A deep learning-based online quality discrimination method for laser-cut steel plates

Abstract

The present application belongs to the technical field of laser processing quality detection, and provides a kind of laser cutting steel plate quality online identification method based on deep learning, comprising: collecting plasma spectrum of laser action area, incision front gas flow pressure fluctuation and back melt acoustic emission response signal, and constructing cross-zone thermal response sequence;Three kinds of sensing signals are detected respectively before and after the local maximum point of the junction position, to form the cross-zone thermal history migration characteristic vector;Cross-zone thermal history migration characteristic vector is input into deep learning inference model, and cutting quality discrimination characteristic vector is constructed;Cutting quality discrimination characteristic vector is input into quality discrimination model, and each defect category probability is output by nonlinear mapping, to determine the defect discrimination result with maximum probability value;According to the quality discrimination result and the corresponding thermal response change trend of cross-zone thermal history migration characteristic vector, the weight updating relationship based on defect feedback constraint is formed.

Description

Technical Field

[0001] This invention belongs to the field of laser processing quality inspection technology, specifically a deep learning-based online method for judging the quality of laser-cut steel plates. Background Technology

[0002] With the increasing application of high-strength steel plates, hot-formed steel plates, and graded performance steel plates in automobile manufacturing, rail transportation, and engineering equipment, laser cutting has become an important method for processing high-performance steel plate parts. Existing technologies typically collect information such as temperature, molten pool state, radiation spectrum, and kerf morphology during the laser cutting process, and use deep learning models to analyze the collected data to determine whether there are quality problems such as incomplete cutting, slag buildup, burrs, and burnt edges based on the current cutting status.

[0003] However, for high-performance steel plates with multiple performance regions, different performance regions have different thermal conductivity and thermal response characteristics. During laser cutting, the heat generated in the current cutting area is not limited to the laser action position, but diffuses along the interior of the material to the surrounding area, so that the thermal response state of the current area is affected by the adjacent performance regions. When the cutting trajectory passes through the boundary of different performance regions, the local temperature change, the stability of the molten pool and the melting behavior of the material will change dynamically with the heat transfer relationship between the regions.

[0004] Existing online quality assessment methods based on deep learning typically use local state data collected at the current moment as the basis for judgment, or establish fixed discrimination models for different performance regions. These methods are insufficient in characterizing the thermal influence transfer relationship between different performance regions during laser cutting, making it difficult to accurately identify dynamic quality changes at the boundary of performance regions. Summary of the Invention

[0005] In order to overcome the shortcomings of the prior art, at least one technical problem raised in the background art is solved.

[0006] The technical solution adopted by this invention to solve its technical problem is:

[0007] The plasma spectrum of the laser-acting area, the airflow pressure fluctuation at the front edge of the cut, and the acoustic emission response signal of the melt on the back side were collected. Based on the difference in thermal diffusivity and material properties, the first and second performance zones and the boundary position were divided. The laser passing through the boundary position was used as the time reference, and the peak response time of the three types of sensor signals was used as the alignment anchor point and propagation delay compensation was performed. The signals were mapped to a unified time coordinate to construct a cross-regional thermal response sequence.

[0008] Based on the cross-regional thermal response sequence, the local maxima of three types of sensor signals before and after the boundary position are detected respectively. The corresponding time is taken as the peak time and the corresponding amplitude is the peak amplitude. The peak time difference and amplitude difference are calculated to form thermal response migration parameters. They are arranged in order of peak time to form a cross-regional thermal history migration feature vector.

[0009] The cross-regional thermal history migration feature vector is input into the deep learning inference model. The weights are calculated inversely based on the ratio of the thermal diffusivity coefficients of adjacent performance regions. The weights are multiplied by the output values ​​of the feature nodes of the preceding regions as the parameters of the influence of the preceding thermal response on the current position. The feature vector for judging cutting quality is constructed by combining the molten pool propulsion speed, the energy of molten material discharge, and the solidification shrinkage displacement parameters of the cut.

[0010] The cutting quality discrimination feature vector is input into the quality discrimination model to extract the thermal accumulation parameter and the thermal response mutation parameter of the boundary region. The two types of parameters are input into the fully connected layer of the quality discrimination model, and the probability of each defect category is output through nonlinear mapping. The defect discrimination result is determined by the maximum probability value.

[0011] Based on the quality discrimination results and the thermal response change trend corresponding to the cross-regional thermal history migration feature vector, the influence of each thermal migration feature on the discrimination results is analyzed and the contribution correction amount is calculated. Accordingly, the weights of thermal migration features in the discrimination model are updated to form a weight update relationship based on defect feedback constraints.

[0012] As a further aspect of the present invention:

[0013] The process of constructing a cross-regional thermal response sequence is as follows:

[0014] The material structure parameters, thermal diffusivity, specific heat capacity and thermal conductivity parameters of the steel plate to be cut are obtained along the cutting trajectory. Based on the difference in thermal diffusivity at different locations, the steel plate is divided into a first performance zone and a second performance zone with different thermal response characteristics.

[0015] Based on the correspondence between the real-time movement position of the laser head and the spatial coordinates of the steel plate, the spatial transformation position of the laser action point entering the first performance zone and the second performance zone is determined, and the moment when the laser beam passes through the boundary between the first performance zone and the second performance zone is marked as the time reference moment.

[0016] The plasma spectral change sequence, the gas flow pressure fluctuation sequence at the cutting edge, and the acoustic emission response sequence of the back melt were obtained within a preset time window before and after the boundary position.

[0017] Using the peak response times of the three types of sensor signals as alignment anchors, propagation delay compensation is performed on different sensor signals to map the peak values ​​of thermal disturbance response, energy change nodes, and abnormal fluctuation locations in each signal to a unified time coordinate.

[0018] Based on the multi-source response data after time synchronization, a thermal response migration sequence centered on the boundary location is established, and the thermal input state of the region before the boundary is correlated with the thermal response state of the region after the boundary to form a cross-regional thermal response sequence.

[0019] The process of constructing thermal response migration parameters is as follows:

[0020] Based on the cross-regional thermal response sequence, the local maxima of three types of sensor signals before and after the boundary are detected respectively. The time corresponding to the maximum value is taken as the peak time and the corresponding amplitude is taken as the peak amplitude. The peak time difference and amplitude difference of each signal before and after the boundary are calculated as the basic parameters of thermal disturbance propagation time sequence and energy amplitude change.

[0021] Extract the pressure peak migration time, pressure decay period, and duration of auxiliary airflow disturbance from the airflow pressure fluctuation signal at the cut front edge, and extract the acoustic energy release amplitude, frequency variation range, and response duration from the acoustic emission response signal of the back melt.

[0022] Based on the peak time difference and amplitude difference of the three types of sensor signals before and after the junction, the time delay parameters between the plasma heat input change and the cut pressure response and the melt release response are calculated.

[0023] Based on the energy change trends and response time delay relationships of each signal, a migration mapping relationship is established during the propagation of thermal disturbance from the previous performance region to the current laser action region, unifying the temporal differences and energy attenuation laws of different physical responses under the same thermal propagation path framework;

[0024] Based on the response lag, energy attenuation ratio, and response differences before and after regional conversion during the thermal response propagation process, thermal response migration parameters characterizing the impact of performance regional conversion are constructed by associating and integrating temporal lag characteristics, energy attenuation characteristics, and cross-regional state change characteristics.

[0025] The process of forming the cross-regional thermal history migration feature vector is as follows:

[0026] Based on the peak time and amplitude difference in the thermal response migration parameters, the thermal response migration parameters before and after the boundary of different performance regions are sorted according to the order of peak time.

[0027] Based on the sequential correspondence of peak times in the sorted thermal response migration parameters, the influence time window during the transfer of thermal accumulation state of the preceding performance region to the current cutting position is determined.

[0028] Within the time window of influence, the cumulative amount of heat input, the degree of response attenuation, the migration delay time, and the amplitude of the region conversion disturbance are extracted, and the parameters of each type are normalized.

[0029] The normalized thermal accumulation parameters, migration delay parameters, energy decay parameters, and boundary disturbance parameters are arranged in chronological order of peak time to form a cross-regional thermal history migration feature vector.

[0030] The process of multiplying the weights by the output values ​​of the preceding region feature nodes as the parameter for the influence of the preceding thermal response on the current location is as follows:

[0031] The cross-regional thermal history migration feature vector is input into a pre-built deep learning inference model. Based on the movement position relationship of the laser along the cutting trajectory, the thermal response features corresponding to different performance regions are assigned region numbers, and the feature association relationship between the previous performance region and the current laser action region is established.

[0032] Obtain the thermal diffusivity between the first performance region and the second performance region, and calculate the weights based on the inverse relationship of the ratio of the thermal diffusivity of adjacent performance regions.

[0033] Based on the calculated weights, the correlation strength between feature nodes in different performance regions in the deep learning inference model is corrected, and the output value of the preceding region feature node is multiplied by the weight as the contribution of the preceding hot response to the current position.

[0034] The contribution of preceding thermal response within different time windows is weighted and fused to obtain thermal response influence parameters that characterize the degree of influence of preceding performance region thermal history.

[0035] The process of constructing the feature vector for judging cutting quality is as follows:

[0036] Based on the influence parameters of the thermal response of the preceding region on the current laser action position, the thermal state change process of the molten pool during laser action is obtained, and the spatial correspondence between the advancing position of the leading edge of the molten pool and the solidification position of the trailing edge is determined according to the laser movement direction.

[0037] Collect information on the boundary changes of the molten pool during the laser cutting process, and calculate the parameters of the molten pool's advance speed along the cutting trajectory, the rate of change of the advance speed, and the duration of molten pool stability.

[0038] Based on the energy release process corresponding to the acoustic emission response signal of the molten material on the back side, the peak acoustic energy, energy decay rate and continuous discharge time of the molten material during the process of detaching from the steel plate are extracted and converted into discharge energy parameters that characterize the discharge capacity of the molten material.

[0039] The edge shrinkage change, the trend of the cut width change, and the solidification shrinkage displacement parameters during the cooling and solidification process of the cut area were obtained, and the correlation between the change of heat input and the evolution of the cut morphology was established.

[0040] By combining the parameters of the influence of the preceding thermal response on the current position, the parameters of the molten pool propulsion speed, the parameters of the molten material discharge energy, and the parameters of the solidification shrinkage displacement of the cut, feature fusion is performed according to the thermal state influence relationship to construct a cutting quality discrimination feature vector for quality judgment.

[0041] The process of extracting thermal accumulation parameters and abrupt change parameters of thermal response in the interface region is as follows:

[0042] Input the cutting quality discrimination feature vector into the quality discrimination model, and accumulate the product of the thermal input intensity and the duration of action at each position along the cutting trajectory based on the thermal response feature components associated with each spatial position in the quality discrimination feature vector.

[0043] The total energy superposition and heat diffusion coverage are calculated separately for the first performance zone and the second performance zone to obtain the heat accumulation parameters during the cross-performance zone migration process.

[0044] For the boundary between the first performance zone and the second performance zone, the multi-source sensing feature components corresponding to the boundary before and after the boundary are extracted from the mass discrimination feature vector, and the difference in the rate of change of plasma spectral energy, the gradient of change of airflow pressure, and the rate of change of acoustic emission energy before and after the boundary are calculated respectively.

[0045] By combining the peak response offset before and after the boundary and the degree of synchronous change of multi-source signals, a thermal response mutation parameter is constructed to characterize the degree of thermal state mutation caused by the switching of performance regions.

[0046] The process of determining the defect identification result based on the maximum probability value is as follows:

[0047] The thermal accumulation parameters and the thermal response mutation parameters of the boundary region are input into the fully connected layer of the quality discrimination model, and the probability of each defect category is output through nonlinear mapping.

[0048] Based on the continuously changing heat accumulation parameters along the cutting trajectory and the abrupt change parameters of the thermal response in the interface region, the heat input state, melt discharge state and solidification shrinkage trend of the current laser-acting area are analyzed to determine the corresponding quality state characteristics.

[0049] The output probabilities of the fully connected layer corresponding to the quality status characteristics are compared. The maximum probability value is used to determine the incomplete cutting, slag, burr or burnt edge defect status of the current cutting area, and the quality judgment result is output.

[0050] The process of forming the weight update relationship based on defect feedback constraints is as follows:

[0051] Based on the defect category results output by the quality discrimination model, obtain the cross-regional thermal history migration feature vectors that participate in the discrimination calculation under the corresponding defect category, and extract the fluctuation amplitude, offset direction and peak time migration position of each peak time difference and amplitude difference in the defect occurrence area.

[0052] Based on the correlation between thermal accumulation parameters, thermal response mutation parameters and peak time differences and amplitude differences under different defect categories, an influence mapping relationship between each feature component in the cross-regional thermal history migration feature vector and the defect category is established.

[0053] Based on the category deviation between the current discrimination result and the actual quality inspection result, the influence of each peak time difference and amplitude difference on the discrimination result is analyzed, and the discrimination residual is allocated according to the influence ratio of each feature component on the defect classification result to obtain the error contribution of each feature component.

[0054] Based on the error contribution, the fluctuation amplitude of each peak time difference and amplitude difference, and the historical discrimination error change trend, the contribution correction of each characteristic component is calculated, and the weight adjustment coefficients corresponding to each peak time difference and amplitude difference are updated according to the contribution correction.

[0055] The updated weight adjustment coefficients are loaded into the thermal history migration feature mapping layer in the quality discrimination model, and the weight values ​​corresponding to each feature dimension in the cross-regional thermal history migration feature vector are recalculated to form a thermal migration feature weight update relationship based on defect feedback constraints.

[0056] The beneficial effects of this invention are as follows:

[0057] By simultaneously acquiring plasma spectra, gas flow pressure fluctuations at the cutting edge, and acoustic emission response signals of the molten material on the back side during laser cutting, and combining this with the differences in thermal diffusion characteristics between different performance regions of the steel plate, a cross-regional thermal response analysis mechanism is established with the boundary position of the performance regions as the time reference. This allows for the effective characterization of the thermal influence process generated by the preceding region and migrating to the current action position during cutting, avoiding the problem of existing methods that only rely on the instantaneous state of the current laser action area for quality judgment while ignoring the influence of the preceding thermal history. By correlating and mapping the response time difference, energy change trend, and thermal propagation characteristics of multi-source sensor signals, a cross-regional thermal history migration feature vector is constructed, realizing a quantitative description of the continuous evolution of the thermal state during the switching of different performance regions. This improves the completeness and accuracy of sensing the cutting state of complex gradient performance steel plates or multi-structure steel plates. Furthermore, by using a deep learning inference model combined with the differences in thermal diffusion parameters between adjacent performance regions to adjust the regional migration weights, the model can dynamically learn the influence law of the preceding region's thermal response on the current cutting region and integrate the molten material... Multi-dimensional parameters such as pool propulsion state, melt discharge characteristics, and cut solidification shrinkage behavior enhance the physical correlation of quality discrimination features and reduce the risk of misjudgment caused by fluctuations in a single detection signal. By establishing a continuous evolution relationship of thermal state along the cutting trajectory and extracting the evolutionary constraint relationship between thermal accumulation parameters and abrupt changes in thermal response parameters in the interface region, the formation process of typical cutting defects such as incomplete cutting, slag adhesion, burrs, and burnt edges can be identified and accurately judged in advance, improving the real-time performance and reliability of online quality detection. At the same time, the contribution of defect discrimination results to different thermal migration characteristics is analyzed, and the feature weights in the discrimination model are updated using the contribution correction amount. This allows the quality discrimination model to be continuously optimized with actual cutting feedback, improving the model's adaptability to different material performance combinations, different cutting parameters, and different thermal migration states. This realizes the transformation from single-point state detection to cross-regional thermal history perception and defect feedback optimization in the laser cutting of steel plates, effectively improving the accuracy, stability, and engineering application value of online quality discrimination for complex performance steel plates. Attached Figure Description

[0058] The invention will now be further described with reference to the accompanying drawings.

[0059] Figure 1 This is a flowchart of the steps in Embodiment 1 of the present invention;

[0060] Figure 2 This is a schematic diagram of the process framework of Embodiment 1 of the present invention. Detailed Implementation

[0061] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0062] Example 1: Please refer to Figure 1-2 As shown in the embodiment of the present invention, a method for online quality judgment of laser-cut steel plates based on deep learning includes the following steps:

[0063] Step S10: Collect the plasma spectrum of the laser-acting area, the airflow pressure fluctuation at the front edge of the cut, and the acoustic emission response signal of the molten material on the back side. Based on the difference in thermal diffusivity and material properties, divide the first and second performance zones and the boundary position. Using the laser passing through the boundary position as the time reference, and the peak response time of the three types of sensor signals as the alignment anchor point and perform propagation delay compensation, map the signals to a unified time coordinate to construct a cross-regional thermal response sequence.

[0064] In step S10, the process of constructing the cross-regional thermal response sequence is as follows:

[0065] The material microstructure, thermal diffusivity, specific heat capacity, and thermal conductivity parameters of the steel plate to be cut are obtained along the cutting trajectory. Based on the differences in thermal diffusivity at different locations, the steel plate is divided into a first performance zone and a second performance zone with different thermal response characteristics. A combination of industry-standard laser flash testing and differential scanning calorimetry is used to accurately detect the thermal properties of the steel plate across the entire cutting path. In actual implementation, a full-area grid-based detection system is pre-installed for non-homogeneous steel materials such as graded-performance steel plates and multi-structure composite steel plates. Multiple detection points are evenly distributed along the preset laser cutting trajectory, with a density sufficient to completely cover the cutting path and the process-affected areas on both sides of the cut, ensuring no blind spots. Stable pulse heating is applied to each detection point using professional thermal diffusivity testing equipment, and data on the temperature rise of the steel plate after heating are collected to measure... The thermal diffusion capacity values ​​at each location are calculated, and the heat storage capacity and thermal conductivity parameters of the steel at each location are obtained through testing. Simultaneously, the metallographic structure, grain structure, material hardness, and other structural properties of the steel at each testing point are recorded. After collecting the physical property parameters for all locations, the plates are divided into regions based on the degree of difference in thermal diffusion capacity at different locations. A continuous plate region with basically consistent thermal conductivity, heat diffusion efficiency, and heat storage performance is designated as the first performance region. A second continuous plate region with significant differences in thermal properties and distinct thermal response speeds and heat transfer patterns is designated as the second performance region. Finally, a clear boundary and well-defined property distinction between the two performance regions are formed on the cutting trajectory, providing an accurate material zoning basis for subsequent cross-regional thermal response time-series analysis and effectively avoiding errors in thermal response analysis caused by uneven plate material composition.

[0066] Based on the correspondence between the real-time movement position of the laser head and the spatial coordinates of the steel plate, the spatial transformation position of the laser action point entering the first and second performance zones is determined, and the moment when the laser beam passes through the boundary between the first and second performance zones is marked as the thermal response migration reference moment. Machine vision positioning combined with laser equipment motion coding calibration is used to achieve precise correspondence between the spatial position of the plate and the cutting time. In actual implementation, an industrial camera is first used to perform overall visual calibration of the placed steel plate, converting the outer contour of the steel plate and the boundary position of the two pre-divided performance zones into spatial coordinate information recognizable by the equipment, and completing the precise matching calibration between the steel plate coordinate system and the laser cutting equipment motion coordinate system. During formal laser cutting... During the laser cutting process, a high-precision motion detection component mounted on the laser head is used to collect the laser head's movement position and trajectory in real time at high frequency. The correspondence between the laser spot's action position and the boundary position of the material's performance area is continuously compared, and the switching process of the laser's action point between the two types of performance areas is monitored in real time. When the center of the laser spot moves precisely to the boundary line between the two types of performance areas, this instant is immediately locked as a unified thermal response migration reference moment. At the same time, the core process parameters such as the laser cutting position, cutting speed, and laser output power corresponding to this moment are recorded simultaneously. This achieves precise binding between the spatial boundary position and the time reference, providing a unified and reliable time zero-point standard for the timing alignment of subsequent multi-type sensor signals.

[0067] The plasma spectral change sequence, the airflow pressure fluctuation sequence at the cutting edge, and the acoustic emission response sequence of the molten material on the back side were acquired within a preset time window before and after the boundary position. A multi-sensor synchronous triggering acquisition mechanism was adopted to achieve synchronous, continuous, and complete acquisition of the three types of thermal response signals. In actual implementation, a complete acquisition time window with symmetrical front and rear was set with the locked thermal response migration reference time as the center. The window duration could completely cover the entire process of thermal response change stages as the laser approaches the boundary, crosses the boundary, and enters a new performance region, ensuring that no thermal state fluctuation details before and after the region switch were missed. During the laser cutting operation, the plasma light generated in the laser action area was continuously acquired by a fiber optic spectral detection device arranged on the side of the cutting nozzle. The entire process of radiation intensity change and spectral characteristic change data is used to form a continuous plasma spectral change sequence. A high-frequency pressure detection device embedded inside the cutting nozzle is used to collect the dynamic pressure change, fluctuation peak and attenuation process of the auxiliary protective airflow at the front end of the cut in real time, forming a complete airflow pressure fluctuation sequence. A magnetic acoustic emission detection device attached to the corresponding position on the back of the steel plate is used to continuously capture the acoustic vibration signal generated during the melting flow, detachment from the plate and cooling solidification of the molten metal, obtaining a continuous acoustic emission response sequence of the molten material. All detection devices are triggered and started by the same industrial control system to ensure that the original acquisition sequence of the three types of signals is completely synchronized, and the original multi-dimensional thermal response data of the entire laser cross-region cutting process is completely preserved.

[0068] Based on the heat transfer paths and response hysteresis characteristics of different sensor signals, time offset compensation is performed on the three types of signals, mapping the peak values ​​of thermal disturbance responses, energy change nodes, and abnormal fluctuation positions in each signal to a unified time coordinate. Differential time-series compensation calibration is implemented based on the inherent transmission hysteresis characteristics of each type of sensor signal to achieve high-precision time alignment of multi-source signals. In actual implementation, extensive calibration experiments are conducted to clarify the inherent response hysteresis patterns of the three types of signals: the plasma spectral signal, generated by direct laser thermal action, has the shortest transmission path, the fastest response speed, and no significant hysteresis, and is used as the reference signal. The airflow pressure signal at the cut edge is affected by the airflow propagation distance and sensor response speed, exhibiting a fixed short-term response lag. The acoustic emission signal from the molten material on the back of the steel plate can only be generated after the molten pool metal has fully melted and flowed away, resulting in the longest heat transfer path and the longest response lag time. Based on the lag times of various signals obtained through calibration, time-series offset compensation is applied to the pressure fluctuation signal and acoustic emission signal for corresponding durations to offset the time delay errors caused by equipment response and heat transfer. Subsequently, key feature nodes in the three types of signals are extracted uniformly, including the response peak generated by thermal disturbance and the transition nodes of energy rise and fall. To address the issue of unclear peak characteristics in actual signals caused by fluctuations and abrupt changes due to abnormal operating conditions, this solution embeds a multi-strategy peak discrimination mechanism to cover feature extraction needs under various signal patterns. When a signal has multiple peaks, the peak with the largest amplitude is selected as the main response peak, and the time position of the second highest peak is recorded simultaneously as an auxiliary reference feature. When the signal has no obvious peak, the direct peak-finding method is abandoned. Instead, a mathematical criterion combining the zero-crossing point of the first derivative of the signal with the minimum value of the second derivative is used to mark the inflection point with the largest rate of change of the signal as the equivalent feature node, replacing the traditional peak value. Subsequent timing calibration: When the peak moment is severely affected by noise, the original signal is first preprocessed with smoothing filtering to effectively suppress high-frequency random noise while preserving the overall trend of the signal. Then, the peak and feature node extraction is re-executed on the filtered clean signal to ensure the accuracy and repeatability of the feature moments. Through the above multi-strategy processing, regardless of the form of the signal, a unified and reliable feature time node can be stably extracted. Subsequently, through timing interpolation calibration, all key feature nodes are mapped to a standard time axis with the reference time as the zero point, realizing accurate time synchronization of multi-source thermal response signals.

[0069] Based on the time-synchronized multi-source response data, a thermal response migration sequence centered on the boundary position is established, and the thermal input state of the region before the boundary is correlated with the thermal response state of the region after the boundary to form a cross-regional thermal response sequence. Through spatiotemporal data reconstruction and state correlation fusion, a complete cross-regional thermal response time-series data sequence is constructed. In actual implementation, the boundary spatial position of the two performance regions of the plate is used as the core spatial reference, and the uniformly calibrated reference time is used as the time reference. The plasma spectral data, cut airflow pressure fluctuation data, and acoustic emission data of the back melt, which are time-synchronized, are structurally reorganized to establish a basic thermal response migration sequence centered on the boundary region. The spectral thermal input state before the reference time, when the laser is in the first performance region cutting stage, is used to establish the basic thermal response migration sequence centered on the boundary region. The stable airflow pressure state and the weak background acoustic emission state are defined as the thermal input state of the preceding region, fully preserving the evolution law of the early thermal accumulation and stable thermal input. The spectral abrupt change, airflow pressure fluctuation, and violent acoustic emission response of the molten material during the laser cutting stage into the second performance region after the reference time are defined as the thermal response state of the following region, accurately capturing the thermal field abrupt change and molten pool state change caused by the performance region switching. By using time-series splicing, state correlation, and coupling and binding of preceding and following conditions, the intrinsic relationship between the early thermal input and the late thermal response of the boundary is established, forming a cross-regional thermal response sequence that can fully characterize the thermal state evolution of the laser across different performance regions. This provides complete, continuous, and accurate basic data support for subsequent thermal response parameter extraction and cutting quality feature analysis.

[0070] Step S20: Based on the cross-regional thermal response sequence, detect the local maxima points of the three types of sensor signals before and after the boundary position, take the corresponding time as the peak time and the corresponding amplitude as the peak amplitude, calculate the peak time difference and amplitude difference to form thermal response migration parameters, and arrange them in order of peak time to form a cross-regional thermal history migration feature vector.

[0071] In step S20, the process of constructing the thermal response migration parameters is as follows:

[0072] Based on cross-regional thermal response sequences, local maxima of three types of sensor signals before and after the boundary are detected. The time corresponding to the maxima is taken as the peak time, and the corresponding amplitude is taken as the peak amplitude. The peak time difference and amplitude difference of each signal before and after the boundary are calculated as the basic parameters of thermal disturbance propagation timing and energy amplitude change. Taking the reference time of the boundary between the two performance regions as the center, three standardized signal sequences of plasma spectrum, airflow pressure fluctuation at the front edge of the notch, and acoustic emission response of the back melt are extracted within a preset time window of equal length before and after the boundary. An adaptive local peak finding algorithm is used to accurately extract peak features, effectively avoiding false peak interference caused by environmental noise and small signal fluctuations. For each type of sensor signal, the first peak before the boundary is independently identified. Local maxima were precisely recorded during the stable cutting phase of the performance zone and the abrupt cutting phase of the second performance zone after the boundary. The timestamps corresponding to the two sets of maxima were used as peak times, and the signal voltage / intensity values ​​were used as peak amplitudes. Subsequently, the peak time difference and peak amplitude difference before and after the boundary of the three types of signals were calculated. The time difference represents the temporal offset of the thermal disturbance response when the laser cuts across the region, and the amplitude difference represents the amplitude of signal energy fluctuation caused by the heat transfer across the region. All difference data were organized into a structured basic parameter set, which completely preserves the temporal characteristics and energy change characteristics of the thermal disturbance propagation throughout the entire process of the laser crossing the performance zone. This provides accurate, original, and quantifiable basic data support for subsequent multi-physics response coupling analysis, ensuring the completeness and accuracy of the thermal disturbance propagation law analysis.

[0073] The pressure peak migration time, pressure decay period, and duration of auxiliary airflow disturbance are extracted from the airflow pressure fluctuation signal at the cut-edge edge. The acoustic energy release amplitude, frequency variation range, and response duration are also extracted from the acoustic emission response signal of the molten material on the back side. For the pre-processed airflow pressure fluctuation signal at the cut-edge edge, the time span of the pressure peak migration from the stable position before the boundary to the abrupt change position after the boundary is tracked using the boundary position as the segmentation node, and the pressure peak migration time is accurately calculated. By fitting the complete waveform curve of the pressure peak falling back to the steady-state reference pressure value, the decay period of a single pressure signal fluctuation is calculated. Simultaneously, the duration of the pressure signal deviating from the steady-state threshold is calculated, which represents the auxiliary airflow. The duration of the disturbance fully characterizes the dynamic fluctuation characteristics of the cutting-assisted airflow affected by the cross-zone thermal field. For the acoustic emission response signal of the molten material on the back side, the time-domain signal is converted to the frequency-domain signal through fast Fourier transform, and the maximum acoustic energy release amplitude, the upper and lower limits of the effective working frequency variation, and the duration of the acoustic emission signal response are extracted to accurately reflect the vibration energy change law of the entire process of steel plate melting, flow, shedding, and solidification. Through the above multi-dimensional feature extraction, the refined state parameters of the two physical processes of airflow-assisted slag discharge and dynamic shedding of molten material are respectively mined, making up for the deficiency of the single peak difference feature dimension, and realizing the comprehensive quantification of the working condition characteristics of pressure and acoustic emission signal.

[0074] Based on the peak time difference and amplitude difference of the three types of sensor signals before and after the interface, the time delay parameters between the plasma heat input change and the cut pressure response and the melt detachment response are calculated. The plasma spectral signal with the fastest response speed and the most direct correlation with heat input is used as the heat input reference signal, and the cross-regional peak change time of this signal is used as the reference trigger time for the switching of the laser thermal field generation area. The cross-regional peak change time of the cut airflow pressure signal and the acoustic emission signal of the back melt are calculated with the reference trigger time to obtain the inherent lag time of the two types of mechanical response and material detachment response relative to the laser heat input change, which are the corresponding time delay parameters. The delay parameters are normalized and corrected for the signal delay differences under different cutting speeds and different sheet thermal properties to eliminate the calculation errors caused by process parameter fluctuations. Through this step, the temporal coupling relationship of the three types of physical behaviors, namely laser heat input, airflow pressure feedback and melt detachment, can be accurately quantified, and the propagation lag characteristics from the generation of thermal disturbance to the triggering of medium response can be clarified, providing a temporal basis for constructing a unified multi-physics thermal propagation model.

[0075] Based on the energy change trends and response time delays of each signal, a migration mapping relationship is established for the thermal disturbance propagation process from the previous performance region to the current laser action region, unifying the temporal differences and energy attenuation laws of different physical responses under the same thermal propagation path framework. First, the energy attenuation curves of the plasma spectrum, the pressure fluctuation of the gas flow at the cut, and the acoustic emission of the molten material are fitted separately to obtain the energy attenuation rates and laws of the three types of signals. Combining the temporal delay parameters obtained in the third step, a unified cross-regional thermal disturbance propagation path framework is constructed with the physical mechanism of laser cutting thermal diffusion as the core, thus unifying the thermal attenuation of the previous performance region. The accumulation and thermal disturbance generation process, the thermal response of the current laser-affected area, and the medium feedback process are coupled temporally and spatially. By combining linear correlation and nonlinear fitting, the physical dimensional differences between spectral thermal signals, pressure mechanical signals, and acoustic emission vibration signals are eliminated. The temporal lag deviation and energy attenuation ratio of different signal types are uniformly mapped to the thermal diffusion propagation logic, establishing a quantized mapping function for the migration of thermal disturbances from the preceding region to the current cutting position. Ultimately, the physical mechanism fusion of multi-source heterogeneous sensor signals is achieved, allowing response characteristics of different dimensions to uniformly characterize cross-regional thermal migration behavior, solving the technical challenge of the inability to collaboratively model multiple physical signals.

[0076] Based on the response lag, energy attenuation ratio, and response differences before and after region switching during the thermal response propagation process, thermal response migration parameters characterizing the impact of performance region switching are constructed by integrating temporal lag characteristics, energy attenuation characteristics, and cross-regional state abrupt change characteristics. The thermal response temporal lag characteristics are determined based on the aforementioned temporal delay parameters, quantifying the time lag degree of thermal disturbance propagation. The energy attenuation ratio per unit propagation distance and per unit time is calculated using the energy attenuation curves of each signal, solidifying the energy attenuation characteristics. Combining the abrupt differences in signal peak value, waveform, and energy before and after the boundary, the thermal field, flow field, and molten pool field caused by performance region switching are extracted. Cross-regional state abrupt change characteristics: By weighted fusion and dimensional compression of three core features—time lag, energy decay, and cross-regional abrupt change—redundant interference features are eliminated, and core quantitative indicators that can accurately reflect the differences in the thermal properties of the plate, the strength of cross-regional heat transfer, and the fluctuation of cutting conditions are retained. Finally, a standardized, multi-dimensional, and physically meaningful thermal response migration parameter is constructed. This parameter can comprehensively quantify the degree of comprehensive influence of switching between different performance regions on the thermal response behavior of laser cutting, providing core quantitative feature support for subsequent feature vector construction, model weight calculation, and accurate defect identification, effectively improving the comprehensiveness and accuracy of the thermal state characterization of complex gradient steel plate cutting.

[0077] In step S20, the process of forming the cross-regional thermal history migration feature vector is as follows:

[0078] Based on the peak time and amplitude difference in the thermal response migration parameters, the thermal response migration parameters before and after the boundary of different performance regions are sorted according to the peak time sequence. The entire thermal response migration parameter dataset calculated in the previous steps is retrieved. This dataset contains all quantized parameters, including peak time difference, peak amplitude difference, energy attenuation coefficient, and temporal delay, for three types of sensor signals: plasma spectrum, incision gas pressure, and molten acoustic emission, before and after the boundary of performance regions. Using the signal peak occurrence time corresponding to each feature parameter as the unique sorting key, and discarding secondary interference dimensions such as amplitude and attenuation coefficient, all thermal response migration parameters of the first performance region before the boundary and the second performance region after the boundary are globally sorted in ascending temporal order. The reorganization and sorting are strictly completed according to the actual physical time sequence of thermal disturbance occurrence, propagation, and response. During the sorting process, the parameter attributes and physical dimensions corresponding to each time node are simultaneously bound to prevent feature misalignment and temporal reversal. This ultimately forms an ordered parameter sequence that fits the actual thermal evolution time sequence of laser cross-region cutting, completely breaking the defects of traditional disordered feature stacking and ensuring that subsequent feature analysis fully conforms to the objective physical laws of thermal disturbance propagation.

[0079] Based on the sequential correspondence of peak times in the sorted thermal response migration parameters, the influence time window during the transfer of thermal accumulation state from the preceding performance region to the current cutting position is determined. Based on the time-series sorted thermal response migration parameter sequence, with the reference moment when the laser crosses the boundary of the performance regions as the zero point, and combined with process property parameters such as the gradient steel plate thermal diffusion rate, laser cutting speed, and laser power, the effective range of influence of the thermal accumulation heat in the preceding first performance region on the current cutting molten pool state is defined. By statistically analyzing the extreme duration of the preceding thermal field decaying to the steady-state threshold under historical cutting conditions, and combining the distribution range of each peak time, the start and end times of effective thermal disturbance propagation and continuous thermal accumulation are precisely locked, forming a fixed influence time window adapted to the current plate characteristics and cutting process. This time window can accurately filter out thermal history features that have a substantial impact on the current cutting quality, eliminating invalid historical thermal response data that has completely decayed and has no interference effect outside the window. This retains core thermal migration influence information while avoiding feature redundancy and increased model computation caused by redundant data, accurately defining the effective influence range of cross-regional thermal history.

[0080] Within the influence time window, the cumulative heat input, response attenuation, migration delay time, and region transition disturbance amplitude are extracted, and each type of parameter is normalized. Within the locked effective influence time window, four types of core quantification parameters are extracted comprehensively: the cumulative superposition value of plasma spectral energy within the window is calculated along the time-series integral as the cumulative heat input, characterizing the total heat input intensity of the preceding region; by comparing the attenuation ratio of the initial peak amplitude to the final steady-state amplitude within the window, the signal response attenuation is calculated, characterizing the loss characteristics of heat energy during cross-regional propagation; the time delay corresponding to each sensor signal within the window is retrieved. The delay parameter, as the migration delay time, quantifies the temporal lag characteristics of thermal disturbance propagation; the abrupt difference in signal amplitude at the moment of switching statistical performance regions is used as the region transition disturbance amplitude, characterizing the intensity of thermal field disturbance caused by the abrupt change in material thermal properties; to eliminate the feature weight imbalance problem caused by the inconsistency of the four types of parameters in terms of dimensions and the large difference in numerical magnitude, the maximum-minimum value normalization algorithm is adopted to uniformly map all parameters to the standard numerical range of [0,1]. Under the premise of fully preserving the relative change characteristics and physical correlation of each parameter, the data standardization processing of multi-dimensional heterogeneous parameters is realized to meet the input data requirements of deep learning models.

[0081] The normalized thermal accumulation parameters, migration delay parameters, energy decay parameters, and boundary disturbance parameters are arranged in chronological order of peak time to form a cross-regional thermal history migration feature vector. Using the peak time sequence within the influence time window as the arrangement logic, the normalized thermal input accumulation parameters, migration delay time parameters, response energy decay parameters, and region transformation disturbance amplitude parameters are sequentially concatenated. The dimensions are arranged according to the temporal evolution order of thermal disturbance occurrence, propagation, decay, and abrupt change, constructing a one-dimensional structured feature vector with fixed dimensions, ordered temporal sequence, and clear physical feature hierarchy. This vector comprehensively encompasses the thermal history accumulation effect, thermal propagation lag effect, energy decay law, and regional abrupt change disturbance characteristics of the entire laser cross-performance zone cutting process. It realizes the transformation of heat dissipation response parameters into a standardized, model-recognizable quantified feature vector, accurately representing the dynamic migration influence of the preceding regional thermal history on the current cutting state. This provides a regular, accurate, and comprehensive pre-feature input for subsequent deep learning model weight calculation and quality discrimination feature fusion.

[0082] Step S30: Input the cross-regional thermal history migration feature vector into the deep learning inference model, calculate the weights based on the ratio of thermal diffusivity coefficients of adjacent performance regions in an inverse relationship, multiply the weights by the output values ​​of the previous region feature nodes as the influence parameters of the previous thermal response on the current position, and combine the molten pool propulsion speed, molten material discharge energy and cut solidification shrinkage displacement parameters to construct the cutting quality discrimination feature vector.

[0083] In step S30, the process of using the weight multiplied by the output value of the previous region feature nodes as the influence parameter of the previous thermal response on the current position is as follows:

[0084] The cross-regional thermal history migration feature vector is input into a pre-constructed deep learning inference model. Based on the positional relationship of the laser along the cutting trajectory, the thermal response features corresponding to different performance regions are assigned region numbers, establishing the feature association between the previous performance region and the current laser action region. The aforementioned normalized and temporally sorted standard cross-regional thermal history migration feature vectors are batch-imported into the training converged and parameter-fixed deep learning inference model. This model pre-builds a region feature mapping layer and a temporal association layer to adapt to multi-dimensional thermal migration feature input. Combining the laser head trajectory coordinates, travel direction, and movement speed fed back by the real-time encoder of the laser cutting equipment, the first performance region and the second performance region on the preset cutting trajectory of the steel plate are matched. The spatial distribution coordinates of the energy region are used to bind each segment of the temporal feature component in the feature vector to its respective physical region. A sequential coding method is used to independently number the two types of performance regions, defining the region that the laser passes through first as the preceding region and the real-time action area of ​​the current spot as the current region, thus completing the regional attribute labeling of all thermal response feature components. A one-to-one mapping relationship between the historical thermal feature nodes of the preceding region and the real-time feature nodes of the current region is established through the model feature correlation matrix, clarifying the correlation path of the temporal features of the preceding thermal field to the current cutting position, and completely solving the problems of ambiguous cross-region feature region attribution and the inability to quantify the coupling relationship of the preceding and following thermal effects. This provides an accurate feature correlation topology foundation for subsequent regional weight calculation and contribution inference.

[0085] The thermal diffusivity between the first and second performance zones is obtained, and weights are calculated based on the inverse relationship of the ratio of thermal diffusivity of adjacent performance zones. The thermal property parameters of the entire steel plate cutting trajectory are retrieved from pre-detection and calibration, and the thermal diffusivity of the first and second performance zones is accurately extracted. Both coefficients are obtained through actual measurement using the laser flare method, which can truly reflect the heat transfer and diffusion capabilities of steel in different regions. Based on the physical mechanism of heat conduction, the larger the thermal diffusivity of a performance zone, the easier it is for the heat field generated in that zone to diffuse to subsequent cutting positions, and the stronger the residual heat accumulation effect. Therefore, the inverse relationship of the thermal diffusivity of adjacent regions is used to calculate the regional correlation weights. Specifically, the ratio of the thermal diffusivity of the preceding region to the thermal diffusivity of the current region is used as the core basis for weight calculation, automatically adapting to the differences in cross-zone thermal influence of steel plates with different gradient materials and different microstructures. Through this inverse weighting method, the preceding region with stronger thermal diffusivity corresponds to a larger thermal influence weight, accurately conforming to the actual cross-zone heat conduction law of laser cutting, overcoming the shortcomings of traditional fixed weights that cannot adapt to material differences, and ensuring the physical authenticity and working condition adaptability of the weight parameters.

[0086] Based on the calculated weights, the correlation strength between feature nodes corresponding to different performance regions in the deep learning inference model is corrected. The output value of the preceding region feature node is multiplied by this weight as the contribution of the preceding thermal response to the current position. The dynamic region weights calculated in the second step are loaded into the region feature connection layer of the deep learning inference model. The initial fixed correlation strength between the preceding performance region feature nodes and the current performance region feature nodes is iteratively corrected, replacing the model's default equal correlation weights. During the model's forward inference process, the output feature values ​​of each temporal feature node in the preceding region are extracted separately. Each set of node output values ​​is multiplied with the corresponding dynamic weight one by one to obtain the contribution of the preceding thermal field to the current laser cutting position in a single temporal dimension. Each contribution strictly corresponds to the characteristics of thermal accumulation, thermal delay, and energy decay in different thermal migration stages. The dynamic interference degree of the residual thermal field in the preceding region on the current molten pool melting, material gasification, and slag flow is accurately quantified, so that the contribution of each feature in the model fits the real physical process of heat transfer, effectively improving the model's quantitative characterization accuracy of cross-regional thermal effects.

[0087] The thermal response contribution amounts within different time windows are weighted and fused to obtain thermal response influence parameters characterizing the degree of thermal history influence of the preceding performance region. Following the temporal evolution logic of laser cross-regional cutting, the complete cross-regional thermal influence process is divided into multiple continuous and non-overlapping time sub-windows, each corresponding to an independent preceding thermal response contribution amount. For the independent contribution amounts within all sub-windows, a temporal attenuation weighting coefficient is set based on the time distance between each time window and the current laser action moment. The closer the time window is to the current moment, the stronger the residual thermal influence, and the larger the corresponding weighting coefficient; conversely, the coefficient decreases as the time window moves further away. The contribution amounts of all sub-windows are fused and calculated using a temporal weighted summation method to eliminate errors caused by the one-sidedness of a single time window feature, ultimately outputting a unique and standardized thermal response influence parameter. This parameter can globally and completely characterize the combined influence of the cumulative effect of the preceding performance region's thermal history, the cross-regional thermal migration attenuation effect, and the temporal lag effect on the current cutting state, providing accurate and reliable core input parameters for subsequent multi-dimensional quality discrimination feature fusion.

[0088] In step S30, the process of constructing the cutting quality discrimination feature vector is as follows:

[0089] Based on the influence parameters of the preceding region's thermal response on the current laser irradiation position, the process of molten pool thermal state change during laser irradiation is obtained, and the spatial correspondence between the leading edge advancement position and the trailing edge solidification position of the molten pool is determined according to the laser movement direction. Using the standardized thermal response influence parameters calculated in the preceding steps as core input data, these parameters quantify the intensity of the continuous thermal interference to the current cutting position caused by historical heat accumulation after switching between different performance regions. Combined with steady-state process parameters such as real-time laser output power, cutting speed, and auxiliary gas pressure, the evolution law of the molten pool thermal state of the steel plate substrate during the entire process of heating, melting, heat storage, and heat dissipation under continuous laser heat input is simultaneously analyzed. Real-time acquisition of the cutting data is achieved using high-speed visual imaging equipment. By combining high-resolution images of the molten pool morphology with the travel coordinates and movement direction of the laser head encoder, a dynamic spatial coordinate system for the molten pool is established. The center of the laser spot is set as the reference origin. The melting and propagation area at the leading edge of the molten pool is defined along the direction of the cutting trajectory, and the cooling and solidification area at the trailing edge of the molten pool is defined along the direction of the reverse extension of the cutting trajectory. Through image pixel calibration and scale conversion, the position of the farthest melting boundary at the leading edge of the molten pool and the position of the latest solidification boundary at the trailing edge are locked in real time at each moment. A dynamic spatiotemporal correspondence between the positions of the leading and trailing edges as the laser travels and the thermal field fluctuates across the region is established, accurately distinguishing the active melting area and the solidification area. This provides a spatial reference and state basis for the accurate calculation of subsequent dynamic parameters of the molten pool and the morphology parameters of the cut.

[0090] Information on the boundary changes of the molten pool during laser cutting is collected, and parameters such as the molten pool's advancing speed along the cutting trajectory, the rate of change of advancing speed, and the molten pool's stable duration are calculated. A millisecond-level high-frequency molten pool visual acquisition scheme is used to continuously acquire a sequence of molten pool boundary contour images throughout the entire cross-region cutting process. The coordinates of the leading edge boundary of the molten pool in each frame are accurately extracted through threshold segmentation, edge extraction, and denoising fitting algorithms. Combining the camera sampling frame rate and pixel physical scale calibration results, the instantaneous advancing speed of the molten pool along the cutting trajectory is calculated in real time by the ratio of the difference in the leading edge displacement of the molten pool in adjacent frames to the time interval. The advancing speed data in the continuous time sequence are differentiated to obtain the rate of change of advancing speed, which is used to characterize the dynamic fluctuations and instability of the molten pool. The duration for which the molten pool boundary contour does not shrink or expand drastically and maintains a continuous and uniform shape is used as the molten pool stable duration parameter. Among them, the advancing speed reflects the efficiency of molten pool melting and following, the rate of change of speed directly corresponds to the degree of molten pool oscillation caused by thermal field abrupt changes, and the stable duration characterizes the stability of the cutting conditions. These three parameters comprehensively quantify the dynamic behavior characteristics of the molten pool under cross-region thermal migration interference.

[0091] Based on the energy release process corresponding to the acoustic emission response signal of the molten material on the back side, the peak acoustic energy, energy decay rate, and continuous discharge time during the process of the molten material detaching from the steel plate are extracted and converted into discharge energy parameters characterizing the discharge capacity of the molten material. The acoustic emission response signal of the molten material on the back side of the steel plate, which is collected and synchronized throughout the entire process, is retrieved. Through a combination of time-domain energy integration and frequency-domain spectral analysis, the peak acoustic energy of the entire process—molten metal melting and flow, wall sliding, detachment from the plate, and impact discharge—is extracted to characterize the instantaneous slag discharge energy intensity. By fitting the decay curve of the peak acoustic energy falling back to the steady-state background noise, the acoustic energy decay amplitude per unit time is calculated to obtain... The energy decay rate reflects the energy dissipation rate and cooling efficiency of the molten pool after the molten material is discharged. The duration for which the effective acoustic emission signal remains above the threshold is statistically analyzed as the continuous discharge time of the molten material, characterizing the continuity and stability of the slag discharge process. Based on three fundamental characteristics—sound energy peak, energy decay rate, and continuous discharge time—and combined with physical properties and process parameters such as steel plate material density, molten metal viscosity, and auxiliary gas purging pressure, a quantitative calculation formula for molten material discharge energy is constructed. This transforms discrete acoustic emission characteristics into molten material discharge energy parameters with clear physical meaning that can directly characterize the smoothness of slag discharge, accurately quantifying the quality of slag discharge capacity under cross-zone thermal disturbance.

[0092] This study acquires the edge shrinkage variation, kerf width variation trend, and solidification shrinkage displacement parameters during the cooling and solidification process of the cut area, establishing a correlation between heat input changes and kerf morphology evolution. During laser cutting, a high-precision side-mounted line scan camera collects real-time contour morphology data of the kerf forming area, continuously tracking the forming position and width of the left and right edges of the kerf. Using the standard kerf width in a stable region as a benchmark, the study calculates the inward or outward expansion offset of the kerf edge during cross-region cutting, obtaining the edge shrinkage variation. Finally, it fits the kerf width data over continuous time. The curves were analyzed to understand the fluctuations in cut width before and after the switching of the cross-regional thermal field, and to obtain the trend of cut width change. Combined with the cooling and solidification time sequence nodes of the molten pool, the overall deformation displacement during the solidification process of the molten metal was statistically analyzed to obtain the solidification shrinkage displacement parameters. The heat input strength, heat accumulation degree, and heat mutation amplitude corresponding to the aforementioned thermal response influence parameters were coupled and correlated with the cut shrinkage, width fluctuation, and solidification deformation data in a time sequence to establish a one-to-one correspondence mapping relationship between the dynamic change of heat input and the deformation of cut morphology and the evolution of forming accuracy, so as to accurately characterize the influence mechanism of cross-regional thermal history interference on the final cut forming quality.

[0093] Combining the parameters of the influence of the preceding thermal response on the current position, the parameters of the molten pool advance speed, the parameters of the molten material discharge energy, and the parameters of the solidification shrinkage displacement of the cut, feature fusion is performed according to the thermal state influence relationship to construct a cutting quality discrimination feature vector for quality judgment. The parameters of the preceding thermal response are used as the underlying core thermal state features, which dominate the basic evolution law of the molten pool thermal field. The dynamic features of the molten pool composed of the molten pool advance speed, the rate of change of the advance speed, and the duration of molten pool stability are successively fused. The slag discharge flow features corresponding to the molten material discharge energy parameters are fused. The forming morphology features corresponding to the shrinkage amount, width change trend, and solidification shrinkage displacement of the cut edge are fused. Following the physical evolution logic of laser cutting—"thermal field input—molten pool response—slag discharge flow—solidification and forming"—multi-dimensional heterogeneous features are systematically spliced ​​and dimensionally fused. Secondary equalization corrections are made for feature parameters of different magnitudes and dimensions, preserving the independent physical meaning and relative change characteristics of each dimension. Finally, a cutting quality discrimination feature vector is formed that is dimensionally regular, hierarchically clear, and covers thermal response, molten pool dynamics, slag discharge characteristics, and cut forming information under all working conditions. This feature vector fully carries the quality disturbance information brought about by cross-regional thermal migration, providing a comprehensive, reliable, and physically highly correlated standardized feature input for accurate subsequent defect classification.

[0094] Step S40: Input the cutting quality discrimination feature vector into the quality discrimination model, extract the thermal accumulation parameter and the thermal response mutation parameter of the boundary region, input the two types of parameters into the fully connected layer of the quality discrimination model, output the probability of each defect category through nonlinear mapping, and determine the defect discrimination result with the maximum probability value.

[0095] In step S40, the process of extracting the thermal accumulation parameters and the thermal response abrupt change parameters of the boundary region is as follows:

[0096] The cutting quality discrimination feature vector is input into the quality discrimination model. Based on the associated thermal response feature components of each spatial location in the quality discrimination feature vector, the product of the corresponding thermal input intensity and duration of action is accumulated along the cutting trajectory. The standardized cutting quality discrimination feature vector, which includes multi-dimensional information such as thermal field response, molten pool dynamics, slag discharge characteristics, and cut morphology, is then fully imported into the pre-trained quality discrimination model input layer. The feature components of each dimension within this feature vector have been bound to spatial coordinates and temporal information, with each set of feature components corresponding to a unique spatial action point and corresponding action time period on the laser cutting trajectory. Within the model, following the actual cutting trajectory of the laser, the thermal input intensity parameters and corresponding laser duration at each spatially discrete location along the trajectory are traversed point by point. The product of the thermal input intensity and duration at each single point is calculated segment by segment using a numerical integration accumulation method. This continuous accumulation yields real-time data on the dynamic superimposed total thermal input along the cutting path. Through this spatial point-by-point accumulation calculation method, the dynamic thermal input superposition pattern of the entire laser cutting process can be completely restored, avoiding the problems of discontinuous thermal accumulation calculation and loss of local thermal accumulation features caused by traditional fixed-point sampling. This provides continuous and accurate raw accumulation calculation data for subsequent regional thermal accumulation statistics.

[0097] The total energy superposition and heat diffusion coverage of the first and second performance zones are statistically analyzed to obtain the heat accumulation parameters during the cross-performance zone migration process. Based on the spatial boundary coordinates of the first and second performance zones pre-divided in the steel plate, the continuous heat input total data accumulated along the cutting trajectory in the first step is divided and classified into regions, and the trajectory heat accumulation data is matched to the performance zone range of the corresponding physical property. For each performance zone, the total heat input energy superposition of all spatial points within the zone is statistically analyzed to characterize the overall heat storage level of the corresponding region during the cutting process. At the same time, combined with the heat diffusion characteristics and thermal field distribution law of the plate, the heat diffusion spatial coverage of each type of performance zone is statistically obtained according to the heat input attenuation threshold boundary at each location, which accurately reflects the diffusion penetration ability and accumulation range of heat in different physical property regions. The total energy superposition and heat diffusion coverage of each region are coupled and integrated to construct a standardized heat accumulation parameter that can quantify the characteristics of cross-performance zone heat migration and accumulation, regional heat storage differences, and uneven heat diffusion. This parameter can intuitively characterize the steady-state influence effect of the heat history accumulation of the preceding and following regions on the cutting conditions during laser cross-zone cutting.

[0098] For the boundary between the first and second performance regions, multi-source sensing feature components corresponding to the boundary before and after the boundary are extracted from the mass discrimination feature vector. The rate of change of plasma spectral energy, the gradient of change of airflow pressure, and the difference in the rate of change of acoustic emission energy before and after the boundary are calculated respectively. Taking the spatial location of the boundary between the two performance regions as the core segmentation node, the three types of multi-source sensing feature components corresponding to the pre-set stable interval before the boundary and the abrupt interval after the boundary are accurately extracted from the mass discrimination feature vector. For the plasma spectral characteristics, the steady-state stage before the boundary and the abrupt interval after the boundary are solved by performing a first-order difference operation on the time-series energy curve. The rate of change of spectral energy over time during the abrupt change phase is calculated, and the rate difference between the two phases is used to characterize the magnitude of the thermal input abrupt change. For the airflow pressure characteristics at the cut, the difference in the gradient of the pressure signal changes on the spatial trajectory before and after the boundary is obtained, reflecting the difference in pressure disturbance caused by the thermal field abrupt change in the auxiliary airflow. For the acoustic emission characteristics of the molten material on the back side, the difference in the rate of change of acoustic emission energy before and after the boundary is calculated using the same differential algorithm, characterizing the degree of abrupt change in the vibration and shedding state of the molten material. Through the multi-dimensional rate difference calculation of the three types of signals, the instantaneous abrupt change characteristics of thermal input, airflow field, and molten pool shedding behavior at the moment of performance zone switching are captured in all aspects, providing refined differential basic data for constructing thermal response abrupt change parameters.

[0099] By combining the peak response offset before and after the boundary and the degree of synchronous change of multi-source signals, a thermal response mutation parameter is constructed to characterize the degree of thermal state mutation caused by the performance region switching. Based on the rate difference of the three types of sensor signals before and after the boundary obtained in the third step, the peak response time offset and peak amplitude offset of the multi-source signals are further extracted to quantify the degree of thermal disturbance response time sequence offset and energy offset. By calculating the correlation coefficient of the mutation trend of plasma spectrum, airflow pressure, and acoustic emission signal, the degree of synchronous change of multi-source signals during the region switching process is evaluated to eliminate the misjudgment interference caused by abnormal fluctuations of a single signal. The three mutation features of signal rate difference, peak response offset, and multi-signal synchronization are weighted, fused, and normalized to form a unified core index that can comprehensively characterize the thermal field mutation, molten pool state mutation, and medium response mutation caused by the thermal property switching of the performance region, and finally form a standardized thermal response mutation parameter. This parameter can accurately characterize the severity of the instantaneous thermal state mutation during laser cross-region cutting, effectively distinguish the quality difference between steady-state cutting conditions and cross-region mutation cutting conditions, and provide key instantaneous mutation feature support for subsequent accurate defect probability identification.

[0100] In step S40, the process of determining the defect discrimination result based on the maximum probability value is as follows:

[0101] The accumulated thermal parameters and the abrupt thermal response parameters of the boundary region are input into the fully connected layer of the quality discrimination model, and the probability of each defect category is output through nonlinear mapping. The previously extracted steady-state quantification features (i.e., accumulated thermal parameters) and instantaneous abrupt change quantification features (i.e., abrupt thermal response parameters of the boundary region) are concatenated to form a dual-core defect discrimination input feature set, which is then completely input into the fully connected layer of the trained quality discrimination model. The fully connected layer is pre-configured with weight matrices and bias parameters adapted to the two parameter dimensions, enabling deep fusion computation of cross-regional accumulated steady-state features and regional switching abrupt thermal change instantaneous features. The model uses ReLU nonlinear mapping. The activation function performs a high-dimensional nonlinear mapping transformation on the fused features, breaking the limitations of linear feature correlation and fully fitting the nonlinear correspondence between excessive heat accumulation, thermal field abrupt imbalance and various laser cutting defects. After multi-layer weighted iterative deduction and normalized probability mapping processing, the model finally outputs independent predicted probability values ​​for four preset defect categories: incomplete cutting defects, slag defects, burr defects and burnt edge defects. The interval of each probability value is uniformly constrained to the range of 0 to 1, and the normalized sum of the four probability values ​​is 1, ensuring the validity, comparability and uniqueness of the output results, and providing accurate probability quantification basis for subsequent defect category determination.

[0102] Based on the continuously changing heat accumulation parameters along the cutting trajectory and the abrupt change parameters of the thermal response in the interface region, the heat input state, molten material discharge state, and solidification shrinkage trend of the current laser-affected area are analyzed to determine the corresponding quality state characteristics. Relying on the time-series continuously changing heat accumulation parameters along the cutting trajectory, the long-term heat input accumulation degree, heat diffusion coverage, and heat storage steady-state level of the current laser-affected area are analyzed in real time to accurately determine whether there are abnormal conditions such as insufficient laser heat input, excessive heat accumulation, or uneven heat diffusion. Combined with the dynamically changing thermal response abrupt change parameters at the interface position, the intensity of thermal field disturbances during performance region switching is captured, and the real-time evolution of the instantaneous melting stability of the molten pool, the smoothness of dynamic molten material discharge, and the cooling solidification shrinkage characteristics of the cut are analyzed in conjunction with these parameters. Through the time-series coupling analysis of the heat accumulation steady-state characteristics and the instantaneous thermal abrupt change characteristics, the real-time quality state characteristics corresponding to the current cutting position are extracted comprehensively from three core dimensions: heat input matching, slag discharge smoothness, and cut solidification and forming stability. This enables pre-analysis of defect causes, ensuring that the model output probability fully matches the actual physical cutting conditions and avoiding the problem of judgment results deviating from the process mechanism due to purely data-driven approaches.

[0103] The output probabilities of the fully connected layer corresponding to the quality status characteristics are compared. The defect status of incomplete cutting, slag, burrs, or burnt edges corresponding to the current cutting area is determined by the maximum probability value, and the quality judgment result is output. Simultaneously, the real-time predicted probability values ​​of the four types of defects output by the fully connected layer are retrieved to establish a defect probability comparison and judgment mechanism. The probability values ​​of the four types of defects (incomplete cutting, slag, burrs, and burnt edges) are horizontally traversed, compared, and sorted by size. The defect category corresponding to the highest probability value is selected as the final defect judgment result of the current laser cutting area. If the probability of each type of defect is lower than the preset qualified threshold, the current cutting area is judged to be in a qualified and defect-free state. While completing the defect judgment, the thermal accumulation state, thermal mutation degree, molten pool condition, slag discharge and solidification state information corresponding to the current frame are simultaneously bound to form a quality judgment result with process traceability information and output it to the outside in real time. This maximum probability selection judgment mechanism can adapt to the dynamic fluctuation scenario of cross-region cutting thermal state, effectively distinguish the differences in the characteristic causes of different defects, and greatly improve the accuracy and stability of cutting defect identification under complex working conditions such as gradient performance steel plates and multi-structure interface areas, so as to realize online accurate judgment of laser cutting quality.

[0104] Step S50: Based on the quality discrimination results and the thermal response change trend corresponding to the cross-regional thermal history migration feature vector, analyze the degree of influence of each thermal migration feature on the discrimination results and calculate the contribution correction amount. Update the thermal migration feature weights in the discrimination model accordingly to form a weight update relationship based on defect feedback constraints.

[0105] In step S50, the process of forming the weight update relationship based on defect feedback constraints is as follows:

[0106] Based on the defect category results output by the quality discrimination model, the cross-regional thermal history migration feature vectors participating in the discrimination calculation under the corresponding defect category are obtained. The fluctuation amplitude, offset direction, and peak time migration position of each peak time difference and amplitude difference in the defect occurrence area are extracted. After the quality discrimination model completes the defect category determination of a single frame segment, the spatiotemporal segment corresponding to the current output defect category is locked in real time, and the original cross-regional thermal history migration feature vectors participating in the model inference calculation within the defect segment are accurately retrieved to prevent interference from non-defect area feature data. For the various core components representing the propagation of thermal disturbances within the feature vector, the plasma spectrum, airflow pressure, and acoustic emission signals are analyzed one by one. The peak time difference and peak amplitude difference parameters are used to accurately extract the fluctuation amplitude of each feature parameter relative to the normal steady-state cutting condition within the defect occurrence interval through time-series differential calculation and baseline comparison. Combined with the feature time-series evolution curve, the parameter offset direction is determined to clarify whether the feature parameter is positively large or negatively small. At the same time, the overall migration position of the thermal response peak on the cross-zone cutting time-series axis is tracked to determine whether the peak is advanced, lagging, or offset within the interval. By comprehensively extracting the fluctuation amplitude, offset trend, and time-series migration position information of the defect interval features, the abnormal source of thermal migration features that cause defects is accurately located, providing accurate defect-related feature data for subsequent feature influence relationship modeling and error analysis.

[0107] Based on the correlation between thermal accumulation parameters, thermal response mutation parameters, and peak time differences and amplitude differences under different defect categories, an influence mapping relationship between each feature component in the cross-regional thermal history migration feature vector and the defect category is established. Using batch historical cutting sample data, the threshold distribution of thermal accumulation parameters, the fluctuation range of thermal response mutation parameters, and the characteristic variation patterns of peak time differences and amplitude differences are statistically analyzed under four typical defect conditions: incomplete cutting, slag adhesion, burrs, and burnt edges. For each type of defect, correlation analysis and gray-scale correlation algorithms are used to quantify and solve for excessive or insufficient thermal accumulation and the severity of thermal field mutations. The coupling correlation between the anomalies of each thermal migration feature component and the induction contribution of different feature components to different defects is clarified. A multi-dimensional feature-defect mapping matrix is ​​established to form a one-to-one quantitative correlation between the abnormal states of different feature dimensions, such as peak time lag, amplitude attenuation anomaly, cross-regional disturbance abrupt change, and thermal accumulation imbalance, and four types of cutting defects, clarifying the core dominant features and secondary influencing features of each type of defect. This mapping relationship can accurately distinguish the defect types corresponding to different thermal migration anomalies, solve the problem that a single feature deviation cannot distinguish the defect induction cause, and provide mechanistic support for subsequent residual allocation and weight correction.

[0108] Based on the category deviation between the current discrimination result and the actual quality inspection result, the influence of each peak time difference and amplitude difference on the discrimination result is analyzed, and the discrimination residual is allocated according to the influence ratio of each feature component on the defect classification result to obtain the error contribution of each feature component; the predicted defect category output by the model in real time is compared with the real quality label obtained by offline manual quality inspection and machine vision re-inspection of steel plates to accurately determine whether there are misjudgments, omissions, and category misclassification problems in the current discrimination result, and the overall classification discrimination residual of the model in a single inference is calculated; based on the features established in the second step... - The defect impact mapping relationship is analyzed quantitatively to determine the weight and influence of each peak time difference and amplitude difference feature component on the current discrimination result, distinguishing between high-contribution core features and low-contribution redundant features. The overall discrimination residual is proportionally split and allocated according to the influence ratio of each feature component, decomposing the global error layer by layer to each heat migration feature dimension, and accurately calculating the independent error contribution corresponding to each peak time difference and amplitude difference. Through the refined residual allocation mechanism, the model error source is traced, and the specific feature dimension that causes the discrimination deviation is accurately located, providing an error basis for subsequent targeted weight correction.

[0109] Based on the error contribution, the fluctuation amplitude of each peak time difference and amplitude difference, and the historical discrimination error trend, the contribution correction of each feature component is calculated, and the weight adjustment coefficients corresponding to each peak time difference and amplitude difference are updated according to the contribution correction. Using the real-time error contribution of each feature component as the core correction basis, combined with the fluctuation amplitude of the corresponding peak parameters within the defect interval, and simultaneously retrieving the convergence or divergence trend of historical discrimination errors in the near-preset batches, a multi-factor contribution correction calculation model is comprehensively constructed. For cases with large error contributions, drastic feature fluctuations, and persistent historical errors... For divergent feature dimensions, the correction amount is increased to achieve targeted correction. For feature dimensions with small error contribution, stable features, and convergent errors, the correction amount is reduced to maintain the steady state of the model. The independent contribution correction amount corresponding to the peak time difference and amplitude difference features of each type is obtained through iterative calculation. The correction amount is superimposed on the original weight adjustment coefficient of the model to complete the dynamic iterative update of the weight coefficients of each heat migration feature dimension. This multi-factor correction method can take into account both instantaneous defect error and long-term model error evolution law, avoid weight overfitting oscillation caused by a single abnormal sample, and improve the stability of model adaptive optimization.

[0110] The updated weight adjustment coefficients are loaded into the thermal history migration feature mapping layer in the quality discrimination model, and the weight values ​​corresponding to each feature dimension in the cross-regional thermal history migration feature vector are recalculated to form a thermal migration feature weight update relationship based on defect feedback constraints. The weight adjustment coefficients of each dimension updated in the fourth step are loaded in batches into the thermal history migration feature mapping layer inside the quality discrimination model to replace the original fixed weight parameters of the model. In the forward inference logic of the model, based on the new weight adjustment coefficients, the weighted fusion calculation of each dimension parameter of the cross-regional thermal history migration feature vector is re-completed to update the weight ratio of each thermal migration feature on the quality discrimination result. Through the closed loop of defect discrimination error reverse constraint and feature weight forward iterative update, a dynamic mapping update mechanism of thermal migration feature anomaly - defect generation - error feedback - weight correction is established, and finally a stable, adaptive, and real-time iterative thermal migration feature weight update relationship based on defect feedback constraints is formed. This enables the model to continuously adapt to the changes in thermal properties of different plates, different cutting process parameters, and different cross-regional thermal migration conditions, continuously reduce discrimination residuals, and continuously improve the accuracy and generalization ability of online laser cutting defect discrimination under complex conditions.

[0111] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for online quality assessment of laser-cut steel plates based on deep learning, characterized in that: Includes the following steps: The plasma spectrum of the laser-acting area, the airflow pressure fluctuation at the front edge of the cut, and the acoustic emission response signal of the melt on the back side were collected. Based on the difference in thermal diffusivity and material properties, the first and second performance zones and the boundary position were divided. The laser passing through the boundary position was used as the time reference, and the peak response time of the three types of sensor signals was used as the alignment anchor point and propagation delay compensation was performed. The signals were mapped to a unified time coordinate to construct a cross-regional thermal response sequence. Based on the cross-regional thermal response sequence, the local maxima of three types of sensor signals before and after the boundary position are detected respectively. The corresponding time is taken as the peak time and the corresponding amplitude is the peak amplitude. The peak time difference and amplitude difference are calculated to form thermal response migration parameters. They are arranged in order of peak time to form a cross-regional thermal history migration feature vector. The cross-regional thermal history migration feature vector is input into the deep learning inference model. The weights are calculated inversely based on the ratio of the thermal diffusivity coefficients of adjacent performance regions. The weights are multiplied by the output values ​​of the feature nodes of the preceding regions as the parameters of the influence of the preceding thermal response on the current position. The feature vector for judging cutting quality is constructed by combining the molten pool propulsion speed, the energy of molten material discharge, and the solidification shrinkage displacement parameters of the cut. The cutting quality discrimination feature vector is input into the quality discrimination model to extract the thermal accumulation parameter and the thermal response mutation parameter of the boundary region. The two types of parameters are input into the fully connected layer of the quality discrimination model, and the probability of each defect category is output through nonlinear mapping. The defect discrimination result is determined by the maximum probability value. Based on the quality discrimination results and the thermal response change trend corresponding to the cross-regional thermal history migration feature vector, the influence of each thermal migration feature on the discrimination results is analyzed and the contribution correction amount is calculated. Accordingly, the weights of thermal migration features in the discrimination model are updated to form a weight update relationship based on defect feedback constraints.

2. The method for online quality judgment of laser-cut steel plates based on deep learning according to claim 1, characterized in that, The process of constructing a cross-regional thermal response sequence is as follows: The material structure parameters, thermal diffusivity, specific heat capacity and thermal conductivity parameters of the steel plate to be cut are obtained along the cutting trajectory. Based on the difference in thermal diffusivity at different locations, the steel plate is divided into a first performance zone and a second performance zone with different thermal response characteristics. Based on the correspondence between the real-time movement position of the laser head and the spatial coordinates of the steel plate, the spatial transformation position of the laser action point entering the first performance zone and the second performance zone is determined, and the moment when the laser beam passes through the boundary between the first performance zone and the second performance zone is marked as the time reference moment. The plasma spectral change sequence, the gas flow pressure fluctuation sequence at the cutting edge, and the acoustic emission response sequence of the back melt were obtained within a preset time window before and after the boundary position. Using the peak response times of the three types of sensor signals as alignment anchors, propagation delay compensation is performed on different sensor signals to map the peak values ​​of thermal disturbance response, energy change nodes, and abnormal fluctuation locations in each signal to a unified time coordinate. Based on the multi-source response data after time synchronization, a thermal response migration sequence centered on the boundary location is established, and the thermal input state of the region before the boundary is correlated with the thermal response state of the region after the boundary to form a cross-regional thermal response sequence.

3. The method for online quality judgment of laser-cut steel plates based on deep learning according to claim 1, characterized in that, The process of constructing thermal response migration parameters is as follows: Based on the cross-regional thermal response sequence, the local maxima of three types of sensor signals before and after the boundary are detected respectively. The time corresponding to the maximum value is taken as the peak time and the corresponding amplitude is taken as the peak amplitude. The peak time difference and amplitude difference of each signal before and after the boundary are calculated as the basic parameters of thermal disturbance propagation time sequence and energy amplitude change. Extract the pressure peak migration time, pressure decay period, and duration of auxiliary airflow disturbance from the airflow pressure fluctuation signal at the cut front edge, and extract the acoustic energy release amplitude, frequency variation range, and response duration from the acoustic emission response signal of the back melt. Based on the peak time difference and amplitude difference of the three types of sensor signals before and after the junction, the time delay parameters between the plasma heat input change and the cut pressure response and the melt release response are calculated. Based on the energy change trends and response time delay relationships of each signal, a migration mapping relationship is established during the propagation of thermal disturbance from the previous performance region to the current laser action region, unifying the temporal differences and energy attenuation laws of different physical responses under the same thermal propagation path framework; Based on the response lag, energy attenuation ratio, and response differences before and after regional conversion during the thermal response propagation process, thermal response migration parameters characterizing the impact of performance regional conversion are constructed by associating and integrating temporal lag characteristics, energy attenuation characteristics, and cross-regional state change characteristics.

4. The method for online quality judgment of laser-cut steel plates based on deep learning according to claim 1, characterized in that, The process of forming the cross-regional thermal history migration feature vector is as follows: Based on the peak time and amplitude difference in the thermal response migration parameters, the thermal response migration parameters before and after the boundary of different performance regions are sorted according to the order of peak time. Based on the sequential correspondence of peak times in the sorted thermal response migration parameters, the influence time window during the transfer of thermal accumulation state of the preceding performance region to the current cutting position is determined. Within the time window of influence, the cumulative amount of heat input, the degree of response attenuation, the migration delay time, and the amplitude of the region conversion disturbance are extracted, and the parameters of each type are normalized. The normalized thermal accumulation parameters, migration delay parameters, energy decay parameters, and boundary disturbance parameters are arranged in chronological order of peak time to form a cross-regional thermal history migration feature vector.

5. The method for online quality judgment of laser-cut steel plates based on deep learning according to claim 1, characterized in that, The process of multiplying the weights by the output values ​​of the preceding region feature nodes as the parameter for the influence of the preceding thermal response on the current location is as follows: The cross-regional thermal history migration feature vector is input into a pre-built deep learning inference model. Based on the movement position relationship of the laser along the cutting trajectory, the thermal response features corresponding to different performance regions are assigned region numbers, and the feature association relationship between the previous performance region and the current laser action region is established. Obtain the thermal diffusivity between the first performance region and the second performance region, and calculate the weights based on the inverse relationship of the ratio of the thermal diffusivity of adjacent performance regions. Based on the calculated weights, the correlation strength between feature nodes in different performance regions in the deep learning inference model is corrected, and the output value of the preceding region feature node is multiplied by the weight as the contribution of the preceding hot response to the current position. The contribution of preceding thermal response within different time windows is weighted and fused to obtain thermal response influence parameters that characterize the degree of influence of preceding performance region thermal history.

6. The method for online quality judgment of laser-cut steel plates based on deep learning according to claim 1, characterized in that, The process of constructing the feature vector for judging cutting quality is as follows: Based on the influence parameters of the thermal response of the preceding region on the current laser action position, the thermal state change process of the molten pool during laser action is obtained, and the spatial correspondence between the advancing position of the leading edge of the molten pool and the solidification position of the trailing edge is determined according to the laser movement direction. Collect information on the boundary changes of the molten pool during the laser cutting process, and calculate the parameters of the molten pool's advance speed along the cutting trajectory, the rate of change of the advance speed, and the duration of molten pool stability. Based on the energy release process corresponding to the acoustic emission response signal of the molten material on the back side, the peak acoustic energy, energy decay rate and continuous discharge time of the molten material during the process of detaching from the steel plate are extracted and converted into discharge energy parameters that characterize the discharge capacity of the molten material. The edge shrinkage change, the trend of the cut width change, and the solidification shrinkage displacement parameters during the cooling and solidification process of the cut area were obtained, and the correlation between the change of heat input and the evolution of the cut morphology was established. By combining the parameters of the influence of the preceding thermal response on the current position, the parameters of the molten pool propulsion speed, the parameters of the molten material discharge energy, and the parameters of the solidification shrinkage displacement of the cut, feature fusion is performed according to the thermal state influence relationship to construct a cutting quality discrimination feature vector for quality judgment.

7. The method for online quality judgment of laser-cut steel plates based on deep learning according to claim 1, characterized in that, The process of extracting thermal accumulation parameters and abrupt change parameters of thermal response in the interface region is as follows: Input the cutting quality discrimination feature vector into the quality discrimination model, and accumulate the product of the thermal input intensity and the duration of action at each position along the cutting trajectory based on the thermal response feature components associated with each spatial position in the quality discrimination feature vector. The total energy superposition and heat diffusion coverage are calculated separately for the first performance zone and the second performance zone to obtain the heat accumulation parameters during the cross-performance zone migration process. For the boundary between the first performance zone and the second performance zone, the multi-source sensing feature components corresponding to the boundary before and after the boundary are extracted from the mass discrimination feature vector, and the difference in the rate of change of plasma spectral energy, the gradient of change of airflow pressure, and the rate of change of acoustic emission energy before and after the boundary are calculated respectively. By combining the peak response offset before and after the boundary and the degree of synchronous change of multi-source signals, a thermal response mutation parameter is constructed to characterize the degree of thermal state mutation caused by the switching of performance regions.

8. The method for online quality judgment of laser-cut steel plates based on deep learning according to claim 1, characterized in that, The process of determining the defect identification result based on the maximum probability value is as follows: The thermal accumulation parameters and the thermal response mutation parameters of the boundary region are input into the fully connected layer of the quality discrimination model, and the probability of each defect category is output through nonlinear mapping. Based on the continuously changing heat accumulation parameters along the cutting trajectory and the abrupt change parameters of the thermal response in the interface region, the heat input state, melt discharge state and solidification shrinkage trend of the current laser-acting area are analyzed to determine the corresponding quality state characteristics. The output probabilities of the fully connected layer corresponding to the quality status characteristics are compared. The maximum probability value is used to determine the incomplete cutting, slag, burr or burnt edge defect status of the current cutting area, and the quality judgment result is output.

9. The method for online quality judgment of laser-cut steel plates based on deep learning according to claim 1, characterized in that, The process of forming the weight update relationship based on defect feedback constraints is as follows: Based on the defect category results output by the quality discrimination model, obtain the cross-regional thermal history migration feature vectors that participate in the discrimination calculation under the corresponding defect category, and extract the fluctuation amplitude, offset direction and peak time migration position of each peak time difference and amplitude difference in the defect occurrence area. Based on the correlation between thermal accumulation parameters, thermal response mutation parameters and peak time differences and amplitude differences under different defect categories, an influence mapping relationship between each feature component in the cross-regional thermal history migration feature vector and the defect category is established. Based on the category deviation between the current discrimination result and the actual quality inspection result, the influence of each peak time difference and amplitude difference on the discrimination result is analyzed, and the discrimination residual is allocated according to the influence ratio of each feature component on the defect classification result to obtain the error contribution of each feature component. Based on the error contribution, the fluctuation amplitude of each peak time difference and amplitude difference, and the historical discrimination error change trend, the contribution correction of each characteristic component is calculated, and the weight adjustment coefficients corresponding to each peak time difference and amplitude difference are updated according to the contribution correction. The updated weight adjustment coefficients are loaded into the thermal history migration feature mapping layer in the quality discrimination model, and the weight values ​​corresponding to each feature dimension in the cross-regional thermal history migration feature vector are recalculated to form a thermal migration feature weight update relationship based on defect feedback constraints.

Images