CN111783616B - Nondestructive testing method based on data-driven self-learning - Google Patents
Nondestructive testing method based on data-driven self-learning Download PDFInfo
- Publication number
- CN111783616B CN111783616B CN202010599968.6A CN202010599968A CN111783616B CN 111783616 B CN111783616 B CN 111783616B CN 202010599968 A CN202010599968 A CN 202010599968A CN 111783616 B CN111783616 B CN 111783616B
- Authority
- CN
- China
- Prior art keywords
- dimensional audio
- audio spectrum
- data
- nondestructive testing
- model
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 45
- 238000009659 non-destructive testing Methods 0.000 title claims abstract description 41
- 238000001228 spectrum Methods 0.000 claims abstract description 62
- 230000007547 defect Effects 0.000 claims abstract description 35
- 238000012549 training Methods 0.000 claims abstract description 22
- 230000035939 shock Effects 0.000 claims abstract description 21
- 230000005284 excitation Effects 0.000 claims abstract description 16
- 238000004422 calculation algorithm Methods 0.000 claims abstract description 15
- 238000007689 inspection Methods 0.000 claims abstract description 15
- 238000001514 detection method Methods 0.000 claims abstract description 12
- 238000013527 convolutional neural network Methods 0.000 claims abstract description 7
- 238000010606 normalization Methods 0.000 claims description 15
- 230000002950 deficient Effects 0.000 claims description 11
- 230000009466 transformation Effects 0.000 claims description 11
- 230000001133 acceleration Effects 0.000 claims description 10
- 238000012545 processing Methods 0.000 claims description 10
- 230000001066 destructive effect Effects 0.000 claims description 9
- 238000013501 data transformation Methods 0.000 claims description 7
- 238000005457 optimization Methods 0.000 claims description 6
- 238000006073 displacement reaction Methods 0.000 claims description 5
- PXFBZOLANLWPMH-UHFFFAOYSA-N 16-Epiaffinine Natural products C1C(C2=CC=CC=C2N2)=C2C(=O)CC2C(=CC)CN(C)C1C2CO PXFBZOLANLWPMH-UHFFFAOYSA-N 0.000 claims description 4
- 238000012216 screening Methods 0.000 claims description 4
- 238000010008 shearing Methods 0.000 claims description 4
- 238000010801 machine learning Methods 0.000 abstract description 14
- 238000007405 data analysis Methods 0.000 abstract description 6
- 238000004458 analytical method Methods 0.000 abstract description 5
- 238000010586 diagram Methods 0.000 description 19
- OKTJSMMVPCPJKN-UHFFFAOYSA-N Carbon Chemical compound [C] OKTJSMMVPCPJKN-UHFFFAOYSA-N 0.000 description 7
- 229910052799 carbon Inorganic materials 0.000 description 7
- 230000000694 effects Effects 0.000 description 4
- 230000005540 biological transmission Effects 0.000 description 3
- 230000008859 change Effects 0.000 description 3
- 230000006870 function Effects 0.000 description 3
- 230000008569 process Effects 0.000 description 3
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 2
- 230000002159 abnormal effect Effects 0.000 description 2
- 238000013528 artificial neural network Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 2
- 238000004364 calculation method Methods 0.000 description 2
- 238000000605 extraction Methods 0.000 description 2
- 210000002569 neuron Anatomy 0.000 description 2
- 238000001028 reflection method Methods 0.000 description 2
- 230000002411 adverse Effects 0.000 description 1
- 230000004075 alteration Effects 0.000 description 1
- XAGFODPZIPBFFR-UHFFFAOYSA-N aluminium Chemical compound [Al] XAGFODPZIPBFFR-UHFFFAOYSA-N 0.000 description 1
- 229910052782 aluminium Inorganic materials 0.000 description 1
- 238000009933 burial Methods 0.000 description 1
- 238000002790 cross-validation Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000007499 fusion processing Methods 0.000 description 1
- 238000009499 grossing Methods 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 238000002310 reflectometry Methods 0.000 description 1
- 238000004088 simulation Methods 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
- G06T7/0004—Industrial image inspection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/02—Preprocessing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/02—Preprocessing
- G06F2218/04—Denoising
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/08—Feature extraction
- G06F2218/10—Feature extraction by analysing the shape of a waveform, e.g. extracting parameters relating to peaks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/12—Classification; Matching
- G06F2218/14—Classification; Matching by matching peak patterns
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T90/00—Enabling technologies or technologies with a potential or indirect contribution to GHG emissions mitigation
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Biophysics (AREA)
- Evolutionary Computation (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Software Systems (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Mathematical Physics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Quality & Reliability (AREA)
- Investigating Or Analyzing Materials By The Use Of Ultrasonic Waves (AREA)
Abstract
The application discloses a nondestructive testing method based on data-driven self-learning, which comprises the following steps: step 1, knocking a detection object through an excitation device to generate a shock elastic wave, and acquiring a two-dimensional audio spectrum corresponding to the shock elastic wave by using a dynamic signal acquisition instrument; step 2, generating a label of the two-dimensional audio spectrum according to the type of the detected object and the number of the main frequency wave peaks and the secondary frequency wave peaks in the two-dimensional audio spectrum; and 3, building an initial nondestructive testing model based on a convolutional neural network, training the initial nondestructive testing model by utilizing a two-dimensional audio map with a label, and marking the trained model as the nondestructive testing model. Through the technical scheme in this application, combine together with machine learning big data analysis algorithm, machine learning carries out the analysis of depth to complicated various data, and inside defect is located in quick accurate automated inspection.
Description
Technical Field
The application relates to the technical field of nondestructive testing, in particular to a nondestructive testing method based on data-driven self-learning.
Background
The existing nondestructive detection method is generally that a frequency domain diagram converted from a time domain diagram accepted by an acceleration sensor is used for judging whether defects exist in a product or not, and the method can be roughly divided into two types: single-sided reflection method, double-sided transmission method. In the single-sided reflection method, an impact echo is adopted for a thick wall, internal defects of a product are detected by identifying a reflection signal, a plate structure with continuously-changing thickness is obtained, and various defects in the product structure can be estimated through the change condition of the reflection time of the bottom of the plate. The double-sided transmission method (also called a correlation method) can penetrate the inside of the structure, when a signal wave encounters a defect, diffraction is generated, and the defect is identified according to the change of the propagation time.
In the prior art, the method has certain requirements on the excitation signal, and the intensity of the excitation signal can directly influence the detection result of the internal defects of the product. The shock elastic wave detection method and the electromagnetic wave detection method and the acoustic wave detection method used for detection are quite different, namely the excitation force cannot be accurately controlled, and the excitation signal strength is different.
The method mainly judges whether the internal defects exist or not by manpower, can be influenced by factors such as the technical level of workers, working states and the like, has low efficiency and is easy to cause manual errors.
Disclosure of Invention
The purpose of the present application is: and combining with a machine learning big data analysis algorithm, performing deep analysis on complex and diverse data by machine learning, replacing a manual experience identification spectrogram to judge whether the inside is defective, and identifying the spectrum picture through the deep algorithm to obtain whether the inside of the detected object is defective and the accurate position of the defect.
The method of the application realizes that whether the detected object has defects or not and the specific positions of the defects are obtained through intelligent recognition of the spectrogram.
The technical scheme of the application is as follows: there is provided a data-driven self-learning based non-destructive inspection method, the method comprising: step 1, a shock elastic wave is generated by knocking a detected object through an automatic exciting hammer, a rebound shock wave is obtained through an acceleration multi-sensor, an analog signal is converted into a digital signal by a dynamic signal acquisition instrument, the digital signal is transmitted to a computer to be displayed in a time domain diagram form, and then the time domain diagram is converted into a frequency domain diagram and is recorded as a two-dimensional audio frequency diagram corresponding to the shock elastic wave; step 2, generating a label of the two-dimensional audio spectrum according to the type of the detected object and the number of the main frequency wave peaks and the secondary frequency wave peaks in the two-dimensional audio spectrum; and 3, building an initial nondestructive testing model based on a convolutional neural network, training the initial nondestructive testing model by utilizing a two-dimensional audio map with a label, and marking the trained model as the nondestructive testing model.
In any of the above technical solutions, further, in step 1, specifically includes: step 11, generating an audio time domain graph according to the impact elastic wave obtained by the dynamic signal acquisition instrument, sliding the audio time domain graph by a preset sliding window, and extracting an audio sequence on the audio time domain graph; step 12, extracting amplitude information in the audio sequence to generate a multi-frame sequence; and step 13, splicing the frame sequence into a two-dimensional audio map by adopting a frame splicing method.
In any of the above technical solutions, further, in step 2, specifically includes: step 21, screening the two-dimensional audio spectrum according to the type of the detected object and the number of the main frequency wave peaks and the secondary frequency wave peaks in the two-dimensional audio spectrum; step 22, when the type of the detected object corresponding to the screened two-dimensional audio spectrum is judged to be non-defective and the number of main frequency peaks is 1, marking the picture category of the two-dimensional audio spectrum as 0; and step 23, when the type of the detected object corresponding to the screened two-dimensional audio spectrum is determined to be defective and the number of the secondary frequency peaks is greater than or equal to 1, marking the picture category of the two-dimensional audio spectrum as 1, and calculating the defect position coordinates in the detected object according to the primary frequency peaks, the secondary frequency peaks and the thickness of the detected object.
And step 24, generating a label of the two-dimensional audio map according to the picture type and the defect position coordinates.
In any of the above solutions, further, the tag further includes a picture ID name.
In any of the above technical solutions, in step 3, after the initial nondestructive testing model is built, the method further includes: optimizing an initial nondestructive testing model, wherein the model optimization method comprises the steps of adjusting the complexity of a hidden layer, increasing a Dropout layer, increasing the training iteration number, adjusting the learning rate of an optimizer and increasing the batch processing size.
In any of the above solutions, further, in step 3, training the initial non-destructive testing model further includes: step 31, carrying out data enhancement on a two-dimensional audio map with a label by utilizing a data transformation algorithm, wherein the data transformation algorithm comprises at least one of data rotation, cutting and stretching, shearing transformation and piecewise affine transformation; and step 32, carrying out normalization processing on the enhanced two-dimensional audio spectrum, and taking the normalized two-dimensional audio spectrum as a training image of the initial nondestructive testing model.
In any of the above solutions, in step 32, the normalization method specifically includes: arranging pixels in the enhanced two-dimensional audio spectrum from small to large according to the pixel value; extracting the maximum value of the pixels in a preset proportion range from the front end of the sequenced pixels, and marking the maximum value as a normalized minimum value; the maximum pixel value of the pixels in the enhanced two-dimensional audio spectrum is recorded as a normalized maximum value; resetting the pixel value of the pixel in the enhanced two-dimensional audio spectrum to the normalized minimum value when the pixel value of the pixel is smaller than the normalized minimum value; and carrying out normalization processing on the two-dimensional audio map after pixel reset according to the normalization minimum value and the normalization maximum value.
In any of the above technical solutions, further, at least one sensor is disposed in the dynamic signal acquisition instrument, and the sensor is one of a displacement sensor, a speed sensor and an acceleration sensor.
The beneficial effects of this application are:
according to the method, by introducing a machine learning big data analysis algorithm, whether the carbon block has defects or not and the accurate temperature of the defects are intelligently detected, and the specific positions of the internal defects can be automatically detected and positioned more accurately by utilizing the frequency spectrum internal characteristic information, so that the internal defects can be detected more accurately and rapidly.
In the method, through combining machine learning and big data analysis, the data are learned and analyzed by utilizing a proper network model, more accurate feature extraction is carried out on the data, such as the influence of errors generated by unstable excitation signals on results, the machine learning also uses the errors as learned information to influence the output results of the model, and further adverse effects of the errors generated by the unstable excitation signals on the results are eliminated.
Drawings
The advantages of the foregoing and/or additional aspects of the present application will become apparent and readily appreciated from the description of the embodiments, taken in conjunction with the accompanying drawings, wherein:
FIG. 1 is a schematic flow chart of a data-driven self-learning based non-destructive inspection method according to one embodiment of the present application;
FIG. 2 is a schematic diagram of a non-destructive inspection process according to one embodiment of the present application;
FIG. 3 is a schematic illustration of a two-dimensional audio map of an internally defect-free inspection object according to one embodiment of the present application;
FIG. 4 is a schematic illustration of a two-dimensional audio map of an internally defective inspection object according to one embodiment of the present application;
FIG. 5 is a schematic illustration of a non-destructive testing model according to one embodiment of the present application;
FIG. 6 is a schematic diagram of a non-destructive inspection model inspection process according to one embodiment of the present application;
fig. 7 is a schematic diagram of a residual module according to one embodiment of the present application.
Detailed Description
In order that the above-recited objects, features and advantages of the present application will be more clearly understood, a more particular description of the application will be rendered by reference to the appended drawings and appended detailed description. It should be noted that, without conflict, the embodiments of the present application and features of the embodiments may be combined with each other.
In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present application, however, the present application may be practiced otherwise than as described herein, and thus the scope of the present application is not limited to the specific embodiments disclosed below.
In this embodiment, the excitation device inputs the impact echo to the sensor through the transmissibility and reflectivity of the impact elastic wave, uses the dynamic signal acquisition instrument to acquire data required by machine learning, constructs a suitable network model, performs feature extraction on the data, performs supervised learning, and encapsulates the data into a black box of a machine learning big data analysis algorithm. When data of a product is input into the black box, specific positions for locating internal defects are automatically detected.
As shown in fig. 1 and 2, the present embodiment provides a data-driven self-learning-based nondestructive testing method, which includes:
step 1, a shock excitation device (an automatic shock hammer) is used for knocking and detecting an object to generate a shock elastic wave, a rebound shock wave is obtained through accelerating multiple sensors, an analog signal is converted into a digital signal by a dynamic signal acquisition instrument, the digital signal is transmitted to a computer to be displayed in a time domain diagram mode, the time domain diagram is converted into a frequency domain diagram and is recorded as a two-dimensional audio map corresponding to the shock elastic wave, at least one sensor is arranged in the dynamic signal acquisition instrument, and the sensor is one of a displacement sensor, a speed sensor and an acceleration sensor.
Specifically, a dynamic signal acquisition instrument is used, and different types of sensors, such as a displacement sensor, a speed sensor and an acceleration sensor, are adopted for different signals to output displacement, speed and acceleration information of vibration signals. In the embodiment, a multi-acceleration sensor is adopted, and a high-frequency data acquisition card is combined, so that different position information can be acquired, and data fusion processing can be performed.
Through setting up many acceleration sensor, realized the automated control collection of data, need not artifical the participation. The automatic control data acquisition can acquire a large amount of data with high efficiency, and the data quality has certain assurance, so that the uniformity of the data is higher than that of manual data acquisition. More beneficial to later machine learning forms a model with better results.
Since noise is present in the signal, some noise reduction means are required to preprocess the data, for example, to perform front-end amplitude on the signal, to introduce a filter device, to introduce a software noise removal technique (moving smoothing filter), and the like.
Further, in step 1, specifically includes:
step 11, generating an audio time domain graph according to the impact elastic wave obtained by the dynamic signal acquisition instrument, sliding the audio time domain graph by a preset sliding window, and extracting an audio sequence on the audio time domain graph; step 12, extracting amplitude information in the audio sequence to generate a multi-frame sequence; and step 13, splicing the frame sequence into a two-dimensional audio map by adopting a frame splicing method.
Specifically, the shock excitation device is used for knocking the object to be detected to generate shock elastic waves, the shock waves return to the shock waves through the object to be detected through the acceleration sensor to perform signal receiving, the collected signals are firstly extracted to obtain a main body part containing the shock excitation signals, pure noise signals are filtered to generate a time domain diagram, a sliding window with a certain duration slides on the original time domain signals, amplitude information of an audio sequence is extracted into a multi-frame sequence, all frames are spliced into a two-dimensional audio map, and the frequency domain diagram is prepared for subsequent model training data.
Step 2, generating a label of the two-dimensional audio spectrum according to the type of the detected object and the number of the main frequency wave peaks and the secondary frequency wave peaks in the two-dimensional audio spectrum;
further, in step 2, the method specifically includes:
step 21, screening the two-dimensional audio spectrum according to the type of the detected object and the number of the main frequency wave peaks and the secondary frequency wave peaks in the two-dimensional audio spectrum;
specifically, when the excitation device knocks the object to be detected, the impact wave generated by the non-ideal knocking will affect the two-dimensional audio spectrum generated later, so that training image screening is required.
Normally, for the internal non-defective detection object, only one main frequency peak exists in the two-dimensional audio spectrum; for a detected object with defects inside, a plurality of secondary frequency peaks exist in the two-dimensional audio spectrum besides the primary frequency peaks.
Therefore, the two-dimensional audio spectrum can be screened according to the type of the detected object and the number of the primary frequency peaks and the secondary frequency peaks in the two-dimensional audio spectrum, so that the abnormal two-dimensional audio spectrum can be deleted. Taking a carbon block as a detection object for nondestructive detection, as shown in fig. 3, a main peak exists in a two-dimensional audio spectrum of the object without defects inside, and a plurality of defective peak exists; as shown in FIG. 4, only one main peak exists in the two-dimensional audio spectrum, a plurality of main peaks appear in the frequency domain image, the two-dimensional audio spectrum has no peak information and the like, and the two-dimensional audio spectrum which does not accord with the rule is screened and removed, so that the rationality of model training data is ensured.
Step 22, when the type of the detected object corresponding to the screened two-dimensional audio spectrum is judged to be non-defective and the number of main frequency peaks is 1, marking the picture category of the two-dimensional audio spectrum as 0;
step 23, when it is determined that the type of the detected object corresponding to the screened two-dimensional audio spectrum is defective and the number of secondary frequency peaks is greater than or equal to 1, marking the picture category of the two-dimensional audio spectrum as 1, and calculating the defect position coordinates inside the carbon block (detected object) by the main frequency peak, the secondary frequency peak and the thickness of the carbon block (detected object thickness) in the frequency domain map, wherein the defect position coordinates are set as (x, y, h), the numerical values of the coordinates x and y are determined by the position of the excitation device on the plane where the surface of the detected object is located, the numerical value of the coordinate h is determined by the defect depth of the detected object, and the calculation formula of the defect depth of the detected object is:
f 1 *2D＝f 2 *2D 1 ＝f 3 *2D 2 ＝……＝f w *2D w-1
wherein f 1 As the main frequency peak, f 2 、f 3 、……f w For the secondary frequency peak, D is the thickness of the detected carbon block, D 1 、D 2 、……D w-1 To detect the defect depth of an object.
And step 24, generating a label of the two-dimensional audio map according to the picture category and the defect position coordinate, wherein the label also comprises a picture ID name.
Specifically, in the present embodiment, the category of the two-dimensional audio map for detecting the defect inside the object is set to 0, and the category of the two-dimensional audio map for detecting the defect inside the object is set to 1. When the label is manufactured, the format of the label is txt format, and each row represents the label of one picture.
For example, for a detected object with a defect inside, the corresponding label is:
picture ID name, picture category 1, detect defect location coordinates (x, y, h) within the object. The exact center of the carbon block surface is the origin of the three-dimensional coordinates. The transverse direction is set as an x-axis, the longitudinal direction is set as a y-axis, the height direction is set as a z-axis, and corresponding coordinates are sequentially set as x, y and h.
And 3, building an initial nondestructive testing model based on a convolutional neural network, training the initial nondestructive testing model by utilizing a two-dimensional audio map with a label, and marking the trained model as the nondestructive testing model.
Specifically, as shown in fig. 5, in this embodiment, a CNN convolutional neural network based on data-driven self-learning for nondestructive testing is built on a deep learning framework Pytorch.
The neural network algorithm is an algorithm of machine learning, is generally similar to an abnormal complex network composed of neurons, and is formed by interconnecting individual units, each unit has a numerical input and output, and the form of each unit can be a real number or a linear combination function. It first uses a learning rule to learn and then can work. When the network judges an error, the probability of making the same error is reduced by learning. The method has strong generalization capability and nonlinear mapping capability, and can perform model processing on a system with small information quantity. The method has parallelism from the function simulation perspective and extremely high information transmission speed.
As shown in fig. 6, the data characteristics required for the artificial neural network to build the prediction model are: test area, survey line information, structure information, excitation signal information, frequency spectrum information and accumulated information. Inputting the AI analysis parameters into an AI learning model, and continuously iterating the model to select an optimal model with the highest model evaluation index, wherein the model outputs information such as whether the product is internally defective, geometric dimensions (thickness, burial depth), positions and the like, and finally, the intelligent, high-precision and high-efficiency internal defect detection of the anode carbon block for aluminum is completed.
Further, in step 3, after the initial nondestructive testing model is built, the method further includes: optimizing an initial nondestructive testing model, wherein the model optimization method comprises the steps of adjusting the complexity of a hidden layer, increasing a Dropout layer, increasing the training iteration number, adjusting the learning rate of an optimizer and increasing the batch processing size.
Specifically, after the initial nondestructive testing model is built, the multi-initial nondestructive testing model can be optimized to improve the performance of the model, and the optimization method comprises the following steps:
1) Increasing the Attention layer, adjusting the neuron number of the hidden layer, and increasing the complexity of the network structure
2) Adding a Dropout layer, optimizing network parameters by using the Dropout principle, and preventing the problem of over-fitting;
3) Different optimizers including SGD, adam and the like are used, and Adam with the best optimization effect is finally selected;
4) The learning rate of the optimizer is adjusted, the learning is too small, the convergence rate of the model is reduced, and a local optimal solution can be obtained, so that the loss function is reduced slowly when the model is trained; too large learning rate can cause the model to fail to reach the optimal solution, and jump back and forth around the optimal solution;
5) Regularization can optimize the problem of overlong running time caused by too complex model, and can prevent the model from being overfitted to a certain extent.
In this embodiment, a residual module may be added to the initial nondestructive testing model, as shown in fig. 7, where s is set as an input value, and F(s) is an output after the first layer is linearly changed and activated. In the residual network, F(s) is added to the input value s of the second layer before the second layer is activated after linear change, and then the second layer is activated and then output.
The effect of the model running on the line directly determines the success or failure of the model. Not only includes the conditions of accuracy, error and the like, but also includes the problems of running speed (time complexity), resource consumption (space complexity), stability and the like. The optimized model meets the requirements on various standards and can be operated on line.
Further, in step 3, training the initial nondestructive testing model further includes:
step 31, carrying out data enhancement on a two-dimensional audio map with a label by utilizing a data transformation algorithm, wherein the data transformation algorithm comprises at least one of data rotation, cutting and stretching, shearing transformation and piecewise affine transformation;
specifically, during training, one or more kinds of transformation is performed on the training two-dimensional audio spectrum with the label, the true value of the label after image transformation is changed along with the transformation, so that the effect of data enhancement is achieved, and the generalization capability of the model can be improved through data enhancement.
The manner of data transformation may include: data rotation (rotation angle is 10 degrees to-10 degrees), cutting and stretching (selected range is x: 0.8-1.2, y: 0.8-1.2), shearing transformation, piecewise affine transformation and the like.
And step 32, carrying out normalization processing on the enhanced two-dimensional audio spectrum, and taking the normalized two-dimensional audio spectrum as a training image of the initial nondestructive testing model.
Further, in step 32, the normalization method specifically includes:
sorting pixels in the enhanced two-dimensional audio spectrum from small to large according to the pixel value;
extracting the maximum value of the pixels in a preset proportion range from the front end of the sequenced pixels, and recording the maximum value as a normalized minimum value img_min, for example, extracting the first 10% of pixels;
the maximum pixel value of the pixels in the enhanced two-dimensional audio spectrum is recorded as a normalized maximum img_max;
when the pixel value of the pixel in the enhanced two-dimensional audio spectrum is less than the normalized minimum value img_min, resetting the pixel value of the pixel to the normalized minimum value img_min;
and carrying out normalization processing on the two-dimensional audio map after pixel reset according to the normalization minimum value img_min and the normalization maximum value img_max, so that a normalization calculation formula of the pixel x is as follows:
wherein X is the pixel value of the pixel X,is the pixel value of the normalized pixel x.
After training the model, evaluating the quality of the model and whether the application problem is solved. The judgment of over fitting and under fitting is a crucial step in model inspection. Common methods such as cross-validation, drawing learning curves, etc. After the training, the phenomenon of over fitting occurs, the basic tuning thought is to increase the training data frequency domain diagram,
error analysis is also a critical step in machine learning. By observing the error sample, the reasons of errors, such as parameter problems, algorithm selection problems, characteristic problems, data problems and the like, are comprehensively analyzed. The diagnosed model needs to be optimized, the new model needs to be diagnosed again after the optimization, and the model needs to be tried continuously through a repeated iteration continuous approximation process, so that the optimal state is achieved.
The technical scheme of the application is explained in detail above with reference to the accompanying drawings, and the application provides a nondestructive testing method based on data-driven self-learning, which comprises the following steps: step 1, knocking a detected object through an excitation device to generate a shock elastic wave, acquiring rebound shock waves through an acceleration sensor, converting analog signals into digital signals through a dynamic signal acquisition instrument, transmitting the digital signals into a computer to be presented in a time domain diagram form, converting the time domain diagram into a frequency domain diagram, and recording the frequency domain diagram as a two-dimensional audio map corresponding to the shock elastic wave; step 2, generating a label of the two-dimensional audio spectrum according to the type of the detected object and the conditions of the primary frequency wave peak and the secondary frequency wave peak in the two-dimensional audio spectrum; and 3, building an initial nondestructive testing model based on a convolutional neural network, training the initial nondestructive testing model by utilizing a two-dimensional audio map with a label, and marking the trained model as the nondestructive testing model. Through the technical scheme in this application, combine together with machine learning big data analysis algorithm, machine learning carries out the analysis of depth to complicated various data, and inside defect is located in quick accurate automated inspection.
The steps in the present application may be sequentially adjusted, combined, and pruned according to actual requirements.
The units in the device can be combined, divided and pruned according to actual requirements.
Although the present application is disclosed in detail with reference to the accompanying drawings, it is to be understood that such descriptions are merely illustrative and are not intended to limit the application of the present application. The scope of the present application is defined by the appended claims and may include various modifications, alterations, and equivalents to the invention without departing from the scope and spirit of the application.
Claims (7)
1. A data-driven self-learning based non-destructive inspection method, the method comprising:
step 1, knocking a detection object through an excitation device to generate a shock elastic wave, and acquiring a two-dimensional audio spectrum corresponding to the shock elastic wave by using a dynamic signal acquisition instrument;
step 2, generating a label of the two-dimensional audio spectrum according to the type of the detected object and the number of the main frequency wave peaks and the secondary frequency wave peaks in the two-dimensional audio spectrum;
step 3, based on convolutional neural network, setting up initial nondestructive testing model, training the initial nondestructive testing model by using the two-dimensional audio spectrum with label, marking the trained model as nondestructive testing model,
in step 2, specifically, the method includes:
step 21, screening the two-dimensional audio spectrum according to the type of the detected object and the number of the primary frequency wave peaks and the secondary frequency wave peaks in the two-dimensional audio spectrum;
step 22, when the type of the detected object corresponding to the screened two-dimensional audio spectrum is determined to be non-defective and the number of the main frequency peaks is 1, marking the picture category of the two-dimensional audio spectrum as 0;
step 23, when the type of the detected object corresponding to the screened two-dimensional audio spectrum is determined to be defect and the number of the secondary frequency peaks is greater than or equal to 1, marking the picture category of the two-dimensional audio spectrum as 1, and calculating the defect position coordinates inside the detected object according to the thicknesses of the primary frequency peaks, the secondary frequency peaks and the detected object;
and step 24, generating a label of the two-dimensional audio map according to the picture category and the defect position coordinates.
2. The method for non-destructive testing based on data-driven self-learning according to claim 1, wherein in step 1, specifically comprising:
step 11, generating an audio time domain graph according to the impact elastic wave acquired by the dynamic signal acquisition instrument, sliding the audio time domain graph by a preset sliding window, and extracting an audio sequence on the audio time domain graph;
step 12, extracting amplitude information in the audio sequence to generate a multi-frame sequence;
and step 13, splicing the frame sequence into the two-dimensional audio map by adopting a frame splicing method.
3. The data-driven self-learning based non-destructive inspection method according to claim 1, wherein said tag further comprises a picture ID name.
4. The method for non-destructive testing based on data-driven self-learning according to claim 1, wherein in step 3, after constructing the initial non-destructive testing model, further comprising:
and optimizing the initial nondestructive testing model, wherein the model optimization method comprises the steps of adjusting the complexity of a hidden layer, increasing a Dropout layer, increasing the training iteration times, adjusting the learning rate of an optimizer and increasing the batch processing size.
5. The data-driven self-learning based non-destructive inspection method according to claim 1, wherein in step 3, training the initial non-destructive inspection model further comprises:
step 31, carrying out data enhancement on the two-dimensional audio spectrum with the label by utilizing a data transformation algorithm, wherein the data transformation algorithm comprises at least one of data rotation, cutting and stretching, shearing transformation and piecewise affine transformation;
and step 32, carrying out normalization processing on the enhanced two-dimensional audio spectrum, and taking the normalized two-dimensional audio spectrum as a training image of the initial nondestructive testing model.
6. The method for non-destructive inspection based on data-driven self-learning according to claim 5, wherein in step 32, the method for normalizing comprises:
arranging pixels in the enhanced two-dimensional audio spectrum from small to large according to the pixel value;
extracting the maximum value of the pixels in a preset proportion range from the front end of the sequenced pixels, and marking the maximum value as a normalized minimum value;
the maximum pixel value of the pixels in the enhanced two-dimensional audio spectrum is recorded as a normalized maximum value;
resetting the pixel value of the pixel in the enhanced two-dimensional audio spectrum to the normalized minimum value when the pixel value of the pixel in the enhanced two-dimensional audio spectrum is smaller than the normalized minimum value;
and carrying out normalization processing on the two-dimensional audio map after pixel reset according to the normalization minimum value and the normalization maximum value.
7. The data-driven self-learning based nondestructive testing method according to any one of claims 1 to 6, wherein at least one sensor is provided in the dynamic signal acquisition instrument, and the sensor is one of a displacement sensor, a speed sensor and an acceleration sensor.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010599968.6A CN111783616B (en) | 2020-06-28 | 2020-06-28 | Nondestructive testing method based on data-driven self-learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010599968.6A CN111783616B (en) | 2020-06-28 | 2020-06-28 | Nondestructive testing method based on data-driven self-learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN111783616A CN111783616A (en) | 2020-10-16 |
| CN111783616B true CN111783616B (en) | 2024-03-26 |
Family
ID=72760724
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010599968.6A Active CN111783616B (en) | 2020-06-28 | 2020-06-28 | Nondestructive testing method based on data-driven self-learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111783616B (en) |
Families Citing this family (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2022175972A1 (en) * | 2021-02-19 | 2022-08-25 | INDIAN INSTITUTE OF TECHNOLOGY MADRAS (IIT Madras) | Method and system for generating time-efficient synthetic non-destructive testing data |
| CN113139464B (en) * | 2021-04-23 | 2022-11-18 | 国网山东省电力公司青岛市即墨区供电公司 | Power grid fault detection method |
| CN114647762B (en) * | 2022-03-23 | 2023-01-24 | 中国水利水电科学研究院 | Dam detection method based on map comparison |
Citations (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105929024A (en) * | 2016-04-21 | 2016-09-07 | 安徽省（水利部淮河水利委员会）水利科学研究院 | Intelligent detection and quantitative recognition method for defect of concrete |
| CN106053597A (en) * | 2016-05-20 | 2016-10-26 | 湖南科技大学 | Detection method of steel pipe concrete cavity defect extracted on basis of HHT characteristics |
| CN106706751A (en) * | 2015-11-13 | 2017-05-24 | 波音公司 | Ultrasonic system for nondestructive testing |
| CN106841403A (en) * | 2017-01-23 | 2017-06-13 | 天津大学 | A kind of acoustics glass defect detection method based on neutral net |
| CN109142523A (en) * | 2018-08-14 | 2019-01-04 | 浙江核力建筑特种技术有限公司 | A kind of metope hollowing recognition quantitative analytic approach based on acoustic imaging |
| JP6527999B1 (en) * | 2018-11-22 | 2019-06-12 | 義昭 新納 | Internal state estimation system |
| CN110082107A (en) * | 2019-04-08 | 2019-08-02 | 四川安和精密电子电器有限公司 | Vibrating motor defect diagonsis device and defect identification method based on audio analysis |
| KR102008973B1 (en) * | 2019-01-25 | 2019-08-08 | (주)나스텍이앤씨 | Apparatus and Method for Detection defect of sewer pipe based on Deep Learning |
| CN110954826A (en) * | 2019-12-17 | 2020-04-03 | 四川安和精密电子电器股份有限公司 | Step screw motor defect diagnosis device and defect identification method based on audio analysis |
| EP3637099A1 (en) * | 2018-10-08 | 2020-04-15 | Ecole Polytechnique Federale de Lausanne (EPFL) | Image reconstruction method based on a trained non-linear mapping |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| MX351680B (en) * | 2012-05-31 | 2017-10-25 | Univ Connecticut | Methods apparatuses for defect diagnosis in a mechanical system. |
-
2020
- 2020-06-28 CN CN202010599968.6A patent/CN111783616B/en active Active
Patent Citations (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106706751A (en) * | 2015-11-13 | 2017-05-24 | 波音公司 | Ultrasonic system for nondestructive testing |
| CN105929024A (en) * | 2016-04-21 | 2016-09-07 | 安徽省（水利部淮河水利委员会）水利科学研究院 | Intelligent detection and quantitative recognition method for defect of concrete |
| CN106053597A (en) * | 2016-05-20 | 2016-10-26 | 湖南科技大学 | Detection method of steel pipe concrete cavity defect extracted on basis of HHT characteristics |
| CN106841403A (en) * | 2017-01-23 | 2017-06-13 | 天津大学 | A kind of acoustics glass defect detection method based on neutral net |
| CN109142523A (en) * | 2018-08-14 | 2019-01-04 | 浙江核力建筑特种技术有限公司 | A kind of metope hollowing recognition quantitative analytic approach based on acoustic imaging |
| EP3637099A1 (en) * | 2018-10-08 | 2020-04-15 | Ecole Polytechnique Federale de Lausanne (EPFL) | Image reconstruction method based on a trained non-linear mapping |
| JP6527999B1 (en) * | 2018-11-22 | 2019-06-12 | 義昭 新納 | Internal state estimation system |
| KR102008973B1 (en) * | 2019-01-25 | 2019-08-08 | (주)나스텍이앤씨 | Apparatus and Method for Detection defect of sewer pipe based on Deep Learning |
| CN110082107A (en) * | 2019-04-08 | 2019-08-02 | 四川安和精密电子电器有限公司 | Vibrating motor defect diagonsis device and defect identification method based on audio analysis |
| CN110954826A (en) * | 2019-12-17 | 2020-04-03 | 四川安和精密电子电器股份有限公司 | Step screw motor defect diagnosis device and defect identification method based on audio analysis |
Also Published As
| Publication number | Publication date |
|---|---|
| CN111783616A (en) | 2020-10-16 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111783616B (en) | Nondestructive testing method based on data-driven self-learning | |
| US20230333533A1 (en) | System and method for generating machine learning model with trace data | |
| CN110992349A (en) | Underground pipeline abnormity automatic positioning and identification method based on deep learning | |
| CN108711148A (en) | A kind of wheel tyre defect intelligent detecting method based on deep learning | |
| CN111783544B (en) | Method for building diamond milling grinding head state monitoring system for processing ceramic mobile phone backboard | |
| CN111275108A (en) | Method for performing sample expansion on partial discharge data based on generation countermeasure network | |
| CN115439458A (en) | Industrial image defect target detection algorithm based on depth map attention | |
| CN113553762A (en) | Neural network for analyzing welding spots based on welding curve and establishing method | |
| CN115294033A (en) | Tire belt layer difference level and misalignment defect detection method based on semantic segmentation network | |
| CN111402249B (en) | Image evolution analysis method based on deep learning | |
| CN113902710A (en) | Method and system for detecting surface defects of industrial parts based on anomaly detection algorithm | |
| CN116597350A (en) | Flotation process fault early warning method based on BiLSTM predictive deviation | |
| CN115410024A (en) | Power image defect detection method based on dynamic activation thermodynamic diagram | |
| CN117197085A (en) | Road rapid-inspection image pavement disease detection method based on improved YOLOv8 network | |
| CN112508946B (en) | Cable tunnel anomaly detection method based on antagonistic neural network | |
| CN114881938A (en) | Grain size detection method and system based on wavelet analysis and neural network | |
| CN113919396A (en) | Vibration signal and image characteristic machine tool cutter wear state monitoring method based on semi-supervised learning | |
| CN113607068A (en) | Method for establishing and extracting recognition model of photoacoustic measurement signal characteristics | |
| CN117173618B (en) | Ground penetrating radar cavity target identification method based on multi-feature sensing Faster R-CNN | |
| CN112926383B (en) | Automatic target identification system based on underwater laser image | |
| CN117152083B (en) | Ground penetrating radar road disease image prediction visualization method based on category activation mapping | |
| CN117541483B (en) | Structural performance evaluation method and system for diffusion plate | |
| CN117689880B (en) | Method and system for target recognition in biomedical images based on machine learning | |
| CN117538863A (en) | Method, device, equipment and medium for improving tunnel lining defect geological radar detection precision | |
| CN113344138A (en) | Image detection method based on attention mechanism |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |