A damage pattern identification method for prestressed concrete beams

Abstract

The present invention discloses a damage pattern recognition method for a prestressed concrete beam, comprising step 1: collecting acoustic emission data of the whole process of a bending failure test of a prestressed concrete beam, and establishing a verification set, a training set, and a test set database required for neural network calculation; step 2: updating the WOA algorithm by introducing a dynamic adaptive weight, generating an IWOA algorithm, and building an IWOA-LSTM neural network model through a fitness function and a peer relationship; step 3: using the verification set data to complete the structural finalization operation of the IWOA-LSTM, and realizing the hyperparameter optimization of the LSTM; step 4: using the training set data to complete the network finalization operation of the optimized LSTM and finally verifying the effect through the test set data. The method provided by the present invention can avoid the local optimization trap caused by the empirical setting of hyperparameters, has a fast training convergence speed, and has a high pattern recognition accuracy.

Description

Technical Field

[0001] The present invention belongs to the intersection of neural network optimization technology and structural health monitoring technology, and is specifically a damage pattern recognition method for prestressed concrete beams. Background Art

[0002] With the extension of the service time of large structures and the increasing sensitivity of the engineering community to structural safety issues, structural health monitoring is gradually replacing structural health testing and becoming a new hot area in civil engineering. Compared with structural health testing, structural health monitoring puts forward higher requirements for non-destructive testing technology, among which the three major technical requirements of dynamic, real-time and online are particularly significant. The technical requirements of structural health monitoring are perfectly matched with the technical advantages of acoustic emission detection technology (AE), which is the technical root of the rapid rise of AE in the past decade. At present, AE technology has been widely used in non-destructive testing of metal, rock and concrete materials, but due to the multi-component and anisotropic characteristics of concrete materials, the breadth and depth of AE technology in the field of concrete materials are far less than those in the fields of metal and rock materials. As a universal signal analysis tool, neural networks naturally fall into the field of vision of AE researchers. The combination of AE+neural network technology for different structural damage pattern recognition targets continues to emerge, and has become an important topic in the intersection of neural network optimization technology and structural health monitoring technology.

[0003] In view of the above technical background, the present invention goes deep into the higher level of structural health monitoring (structural damage quantitative level), and indirectly realizes the quantitative judgment of structural damage by damage pattern recognition based on the time series of component damage. The biggest technical barrier for AE in quantitative judgment of concrete structure damage is the existence of Kaiser effect, which directly causes historical stress interference, thereby limiting the accuracy of AE's quantitative judgment of structural damage. Long short-term memory neural network LSTM belongs to deep learning neural network, which retains the temporal history relationship between signals in principle and is an excellent tool option for solving historical stress interference. However, the hyperparameters used in traditional LSTM training are often set according to experience, which leads to the network stereotyped training of LSTM easily falling into the state of underfitting or overfitting when facing complex networks. The present invention proposes an IWOA-LSTM neural network architecture that assists LSTM in hyperparameter optimization by improving the whale optimization algorithm IWOA, which not only solves the training problem of the above stereotyped network, but also enhances the global optimization characteristics of the classic WOA algorithm when searching for LSTM hyperparameters by introducing dynamic adaptive weights, and has a higher recognition accuracy for damage pattern recognition of prestressed concrete. Summary of the invention

[0004] The purpose of the present invention is to provide a damage pattern recognition method for prestressed concrete beams, so as to simplify the recognition process of damage of prestressed concrete beams and improve the recognition accuracy.

[0005] In order to solve the above technical problems, the specific technical solutions adopted by the present invention are as follows.

[0006] A damage pattern identification method for a prestressed concrete beam is characterized by comprising the following steps.

[0007] Step 1: Collect the acoustic emission data of the whole process of the prestressed concrete beam bending failure test. Under the principle of ensuring the rationality of the spectrum and giving priority to large-value signals, use random selection to screen the data and establish the validation set, training set and test set databases respectively;

[0008] Ensuring the rationality of the spectrum means: eliminating the acoustic emission signals with a main frequency below 20kHz or above 200kHz. The selection of the upper and lower limits of the main frequency is based on the frequency band characteristics of the acoustic emission signal of the prestressed concrete component and is appropriately extended;

[0009] Prioritizing large-amplitude signals means that the signals that have been screened for spectral rationality are rescreened, and the acoustic emission signals that exceed 80% of the maximum amplitude of the typical damage stage to which the signal belongs are directly included in the database;

[0010] The specific random selection method is as follows: re-screen the signals that have been screened by the spectrum rationality and large-value signal priority principle, generate random data labels of [0, 1], sort them in descending order according to the label values, and take the top 50% of the signals into the database;

[0011] The training set and validation set data come from the No. 1 test beam. After screening and storage according to the above method, 30% of the data are randomly selected to establish the validation set database, and 70% of the data are randomly selected to establish the training set database;

[0012] The test set data comes from the No. 2 test beam with the same working condition as the No. 1 test beam. When screening data, the spectrum rationality and the principle of large-value signal priority are no longer considered. Random data labels of [0, 1] are directly generated, and the labels are arranged in descending order. The first 50% of the signals are taken into the test set database.

[0013] Step 2: By introducing dynamic adaptive weights, the whale position update formula under different behavior modes in the classic whale optimization algorithm WOA is changed to obtain the whale position update formula of the improved whale optimization algorithm IWOA; through the fitness function and the identity relationship between the whale position update formula and the LSTM hyperparameter iteration formula, a reverse connection is built from the long short-term memory neural network LSTM to IWOA to generate an IWOA-LSTM neural network model;

[0014] Step 3: Use the validation set database to complete the finalization operation of the IWOA-LSTM neural network and obtain the optimal hyperparameter values ​​of the LSTM neural network;

[0015] Step 4: Use the training set database to complete the finalization operation of the LSTM neural network under the optimal hyperparameter value, and use the test set database to complete the recognition accuracy test.

[0016] The step 2 includes the following key processes.

[0017] Process 1, the dynamic adaptive weight expression is as follows

[0018]

[0019] t is the current iteration number, ω t is the dynamic adaptive weight in the current iteration step, T is the maximum number of iteration steps set in advance, ω min is the lower limit of the dynamic adaptive weight (take ω min =0.1),ω max is the upper limit of the dynamic adaptive weight, (take ω max =0.9);

[0020] Process 2: Classic WOA simulates three behavior modes of whales in the predation process, namely, search and prey behavior, prey encirclement behavior, and foam net attack behavior. After the dynamic adaptive weight is introduced, the whale position update formula in the three behavior modes of classic WOA will be replaced by the IWOA whale position update formula. For other algorithm details, please refer to classic WOA.

[0021] In the IWOA-LSTM neural network, the whale position update formula of IWOA is set to be identical to the hyperparameter iteration formula of LSTM. This is one of the two key points in the linkage between IWOA and LSTM generation algorithms. The whale position update formula of IWOA is also the hyperparameter iteration formula of LSTM.

[0022] Considering the dynamic adaptive weight correction, based on the WOA search and predation behavior pattern, the IWOA whale position update formula, that is, the LSTM hyperparameter iteration formula, is as follows:

[0023]

[0024] Take the value of the kth hyperparameter of LSTM in the next iteration (such as the number of core hidden layer nodes, initial learning rate, etc.), Take the value of the kth hyperparameter of LSTM in the current iteration step, is the random value of the kth hyperparameter of LSTM in the current iteration step. The operator |·| indicates the absolute value. A1 and C1 are coefficients, and the value selection method of both is the same as WOA.

[0025] Considering the WOA-based encirclement behavior pattern after dynamic adaptive weight correction, the IWOA whale position update formula, i.e. the LSTM hyperparameter iteration formula, is as follows:

[0026]

[0027] is the optimal value of the kth hyperparameter of LSTM in the current iteration step;

[0028] Considering the WOA-based foaming network attack behavior pattern after dynamic adaptive weight correction, the IWOA whale position update formula, i.e. the LSTM hyperparameter iteration formula, is as follows:

[0029]

[0030] b is the logarithmic spiral constant, l is a random number in the interval [-1,1], and p is a random number in the interval [0,1];

[0031] Process 3, in the IWOA-LSTM neural network, the fitness function is the second of the two key points of the linkage between the IWOA and LSTM generation algorithms. The fitness function expression is as follows:

[0032]

[0033] n is the number of samples in the validation set database, is the damage pattern represented by the sample, which is always equal to 1; Y is the damage pattern obtained by converting the four-dimensional vector output by the IWOA-LSTM neural network. When the sample damage pattern is incorrectly identified, Y=0; when the sample damage pattern is correctly identified, Y=1.

[0034] The present invention has beneficial effects.

[0035] The present invention adopts LSTM as the neural network architecture, which retains the temporal history relationship between signals in principle and solves the interference of historical stress paths on pattern recognition for the purpose of quantitative judgment of structural damage in actual engineering. The present invention proposes an IWOA-LSTM neural network architecture that uses an improved whale optimization algorithm IWOA to assist LSTM in hyperparameter optimization. This not only solves the training problem of the above-mentioned stereotyped network, but also enhances the global optimization characteristics of the classic WOA algorithm when searching for LSTM hyperparameters by introducing dynamic adaptive weights, with stable convergence and high recognition accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is the overall architecture diagram of the present invention;

[0037] Figure 2Iterative convergence of the LSTM neural network after hyperparameter optimization using IWOA in the present invention;

[0038] Figure 3 The figure is a comparison of the recognition accuracy of four damage patterns in the test set before and after the improvement of the present invention. DETAILED DESCRIPTION

[0039] The technical solution of the present invention is described in detail below, but the protection scope of the present invention is not limited to the embodiments.

[0040] Embodiment: The overall structure of the present invention is shown in Figure 1 .

[0041] The cross-sectional dimensions of the test beam used in the present invention are 1200 mm×200 mm×250 mm.

[0042] Taking into account the small number of acoustic emission signals in the early stage of loading, the test beam used less reinforcement within the appropriate reinforcement range. The longitudinal tensile reinforcement used two secondary steel bars with a diameter of 16 mm, the prestressed reinforcement used 1×3 steel strands with a diameter of 8.6 mm, fixed with anchors, the frame reinforcement used two primary steel bars with a diameter of 10 mm, and the stirrups used primary steel bars with a diameter of 8 mm, a spacing of 100 mm in the dense area, and a spacing of 1500 mm in the non-densified area.

[0043] The concrete adopts C50 commercial concrete, and its mix ratio is 1:1.06:1.94:0.3 of cement:sand:gravel:water. The three-point bending loading method is adopted, and the loading is carried out by YCW400 hydraulic jack. The contact surface between the jack and the beam, and the contact surface between the support and the beam are all made of asbestos to isolate the friction noise.

[0044] The loading steps adopt graded loading, and finally merge into four damage modes according to the time sequence and crack morphology, namely slight damage, surface texture development, through crack formation, and imminent instability.

[0045] Table 1 Overall distribution of samples in the sample library

[0046]

[0047] The test uses SAEU2S multi-channel acoustic emission detector to collect acoustic emission signals. The detector consists of four main parts: sensor, amplifier, acquisition software, and host.

[0048] The dynamic adaptive weight expression is as follows:

[0049]

[0050] t is the current iteration number, ω t is the dynamic adaptive weight in the current iteration step, T is the maximum number of iteration steps set in advance, ωmin is the lower limit of the dynamic adaptive weight (take ω min =0.1),ω max is the upper limit of the dynamic adaptive weight, (take ω max =0.9).

[0051] After the dynamic adaptive weight correction, based on the search and predation behavior mode of WOA, the whale position update formula of IWOA, that is, the hyperparameter iteration formula of LSTM, is as follows:

[0052]

[0053] Take the value of the kth hyperparameter of LSTM in the next iteration (such as the number of core hidden layer nodes, initial learning rate, etc.), Take the value of the kth hyperparameter of LSTM in the current iteration step, is the random value of the kth hyperparameter of LSTM in the current iteration step. The operator |·| represents the absolute value. A1 and C1 are coefficients, and the method of taking values ​​for both is the same as that of the classic WOA.

[0054] After the dynamic adaptive weight correction, based on the WOA's encirclement behavior pattern, the IWOA whale position update formula, that is, the LSTM hyperparameter iteration formula, is as follows:

[0055]

[0056] is the optimal value of the kth hyperparameter of LSTM in the current iteration.

[0057] After the dynamic adaptive weight correction, based on the WOA's foaming network attack behavior pattern, the IWOA whale position update formula, that is, the LSTM hyperparameter iteration formula, is as follows:

[0058]

[0059] b is the logarithmic spiral constant, l is a random number in the interval [-1,1], and p is a random number in the interval [0,1].

[0060] The fitness function expression is as follows:

[0061]

[0062] n is the number of samples in the validation set database, is the damage mode target represented by the sample, the target value is always correct, Y is the damage pattern recognition obtained by digital conversion of the output vector of the IWOA-LSTM neural network. When the sample damage pattern recognition is wrong, Y=0 is taken, and when the sample damage pattern recognition is correct, Y=1 is taken.

[0063] Specifically, the digital conversion process, if the IWOA-LSTM neural network output vector is:

[0064] [α1 α2 α3 α4]

[0065] There are four possible positions for the maximum component of the test set output vector:

[0066] [α max α2 α3 α4] is judged as slight damage. If it is judged to meet the expected target, Y=1, otherwise Y=0.

[0067] [α1 α max α3 α4] is judged as surface pattern development. When it is judged to meet the expected goal, Y=1, otherwise Y=0.

[0068] [α1 α2 α max α4] When it is judged that a through seam is formed and it meets the expected target, Y=1, otherwise Y=0.

[0069] [α1 α2 α3 α max ] When it is judged to be on the verge of instability and meets the expected target, Y=1, otherwise Y=0.

[0070] The input signal includes eight characteristic parameters of AE: amplitude, ring count, duration, energy, rise count, rise time, RMS, and ASL.

[0071] The initialization population of IWOA in IWOA-LSTM is selected as 10, and the maximum number of iterations is set to 25.

[0072] There are two hyperparameters for LSTM in IWOA-LSTM, one is the number of neurons in the core hidden layer, and the other is the initial learning rate. The search range of the former is [6-20], and the optimal number of neurons in the hidden layer is finally determined to be 14 after the optimization operation. The search range of the latter is [0.00001-0.1], and the optimal initial learning rate is finally determined to be 0.001 after the optimization operation.

[0073] After the hyperparameters of IWOA-LSTM are optimized according to the present invention, the LSTM neural network does not have underfitting or overfitting in the training process of different samples. Figure 2 is the network convergence situation when training with all training set data as samples, Figure 2It can be seen that the training of the fixed network can basically ensure that it reaches a stable state after 1000 cycles of iteration. Compared with the conventional LSTM network, the fixed network after hyperparameter optimization according to IWOA-LSTM has significantly improved the damage pattern recognition accuracy of prestressed concrete, and the four damage modes have been improved by 8.4%, 9.1%, 8.6%, and 8.3% respectively.

[0074] As described above, although the present invention has been shown and described with reference to specific preferred embodiments, it should not be construed as limiting the present invention itself. Various changes in form and details may be made without departing from the spirit and scope of the present invention as defined in the appended claims.

Claims

1. A damage pattern recognition method for prestressed concrete beams, characterized in that The following steps are included Step 1: Collect the acoustic emission data of the whole process of the prestressed concrete beam bending failure test. Under the principle of ensuring the rationality of the spectrum and giving priority to large-value signals, use random selection to screen the data and establish the validation set, training set and test set databases respectively; Ensuring the rationality of the spectrum means: eliminating the acoustic emission signals with a main frequency below 20kHz or above 200kHz. The selection of the upper and lower limits of the main frequency is based on the frequency band characteristics of the acoustic emission signal of the prestressed concrete component and is appropriately extended; Prioritizing large-amplitude signals means that the signals that have been screened for spectral rationality are rescreened, and the acoustic emission signals that exceed 80% of the maximum amplitude of the typical damage stage to which the signal belongs are directly included in the database; The specific random selection method is as follows: re-screen the signals that have been screened by the spectrum rationality and large-value signal priority principle, generate random data labels of [0, 1], sort them in descending order according to the label values, and take the top 50% of the signals into the database; The training set and validation set data come from the No. 1 test beam. After screening and storage according to the above method, 30% of the data are randomly selected to establish the validation set database, and 70% of the data are randomly selected to establish the training set database; The test set data comes from the No. 2 test beam with the same working condition as the No. 1 test beam. When screening data, the spectrum rationality and the principle of large-value signal priority are no longer considered. Random data labels of [0, 1] are directly generated, and the labels are arranged in descending order. The first 50% of the signals are taken into the test set database. Step 2: By introducing dynamic adaptive weights, the whale position update formula under different behavior modes in the classic whale optimization algorithm WOA is changed to obtain the whale position update formula of the improved whale optimization algorithm IWOA; through the fitness function and the identity relationship between the whale position update formula and the LSTM hyperparameter iteration formula, a reverse connection is built from the long short-term memory neural network LSTM to IWOA to generate an IWOA-LSTM neural network model; Step 3: Use the validation set database to complete the finalization operation of the IWOA-LSTM neural network and obtain the optimal hyperparameter values ​​of the LSTM neural network; Step 4: Use the training set database to complete the finalization operation of the LSTM neural network under the optimal hyperparameter value, and use the test set database to complete the recognition accuracy test; The step 2 includes the following key processes: Process 1, the dynamic adaptive weight expression is as follows t is the current iteration number, ω t is the dynamic adaptive weight in the current iteration step, T is the maximum number of iteration steps set in advance, ω min is the lower limit of the dynamic adaptive weight, taking ω min =0.1,ω max is the upper limit of the dynamic adaptive weight, taking ω max =0.9; Process 2: Classic WOA simulates three behavior modes of whales in the predation process, namely, search and prey behavior, encirclement and prey attack behavior, and foam net attack behavior. After the dynamic adaptive weight is introduced, the whale position update formula in the three behavior modes of classic WOA will be replaced by the IWOA whale position update formula. In the IWOA-LSTM neural network, the whale position update formula of IWOA is set to be identical to the hyperparameter iteration formula of LSTM. This is one of the two key points in the linkage between IWOA and LSTM generation algorithms. The whale position update formula of IWOA is also the hyperparameter iteration formula of LSTM. Considering the dynamic adaptive weight correction, based on the WOA search and predation behavior pattern, the IWOA whale position update formula, that is, the LSTM hyperparameter iteration formula, is as follows: Take the value of the kth hyperparameter of LSTM in the next iteration step, Take the value of the kth hyperparameter of LSTM in the current iteration step, is the random value of the kth hyperparameter of LSTM in the current iteration step. The operator |·| represents the absolute value. A1 and C1 are coefficients. Considering the WOA-based encirclement behavior pattern after dynamic adaptive weight correction, the IWOA whale position update formula, i.e. the LSTM hyperparameter iteration formula, is as follows: is the optimal value of the kth hyperparameter of LSTM in the current iteration step; Considering the WOA-based foaming network attack behavior pattern after dynamic adaptive weight correction, the IWOA whale position update formula, i.e. the LSTM hyperparameter iteration formula, is as follows: b is the logarithmic spiral constant, l is a random number in the interval [-1,1], and p is a random number in the interval [0,1]; Process 3, in the IWOA-LSTM neural network, the fitness function is the second of the two key points of the linkage between the IWOA and LSTM generation algorithms. The fitness function expression is as follows: n is the number of samples in the validation set database, is the damage pattern represented by the sample, which is always equal to 1; Y is the damage pattern obtained by converting the four-dimensional vector output by the IWOA-LSTM neural network. When the sample damage pattern is incorrectly identified, Y=0; when the sample damage pattern is correctly identified, Y=1.

Images