Method for detecting energy decay of laser of laser cutting machine

Abstract

The application discloses a laser energy recession detection method for a laser cutting machine, and belongs to the technical field of intelligent monitoring of the state of a laser processing device. Temperature data of a laser module during laser cutting is acquired in real time, and a temperature rise rate per unit time is calculated. A slit image is acquired synchronously, and a slit stability index is extracted. The two types of features are input into a multi-modal fusion neural network model, wherein a first channel extracts a thermal load feature, and a second channel extracts a processing consistency feature. A normal, energy recession or pseudo-normal state label is output by fusion discrimination. If the pseudo-normal state is determined, the participation weight in subsequent model training is reduced, and the model is periodically retrained based on an updated sample set. The method can realize early identification of the energy state of the laser, especially the pseudo-normal state hidden by artificial compensation behavior, and has high precision, high robustness and industrial practical value.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology for the status of laser processing equipment, specifically to a method for detecting laser energy decay in laser cutting machines. Background Technology

[0002] Laser energy degradation detection for laser cutting machines involves monitoring the change in laser output power over time to determine if its energy has decreased or become unstable. This detection helps to promptly identify problems such as laser performance degradation, optical component contamination, or aging, thereby ensuring cutting quality and production efficiency, and avoiding equipment failure or material waste.

[0003] In laser cutting energy decay detection, if the model training data comes from "surface normal" conditions maintained by operator-manually adjusted parameters (such as increasing power and decreasing speed) during equipment aging, it may lead to "label contamination" problems. During the learning process, the model mistakenly treats these artificially compensated abnormal states as normal, resulting in pseudo-learning. Consequently, the model loses sensitivity to slight energy decay, has an excessively high warning tolerance, and is prone to missing early fault identification opportunities. Ultimately, equipment problems accumulate to a severe stage before being discovered, affecting cutting quality and equipment reliability. Summary of the Invention

[0004] The purpose of this invention is to provide a method for detecting laser energy decay in laser cutting machines, thereby addressing the shortcomings of the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for detecting laser energy decay in a laser cutting machine, comprising:

[0006] Real-time acquisition of temperature data of the laser module during laser cutting, and calculation of its temperature rise rate per unit time;

[0007] Simultaneously acquire kerf images during the cutting process, and extract kerf stability indicators based on kerf edge morphology, width fluctuation, or structural texture consistency;

[0008] The temperature rise rate and the kerf stability index are respectively input into a preset multimodal fusion neural network model, wherein the first channel is used to extract potential abnormal thermal load features from the temperature rise rate; and the second channel is used to extract abnormal processing consistency features from the kerf stability index.

[0009] The abnormal heat load characteristics and abnormal processing consistency characteristics are jointly nested and encoded by the fusion discriminant layer, and a discriminant label is output. The discriminant label includes three categories: normal state, energy decay state and pseudo-normal state.

[0010] If the judgment result is a pseudo-normal state, it is determined that there may be a situation where the laser energy decay is artificially compensated and masked.

[0011] For samples identified as pseudo-normal states, their participation weight in subsequent iterative training of the multimodal fusion neural network model is reduced, and the multimodal fusion neural network model is periodically retrained based on the updated training sample set.

[0012] Preferably, the calculation of the temperature rise rate includes:

[0013] Temperature values ​​at consecutive time points are collected using a temperature sensor located inside the laser module near the gain medium at a sampling frequency of not less than 1 Hz. At any two consecutive time points, the difference between the current temperature and the previous temperature is calculated and divided by the time interval between the two time points to obtain the rate of temperature change per unit time, which is used as the temperature rise rate.

[0014] Preferably, the temperature rise rate is smoothed using a sliding window before being input into the neural network, including:

[0015] A sliding window is formed by combining multiple consecutive temperature rise rate samples;

[0016] The arithmetic average or weighted average of all temperature rise rates within the window is calculated as the smoothed temperature rise rate feature for the current time period; the smoothed feature is used as the time series data input to the first channel to extract potential abnormal heat load features.

[0017] Preferably, the extraction of kerf stability indicators includes: acquiring kerf images of the cutting area using an industrial camera; performing grayscale conversion, filtering, and edge enhancement processing on the images; extracting the kerf edges using an edge detection algorithm and performing morphological restoration on the edge data; calculating three sub-indices: kerf width fluctuation, edge smoothness index, and structural texture consistency score; and normalizing the three sub-indices to form a kerf stability vector as the second channel input.

[0018] The preferred method for obtaining the kerf width fluctuation is as follows:

[0019] Multiple cross sections were sampled at equal intervals along the cut path;

[0020] Measure the distance between the left and right edges of each cross section to obtain the kerf width sequence;

[0021] Calculate the standard deviation of the kerf width sequence as the kerf width variability.

[0022] Preferably, the participation weight of samples judged as pseudo-normal states in subsequent iterative training of multimodal fusion neural network models is reduced, including:

[0023] Based on the classification confidence score output by the model, a weight coefficient is assigned to each pseudo-normal sample;

[0024] If the maximum classification probability of a pseudo-normal sample is close to 1 / 3, its confidence level is low.

[0025] The exponential decay function is used to compress the weights of low-confidence samples to below 0.1.

[0026] Preferably, the periodic retraining of the multimodal fusion neural network model based on the updated training sample set includes:

[0027] Set a fixed training period. After each period, construct a training dataset from three types of samples: normal, decaying, and pseudo-normal. Input each sample and its corresponding weight into the model for weighted training. The updated model is used for energy decay discrimination in the next period.

[0028] Preferably, the loss function used in the training process is the weighted cross-entropy loss function, calculated as follows:

[0029] Multiply the prediction loss of each training sample by its corresponding participation weight;

[0030] The sum of all weighted losses divided by the number of samples yields the overall loss value.

[0031] Use this loss value to update the model weights.

[0032] Preferably, pseudo-normal samples are recorded in a pseudo-sample pool, which stores all sample data marked as pseudo-normal by the model, confidence scores, and label information; the pseudo-sample weights are redistributed from the sample pool before each round of training to adapt to changes in operating conditions.

[0033] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0034] 1. This invention integrates key features from two complementary dimensions—the laser module's temperature rise rate and the stability of the kerf image—to construct a multimodal fusion neural network model. This model enables intelligent identification of the laser's energy state during laser cutting, exhibiting a significant advantage in identifying pseudo-normal states where "power parameters are normal but actual energy has degraded." Compared to traditional detection methods relying on a single power parameter, this invention not only provides early warning of energy degradation but also effectively avoids misjudgments and missed judgments caused by human parameter compensation.

[0035] 2. This invention introduces a confidence-based sparse weighting mechanism and a periodic retraining strategy to dynamically compress the influence of pseudo-normal state samples during model training, suppressing the interference of label contamination on model accuracy and improving the model's long-term stability and adaptability. The overall system possesses high accuracy, high robustness, and good scalability, making it suitable for various laser cutting processes and complex production scenarios. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0037] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] For examples, please refer to Figure 1 As shown in this embodiment, the laser energy decay detection method for laser cutting machines includes:

[0040] Real-time acquisition of temperature data of the laser module during laser cutting, and calculation of its temperature rise rate per unit time;

[0041] Simultaneously acquire kerf images during the cutting process, and extract kerf stability indicators based on kerf edge morphology, width fluctuation, or structural texture consistency;

[0042] The temperature rise rate and the kerf stability index are respectively input into a preset multimodal fusion neural network model, wherein the first channel is used to extract potential abnormal thermal load features from the temperature rise rate; and the second channel is used to extract abnormal processing consistency features from the kerf stability index.

[0043] The abnormal heat load characteristics and abnormal processing consistency characteristics are jointly nested and encoded by the fusion discriminant layer, and a discriminant label is output. The discriminant label includes three categories: normal state, energy decay state and pseudo-normal state.

[0044] If the judgment result is a pseudo-normal state, it is determined that there may be a situation where the laser energy decay is artificially compensated and masked.

[0045] For samples identified as pseudo-normal states, their participation weight in subsequent iterative training of the multimodal fusion neural network model is reduced, and the multimodal fusion neural network model is periodically retrained based on the updated training sample set.

[0046] During the operation of laser cutting equipment, the laser module, as the core component for generating laser energy, directly affects the stability and quality of laser output due to its thermal load. Especially in the early stages of energy decay, traditional methods of assessing the health status of equipment through process parameters such as laser power and cutting speed often fail due to human compensation adjustments (such as increasing power or decreasing speed), thus masking potential performance degradation problems and creating a so-called "pseudo-normal" operating condition.

[0047] To achieve a more objective, independent, and early determination of the energy state of a laser, this invention proposes a physical feature extraction method based on the temperature rise rate of the laser module, and uses it as one of the important input features of the cross-domain fusion model. This method utilizes the dynamic response to temperature changes to reveal the trend of internal thermal load changes in the laser, thereby indirectly assessing the authenticity of the energy output.

[0048] Laser modules are typically equipped with multiple temperature sensing units, commonly found in key components such as the laser resonator, electro-optic module, heat sink, and power driver module within the laser. This invention selects a temperature sensor located closest to the amplification and gain medium along the laser output path as a representative sampling point.

[0049] The sensor has high time resolution (sampling frequency of not less than 1Hz, preferably 10Hz or higher) and temperature measurement accuracy within ±0.1°C, which can meet the requirements for accurate identification of minute temperature fluctuations.

[0050] The acquisition and control module acquires the sensor's temperature data stream in real time through a data interface with the laser control system (such as RS-485, CAN, or Modbus-TCP), and the system records the laser module's temperature value at each moment.

[0051] Based on continuous temperature sampling values, the temperature rise rate is defined by calculating the change in temperature per unit time. Let two consecutive moments t1 and t2 (where t2 > t1) correspond to temperature values ​​T1 and T2, respectively. The temperature rise rate R is expressed as follows: subtract the temperature value of the laser module at the current moment from the temperature value at the previous moment, and then divide by the time difference between the two moments. The result is the rate of temperature change per unit time, called the temperature rise rate.

[0052] In practical systems, a sliding window approach is used to calculate the average of multiple continuous temperature rise rates in order to offset single-point abnormal fluctuations caused by environmental noise or instantaneous disturbances.

[0053] For example, at a sampling frequency of 1 second, five consecutive sampling points form a sliding window. The temperature rise rate within this window is calculated for each second, and the average value is obtained as the temperature rise rate characterization index for the current period.

[0054] This moving average can be further used for trend analysis, for example:

[0055] If the temperature rise rate of multiple consecutive windows shows a monotonically increasing trend, it can be preliminarily judged that the laser module is in a state of continuous temperature rise.

[0056] If the rate of temperature rise shows periodic peaks or jumps, there may be problems such as abnormal modulation of the driving current or uneven heating of the optical system.

[0057] If the temperature rise rate is abnormally higher than the historical benchmark under constant cutting power, it may indicate a decrease in the internal efficiency of the laser, i.e., there is a risk of energy degradation.

[0058] In laser cutting, the cutting quality is directly reflected in the geometry and consistency of the kerf (hereinafter referred to as "kerf"). The continuity of the kerf edge, the uniformity of its width, and the structural stability of the texture can all serve as important indirect parameters reflecting the laser energy state. However, most traditional laser cutting energy monitoring methods only focus on the laser's electrical power, pulse parameters, or processing time, while neglecting the feedback signal of the processing result itself on the energy state.

[0059] To address this deficiency, this invention proposes a method for extracting kerf stability indicators based on image vision, serving as a second type of key feature parameter in the multimodal fusion model that complements the laser's temperature rise rate. This method utilizes industrial image acquisition and computer vision technology to quantitatively analyze the actual cutting morphology of the kerf, thereby providing a more visual and result-oriented evaluation dimension for the laser's energy state.

[0060] The image acquisition system consists of a high-resolution industrial camera, a lens module, a light source module, and a protective system. The camera is preferably a 5-megapixel or higher linear or area-scan industrial camera, possessing high shutter speed and anti-interference capabilities. The light source can employ high-brightness ring LEDs or coaxial backlighting to ensure high contrast at image edges.

[0061] The camera is mounted behind or to the side of the laser cutting head, forming a certain angle (e.g., 45 degrees) with the kerf. It moves synchronously with the laser cutting head via an industrial arm or linear guide, or it can be fixed in the cutting area. It records the kerf state image after the cutting path is completed using a triggered acquisition method. The system synchronizes with the laser or CNC system via a controller to ensure that the kerf corresponding to the image and the temperature rise rate of the sampled image are within the same process time period.

[0062] The acquired raw images are first subjected to grayscale conversion and filtering to remove background noise and uneven lighting. Then, edge detection algorithms (such as Canny, Sobel, or adaptive gradient algorithms) are used to extract the kerf edge contours.

[0063] After extracting the kerf edge region, morphological operations (such as erosion, dilation, and opening / closing operations) are used to clean and smooth the edge lines to obtain a continuous kerf path. Then, the kerf principal axis and its edge point cloud are determined by centerline fitting or edge envelope.

[0064] This invention defines the kerf stability index as a quantitative value that comprehensively reflects the continuity of the cut, the clarity of the edge, and the uniformity of the width, specifically including the following three dimensions:

[0065] The kerf width fluctuation is defined as the standard deviation of the kerf width per unit length. The extraction method is as follows: sample multiple cross-sections at equal intervals along the kerf path, measure the distance between the left and right edges of each cross-section to obtain the kerf width sequence, and calculate the standard deviation of the kerf width sequence. If the laser energy is unstable, the kerf width will fluctuate periodically or suddenly, and the standard deviation will increase, indicating poor cutting consistency.

[0066] The edge smoothness index is defined as: the rate of change of curvature of the two sides of the kerf or the average fitting residual; the extraction method is: to perform polynomial fitting on the left and right edges of the kerf and calculate the average deviation of the point set from the fitting line; insufficient energy or overheating will lead to defects such as edge ablation and slag adhesion, which will distort the edge contour.

[0067] The structural texture consistency score is defined as: a feature similarity score between local texture blocks in an image; the extraction method is: to calculate the mean texture similarity between image blocks using gray-level co-occurrence matrix, LBP (Local Binary Pattern), or depth feature extraction methods; if the energy state is unstable, the texture distribution of the cut surface will show obvious non-uniformity.

[0068] After normalizing the kerf width fluctuation, edge smoothness index, and structural texture consistency score, the scores are weighted and summed to obtain the total kerf stability score index, which is used to characterize the processing consistency of the cutting kerf segment.

[0069] To achieve accurate monitoring and early decay identification of the energy state of a laser in a laser cutting machine, this invention designs an energy state discrimination method based on a multimodal fusion neural network. This method combines two complementary key feature parameters: the temperature rise rate (representing changes in the laser's thermal load) and the kerf stability index (representing actual cutting quality performance). These parameters are input into two independent sub-channels of the fusion model for deep semantic extraction, ultimately achieving cross-physical domain feature combination and state discrimination in the fusion layer.

[0070] To ensure accurate and efficient model operation, both types of raw data need to be structured and feature extracted. The following sections describe the feature extraction steps for the temperature rise rate channel and the kerf stability index channel, respectively.

[0071] The rate of temperature rise reflects the thermal change trend of the laser module per unit time, and can indirectly determine whether there is a decrease in laser energy output or efficiency reduction. This invention uses the following steps to extract this feature for a neural network model:

[0072] The laser module integrates multiple temperature sensors, preferably located near the laser resonant cavity, drive module, or gain medium. The sensors output the current temperature value in real time via a bus interface with the control system (such as RS485, CAN, or Ethernet). The sampling time interval is Δt, in seconds (e.g., 1 second).

[0073] The system records the temperature values ​​at two consecutive time points: the previous temperature: T1; and the current temperature: T2.

[0074] The rate of temperature rise, R, is defined as the rate of temperature change per unit time. It can be calculated as follows: subtract the temperature from the previous temperature, and then divide the result by the time interval between the two times. This calculation is performed immediately upon the arrival of new temperature data, generating a real-time temperature rise rate stream.

[0075] To suppress instantaneous peak values ​​caused by environmental fluctuations or sampling errors, this invention introduces a sliding window mechanism. A certain number of historical temperature rise rate data (e.g., the past 5 seconds) are selected to form a time series, and the average or weighted average of this series is calculated.

[0076] For example, if the temperature rise rates at five time points are R1, R2, R3, R4, and R5, then the temperature rise rate characteristic can be defined as:

[0077] Average temperature rise rate = (R1 + R2 + R3 + R4 + R5) ÷ 5; a weighted algorithm can also be used, giving higher weight to the most recent time point to increase response sensitivity.

[0078] The processed temperature rise rate data is input as a fixed-length numerical sequence (such as 5-dimensional or 10-dimensional) into the first channel of the neural network. The network structure can be a one-dimensional convolutional network (1D-CNN) or LSTM (Long Short-Term Memory network), used to learn the thermal load change pattern and output thermal semantic vectors.

[0079] The kerf stability index is a structured quantification of cutting quality using a visual method, reflecting the consistent energy output capability of the laser during actual processing. This invention extracts this index through the following steps:

[0080] Industrial cameras acquire kerf images at a set frame rate (e.g., 5 fps) or by triggering. The images are enhanced with preset light sources (e.g., coaxial illumination, backlighting) to improve the contrast of the kerf edges.

[0081] After acquiring the image, the following preprocessing steps are performed: grayscale conversion (to remove color interference), noise filtering (such as Gaussian filtering), and edge enhancement (such as Sobel or Laplacian operators).

[0082] The Canny edge detection algorithm is used to extract the seam edges from the image, forming left and right edge point sets. If the edges are discontinuous, morphological operations (such as dilation-erosion) are used for smoothing and repair.

[0083] The kerf is divided into several cross-sections, and the distance between the left and right edges is measured at each cross-section to obtain the kerf width at that point. All kerf width values ​​within the entire image area are statistically analyzed, and their standard deviation is calculated as the kerf width variability. The kerf width variability, which is the standard deviation of all kerf width values, represents the uniformity of the kerf during processing. A larger standard deviation indicates unstable laser output, causing kerf jitter.

[0084] The kerf edge curve is fitted using a quadratic or cubic polynomial for each side of the edge. The vertical distance from the original edge point to the fitted curve is calculated, and the average value is taken as the edge residual. The smoothness index indicates the average deviation of all edge points from the fitted edge. The larger the deviation, the more uneven the edge, which may be caused by uneven ablation or fluctuations in heat input.

[0085] The image is divided into local blocks (e.g., 8×8 pixels), and the texture features of each block are extracted, such as gray-level co-occurrence matrix, LBP (Local Binary Pattern), or depth image features (e.g., ResNet extraction layers).

[0086] The texture similarity between all local blocks is calculated as follows: the kerf texture consistency score = the average similarity between the texture features of all image blocks; if the texture distribution is messy or there are abrupt changes, it indicates that there are fluctuations in the laser cutting process, reflecting the instability of the energy state.

[0087] The three sub-indicators—kerf width fluctuation, edge smoothness index, and texture consistency score—are normalized and then arranged in a fixed order to form a feature vector, which is then input into the second channel of the neural network. Lightweight image coding networks (such as MobileNet, EfficientNet) or MLPs can be used for processing to extract visual semantic feature vectors.

[0088] The two channels output thermal and visual feature vectors respectively, which are then concatenated in a fusion layer and finally classified using a multi-layer fully connected network. Output labels include, but are not limited to, the following three categories: normal state, energy decay state, and pseudo-normal state (compensation actions masking energy decline). If the classification result is a pseudo-normal state, it is determined that there may be a situation where laser energy decay is artificially masked by compensation. During model training, a label correction mechanism is introduced to label and reduce the weight of pseudo-normal state samples to prevent contamination from affecting the long-term performance of the model.

[0089] In the intelligent detection of energy status in laser cutting equipment, the multimodal fusion neural network model proposed in this invention classifies and distinguishes the energy decay state of the laser by jointly analyzing the temperature rise rate of the laser module and the kerf stability index. In actual operation, some "compensatory normal state" samples—that is, conditions where the surface power is normal but the actual energy is decreasing—may be masked by artificially increasing the laser power or artificially decreasing the cutting speed, thus being mistakenly recorded as "normal." If these samples are included in subsequent model training without differentiation, the model will learn incorrect category boundaries, reducing its sensitivity to early signs of energy decay.

[0090] In this invention, a pseudo-normal state refers to a situation where the laser module's temperature rise rate is significantly higher than historical process standards, and the kerf stability indicators (such as kerf width fluctuation and edge smoothness) show a deteriorating trend, but the laser's apparent power parameters remain within the normal range. When the multimodal fusion model determines the result to be in the "pseudo-normal" category, it automatically labels the sample as a "low-confidence sample" and records its feature vector and initial label.

[0091] The system stores this type of sample in a "pseudo-normal sample pool" for use in subsequent training processes and assigns it initial weight coefficients.

[0092] During training, each input sample xi corresponds to a weight coefficient wi, which determines the actual contribution of the sample to the model's loss function.

[0093] For samples marked as pseudo-normal states, this invention employs the following weight control strategy:

[0094] The baseline confidence score ci is set based on the maximum value of the model's output classification distribution on that sample. If the model output is close to a uniform distribution (e.g., the probability of each class is close to 1 / 3), the confidence score is low.

[0095] A sparse regularization term is introduced so that only samples with high confidence scores receive larger weights. The weight wi is obtained by combining the confidence score ci with a suppression function (such as an exponential decay function). That is, if ci approaches 1, it means that the model's judgment on the sample is stable, and wi is close to 1; if ci approaches 0.33 (completely uncertain), wi drops sharply to 0.1 or lower.

[0096] Samples identified as pseudo-normal generally have low confidence levels and therefore have a lower influence on the training loss function, thus reducing the risk of overfitting the model on these potentially labeled samples.

[0097] This invention combines the aforementioned weight adjustment methods to design a periodic model retraining framework to dynamically adapt to changes in actual working conditions and the evolution of data quality.

[0098] Set a fixed time window (e.g., every 24 hours or every 1000 artifacts) as one retraining cycle. Collect all samples (including normal, energy decay, and pseudo-normal samples) within the cycle.

[0099] Confidence assessment and weight allocation for pseudo-normal samples;

[0100] High-confidence normal and abnormal samples are given high weights;

[0101] Construct an updated training sample set, including weight annotations.

[0102] The model is retrained based on the updated sample set, using a weighted loss function. The loss function is expressed as: Weighted Loss Function = Weighted Average of the Losses of All Samples, where the loss of each sample is multiplied by its corresponding weight coefficient.

[0103] Through the above mechanism, the impact of pseudo-normal samples on model training is effectively controlled, preventing them from shifting the decision boundary.

[0104] After model retraining, the system performs a self-validation process. If the model exhibits a large number of "pseudo-normal" misclassifications on new samples, the confidence calculation strategy is adjusted or some boundary samples are relabeled.

[0105] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for detecting laser energy decay in a laser cutting machine, characterized in that: include: Real-time acquisition of temperature data of the laser module during laser cutting, and calculation of its temperature rise rate per unit time; Simultaneously acquire kerf images during the cutting process, normalize the kerf width fluctuation, edge smoothness index, and structural texture consistency score respectively, and then sum them by weight to obtain the kerf stability index. The kerf width fluctuation is defined as the standard deviation of the kerf width per unit length. The extraction method involves sampling multiple cross-sections at equal intervals along the kerf path, measuring the distance between the left and right edges of each cross-section to obtain a kerf width sequence, and calculating the standard deviation of the kerf width sequence. The edge smoothness index is defined as the rate of change of curvature of the edges on both sides of the kerf or the average fitting residual. The extraction method involves performing polynomial fitting on the left and right edges of the kerf and calculating the average deviation of the point set from the fitted line. The structural texture consistency score is defined as the mean texture similarity between image patches based on the feature similarity score between local texture patches in the image. The temperature rise rate and the kerf stability index are respectively input into a preset multimodal fusion neural network model, wherein the first channel is used to extract potential abnormal thermal load features from the temperature rise rate; and the second channel is used to extract abnormal processing consistency features from the kerf stability index. The processed temperature rise rate data is input into the first channel of the neural network as a fixed-length numerical sequence. The kerf width fluctuation, edge smoothness index and structural texture consistency score are normalized and then formed into a feature vector in a fixed order, which is then input into the second channel of the neural network. The two channels output thermal feature vectors and visual feature vectors respectively, which are then connected to the fusion layer for splicing and finally classified through a multi-layer fully connected network. The abnormal heat load characteristics and abnormal processing consistency characteristics are jointly nested and encoded by the fusion discriminant layer, and a discriminant label is output. The discriminant label includes three categories: normal state, energy decay state and pseudo-normal state. If the judgment result is a pseudo-normal state, it is determined that there may be a situation where the laser energy decay is artificially compensated and masked. For samples identified as pseudo-normal states, their participation weight in subsequent iterative training of the multimodal fusion neural network model is reduced, and the multimodal fusion neural network model is periodically retrained based on the updated training sample set.

2. The laser energy decay detection method for laser cutting machines according to claim 1, characterized in that: The calculated temperature rise rate includes: Temperature values ​​at consecutive time points are collected using a temperature sensor located inside the laser module near the gain medium at a sampling frequency of not less than 1 Hz. At any two consecutive time points, the difference between the current temperature and the previous temperature is calculated and divided by the time interval between the two time points to obtain the rate of temperature change per unit time, which is used as the temperature rise rate.

3. The laser energy decay detection method for laser cutting machines according to claim 2, characterized in that: The temperature rise rate is smoothed using a sliding window before being input into the neural network, including: A sliding window is formed by combining multiple consecutive temperature rise rate samples; The arithmetic average or weighted average of all temperature rise rates within the window is calculated as the smoothed temperature rise rate feature for the current time period; the smoothed feature is used as the time series data input to the first channel to extract potential abnormal heat load features.

4. The laser energy decay detection method for laser cutting machines according to claim 1, characterized in that: Extracting kerf stability indicators also includes: acquiring kerf images of the cutting area using an industrial camera; performing grayscale conversion, filtering, and edge enhancement processing on the images; extracting the kerf edges using an edge detection algorithm; and performing morphological restoration on the edge data.

5. The laser energy decay detection method for laser cutting machines according to claim 1, characterized in that: For samples identified as pseudo-normal states, their participation weight in subsequent iterative training of the multimodal fusion neural network model is reduced, including: Based on the classification confidence score output by the model, a weight coefficient is assigned to each pseudo-normal sample; If the maximum classification probability of a pseudo-normal sample is close to 1 / 3, its confidence level is low. The weights of low-confidence samples are compressed to below 0.1 using an exponential decay function.

6. The laser energy decay detection method for laser cutting machines according to claim 1, characterized in that: The periodic retraining of the multimodal fusion neural network model based on the updated training sample set includes: Set a fixed training period. After each period, construct a training dataset from three types of samples: normal, decaying, and pseudo-normal. Input each sample and its corresponding weight into the model for weighted training. The updated model is used for energy decay discrimination in the next period.

7. The laser energy decay detection method for laser cutting machines according to claim 1, characterized in that: The loss function used during training is the weighted cross-entropy loss function, which is calculated as follows: Multiply the prediction loss of each training sample by its corresponding participation weight; The sum of all weighted losses divided by the number of samples yields the overall loss value. Use this loss value to update the model weights.

8. The laser energy decay detection method for laser cutting machines according to claim 7, characterized in that: Pseudo-normal samples are recorded in a pseudo-sample pool, which stores all sample data marked as pseudo-normal by the model, confidence scores, and label information; pseudo-sample weights are redistributed from the sample pool before each round of training to adapt to changes in operating conditions.

Images