Welding machine following type welding seam pre-detection system based on light field energy collection and welding seam defect identification method

Abstract

The invention provides a welding machine following type weld joint pre-detection system based on light field energy collection. The welding machine following type weld joint pre-detection system comprises a mounting frame, and the advancing end of the mounting frame is fixed to the rear end of a welding machine; the energy collecting cover is fixed on the mounting frame and is of a conical structure, and the energy collecting cover is kept in an attached state with the edge of the welding seam through a spring locking device; the laser source is fixed in the energy collecting cover and is provided with a power amplifier and a combined lens; the laser radar is fixed above the energy collecting cover; the upper computer receives information collected by the laser radar and the energy collection cover; the weld defect identification system is in communication connection with the upper computer; the sample data formed by combining the welding seam position information and the light field energy information collected by the energy collecting cover is trained, and due to the fact that part of light field energy leaks through the welding seam when the defect welding seam exists, the light field energy is obviously reduced compared with the light field energy collected through the non-defect welding seam, the defect welding seam can be efficiently recognized and confirmed, and the detection accuracy is improved. And a technical guarantee is provided for effectively searching helium detection.

Description

Technical Field

[0001] The present invention relates to the technical field of ship helium detection, and in particular to a welder-following weld pre-detection system and a weld defect identification method based on light field energy collection. Background Art

[0002] Since the membrane containment system is formed by welding, even the smallest defect on the weld may cause natural gas leakage and explosion, so it is very important to test the tightness of the weld. Currently, the inspection of welds mainly relies on subsequent manual handheld suction guns to conduct multi-dimensional inspection of the entire weld through a helium detector.

[0003] Although the tightness test procedures for LNG ships emphasize the need to conduct multiple full inspections on all welds, conducting follow-up pre-inspections when the welds are just formed can increase the focus of subsequent helium inspections. Operators can increase the number of inspections or slow down the inspection speed at key points to reduce the re-inspection rate and improve the efficiency of tightness testing.

[0004] Therefore, it is necessary to design a welder-following weld pre-detection technology based on light field energy collection. Summary of the invention

[0005] The purpose of the present invention is to provide a welding machine following type weld pre-detection system and a weld defect identification method based on light field energy collection.

[0006] In order to achieve the above object, the technical solution of the present invention is:

[0007] The welder-following weld pre-detection system based on light field energy collection is characterized in that the pre-detection system includes

[0008] A mounting frame, the forward end of which is fixed to the rear end of the welding machine;

[0009] An energy collection cover is fixed on the mounting frame, and the energy collection cover is a conical structure;

[0010] A laser source is fixed inside the energy collection cover, and the laser source is provided with a power amplifier and a combined lens;

[0011] A laser radar is fixed above the energy collection cover;

[0012] A host computer receives information collected by the laser radar and the energy collection cover;

[0013] A weld defect identification system, which is in communication connection with the host computer;

[0014] Furthermore, the energy collection cover is kept in close contact with the edge of the weld by a spring locking device.

[0015] Furthermore, the mounting frame is a square hollow structure.

[0016] A pre-detection method using the welding machine following type weld pre-detection system based on light field energy collection is characterized by comprising the following steps:

[0017] Step S1) the weld defect recognition system inputs the position information of the defect-free weld to obtain training data, i.e., early normal samples;

[0018] Step S2) the training data is automatically encoded by a first encoder;

[0019] Step S3) The encoded data enters the generative adversarial network and is decoded, and the generative network and the discriminative network are used for adversarial training;

[0020] Step S4) generating adversarial network outputs adversarial trained data, which is encoded by a second encoder;

[0021] Step S5) determines whether the training is completed. If not, returns to step S2, and the data after this training enters the first encoder for iterative optimization. If completed, jumps to step S6.

[0022] Step S6) obtaining the trained first encoder, the generator network and the second encoder;

[0023] Step S7) the weld defect identification system inputs the light field energy information collected by the energy collection cover and the laser radar and its corresponding detection position to obtain test data;

[0024] Step S8) the test data is passed through the first encoder, the generator network and the second encoder, encoding-decoding-decoding again;

[0025] Step S9) Analyze the potential feature differences between the test data and the training data to complete defect detection.

[0026] The present invention trains on sample data composed of weld position information and light field energy information collected by an energy collection cover. Since part of the light field energy will leak through the weld when there is a defective weld, the light field energy collected is significantly reduced compared to that of a non-defective weld. Therefore, defective welds can be efficiently identified and confirmed, providing technical support for effective helium detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is a schematic diagram of the system structure of the present invention;

[0028] Figure 2 A weld defect identification method according to an embodiment of the present invention.

[0029] Figure 3 Schematic diagram of anomaly detection based on training model.

[0030] Reference numerals:

[0031] 1 mounting frame, 2 energy collection cover, 3 laser source, 4 laser radar, 5 weld defect identification system,

[0032] 6 Host computer. DETAILED DESCRIPTION

[0033] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0034] This embodiment discloses a welder-following weld pre-detection system based on light field energy collection. The pre-detection system includes a mounting frame 1, an energy collection cover 2, a laser source 3, a laser radar 4, a weld defect identification system 5 and a host computer 6.

[0035] The forward end of the mounting frame 1 is fixed to the rear end of the welding machine by fasteners, and provides a mounting position for the energy collection cover. Preferably, the mounting frame 1 is a steel square hollow structure.

[0036] The energy collection cover 2 is made of non-reflective energy-absorbing material. Preferably, the energy collection cover 2 is made of a multifunctional metamaterial. It is fixed on the mounting frame 1 and is always kept in close contact with the edge of the weld through a spring locking device. The energy collection cover 2 is connected to the collection device to realize real-time transmission of light field energy data signals.

[0037] Preferably, the energy collection cover 2 is a conical structure, which can cover a circular area.

[0038] The laser source 3 is fixedly arranged inside the energy collection cover 2. The laser source 3 focuses the ultra-strong light on a small weld area through a power amplifier and a combined lens. The laser source 3 moves forward at a uniform speed following the welder.

[0039] The laser radar 4 is fixedly arranged inside the energy collection cover 2. By reconstructing the model of the weld based on the point cloud data, it communicates with the communication and collection end to record the position of the weld detection point. The laser radar 4 remains in a working state and saves each sampling point as an independent position mark corresponding to the independent light intensity information according to the sampling frequency. In this embodiment, the communication and collection end of the laser radar 4 is an industrial computer.

[0040] This embodiment also discloses a weld defect identification method of a welder-following weld pre-detection system based on light field energy collection, such as Figure 2As shown, the weld defect identification system 5 is implemented based on a deep neural network. However, the establishment of a traditional deep learning model requires a large amount of historical data and corresponding labels, and in actual diagnosis work, only data samples under the normal state of the machine can be obtained, so it is difficult to establish a traditional intelligent diagnosis model.

[0041] Preferably, in one embodiment, the deep neural network is a network structure of encoding-decoding-re-encoding constructed by combining the adversarial neural network GAN and the autoencoder AE. The host computer 6 is trained by combining the sample data composed of a sufficient number of defect-free weld position information and the light field energy information collected by the energy collection cover 2. The normalized spectrum is used as the input data of the model. The decoder in the AE, that is, the generating network, obtains the generated sample. The AE learns the potential features of the normal sample and the generated sample through the mutual game between the generating network and the discriminating network, so that the difference between the potential features obtained by the two encodings is extremely small. After the model training is completed, the unknown weld state collected online in real time is tested, and finally the difference between the two potential features of the output is calculated, which is used as an indicator to determine whether there is a defect.

[0042] It should be noted that the adversarial neural network GAN learns the data distribution of training samples through the game process of the generator network G and the discriminant network D, so that the generator network outputs training samples that are indistinguishable from the real ones, thereby solving the data imbalance problem in actual fault diagnosis where the number of fault samples is less than the number of normal samples. The ultimate goal is: when the discriminant network inputs x, D(x) is close to 1, and when G(z) is input, D(G(z)) is close to 0. The objective function is expressed as:

[0043]

[0044] For welding systems, since the collected data are normal samples in most cases, the generative adversarial network that learns real samples and generated samples will have large errors when used to reconstruct the data distribution of defective samples, which also provides the possibility for mechanical system anomaly detection under defect-free sample training.

[0045] The anomaly detection model in this embodiment is an encoding-decoding-re-encoding structure. The encoder and decoder form an AE, where the encoder is used to learn the low-dimensional features of the input signal, and the decoder reproduces the input signal as much as possible. In this embodiment, the decoder is the generator network G; the specific implementation plan is as follows:

[0046] AE learns the latent feature z of the input sample x through encoder 1, and then restores the input sample x through the decoder. The input of the generative network is the latent feature of the input sample learned by the encoder, which ensures the consistency of the input data and reduces the learning difficulty of GAN. The game between the generative network and the discriminative network enables the latent feature z to better reflect the effective information of the input sample x. However, since the generated sample G(z) is generated by the latent feature z excluding interference factors, the gap with the input sample x makes the direct comparison have a large error fluctuation. In order to obtain the indicator for detecting abnormal samples, the generated sample G(z) is further encoded by the encoder E2 to obtain its latent feature

[0047] The training process of the anomaly detection model is the process of learning the data distribution and potential features of normal samples, such as Figure 3 As shown:

[0048] 1) Input the normalized spectrum x of the normal sample into the model, and obtain the potential feature z of x through encoder 1. Then, the potential feature z is passed through the generative network G to obtain the generated sample G(z). In order to make the data distribution of G(z) similar to that of x, optimize the objective function L con :

[0049]

[0050] Wherein, N is the data length of x. In addition, the objective function can also optimize the encoder 1 to obtain reliable potential features z.

[0051] 2) The discriminant network D distinguishes the normal sample x and the generated sample G(z). In order to improve the feature restoration ability of the generated network and the stability of model training, the error of matching features between x and G(z) in the discriminant network is optimized. The objective function L adv for:

[0052]

[0053] Where D f (·) represents the output of the middle layer of the discriminant network. To improve the discrimination ability of the discriminant network, the objective function L is optimized. d :

[0054]

[0055] 3) Through the objective function, the adversarial training of the generating network and the discriminative network is formed, and the generated sample G(z) is input into the encoder 2 to obtain the potential code of G(z) At this time, the optimized encoding features z and The distribution in the latent space, the objective function L enc for:

[0056]

[0057] Wherein, M is the data length of z.

[0058] 4) Fix the weights in the generation network and the two encoders, and optimize the discriminant network according to the L d objective function, then fix the discriminant network and optimize the following objective function:

[0059] L = w con L con + w adv L adv + w enc L enc

[0060] Wherein, w con 、w adv and w enc respectively represent the weights of L con 、L adv and L enc .

[0061] After the model training is completed (i.e., i < n, where n is the number of training samples), the real-time monitoring samples collected online are tested, as Figure 3 shown. Fix the weight parameters in the trained Encoder 1, Generation Network G, and Encoder 2 network. The normalized spectrum x of the monitoring sample is obtained as the latent feature z through Encoder 1, and then the latent feature is obtained through G and Encoder 2. Since the model is trained with normal samples and can only correctly learn the data distribution and latent features of normal samples, when the monitoring sample is a normal sample, the difference between the latent features z and output by the two encoders is extremely small. When there is a defective weld, part of the light field energy will leak through the weld, and the energy collected relative to the non-defective weld is significantly reduced. Therefore, when there is a defective weld sample, the difference between the two latent features z and increases, thus realizing defect detection under the training of non-defective samples. The difference between the latent features is reflected by the mean square error, and the calculation formula is as follows. Continuously observing the change trend of MSE can detect the abnormality of the mechanical system in time:

[0062]

[0063] In another embodiment, the deep neural network of the present invention is built based on a convolutional neural network CNN, and the upper computer 6 is trained by combining a sufficient number of sample data composed of defective weld position information, non-defective weld position information and light field energy information collected by the energy collection cover 2. During the training process, the corresponding features of the defective weld sample data and the non-defective weld sample data are extracted respectively. After the training is completed, when the input sample is a defective weld sample, the defect diagnosis is achieved by comparing it with its characteristic value.

[0064] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A welding machine following weld pre-detection system based on light field energy collection, characterized in that: The pre-detection system comprises A mounting frame (1), the forward end of the mounting frame (1) being fixed to the rear end of the welding machine; An energy collection cover (2) is fixed on the mounting frame (1), and the energy collection cover (2) is a conical structure; A laser source (3) is fixed inside the energy collection cover (2), wherein the laser source (3) is provided with a power amplifier and a combined lens; A laser radar (4) is fixed above the energy collection cover (2); A host computer (6) receives information collected by the laser radar (4) and the energy collection cover (2); The weld defect identification system (5) is communicatively connected to the host computer (6).

2. The welding machine following weld pre-detection system based on light field energy collection according to claim 1 is characterized in that: The energy collection cover (2) is kept in close contact with the edge of the weld by means of a spring locking device.

3. The welding machine following weld pre-detection system based on light field energy collection according to claim 1 is characterized in that: The mounting frame (1) is a square hollow structure.

4. The weld defect identification method of the welder following weld pre-detection system based on light field energy collection according to claim 1 is characterized in that: The following steps are involved: Step S1) The weld defect recognition system (5) inputs the position information of the defect-free weld to obtain training data, i.e., early normal samples; Step S2) the training data is automatically encoded by the first encoder; Step S3) The encoded data enters the generative adversarial network and is decoded, and the generative network and the discriminative network are used for adversarial training; Step S4) Generate adversarial network outputs adversarially trained data, which is encoded by a second encoder; Step S5) Determine whether the training is completed. If not, return to step S2. The data after this training enters the first encoder for iterative optimization. If completed, jump to step S6. Step S6) obtaining the trained first encoder, the generator network and the second encoder; Step S7) the weld defect identification system (5) inputs the light field energy information and the corresponding detection position collected by the energy collection cover (2) and the laser radar (4) to obtain test data; Step S8) The test data is passed through the first encoder, the generator network and the second encoder, and encoded-decoded-decoded again; Step S9) Analyze the potential feature differences between the test data and the training data to complete defect detection.

Images