formula: The original signal; The expected value is 0; denoted as the variance of independent and identically distributed Gaussian white noise.

Images