WO2024108717A1 - 一种深度神经网络裂纹量化方法及装置、存储介质 - Google Patents

一种深度神经网络裂纹量化方法及装置、存储介质 Download PDF

Info

Publication number
WO2024108717A1
WO2024108717A1 PCT/CN2022/141622 CN2022141622W WO2024108717A1 WO 2024108717 A1 WO2024108717 A1 WO 2024108717A1 CN 2022141622 W CN2022141622 W CN 2022141622W WO 2024108717 A1 WO2024108717 A1 WO 2024108717A1
Authority
WO
WIPO (PCT)
Prior art keywords
crack
neural network
deep neural
cracks
eigenvalue
Prior art date
Application number
PCT/CN2022/141622
Other languages
English (en)
French (fr)
Inventor
郭静波
王艺钊
胡铁华
Original Assignee
清华大学
清华四川能源互联网研究院
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by 清华大学, 清华四川能源互联网研究院 filed Critical 清华大学
Publication of WO2024108717A1 publication Critical patent/WO2024108717A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/10—Pre-processing; Data cleansing
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • 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
    • 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

Definitions

  • the embodiments of the present disclosure relate to, but are not limited to, electromagnetic nondestructive testing, oil and gas storage and transportation safety, machine learning and other fields, and in particular to a deep neural network crack quantification method based on wavelet transform and eigenvalue screening, and the applicable objects include, but are not limited to, cracks on ferromagnetic materials such as ring weld cracks in oil and gas pipelines, pipe body cracks, and tank bottom plate weld cracks.
  • Electromagnetic nondestructive testing technologies such as magnetic flux leakage, dynamic magnetic field, and pulsed eddy current are the mainstream technologies for detecting cracks in oil and gas pipelines and oil storage tanks. After the cracks are detected, there are two difficulties in accurately quantifying the three-dimensional size of the cracks: first, the crack size is small and the inclination angles are different; second, crack quantification is an inverse problem of the electromagnetic field, which is an ill-posed problem and does not have a unique solution. At most, it seeks the optimal solution that conforms to reality.
  • the present disclosure provides a deep neural network crack quantification method, including:
  • Deep neural network models are trained for axial cracks, circumferential cracks, and inclined cracks respectively;
  • the axial cracks, annular cracks and inclined cracks are quantified using the corresponding trained deep neural network models respectively.
  • An embodiment of the present disclosure also provides a deep neural network crack quantization device, including a display for displaying quantization results, a memory for storing instructions, and a processor connected to the memory, wherein the processor is configured to execute the steps of the deep neural network crack quantization method described in any embodiment of the present disclosure based on the instructions stored in the memory.
  • An embodiment of the present disclosure also provides a storage medium on which a computer program is stored.
  • the program is executed by a processor, the deep neural network crack quantization method described in any embodiment of the present disclosure is implemented.
  • the deep neural network crack quantization method, detection device, and storage medium of the embodiments of the present disclosure can, through crack classification, strengthen the axial characteristics of axial cracks, the circumferential characteristics of circumferential cracks, and the inclination angle characteristics of inclined cracks in their respective categories; further, by aligning the wave peaks of inclined cracks, the quantization problem of inclined cracks can be converted into a quantization problem similar to that of circumferential cracks; by performing continuous wavelet transform on the time domain waveform of the crack, the time domain characteristics and frequency domain characteristics of the crack can be accurately extracted to improve the quantization accuracy of the crack; through the use of the eigenvalue screening mechanism, the eigenvalues that have a greater impact on the model quantization results can be classified and screened to participate in model training, which not only strengthens the effect of influential eigenvalues on specific types of quantization models, but also improves the computational efficiency of crack quantization by reducing the number of inefficient eigenvalues.
  • FIG1 is a schematic diagram of a process of a deep neural network crack quantification method according to an exemplary embodiment of the present disclosure
  • FIG2A is a schematic diagram of a framework structure of a deep neural network crack quantization method in a training phase according to an exemplary embodiment of the present disclosure
  • FIG2B is a schematic diagram of a framework structure of a deep neural network crack quantization method at a quantization stage according to an exemplary embodiment of the present disclosure
  • FIG3A is a schematic diagram showing the comparison results of the axial crack model training process with and without the eigenvalue screening mechanism
  • FIG3B is a schematic diagram showing the comparison results of a model training process for a circumferential crack with and without an eigenvalue screening mechanism
  • FIG3C is a schematic diagram showing the comparison results of the model training process of the inclined crack with and without the eigenvalue screening mechanism
  • FIG. 4A is a (X80 steel) Real natural gas pipeline magnetic leakage axial component MFLX channel crack distribution map
  • FIG. 4B shows a (X80 steel) Real natural gas pipeline magnetic flux leakage radial component MFLY channel crack distribution map
  • FIG. 4C shows a (X80 steel) Real natural gas pipeline magnetic flux leakage circumferential component MFLZ channel crack distribution map
  • Figure 5 shows a (X80 steel) Real natural gas pipeline dynamic magnetic channel crack distribution map
  • FIG6 is a schematic diagram of the structure of a deep neural network crack quantization device according to an exemplary embodiment of the present disclosure.
  • the present invention performs continuous wavelet transform on the time domain waveform of the crack to accurately extract the time domain characteristics and frequency domain characteristics of the crack and improve the quantization accuracy of the crack.
  • the eigenvalues that have a greater impact on the model quantization results are classified and screened to participate in model training. This can not only enhance the effect of influential eigenvalues on specific types of quantization models, but also improve the computational efficiency of the model by reducing the number of inefficient eigenvalues.
  • the embodiment of the present disclosure provides a deep neural network crack quantization method, comprising the following steps:
  • the axial cracks, circumferential cracks and inclined cracks are quantified using the trained corresponding deep neural network models respectively.
  • the output of the deep neural network model is the length, width and depth dimensions of the cracks; for inclined cracks, the output of the deep neural network model is the oblique length, oblique width and depth dimensions of the cracks.
  • the method when training a deep neural network model for an inclined crack, or when quantifying an inclined crack, the method further includes:
  • a crack data matrix containing the inclined crack peak is selected, the lateral dimension of the crack data matrix represents the sampling point, the longitudinal dimension of the crack data matrix represents the number of channels, and the element value of the crack data matrix represents the detection data value;
  • the peaks of other rows in the crack data matrix are translated to the left or right and aligned with the first peak;
  • a new crack data matrix is formed by selecting intervals of equal width on the left and right sides of the aligned wave peak as the symmetry axis, and the eigenvalue vector of the inclined crack is extracted from this new crack data matrix.
  • deep neural network models are trained separately, including:
  • the crack types include axial cracks, circumferential cracks and inclined cracks, and extract the original eigenvalue vector of each crack original sample data file;
  • the optimal eigenvalue vector corresponding to each classification is screened out from the original eigenvalue vector corresponding to each classification, and the optimal eigenvalue screening model and the optimal deep neural network hidden layer model of each classification are generated at the same time.
  • the number of eigenvalues contained in the optimal eigenvalue vector is less than or equal to the number of eigenvalues contained in the original eigenvalue vector.
  • the optimal eigenvalue screening model is used to screen the corresponding optimal eigenvalue vector for each classification, and the optimal deep neural network hidden layer model is used as the hidden layer of the deep neural network to quantify cracks.
  • step 101 may include the following steps:
  • the original crack sample data files can be obtained through pulling experiments; for tank bottom plates, the original crack sample data files can be obtained through steel plate experiments, and so on; these original crack sample data files are the original materials for training neural network models.
  • a section of artificial crack defective pipe with the same material and diameter as the actual detection pipe can be prepared in advance; a large number of artificial cracks with different lengths, widths, depths and inclination angles are processed at different positions of the artificial crack defective pipe; then, a detector pulling experiment is carried out in the oil and gas pipeline on this section of artificial crack defective pipe to obtain the pulling data of the artificial crack; the above pulling data is imported into the crack analysis software, the suspected crack area is determined by the crack automatic detection algorithm, and the data matrix contained in all suspected crack areas is saved as a text file (.txt), which constitutes the crack original sample data file set.
  • each crack original sample data file may include one or more crack data matrices.
  • the multiple crack data matrices may include crack data matrices of the leakage magnetic axial component MFLX channel, the leakage magnetic radial component MFLY channel, the leakage magnetic circumferential component MFLZ channel, and the dynamic magnetic channel.
  • each crack original sample data file may also include data such as the detector movement speed and the nominal wall thickness.
  • the material will affect physical parameters such as magnetization intensity and magnetic permeability, which in turn will affect the non-destructive testing measurement results of cracks. Therefore, the first level of classification is based on the "material" of the object being tested.
  • step 2) the crack original sample data file set is further classified according to the crack type
  • the classification is based on the inclination angle of the crack.
  • the crack original sample data file set is divided into three types: axial crack original sample data file set, circumferential crack original sample data file set and inclined crack original sample data file set.
  • an axial crack refers to a crack with an inclination angle equal to or close to 0° and an inclination angle close to 180°
  • a circumferential crack refers to a crack with an inclination angle equal to or close to 90°
  • an inclined crack refers to a crack with an inclination angle not equal to or close to 0° and not equal to or close to 90°.
  • an axial crack may be a crack with an inclination angle between [0°, ⁇ °) and [180°- ⁇ °, 180°)
  • a circumferential crack may be a crack with an inclination angle between [90°- ⁇ °, 90°+ ⁇ °)
  • an inclined crack refers to a crack with an inclination angle between [ ⁇ 0°, 90°- ⁇ °) and [90°+ ⁇ °, 180°- ⁇ °) and between.
  • axial cracks refer to cracks with inclination angles between 0 and 10° and between 170° and 180°
  • circumferential cracks refer to cracks with inclination angles between 80 and 100°
  • inclined cracks refer to cracks with inclination angles between 10° and 80° and between 100° and 170°.
  • crack original eigenvalue vectors may come from one or more of the following data: three-axis Hall element detection data in magnetic flux leakage detection, dynamic magnetic detection data, detector movement speed, nominal wall thickness, etc.
  • the crack original eigenvalue vector includes multiple eigenvalues, and illustratively, the eigenvalues include: the maximum transformation coefficient of the wavelet transform, the optimal scale factor of the wavelet transform, etc.
  • the maximum transformation coefficient of the wavelet transform and the optimal scale factor of the wavelet transform are obtained by performing continuous wavelet transform on the crack time domain waveform with the largest peak value in the crack data matrix.
  • the crack original sample data files may be classified first, and then the crack original eigenvalue vectors may be extracted from each crack original sample data file in the classification to form crack original eigenvalue vector sets for different materials and different crack types.
  • the optimal eigenvalue vector of the crack sample is screened out through the eigenvalue screening mechanism. While screening out the optimal eigenvalue vector of the crack sample, a trained deep neural network model can be obtained.
  • the deep neural network model includes: an input layer, a hidden layer, and an output layer. The number of nodes in the input layer is equal to the number of eigenvalues contained in the screened optimal eigenvalue vector.
  • the eigenvalue screening mechanism selects part or all of the eigenvalues that minimize the overall eigenvalue screening evaluation function of the deep neural network model from the eigenvalues contained in the extracted original crack eigenvalue vector to form the optimal eigenvalue vector of the crack sample.
  • the mapping speed of the subsequent deep neural network model will be improved, and the calculation accuracy will also be improved.
  • the eigenvalue vector is not directly extracted from the original sample data file, but the peaks of the inclined crack are first aligned, and then the eigenvalue vector is extracted from the aligned sample data.
  • the process of “tilted crack peak alignment” includes the following steps:
  • step 402 when performing peak alignment on inclined cracks, in step 402, the first peak appearing from the right in the crack data matrix can also be found, then, in step 403, the peaks of other rows in the crack data matrix are translated to the right and aligned with the first peak in step 402.
  • the step of extracting the maximum transform coefficient of the eigenvalue wavelet transform and the optimal scale factor of the wavelet transform includes:
  • x[k] is the observed value of the kth point of the signal channel with the largest peak value
  • ⁇ [k] is the basic wavelet or mother wavelet
  • ⁇ * [k] represents the conjugate operation of ⁇ [k]
  • dj is the sampling step
  • dj is a real number
  • the * in ⁇ *a is a multiplication sign
  • ⁇ is the half width of ⁇ [k]
  • ⁇ is a real number
  • ⁇ can be 5, a ⁇ [a 1 , a max1 ], called the scale factor, which characterizes the width (oblique width) of the crack;
  • ⁇ a, b [k] are wavelet basis functions;
  • ⁇ x[k], ⁇ a, k [k]> are wavelet transform coefficients, which characterize the length (o
  • the transformation coefficient matrix obtained after wavelet transformation is:
  • WT is a matrix of max1*max2 dimensions.
  • selecting an optimal eigenvalue vector corresponding to a classification from an original eigenvalue vector corresponding to the classification includes:
  • the number of nodes in the input layer of the deep neural network model is determined according to the current number of eigenvalue screening
  • the deep neural network model is constructed according to the number of nodes in the input layer, the number of nodes in the output layer, and the number of hidden layers
  • m training sets and n test sets are input into the deep neural network model to obtain the training set error matrix and the test set error matrix
  • the eigenvalue screening evaluation function is calculated according to the training set error matrix and the test set error matrix, where p is the number of eigenvalues contained in each original eigenvalue vector, 0 ⁇ q ⁇ p;
  • the number of eigenvalue screening that minimizes the corresponding eigenvalue screening evaluation function is taken as the optimal number of input eigenvalues screened out, and the eigenvalue screening model and the deep neural network hidden layer model corresponding to the optimal number of input eigenvalues are taken as the optimal eigenvalue screening model and the optimal deep neural network hidden layer model corresponding to the classification.
  • an optimal eigenvalue vector of a crack sample is screened out through an eigenvalue screening mechanism to obtain a trained deep neural network model, and the eigenvalue screening mechanism includes the following steps:
  • the initial value of the eigenvalue screening evaluation function of the crack type is set to Gi_min ; in the embodiment of the present disclosure, the initial value of Gi_min can be set to a larger value, so that there is at least one value in the values of the eigenvalue screening evaluation function Gi corresponding to the subsequent eigenvalue screening numbers q to p that is smaller than the initial value of Gi_min , so that the initial value of Gi_min can be replaced by the value of Gi , or the initial value of Gi_min can be directly assigned with the value of the eigenvalue screening evaluation function Gi corresponding to the eigenvalue screening number q calculated subsequently.
  • a “eigenvalue screening model” is established by using the recursive feature elimination method, and q eigenvalues are screened out from the p original eigenvalues to form a new training set eigenvalue matrix Xi_train_s and a new test set eigenvalue matrix Xi_test_s , where the dimension of Xi_train_s is m ⁇ q and the dimension of Xi_test_s is n ⁇ q;
  • the current “eigenvalue screening model” saved in the cache RAMi is used as the “optimal eigenvalue screening model”;
  • the current “deep neural network hidden layer model” saved in the cache RAMi is used as the “optimal deep neural network hidden layer model”;
  • a recursive feature elimination method may be called through the Python package sklearn.feature_selection() to establish a feature value screening model, which may screen out q feature values from p original feature values.
  • a deep neural network hidden layer model may be constructed by using the Python packages keras.models.Sequential() and keras.layers.Dense().
  • the characteristic value screening evaluation function Gi is calculated according to the following method:
  • the subscript train represents the training set, and the subscript test represents the test set.
  • the single precision vector of the error interval estimate and the median vector of the confidence interval are:
  • the coefficient ⁇ can be set according to the value of the confidence level. For example, when the confidence level is 90%, the coefficient ⁇ is 3.28.
  • the eigenvalue screening evaluation function of the training set and the test set is calculated by formula (8):
  • Both ⁇ 1 and ⁇ 2 are between [0, 1].
  • the first weight coefficient ⁇ 1 may be 0.5; and the second weight coefficient ⁇ 2 may be 0.5.
  • the eigenvalue screening mechanism of the disclosed embodiment compares the Gi values corresponding to different numbers of eigenvalue combinations, and selects the set of eigenvalue vectors with the smallest Gi value to construct a deep neural network model.
  • axial cracks, annular cracks, and inclined cracks are quantified using the corresponding trained deep neural network models, respectively, including:
  • the original eigenvalue vector is input into the optimal eigenvalue screening model corresponding to the classification of the crack original data file to screen out the optimal eigenvalue vector;
  • the selected optimal eigenvalue vector is used as the input layer of the deep neural network; the optimal deep neural network hidden layer model corresponding to the classification of the crack original data file is used as the hidden layer of the deep neural network; the length, width, and depth, or oblique length, oblique width and depth of the crack are used as the output of the deep neural network.
  • step 102 may include the following steps:
  • the detection data of the pipeline detector can be imported into the crack analysis software, the suspected crack area can be determined by the "crack automatic detection algorithm", and the crack data matrix contained in the suspected crack area can be saved as a text file (.txt), which constitutes the crack raw data file;
  • step II classifying the crack raw data file of step I) according to the material of the object to be tested;
  • the crack raw data files are further classified according to the crack type. According to the inclination angle of the crack, the crack raw data files are classified into one of the three types: axial crack raw data files, circumferential crack raw data files, or inclined crack raw data files;
  • the original eigenvalue vector may come from one or more of the following data: three-axis Hall element detection data in magnetic flux leakage detection, dynamic magnetic detection data, detector movement speed, nominal wall thickness, etc.
  • the eigenvalues contained in the original eigenvalue vector may include the maximum transformation coefficient of wavelet transform, the optimal scale factor of wavelet transform, etc.;
  • step V Substituting the original eigenvalue vector of step IV) into the optimal eigenvalue screening model for the corresponding material and the corresponding crack type in step 101, and screening out the optimal eigenvalue vector therefrom;
  • step VI) Input the optimal eigenvalue vector selected in step V) into a deep neural network, the hidden layer of which is the optimal deep neural network hidden layer model for the corresponding material and the corresponding crack type in step 101, and the output of which is the length (oblique length), width (oblique width), and depth data of the crack.
  • the inclined cracks instead of directly extracting the original eigenvalue vector from the original sample data file of the crack, the inclined cracks are firstly peak aligned, and then the original eigenvalue vector is extracted from the aligned data.
  • the specific process please refer to the relevant steps of the training phase.
  • the above step VI) for inclined cracks, a mapping relationship is established between the sample data of the peak alignment and the oblique length, oblique width, and depth of the crack, and finally the oblique length, oblique width, depth and other dimensions of the inclined crack are used as the output of the deep neural network.
  • the deep neural network crack quantization method disclosed in the present invention can, through crack classification, strengthen the axial characteristics of axial cracks, the circumferential characteristics of circumferential cracks, and the inclination angle characteristics of inclined cracks in their respective categories; by aligning the wave peaks of inclined cracks, the quantization problem of inclined cracks can be converted into a quantization problem similar to that of circumferential cracks; by performing continuous wavelet transform on the time domain waveform of the crack, the time domain characteristics and frequency domain characteristics of the crack can be accurately extracted to improve the quantization accuracy of the crack; through the use of the eigenvalue screening mechanism, the eigenvalues that have a greater impact on the model quantization results can be classified and screened to participate in the model training; this can not only strengthen the effect of influential eigenvalues on specific types of quantization models, but also improve the computational efficiency of crack quantization by reducing the number of inefficient eigenvalues.
  • This embodiment uses a crack detection method based on the fusion of magnetic flux leakage and dynamic magnetic data to determine the suspected crack area, and obtains the crack original sample data file set and crack original data file set.
  • magnetic flux leakage detection technology has many advantages such as simple principle, easy engineering implementation, and high detection efficiency; dynamic magnetic detection technology has high sensitivity to detecting crack defects in any direction, which complements the advantages of magnetic flux leakage detection.
  • the implementation steps of the method are introduced in detail below.
  • the entire implementation process is divided into two stages.
  • the first stage is the establishment stage of the "optimal eigenvalue screening model” and the “optimal deep neural network hidden layer model” (model training stage); the second stage is the use stage of the deep neural network crack quantification method.
  • the steps to establish the “optimal feature value screening model” and the “optimal deep neural network hidden layer model” include:
  • step 2) the crack original sample data file set is classified according to the material of the object to be tested.
  • the material level there is only one type of X80 steel;
  • the crack original sample data file set is further classified according to the crack type to obtain three crack original sample data file sets.
  • the crack original sample data file set of material X80 is divided into an axial crack original sample data file set (90 axial crack original sample data files), a circumferential crack original sample data file set (120 circumferential crack original sample data files) and an inclined crack original sample data file set (64 inclined crack original sample data files);
  • step 4 extract the original eigenvalue vector for each data file in the three crack original sample data file sets, and form the original eigenvalue vector set of axial crack, the original eigenvalue vector set of circumferential crack, and the original eigenvalue vector set of inclined crack (for inclined crack, first perform peak alignment on it).
  • a total of 58 original eigenvalues are extracted here to form the original eigenvalue vector, which are: 15 eigenvalues extracted from the waveform data of the axial component of the leakage magnetic field MFLX; 15 eigenvalues extracted from the waveform data of the radial component of the leakage magnetic field MFLY; 15 eigenvalues extracted from the waveform data of the circumferential component of the leakage magnetic field MFLZ; 5 eigenvalues extracted from the waveform data of the dynamic magnetic signal DM, as well as the detector running speed value and the nominal wall thickness value.
  • gaus2 wavelet, gaus1 wavelet, and gaus5 wavelet are selected from the Python wavelet library function to represent the MFLX signal, MFLY signal, and DM signal respectively; through continuous wavelet transform, the largest wavelet transform coefficients are extracted from the MFLX channel signal with the maximum peak value, the MFLY channel signal, and the DM channel signal. And the corresponding wavelet optimal scaling factor As the other 6 eigenvalues, these 6 wavelet eigenvalues can very finely characterize the length (oblique length), width (oblique width), and depth information of the crack.
  • step 5 for the original crack eigenvalue vector sets in the three crack types under the X80 steel catalog, the "optimal eigenvalue screening model” and “optimal deep neural network hidden layer model” are established according to the eigenvalue screening mechanism, and the specific operations are as follows:
  • 5d Construct a "deep neural network hidden layer model"; by importing the keras.models.Sequential() and keras.layers.Dense() packages in Python, a 6-layer fully connected neural network model is built, for example.
  • the first layer is the input layer, and the dimension of the input layer is equal to the number of filtered eigenvalues.
  • the middle 4 layers are hidden layers (in other exemplary embodiments, the number of hidden layers can also be 5 or 6 layers), and the number of nodes is 30, 18, 20, and 20 respectively.
  • the last layer is the output layer, which has 3 nodes, corresponding to the length (oblique length), width (oblique width), and depth results of crack quantization;
  • the current “eigenvalue screening model” saved in the cache RAMi is used as the “optimal eigenvalue screening model”;
  • the current “deep neural network hidden layer model” saved in the cache RAMi is used as the “optimal deep neural network hidden layer model”;
  • step 5j Return to step 5a) until all the original crack eigenvalue vector sets in the three crack types under the X80 steel catalog are screened, and the "optimal eigenvalue screening model” and “optimal deep neural network hidden layer model” in each crack type under each material catalog are output and saved.
  • the "optimal eigenvalue screening model” can be saved as a .pkl file
  • the "optimal deep neural network hidden layer model” can be saved as a .h5 file.
  • the optimal eigenvalue vector selected by the "optimal eigenvalue screening model” is used as the input layer of the deep neural network.
  • the "optimal deep neural network hidden layer model” corresponding to each classification is also obtained. These two models are part of the deep neural network model.
  • the trained corresponding deep neural network models can be used for quantification, and the optimal DNN quantification models of the three cracks and the crack length (oblique length), width (oblique width) and depth results can be output.
  • Table 1 shows the crack quantification effect achieved by the above method for 274 cracks in the pulling field. It can be seen that at a confidence level of 90%, the overall confidence intervals of the length and width errors of all cracks, whether axial cracks, circumferential cracks, or inclined cracks, are within the range of ⁇ 10mm; the overall confidence intervals of the depth errors are within the range of ⁇ 10%wt, where %wt represents the percentage of the crack depth relative to the nominal wall thickness (15.30mm).
  • Figures 3A to 3C are schematic diagrams of the comparison results of the optimal deep neural network hidden layer model with and without the eigenvalue screening mechanism in the training process of the axial crack, the circumferential crack and the inclined crack, respectively, wherein t_loss_new, v_loss_new are the training set loss function and the validation set loss function of the method using the eigenvalue screening mechanism (the method in the present disclosure); t_loss_old, v_loss_old are the training set loss function and the validation set loss function of the traditional method without the eigenvalue screening mechanism.
  • the crack quantization model trained by the eigenvalue screening mechanism has a lower loss function in both the training set and the validation set. This shows that the present disclosure has a higher crack quantization accuracy than the traditional neural network method.
  • the deep neural network crack quantification method based on wavelet transform and eigenvalue screening is used in the following steps:
  • step I the original crack data file of the object under test is obtained.
  • X80 steel, nominal wall thickness 15.30mm An internal inspection experiment was carried out on a real natural gas pipeline. Several artificial cracks with different inclination angles were processed on the pipe body. The length (oblique length), width (oblique width), depth and inclination angle of these artificial cracks were measured in advance by vernier calipers and protractors. It is particularly pointed out that the artificial cracks on this section of retired pipeline do not participate in the first stage of deep neural network crack quantification model training, that is, they do not participate in the establishment process of "optimal eigenvalue screening model” and "optimal deep neural network hidden layer model", but only participate in the second stage of the quantification method test.
  • the detection data of the pipeline detector is imported into the crack analysis software, and the suspected crack area is determined by the crack detection method based on the fusion of leakage magnetic and dynamic magnetic data.
  • the crack data matrix contained in the suspected crack area is saved as a text file (.txt), which constitutes the crack original data file.
  • .txt constitutes the crack original data file.
  • 12 real cracks were detected in a certain area of the test data by the crack analysis software.
  • the area framed by the box in the figure is the suspected crack area automatically detected by the crack analysis software.
  • Figures 4A to 4C and L1-L12 in Figure 5 represent the common 12 cracks, but they are displayed separately using the data of the four channels MFLX/MFLY/MFLZ/DM.
  • each crack raw data file For each crack, a corresponding crack raw data file is established, a total of 12 crack raw data files, each crack raw data file includes the crack data matrix of the four channels MFLX/MFLY/MFLZ/DM, and also includes data such as the detector movement speed and nominal wall thickness.
  • the crack raw data files are classified according to the material of the object to be tested; in this embodiment, at the material level, there is only one type of X80 steel, so the 12 crack raw data files belong to the same category in terms of material;
  • the crack raw data files are further classified according to the crack type.
  • cracks with an inclination angle of [0°, 10°) and [170°, 180°) are defined as axial cracks
  • cracks with an inclination angle of [80°, 100°) are defined as circumferential cracks
  • cracks with an inclination angle of [10°, 80°) and [100°, 170°) are defined as inclined cracks; as shown in Figures 4A to 4C and Figure 5, L1-L4 are inclined cracks, L5-L8 are axial cracks, and L9-L12 are circumferential cracks, that is, the 12 crack raw data files are divided into 3 categories: X80 inclined crack raw data files, X80 axial crack raw data files, and X80 circumferential crack raw data files;
  • step IV the original eigenvalue vector of the crack original data file is extracted (for inclined cracks, it is first peak aligned).
  • the eigenvalue types contained in the original eigenvalue vector are the same as the 58 eigenvalue types in step 4) of the first stage;
  • step V) the original eigenvalue vector is substituted into the "optimal eigenvalue screening model" under the corresponding material and crack type in the first stage, and the optimal eigenvalue vector corresponding to each crack original data file is screened out;
  • each optimal eigenvalue vector screened out in step V) is used as the input layer of the deep neural network;
  • the hidden layer of the deep neural network is the optimal deep neural network hidden layer model under the corresponding material and crack type in the first stage, and the output of the deep neural network is the length (oblique length), width (oblique width), and depth data of the crack.
  • Table 2 is the above (X80 steel, nominal wall thickness 15.30 mm)
  • the vernier caliper and protractor measurement values of 12 cracks in the real natural gas pipeline are shown in Table 3.
  • Quantitative estimation values of 12 cracks in a real natural gas pipeline are shown in Table 3.
  • Table 4 shows the error between the quantitative estimation values in Table 3 and the measured values in Table 2. It can be seen that no matter it is an axial crack, annular crack, or inclined crack, the length (oblique length) and width (oblique width) errors of all cracks are within the range of ⁇ 10mm; the depth error is within the range of ⁇ 10%wt, where %wt represents the percentage of the crack depth relative to the nominal wall thickness (15.30mm). It can be seen from Tables 2 to 4 that the deep neural network crack quantification method based on wavelet transform and eigenvalue screening provided by the present disclosure has a high crack quantification accuracy.
  • An embodiment of the present disclosure also provides a deep neural network crack quantization device, comprising a display for displaying crack quantization results, a memory for storing crack quantization methods and temporary results; and a processor connected to the memory, wherein the processor executes the steps of the deep neural network crack quantization method as described in any of the preceding items based on instructions stored in the memory.
  • a deep neural network crack quantification device may include: a processor 610, a memory 620, a bus system 630, and a display 640, wherein the processor 610, the memory 620, and the display 640 are connected via the bus system 630, the memory 620 is used to store instructions and optimal eigenvalue vectors, optimal eigenvalue screening models, and optimal deep neural network hidden layer models, etc., and the processor 610 is used to execute the instructions stored in the memory 620 to perform crack quantification through a deep neural network.
  • the processor 610 can train the "optimal eigenvalue screening model” and the “optimal deep neural network hidden layer model” for axial cracks, circumferential cracks, and inclined cracks, respectively; and quantify the axial cracks, circumferential cracks, and inclined cracks using the trained corresponding models, respectively; and finally display the quantization results through the display 640.
  • processor 610 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.
  • DSP digital signal processors
  • ASIC application-specific integrated circuits
  • FPGA field-programmable gate arrays
  • a general-purpose processor may be a microprocessor or the processor 610 may be any conventional processor, etc.
  • the memory 620 may include a read-only memory and a random access memory, and provide instructions and data, including the optimal eigenvalue vector, the deep neural network model, etc., to the processor 610.
  • a portion of the memory 620 may also include a non-volatile random access memory.
  • the memory 620 may also store information about the device type.
  • the bus system 630 may also include a power bus, a control bus, a status signal bus, and the like.
  • the display 640 can also display the crack detection data in the memory 620 through the crack analysis software.
  • the processing performed by the deep neural network crack quantization device can be completed by the hardware integrated logic circuit in the processor 610 or the instructions in the form of software. That is, the steps of the deep neural network crack quantization method of the embodiment of the present disclosure can be executed by a hardware processor, or by a combination of hardware and software modules in the processor 610.
  • the software module can be located in a storage medium such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register, etc.
  • the storage medium is located in the memory 620, and the processor 610 reads the information in the memory 620 and completes the steps of the above method in combination with its hardware. To avoid repetition, it will not be described in detail here.
  • the embodiments of the present disclosure also provide a storage medium storing executable instructions.
  • the deep neural network crack quantification method can train deep neural network models for axial cracks, circumferential cracks, and inclined cracks, respectively; and the axial cracks, circumferential cracks, and inclined cracks can be quantified using the trained corresponding deep neural network models, respectively.
  • the method for implementing crack detection by executing executable instructions is basically the same as the deep neural network crack quantification method provided in the above embodiments of the present disclosure, and will not be described in detail here.
  • Such software may be distributed on a computer-readable medium, which may include a computer storage medium (or non-transitory medium) and a communication medium (or temporary medium).
  • a computer storage medium includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data).
  • Computer storage media include, but are not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital versatile disks (DVD) or other optical disk storage, magnetic cassettes, magnetic tapes, magnetic disk storage or other magnetic storage devices, or any other medium that can be used to store the desired information and can be accessed by a computer.
  • communication media typically contain computer-readable instructions, data structures, program modules, or other data in a modulated data signal such as a carrier wave or other transport mechanism, and may include any information delivery media.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Physics & Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Computational Linguistics (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Health & Medical Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Investigating Or Analyzing Materials By The Use Of Magnetic Means (AREA)

Abstract

一种深度神经网络裂纹量化方法，包括：针对轴向裂纹、环向裂纹和倾斜裂纹，分别训练深度神经网络模型；对所述轴向裂纹、环向裂纹和倾斜裂纹，分别使用训练好的对应的深度神经网络模型进行量化。本公开通过裂纹分类，将轴向裂纹、环向裂纹、倾斜裂纹的特征在各自的类别下进行强化；进一步地，通过对齐倾斜裂纹的波峰，将倾斜裂纹的量化问题转化为类似环向裂纹的量化问题；通过对裂纹的时域波形进行连续小波变换，准确提取裂纹的时域特征和频域特征，提高裂纹的量化精度；通过特征值筛选机制的使用，不仅可以强化有影响力的特征值对特定类型量化模型的作用效果，还能通过减少低效特征值来提升裂纹量化的计算效率。

Description

一种深度神经网络裂纹量化方法及装置、存储介质
本申请要求于2022年11月21日提交中国专利局、申请号为2022114532446、发明名称为“一种深度神经网络裂纹量化方法及装置、存储介质”的中国专利申请的优先权，其内容应理解为通过引用的方式并入本申请中。
技术领域
本公开实施例涉及但不限于电磁无损检测、石油天然气储运安全、机器学习等领域，尤其涉及一种基于小波变换和特征值筛选的深度神经网络裂纹量化方法，适用对象包括但不限于油气管道环焊缝裂纹、管体裂纹，储罐底板焊缝裂纹等铁磁性材料上的裂纹。
背景技术
漏磁、动磁、脉冲涡流等电磁无损检测技术是检测油气管道、石油储罐裂纹的主流技术。在检测到裂纹之后，如何精确地量化裂纹的三维尺寸具有两个难点：其一是裂纹尺寸较小、倾斜角各异；其二是裂纹量化属于电磁场的反问题，它属于不适定问题，不具有唯一解，至多是寻求符合实际的最优解。
发明内容
以下是对本文详细描述的主题的概述。本概述并非是为了限制权利要求的保护范围。
本公开实施例提供了一种深度神经网络裂纹量化方法，包括：
针对轴向裂纹、环向裂纹和倾斜裂纹，分别训练深度神经网络模型；
对所述轴向裂纹、环向裂纹和倾斜裂纹，分别使用训练好的对应的深度神经网络模型进行量化。
本公开实施例还提供了一种深度神经网络裂纹量化装置，包括显示量化结果的显示器、存储指令的存储器和连接至所述存储器的处理器，所述处理 器被配置为基于存储在所述存储器中的指令，执行本公开任一实施例所述的深度神经网络裂纹量化方法的步骤。
本公开实施例还提供了一种存储介质，其上存储有计算机程序，该程序被处理器执行时实现本公开任一实施例所述的深度神经网络裂纹量化方法。
本公开实施例的深度神经网络裂纹量化方法及检测装置、存储介质，通过裂纹分类，可以使轴向裂纹的轴向特征，环向裂纹的环向特征，倾斜裂纹的倾斜角特征在各自的类别下进行强化；进一步地，通过对齐倾斜裂纹的波峰，可以将倾斜裂纹的量化问题转化为类似环向裂纹的量化问题；通过对裂纹的时域波形进行连续小波变换，可以准确提取裂纹的时域特征和频域特征，提高裂纹的量化精度；通过特征值筛选机制的使用，可以分类筛选出对模型量化结果影响比较大的特征值参与模型训练，这样不仅可以强化有影响力的特征值对特定类型量化模型的作用效果，还能通过减少低效特征值的数量提升裂纹量化的计算效率。
在阅读理解了附图和详细描述后，可以明白其他方面。
附图说明
附图用来提供对本公开技术方案的理解，并且构成说明书的一部分，与本公开的实施例一起用于解释本公开的技术方案，并不构成对本公开技术方案的限制。
图1为本公开示例性实施例一种深度神经网络裂纹量化方法的流程示意图；
图2A为本公开示例性实施例一种深度神经网络裂纹量化方法在训练阶段的框架结构图；
图2B为本公开示例性实施例一种深度神经网络裂纹量化方法在量化阶段的框架结构图；
图3A为一种对轴向裂纹进行模型训练过程中有特征值筛选机制和无特征值筛选机制的对比结果示意图；
图3B为一种对环向裂纹进行模型训练过程中有特征值筛选机制和无特征值筛选机制的对比结果示意图；
图3C为一种对倾斜裂纹进行模型训练过程中有特征值筛选机制和无特征值筛选机制的对比结果示意图；
图4A为一种
Figure PCTCN2022141622-appb-000001
(X80钢材)真实天然气管道漏磁轴向分量MFLX通道裂纹分布图；
图4B为一种
Figure PCTCN2022141622-appb-000002
(X80钢材)真实天然气管道漏磁径向分量MFLY通道裂纹分布图；
图4C为一种
Figure PCTCN2022141622-appb-000003
(X80钢材)真实天然气管道漏磁周向分量MFLZ通道裂纹分布图；
图5为一种
Figure PCTCN2022141622-appb-000004
(X80钢材)真实天然气管道动磁通道裂纹分布图；
图6为本公开示例性实施例一种深度神经网络裂纹量化装置的结构示意图。
具体实施方式
为使本公开的目的、技术方案和优点更加清楚明白，下文中将结合附图对本公开的实施例进行详细说明。需要说明的是，在不冲突的情况下，本公开中的实施例及实施例中的特征可以相互任意组合。
除非另外定义，本公开实施例公开使用的技术术语或者科学术语应当为本公开所属领域内具有一般技能的人士所理解的通常意义。本公开实施例中使用的“第一”、“第二”以及类似的词语并不表示任何顺序、数量或者重要性，而只是用来区分不同的组成部分。“包括”或者“包含”等类似的词语意指出该词前面的元件或物件涵盖出现在该词后面列举的元件或者物件及其等同，而不排除其他元件或者物件。
目前在裂纹量化方面主要采用神经网络等机器学习的方法。传统的神经网络算法量化裂纹对裂纹类型不加区分，直接建立特征向量与输出向量(例如几何尺寸)之间的映射关系。然而，不同类型裂纹的信号特征差别非常大， 而同类型裂纹的信号特征具有相似的变化规律，因此有必要分类别建立裂纹的量化模型。本公开提出“分类量化”的思想，将裂纹分为轴向裂纹、环向裂纹、倾斜裂纹3种类型，对这三种类型分别建立深度神经网络(Deep Neural Network，DNN)量化模型，通过裂纹分类，将轴向裂纹、环向裂纹、倾斜裂纹的特征在各自的类别下得到强化。
在模型训练之前，本公开对裂纹的时域波形进行连续小波变换，准确提取裂纹的时域特征和频域特征，提高裂纹的量化精度。另外，通过加入特征值筛选机制，分类筛选出对模型量化结果影响比较大的特征值参与模型训练。这样不仅可以强化有影响力的特征值对特定类型量化模型的作用效果，还能通过减少低效特征值的数量提升模型的计算效率。
如图1所示，本公开实施例提供了一种深度神经网络裂纹量化方法，包括如下步骤：
101、针对轴向裂纹、环向裂纹和倾斜裂纹，分别训练深度神经网络模型；
102、对所述轴向裂纹、环向裂纹和倾斜裂纹，分别使用训练好的对应的深度神经网络模型进行量化。
在一些示例性实施方式中，对轴向裂纹和环向裂纹，深度神经网络模型的输出为裂纹的长度、宽度和深度尺寸；对倾斜裂纹，深度神经网络模型的输出为裂纹的斜长、斜宽和深度尺寸。
在一些示例性实施方式中，在针对倾斜裂纹训练深度神经网络模型时，或者在对倾斜裂纹进行量化时，所述方法还包括：
对每个倾斜裂纹，选取包含倾斜裂纹波峰的裂纹数据矩阵，裂纹数据矩阵的横向维度表示采样点，裂纹数据矩阵的纵向维度表示通道数，裂纹数据矩阵的元素值表示检测数据值；
确定裂纹数据矩阵中从左数或从右数出现的第一个波峰；
将裂纹数据矩阵中其他行的波峰向左平移或向右平移，并与第一个波峰对齐；
以对齐后的波峰为对称轴左右选取等宽的区间形成新的裂纹数据矩阵，以此新的裂纹数据矩阵提取倾斜裂纹的特征值向量。
在一些示例性实施方式中，针对轴向裂纹、环向裂纹和倾斜裂纹，分别训练深度神经网络模型，包括：
获取裂纹原始样本数据文件集，所述裂纹原始样本数据文件集包括多个裂纹原始样本数据文件；
按照被测对象的材质和裂纹类型对裂纹原始样本数据文件进行分类，裂纹类型包括轴向裂纹、环向裂纹和倾斜裂纹，提取每个裂纹原始样本数据文件的原始特征值向量；
从每个分类对应的原始特征值向量中筛选出每个分类对应的最优特征值向量，同时生成每个分类的最优特征值筛选模型和最优深度神经网络隐藏层模型，最优特征值向量包含的特征值数目小于或等于原始特征值向量包含的特征值数目，最优特征值筛选模型用于为每个分类筛选对应的最优特征值向量，最优深度神经网络隐藏层模型用于作为深度神经网络的隐藏层，对裂纹进行量化。
在一些示例性实施方式中，参考图2A，步骤101可以包括如下步骤：
1)获得裂纹原始样本数据文件，多个裂纹原始样本数据文件组成裂纹原始样本数据文件集；
对于油气管道，可以通过牵拉实验获得裂纹原始样本数据文件；对于储罐底板，可以通过钢板实验获得裂纹原始样本数据文件，等等；这些裂纹原始样本数据文件是训练神经网络模型的原始材料。
以油气管道裂纹检测为例，可以根据实际检测管道的需求，事先准备一段与实际检测管道相同材质、相同管径的人工裂纹缺陷管；在人工裂纹缺陷管的不同位置加工有大量不同长度、宽度、深度、倾斜角的人工裂纹；然后在这段人工裂纹缺陷管上开展油气管道内检测器牵拉实验，由此获得人工裂纹的牵拉数据；将上述牵拉数据导入裂纹分析软件，通过裂纹自动检测算法确定裂纹可疑区域，将所有裂纹可疑区域包含的数据矩阵保存为文本文件(.txt)，即构成裂纹原始样本数据文件集。
本公开实施例中，每个裂纹原始样本数据文件可以包括一个或多个裂纹数据矩阵，示例性的，多个裂纹数据矩阵可以包括漏磁轴向分量MFLX通道、 漏磁径向分量MFLY通道、漏磁周向分量MFLZ通道及动磁通道的裂纹数据矩阵。另外，每个裂纹原始样本数据文件还可以包括检测器运动速度、标称壁厚等数据。
2)按照被测对象的材质对裂纹原始样本数据文件集进行分类；
材质会影响磁化强度、磁导率等物理参数，进而影响裂纹的无损检测测量结果，因此第一级别的分类依据是被测对象的“材质”。
3)在步骤2)的基础上，按照裂纹类型对裂纹原始样本数据文件集进一步进行分类；
分类依据是裂纹的倾斜角，根据不同的裂纹倾斜角，将裂纹原始样本数据文件集分为轴向裂纹原始样本数据文件集、环向裂纹原始样本数据文件集和倾斜裂纹原始样本数据文件集三种类型。
本公开实施例中，轴向裂纹指的是倾斜角等于或接近0°以及倾斜角接近180°的裂纹，环向裂纹指的是倾斜角等于或接近90°的裂纹，倾斜裂纹指的是倾斜角不等于或接近0°且不等于或接近90°的裂纹。假设误差角度为δ，轴向裂纹可以为倾斜角在[0°，δ°)以及[180°-δ°，180°)之间的裂纹，环向裂纹可以为倾斜角在[90°-δ°，90°+δ°)之间的裂纹，倾斜裂纹指的是倾斜角在[δ0°，90°-δ°)以及[90°+δ°，180°-δ°)以及之间的裂纹。示例性的，以δ＝10°为例，轴向裂纹指的是倾斜角在0到10°之间以及170°到180°之间的裂纹，环向裂纹指的是倾斜角在80到100°之间的裂纹，倾斜裂纹指的是倾斜角在10°到80°之间以及100°到170°之间的裂纹。
4)根据裂纹分类，对每个裂纹原始样本数据文件集的分类中的每个裂纹原始样本数据文件提取裂纹原始特征值向量，分别组成不同材质不同裂纹类型下的裂纹原始特征值向量集(假设全体共有Cnt个裂纹原始特征值向量集，示例性的，Cnt可以为材质数*裂纹类型数，裂纹类型数＝3)；
这些裂纹原始特征值向量可以来自以下一个或多个数据：漏磁检测中的三轴霍尔元件检测数据，动磁检测数据，检测器运动速度，标称壁厚等。
本公开实施例中，裂纹原始特征值向量包括多个特征值，示例性的，特 征值包括：小波变换的最大变换系数，小波变换的最优尺度因子等等。本公开实施例中，小波变换的最大变换系数以及小波变换的最优尺度因子均通过对裂纹数据矩阵中峰值最大的裂纹时域波形进行连续小波变换得到。
本公开实施例中，可以先对裂纹原始样本数据文件进行分类，然后对分类中的每个裂纹原始样本数据文件提取裂纹原始特征值向量，分别组成不同材质不同裂纹类型下的裂纹原始特征值向量集。
5)对每个材质目录下的每个裂纹类型的裂纹原始特征值向量集，分别通过特征值筛选机制筛选出裂纹样本的最优特征值向量，在筛选出裂纹样本的最优特征值向量的同时，可以得到训练好的深度神经网络模型，深度神经网络模型包括：输入层、隐藏层和输出层，输入层的节点数等于筛选出的最优特征值向量包含的特征值数目。所述特征值筛选机制从提取的裂纹原始特征值向量包含的特征值中，选择使该深度神经网络模型整体特征值筛选评价函数最小的部分或全部特征值，构成所述裂纹样本的最优特征值向量。
假设原始特征值向量数量为p，经过筛选的最优特征值向量数量为q，q<p，则后续深度神经网络模型的映射速度将会提高，计算精度也将提高。
在上述步骤4)中，对于倾斜裂纹，并非直接从原始样本数据文件中提取特征值向量，而是首先对倾斜裂纹进行波峰对齐，然后从对齐的样本数据中提取特征值向量。
在一些示例性实施方式中，“倾斜裂纹波峰对齐”的过程包括如下步骤：
401、选取包含倾斜裂纹波峰的裂纹数据矩阵，裂纹数据矩阵的横向表示采样点，纵向表示通道，元素值表示检测数据值；
402、找到裂纹数据矩阵中从左数出现的第一个波峰；
403、将裂纹数据矩阵中其他行的波峰向左平移并与步骤402中第一个波峰对齐；
404、以对齐后的波峰为对称轴左右选取等宽的区间形成新的裂纹数据矩阵，以此新的裂纹数据矩阵参与后续的特征值向量提取。
在另一些示例性实施方式中，对倾斜裂纹进行波峰对齐时，在步骤402 中，也可以找到裂纹数据矩阵中从右数出现的第一个波峰，那么，在步骤403中，将裂纹数据矩阵中其他行的波峰向右平移并与步骤402中第一个波峰对齐。
在一些示例性实施方式中，在上述步骤4)中，提取特征值小波变换的最大变换系数与小波变换的最优尺度因子的步骤包括：
411、从裂纹数据矩阵中选取峰值最大的通道信号；
412、对峰值最大的信号通道进行连续小波变换，变换公式为：
Figure PCTCN2022141622-appb-000005
其中，x[k]为峰值最大的信号通道的第k点的观测值，ψ[k]为基本小波或母小波，ψ *[k]表示对ψ[k]求共轭运算，dj为采样步长，dj为实数，Δ*a中的*为乘号，Δ为ψ[k]的半宽度，Δ为实数，示例性的，Δ可以为5，a∈[a 1，a max1]，称为尺度因子，它表征裂纹的宽度(斜宽)尺寸；b∈[b 1， bmax2]，称为平移量因子，它表征裂纹的空间位置，max1和max2均为大于1的自然数；ψ a，b[k]为小波基函数；<x[k]，ψ a，k[k]>为小波变换系数，表征裂纹的长度(斜长)和深度；
小波变换之后得到的变换系数矩阵为：
Figure PCTCN2022141622-appb-000006
WT为max1*max2维的矩阵。
413、从变换系数矩阵中找到最大的变换系数WT max作为提取出的小波变换的最大变换系数，选择与最大变换系数WT max对应的小波变换的最优尺度因子a best作为提取出的小波变换的最优尺度因子。
在一些示例性实施方式中，从一个分类对应的原始特征值向量中筛选出 该分类对应的最优特征值向量，包括：
将一个分类对应的原始特征值向量集分为m个训练集和n个测试集，m+n＝NUM，NUM为该分类对应的原始特征值向量集中包括的原始特征值向量数量；
对特征值筛选数量从初始值q到最大值p，分别执行如下操作：根据当前的特征值筛选数量确定深度神经网络模型的输入层的节点数，根据输入层的节点数、输出层的节点数以及隐藏层层数构建深度神经网络模型，将m个训练集和n个测试集输入深度神经网络模型，得到训练集误差矩阵和测试集误差矩阵，并根据训练集误差矩阵和测试集误差矩阵，计算特征值筛选评价函数，其中，p为每个原始特征值向量包含的特征值数量，0<q<p；
以使对应的特征值筛选评价函数最小的特征值筛选数量作为筛选出的最优输入特征值数量，并将最优输入特征值数量对应的特征值筛选模型和深度神经网络隐藏层模型作为该分类对应的最优特征值筛选模型和最优深度神经网络隐藏层模型。
在一些示例性实施方式中，在上述步骤5)中，对某个材质目录下的一个裂纹类型的裂纹原始特征值向量集，通过特征值筛选机制筛选出裂纹样本的最优特征值向量，得到训练好的深度神经网络模型，特征值筛选机制包括如下步骤：
511)将该裂纹原始特征值向量集随机拆分为m个训练集和n个测试集，m+n＝NUM，NUM表示该材质目录下该裂纹类型对应的裂纹原始样本数据文件数量；
512)假设每个原始特征值向量中包含的特征值数目为p，特征值筛选数量的初始值为q，并满足q<p，设定该裂纹类型的特征值筛选评价函数初始值为G i_min；本公开实施例中，G i_min的初始值可以设置为一个较大的值，以使得后续特征值筛选数量q到p对应的特征值筛选评价函数G i的值中至少存在一个值小于G i_min的初始值，以用该G i的值替换G i_min的初始值，或者，G i_min的初始值可以用后续计算出的特征值筛选数量q对应的特征值筛选评价函数 G i的值直接赋值。
513)根据本次特征值筛选数量q采用递归特征消除法建立“特征值筛选模型”，并从p个原始特征值中筛选出q个特征值，组成新的训练集特征值矩阵X i_train_s和新的测试集特征值矩阵X i_test_s，其中X i_train_s的维数为m×q，X i_test_s的维数为n×q；
514)根据特征值筛选数量q构建“深度神经网络隐藏层模型”；
515)将步骤513)中的X i_train_s和X i_test_s输入到步骤514)“深度神经网络隐藏层模型”，然后将输出代入公式(3)，求得当前该裂纹类型的训练集误差矩阵和测试集误差矩阵E i_train和E i_test，其中E i_train的维数为m×3，E i_test的维数为n×3；
516)根据公式(9)，计算当前第i类(i＝1,2,…,Cnt)裂纹的特征值筛选评价函数G i；
517)当G i≥G i_min时，直接跳转步骤518)；当G i＜G i_min时，替换G i_min(用G i替换G i_min)，并将当前G i对应的“特征值筛选模型”以及“深度神经网络隐藏层模型”作为当前最新的结果，并将它们保存到缓存RAMi中，之后跳转步骤518)；
518)更新特征值筛选数量q＝q+1，更新之后如果q>p则跳转步骤519)，否则跳转步骤513)；
519)以缓存RAMi中保存的当前“特征值筛选模型”为“最优特征值筛选模型”；以缓存RAMi中保存的当前“深度神经网络隐藏层模型”为“最优深度神经网络隐藏层模型”；
按照该方法，分别对每个材质目录下的每个裂纹类型中的所有裂纹原始特征值向量集进行筛选，输出并保存每个材质目录下的每个裂纹类型中“最优特征值筛选模型”和“最优深度神经网络隐藏层模型”。
在一些示例性实施方式中，在上述步骤513)中，可以通过Python程序包sklearn.feature_selection()调用递归特征消除法建立特征值筛选模型，该特征值筛选模型可以从p个原始特征值中筛选出q个特征值。
在一些示例性实施方式中，在上述步骤514)中，可以通过Python程序包keras.models.Sequential()和keras.layers.Dense()构建深度神经网络隐藏层模型。
在一些示例性实施方式中，在上述步骤5)中，特征值筛选评价函数G i根据如下方法进行计算：
501、定义训练集误差矩阵和测试集误差矩阵，分别表示为(3)式：
Figure PCTCN2022141622-appb-000007
其中，E il、E iw、E id分别表示第i类(i＝1,2,…,Cnt)裂纹的长(斜长)(l)、宽(斜宽)(w)、深度(d)的误差向量，下角标train表示训练集，下角标test表示测试集；
502、计算置信度为(1-α)的条件下，其中0＜α＜1，误差区间估计的单体精确度向量以及置信区间的中值向量分别为：
Figure PCTCN2022141622-appb-000008
Figure PCTCN2022141622-appb-000009
其中，A i_train，A i_test，M i_train和M i_test的维数都是1×3；
Figure PCTCN2022141622-appb-000010
和
Figure PCTCN2022141622-appb-000011
分别表示第i类(i＝1，2，……，Cnt)裂纹的训练集和测试集误差矩阵的均值；
Figure PCTCN2022141622-appb-000012
和
Figure PCTCN2022141622-appb-000013
分别表示置信度为(1-α)，m个样本和n个样本的t分布的上 分位点；
Figure PCTCN2022141622-appb-000014
和
Figure PCTCN2022141622-appb-000015
分别表示置信度为(1-α)，m个样本和n个样本的χ 2分布的上分位点；S m和S n分别表示训练集和测试集误差矩阵的样本标准差；
Figure PCTCN2022141622-appb-000016
其中φ -1(·)表示标准正态累积分布函数的逆函数。
Figure PCTCN2022141622-appb-000017
Figure PCTCN2022141622-appb-000018
系数γ可以根据置信度的值进行设置，示例性的，置信度为90％的条件下，系数γ为3.28。
503、将(4)式和(5)式用相同的权重向量合并处理为：
Figure PCTCN2022141622-appb-000019
Figure PCTCN2022141622-appb-000020
其中，假设第i类(i＝1，2，……，Cnt)裂纹长(斜长)、宽(斜宽)、深度的权重向量为ω i＝[ω il，ω iw，ω id]，满足权重向量ω i的每个元素都介于[0,1]之间，且ω il+ω iw+ω id＝1；
Figure PCTCN2022141622-appb-000021
表示对A i_train进行转置；
504、通过(8)式计算训练集和测试集的特征值筛选评价函数为：
Figure PCTCN2022141622-appb-000022
λ1和λ2均介于[0,1]之间；示例性的，第一权重系数λ1可以为0.5；第二权重系数λ2可以为0.5。
505、假设在第i类(i＝1，2，……，Cnt)裂纹信号样本集中，随机分为m个训练集样本以及n个测试集样本，所以通过(9)式可以计算针对第i类(i＝1，2，……，Cnt)裂纹信号的最终特征值筛选评价函数为：
Figure PCTCN2022141622-appb-000023
通过(9)式，可以定量评价每一种特征值组合对模型训练的影响效果，G i越小表明模型训练的效果越好。本公开实施例的特征值筛选机制，通过对比不同数量的特征值组合对应的G i值，从中选择G i值最小的那组特征值向量构建深度神经网络模型。
在一些示例性实施方式中，对轴向裂纹、环向裂纹和倾斜裂纹，分别使用训练好的对应的深度神经网络模型进行量化，包括：
获取被测对象的一个或多个裂纹原始数据文件；
按照被测对象的材质和裂纹类型对裂纹原始数据文件进行分类；提取裂纹原始数据文件的原始特征值向量；
将原始特征值向量输入裂纹原始数据文件分类对应的最优特征值筛选模型，筛选出最优特征值向量；
将筛选出的最优特征值向量作为深度神经网络的输入层；将裂纹原始数据文件分类对应的最优深度神经网络隐藏层模型作为深度神经网络的隐藏层；将裂纹的长度、宽度、和深度，或者斜长、斜宽和深度作为深度神经网络的输出。
在一些示例性实施方式中，参考图2B，步骤102可以包括如下步骤：
I)获得被测对象的裂纹原始数据文件。以油气管道裂纹检测为例，可以将管道检测器的检测数据导入裂纹分析软件，通过“裂纹自动检测算法”确定裂纹可疑区域，将裂纹可疑区域包含的裂纹数据矩阵保存为文本文件(.txt)，即构成裂纹原始数据文件；
II)按照被测对象的材质将步骤I)的裂纹原始数据文件进行分类；
III)在步骤II)的基础上，按照裂纹类型对裂纹原始数据文件进一步进行分类。根据裂纹的倾斜角，将裂纹原始数据文件分为轴向裂纹原始数据文件、环向裂纹原始数据文件或倾斜裂纹原始数据文件三种类型中的某一种；
IV)提取步骤III)的裂纹原始数据文件的原始特征值向量。原始特征值向量可以来自以下数据中的一个或多个：漏磁检测中的三轴霍尔元件检测数据，动磁检测数据，检测器运动速度，标称壁厚等。原始特征值向量包含的特征值可以包括小波变换的最大变换系数，小波变换的最优尺度因子等等；
V)将步骤IV)的原始特征值向量代入步骤101中对应材质以及对应裂纹类型下的最优特征值筛选模型，从中筛选出最优特征值向量；
VI)将步骤V)筛选出的最优特征值向量输入深度神经网络，该深度神经网络的隐藏层为步骤101中对应材质以及对应裂纹类型下的最优深度神经网络隐藏层模型，该深度神经网络的输出为裂纹的长(斜长)、宽(斜宽)、深数据。
在一些示例性实施方式中，在上述步骤IV)中，对于倾斜裂纹，并非直接从裂纹原始样本数据文件中提取原始特征值向量，而是首先对倾斜裂纹进行波峰对齐，然后从对齐的数据中提取原始特征值向量。具体过程可参见训练阶段的相关步骤。相应的，在上述步骤VI)中，对于倾斜裂纹，是将波峰对齐的样本数据与裂纹的斜长、斜宽、深度建立映射关系，最后将倾斜裂纹的斜长、斜宽、深等尺寸作为深度神经网络的输出。
本公开的深度神经网络裂纹量化方法，通过裂纹分类，可以使轴向裂纹的轴向特征，环向裂纹的环向特征，倾斜裂纹的倾斜角特征在各自的类别下得到强化；通过对齐倾斜裂纹的波峰，可以将倾斜裂纹的量化问题转化为类似环向裂纹的量化问题；通过对裂纹的时域波形进行连续小波变换，可以准确提取裂纹的时域特征和频域特征，提高裂纹的量化精度；通过特征值筛选机制的使用，可以分类筛选出对模型量化结果影响比较大的特征值参与模型训练；这样不仅可以强化有影响力的特征值对特定类型量化模型的作用效果，还能通过减少低效特征值的数量提升裂纹量化的计算效率。
为了更好地理解本公开提供的深度神经网络裂纹量化方法，下面结合
Figure PCTCN2022141622-appb-000024
(X80钢材，15.30mm壁厚)油气管道裂纹量化这一示例性实施例对本公开的技术方案做进一步说明。
本实施例使用基于漏磁与动磁数据融合的裂纹检测方法确定裂纹可疑区域，并从中获得裂纹原始样本数据文件集和裂纹原始数据文件集。漏磁检测技术相比其他电磁无损检测技术具有原理简单、工程易实现、检测效率高等诸多优点；动磁检测技术对检测任意方向的裂纹缺陷具有很高的灵敏度，与漏磁检测形成优势互补。下面对方法的实施步骤进行详细介绍。
整个实施步骤共分为2个阶段，第1阶段为“最优特征值筛选模型”和“最优深度神经网络隐藏层模型”的建立阶段(模型训练阶段)；第2阶段为深度神经网络裂纹量化方法的使用阶段。
在第1阶段，“最优特征值筛选模型”和“最优深度神经网络隐藏层模型”的建立步骤包括：
按照步骤1)，开展对
Figure PCTCN2022141622-appb-000025
(X80钢材，15.30mm壁厚)人工裂纹管道的牵拉实验，本次获得274个裂纹原始样本数据文件，由此组成裂纹原始样本数据文件集；
按照步骤2)，依据被测对象材质将裂纹原始样本数据文件集进行分类。本实施例中在材质层面，只有X80钢材这一种类型；
按照步骤3)，依据裂纹类型对裂纹原始样本数据文件集进行进一步地分类，得到三个裂纹原始样本数据文件集。本实施例将材质X80的裂纹原始样本数据文件集分为轴向裂纹原始样本数据文件集(90个轴向裂纹原始样本数据文件)、环向裂纹原始样本数据文件集(120个环向裂纹原始样本数据文件)和倾斜裂纹原始样本数据文件集(64个倾斜裂纹原始样本数据文件)；
按照步骤4)，对三个裂纹原始样本数据文件集中的每个数据文件提取原始特征值向量，并分别组成轴向裂纹原始特征值向量集，环向裂纹原始特 征值向量集，倾斜裂纹原始特征值向量集(对于倾斜裂纹，首先对它进行波峰对齐)。此处共提取58个原始特征值组成原始特征值向量，它们分别为：从漏磁轴向分量MFLX波形数据中提取的15个特征值；从漏磁径向分量MFLY波形数据中提取的15个特征值；从漏磁周向分量MFLZ波形数据中提取的15个特征值；从动磁信号DM波形数据中提取的5个特征值，以及检测器运行速度值和标称壁厚值。此外，从Python小波库函数中选取gaus2小波、gaus1小波、gaus5小波分别表征MFLX信号、MFLY信号以及DM信号；通过连续小波变换，分别从峰值最大MFLX通道信号、MFLY通道信号、DM通道信号中提取出小波变换最大的变换系数
Figure PCTCN2022141622-appb-000026
Figure PCTCN2022141622-appb-000027
以及对应的小波最优尺度因子
Figure PCTCN2022141622-appb-000028
作为另外6个特征值。这6个小波特征值可以很精细地表征裂纹的长(斜长)、宽(斜宽)、深信息。
按照步骤5)，对X80钢材目录下的3个裂纹类型中的裂纹原始特征值向量集，分别依据特征值筛选机制建立“最优特征值筛选模型”和“最优深度神经网络隐藏层模型”，具体执行如下操作：
5a)设定所有裂纹量化结果中长(斜长)、宽(斜宽)、深3个指标相同的权重系数，即设定权重向量为ω i＝[ω i1，ω iw，ω id]＝[0.333，0.333，0.334]。，其中i＝1,2,3；设定置信度为90％，
Figure PCTCN2022141622-appb-000029
另外设定样本集包含75％的训练集，包含25％的测试集，即m:n＝3:1。
5b)原始特征值数目p为58，特征值筛选数量的初始值q为20，设定第i类(i＝1,2,3)裂纹的特征值筛选评价函数初始值G i_min为1000；
5c)通过导入Python程序包sklearn.feature_selection()调用“递归特征消除法”，根据本次特征值筛选数量q，采用“递归特征消除法”从58个原始特征值中筛选出q个特征值，组成新的特征值矩阵X i_train_s和X i_test_s；，其中X i_train_s的维数为m×q，X i_test_s的维数为n×q；
5d)构建“深度神经网络隐藏层模型”；通过导入Python中keras.models.Sequential()和keras.layers.Dense()程序包，示例性地，搭建了6层全连接神经网络模型。其中，第一层为输入层，输入层的维数等于筛选出的特征值的个数。中间4层为隐藏层(在另一些示例性实施例中，隐藏层的层数也可以为5层或6层)，节点数分别为30、18、20、20。最后一层为输出层，有3个节点，分别对应裂纹量化的长(斜长)、宽(斜宽)、深结果；
5e)通过DNN模型求得第i类(i＝1,2,3)裂纹的训练集误差矩阵和测试集误差矩阵E i_train和E i_test；
5f)根据公式(9)计算第i类(i＝1,2,3)裂纹的特征值筛选评价函数G i；
5g)如果G i＜G i_min，则用G i替换G i_min，并将当前的“特征值筛选模型”以及“深度神经网络隐藏层模型”作为当前最新的结果，保存到缓存RAMi中，跳转5h)；如果G i≥G i_min，直接跳转5h)；
5h)更新特征值筛选数量q＝q+1，更新之后如果q>58则跳转5i)，如果q≤58跳转5c)；
5i)以缓存RAMi中保存的当前“特征值筛选模型”为“最优特征值筛选模型”；以缓存RAMi中保存的当前“深度神经网络隐藏层模型”为“最优深度神经网络隐藏层模型”；
5j)返回步骤5a)，直到X80钢材目录下的3个裂纹类型中的所有裂纹原始特征值向量集全部被筛选完毕，输出并保存每个材质目录下的每个裂纹类型中“最优特征值筛选模型”和“最优深度神经网络隐藏层模型”。示例性的，在本实施例中，“最优特征值筛选模型”可以保存为.pkl文件，“最优深度神经网络隐藏层模型”可以保存为.h5文件。
针对每个分类，用“最优特征值筛选模型”筛选出的最优特征值向量作为深度神经网络的输入层，并在建立“最优特征值筛选模型”的同时也得到 了每个分类对应的“最优深度神经网络隐藏层模型”，这两个模型是深度神经网络模型的一部分。对轴向裂纹、环向裂纹和倾斜裂纹，后续可以分别使用训练好的对应的深度神经网络模型进行量化，输出3种裂纹的最优DNN量化模型和裂纹的长(斜长)、宽(斜宽)、深结果。
表1是通过上述方法对牵拉场274个裂纹实现的裂纹量化效果。可见，在90％的置信度下，无论是轴向裂纹，环向裂纹，还是倾斜裂纹，所有裂纹的长、宽误差的整体置信区间都在±10mm范围内；深度误差的整体置信区间都在±10％wt范围内，其中，％wt表示裂纹深度相对标称壁厚(15.30mm)的百分比。
Figure PCTCN2022141622-appb-000030
表1
图3A至图3C分别是在轴向裂纹、环向裂纹和倾斜裂纹的最优深度神经网络隐藏层模型的训练过程中有特征值筛选机制和无特征值筛选机制的对比结果示意图，其中，t_loss_new，v_loss_new是采用特征值筛选机制方法(本公开中的方法)的训练集损失函数以及验证集的损失函数；t_loss_old，v_loss_old是传统的无特征值筛选机制方法的训练集损失函数以及验证集的 损失函数。可以看出，无论是轴向裂纹，环向裂纹，还是倾斜裂纹，采用特征值筛选机制训练出的裂纹量化模型，无论在训练集中还是在验证集中，都具有更低的损失函数。这表明本公开相比于传统的神经网络方法具有更高的裂纹量化精度。
在第2阶段，基于小波变换和特征值筛选的深度神经网络裂纹量化方法的使用步骤为：
按照步骤I)，获得被测对象的裂纹原始数据文件。对某段退役的
Figure PCTCN2022141622-appb-000031
(X80钢材，标称壁厚15.30mm)真实天然气管道开展内检测实验，管体上加工有若干不同倾斜角的人工裂纹，事先通过游标卡尺、量角器测量得到这些人工裂纹长(斜长)、宽(斜宽)、深、倾斜角。特别指出的是，该段退役管道上的人工裂纹不参与第一阶段的深度神经网络裂纹量化模型训练，即不参与“最优特征值筛选模型”和“最优深度神经网络隐藏层模型”的建立过程，而只参与第二阶段对量化方法的检验。检测结束后，将管道检测器的检测数据导入裂纹分析软件，通过基于漏磁与动磁数据融合的裂纹检测方法确定裂纹可疑区域，将裂纹可疑区域包含的裂纹数据矩阵保存为文本文件(.txt)，即构成裂纹原始数据文件。如图4A至图4C以及图5所示，通过裂纹分析软件在检测数据的某个区域共检测到12个真实裂纹，图中方框框住的区域为裂纹分析软件自动检测到的裂纹可疑区域，图4A至图4C以及图5中L1-L12表示的是共同的12个裂纹，只不过是用MFLX/MFLY/MFLZ/DM四个通道的数据分别展示而已。针对每个裂纹，建立相应的裂纹原始数据文件，共12个裂纹原始数据文件，每个裂纹原始数据文件包括MFLX/MFLY/MFLZ/DM四个通道的裂纹数据矩阵，还包括检测器运动速度、标称壁厚等数据。
按照步骤II)，根据被测对象的材质将裂纹原始数据文件进行分类；本实施例中在材质层面，只有X80钢材这一种类型，因此，该12个裂纹原始数据文件在材质上属于同一大类；
按照步骤III)，按照裂纹类型对裂纹原始数据文件进一步进行分类。在工程上，将倾斜角为[0°，10°)和[170°，180°)的裂纹定义为轴向裂纹， 将倾斜角为[80°，100°)的裂纹定义为环向裂纹，将倾斜角为[10°，80°)和[100°，170°)的裂纹定义为倾斜裂纹；如图4A至图4C以及图5所示，其中L1-L4为倾斜裂纹，L5-L8为轴向裂纹，L9-L12为环向裂纹，即将12个裂纹原始数据文件分为3类：X80倾斜裂纹原始数据文件、X80轴向裂纹原始数据文件和X80环向裂纹原始数据文件；
按照步骤IV)，提取裂纹原始数据文件的原始特征值向量(对于倾斜裂纹，首先对它进行波峰对齐)。该原始特征值向量所包含的特征值类型与第一阶段的步骤4)的58个特征值类型相同；
按照步骤V)，将原始特征值向量代入第1阶段的对应材质与裂纹类型下的“最优特征值筛选模型”，从中筛选出每个裂纹原始数据文件对应的最优特征值向量；
按照步骤VI)，将步骤V)筛选出的每个最优特征值向量作为深度神经网络的输入层；该深度神经网络的隐藏层为第1阶段对应材质以及裂纹类型下的最优深度神经网络隐藏层模型，该深度神经网络的输出为裂纹的长(斜长)、宽(斜宽)、深数据。
表2是上述
Figure PCTCN2022141622-appb-000032
(X80钢材，标称壁厚15.30mm)真实天然气管道12个裂纹的游标卡尺、量角器的测量值，表3是上述
Figure PCTCN2022141622-appb-000033
(X80钢材，标称壁厚15.30mm)真实天然气管道12个裂纹的量化估计值，表4是表3的量化估计值与表2的测量值之间的误差，可见，无论是轴向裂纹，环向裂纹，还是倾斜裂纹，所有裂纹的长(斜长)、宽(斜宽)误差都在±10mm范围内；深度误差都在±10％wt范围内，其中％wt表示裂纹深度相对标称壁厚(15.30mm)的百分比。由表2至表4看出，本公开提供的基于小波变换和特征值筛选的深度神经网络裂纹量化方法具有较高的裂纹量化精度。
Figure PCTCN2022141622-appb-000034
Figure PCTCN2022141622-appb-000035
表2
Figure PCTCN2022141622-appb-000036
表3
Figure PCTCN2022141622-appb-000037
Figure PCTCN2022141622-appb-000038
表4
本公开实施例还提供了一种深度神经网络裂纹量化装置，包括显示裂纹量化结果的显示器、存储裂纹量化方法和临时结果的存储器；和连接至所述存储器的处理器，所述处理器执行基于存储在所述存储器中的指令，执行如前任一项所述的深度神经网络裂纹量化方法的步骤。
在一个示例中，如图6所示，深度神经网络裂纹量化装置可包括：处理器610、存储器620、总线系统630和显示器640，其中，处理器610、存储器620、显示器640通过该总线系统630相连，存储器620用于存储指令及最优特征值向量、最优特征值筛选模型和最优深度神经网络隐藏层模型等，处理器610用于执行存储器620存储的指令，以通过深度神经网络进行裂纹量化。具体地，处理器610可以针对轴向裂纹、环向裂纹和倾斜裂纹，分别训练“最优特征值筛选模型”和“最优深度神经网络隐藏层模型”；并对轴向裂纹、环向裂纹和倾斜裂纹，分别使用训练好的对应的模型进行量化；最后通过显示器640将量化结果显示出来。
应理解，处理器610可以是中央处理单元(Central Processing Unit，CPU)，处理器610还可以是其他通用处理器、数字信号处理器(DSP)、专用集成电路(ASIC)、现成可编程门阵列(FPGA)或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件等。通用处理器可以是微处理器或者该处理器610也可以是任何常规的处理器等。
存储器620可以包括只读存储器和随机存取存储器，并向处理器610提供指令和数据，包括所述的最优特征值向量、深度神经网络模型等。存储器620的一部分还可以包括非易失性随机存取存储器。例如，存储器620还可以存储设备类型的信息。
总线系统630除包括数据总线之外，还可以包括电源总线、控制总线和状态信号总线等。
显示器640除了显示裂纹的量化结果之外，还可以通过裂纹分析软件显示存储器620中的裂纹检测数据。
在实现过程中，该深度神经网络裂纹量化装置所执行的处理可以通过处理器610中的硬件的集成逻辑电路或者软件形式的指令完成。即本公开实施例的深度神经网络裂纹量化方法步骤可以由硬件处理器执行完成，或者用处理器610中的硬件及软件模块组合执行完成。软件模块可以位于随机存储器，闪存、只读存储器，可编程只读存储器或者电可擦写可编程存储器、寄存器等存储介质中。该存储介质位于存储器620，处理器610读取存储器620中的信息，结合其硬件完成上述方法的步骤。为避免重复，这里不再详细描述。
本公开实施例还提供了一种存储介质，该存储介质存储有可执行指令，该可执行指令被处理器执行时可以实现本公开上述任一实施例提供的深度神经网络裂纹量化方法，该深度神经网络裂纹量化方法可以针对轴向裂纹、环向裂纹和倾斜裂纹，分别训练深度神经网络模型；对所述轴向裂纹、环向裂纹和倾斜裂纹，分别使用训练好的对应的深度神经网络模型进行量化。通过执行可执行指令实现裂纹检测的方法与本公开上述实施例提供的深度神经网络裂纹量化方法基本相同，在此不做赘述。
本领域普通技术人员可以理解，上文中所公开方法中的全部或某些步骤、系统、装置中的功能模块/单元可以被实施为软件、固件、硬件及其适当的组合。在硬件实施方式中，在以上描述中提及的功能模块/单元之间的划分不一定对应于物理组件的划分；例如，一个物理组件可以具有多个功能，或者一个功能或步骤可以由若干物理组件合作执行。某些组件或所有组件可以被实施为由处理器，如数字信号处理器或微处理器执行的软件，或者被实施为硬件，或者被实施为集成电路，如专用集成电路。这样的软件可以分布在计算机可读介质上，计算机可读介质可以包括计算机存储介质(或非暂时性介质)和通信介质(或暂时性介质)。如本领域普通技术人员公知的，术语计算机存储介质包括在用于存储信息(诸如计算机可读指令、数据结构、程序模块或其他数据)的任何方法或技术中实施的易失性和非易失性、可移除和不可 移除介质。计算机存储介质包括但不限于RAM、ROM、EEPROM、闪存或其他存储器技术、CD-ROM、数字多功能盘(DVD)或其他光盘存储、磁盒、磁带、磁盘存储或其他磁存储装置、或者可以用于存储期望的信息并且可以被计算机访问的任何其他的介质。此外，本领域普通技术人员公知的是，通信介质通常包含计算机可读指令、数据结构、程序模块或者诸如载波或其他传输机制之类的调制数据信号中的其他数据，并且可包括任何信息递送介质。
虽然本公开所揭露的实施方式如上，但所述的内容仅为便于理解本公开而采用的实施方式，并非用以限定本公开。任何本公开所属领域内的技术人员，在不脱离本公开所揭露的精神和范围的前提下，可以在实施的形式及细节上进行任何的修改与变化，但本公开的保护范围，仍须以所附的权利要求书所界定的范围为准。

Claims (12)

  1. 一种深度神经网络裂纹量化方法，包括：
    针对轴向裂纹、环向裂纹和倾斜裂纹，分别训练深度神经网络模型；
    对所述轴向裂纹、环向裂纹和倾斜裂纹，分别使用训练好的对应的深度神经网络模型进行量化。
  2. 根据权利要求1所述的深度神经网络裂纹量化方法，其中，对所述轴向裂纹和环向裂纹，所述深度神经网络模型的输出为裂纹的长度、宽度和深度尺寸，对所述倾斜裂纹，所述深度神经网络模型的输出为裂纹的斜长、斜宽和深度尺寸。
  3. 根据权利要求2所述的深度神经网络裂纹量化方法，在针对所述倾斜裂纹训练深度神经网络模型时，或者在对所述倾斜裂纹进行量化时，所述方法还包括：
    对每个倾斜裂纹，选取包含倾斜裂纹波峰的裂纹数据矩阵，所述裂纹数据矩阵的横向维度表示采样点，所述裂纹数据矩阵的纵向维度表示通道数，所述裂纹数据矩阵的元素值表示检测数据值；
    确定所述裂纹数据矩阵中从左数或从右数出现的第一个波峰；
    将所述裂纹数据矩阵中其他行的波峰向左平移或向右平移，并与第一个波峰对齐；
    以对齐后的波峰为对称轴左右选取等宽的区间形成新的裂纹数据矩阵，以此新的裂纹数据矩阵提取所述倾斜裂纹的特征值向量。
  4. 根据权利要求1所述的深度神经网络裂纹量化方法，其中，所述针对轴向裂纹、环向裂纹和倾斜裂纹，分别训练深度神经网络模型，包括：
    获取裂纹原始样本数据文件集，所述裂纹原始样本数据文件集包括多个裂纹原始样本数据文件；
    按照被测对象的材质和裂纹类型对裂纹原始样本数据文件进行分类，所述裂纹类型包括轴向裂纹、环向裂纹和倾斜裂纹，提取每个裂纹原始样本数 据文件的原始特征值向量；
    从每个分类对应的原始特征值向量中筛选出每个分类对应的最优特征值向量，同时生成每个分类的最优特征值筛选模型和最优深度神经网络隐藏层模型，所述最优特征值向量包含的特征值数目小于或等于所述原始特征值向量包含的特征值数目，所述最优特征值筛选模型用于为每个分类筛选对应的最优特征值向量，所述最优深度神经网络隐藏层模型用于作为深度神经网络的隐藏层，对裂纹进行量化。
  5. 根据权利要求4所述的深度神经网络裂纹量化方法，其中，所述原始特征值向量来自以下数据中的一个或多个：漏磁检测信号中的三轴霍尔元件检测数据、动磁检测数据、检测器运动速度、标称壁厚。
  6. 根据权利要求4所述的深度神经网络裂纹量化方法，其中，所述裂纹原始样本数据文件包括一个或多个裂纹数据矩阵，所述原始特征值向量包含的特征值包括：小波变换的最大变换系数以及小波变换的最优尺度因子，所述小波变换的最大变换系数以及小波变换的最优尺度因子均通过对所述裂纹数据矩阵中峰值最大的裂纹时域波形进行连续小波变换得到。
  7. 根据权利要求6所述的深度神经网络裂纹量化方法，其中，提取所述小波变换的最大变换系数与小波变换的最优尺度因子，包括：
    从所述裂纹数据矩阵中选取峰值最大的信号通道；
    依据下式对峰值最大的信号通道进行连续小波变换：
    Figure PCTCN2022141622-appb-100001
    其中，x[k]为峰值最大的信号通道的第k点的观测值，ψ[k]为基本小波或母小波，ψ *[k]表示对ψ[k]求共轭运算；dj为采样步长，dj为实数；Δ*a中的*为乘号，Δ为ψ[k]的半宽度，Δ为实数；a为尺度因子，a∈[a 1，a max1]；b为平移量因子，b∈[b 1，b max2]，max1和max2均为大于1的自然数；ψ a，b[k]为小波基函数；<x[k]，ψ a，b[k]>为小波变换系数；
    依据下式得到变换系数矩阵WT：
    Figure PCTCN2022141622-appb-100002
    从变换系数矩阵WT中选择小波变换的最大变换系数WT max作为提取出的小波变换的最大变换系数，选择与最大变换系数WT max对应的小波变换的最优尺度因子a best作为提取出的小波变换的最优尺度因子。
  8. 根据权利要求4所述的深度神经网络裂纹量化方法，其中，从一个分类对应的原始特征值向量中筛选出该分类对应的最优特征值向量，包括：
    将一个分类对应的原始特征值向量集分为m个训练集和n个测试集，m+n＝NUM，NUM为该分类对应的原始特征值向量集中包括的原始特征值向量数量；
    对特征值筛选数量从初始值q到最大值p，分别执行如下操作：根据当前的所述特征值筛选数量确定所述深度神经网络模型的输入层的节点数，根据所述输入层的节点数、输出层的节点数以及隐藏层层数构建所述深度神经网络模型，将所述m个训练集和n个测试集输入所述深度神经网络模型，得到训练集误差矩阵和测试集误差矩阵，并根据所述训练集误差矩阵和测试集误差矩阵，计算特征值筛选评价函数，其中，p为每个原始特征值向量包含的特征值数量，0<q<p；
    以使对应的特征值筛选评价函数最小的特征值筛选数量作为筛选出的最优输入特征值数量，并将所述最优输入特征值数量对应的特征值筛选模型和深度神经网络隐藏层模型作为该分类对应的最优特征值筛选模型和最优深度神经网络隐藏层模型。
  9. 根据权利要求8所述的深度神经网络裂纹量化方法，其中，所述训练集误差矩阵和测试集误差矩阵分别表示为：
    Figure PCTCN2022141622-appb-100003
    其中，E il、E iw、E id分别表示第i类裂纹的长、宽、深的误差向量，下角标train表示训练集，下角标test表示测试集，i为1至Cnt之间的自然数，Cnt为所述裂纹原始样本数据文件的分类数目；
    所述根据所述训练集误差矩阵和测试集误差矩阵，计算特征值筛选评价函数，包括：
    依据下式分别计算置信度为(1-α)的条件下，其中0＜α＜1，误差区间估计的单体精确度矩阵以及置信区间的中值矩阵：
    Figure PCTCN2022141622-appb-100004
    Figure PCTCN2022141622-appb-100005
    其中，A i_train、A i_test、M i_train和M i_test的维数均为1×3，系数γ根据置信度的值进行设置，
    Figure PCTCN2022141622-appb-100006
    其中φ -1(·)表示标准正态累积分布函数的逆函数，
    Figure PCTCN2022141622-appb-100007
    和
    Figure PCTCN2022141622-appb-100008
    分别表示训练集误差矩阵和测试集误差矩阵的均值；
    Figure PCTCN2022141622-appb-100009
    和
    Figure PCTCN2022141622-appb-100010
    分别表示置信度为(1-α)，m个训练集和n个测试集的t分布的上分位点；
    Figure PCTCN2022141622-appb-100011
    和
    Figure PCTCN2022141622-appb-100012
    分别表示置信度为(1-α)，m个训练集和n个测试集的χ 2分布的上分位点；S m和S n分别表示训练集误差矩阵和测试集误差矩阵的样本标准差；
    Figure PCTCN2022141622-appb-100013
    Figure PCTCN2022141622-appb-100014
    依据下式对计算出的单体精确度矩阵和中值矩阵分别进行合并处理：
    Figure PCTCN2022141622-appb-100015
    Figure PCTCN2022141622-appb-100016
    其中，第i类裂纹长、宽、深度的权重向量为ω i＝[ω il，ω iw，ω id]，ω il、ω iw和ω id均介于[0,1]之间，且ω il+ω iw+ω id＝1；
    Figure PCTCN2022141622-appb-100017
    表示对A i_train进行转置；
    依据下式分别计算训练集的特征值筛选评价函数G i_train和测试集的特征值筛选评价函数G i_test：
    Figure PCTCN2022141622-appb-100018
    其中，λ1为第一权重系数；λ2为第二权重系数，λ1和λ2均介于[0,1]之间；
    依据下式计算第i类裂纹信号的特征值筛选评价函数G i：
    Figure PCTCN2022141622-appb-100019
  10. 根据权利要求4所述的深度神经网络裂纹量化方法，其中，对所述轴向裂纹、环向裂纹和倾斜裂纹，分别使用训练好的对应的深度神经网络模型进行量化，包括：
    获取被测对象的一个或多个裂纹原始数据文件；
    按照被测对象的材质和裂纹类型对所述裂纹原始数据文件进行分类；提取所述裂纹原始数据文件的原始特征值向量；
    将所述原始特征值向量输入所述裂纹原始数据文件分类对应的最优特征值筛选模型，筛选出最优特征值向量；
    将筛选出的最优特征值向量作为深度神经网络的输入层；将所述裂纹原始数据文件分类对应的最优深度神经网络隐藏层模型作为深度神经网络的隐藏层；将裂纹的长度、宽度、和深度，或者斜长、斜宽和深度作为深度神经网络的输出。
  11. 一种深度神经网络裂纹量化装置，包括显示量化结果的显示器、存 储指令的存储器和连接至所述存储器的处理器，所述处理器可以执行基于存储在所述存储器中的指令，执行如权利要求1至10中任一项所述的深度神经网络裂纹量化方法的步骤。
  12. 一种存储介质，其上存储有裂纹量化的程序，该程序被处理器执行时实现如权利要求1至10中任一项所述的深度神经网络裂纹量化方法。
PCT/CN2022/141622 2022-11-21 2022-12-23 一种深度神经网络裂纹量化方法及装置、存储介质 WO2024108717A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202211453244.6A CN115758084B (zh) 2022-11-21 2022-11-21 一种深度神经网络裂纹量化方法及装置、存储介质
CN202211453244.6 2022-11-21

Publications (1)

Publication Number Publication Date
WO2024108717A1 true WO2024108717A1 (zh) 2024-05-30

Family

ID=85333858

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/141622 WO2024108717A1 (zh) 2022-11-21 2022-12-23 一种深度神经网络裂纹量化方法及装置、存储介质

Country Status (2)

Country Link
CN (1) CN115758084B (zh)
WO (1) WO2024108717A1 (zh)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN118385777A (zh) * 2024-06-24 2024-07-26 长沙韶光芯材科技有限公司 一种用于半导体基底的基板形成方法
CN118698995A (zh) * 2024-08-28 2024-09-27 克拉玛依市三达有限责任公司 抽油杆清洗修复方法

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116754632B (zh) * 2023-08-16 2023-11-21 清华大学 基于正交孪生的油气管道裂纹量化方法及装置、存储介质
CN117007673B (zh) * 2023-08-16 2024-01-23 清华大学 一种油气管道裂纹信号的正交孪生方法及装置、存储介质

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107909564A (zh) * 2017-10-23 2018-04-13 昆明理工大学 一种基于深度学习的全卷积网络图像裂纹检测方法
CN111311542A (zh) * 2020-01-15 2020-06-19 歌尔股份有限公司 一种产品质量检测方法及装置
CN112857197A (zh) * 2021-04-09 2021-05-28 中国石油大学(华东) 一种旋转电磁场非铁磁性材料裂纹尺寸测量方法
US20220092856A1 (en) * 2020-09-22 2022-03-24 Bentley Systems, Incorporated Crack detection, assessment and visualization using deep learning with 3d mesh model

Family Cites Families (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9377548B2 (en) * 2011-11-09 2016-06-28 Chevron U.S.A. Inc. Wavelet-transform based system and method for analyzing characteristics of a geological formation
CN105334269A (zh) * 2015-10-19 2016-02-17 江苏大学 一种基于神经网络和导波特征数据库的管道缺陷类别判别方法
CN109376773A (zh) * 2018-09-30 2019-02-22 福州大学 基于深度学习的裂纹检测方法
CN112116587A (zh) * 2020-09-29 2020-12-22 西安热工研究院有限公司 基于孪生支持向量机的水轮机转轮叶片裂纹识别方法、系统、设备及存储介质
CN112633328A (zh) * 2020-12-04 2021-04-09 北京科技大学 一种基于深度学习的致密油藏改造效果评价方法
CN112686887A (zh) * 2021-01-27 2021-04-20 上海电气集团股份有限公司 混凝土表面裂纹的检测方法、系统、设备和介质
CN113705567B (zh) * 2021-08-25 2024-02-02 浙江国际海运职业技术学院 船舶裂纹检测方法、系统、设备及计算机可读存储介质
CN114332075A (zh) * 2022-01-21 2022-04-12 广州大学 基于轻量化深度学习模型的结构缺陷快速识别与分类方法
CN114791460B (zh) * 2022-04-08 2023-12-29 清华大学 一种基于数据融合的裂纹检测方法及检测装置、存储介质
CN114897909B (zh) * 2022-07-15 2022-09-20 四川大学 基于无监督学习的曲轴表面裂纹监测方法及系统

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107909564A (zh) * 2017-10-23 2018-04-13 昆明理工大学 一种基于深度学习的全卷积网络图像裂纹检测方法
CN111311542A (zh) * 2020-01-15 2020-06-19 歌尔股份有限公司 一种产品质量检测方法及装置
US20220092856A1 (en) * 2020-09-22 2022-03-24 Bentley Systems, Incorporated Crack detection, assessment and visualization using deep learning with 3d mesh model
CN112857197A (zh) * 2021-04-09 2021-05-28 中国石油大学(华东) 一种旋转电磁场非铁磁性材料裂纹尺寸测量方法

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
CHUAN YANG: "Eddy Current Treatment Method and Mechanics Research on Microcrack Healing in the Metal Tubes", DISSERTATION FOR THE DOCTORAL DEGREE IN ENGINEERING, SCHOOL OF MATERIALS SCIENCE AND ENGINEERIN, 1 September 2019 (2019-09-01), School of Materials Science and Engineerin, XP093173006 *

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN118385777A (zh) * 2024-06-24 2024-07-26 长沙韶光芯材科技有限公司 一种用于半导体基底的基板形成方法
CN118698995A (zh) * 2024-08-28 2024-09-27 克拉玛依市三达有限责任公司 抽油杆清洗修复方法

Also Published As

Publication number Publication date
CN115758084B (zh) 2023-11-14
CN115758084A (zh) 2023-03-07

Similar Documents

Publication Publication Date Title
WO2024108717A1 (zh) 一种深度神经网络裂纹量化方法及装置、存储介质
CN107909564B (zh) 一种基于深度学习的全卷积网络图像裂纹检测方法
Dutta et al. Simulation and analysis of 3-D magnetic flux leakage
WO2016107009A1 (zh) 检查货物的方法和系统
Qiao et al. A crack identification method for concrete structures using improved U‐Net convolutional neural networks
CN109298464B (zh) 致密砂岩储层成岩相测井识别方法及装置
Liu et al. Pulsed eddy current data analysis for the characterization of the second-layer discontinuities
Lin et al. A spectrum-domain instance segmentation model for casting defects
CN108344795B (zh) 油气管道缺陷识别方法、装置及电子设备
CN111581197B (zh) 对数据集中的数据表进行抽样和校验的方法及装置
CN103196356A (zh) 基于支持向量机的油管缺陷定量识别方法
Pan et al. A novel method for defects marking and classifying in MFL inspection of pipeline
CN106404893B (zh) 一种轴向漏磁缺陷信号自动检测方法
Liu et al. Steel coiled tubing defect evaluation using magnetic flux leakage signals
KR102272879B1 (ko) 공간 밀도 분석에 기초한 반도체 웨이퍼의 유도식 검사
van Sebille et al. Nanocrystal size distribution analysis from transmission electron microscopy images
Pavo et al. Adaptive inversion database for electromagnetic nondestructive evaluation
Wang et al. Multi-angle excited MOI and image processing strategies specified for detection of orthogonal weld defects
Jia et al. An effective non-HT circle detection for centers and radii
Mu et al. The application of RBF neural network in the wood defect detection
Zach et al. Automated texture analysis and determination of fibre orientation of heart tissue: a morphometric study
JP6899644B2 (ja) 損傷検査システム、損傷検査方法、およびプログラム
Yang et al. The method of the pipeline magnetic flux leakage detection image formation based on the artificial intelligence
CN116754632B (zh) 基于正交孪生的油气管道裂纹量化方法及装置、存储介质
CN106600604B (zh) 一种磁光成像缺陷检测方法的阀值优化方法

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 22966396

Country of ref document: EP

Kind code of ref document: A1