CN105184364B - A kind of pulling force compensation method for non-destructive tests - Google Patents

Abstract

The invention discloses a tension compensation method for damage identification, which comprises the following steps: (1) taking frequency sweep frequency and different stress states as input layer, and admittance information as output layer to form RBF neural network structure; ( 2) Use sample data including frequency, tension and admittance to train the RBF neural network; (3) take the frequency and tension of the test data after the training as input, output the corresponding admittance data through simulation, and compare the measured Comparing the admittance data obtained from the simulation with the admittance data obtained by simulation, the compensation effect is measured according to the RMSD damage index; the influence of external tension on the admittance data is compensated; the method passes a non-destructive tensile steel beam test and a The beam is verified; through the method provided by the invention, the influence of the tensile force on the structure on the damage identification of EMI technology is eliminated, and it has good practicability.

Description

A Tension Compensation Method for Damage Identification

technical field

The invention belongs to the technical field of structural health monitoring, and more particularly relates to a tension compensation method for damage identification.

Background technique

Civil engineering structures are damaged due to continuous use, material degradation, environmental impact, earthquakes, etc., and the potential micro damages of most structures are difficult to identify with the naked eye. If the small damage in the early stage of the structure is not found in time, it will cause the cumulative damage of the structure, which will lead to the sudden failure of the structure. It is a difficult problem in the field of structural engineering to identify the weak and potential damage in the early stage of the structure in time and accurately; the existing structure is healthy Monitoring technologies include visual inspection method, low-frequency vibration technology, static structural response technology, and local non-destructive testing technology; the defects of these technologies mainly lie in: the engineering structure system is huge, and it is difficult to reach every part of the inspection, which is time-consuming and labor-intensive; due to its low-frequency characteristics It is ineffective for local damage and is susceptible to environmental noise; it is not operable to measure the static characteristics (such as displacement, velocity, acceleration, etc.) of huge structures.

PZT-based piezoelectric impedance (EMI) technology is mainly based on local high-frequency excitation, using PZT as a sensor and driver at the same time, to obtain information on structural performance changes for local excitation of the structure, so as to realize the identification of small damage to the structure; the basic principle is to use The high-strength adhesive pastes PZT on the surface of the structure or implants it inside the structure. Using the positive and negative piezoelectric effect of PZT, the voltage is applied to the structure through the piezoelectric impedance meter to locally excite the structure, and the structure performance (mass, stiffness, damping, etc.) is obtained. Monitoring signal, which is used as a benchmark for structural health measurement, will identify whether the structure is damaged by observing the change of the signal in the future. Due to its high-frequency characteristics, it has the advantages of being sensitive to small structural damage and avoiding the influence of environmental noise; however, this technology is still limited to the test piece and does not take into account the influence of external tension on the workpiece.

In actual working conditions, the engineering structure is always under stress; and during the entire service process of the engineering structure, the load it bears is constantly changing; when the structure is damaged, the local stress will change greatly; EMI technology can measure The admittance signal and realize structural damage identification by judging the offset of the admittance curve; the test shows that the piezoelectric impedance admittance signal of the structure under the state of tension is different from that of the state without tension, and the tension of the structure will cause the admittance curve Therefore, the establishment of an effective tension compensation method needs to be solved to eliminate the interference of tension changes on the identification of engineering structure damage.

Contents of the invention

In view of the above defects or improvement needs of the prior art, the present invention provides a tension compensation method for damage identification, the purpose of which is to train the RBF neural network by using sample data including frequency, tension and admittance to realize the The compensation of the tensile force on the structure eliminates interference for the accurate damage identification of the structure.

In order to achieve the above object, according to one aspect of the present invention, a kind of tension compensation method for damage identification is provided, the tension compensation method is based on RBF (radial basis function of multivariable interpolation, Radical Basis Function) neural network specifically includes as follows step:

(1) The sweep frequency and pulling force are used as the input layer, the admittance value is used as the output layer, and the RBF neural network is formed together with the hidden layer;

(2) adopt the sample data that comprises sweeping frequency, pulling force and admittance value to train described RBF neural network, until the difference between output admittance value and the admittance value of corresponding sample data is within sample data admittance value Within ±5% of , end RBF neural network training;

(3) With the sweep frequency and pulling force of test data as input, adopt the RBF neural network that step (2) training obtains, obtain the admittance value corresponding with described sweep frequency and pulling force; Described admittance value has included external Pull force information, which compensates the influence of pull force on admittance data;

(4) Obtain the compensated RMSD damage index according to the admittance value obtained in step (3);

The above tension compensation method for damage identification first uses the sample data including sweep frequency, tension and admittance value to train the RBF neural network, and after the training is completed, the frequency sweep frequency and tension of the test data are used as the input of the RBF neural network. Obtain the admittance value of the corresponding output, the admittance value contains the pulling force information, and compensates the influence of the pulling force on the admittance data.

Preferably, in the RBF neural network training described in step (2), determine the quantity of hidden layer neurons according to the following method:

a. The input vector corresponding to the mean square error (MSE) generated by the RBF neural network is used as a weight vector to generate a new hidden layer neuron;

b. Taking the admittance value obtained from the test as an input, and taking the admittance value obtained from the simulation as an output, and the new hidden layer neurons generated in step a to form a new RBF neural network;

c, obtaining the mean square error of the new RBF neural network described in step b;

d. Repeat steps a to c until the mean square error of the RBF neural network reaches the preset mean square error, and the number of neurons in the hidden layer at this time is used as the number of neurons in the hidden layer of the RBF neural network; The square error is between 5% and 10%;

In RBF neural network training, the existing technology is to make the number of hidden layer neurons equal to the elements of the input vector; but when there are many input vectors, too many hidden layer units will increase the computational complexity; using this The above-mentioned method provided by the invention gradually approaches the optimal value of the number of neurons in the hidden layer, and can obtain the optimal compensation effect within the minimum amount of calculation.

Preferably, the step of training the RBF neural network described in step (2) includes the following sub-steps:

(2.1) The admittance value of the given sweep frequency, pulling force and expected output;

(2.2) the above-mentioned sweep frequency and pulling force are used as input vectors, and the RBF neural network obtained through step (1) is processed to obtain the admittance value;

(2.3) making a difference between the admittance value obtained in step (2.2) and the admittance value of the desired output to obtain a difference; judge whether the difference is within ±5% of the desired output admittance value; if so, Then end the training; if not, add neurons and return to step (2.2).

Preferably, after the above-mentioned steps (2.3) finish the training to the RBF neural network, adopt the following steps to determine the RBF neural network training effect, specifically:

(2.4) Using the data different from the training sample as the input vector, simulate with the trained RBF neural network, and obtain the admittance value of the simulation output;

(2.5) Obtain the RMSD value according to the admittance value output by the simulation in step (2.4). If the RMSD value obtained under each tension condition fluctuates within 5%, go to step (3); otherwise, return to step (2.1) , using new sample data to retrain the RBF neural network;

Among them, RMSD (root mean-square deviation) is the root mean square index of the signal measured under different states of the structure. According to this statistical index, the electrical signal of the PZT (piezoelectric ceramics, piezoelectric ceramics) sensor before and after the change of the admittance signal can be measured. The degree of change is obtained according to the following formula:

Among them, n refers to the number of sampling frequencies, i refers to the admittance value of each frequency sampling point, _xi represents the impedance value measured before the structure damage; y _i represents the impedance value measured after the structure damage; where, i= 1, 2, 3, ..., n.

Generally speaking, compared with the prior art, the above technical solutions conceived by the present invention can achieve the following beneficial effects:

(1) The tension compensation method for damage identification provided by the present invention adopts RBF neural network to compensate the impact of tension; frequency sweep frequency and different stressed states are used as input layer, and admittance information is used as output layer; neural network A specific mapping relationship is formed between the input layer and the output layer; through training, the neural network that has been formed is used for simulation; after a certain input is given, the corresponding output is simulated through the neural network; thereby eliminating the pull force on the admittance The influence of the value can reduce the interference of the pulling force on the piezoresistive impedance technology.

(2) The tension compensation method provided by the present invention, in its preferred solution, also includes the step of analyzing whether the network training effect is reasonable, and has the flexibility to continuously adjust the simulation data until the compensation effect in line with the actual working conditions is achieved.

Description of drawings

Fig. 1 is the schematic diagram of RBF neural network;

Fig. 2 is a schematic diagram of the RBF neural network training process;

Figure 3 is the admittance curve and partial enlarged view of PZT2 under various tension conditions of the test piece 1;

Figure 4 is the admittance curve and partial enlarged view of PZT3 under various tension conditions of the test piece 2;

Fig. 5 is a comparison chart of PZT2RBF simulation and measured admittance information of test piece 1;

Fig. 6 is a comparison chart of the PZT3RBF simulation and measured admittance information of the second test piece;

Figure 7 is a comparison chart of the RMSD index of PZT2 admittance of specimen 1 before and after RBF compensation;

Figure 8 is a comparison chart of the RMSD index of the PZT3 admittance before and after RBF compensation when the second specimen is undamaged;

Figure 9 is a comparison chart of the RMSD index of the PZT3 admittance before and after RBF compensation when the second specimen is damaged.

detailed description

In order to make the object, technical solution and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described here are only used to explain the present invention, not to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not constitute a conflict with each other.

The tension compensation method for damage identification provided by the present invention is used for damage detection when the structure is affected by tension. It is based on a three-layer forward RBF neural network. The first layer is the input layer, which is composed of signal source nodes. The second layer is the hidden layer, and the third layer is the output layer; the number of neurons in the hidden layer is determined according to the needs of describing the problem, and the transformation function of the hidden layer is a non-negative nonlinear function with radial symmetry and attenuation at the center point; The transformation from the input layer to the hidden layer control is nonlinear, while the change from the hidden layer control to the output layer is linear;

Use RBF as the "base" of the hidden layer to form the hidden layer control, and map the output vector directly (without connecting through weights) to the hidden space; while the mapping relationship between the hidden layer space and the output space is linear, the output of the network is the linear weighted sum of the hidden unit output; the weight here is the adjustable parameter of the network. On the whole, the mapping of the network from input to output is nonlinear, while the output of the network is linear to the adjustable parameter; the weight of the network is based on Linear equation obtains, thus greatly provided the learning speed of RBF neural network; The radial basis function of RBF neural network adopts Gaussian function in the embodiment; The RBF neural network structure that it adopts of embodiment 1 as shown in Figure 1 schematically, with frequency sweep The frequency and tension are used as the input layer to output the admittance data, and the number of neurons in the hidden layer is determined according to the degree of approach between the mean square error of the RBF neural network and the preset error.

As shown in Figure 2, it is a schematic diagram of the RBF neural network training process, which includes the following steps:

(2.1) Given sweep frequency and pulling force as input vector, given admittance of desired output;

(2.2) Obtain the output of each unit of the hidden layer and the output layer;

(2.3) Obtain the output error; judge whether the output error meets the requirements of the predetermined target, if so, end the training; if not, increase the neurons, and return to step (2.3).

In the embodiment, test piece 1 is a Q235 steel beam with a size of 500mm*35mm*4mm and no damage; test piece 2 is a Q235 steel beam with a size of 500mm*35mm*4mm and a crack of 8mm long Damage: Tensile force is applied to steel beam specimen 1 and specimen 2 with a tensile instrument, and the admittance signal of PZT under different tension conditions is tested with an impedance analyzer.

Adopt the tension compensation method that is used for damage identification provided by the present invention, carry out tension test to test piece 1 and test piece 2; As shown in Figure 3, be the PZT2 each tension working conditions of test piece 1 0kN, 2kN, 4kN, 6kN, Admittance curve and local enlarged diagram measured at 8kN;

As shown in Figure 4, it is the admittance curve and partial enlarged view measured under the tension conditions of PZT3 of test piece 2: 0kN, 4kN, 8kN, 12kN, 16kN;

It can be seen from Figure 3 that the admittance curve of specimen 1 measured in the frequency range from 300kHz to 400kHz shows a downward trend as the tension gradually increases, especially at the peak of the admittance curve.

The frequency sweep frequency of the measured admittance signal and different stress states are used as the input layer, and the admittance information is used as the output layer to form a RBF neural network structure; the number of neurons in the hidden layer is 10.

Take 1/2 of the data measured by PZT2 in Specimen 1 and PZT3 in Specimen 2 as training samples, and 1/2 as test samples; The admittance information under the working condition is compared with the actual measured admittance information; as shown in Figure 5, it is the data comparison of the test piece 1PZT2 under the tensile force values of 0kN, 2kN, 4kN, and 6kN; as shown in Figure 6, it is the test The data comparison of piece 2PZT3 under the tension values of 0kN, 4kN, 8kN, and 12kN tension conditions shows that the simulated admittance data under different tension conditions are basically consistent with the measured admittance values;

It can be seen that after data training under different tension conditions, the RBF neural network can simulate the admittance data under different tension conditions; compared with the measured data, it can be seen that the simulated admittance data is basically consistent with the measured admittance data , which can well predict the admittance data under different tension conditions in a healthy state.

Then use the measured admittance data and the predicted admittance data for comparative analysis, and judge whether the difference between the predicted admittance data and the measured data meets the preset difference expectation, and the preset difference expectation is 5% to 10%. , it is 6% in the embodiment to compensate the influence of tension on the admittance data.

Quantitative analysis of the RBF compensation effect of the specimen under tension without damage and with crack damage; Fig. 7 shows the RMSD damage index of No. Under different tension conditions, they are all around 0.1, and the RMSD index relative to the original data is basically stable;

Therefore, it can be determined that the health state of specimen 1 has not changed, that is, no damage has occurred; the misjudgment of the structural health state due to the influence of tension in the original data has been corrected after RBF compensation.

Figure 8 shows the comparison chart of the RMSD index before and after RBF compensation of the No. 3 piezoelectric film of specimen 2 in the non-destructive state; after the specimen 2 is damaged, as shown in Figure 9, the damage index is compared with the RMSD index after compensation in the non-destructive state From 0.26 to 0.44, it increases obviously, indicating that the structure is damaged;

As the tension increases, the RMSD tends to be stable without significant changes, indicating that the degree of structural damage has not changed; it can be seen that the judgment of the health status of the structure and the degree of damage after compensation have been corrected, and the EMI damage detection method The tension compensation is effective; the method provided by the invention can compensate the tension on the structure and then identify whether the structure is damaged, eliminating the influence of the tension on the accuracy of damage identification.

It is easy for those skilled in the art to understand that the above descriptions are only preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, equivalent replacements and improvements made within the spirit and principles of the present invention, All should be included within the protection scope of the present invention.

Claims (3)

1. a kind of pulling force compensation method for non-destructive tests, it is characterised in that the pulling force compensation method is based on RBF nerve nets Network, comprise the following steps：

(1) using swept frequency and pulling force as input layer, using admittance value as output layer, RBF nerves are formed together with hidden layer Network；

(2) RBF neural is trained using the sample data comprising swept frequency, pulling force and admittance value, until Difference between the admittance value of output admittance value and respective sample data terminates RBF within ± the 5% of sample data admittance value Neural metwork training；

(3) using the swept frequency of test data and pulling force as input, the RBF neural obtained is trained using step (2), is obtained Take admittance value corresponding with the swept frequency and pulling force；The admittance value contains external pull information, compensate for pulling force pair The influence of admittance data；

(4) admittance value obtained according to step (3), the RMSD damage criterions after compensation are obtained；

In RBF neural training described in step (2), the quantity of hidden layer neuron is determined according to following methods：

A, using the input vector corresponding to mean square error caused by RBF neural as weight vector, generate one it is new hidden Neuron containing layer；

B, using test gained admittance value as input, using emulation gained admittance value as output, with step a generate newly it is hidden Neuron containing layer, form new RBF neural；

C, the mean square error of RBF neural new described in obtaining step b；

D, repeat step a~c, until the mean square error of RBF neural reaches default mean square error, with hidden layer god now Quantity through first quantity as RBF neural hidden layer neuron；Wherein, mean square error is preset 5%~10%.

2. the pulling force compensation method described in claim 1, it is characterised in that being instructed to RBF neural described in step (2) Experienced step, including following sub-step：

(2.1) admittance value of swept frequency, pulling force and desired output is given；

(2.2) using the swept frequency and pulling force as input vector, the RBF neural obtained through step (1) is handled, and is obtained Admittance value；

(2.3) admittance value of the admittance value and the desired output that are obtained in step (2.2) is made the difference, obtains difference；Judge institute Difference is stated whether within ± the 5% of desired output admittance value；If so, then terminate to train；If it is not, then increasing neuron, and return Return step (2.2).

3. pulling force compensation method as claimed in claim 2, it is characterised in that after step (2.3) terminates training, using as follows Step determines RBF neural training effect, is specially：

(2.4) data of training sample be will differ from as input vector, emulated, obtained with the RBF neural trained Take the admittance value of simulation data；

(2.5) RMSD values are obtained according to the admittance value of step (2.4) inner simulation data, if the RMSD obtained under each pulling force operating mode Value fluctuates within 5%, then into step (3)；Otherwise, return to step (2.1), using new sample data re -training RBF Neutral net.