A method for ship structure damage identification

Abstract

The present invention discloses a method for identifying damage to a ship structure. The method uses multiple acoustic emission sensors to acquire acoustic emission signals generated by structural deformation and converts them into multiple acoustic emission electrical signals. The multiple acoustic emission electrical signals are sequentially passed through corresponding preamplifiers and signal separators to output multiple amplified acoustic emission signals. The amplified acoustic emission signals are then transmitted to a signal acquisition and analysis system for digitization, where the acoustic emission characteristic signals are extracted. The acoustic emission sensors acquire damage characteristic parameter information related to the stiffened plate structure, ultimately outputting the damage generation location, time t, and the actual values ​​of the damage length (imgabs0), average damage width (imgabs1), damage extension velocity (imgabs2), and damage characteristic parameters after data fusion. The present invention achieves reliable identification of damage to a ship structure by fusing acoustic emission technology with image recognition technology.

Description

Technical Field

[0001] The present invention belongs to the technical field of non-destructive testing of ship structures, and in particular relates to a method for identifying damage to a ship structure. Background Art

[0002] Ships are large, complex, and integrated systems. Under the influence of factors such as harsh, periodic environmental loads and difficulties in maintaining welding process quality, ship structures are prone to damage. Failure to detect and repair them promptly can lead to serious maritime accidents. Against this backdrop, the identification of ship structural damage is an urgent issue that needs to be addressed.

[0003] Methods for detecting damage to ship structures are primarily divided into traditional manual methods and emerging machine-based methods. The traditional method involves regular dry docking for inspection and review by professional ship inspectors. This method is subject to significant subjective influence from inspectors, resulting in poor efficiency and reliability. Furthermore, the complex and narrow nature of ship structures, such as double bottoms and double sides, poses significant challenges to inspectors, making it easy for minor structural damage to be concealed, posing a safety hazard to the vessel. Emerging methods include eddy current crack detection and fiber Bragg grating (FBG) sensors. The FBG sensor uses its own coils to measure physical quantities, acquiring magnetic field information for crack identification, but exhibits strong electromagnetic interference and consumes significant power. The FBG sensor utilizes the strain-sensitive wavelength characteristics of the fiber Bragg grating to detect structural responses. However, the sensor is susceptible to other factors, such as temperature and humidity, and is significantly affected by harsh environments in practical engineering applications. Summary of the Invention

[0004] In order to overcome the defects of the above-mentioned background technology, the present invention provides a method for identifying damage to ship structure, which realizes reliable identification of damage to the hull structure by fusing information of acoustic emission technology and image recognition technology.

[0005] In order to solve the above technical problems, the technical solutions adopted by the present invention are:

[0006] A ship structure damage identification method, comprising:

[0007] Step 1: Acquire acoustic emission signals generated by structural deformation through a plurality of acoustic emission sensors and convert them into a plurality of acoustic emission electrical signals;

[0008] Step 2: several acoustic emission electrical signals are sequentially passed through corresponding preamplifiers and signal separators to output several amplified acoustic emission signals;

[0009] Step 3: The amplified acoustic emission signal in step 2 is transmitted to the signal acquisition and analysis system for signal digitization, and the acoustic emission characteristic signal is extracted. The damage characteristic parameter information of the stiffened plate structure is obtained through the acoustic emission sensor. The damage characteristic parameter information includes the damage generation position, damage generation time t, damage length y l1 , average damage width y b1 and damage extension rate y v1 Damage characteristic parameter measurement values;

[0010] Step 4: Using the damage generation position obtained in step 3 as the target position, the control system drives the camera to obtain the damage image signal of the target position;

[0011] Step 5: The damage image signal collected in step 4 is transmitted to the signal acquisition and analysis system. Based on the target area damage detection system, the model training and damage area solution are realized. Then, damage feature extraction is completed through image operation, and the damage length y is output. l2 , average damage width y b2 and damage extension rate y v2 Damage characteristic parameter measurement values;

[0012] Step 6: Obtain the damage generation position, damage generation time t, and damage length y output by the signal acquisition and analysis system 5 in steps 3 and 5. l1 and y l2 , average damage width y b1 and y b2 and damage extension rate y v1 and y v2 The measured values ​​of damage characteristic parameters are processed by data fusion of damage length, average damage width and damage extension speed, and the damage generation position, time t and damage length after data fusion are finally output. Average damage width Damage propagation speed The actual value of the damage characteristic parameter.

[0013] Preferably, at least three acoustic emission sensors are provided, each of which is disposed at the end of the joint between the T-section and the plate of the ship's stiffened plate; the three acoustic emission sensors are installed on the surface of the hull structure in an isosceles triangle probe array.

[0014] Preferably, the method is characterized in that: Step 3 obtains the damage generation position, including the distance data r1 between the first acoustic emission sensor and the damage signal source position, and the angle data θ between the line connecting the first acoustic emission sensor to the damage signal source position and a preset reference line, and the specific method includes:

[0015]

[0016] Where A=a sinθ1+b sinθ2, B=a cosθ1+b cosθ2, C=D1(D2 2 -△t2 2 v 2 ),D=D2(D1 2 -△t1 2 v 2 ), a=D1(D2 2 -△t2 2 v 2 ), b=D2(D1 2 -Δt1 2 v 2 ), v is the propagation speed of the acoustic emission signal in the hull structure, r1 is the distance between the first acoustic emission sensor 11 and the damage signal source, r2 is the distance between the second acoustic emission sensor 12 and the damage signal source, r3 is the distance between the third acoustic emission sensor 13 and the damage signal source, D1 is the distance between the first acoustic emission sensor 11 and the second acoustic emission sensor 12, D2 is the distance between the first acoustic emission sensor 11 and the third acoustic emission sensor 13, θ is the angle between the line connecting the first acoustic emission sensor 11 and the damage signal source and the reference line, θ1 is the angle between the line connecting the first acoustic emission sensor 11 and the second acoustic emission sensor 12 and the reference line, θ2 is the angle between the line connecting the first acoustic emission sensor 11 and the third acoustic emission sensor 13 and the reference line; the preset reference line is a straight line passing through the first acoustic emission sensor 11 and parallel to the transverse frame on the stiffener.

[0017] Preferably, in step 1, three acoustic emission sensors are set at the ends of the joints between the T-section and the plate of the stiffened plate at the inner bottom plate of the ship structure; in step 1, three acoustic emission sensors are set, and the three acoustic emission sensors are set on the surface of the hull structure in an isosceles triangle probe array.

[0018] Preferably, the number of preamplifiers and signal separators in step 2 is the same as the number of acoustic emission sensors, and the voltage gain of the preamplifier is 40 dB.

[0019] Preferably, the method of extracting damage features in step 5 includes:

[0020] Step 51, define the network structure of the improved AlexNet architecture,

[0021] The network structure of the AlexNet architecture in the model training unit was optimized and adjusted. The high-level semantic features of the hull damage image were extracted through the first five layers of the network. The first five layers of the network include convolutional layers and pooling layers, and the last three layers of the network are fully connected layers. The crack images were classified using the Softmax classifier, and the classification results were used to determine whether the image contained cracks.

[0022] Step 52: The model training unit completes the model training;

[0023] A data set is constructed based on the damage image signal. The damage image is annotated with a box using the LabelImg software to construct the data set. The constructed data set is randomly divided into a training set and a validation set in a ratio of 7:3. After the data set is divided, the data set is subjected to random cropping, random horizontal flipping, brightness change, noise addition, and regularization preprocessing operations in sequence. The preprocessed training set image signal is passed through the improved AlexNet network of the model training unit, and a feature map is formed through the convolution layer and the pooling layer, and then a feature vector is formed. Classification is achieved through the fully connected layer, and the classifier result determines whether it is a crack image. After the model training unit trains the model, it will save the training model.

[0024] Step 53: The damage detection unit completes the solution of the damage target area;

[0025] The damage image signal is input into the damage detection unit, and the damage target area of ​​the stiffened plate is solved through the training model trained by the model training unit; the trained model is loaded into the prepared prediction script to complete the damage area prediction process;

[0026] The implementation method of the prediction script is as follows: by initializing the improved AlexNet network, loading the trained training model, and using the forward propagation and backpropagation algorithms of the neural network to calculate the model output of the image numerical matrix, the output matrix is ​​then passed through the Softmax classifier to obtain the maximum value in the probability distribution, that is, to obtain the crack category and predicted probability, and complete the damage area solution process.

[0027] Step 54, image pre-processing for damage feature parameter extraction;

[0028] For the damaged target area obtained in step 53, image pre-processing operations are performed for subsequent damage feature parameter extraction through image processing. Feature enhancement and filtering are performed on the damaged target area solution in step 53, and the image is converted into a binary image. The structural damaged area has a value of 1, which is recorded as a white point, and the other areas have a value of 0, which is recorded as a black point, to obtain a binary damage image.

[0029] Step 55, damage skeleton extraction

[0030] The main steps of damage skeleton extraction are as follows: to avoid the deviation of the extraction results and take into account the geometric characteristics of the slender cracks in steel structures, first create a 3×3 symmetrical rectangular image structure element, perform corrosion and opening operations on the damage binary image in sequence, and then subtract the two operation results to obtain the damage skeleton edge feature points. This completes a cycle. Based on the corrosion operation results of the previous cycle, the above cycle is continued until the number of white points in the eight-neighborhood of the structure element at each point in the image is less than 3. Finally, a union operation is performed on the damage skeleton edge feature points obtained by the subtraction operation in each cycle. The point with the number of white points in the eight-neighborhood of the structure element is 1 as the starting and end point of the structural damage, and the point with the number of white points in the eight-neighborhood of the structure element and no white points are connected is taken as the middle point of the damage skeleton to complete the extraction of the damage skeleton.

[0031] Step 56, extracting damage characteristic parameters;

[0032] The damage binary image and the number of white dots in the damage skeleton extraction results in step 54 and step 55 are summed up to obtain the damage area and damage length y l2 The ratio of the two is the average damage width y b2 , the damaged skeleton length y l2 Taking the derivative with respect to time, we can get the damage extension rate y v2 , in summary, the final output damage length y l2 , average damage width y b2 and damage extension rate y v2 .

[0033] Preferably, step 51 defines the network structure of the improved AlexNet architecture:

[0034] The AlexNet architecture was optimized by modifying the number and size of convolution kernels in the AlexNet feature extraction network, as well as the classifier's dropout parameters and the number of classification categories. To reduce the number of computational parameters, the convolution kernel size was set to 3x3; the number of convolution kernels was reduced based on the number of predicted categories; and to enhance the predictive power of neurons, the dropout method's neuron inactivation probability was increased.

[0035] The method for extracting high-level semantic features of the hull damage image is as follows: convolution operations are performed on the input damage image using different convolution kernels in the convolution layer to extract features of different frequency bands of the image; and maximum pooling operations are performed in the pooling layer to extract features of different granularities of the image.

[0036] The crack image is classified using the Softmax classifier. The method of judging whether the image contains cracks based on the classification results is as follows: the Softmax classifier calculates the probability of the predicted category. If the obtained probability is greater than a given threshold, it is judged to be a crack. If not, it is not judged to be a crack.

[0037] Preferably, the method for extracting damage features in step 5 includes: the optimization target of the model training unit in step 5 solves the problem of difficulty in identifying complex-shaped cracks and difficulty in ensuring recognition accuracy in actual engineering during the model training process, which is specifically achieved by the following method:

[0038] To address the problem of complex crack identification in actual engineering, the classification loss function of the model training unit is set to the Focal Loss-a function, reducing the weight of samples containing hull cracks. This allows the model to focus more on complex crack areas in the hull structure, which are difficult to classify, during training, thereby improving the detection accuracy of complex crack identification in complex hull structures. The Focal Loss-a function used is as follows:

[0039] FL(P m )=-a(1-P m ) γ log(P m )

[0040] Among them, P m is the probability that the sample belongs to the crack category, P m The closer it is to 1, the more accurate the classification is. γ is the concentration parameter and a is the shared weight coefficient.

[0041] Preferably, the information fusion system obtains the damage generation position, damage generation time t, damage length y output by the signal acquisition and analysis system 5 in step 3 and step 5. l1 and y l2 , average damage width y b1 and y b2 and damage extension rate y v1 and y v2 The measured values ​​of damage characteristic parameters are processed by data fusion of damage length, average damage width and damage extension speed, and the damage generation position, time t and damage length after data fusion are finally output. Average damage width Damage propagation speed The actual value of the damage characteristic parameter is obtained by preprocessing the collected damage characteristic information to a time reference system, with a damage characteristic sampling period of T and a sampling number of k. For the damage characteristic parameters of damage length, average damage width and damage extension speed, there is a measurement value y of the damage characteristic parameter i from the acoustic emission technology and image recognition technology at each kT moment.i1 and y i2 , the observation model used is:

[0042] y i =H i x i +v i

[0043] Among them, y i is a two-dimensional measurement value matrix of the damage characteristic parameter i (such as damage length) at time kT, and its main diagonal elements are the measurement values ​​y of the damage characteristic parameter i from acoustic emission recognition technology and image recognition technology i1 and y i2 , H i is the observation matrix of damage characteristic parameter i at time kT, let H i =[1,1] T , x i is the one-dimensional true value matrix of damage characteristic parameter i at time kT, v i is the two-dimensional random interference matrix of the damage characteristic parameter i at time kT, which is used to characterize the interference noise that affects the measurement effect of the sensor;

[0044] In order to fuse the feature information of different precision, the experimental statistical results of the damage characteristic parameter i based on the acoustic emission recognition technology and the image recognition technology should be compared with the true value of the damage characteristic parameter i, and the recognition accuracy of the two recognition methods for different damage characteristic parameters should be obtained, and then the weight w of the corresponding sample in the regression should be assumed. i1 and w i2 , construct a two-dimensional weight matrix W i . Two-dimensional weight matrix W i The specific construction method is as follows:

[0045] 1①Measure the true value of the damage characteristic parameter i of a structurally damaged component

[0046] 2②Use acoustic emission recognition technology to conduct multiple measurement tests on the damage characteristic parameter i of the structural damage component in ① to obtain the measured value i of the damage characteristic parameter i 11 、i 12 、i 13 、……、i 1n ;

[0047] 3③ Repeatedly measure the damage characteristic parameter i of the structural damage component in ① through image recognition technology to obtain the measured value i of the damage characteristic parameter i 21 、i 22 、i 23 、……、i 2n ;

[0048] 4④ Calculate the measurement variance of the damage characteristic parameter i of the two identification methods respectively to characterize their identification accuracy, namely:

[0049]

[0050] ⑤ Determine the corresponding weight w according to the recognition accuracy of the two recognition methods i1 and w i2 :

[0051]

[0052] ⑥ Take the w obtained in ⑤ i1 and w i2 Construct a two-dimensional weight matrix W i .

[0053] Complete the two-dimensional weight matrix W i After the construction, data fusion processing is performed, and according to the principle of weighted least squares method, the generalized residual sum of squares of the model is minimized, that is:

[0054]

[0055] Solve for the true value of the damage characteristic parameter i at time kT for:

[0056]

[0057] The damage length after data fusion Average damage width Damage propagation speed The true value of is:

[0058]

[0059]

[0060] The beneficial effects of the present invention are as follows: by arranging acoustic emission sensors and industrial cameras on the surface of the internal structure of the hull, the present invention conducts joint detection of hull structure damage, and combines information fusion methods to greatly improve the fault tolerance and working stability of the ship structure damage identification method, realize global detection of the ship structure, and assist ship operators to take early warning measures for developing damage in a timely manner. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 Schematic diagram of the overall structure of an embodiment of the present invention;

[0062] Figure 2 This is a schematic diagram of the sensor arrangement position in an embodiment of the present invention;

[0063] Figure 3 This is a flow chart of a method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0064] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0065] A ship structure damage identification method, comprising:

[0066] Step 1: Acquire acoustic emission signals generated by structural deformation using multiple acoustic emission sensors and convert them into multiple acoustic emission electrical signals. In this embodiment, three acoustic emission sensors are set at the ends of the joints between the T-section and the plate of the stiffened plate at the inner bottom plate of the ship structure. The three acoustic emission sensors are set on the surface of the hull structure in an isosceles triangle probe array.

[0067] Step 2: Several of the acoustic emission electrical signals are sequentially passed through corresponding preamplifiers and signal separators to output several amplified acoustic emission signals; the number of preamplifiers and signal separators is the same as the number of acoustic emission sensors, the sensor frequency range is 100kHz-450kHz, the center frequency is 150kHz, and the voltage gain of the preamplifier is 40dB.

[0068] Step 3: The amplified acoustic emission signal in step 2 is transmitted to the signal acquisition and analysis system for signal digitization, and the acoustic emission characteristic signal is extracted. The extracted signal is analyzed and processed by the acoustic emission signal acquisition and analysis software, and the damage characteristic parameter information is directly output, including the damage generation location, damage generation time t, and damage length y. l1 , average damage width y b1 and damage extension rate y v1 Damage characteristic parameter measurement values;

[0069] The damage generation location is obtained, including the distance data r1 between the first acoustic emission sensor and the damage signal source location, and the angle data θ between the line connecting the first acoustic emission sensor to the damage signal source location and a preset reference line. The specific method includes:

[0070]

[0071] Among them, A=a sinθ1+b sinθ2, B=a cosθ1+b cosθ2, C=D1(D2 2 -Δt2 2 v 2 ), D=D2(D1 2 -Δt1 2 v 2 ), a=D1(D2 2 -Δt2 2 v 2), b=D2(D1 2 -Δt1 2 v 2 ), v is the propagation speed of the acoustic emission signal in the hull structure, r1 is the distance between the first acoustic emission sensor 11 and the damage signal source position, r2 is the distance between the second acoustic emission sensor 12 and the damage signal source position, r3 is the distance between the third acoustic emission sensor 13 and the damage signal source position, D1 is the distance between the first acoustic emission sensor 11 and the second acoustic emission sensor 12, D2 is the distance between the first acoustic emission sensor 11 and the third acoustic emission sensor 13, θ is the angle between the line connecting the first acoustic emission sensor 11 and the damage signal source position and the reference line, θ1 is the angle between the line connecting the first acoustic emission sensor 11 and the second acoustic emission sensor 12 and the reference line, θ2 is the angle between the line connecting the first acoustic emission sensor 11 and the third acoustic emission sensor 13 and the reference line; the preset reference line is a straight line passing through the first acoustic emission sensor 11 and parallel to the transverse frame on the stiffened plate.

[0072] Step 4: Using the damage generation position obtained in step 3 as the target position, the control system drives the camera to obtain the damage image signal of the target position;

[0073] Step 5: The damage image signal collected in step 4 is transmitted to the signal acquisition and analysis system. Based on the target area damage detection system, it includes a model training unit and a damage detection unit based on the improved AlexNet architecture, which are used to implement model training and damage area solution respectively. Then, damage feature extraction is completed through image operation, and the damage length y is output. l2 , average damage width y b2 and damage extension rate y v2 Damage characteristic parameter measurements, including:

[0074] Step 51, defining the network structure of the improved AlexNet architecture;

[0075] The network structure of the AlexNet architecture in the model training unit was optimized. The first five layers of the network are convolutional and pooling layers, which are used to extract high-level semantic features from hull damage images. The last three layers are fully connected layers, which use a Softmax classifier to classify crack images and determine whether an image contains cracks based on the classification results. The purpose of defining the improved AlexNet architecture is to customize a damage detection model specifically for crack detection scenarios.

[0076] The AlexNet architecture was optimized by modifying the number and size of convolution kernels in the AlexNet feature extraction network, as well as the classifier's dropout parameters and the number of classification categories. To reduce the number of computational parameters, the convolution kernel size was set to 3x3; the number of convolution kernels was reduced based on the number of predicted categories; and to enhance the predictive power of neurons, the dropout method's neuron inactivation probability was increased.

[0077] The method for extracting high-level semantic features of the hull damage image is as follows: convolution operations are performed on the input damage image using different convolution kernels in the convolution layer to extract features of different frequency bands of the image; and maximum pooling operations are performed in the pooling layer to extract features of different granularities of the image.

[0078] The crack image is classified using the Softmax classifier. The method of judging whether the image contains cracks based on the classification results is as follows: the Softmax classifier calculates the probability of the predicted category. If the probability is greater than a given threshold, it is judged to be a crack. If not, it is not judged to be a crack.

[0079] Step 52: The model training unit completes the model training;

[0080] A data set is constructed based on the damage image signal. The damage image is annotated with a box using the LabelImg software to achieve data set construction. The constructed data set is randomly divided into a training set and a validation set in a ratio of 7:3. The damage training set is used to train the model architecture, and the validation set is used to evaluate the training effect of the model architecture. After the data set is divided, the data set is subjected to preprocessing operations such as random cropping, random horizontal flipping, brightness change, noise addition, and regularization. The training set image signal that has undergone the preprocessing operation is passed through the improved AlexNet network of the model training unit. A feature map is formed through the convolution layer and the pooling layer, and then a feature vector is formed. Classification is achieved through the fully connected layer. The softmax classifier maps the output results of the neurons in the fully connected layer to the range of 0-1, calculates the predicted classification probability, and compares the probability result with the threshold value to determine whether it is a crack image. After the model is trained, the model training unit will save the training model; the softmax classifier calculates the probability of the predicted category. If the probability is greater than the given threshold, it is determined to be a crack; if it is not greater than, it is not determined to be a crack.

[0081] Step 53: The damage detection unit completes the solution of the damage target area;

[0082] The damage image signal is input into the damage detection unit. The model trained in the model training unit is used to determine the target damage area of ​​the stiffened plate. The trained model is then loaded into the prediction script to complete the damage area prediction process. The prediction script must be written by the user. The prediction script draws the prediction box and the corresponding category, indicating whether the area is a crack or not.

[0083] The implementation method of the prediction script is as follows: by initializing the improved AlexNet network, loading the trained training model, and using the forward propagation and backpropagation algorithms of the neural network to calculate the model output of the image numerical matrix, the output matrix is ​​then passed through the Softmax classifier to obtain the maximum value in the probability distribution, that is, to obtain the crack category and predicted probability, and complete the damage area solution process.

[0084] Step 54, image pre-processing for damage feature parameter extraction;

[0085] The image pre-processing operation is completed for the subsequent damage feature parameter extraction through image processing. The specific steps are as follows: the damage target area solution in step 53 is sequentially subjected to feature enhancement, filtering and noise reduction, and converted into a binary image. The structural damage area is set to 1 and recorded as a white point, and the other areas are set to 0 and recorded as a black point, thereby obtaining a damage binary image;

[0086] Step 55, damage skeleton extraction

[0087] The main steps of damage skeleton extraction are as follows: to avoid the deviation of the extraction results and take into account the geometric characteristics of the slender cracks in steel structures, first create a 3×3 symmetrical rectangular image structure element, perform corrosion and opening operations on the damage binary image in sequence, and then subtract the two operation results to obtain the damage skeleton edge feature points. This completes a cycle. Based on the corrosion operation results of the previous cycle, the above cycle is continued until the number of white points in the eight-neighborhood of the structure element at each point in the image is less than 3. Finally, a union operation is performed on the damage skeleton edge feature points obtained by the subtraction operation in each cycle. The point with the number of white points in the eight-neighborhood of the structure element is 1 as the starting and end point of the structural damage, and the point with the number of white points in the eight-neighborhood of the structure element and no white points are connected is taken as the middle point of the damage skeleton to complete the extraction of the damage skeleton.

[0088] Step 56, extracting damage characteristic parameters;

[0089] The damage binary image and the number of white dots in the damage skeleton extraction results in step 54 and step 55 are summed up to obtain the damage area and damage length y l2 The ratio of the two is the average damage width y b2 , the damaged skeleton length y l2 Taking the derivative with respect to time, we can get the damage extension rate y v2, in summary, the final output damage length y l2 , average damage width y b2 and damage extension rate y v2 .

[0090] Step 6: The information fusion system obtains the damage generation position, damage generation time t, and damage length y output by the signal acquisition and analysis system 5 in steps 3 and 5. l1 and y l2 , average damage width y b1 and y b2 and damage extension rate y v1 and y v2 The measured values ​​of damage characteristic parameters are processed by data fusion of damage length, average damage width and damage extension speed, and the damage generation position, time t and damage length after data fusion are finally output. Average damage width Damage propagation speed The actual value of the damage characteristic parameter is obtained by preprocessing the collected damage characteristic information to a time reference system, with a damage characteristic sampling period of T and a sampling number of k. For the damage characteristic parameters of damage length, average damage width and damage extension speed, there is a measurement value y of the damage characteristic parameter i from the acoustic emission technology and image recognition technology at each kT moment. i1 and y i2 , the observation model used is:

[0091] y i =H i x i +v i

[0092] Among them, y i is a two-dimensional measurement value matrix of the damage characteristic parameter i (such as damage length) at time kT, and its main diagonal elements are the measurement values ​​y of the damage characteristic parameter i from acoustic emission recognition technology and image recognition technology i1 and y i2 , H i is the observation matrix of damage characteristic parameter i at time kT, let H i =[1,1] T , x i is the one-dimensional true value matrix of damage characteristic parameter i at time kT, v i is the two-dimensional random interference matrix of the damage characteristic parameter i at time kT, which is used to characterize the interference noise that affects the measurement effect of the sensor;

[0093] In order to fuse the feature information of different precision, the experimental statistical results of the damage characteristic parameter i based on the acoustic emission recognition technology and the image recognition technology should be compared with the true value of the damage characteristic parameter i, and the recognition accuracy of the two recognition methods for different damage characteristic parameters should be obtained, and then the weight w of the corresponding sample in the regression should be assumed. i1 and w i2 , construct a two-dimensional weight matrix W i . Two-dimensional weight matrix W i The specific construction method is as follows:

[0094] ⑦Measure the true value of the damage characteristic parameter i of a structurally damaged component

[0095] ⑧ Through the acoustic emission recognition technology, the damage characteristic parameter i of the structural damage component in ① is measured multiple times to obtain the measured value i of the damage characteristic parameter i 11 、i 12 、i 13 、……、i 1n ;

[0096] ⑨ Repeatedly measure the damage characteristic parameter i of the structural damage component in ① through image recognition technology to obtain the measured value i of the damage characteristic parameter i 21 、i 22 、i 23 、……、i 2n ;

[0097] ⑩ Calculate the measurement variance of the damage characteristic parameter i of the two identification methods respectively to characterize their identification accuracy, namely:

[0098]

[0099] Determine the corresponding weight w according to the recognition accuracy of the two recognition methods i1 and w i2 :

[0100]

[0101] With the w obtained in ⑤ i1 and w i2 Construct a two-dimensional weight matrix W i .

[0102] Complete the two-dimensional weight matrix W i After the construction, data fusion processing is performed, and according to the principle of weighted least squares method, the generalized residual sum of squares of the model is minimized, that is:

[0103]

[0104] Solve for the true value of the damage characteristic parameter i at time kT for:

[0105]

[0106] For the damage length after data fusion Average damage width Damage propagation speed The true value of is:

[0107]

[0108] It should be understood that those skilled in the art can make improvements or changes based on the above description, and all such improvements and changes should fall within the scope of protection of the appended claims of the present invention.

Claims

1. A method for identifying damage to a ship structure, characterized in that: include: Step 1: Acquire acoustic emission signals generated by structural deformation through a plurality of acoustic emission sensors and convert them into a plurality of acoustic emission electrical signals; Step 2: the plurality of acoustic emission electrical signals are sequentially passed through corresponding preamplifiers and signal separators to output a plurality of amplified acoustic emission signals; Step 3: The amplified acoustic emission signal in step 2 is transmitted to the signal acquisition and analysis system for signal digitization, and the acoustic emission characteristic signal is extracted. The damage characteristic parameter information of the stiffened plate structure is obtained through the acoustic emission sensor. The damage characteristic parameter information includes the damage generation position, damage generation time t, damage length y l1 , average damage width y b1 and damage extension rate y v1 Damage characteristic parameter measurement values; Step 4: Using the damage generation position obtained in step 3 as the target position, the control system drives the camera to obtain the damage image signal of the target position; Step 5: The damage image signal collected in step 4 is transmitted to the signal acquisition and analysis system. Based on the target area damage detection system, the model training and damage area solution are realized. Then, damage feature extraction is completed through image operation, and the damage length y is output. l2 , average damage width y b2 and damage extension rate y v2 Damage characteristic parameter measurement values; Step 6: Obtain the damage generation position, damage generation time t, and damage length y output by the signal acquisition and analysis system 5 in steps 3 and 5. l1 and y l2 , average damage width y b1 and y b2 and damage extension rate y v1 and y v2 The measured values ​​of damage characteristic parameters are processed by data fusion of damage length, average damage width and damage extension speed, and the damage generation position, time t and damage length after data fusion are finally output. Average damage width Damage propagation speed The actual value of the damage characteristic parameter.

2. A ship structure damage identification method according to claim 1, characterized in that: At least three acoustic emission sensors are set up, and each acoustic emission sensor is set at the end of the joint between the T-shaped bar and the plate of the ship's stiffened plate; the three acoustic emission sensors are installed on the surface of the hull structure in an isosceles triangle probe array.

3. A ship structure damage identification method according to claim 2, characterized in that: The step 3 obtains the damage generation position, including the distance data r1 between the first acoustic emission sensor and the damage signal source position, and the angle data θ between the line connecting the first acoustic emission sensor to the damage signal source position and the preset reference line. The specific method includes: Among them, A=a sinθ1+b sinθ2, B=a cosθ1+b cosθ2, C=D1(D2 2 -Δt2 2 v 2 ), D=D2(D1-Δt1 2 v 2 ), a=D1(D2 2 -Δt2 2 V 2 ), b=D2(D1 2 -Δt1 2 V 2 ), v is the propagation speed of the acoustic emission signal in the hull structure, r1 is the distance between the first acoustic emission sensor (11) and the position of the damage signal source, r2 is the distance between the second acoustic emission sensor (12) and the position of the damage signal source, r3 is the distance between the third acoustic emission sensor (13) and the position of the damage signal source, D1 is the distance between the first acoustic emission sensor (11) and the second acoustic emission sensor (12), D2 is the distance between the first acoustic emission sensor (11) and the third acoustic emission sensor (13), θ is the angle between the line connecting the first acoustic emission sensor (11) and the position of the damage signal source and the reference line, θ is the angle between the line connecting the first acoustic emission sensor (11) and the second acoustic emission sensor (12) and the reference line, θ2 is the angle between the line connecting the first acoustic emission sensor (11) and the third acoustic emission sensor (13) and the reference line; the preset reference line is a straight line passing through the first acoustic emission sensor (11) and parallel to the transverse frame on the stiffened plate.

4. A ship structure damage identification method according to claim 1, characterized in that: In step 1, three acoustic emission sensors are set at the ends of the joints between the T-section and the plate of the stiffened plate at the inner bottom plate of the ship structure; in step 1, three acoustic emission sensors are set, and the three acoustic emission sensors are set on the surface of the hull structure in an isosceles triangle probe array.

5. A ship structure damage identification method according to claim 1, characterized in that: The number of the preamplifiers and the signal separators in step 2 is the same as the number of the acoustic emission sensors, and the voltage gain of the preamplifier is 40 dB.

6. A ship structure damage identification method according to claim 1, characterized in that: The method for extracting damage features in step 5 includes: Step 51, define the network structure of the improved AlexNet architecture, The network structure of the AlexNet architecture in the model training unit was optimized and adjusted. The high-level semantic features of the hull damage image were extracted through the first five layers of the network. The first five layers of the network include convolutional layers and pooling layers, and the last three layers of the network are fully connected layers. The crack images were classified using the Softmax classifier, and the classification results were used to determine whether the image contained cracks. Step 52: The model training unit completes the model training; A data set is constructed based on the damage image signal. The damage image is annotated with a box using the LabelImg software to achieve data set construction. The constructed data set is randomly divided into a training set and a validation set in a ratio of 7:

3. After the data set division is completed, the data set is subjected to random cropping, random horizontal flipping, brightness change, noise addition, and regularization preprocessing operations in sequence. The training set image signal that has undergone the preprocessing operations is passed through the improved AlexNet network of the model training unit. A feature map is formed through the convolution layer and the pooling layer, and then a feature vector is formed. Classification is achieved through the fully connected layer. The classifier result determines whether it is a crack image. After the model training unit has trained the model, it will save the training model. Step 53: The damage detection unit completes the solution of the damage target area; The damage image signal is input into the damage detection unit, and the damage target area of ​​the stiffened plate is solved through the training model trained by the model training unit; the trained model is loaded into the prepared prediction script to complete the damage area prediction process; The prediction script is implemented as follows: by initializing the improved AlexNet network, loading the trained model, and using the neural network's forward and backpropagation algorithms to calculate the model output from the image numerical matrix. The output matrix is ​​then passed through the Softmax classifier to obtain the maximum value in the probability distribution, that is, to obtain the crack category and predicted probability, completing the damage area solution process; Step 54, image pre-processing for damage feature parameter extraction; For the damaged target area obtained in step 53, image pre-processing operations are performed for subsequent damage feature parameter extraction through image processing. Feature enhancement and filtering are performed on the damaged target area solution in step 53, and the image is converted into a binary image. The structural damaged area has a value of 1, which is recorded as a white point, and the other areas have a value of 0, which is recorded as a black point, to obtain a binary damage image. Step 55, damage skeleton extraction The main steps of damage skeleton extraction are as follows: to avoid the deviation of the extraction results and take into account the geometric characteristics of the slender cracks in steel structures, first create a 3×3 symmetrical rectangular image structure element, perform corrosion and opening operations on the damage binary image in sequence, and then subtract the two operation results to obtain the damage skeleton edge feature points. This completes a cycle. Based on the corrosion operation results of the previous cycle, the above cycle is continued until the number of white points in the eight-neighborhood of the structure element at each point in the image is less than 3. Finally, a union operation is performed on the damage skeleton edge feature points obtained by the subtraction operation in each cycle. The point with the number of white points in the eight-neighborhood of the structure element is 1 as the starting and end point of the structural damage, and the point with the number of white points in the eight-neighborhood of the structure element and no white points are connected is taken as the middle point of the damage skeleton to complete the extraction of the damage skeleton. Step 56, extracting damage characteristic parameters; The damage binary image and the number of white dots in the damage skeleton extraction results in step 54 and step 55 are summed up to obtain the damage area and damage length y l2 The ratio of the two is the average damage width y b2 , the damaged skeleton length y l2 Taking the derivative with respect to time, we can get the damage extension rate y v2 , in summary, the final output damage length y l2 , average damage width y b2 and damage extension rate y v2 .

7. A ship structure damage identification method according to claim 6, characterized in that: In step 51, the network structure of the improved AlexNet architecture is defined as follows: The method for optimizing the network structure of the AlexNet architecture is as follows: the number and size of convolution kernels in the AlexNet feature extraction network structure, the parameters of the classifier's dropout method, and the number of classification categories are modified; to reduce the number of parameters in the operation, the size of the convolution kernel is set to 3x3; the number of convolution kernels is reduced according to the number of predicted categories; to enhance the prediction ability of neurons, the probability of neuron deactivation in the dropout method is increased; The method for extracting high-level semantic features of the hull damage image is as follows: different convolution kernels of the convolution layer are used to perform convolution operations on the input damage image to extract features of different frequency bands of the image; And through the maximum pooling operation of the pooling layer, features of different granularities of the image are extracted; The crack image is classified using the Softmax classifier. The method of judging whether the image contains cracks based on the classification results is as follows: the Softmax classifier calculates the probability of the predicted category. If the obtained probability is greater than a given threshold, it is judged to be a crack. If not, it is not judged to be a crack.

8. A ship structure damage identification method according to claim 6, characterized in that: The method for extracting damage features in step 5 includes: the optimization target of the model training unit in step 5 solves the problem of difficulty in identifying complex-shaped cracks and difficulty in ensuring recognition accuracy in actual engineering during the model training process, which is specifically achieved by the following method: To address the problem of complex shape crack identification in actual engineering, the classification loss function of the model training unit is set to the Focal loss-a function, reducing the weight of samples containing hull cracks. This allows the model to focus more on difficult-to-classify samples such as complex crack areas in the hull structure during training, thereby improving the detection accuracy of complex crack identification in complex hull structures. The Focal loss-a function used is as follows: FL(P m )=-a(1-P m ) γ log(P m ) Among them, P m is the probability that the sample belongs to the crack category, P m The closer it is to 1, the more accurate the classification is. γ is the concentration parameter and a is the shared weight coefficient.

9. A ship structure damage identification method according to claim 1, characterized in that: The information fusion system obtains the damage generation position, damage generation time t, and damage length y output by the signal acquisition and analysis system 5 in step 3 and step 5 l1 and y l2 , average damage width y b1 and y b2 and damage extension rate y v1 and y v2 The measured values ​​of damage characteristic parameters are processed by data fusion of damage length, average damage width and damage extension speed, and the damage generation position, time t and damage length after data fusion are finally output. Average damage width Damage propagation speed The actual value of the damage characteristic parameter is obtained by preprocessing the collected damage characteristic information to a time reference system, with a damage characteristic sampling period of T and a sampling number of k. For the damage characteristic parameters of damage length, average damage width and damage extension speed, there is a measurement value y of the damage characteristic parameter i from the acoustic emission technology and image recognition technology at each kT moment. i1 and y i2 , the observation model used is: y i =H i x i +v i Among them, y i is the two-dimensional measurement value matrix of the damage characteristic parameter damage length i at time kT, and its main diagonal elements are the measurement values ​​y of the damage characteristic parameter i from acoustic emission recognition technology and image recognition technology i1 and y i2 , H i is the observation matrix of damage characteristic parameter i at time kT, let H i =[1,1] T , x i is the one-dimensional true value matrix of damage characteristic parameter i at time kT, v i is the two-dimensional random interference matrix of the damage characteristic parameter i at time kT, which is used to characterize the interference noise that affects the measurement effect of the sensor; In order to fuse the feature information of different precision, the experimental statistical results of the damage characteristic parameter i based on the acoustic emission recognition technology and the image recognition technology should be compared with the true value of the damage characteristic parameter i, and the recognition accuracy of the two recognition methods for different damage characteristic parameters should be obtained, and then the weight w of the corresponding sample in the regression should be assumed. i1 and w i2 , construct a two-dimensional weight matrix W i； Two-dimensional weight matrix W i The specific construction method is as follows: ①Measure the true value of the damage characteristic parameter i of a structurally damaged component ② Through the acoustic emission recognition technology, the damage characteristic parameter i of the structural damage component in ① is measured multiple times to obtain the measured value i of the damage characteristic parameter i 11 、i 12 、i 13 、……、i 1n ; ③ Repeatedly measure the damage characteristic parameter i of the structural damage component in ① through image recognition technology to obtain the measured value i of the damage characteristic parameter i 21 、i 22 、i 23 、……、i 2n ; ④ Calculate the measurement variance of the damage characteristic parameter i of the two identification methods respectively to characterize their identification accuracy, namely: ⑤ Determine the corresponding weight w according to the recognition accuracy of the two recognition methods i1 and w i2 : ⑥ Take the w obtained in ⑤ i1 and w i2 Construct a two-dimensional weight matrix W i , Complete the two-dimensional weight matrix W i After the construction, data fusion processing is performed, and according to the principle of weighted least squares method, the generalized residual sum of squares of the model is minimized, that is: Solve for the true value of the damage characteristic parameter i at time kT for: For the damage length after data fusion Average damage width Damage propagation speed The true value of is:

Images