A neural network-based laser welding defect detection method and system

Abstract

The present application relates to a kind of neural network-based laser welding defect detection method and system, the method comprises: step S1: the data acquisition environment of laser welding machine is built;Step S2: data is collected and the data collected is constantly put into buffer queue;Step S3: input parameter is constructed based on the acquisition data;Step S4: neural network model is constructed, and sample set is used to train neural network model until reaching training target;Step S5: based on the N long triplet parameter matrix of continuous N triplet parameters as input parameter is constructed;Step S6: input parameter is input into neural network model to obtain detection result.The present application realizes the intelligent detection of laser welding defect, and can use laser welding process data to judge welding defect with high recognition accuracy.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of intelligent defect detection, and particularly relates to a laser welding defect detection method and system based on a neural network. BACKGROUND

[0002] As an important technology in industrial production and processing, laser welding has been concerned by researchers at home and abroad. In the production process of metal products, laser welding as an important processing technology is increasingly widely used in industrial production. The light and heat change in the laser welding process will cause the quality of laser welding to deviate. Due to the small laser beam spot and the intense heat conversion effect in the welding process, the precision requirements of welding process parameters and workpiece fixation are very high, and a slight welding deviation can cause serious welding defects, thereby affecting the quality of the welded product. Therefore, it is particularly important to quickly and effectively detect defects in the welding product in the mass production process. Laser welding technology has been developed for a long time since it was proposed in the 1960s, but the intelligent detection of welding quality has not kept pace with the research progress. The existing laser welding process monitoring technology mostly focuses on detecting post-welding defects. Conventional non-destructive testing has certain limitations. For example, visual inspection is prone to missed detection, radiation detection is dangerous and expensive, magnetic powder detection and penetration detection mainly locate the defect position but cannot specifically display the shape, and the detection results of eddy current detection and other signal detection technologies cannot directly reflect the defect morphology.

[0003] In the prior art, image processing is often used for defect recognition, but an image often contains a large amount of information, and when a large amount of information is used, the timeliness cannot be estimated, that is, the defect recognition based on the image is large in cost and is not suitable for real-time detection or light detection; on the other hand, a complex model often indicates a complex training and simple environmental conversion, and the object of the welding technology is often complex, and the environment will also change with the climate and other factors, so the neural network analysis method depending on the image cannot quickly change the recognition environment with the change of the training environment, and when the operation target changes, the model without sufficient training is also useless, that is, the prior art has the problems of poor timeliness and low recognition accuracy, which are technical problems to be solved.

[0004] However, in fact, the change of physical quantity in the laser welding process has reflected the difference of welding quality, the detection model of physical quantity in the welding process needs to be able to extract the subtle features of the change of process quantity, as long as these features can be found and used, the defects can be effectively detected; the neural network model is used for space feature extraction, the long short-term memory network model is used for extracting the front and rear dependence relationship and time sequence features of the time sequence data in the process time, and the multi-dimensional information is used for welding defect detection; SUMMARY

[0005] In order to solve the above problems in the prior art, the present application provides a laser welding defect detection method and system based on a neural network,

[0006] The method comprises:

[0007] Step S1: build a laser welding machine data acquisition environment, set a photodiode, a temperature sensor prism to detect light intensity data and temperature data;

[0008] Step S2: collect data and continuously put the collected data into a buffer queue; specifically, collect plasma intensity ItP t , light intensity information LP t , temperature TmP t These three types of parameters are sent to the data acquisition terminal, the data acquisition terminal forms a three-tuple parameter at time t from the parameters collected at time t, and puts the three-tuple parameter at time t into the buffer queue;

[0009] Step S3: build input parameters based on collected data; the matrix composed of N three-tuple parameters obtained at N continuous time points is taken as an input parameter and put into a sample set;

[0010] Step S4: build a neural network model, and train the neural network model with the sample set until the training target is reached;

[0011] Step S5: receive the preprocessed laser welding data to form a three-tuple parameter, and build an N-long three-tuple parameter matrix as an input parameter based on continuous N three-tuple parameters;

[0012] Step S6: input the input parameter into the neural network model to obtain a detection result.

[0013] Further, the front-end signal acquisition mainly consists of two different photodiode sensors, a thermometer sensor and a 45-degree prism; the first photodiode sensor is installed behind the 45-degree prism to obtain the plasma intensity ItP; behind the 45-degree prism, the second photodiode sensor is arranged to capture the light intensity information LP; the end pyrometer sensor is used to detect the temperature TmP in the welding process.

[0014] Further, the N is a preset value.

[0015] Further, N = 8.

[0016] Further, the step S3 specifically comprises the following steps:

[0017] Step S31: determining a length-N triad parameter starting from a head pointer as a target input parameter; wherein the head pointer points to a head of a buffer queue;

[0018] Step S32: obtaining a length-N triad parameter starting from an active pointer as a candidate input parameter; wherein the active pointer is used to determine a next target input parameter, that is, a next position of the head pointer;

[0019] Step S33: determining a continuous change condition of same-type parameters between the candidate input parameter and the target input parameter and a relationship change condition between different-type parameters; if both the continuous change condition and the relationship change condition meet a preset condition, proceeding to step S34; otherwise, proceeding to step S35;

[0020] Step S34: judging whether a distance between the active pointer and the head pointer is less than N, if yes, incrementing the active pointer and returning to step S32; if no, proceeding to the next step;

[0021] Step S35: taking the target input parameter of the length-N triad parameter position starting from the head pointer as an input parameter and putting it into a sample set; deleting triad data between the head pointer and a position before the active pointer, resetting the head pointer position to the active pointer position, and incrementing the active pointer position; if the sample set has been constructed completely, proceeding to step S4; otherwise, returning to step S31.

[0022] A laser welding defect detection system based on a neural network, the system comprising:

[0023] A collection module: building a laser welding machine data collection environment, setting a photodiode, a temperature sensor prism to detect light intensity data and temperature data; collecting data and constantly putting the collected data into a buffer queue; specifically: collecting plasma intensity ItP t , light intensity information LP t , temperature TmP t These three types of parameters and send them to a data collection terminal, the data collection terminal forms a t-time triad parameter from the parameters collected at t time, and puts the t-time triad parameter into the buffer queue;

[0024] ​​Parameter construction module: based on the collected data to construct input parameters; the matrix composed of N triple parameters obtained at continuous N time points is taken as input parameters and put into the sample set;

[0025] Model building module: build a neural network model, and train the neural network model with the sample set until the training target is reached;

[0026] Real-time acquisition module: after receiving the preprocessed laser welding data, form triple parameters, and based on continuous N triple parameters, build N long triple parameter matrix as input parameters;

[0027] Detection module: input the input parameters into the neural network model to obtain the detection result.

[0028] Further, the acquisition module includes two different photodiode sensors, a thermometer sensor and a 45-degree prism.

[0029] A processor for running a program, wherein the program performs the neural network-based laser welding defect detection method when running.

[0030] A computer-readable storage medium comprising a program that, when executed on a computer, causes the computer to perform the neural network-based laser welding defect detection method.

[0031] An execution device comprising a processor, wherein the processor and the memory are coupled, and the memory stores program instructions, and when the program instructions stored in the memory are executed by the processor, the neural network-based laser welding defect detection method is realized.

[0032] The beneficial effects of the present application include:

[0033] (1) Through the rapid acquisition of different samples, the rapid establishment of different detection models when the data acquisition environment changes significantly is supported, the acquisition data amount of effective samples is fully expanded on the basis of not losing the time characteristics of the collected data, sufficient effective samples under the target experimental environment are acquired in a short time, and through N length adjustment, a good balance between model complexity and prediction accuracy can be achieved; (2) Different complexity input parameter determination methods are set in the training sample acquisition and model use process, so that sufficient training during training and rapid acquisition of input parameters during use are realized, and the detection efficiency is fully improved;

[0034] (3) By expanding the parameters in the time dimension, the time sequence characteristics are added on the basis of simple detection parameters, compared with the image-based recognition method, the data acquisition is simplified while the different defect detection environments and detection targets can be quickly adapted, and the method has the advantages of strong timeliness and high recognition accuracy; BRIEF DESCRIPTION OF DRAWINGS

[0035] The accompanying drawings, which are included to provide a further understanding of the application and are incorporated in and constitute a part of this application, illustrate embodiments of the application and together with the description serve to explain the principles of the application. In the drawings:

[0036] Figure 1 It is a schematic diagram of the laser welding machine data acquisition environment of the present application.

[0037] Figure 2 It is a schematic diagram of the laser welding defect detection method based on neural network of the present application. DETAILED DESCRIPTION

[0038] The application will be described in greater detail with reference to the accompanying drawings and specific embodiments, which are included to provide a further understanding of the application and are incorporated in and constitute a part of this application, illustrate embodiments of the application and together with the description serve to explain the principles of the application. In the drawings:

[0039] As shown in the accompanying drawings Figure 1 , the present application proposes a laser welding defect detection method based on neural network, which comprises the following steps:

[0040] Step S1: build a laser welding machine data acquisition environment, set up a photodiode, a temperature sensor prism to detect light intensity data and temperature data; specifically: as shown in the accompanying drawings Figure 1 . The front-end signal acquisition mainly consists of two different photodiode sensors, a thermometer sensor and a 45-degree prism. The first photodiode sensor is installed behind the 45-degree prism to obtain the plasma intensity A second photodiode sensor is set behind the 45-degree prism to capture the light intensity information LP; the end pyrometer sensor is used to detect the temperature TmP during the welding process; the detected plasma intensity ItP t , light intensity information LP t , temperature TmP t is sent to the data acquisition terminal in real time;

[0041] Preferably: in the process of laser welding, photodiode and other devices are used to collect photoelectric signals, and the photoelectric signals of the plasma collected in the process of laser welding are converted into electrical signal data with time sequence;

[0042] Step S2: collect data and constantly put the collected data into the buffer queue; specifically: collect the plasma intensity ItP t , light intensity information LP t , temperature TmP t These three types of parameters and send them to the data acquisition terminal, and the data acquisition terminal will The parameters collected at the moment constitute a three-tuple parameter at the moment t, and the three-tuple parameter at the moment t is put into a buffer queue; the three-tuple parameters in the buffer queue are stored in time sequence;

[0043] Preferably, the collected data is pre-processed before being put into the buffer queue, and the collected data is standardized to eliminate the influence of dimension and value range difference;

[0044] Step S3: constructing input parameters based on the collected data; a matrix composed of N three-tuple parameters acquired at N moments is taken as an input parameter and put into a sample set; that is, the input parameter is an N*3 matrix; the sample set stores a plurality of input parameters constructed based on the collected data;

[0045] The step S3 specifically includes the following steps:

[0046] Step S31: determining the three-tuple parameters with a length of N starting from the head pointer as target input parameters; wherein, the head pointer points to the head of the buffer queue;

[0047] Step S32: acquiring the three-tuple parameters with a length of N starting from the active pointer as candidate input parameters; wherein, the active pointer is used to determine the next target input parameter, that is, the next position of the head pointer;

[0048] Step S33: determining the continuous change of the same type parameters between the candidate input parameters and the target input parameters and the relationship change between different type parameters; if the continuous change and the relationship change both satisfy a preset condition, step S34 is entered; otherwise, step S35 is entered;

[0049] The step S33 specifically includes the following steps:

[0050] Step S331: acquiring the candidate input parameters CDP and the target input parameters TGP;

[0051] CDP=(Cdp td ) =(ItP td , LP td , TmP td ); td∈CDP time range (td1~tdN);

[0052] TGP=(Tgp tt ) =(ItP tt , LP tt , TmP tt ); tt∈TGP time range (tt1~ttN);

[0053] Step S232: sorting the same type of parameters of the candidate input parameters and the target input parameters respectively, obtaining a parameter sequence corresponding to the candidate input parameters (S_Cdp tti ,(S_Cdp tti for ItP tt , LP tt or TmP tt ) and a parameter sequence corresponding to the target input parameters (S_Tgp tdi ,(S_Tgp tdi for ItP td , LP td or TmP td ); i = 1 ~ N;

[0054] Step S333: calculating the parameter sequence difference degree CTNS and the difference fluctuation range RCTNS = (VCTNS-D1, VCTNS+D2) of each type respectively; wherein: VCTNS is the parameter mean value; D1 is the lower limit fluctuation value, and D2 is the upper limit fluctuation value; σ is the fluctuation constant value;

[0055]

[0056]

[0057]

[0058] The sorting sequence comparison allows the occurrence and continuous occurrence of the burr value within the N window, allows the common fluctuation range of the collected data, filters the repeated collection of invalid samples and subsequent invalid training; the right side of formula (2) is the parameter mean value before updating, and the left side is the parameter mean value after updating;

[0059] Preferably:

[0060] Step S334: determining whether the parameter sequence difference degree of each type of parameter is within the difference fluctuation range, if yes, entering the next step, otherwise, entering step S35;

[0061] Step S335: determining the type parameter relationship f(ItP t , LP t , TmP t ) between different types of parameters based on the historical collected data; calculating the relationship value standard deviation σ; is the relationship value mean value calculated based on the historical collected data;

[0062]

[0063] Preferred: the type parameter relationship f is calculated in a fitting manner;

[0064] Alternative: the type parameter relationship is the relationship between two types of parameters, and then the parameter relationship of each two types is judged respectively;

[0065] Step S336: judge whether the relationship value standard deviation is less than the reference standard deviation, if yes, go to step S34, otherwise, go to step S35;

[0066] Preferred: the reference standard deviation is a preset value; it can be flexibly set according to the detection environment and the detection object;

[0067] Step S34: judge whether the distance between the active pointer and the head pointer is less than N, if yes, increment the active pointer, and return to step S32; if not, go to the next step;

[0068] Step S35: input the target input parameters of the N-length triple parameter position starting from the head pointer as input parameters and put them into the sample set; delete the triple data between the head pointer and the position one before the active pointer, reset the head pointer position to the active pointer position, and increment the active pointer position; at this time, if the buffer queue has been processed or the sample set has been constructed, go to step S4, otherwise, return to step S31; that is, although there is no meaningful change in the N-length parameters, the N-length triple parameters are still intercepted in the case of N-dimensional expansion failure to maintain the acquisition of regular samples, so as to ensure the training of the basic detection environment;

[0069] The target input parameters of the N-length triple parameter position are put into the sample set as input parameters, specifically: based on the time range involved by the input parameters, the laser welding product quality data is searched, the laser welding product quality data with defect labels falling into the same time range is used to associate the input parameters to form a sample, and the sample is put into the sample set;

[0070] Through such a way of constructing the sample set, the dimension of the sample is expanded in an iterative cycle without losing the time characteristics of the collected data, which fully expands the amount of effective sample data, and for the same N-length sample, the effective sample data can be expanded by N times in the best case, in addition, through the adjustment of N length, a good balance between model complexity and prediction accuracy can be achieved;

[0071] Step S4: construct a neural network model, and train the neural network model with the sample set until the training target is reached; for example, a convolutional neural network model;

[0072] Preferably, the neural network model is a deep neural network model, which adopts a bidirectional long short-term memory network and an attention mechanism;

[0073] Preferably, an LSTM network model is adopted, and a self-attention mechanism is added after the BiLSTM layer outputs, so as to filter out a small amount of important information from a large amount of information, focus on the important information, and ignore a large amount of unimportant information;

[0074] Preferably, the cosine annealing learning rate can jump out of the local minimum value and find a path to the global minimum value by suddenly increasing the learning rate, so that the model can quickly converge to the optimal solution;

[0075] Step S5: receiving the preprocessed laser welding data to form a triple parameter, and constructing an N-long triple parameter matrix based on the continuous N triple parameters as an input parameter;

[0076] The step S5 further comprises: determining the arithmetic difference degree Sub between the continuous two N-long triple parameter matrices; if the arithmetic difference degree is greater than a preset arithmetic difference degree, the latter triple parameter matrix of the continuous two N-long triple parameter matrices is taken as an input parameter and then enters step S6, otherwise, returns to step S5, so as to continue to receive welding data to obtain an input parameter with an arithmetic difference degree;

[0077] Alternatively, the continuous two N-long triple parameter matrices are both taken as input parameters and then enter step S6,

[0078] Here, the continuous two N-long triple parameter matrices include a newly acquired parameter matrix (second parameter matrix) and an old parameter matrix (first parameter matrix); wherein: the old parameter matrix is an input parameter that must be detected at a fixed time interval and obtains a detection result without quality problems in step S6; the newly acquired parameter matrix is newly constructed based on real-time acquired welding data; when the difference degree is not enough, the newly acquired parameter matrix is ignored and a new parameter matrix is continuously constructed to obtain an input parameter with a difference degree, that is, a detection value;

[0079] Preferably, the triple parameters contained in the continuous two N-long triple parameter matrices are not repeated, so that here the continuous can also be an N-long triple parameter matrix constructed based on continuous triple parameters;

[0080] Preferably, 130 samplings are performed, and two N-long triple parameter matrices are constructed based on the acquisition data obtained by the 130 samplings;

[0081] The arithmetic difference Sub between two consecutive N-length triplet parameter matrices is specifically determined as follows: the arithmetic difference Sub is calculated using the following formula (5);

[0082] Sub=w1×|∑ t∈CT ItP t -∑ t∈TT ItP t |+w2×|∑ t∈CT LP t -∑ t∈TT LP t |+w3×|∑ t∈CT TmP t -∑ t∈TT TmP t |; (5)

[0083] Where: two consecutive N-length triplet parameter matrices are the first parameter matrix and the second parameter matrix; CT is the time range involved in the first parameter matrix; TT is the time range involved in the second parameter matrix; w1, w2, w3 are weight values; the arithmetic difference can be set according to the on-site conditions of laser welding;

[0084] Step S6: Input the input parameters into the neural network model to obtain the test results. Obviously, one possibility is that the input parameters here are the N-length parameter matrix composed of the latest N triplet parameters. In this case, the first parameter matrix may have been used as the input parameter to obtain a test result without quality problems.

[0085] Alternatively, the two consecutive N-length triplet parameter matrices are input into a neural network model to obtain two detection results, and a final detection result is determined based on the two detection results; for example, the final detection result is determined by manual feedback;

[0086] Preferably, the latest N triplet parameters are obtained from an online detection process of welding defects during laser welding of a SUS304 material workpiece;

[0087] Based on the same inventive concept, the present invention also provides a neural network-based laser welding defect detection system, the system comprising:

[0088] Acquisition module: build laser welding machine data acquisition environment, set up photodiode, temperature sensor prism to detect light intensity data and temperature data; collect data and continuously put the collected data into the buffer queue; specifically: collect plasma intensity ItP t , Guangqiang Information LP t , temperature TmP tThe three types of parameters are sent to the data acquisition terminal, and the data acquisition terminal forms a t-time triplet parameter from the parameters collected at t time, and puts the t-time triplet parameter into a buffer queue;

[0089] The parameter construction module: constructs input parameters based on the collected data; and puts a matrix formed by N triplet parameters acquired at N continuous time points into a sample set as input parameters;

[0090] The model construction module: constructs a neural network model, and trains the neural network model using the sample set until a training target is reached;

[0091] The real-time acquisition module: forms a triplet parameter after receiving preprocessed laser welding data, and constructs an N-long triplet parameter matrix as input parameters based on continuous N triplet parameters;

[0092] The detection module: inputs the input parameters into the neural network model to obtain a detection result.

[0093] The terms "data processing apparatus", "data processing system", "user device" or "computing device" encompass all kinds of apparatus, devices, and machines for processing data, including by way of example a programmable processor, a computer, a system on a chip, or multiple ones of the same or a combination thereof. The apparatus can include special purpose logic circuitry, e.g., an FPGA (field programmable gate array) or an ASIC (application specific integrated circuit). The apparatus can also include, in addition to hardware, code that creates an execution environment for the computer program in question, e.g., code that constitutes processor firmware, a protocol stack, a database management system, an operating system, a cross-platform runtime environment, a virtual machine, or a combination of one or more of them. The apparatus and execution environment can realize various different computing model infrastructures, such as web services, distributed computing, and grid computing infrastructures.

[0094] A computer program (also known as a program, software, software application, script, or code) can be written in any form of programming language, including compiled or interpreted languages, declarative or procedural languages, and it can be deployed in any form, including as a stand-alone program or as a module, component, subroutine, object, or other unit suitable for use in a computing environment. A computer program may, but need not, correspond to a file in a file system. A program can be stored in a portion of a file that holds other programs or data (e.g., one or more scripts stored in a markup language document), in a single file dedicated to the program in question, or in multiple coordinated files (e.g., files that store one or more modules, sub programs, or portions of code). A computer program can be deployed to be executed on one computer or on multiple computers that are located at one site or distributed across multiple sites and interconnected by a communication network.

[0095] Those skilled in the art will appreciate that embodiments of the application can be devised for a method, a system, or a computer program product. Accordingly, the present application can be embodied in the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present application can take the form of a computer program product on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, etc.) embodying computer readable program code.

[0096] The present application is described in reference to the flowchart illustrations and / or block diagrams of the methods, apparatus (systems) and computer program products according to embodiments of the application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart illustrations and / or block diagrams block or blocks. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams block or blocks. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams block or blocks.

[0097] These computer program instructions can also be stored in a computer- readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the functions specified in the flowchart illustrations and / or block diagrams block or blocks. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams block or blocks. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams block or blocks.

[0098] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart illustrations and / or block diagrams block or blocks. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams block or blocks. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams block or blocks.

[0099] Finally, it should be noted that the above-mentioned embodiments are only intended to illustrate the technical solutions of the present application, and are not intended to limit the present application. Although the present application has been described in detail with reference to the above-mentioned embodiments, those skilled in the art should understand that the specific embodiments of the present application can be modified or replaced by equivalents without departing from the spirit and scope of the present application, and any modifications or equivalent replacements without departing from the spirit and scope of the present application should be covered within the protection scope of the claims of the present application.

Claims

1. A neural network-based laser welding defect detection method, characterized by, The application relates to a laser welding data preprocessing method and system. Step S1: a laser welding machine data collection environment is built, and a photodiode, a temperature sensor and a prism are arranged to detect light intensity data and temperature data; Step S2: data is collected and the collected data is continuously put into a buffer queue; Specifically: collecting plasma intensity ItP t , light intensity information LP t , temperature TmP t The three types of parameters are sent to the data acquisition terminal, which forms a t-time three-tuple parameter by collecting the parameters at t time and puts the t-time three-tuple parameter into a buffer queue. Step S3: input parameters are constructed based on the collected data; a matrix formed by N three-tuple parameters obtained at continuous N time points is taken as an input parameter and is put into a sample set; The step S3 specifically comprises the following steps: Step S31: three-tuple parameters with a length of N starting from a head pointer are determined as target input parameters; wherein the head pointer points to the head of the buffer queue; Step S32: three-tuple parameters with a length of N starting from an active pointer are obtained as candidate input parameters; wherein the active pointer is used to determine the next target input parameter, that is, the next position of the head pointer; Step S33: the continuous change condition of the same type parameters and the relationship change condition between different type parameters between the candidate input parameters and the target input parameters are determined; if the continuous change condition and the relationship change condition both satisfy preset conditions, the step S34 is entered; otherwise, the step S35 is entered; Step S34: whether the distance between the active pointer and the head pointer is less than N is judged; if yes, the active pointer is incremented, and the step S32 is returned; if no, the next step is entered; Step S35: the target input parameters of the three-tuple parameters with a length of N starting from the position of the head pointer are taken as input parameters and are put into the sample set; the three-tuple data between the head pointer and the position before the active pointer are deleted, the position of the head pointer is reset to the position of the active pointer, and the position of the active pointer is incremented; if the sample set has been constructed, the step S4 is entered; otherwise, the step S31 is returned; Step S4: a neural network model is constructed, and the neural network model is trained by using the sample set until a training target is reached; Step S5: three-tuple parameters are formed after receiving preprocessed laser welding data, and an N-long three-tuple parameter matrix is constructed based on continuous N three-tuple parameters as input parameters; Step S6: the input parameters are input into the neural network model to obtain a detection result.

2. The neural network-based laser welding defect detection method of claim 1, wherein, The front-end signal collection mainly comprises two different photodiode sensors, a pyrometer sensor and a 45-degree prism; the first photodiode sensor is installed behind the 45-degree prism to obtain plasma intensity ItP; the second photodiode sensor is arranged behind the 45-degree prism to capture light intensity information LP; and the terminal pyrometer sensor is used to detect temperature TmP in the welding process.

3. The neural network-based laser welding defect detection method of claim 2, wherein, The N is a preset value.

4. The neural network-based laser welding defect detection method of claim 3, wherein, N＝8。 5. A neural network-based laser welding defect detection system, characterized by, The system comprises: a collection module: a laser welding machine data collection environment is built, and a photodiode, a temperature sensor and a prism are arranged to detect light intensity data and temperature data; data is collected and the collected data is continuously put into a buffer queue; Specifically: collecting plasma intensity ItP t , light intensity information LP t , temperature TmP t The three types of parameters are sent to the data acquisition terminal, which forms a t-time three-tuple parameter by collecting the parameters at t time and puts the t-time three-tuple parameter into a buffer queue. a parameter construction module: input parameters are constructed based on the collected data; a matrix formed by N three-tuple parameters obtained at continuous N time points is taken as an input parameter and is put into a sample set; specifically: Step S31: three-tuple parameters with a length of N starting from a head pointer are determined as target input parameters; wherein the head pointer points to the head of the buffer queue; Step S32: Obtain the length-N triplets starting from the active pointer as candidate input parameters; wherein the active pointer is used to determine the next target input parameter, that is, the next position of the head pointer; Step S33: Determine the continuous change of the same type parameters between the candidate input parameters and the target input parameters and the relationship change between different type parameters; if both the continuous change and the relationship change meet the preset conditions, proceed to step S34; otherwise, proceed to step S35; Step S34: Determine whether the distance between the active pointer and the head pointer is less than N, if yes, increment the active pointer, and return to step S32; if no, proceed to the next step; Step S35: Take the target input parameters at the position of the length-N triplets starting from the head pointer as input parameters and put them into the sample set; delete the triplet data between the head pointer and the position before the active pointer, reset the position of the head pointer to the position of the active pointer, and increment the position of the active pointer; if the sample set has been constructed, proceed to step S4, otherwise return to step S31; The model building module: constructs a neural network model, and trains the neural network model using the sample set until the training target is reached; The real-time acquisition module: receives the preprocessed laser welding data to form triplets, and constructs an N-long triplet parameter matrix as input parameters based on the continuous N triplets; The detection module: inputs the input parameters into the neural network model to obtain the detection result.

6. The neural network-based laser weld defect detection system of claim 5, wherein, The acquisition module includes two different photodiode sensors, a thermometer sensor, and a 45-degree prism.

7. A processor, comprising: The processor is configured to execute a program, and the program, when executed, performs the neural network-based laser welding defect detection method of any one of claims 1-4.

8. A computer-readable storage medium, characterized in that, A program that, when executed on a computer, causes the computer to perform the neural network-based laser welding defect detection method of any one of claims 1-4.

9. An execution device, characterized by A processor coupled to a memory, the memory storing program instructions that, when executed by the processor, implement the neural network-based laser welding defect detection method of any one of claims 1-4.

Images