Method and device for quality detection of laser cutting and storage medium

Abstract

The application discloses a laser cutting quality detection method and device and a storage medium, and relates to the technical field of artificial intelligence manufacturing. The application constructs a training data set based on historical data of a laser cutting process of the same type of target workpiece, iteratively optimizes the initial quality detection model super parameter in combination with a swarm intelligence optimization algorithm, makes the model accurately adapt to the laser cutting quality detection requirements of the type of workpiece, and effectively reduces the false alarm and missed alarm rates. When an abnormal state is detected, the deviation characteristics of real-time process parameters and preset standard values are matched with a root cause knowledge base, abnormal root cause descriptions and treatment measures are quickly output, the time consumption of abnormal troubleshooting is reduced, and the production efficiency is improved. Moreover, in response to newly generated real-time data, the training data set is continuously updated and the model is iteratively optimized, so that the quality detection model can continuously evolve with the cutting process data increment accumulation of the same type of target workpiece, adapt to dynamic scenes such as equipment degradation and incoming material fluctuation, and long-term maintain stable detection performance.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence manufacturing technology, specifically to a quality inspection method, device, and storage medium for laser cutting. Background Technology

[0002] With the development of technologies such as intelligent sensors, information technology, big data, and artificial intelligence, a large amount of data has been accumulated in the field of laser cutting in automotive equipment manufacturing. However, most of this data is still collected through manual visual inspection. Although a few papers have developed data-driven and knowledge-driven algorithms, these algorithms undergo manual data cleaning and fusion, manual algorithm parameter tuning, and manual algorithm training and evaluation before being used to inspect the quality of laser-cut workpieces. As the cutting process data, quality inspection defect data, and cutting quality anomaly investigation data accumulate incrementally, the algorithms become increasingly prone to false alarms and false negatives due to factors such as gradual equipment degradation and fluctuations in incoming workpiece materials. Furthermore, retraining the algorithm based on the incrementally accumulated data is extremely time-consuming. Summary of the Invention

[0003] The purpose of this application is to provide a laser cutting quality inspection method, device, and storage medium to solve the problems of low accuracy and long inspection time in traditional laser cutting in the manufacturing industry.

[0004] To achieve the above objectives, the first aspect of this application provides a method for quality inspection of laser cutting, comprising: Historical data of laser cutting processes for the same type of target workpiece are acquired, and a training dataset is constructed based on the historical data, which includes historical time-series data, workpiece quality inspection result data, and workpiece material data. Based on the training dataset, the hyperparameters of the initial quality detection model are iteratively optimized using a swarm intelligence optimization algorithm to determine the target hyperparameter combination, and the initial quality detection model is updated based on the target hyperparameter combination to obtain the current quality detection model. The real-time process parameter data of the laser cutting process of the target workpiece is obtained and input into the current quality inspection model to obtain the real-time quality inspection result of the target workpiece. If the real-time quality detection result is abnormal, then based on the deviation characteristics between the real-time process parameter data and the preset standard value, the abnormal root cause description and handling measures corresponding to the real-time quality detection result are matched from the pre-constructed root cause knowledge base. In response to newly generated real-time data, the real-time data is incorporated into the historical data as new data to reconstruct the training dataset, and the iterative optimization and model update steps are repeatedly performed to update the current quality inspection model for use in detecting real-time process parameter data collected from subsequent target workpieces of the same type.

[0005] A second aspect of this application provides a quality inspection device for laser cutting, comprising: The acquisition module is used to acquire historical data of the laser cutting process of the same type of target workpiece, and to construct a training dataset based on the historical data. The historical data includes historical time series data, workpiece quality inspection result data, and workpiece material data. The training module is used to iteratively optimize the hyperparameters of the initial quality detection model based on the training dataset using a swarm intelligence optimization algorithm, determine the target hyperparameter combination, and update the initial quality detection model based on the target hyperparameter combination to obtain the current quality detection model. The detection module is used to acquire real-time process parameter data of the laser cutting process of the target workpiece, input it into the current quality detection model, and obtain the real-time quality detection result of the target workpiece. The output module is used to, if the real-time quality detection result is abnormal, match and output the abnormal root cause description and handling measures corresponding to the real-time quality detection result from the pre-built root cause knowledge base based on the deviation characteristics between the real-time process parameter data and the preset standard value. The update module is used to respond to newly generated real-time data, integrate the real-time data as new data into the historical data to reconstruct the training dataset, and repeatedly perform iterative optimization and model update steps to update the current quality inspection model for use in detecting real-time process parameter data collected from subsequent target workpieces of the same type.

[0006] A third aspect of this application provides a computer-readable storage medium storing a program that can be loaded by a processor and executed using the laser cutting quality inspection method described above.

[0007] The beneficial effects of this application are: This application constructs a training dataset based on historical data of laser cutting processes for the same type of target workpiece. It then uses a swarm intelligence optimization algorithm to iteratively optimize the hyperparameters of the initial quality inspection model, enabling the model to accurately adapt to the laser cutting quality inspection requirements of this type of workpiece and effectively reduce false alarm and false negative rates. When an abnormal state is detected, the model quickly outputs a description of the root cause and handling measures by matching the deviation features between real-time process parameters and preset standard values ​​to a root cause knowledge base, reducing the time spent on anomaly investigation and improving production efficiency. Furthermore, in response to newly generated real-time data, the training dataset is continuously updated and the model is iteratively optimized, allowing the quality inspection model to continuously evolve with the accumulation of incremental data from the cutting process of the same type of target workpiece, adapting to dynamic scenarios such as equipment degradation and material fluctuations, and maintaining stable detection performance over the long term.

[0008] Other features and advantages of this application will be described in detail in the following detailed description section. Attached Figure Description

[0009] Figure 1 This is a flowchart illustrating a laser cutting quality inspection method provided in an embodiment of this application. Figure 2 This is a flowchart illustrating a laser cutting quality inspection method provided in a specific embodiment of this application; Figure 3 This is a schematic diagram of the structure of a laser cutting quality inspection device provided in the embodiments of this application. Detailed Implementation

[0010] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0011] In the description of this application, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "a plurality of" means two or more, unless otherwise explicitly specified. Details are set forth in the following description for illustrative purposes. It should be understood that those skilled in the art will recognize that this application can be implemented without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid unnecessarily obscuring the description of this application. Therefore, this application is not intended to be limited to the embodiments shown, but rather to be consistent with the broadest scope of the principles and features disclosed herein.

[0012] Figure 1 This is a schematic flowchart of a laser cutting quality inspection method provided in an embodiment of this application. Figure 1 As shown, this quality inspection method may include steps 101-105, which will be described in detail below.

[0013] Step 101: Obtain historical data on the laser cutting process of the same type of target workpiece, and construct a training dataset based on the historical data. The historical data may include historical time-series data, workpiece quality inspection results data, and workpiece material data.

[0014] The same type of target workpiece refers to a batch of workpieces with the same material properties (such as material, model, thickness) and the same cutting process requirements (such as cutting accuracy, edge quality standards) (such as Class A 4~6mm cast aluminum parts, Class B 10mm stainless steel parts, etc.), to ensure the relevance of the data and the adaptability of the model.

[0015] Taking a target workpiece of type A as an example, the historical data of the laser cutting process of the target workpiece of type A is collected first. The historical data can include three types of data accumulated in the past laser cutting processes of the same type of target workpiece: historical time-series data (such as continuous data on the changes of laser power, protective lens temperature, cutting gas pressure, tool center point speed (TCP) and other speeds over time), workpiece quality inspection result data (such as discrete inspection labels such as normal, incomplete cut, slag and other indicators), and workpiece material data (such as basic attribute data such as material type, model, and thickness).

[0016] Then, the collected data undergoes preprocessing, such as removing missing values ​​and eliminating abnormal noise data. Based on the segmented time intervals, the historical time-series data is precisely aligned with the quality inspection results data to ensure that each set of time-series data corresponds to a clear quality label, ultimately forming the initial training dataset. The training dataset is a collection of basic data for model training, formed by cleaning and aligning historical data of the same type of target workpieces. It includes data on normal workpieces and faulty workpieces and is continuously updated as new data accumulates.

[0017] Similarly, if it is a new type of target workpiece (such as type B or type C target workpiece), then a dedicated training dataset is constructed according to the above logic to avoid cross-type data interference.

[0018] Step 102: Based on the training dataset, use the swarm intelligence optimization algorithm to iteratively optimize the hyperparameters of the initial quality detection model, determine the target hyperparameter combination, and update the initial quality detection model based on the target hyperparameter combination to obtain the current quality detection model.

[0019] The initial quality inspection model is an algorithmic model (such as a gradient boosting tree model) used to determine the quality status of laser-cut workpieces. Its hyperparameters (such as tree depth, learning rate, etc.) need to be optimized to adapt to the inspection of specific types of workpieces. First, determine the backbone algorithm of the initial quality model, such as Lightweight Gradient Boosting Machine (LightGBM), gradient boosting tree, etc., and clarify the core hyperparameters that need to be optimized (such as the maximum depth of the tree, learning rate, subsample ratio, etc.).

[0020] Then, a swarm intelligence optimization algorithm is initiated, aiming to optimize the detection evaluation metrics of the training dataset. It iteratively optimizes the core hyperparameters to determine the target hyperparameter combination. The swarm intelligence optimization algorithm (such as the Pathfinder optimization algorithm) is used to automatically search for the optimal hyperparameter combination of the initial quality inspection model, improving the model's detection accuracy through iterative optimization. The target hyperparameter combination is the set of hyperparameters that optimizes the model's evaluation metrics after iterative optimization. Updating the initial quality inspection model using the target hyperparameter combination yields a current quality inspection model that fully adapts to the current type of target workpiece, while simultaneously determining the model's detection control limits (threshold standards used to judge normal / abnormal quality).

[0021] Step 103: Obtain real-time process parameter data of the laser cutting process of the target workpiece, input it into the current quality inspection model, and obtain the real-time quality inspection results of the target workpiece.

[0022] The system collects real-time process parameter data (such as real-time laser power, protective mirror temperature, and other time-series data) during the laser cutting process of the target workpiece of the current type. This real-time process parameter data is input into the current quality inspection model obtained in step 102. The model quickly judges based on preset detection control limits and outputs real-time quality inspection results. Real-time quality inspection results can include at least one of the following: normal, incomplete cutting, or slag. For example, the number 0 can represent a normal real-time quality inspection result, 1 represents an incomplete cutting result, and 2 represents a slag result. This approach allows the quality inspection results to better reflect the actual defect types in laser cutting, improving the targeting and accuracy of detection. Furthermore, it facilitates accurate matching with the root cause knowledge base, enabling different root causes and treatment measures to correspond to different anomaly types, thus improving anomaly location efficiency and processing reliability.

[0023] Step 104: If the real-time quality inspection result is abnormal, then based on the deviation characteristics between the real-time process parameter data and the preset standard value, match and output the abnormal root cause description and handling measures corresponding to the real-time quality inspection result from the pre-constructed root cause knowledge base.

[0024] The root cause knowledge base is a pre-built structured database linking "process parameter deviation - anomaly type - root cause description - handling measures" for rapid source tracing and resolution of abnormal states. If the real-time quality inspection result indicates an abnormal state, deviation features exceeding preset standard values ​​are extracted from the real-time process parameter data (e.g., protective lens temperature higher than the standard value, cutting gas pressure lower than the standard value, etc.). Based on these deviation features, the pre-built root cause knowledge base is searched to match the corresponding abnormal root cause description (e.g., the root cause of a high protective lens temperature is dirt on the protective lens) and specific handling measures (e.g., disassembling and inspecting the protective lens and observing its cleanliness with a 10x magnifying glass), and this information is fed back to the production site in real time.

[0025] Step 105: In response to the newly generated real-time data, the real-time data is integrated into the historical data as new data to reconstruct the training dataset, and the iterative optimization and model update steps are repeated to update the current quality inspection model for use in inspecting the real-time process parameter data collected on subsequent target workpieces of the same type.

[0026] If the cutting process of the current type of target workpiece generates new real-time data (including newly added normal workpiece data and newly added faulty workpiece data), it is merged into the original historical data as new data to reconstruct the updated training dataset (for example, the original historical data of type A target workpiece includes 100 normal workpiece data and 5 faulty workpiece data, and 200 new normal workpiece data and 10 faulty workpiece data are added, and the updated training set contains 300 normal workpiece data and 15 faulty workpiece data).

[0027] Next, based on the updated training dataset, steps 102 (swarm intelligence optimization algorithm hyperparameter optimization), 103 (online detection), and 104 (anomaly handling) are repeated to achieve automatic updating of the hyperparameters and detection control limits of the current quality detection model.

[0028] As data continues to accumulate (for example, if 1000 normal workpiece data and 20 faulty workpiece data are subsequently added to the target workpiece type A, the updated training set will contain 1300 normal workpiece data and 35 faulty workpiece data), the above iterative process is repeated to continuously optimize the model. When the training dataset has covered all possible data scenarios for the current type of workpiece (such as normal operating conditions and various fault conditions), and the model's detection accuracy is stable and the false positive and false negative rates are reduced to a minimum, the iteration is stopped. At this point, the model and detection control limits are in their optimal state and can be used for subsequent routine detection of this type of workpiece.

[0029] In addition, for other types of target workpieces, such as type B, type C, and type D, steps 101 to 105 are repeated for each type to construct a dedicated training dataset, optimize the model, and establish a detection system for each type of target workpiece, thereby achieving accurate detection of multiple types of workpieces.

[0030] This application's embodiments construct a training dataset based on historical data of laser cutting processes for the same type of target workpiece. A swarm intelligence optimization algorithm is used to iteratively optimize the hyperparameters of the initial quality inspection model, enabling the model to accurately adapt to the laser cutting quality inspection requirements of this type of workpiece and effectively reduce false alarm and false negative rates. When an abnormal state is detected, the root cause description and handling measures are quickly output by matching the deviation features between real-time process parameters and preset standard values ​​to a root cause knowledge base, reducing the time spent on anomaly investigation and improving production efficiency. Furthermore, in response to newly generated real-time data, the training dataset is continuously updated and the model is iteratively optimized, allowing the quality inspection model to continuously evolve with the accumulation of incremental data from the cutting process of the same type of target workpiece, adapting to dynamic scenarios such as equipment degradation and incoming material fluctuations, and maintaining stable detection performance over the long term.

[0031] In step 101, workpiece material data, historical time-series data of the laser cutting process, and workpiece quality inspection results data within a historical preset time range for the same type of target workpiece can be extracted first. Historical time-series data refers to process parameters continuously collected over time during the laser cutting process, including but not limited to laser power, protective mirror temperature, cutting gas pressure, and TCP speed. Workpiece quality inspection results data can include the start and end times of cutting for each workpiece. The start and end times of cutting refer to the start and end times of completing one laser cut for a single workpiece, used to determine the time window for data extraction. The time window is a time period bounded by the start and end times of cutting for a single workpiece, used to accurately extract the cutting process data corresponding to that workpiece from continuous time-series data. For example, extracting from the data storage module... The data includes workpiece material data within a given time frame, real-time historical data of the laser cutting process, and workpiece quality inspection results after laser cutting. This refers to the latest result of the workpiece quality inspection after laser cutting. As an example, the workpiece quality inspection result data after laser cutting mainly includes the workpiece number, workpiece type, start and end times of cutting the workpiece, and the inspection result (including three types: normal, incomplete cut, and slag, labeled 0, 1, and 2 respectively). Workpiece material data mainly includes workpiece type, material, and model. Historical time-series data of the laser cutting process mainly includes laser power, protective mirror temperature, high back voltage, high voltage, cutting gas pressure, and TCP speed.

[0032] Then, the extracted historical time-series data, workpiece quality inspection result data, and workpiece material data are cleaned to remove missing values ​​and ensure data validity. For each target workpiece cutting task, a time window is determined based on the start and end times of the cutting. The process parameter sequence within the time window is extracted from the historical time-series data to form a time-series data vector. The time-series data vector is a vector formed by combining multiple process parameters within the time window in chronological order, and serves as the input feature of the quality inspection model. Quality labels are classification labels marked according to the workpiece quality inspection results, used to indicate whether the cutting quality is normal, whether there are incomplete cuts, slag, or other abnormalities. The time-series data vector is then associated and mapped with the corresponding workpiece material data and the quality labels determined based on the quality inspection result data to construct a single training sample. Finally, the single training samples corresponding to all target workpieces are aggregated to form a training dataset. The training dataset can include a time-series dataset composed of time-series data vectors and a quality label dataset composed of quality labels.

[0033] For example, based on the quality inspection data of the workpiece after laser cutting, in The time period was used to cut the cast aluminum parts The quality inspection result was that it was not cut through (i.e., the serial number was marked as...). ),according to Historical data of the laser cutting process for the corresponding time period can be extracted, which is a time-series data vector. Similarly, in The time period also saw the cutting of the cast aluminum parts. The quality inspection result was slag (i.e., the quality label was...). ),according to Historical data of the laser cutting process for the corresponding time period can be extracted, which is a time-series data vector. And so on, in The time period also saw the cutting of the cast aluminum parts. A time-series data vector can be obtained as follows: The quality inspection result was that it was not cut through (i.e., the quality label showed...). ); and thus we can conclude that, During the time period, for cast aluminum parts The time series dataset can be obtained. Quality label dataset , .

[0034] By constructing a time window based on the start and end times of cutting, strict alignment is achieved between a single workpiece, a cutting process, and a quality result, reducing data contamination across different workpieces and time periods and significantly improving sample reliability. Cleaning and removing missing values ​​and outliers reduces noise interference, allowing the subsequent model to learn the true process rules and improving detection accuracy. Dedicated modeling for the same type of workpiece results in higher detection accuracy. Using only data from the same type of workpiece to build the training set allows the model to deeply adapt to the cutting characteristics of that type of workpiece, reducing generalization bias caused by data from multiple workpiece types. Datasets are built using individual samples as units; new workpiece data can be directly appended without reconstructing the entire dataset, providing a foundation for subsequent automatic model updates and parameter optimization. The time-series data vectors retain complete trends in process parameter changes, useful not only for quality inspection but also for subsequent root cause analysis and troubleshooting.

[0035] In step 102, an initial quality detection model based on the Lightweight Gradient Boosting Machine (LTM) algorithm can be constructed. The time-series dataset from the training dataset is used as input, and the quality label dataset is used as output, establishing a mapping relationship between laser cutting process parameters and workpiece quality status. The Lightweight Gradient Boosting Machine (LTM) algorithm is a highly efficient gradient boosting decision tree algorithm suitable for classification and recognition tasks in industrial scenarios. It features fast training speed, high accuracy, and low resource consumption.

[0036] First, initialize the set of hyperparameters to be optimized. The set of hyperparameters is a combination of key parameters that affects the model structure and training performance and needs to be determined through optimization. These parameters include tree depth, learning rate, subsample ratio, and minimum leaf node sample weights. The set of hyperparameters should at least include the maximum tree depth, learning rate, subsample ratio of the entire sample set, and minimum leaf node sample weights in the lightweight gradient boosting machine algorithm.

[0037] Then, in each iteration using the swarm intelligence optimization algorithm, the initial quality detection model is trained based on the currently generated hyperparameter combination, and an evaluation index value is calculated. The evaluation index value is the area under the receiver operating characteristic (ROC) curve. The ROC curve is a core indicator for evaluating the performance of the classification model; a larger value indicates a stronger ability to distinguish between normal and abnormal samples. After multiple iterations, the hyperparameter combination corresponding to the maximum evaluation index value is selected as the target hyperparameter combination.

[0038] Finally, the initial quality inspection model is updated based on the target hyperparameter combination to obtain the current quality inspection model, and the target hyperparameter combination, corresponding evaluation index values, and training date are stored. The current quality inspection model is the final online inspection model suitable for the current type of workpiece after being updated using the target hyperparameter combination.

[0039] For example, with Data within a time period is the object, in the form of a time series dataset. As input, the quality label dataset For output, an online quality inspection algorithm for laser-cut workpieces is constructed based on LightGBM. The four main parameters of the LightGBM algorithm are initialized, including the maximum depth of the tree. Learning rate The proportion of subsamples to the entire sample set Minimum leaf node sample weights The algorithm evaluation metrics are ;go through Round training will evaluate indicators The set of parameters corresponding to the maximum value is used as the final parameter combination. Date of this training Evaluation index values They are stored together in the online quality detection algorithm parameter storage module.

[0040] A dedicated quality inspection model is built based on a lightweight gradient booster algorithm, using the area under the receiver operating characteristic (ROC) curve as the optimization objective. This significantly improves the model's ability to classify cut quality as normal or abnormal. A swarm intelligence optimization algorithm automatically searches and selects hyperparameters, replacing manual periodic parameter tuning and greatly reducing deployment and maintenance costs. Using the ROC curve as a unified evaluation metric, the optimal model can be clearly quantified. Hyperparameters, evaluation metrics, and training dates are stored, enabling model version traceability and comparison. The lightweight gradient booster algorithm offers fast training speed and short inference time, making it suitable for online real-time laser cutting inspection scenarios without consuming excessive computing resources. The model and optimal hyperparameters can be continuously updated and overwritten, supporting subsequent data-triggered re-optimization, allowing the model to continuously evolve with data accumulation.

[0041] In step 104, the real-time process parameter sequence within the current cutting time window can be obtained. The cutting time window is a time period bounded by the actual start and end times of the current workpiece cutting, used to accurately extract the process parameter changes corresponding to this cutting. The real-time process parameter sequence consists of process parameters continuously acquired over time within the cutting time window, and may include at least high back voltage, laser power, focal position, and auxiliary gas pressure.

[0042] Then, the real-time process parameter sequence is compared with the preset standard threshold range of the corresponding process parameters to calculate a deviation feature vector containing the deviation direction and magnitude. Statistical feature values ​​reflect the overall trend of process parameters during cutting, such as mean, peak, trough, and fluctuation amplitude. The preset standard threshold range is the allowable range of process parameters during normal operation, preset by process specifications or equipment requirements. The deviation feature vector, composed of the deviation direction (higher / lower) and magnitude of each process parameter relative to the standard threshold, is used to characterize abnormal patterns.

[0043] Next, the pre-built root cause knowledge base for laser cutting quality anomalies is accessed. This root cause knowledge base is a pre-established structured rule library that stores the corresponding relationships between various laser cutting anomalies, parameter deviations, fault causes, and handling measures. The root cause knowledge base stores multiple root cause rules, each a complete matching entry. Each root cause rule includes an anomaly type identifier, a key parameter deviation pattern, a corresponding anomaly root cause description, and handling measures. The anomaly type identifier is used to distinguish different quality anomaly categories, such as incomplete cutting or slag formation. The anomaly type identifier and deviation feature vector determined by the real-time quality detection results are matched against the root cause rules in the root cause knowledge base. If a target root cause rule matches the real-time quality detection results, the anomaly root cause description and handling measures are extracted from the target root cause rule. The target root cause rule is the root cause rule that successfully matches the current anomaly type and deviation feature vector.

[0044] For example, the trained LightGBM algorithm can be retrieved from the online quality detection algorithm parameter storage module with the highest evaluation index. Corresponding algorithm parameter combinations Therefore, as an online application module for quality inspection of laser-cut workpieces, it is used for cast aluminum parts. Online quality inspection is performed. First, the cut aluminum castings are extracted from the real-time data storage module of the laser cutting process. Real-time data is used to detect the cutting quality through the online quality detection algorithm parameter storage module. If the detection result is normal, it is directly returned to the laser-cut workpiece quality detection result storage module. If the detection result is incomplete cutting or slag, the quality anomaly root cause localization and recommendation module is called. This module extracts root cause knowledge and recommended measures from the laser cutting quality anomaly root cause knowledge storage module. Finally, the detection result, quality anomaly root cause knowledge, and recommended measures are returned to the laser-cut workpiece quality detection result storage module.

[0045] By quantifying the direction and magnitude of process parameter deviations from standards using deviation feature vectors, and combining this with rule matching based on anomaly types, misjudgments caused by relying solely on experience are avoided. Root causes are automatically matched directly based on real-time parameter deviations and a knowledge base, eliminating the need for manual analysis and significantly reducing fault location time. The output includes clear descriptions of the anomaly's root cause and actionable solutions, allowing on-site personnel to implement them directly and reducing reliance on professional experience. The correspondence between anomaly patterns and root causes is stored in a rule-based and structured manner, facilitating subsequent traceability, process optimization, and equipment maintenance. Linking detection results with root cause analysis achieves a complete industrial closed loop of "anomaly detection - cause location - solution provision," improving production stability.

[0046] To make the root cause localization of laser cutting quality anomalies more accurate, comprehensive, and adaptable to different data sources, embodiments of this application provide a root cause knowledge base that can be built based on text semantics to form a standardized, automatically matching root cause knowledge system, providing support for real-time anomaly diagnosis.

[0047] In one example, firstly, historical maintenance records and fault reports containing unstructured text can be obtained as raw fault data. Unstructured text refers to natural language text without a fixed format, such as maintenance records, fault reports, and log descriptions. Then, the raw fault data is cleaned to remove redundant fields and noisy data. The cleaned data is then standardized to convert fault data from different sources into a unified data format. Standardization involves unifying text data with different formats and field names into a processable, standardized format.

[0048] Next, natural language processing (NLP) techniques are used to perform entity recognition and relation extraction on the unstructured text in the standardized fault data. Entity recognition automatically identifies key entities such as fault phenomena, fault codes, fault locations, and handling methods from the text. Relationship extraction identifies the logical relationships between entities, such as "fault code - causing - fault phenomenon" or "fault phenomenon - corresponding - solution". Based on this, the associations between fault phenomenon entities, fault code entities, and solution entities are extracted.

[0049] The processed structured data and the associated relationships are stored in a pre-defined graph database to generate a root cause knowledge graph, which serves as the core data model of the root cause knowledge base. A graph database stores knowledge in the form of nodes, relationships, and attributes, facilitating association retrieval and path reasoning. A root cause knowledge graph is a knowledge model that organizes the relationships between faults, root causes, parameters, and measures in a graph format.

[0050] In the anomaly diagnosis in step 104, the target fault code entity can be determined based on real-time process parameter data. The target fault code entity is a coded entity used to uniquely identify a certain type of quality anomaly. Then, solution entities with associated paths to the target fault code entity are retrieved from the root cause knowledge graph. The solution entities are the handling, maintenance, and adjustment measures entities corresponding to the fault code / fault phenomenon. The content corresponding to the retrieved solution entities is then used as the handling measures.

[0051] Generating knowledge directly from on-site text-based fault data can significantly reduce knowledge organization costs. Knowledge graphs enable correlated reasoning about faults, root causes, and corrective actions, supporting chain-like localization of complex anomalies. Fuzzy matching and relational retrieval are supported, providing greater adaptability to faults with non-standard descriptions and multiple causes. Furthermore, it can be used in conjunction with numerical rule bases to improve overall root cause diagnosis coverage and robustness.

[0052] In step 105, the newly generated real-time quality inspection results and corresponding real-time process parameter data can be combined as new data and integrated into the historical data. It is then determined whether the real-time quality inspection results are in a normal state. If the real-time quality inspection results are in a normal state, the real-time data is marked as a high-confidence positive sample and stored in the sample buffer. High-confidence positive samples refer to cut samples that are judged to be normal, complete, and without obvious abnormalities by online detection. These can be safely used for incremental learning to enhance the model's learning ability under normal operating conditions. The sample buffer is a storage area used to temporarily cache newly added high-confidence positive samples, accumulating a certain amount of data before uniformly participating in model retraining to avoid frequent training.

[0053] The system continuously monitors the accumulated data in the sample buffer. When the accumulated data reaches a preset incremental update threshold, a model retraining command is triggered. The incremental update threshold is a pre-set threshold for the number of samples. When the amount of data in the buffer reaches this threshold, the model is incrementally updated, balancing real-time performance and computational efficiency.

[0054] Model retraining, based on a dataset merged with newly added samples and historical data, re-executes hyperparameter optimization and model training to achieve self-learning iteration of the detection model. In response to the model retraining command, high-confidence positive samples, historical data, and newly added abnormal data in the sample buffer are merged to construct an updated training dataset. The process of iteratively optimizing the hyperparameters of the initial quality detection model using a swarm intelligence optimization algorithm is then re-executed to generate a new generation of the current quality detection model. This new generation of the current quality detection model is an updated version obtained through incremental data retraining and hyperparameter re-optimization, possessing stronger adaptability to various operating conditions and improved detection stability.

[0055] For example, as abnormal results accumulate in the data storage module, on dates such as... At that time, the algorithm self-learning module is invoked, and a swarm intelligence optimization algorithm is used to automatically optimize the four parameters of the LightGBM algorithm; the swarm intelligence optimization algorithm adopts the Pathfinder optimization algorithm, and the population size of the Pathfinder optimization algorithm is initialized first. The maximum number of iterations of the algorithm is Then, the algorithm performs self-learning. When the iteration termination condition is met (i.e., the maximum number of iterations is reached), the maximum evaluation metric obtained in this training is used. A corresponding set of parameter combinations Date of this training These parameters are then stored in the online quality detection algorithm parameter storage module. The above steps are executed periodically to continuously update the algorithm parameters and improve the algorithm's detection performance. , , , These are adjustable parameters.

[0056] By automatically accumulating normal samples and triggering retraining, model iteration can be completed without manual intervention, truly achieving self-learning and self-optimization of the detection algorithm. Employing a sample buffer and incremental update threshold mechanism, retraining is only triggered after data accumulation reaches a certain scale, reducing unnecessary computational overhead and not affecting the real-time performance of online detection. As high-confidence positive samples are continuously added, the model learns more thoroughly from normal operating conditions, resulting in more accurate detection boundaries and a sustained reduction in false positive and false negative rates. With continuous accumulation of production data, the model can adapt to operating condition drifts such as equipment aging, material fluctuations, and environmental changes, maintaining high-precision detection over the long term. Furthermore, the same incremental update logic can be used for different types of target workpieces (A, B, C, D), maintaining their own sample buffers and iteration cycles, making it easy to expand and highly versatile.

[0057] Figure 2 This is a flowchart illustrating a laser cutting quality inspection method provided in a specific embodiment of this application. In one specific embodiment, data preparation is first performed. Workpiece material information, historical laser cutting process data, and quality inspection result data for the same type of target workpiece are collected. The above data is then aligned (correlated by time dimension) and cleaned (removing missing and outlier values). The processed data serves as the base dataset for subsequent training.

[0058] Then, perform algorithm initialization. Complete the basic initialization of the quality detection algorithm (such as the lightweight gradient booster algorithm), and initialize the relevant parameters of the swarm intelligence optimization algorithm to prepare for subsequent automatic optimization.

[0059] Next, model training and evaluation are performed. Based on the processed base dataset, a quality detection algorithm is used for model training. Simultaneously, a swarm intelligence optimization algorithm is used to iteratively search and optimize the model's hyperparameters. After training is complete, the quality detection result evaluation index of the model is calculated.

[0060] Finally, the termination condition is determined. It is checked whether the current evaluation metric meets the preset termination condition (e.g., the metric accuracy reaches the requirement or the iteration count is exhausted). If not, the algorithm parameters are updated based on the feedback from the swarm intelligence optimization algorithm, and the training of the quality detection algorithm is resumed for retraining and evaluation. If the condition is met, the final output stage is entered, and the latest algorithm parameter combination, training date, and corresponding quality detection result evaluation metric are output, completing the self-learning training of this model. The updated parameters can then be applied to online detection operations.

[0061] This embodiment achieves automated iterative updates of the quality inspection model parameters through a closed-loop process of data alignment and cleaning, algorithm initialization, round-training optimization, and conditional judgment. Compared with traditional fixed-parameter models, this method can continuously optimize model parameters based on accumulated data, thereby continuously improving detection accuracy. At the same time, it reduces the workload of manual parameter tuning, effectively adapts to dynamic factors such as equipment aging and changes in operating conditions during laser cutting, and ensures long-term stable high-precision detection.

[0062] The following provides a specific example of a self-learning-based laser cutting quality inspection and root cause analysis method, including the following steps.

[0063] Step 1: Extract cut cast aluminum parts from the data storage module for the time range of January 2025 to November 2025. The data includes material data, real-time historical data of the laser cutting process, and quality inspection results of the workpiece after laser cutting. The latest result for the quality inspection of the workpiece after laser cutting is from November 2025.

[0064] (1) Data on the quality inspection results of laser-cut workpieces from January 1, 2025 to November 20, 2025, mainly including workpiece number, workpiece type, start and end time of cutting the workpiece, and inspection results (including three types: normal, incomplete cut, and slag, marked as 0, 1, and 2 respectively), as shown in Table 1. Table 1 shows the quality inspection results of laser-cut workpieces. Sample table of quality inspection results data.

[0065] Table 1 (2) Workpiece material data from January 2025 to November 2025, mainly including workpiece type, workpiece material and model, as shown in Table 2. Table 2 is a sample table of workpiece material data after laser cutting.

[0066] Table 2 (3) Historical data of laser cutting process from January 1, 2025 to November 20, 2025 are shown in Table 3. Table 3 shows the workpieces after laser cutting. Sample table of quality inspection results data.

[0067] Table 3 Step 2: First, clean the extracted data, such as removing missing NaN values ​​from Table 3; then align the cleaned data. The alignment result is shown in Table 4, which is a sample table of data alignment.

[0068] Table 4 Therefore, within the time periods of 2025-01-01 09:00:00.000 and 2025-11-20 16:01:30.100, the quality label dataset is as follows: Time series dataset .

[0069] Step 3: Using data from January 1, 2025 to November 20, 2025 as the object, and using a time series dataset... As input, the quality label dataset For output, an online quality inspection algorithm for laser-cut workpieces is constructed based on LightGBM. The four main parameters of the LightGBM algorithm are initialized, including the maximum depth of the tree. Learning rate The proportion of subsamples to the entire sample set Minimum leaf node sample weights The evaluation index value obtained based on the initialization parameters and algorithm is .go through After multiple iterations, 10 evaluation index values ​​are obtained. Maximize the evaluation indicators ( The corresponding set of parameters () The final parameter combination will be used to combine the parameters. The training date (2025-11-20) and the evaluation index value (0.82) are stored together in the online quality detection algorithm parameter storage module.

[0070] Step 4: The trained LightGBM algorithm retrieves the parameter with the highest evaluation index from the online quality detection algorithm parameter storage module. Corresponding algorithm parameter combinations Therefore, it serves as an online application module for quality inspection of laser-cut workpieces, specifically for 4-6mm thick cast aluminum parts. Conduct online quality testing.

[0071] Step 5: First, extract the latest cut of the cast aluminum part from the real-time data storage module of the laser cutting process. Real-time data (as a time-series data vector) ), based on the online application module for quality inspection of laser-cut workpieces (where the algorithm is LightGBM, and the algorithm parameter combination is ), ) for this time series data vector If the test result is normal (i.e., 0), the result is directly returned to the laser-cut workpiece quality inspection result storage module. If the test result is abnormal (i.e., 1 or 2), the quality abnormality root cause location and recommendation are called. That is, the real-time value of each process parameter is compared with the set value. If it is higher or lower than the set value, the root cause knowledge and recommended measures are extracted from the laser cutting quality abnormality root cause knowledge storage module (as shown in Table 5, which is a sample table of laser cutting quality abnormality root cause knowledge storage). Finally, the test result, quality abnormality root cause knowledge, investigation direction and investigation measures are returned to the laser-cut workpiece quality inspection result storage module.

[0072] Table 5 Step 6: As abnormal results accumulate in the data storage module, on December 31, 2025, the algorithm self-learning module is invoked, and the Pathfinder optimization algorithm is used to automatically optimize the four parameters of the LightGBM algorithm; first, the population size of the Pathfinder optimization algorithm is initialized. The maximum number of iterations of the algorithm is Then the algorithm self-learning module is executed. The algorithm self-learning process is as follows: Figure 2 As shown, when the maximum number of iterations is reached, the maximum evaluation metric obtained in this training is obtained. A corresponding set of parameter combinations That is, [10, 0.0015, 0.65, 0.75], the date of this training. They are stored together in the online quality detection algorithm parameter storage module.

[0073] Step 7: Execute Step 6 periodically to continuously update the algorithm parameters and improve the algorithm's detection performance.

[0074] This application embodiment is a self-learning-based laser cutting quality inspection and root cause analysis method. Its advantages are: through the algorithm self-learning module, it makes full use of the gradually accumulated negative sample data (i.e., incomplete cutting and slag) to automatically optimize the parameters of the online laser cutting workpiece quality inspection algorithm and gradually improve the quality inspection effect.

[0075] Figure 3 This is a schematic diagram of the structure of a laser cutting quality inspection device provided in an embodiment of this application. Figure 3 As shown, the laser cutting quality inspection device 300 may include an acquisition module 301, a training module 302, a detection module 303, an output module 304, and an update module 305.

[0076] The acquisition module 301 is used to acquire historical data of the laser cutting process of the same type of target workpiece, and to build a training dataset based on the historical data. The historical data includes historical time series data, workpiece quality inspection result data, and workpiece material data.

[0077] The training module 302 is used to iteratively optimize the hyperparameters of the initial quality detection model based on the training dataset using a swarm intelligence optimization algorithm, determine the target hyperparameter combination, and update the initial quality detection model based on the target hyperparameter combination to obtain the current quality detection model.

[0078] The detection module 303 is used to acquire real-time process parameter data of the laser cutting process of the target workpiece, input it into the current quality detection model, and obtain the real-time quality detection result of the target workpiece.

[0079] The output module 304 is used to match and output the abnormal root cause description and handling measures corresponding to the real-time quality detection result if the real-time quality detection result is abnormal, based on the deviation characteristics between the real-time process parameter data and the preset standard value.

[0080] The update module 305 is used to respond to newly generated real-time data, integrate the real-time data as new data into the historical data to reconstruct the training dataset, and repeatedly perform iterative optimization and model update steps to update the current quality inspection model for use in detecting real-time process parameter data collected on subsequent target workpieces of the same type.

[0081] The acquisition module 301, training module 302, detection module 303, output module 304 and update module 305 can be used to execute steps 101-105 in the embodiments of the laser cutting quality detection method described above. For the specific implementation of these modules and more details, please refer to the corresponding method section, which will not be elaborated here.

[0082] This application also provides a computer-readable storage medium storing a program that can be loaded by a processor and executed by any of the laser cutting quality inspection methods described in this application.

[0083] Those skilled in the art will understand that all or part of the functions of the various methods in the above embodiments can be implemented by hardware or by computer programs. When all or part of the functions in the above embodiments are implemented by computer programs, the program can be stored in a computer-readable storage medium, which may include: read-only memory, random access memory, disk, optical disk, hard disk, etc., and the program is executed by a computer to achieve the above functions. For example, the program can be stored in the memory of a device, and when the program in the memory is executed by the processor, all or part of the above functions can be achieved. In addition, when all or part of the functions in the above embodiments are implemented by computer programs, the program can also be stored in a server, another computer, disk, optical disk, flash drive, or external hard drive, etc., and can be downloaded or copied to the memory of a local device, or the system of the local device can be updated. When the program in the memory is executed by the processor, all or part of the functions in the above embodiments can be achieved.

[0084] The above examples illustrate this application only to aid understanding and are not intended to limit its scope. Those skilled in the art to which this application pertains can make various simple deductions, modifications, or substitutions based on the ideas presented.

Claims

1. A method for quality inspection of laser cutting, characterized in that, include: Historical data of laser cutting processes for the same type of target workpiece are acquired, and a training dataset is constructed based on the historical data, which includes historical time-series data, workpiece quality inspection result data, and workpiece material data. Based on the training dataset, the hyperparameters of the initial quality detection model are iteratively optimized using a swarm intelligence optimization algorithm to determine the target hyperparameter combination, and the initial quality detection model is updated based on the target hyperparameter combination to obtain the current quality detection model. The real-time process parameter data of the laser cutting process of the target workpiece is obtained and input into the current quality inspection model to obtain the real-time quality inspection result of the target workpiece. If the real-time quality detection result is abnormal, then based on the deviation characteristics between the real-time process parameter data and the preset standard value, the abnormal root cause description and handling measures corresponding to the real-time quality detection result are matched from the pre-constructed root cause knowledge base. In response to newly generated real-time data, the real-time data is incorporated into the historical data as new data to reconstruct the training dataset, and the iterative optimization and model update steps are repeatedly performed to update the current quality inspection model for use in detecting real-time process parameter data collected from subsequent target workpieces of the same type.

2. The quality inspection method for laser cutting according to claim 1, characterized in that, The acquisition of historical data on laser cutting processes of the same type of target workpiece, and the construction of a training dataset based on the historical data, includes: Extract workpiece material data, historical time sequence data of laser cutting process, and workpiece quality inspection result data after laser cutting from the target workpiece of the same type within a historical preset time range. The workpiece quality inspection result data includes the cutting start and end time. The extracted historical time-series data, workpiece quality inspection results data, and workpiece material data are cleaned to remove missing values; For each target workpiece cutting task, a time window is determined based on the start and end times of the cutting of the target workpiece. The process parameter sequence within the time window is extracted from the historical time series data to form a time series data vector. The time series data vector is then associated and mapped with the corresponding workpiece material data and the quality label determined based on the quality inspection result data to construct a single training sample. The training dataset is formed by aggregating the individual training samples corresponding to all the target workpieces. The training dataset includes a time series dataset composed of the time series data vectors and a quality label dataset composed of the quality labels.

3. The quality inspection method for laser cutting according to claim 2, characterized in that, The process of iteratively optimizing the hyperparameters of the initial quality detection model based on the training dataset using a swarm intelligence optimization algorithm to determine a target hyperparameter combination, and updating the initial quality detection model based on the target hyperparameter combination to obtain the current quality detection model includes: An initial quality detection model based on a lightweight gradient boosting machine algorithm is constructed, with the time series dataset in the training dataset as input and the quality label dataset as output. Initialize the set of hyperparameters to be optimized, which includes at least the maximum depth of the tree, the learning rate, the proportion of subsamples to the entire sample set, and the minimum sum of leaf node sample weights in the lightweight gradient boosting machine algorithm. In each iteration using the swarm intelligence optimization algorithm, the initial quality detection model is trained based on the currently generated hyperparameter combination, and an evaluation index value is calculated, which is the area under the receiver operating characteristic curve. After multiple iterations, the hyperparameter combination corresponding to the maximum value of the evaluation index is selected as the target hyperparameter combination. The initial quality detection model is updated based on the target hyperparameter combination to obtain the current quality detection model, and the target hyperparameter combination, the corresponding evaluation index value, and the training date are stored.

4. The quality inspection method for laser cutting according to claim 1, characterized in that, If the real-time quality detection result is abnormal, then based on the deviation characteristics between the real-time process parameter data and the preset standard value, a description of the abnormal root cause and handling measures corresponding to the real-time quality detection result are matched and output from a pre-constructed root cause knowledge base, including: Obtain the real-time process parameter sequence within the current cutting time window, compare the real-time process parameter sequence with the preset standard threshold range of the corresponding process parameters, and calculate the deviation feature vector containing the deviation direction and deviation amplitude. The real-time process parameter sequence includes at least high back voltage, laser power, focal position and auxiliary gas pressure. Access a pre-built root cause knowledge base for laser cutting quality anomalies. The root cause knowledge base stores multiple root cause rules. Each root cause rule includes an anomaly type identifier, a key parameter deviation pattern, a corresponding anomaly root cause description, and a handling measure. The anomaly type identifier and deviation feature vector determined by the real-time quality detection results are matched with the root cause rules in the root cause knowledge base; If a target root cause rule exists that matches the real-time quality detection result, then the abnormal root cause description and handling measures are extracted from the target root cause rule.

5. The quality inspection method for laser cutting according to claim 1, characterized in that, It also includes the step of building a root cause knowledge base through textual semantics, which includes: Obtain historical maintenance records and fault reports containing unstructured text as raw fault data; The original fault data is cleaned to remove redundant fields and noisy data, and the cleaned data is standardized to convert fault data from different sources into a unified data format. Natural language processing technology is used to perform entity recognition and relation extraction on unstructured text in standardized fault data, and to extract the relationships between fault phenomenon entities, fault code entities and solution entities. The processed structured data and the associated relationships are stored in a preset graph database to generate a root cause knowledge graph, which serves as the core data model of the root cause knowledge base.

6. The laser cutting quality monitoring method according to claim 5, characterized in that, The step of matching and outputting abnormal root cause descriptions and handling measures corresponding to the real-time quality detection results from a pre-built root cause knowledge base includes: The target fault code entity is determined based on the real-time process parameter data; Retrieve solution entities that have an associated path with the target fault code entity from the root cause knowledge graph, and use the content corresponding to the retrieved solution entities as the processing measures.

7. The quality inspection method for laser cutting according to claim 1, characterized in that, The steps of responding to newly generated real-time data, incorporating the real-time data as new data into the historical data to reconstruct the training dataset, and repeatedly performing iterative optimization and model updates include: The newly generated real-time quality detection results and the corresponding real-time process parameter data are combined as new data and integrated into the historical data. Determine whether the real-time quality detection result is in a normal state; If the real-time quality detection result is normal, the real-time data is marked as a high-confidence positive sample and stored in the sample buffer. Monitor the accumulated amount of data in the sample buffer, and trigger a model retraining instruction when the accumulated amount of data reaches a preset incremental update threshold; In response to the model retraining instruction, the high-confidence positive samples in the sample buffer, the historical data, and the newly added abnormal data are merged to construct an updated training dataset. The step of iteratively optimizing the hyperparameters of the initial quality detection model using a swarm intelligence optimization algorithm is then re-executed to generate a new generation of the current quality detection model.

8. The laser cutting quality inspection method according to any one of claims 1 to 7, characterized in that, The real-time quality inspection results include at least one of the following: normal, incomplete cut, and slag.

9. A quality inspection device for laser cutting, characterized in that, include: The acquisition module is used to acquire historical data of the laser cutting process of the same type of target workpiece, and to construct a training dataset based on the historical data. The historical data includes historical time series data, workpiece quality inspection result data, and workpiece material data. The training module is used to iteratively optimize the hyperparameters of the initial quality detection model based on the training dataset using a swarm intelligence optimization algorithm, determine the target hyperparameter combination, and update the initial quality detection model based on the target hyperparameter combination to obtain the current quality detection model. The detection module is used to acquire real-time process parameter data of the laser cutting process of the target workpiece, input it into the current quality detection model, and obtain the real-time quality detection result of the target workpiece. The output module is used to, if the real-time quality detection result is abnormal, match and output the abnormal root cause description and handling measures corresponding to the real-time quality detection result from the pre-built root cause knowledge base based on the deviation characteristics between the real-time process parameter data and the preset standard value. The update module is used to respond to newly generated real-time data, integrate the real-time data as new data into the historical data to reconstruct the training dataset, and repeatedly perform iterative optimization and model update steps to update the current quality inspection model for use in detecting real-time process parameter data collected from subsequent target workpieces of the same type.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that can be loaded by a processor and executed as described in any one of claims 1 to 8, the laser cutting quality inspection method.

Images