Method and system for predicting the fatigue life of a draw die profile coating and substrate

Abstract

This invention relates to the field of mold monitoring and life assessment technology, and discloses a method and system for predicting the life of coating and substrate fatigue on the surface of a drawing die. The method includes acquiring mold surface image data and mechanical vibration signal data; processing the surface image data to obtain a coating degradation characteristic distribution map; calculating a coating integrity quantization score based on the distribution map; if the quantization score is lower than a preset quantization score threshold, extracting high-frequency components from the mechanical vibration signal data to obtain a substrate fatigue vibration spectrum; classifying and retrieving the spectrum and calculating estimated microcrack sizes; if the estimated microcrack sizes exceed a preset critical value, normalization processing is performed to obtain a coupled failure risk index; fitting the coupled failure risk index to obtain the remaining life prediction cycle number; if the remaining life prediction cycle number is less than a preset safe cycle number, a maintenance timing alarm signal is output, providing the final decision basis. This method can improve the accuracy of mold failure risk prediction.

Description

Technical Field

[0001] This invention relates to the field of mold monitoring and life assessment technology, and in particular to a method and system for predicting the life of fatigue of the coating and substrate on the surface of a drawing die. Background Technology

[0002] Currently, drawing dies are the core process equipment in the stamping production of automotive body panels, and their service life directly determines the surface quality and production efficiency of the parts. With the large-scale application of high-strength steel plates, dies are subjected to harsh alternating loads and intense friction during service, and the synergistic failure problem of surface coating wear and substrate fatigue is becoming increasingly prominent.

[0003] In existing technologies, mold life extension and failure early warning schemes mostly adopt a separate approach, that is, researching and controlling coating wear or substrate fatigue separately. For coating wear, existing technologies often attempt to improve the wear resistance of the coating by optimizing the coating material composition, improving the coating process, or periodically re-coating. However, they lack consideration of the correlation between the coating degradation state and the degree of substrate fatigue, failing to establish an effective linkage with data acquisition and control, and unable to predict the risk state of the substrate based on the real-time degradation of the coating. For substrate fatigue, existing technologies typically analyze the fatigue life of the substrate under stress based on materials mechanics theory or finite element simulation methods, but ignore the significant impact of coating degradation on the stress distribution and fatigue process of the substrate, making it difficult to reflect the synergistic effect of the two during actual service. Some existing technologies attempt to predict failure by collecting single types of monitoring data, analyzing the substrate fatigue state only through vibration signals, or detecting coating wear only through surface images. However, because a single data source cannot fully capture the synergistic failure characteristics of the coating and the substrate, and lacks effective fusion and in-depth mining of multi-source data, the predictive model is difficult to adapt to complex and ever-changing stamping conditions. It is unable to scientifically determine the optimal time for mold continued use, partial repair, or production line shutdown for replacement, often resulting in "over-maintenance" leading to resource waste, or "untimely maintenance" causing sudden mold failure.

[0004] Existing technologies suffer from low accuracy in predicting mold failure risks. Summary of the Invention

[0005] This invention provides a method and system for predicting the fatigue life of the coating and substrate on the surface of a drawing die, in order to solve the problem of low accuracy in predicting die failure risk.

[0006] In a first aspect, to solve the above-mentioned technical problems, the present invention provides a method for predicting the fatigue life of the coating on the surface of a drawing die and the substrate, comprising: Collect surface image data and mechanical vibration signal data of the mold, obtain historical production cycle data of the mold, and perform image processing on the surface image data to obtain a coating degradation feature distribution map; A multidimensional wear feature vector is extracted based on the coating degradation feature distribution map. The multidimensional wear feature vector is then input into a preset coating wear model to obtain a coating integrity quantification score. If the coating integrity quantization score is lower than the preset quantization score threshold, then high-frequency vibration components are filtered and extracted from the mechanical vibration signal data, and the high-frequency vibration components are subjected to spectral analysis to obtain the substrate fatigue vibration spectrum. The peak frequency offset and amplitude attenuation value are calculated based on the fatigue vibration spectrum of the matrix. The peak frequency offset and amplitude attenuation value are then classified and retrieved based on a preset support vector machine model to obtain the estimated value of the microcrack size. If the estimated microcrack size exceeds the preset critical size, the coating integrity quantification score and the estimated microcrack size are normalized and weighted to obtain the coupling failure risk index. The coupling failure risk index is matched with the historical risk index sequence stored in the historical production cycle data to screen out similar historical decline trajectories. Linear regression fitting is performed based on the similar historical decline trajectories to obtain the life prediction model. The remaining life prediction cycle number is calculated according to the life prediction model and the preset critical failure threshold. If the remaining lifespan prediction cycle number is less than the preset safe lifespan cycle number, then the maintenance timing is determined based on the remaining lifespan prediction cycle number, a maintenance timing alarm signal is generated and output, all calculation data is recorded, and the final decision basis is obtained.

[0007] Secondly, the present invention provides a fatigue life prediction system for the coating and substrate of a drawing die surface, comprising: The data acquisition and preprocessing module is used to acquire surface image data and mechanical vibration signal data of the mold, obtain historical production cycle data of the mold, and perform image processing on the surface image data to obtain a coating degradation feature distribution map. The coating integrity assessment module is used to extract a multi-dimensional wear feature vector based on the coating degradation feature distribution map, and input the multi-dimensional wear feature vector into the training preset coating wear model to obtain a coating integrity quantification score. The substrate fatigue spectrum analysis module is used to filter and extract high-frequency vibration components from the mechanical vibration signal data if the coating integrity quantization score is lower than a preset quantization score threshold, and perform spectrum analysis on the high-frequency vibration components to obtain a substrate fatigue vibration spectrum diagram. The microcrack size estimation module is used to calculate the peak frequency offset and amplitude attenuation value based on the fatigue vibration spectrum of the matrix, and to classify and retrieve the peak frequency offset and amplitude attenuation value based on a preset support vector machine model to obtain the microcrack size estimate. The coupling failure risk calculation module is used to normalize the coating integrity quantification score and the microcrack size estimate if the estimated value of the microcrack size exceeds the preset critical size, and then calculate the coupling failure risk index by weighting. The remaining life prediction module is used to match the coupled failure risk index with the historical risk index sequence stored in the historical production cycle data, filter out similar historical decline trajectories, perform linear regression fitting based on the similar historical decline trajectories to obtain the life prediction model, and calculate the remaining life prediction cycle number based on the life prediction model and the preset critical failure threshold. The alarm and decision module is used to determine the maintenance timing based on the predicted remaining lifespan if the predicted remaining lifespan is less than the preset safe lifespan, generate and output a maintenance timing alarm signal, record all calculation data, and obtain the final decision basis.

[0008] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention integrates heterogeneous data from multiple sources, such as surface images and vibration signals, and employs image processing, support vector machines, spectrum analysis, and data acquisition and control technologies to achieve the collaborative extraction of coating integrity quantification scores and microcrack size estimates. Existing technologies typically treat coating wear and substrate fatigue separately, relying on only a single data source, making it difficult to comprehensively capture the coupled failure characteristics between the two. This invention uses image data to objectively assess coating degradation and dynamically triggers in-depth analysis of vibration signals based on quantification scores, achieving a shift from "single monitoring" to "collaborative sensing." This effectively solves the problems of prediction blind spots and insufficient accuracy caused by single data sources and one-sided features in traditional methods, and improves the ability to capture early failure signs.

[0009] (2) This invention constructs a cascade analysis and triggering mechanism of coating integrity quantification score, vibration spectrum characteristics, and microcrack size estimation, and introduces entropy weighting method for weighted fusion to generate a coupled failure risk index. Traditional methods lack effective quantification of the mutual acceleration mechanism of coating and substrate failure, making it difficult to assess the sudden risk caused by their superposition. This invention uses the coating state as a dynamic triggering condition and weighting factor for substrate fatigue analysis, objectively reflecting the aggravating effect of coating degradation on substrate stress distribution and crack propagation, thereby solving the problem of insufficient modeling of the coating-substrate synergistic failure process in existing technologies, and realizing accurate and dynamic assessment of the overall failure risk of the mold.

[0010] (3) This invention achieves personalized and adaptive prediction of remaining life by matching the real-time coupled failure risk index with historical production cycle data through trajectory matching and linear regression fitting. Existing life prediction methods are mostly based on fixed theoretical models or general historical data, which are difficult to adapt to individual differences and changes in operating conditions of different molds. This invention uses historical similar decay trajectories to construct a regression model, realizing data-driven prediction of current state, historical trajectory, and future trend. It solves the problems of large deviation and poor universality in life prediction caused by individual differences in molds and process fluctuations, and significantly improves the accuracy of remaining life prediction and its guiding value for production scheduling.

[0011] (4) This invention forms a closed-loop management chain by setting graded early warning thresholds (quantitative score threshold, critical size, number of safe cycles) and outputting maintenance timing alarm signals and structured decision-making basis. Traditional maintenance decisions rely on experience judgment and lack data support, which can easily lead to over-maintenance or untimely maintenance. This invention combines the prediction results with the actual production (maintenance timing cutoff point, production plan) and automatically outputs operable maintenance alarms and log basis, solving the problems of strong subjectivity and delayed response in maintenance decisions. It realizes the intelligent upgrade from post-maintenance to predictive maintenance, and optimizes maintenance costs and production continuity while ensuring mold safety. Attached Figure Description

[0012] Figure 1 This is a schematic flowchart of the fatigue life prediction method for the coating and substrate of the drawing die surface provided in the first embodiment of the present invention. Figure 2 This is a schematic diagram of the fatigue life prediction system for the coating and substrate of the drawing die provided in the second embodiment of the present invention. Detailed Implementation

[0013] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0014] Reference Figure 1 The first embodiment of the present invention provides a method for predicting the fatigue life of the coating on the surface of a drawing die and the substrate, comprising the following steps: S11, Collect surface image data and mechanical vibration signal data of the mold, obtain historical production cycle data of the mold, perform image processing on the surface image data, and obtain a coating degradation feature distribution map; S12, a multidimensional wear feature vector is extracted based on the coating degradation feature distribution map, and the multidimensional wear feature vector is input into a preset coating wear model to obtain a coating integrity quantification score; S13, if the coating integrity quantization score is lower than the preset quantization score threshold, then the high-frequency vibration component is filtered and extracted from the mechanical vibration signal data, and the high-frequency vibration component is subjected to spectrum analysis to obtain the substrate fatigue vibration spectrum. S14, calculate the peak frequency offset and amplitude attenuation value based on the fatigue vibration spectrum of the matrix, and classify and retrieve the peak frequency offset and amplitude attenuation value based on the preset support vector machine model to obtain the microcrack size estimate. S15, if the estimated value of the microcrack size exceeds the preset critical size, the coating integrity quantification score and the estimated value of the microcrack size are normalized and weighted to obtain the coupling failure risk index. S16, Match the coupling failure risk index with the historical risk index sequence stored in the historical production cycle data, screen out similar historical decline trajectories, perform linear regression fitting based on the similar historical decline trajectories to obtain the life prediction model, and calculate the remaining life prediction cycle number based on the life prediction model and the preset critical failure threshold. S17. If the remaining lifespan prediction cycle number is less than the preset safe cycle number, then the maintenance timing is determined based on the remaining lifespan prediction cycle number, a maintenance timing alarm signal is generated and output, all calculation data is recorded, and the final decision basis is obtained.

[0015] In step S11, surface image data and mechanical vibration signal data of the mold are acquired, historical production cycle data of the mold are obtained, and image processing is performed on the surface image data to obtain a coating degradation feature distribution map, including: Acquire surface image data of the mold, mechanical vibration signal data, and historical production cycle data containing historical risk index evolution sequences and corresponding cumulative production cycle counts; The surface image data is convolved to extract horizontal and vertical gradients and synthesize a gradient magnitude map. Extract the high-frequency components from the mechanical vibration signal data. If the amplitude of the high-frequency components exceeds a preset amplitude threshold, an adaptive dynamic threshold coefficient is obtained by weighted calculation. The adaptive dynamic threshold coefficient is applied to perform double threshold hysteresis segmentation on the gradient magnitude map to obtain a surface defect profile map; Based on the feature quantification and scoring of the area, perimeter, and shape complexity of each defect region in the surface defect profile map, a local degradation index matrix is ​​constructed, and the local degradation index matrix is ​​used to obtain a coating degradation feature distribution map through pseudo-color mapping.

[0016] In one implementation, this embodiment constructs a multi-source data collaborative acquisition system. Through standardized deployment of acquisition equipment, parameter calibration, and data synchronization mechanisms, it ensures the integrity, accuracy, and temporal consistency of surface images, vibration signals, and historical data, providing a reliable data foundation for the accurate extraction of coating degradation characteristics. For mold surface image data acquisition, this embodiment selects an industrial-grade high-speed area array camera as the core acquisition device. Its resolution is set to 2048×1536 pixels, covering the complete grayscale levels from 0 to 255, clearly capturing micron-level scratches, peeling, bubbles, and other minute defects on the coating surface. The camera is mounted above the mold using a custom bracket, and after precise calibration, maintains a fixed shooting distance of 500mm from the key surfaces. The lens field of view is adjusted to 30° to ensure that a single shot completely covers the mold's stamping working area, with no monitoring blind spots. Considering the complex lighting environment in the stamping workshop, four sets of ring-shaped LED cold light source supplementary lights are configured, with an illuminance adjustment range of 5000-15000 lux. These lights can adaptively adjust according to the ambient light intensity, avoiding interference from coating reflections and shadows, and ensuring uniform brightness across all areas of the image. The image acquisition frequency is synchronized with the die stamping cycle; one image is captured after each stamping cycle is completed. Data is transmitted in real-time to the edge computing unit via the GigE Vision interface, with transmission latency controlled within 50ms, achieving real-time linkage between the production process and data acquisition.

[0017] It should be noted that, in terms of mechanical vibration signal acquisition, this embodiment deploys a total of six high-precision accelerometers at the mold base, upper mold base, and key load-bearing components. The sensors have a measurement range of ±50g, a sensitivity set at 100mV / g, and a fixed sampling frequency of 10kHz. They can comprehensively capture low-frequency vibrations below 10Hz to high-frequency impact signals above 5kHz, exhibiting particularly high response sensitivity to high-frequency vibration anomalies caused by coating peeling and substrate micro-crack propagation. The sensors are installed using a threaded fastening method, and the mounting surface is polished, with a flatness error controlled within 0.02mm to ensure tight contact with the mold surface and reduce vibration signal transmission loss. The signal is connected to the data acquisition card via a shielded cable using differential transmission, effectively resisting electromagnetic interference in the workshop and improving the signal-to-noise ratio to over 45dB. After the acquisition card converts the analog signal into a 16-bit digital signal, it is transmitted to the edge computing unit through the PCIe interface. A timestamp synchronization mechanism is used with the surface image data, with a time synchronization error not exceeding 1ms, ensuring time consistency of multi-source data.

[0018] It should be noted that historical production cycle data is acquired through an industrial Ethernet interface that establishes a communication connection with the stamping production line MES system. Data is periodically read based on the OPC UA protocol, with an acquisition cycle set at 1 hour per read to balance data timeliness and network bandwidth usage. The acquired data includes a historical risk index evolution sequence and the corresponding cumulative production cycle count. The former records the coupled failure risk index of each inspection cycle during the mold's past service life, forming a time series; the latter corresponds to the total number of stamping cycles completed by the mold at the time each risk index is recorded. Historical data is managed using a time-series database, supporting high-concurrency read / write and fast querying. Complete data for the most recent 3 years is retained, and invalid data (such as abnormal changes in risk index, mismatches between cycle counts and timestamps) is periodically cleaned to ensure data quality.

[0019] In another implementation, after completing multi-source data acquisition, this embodiment performs convolution operations on the mold surface image data to extract a gradient magnitude map. First, a 5×5 Gaussian filter is used to preprocess the original image, smoothing random noise while preserving defect details. Then, the Sobel operator is used to extract horizontal and vertical gradients respectively. Pixel-by-pixel convolution operations are used to obtain horizontal and vertical gradient maps, which are then synthesized to obtain the gradient magnitude map. In the map, the coating defect area exhibits high gradient magnitude bright spots or bands due to drastic changes in grayscale values, while normal areas maintain low gradient magnitudes, achieving a preliminary distinction between defects and the background. For example, the scratch edge on the coating surface of a mold has a gradient magnitude much higher than the surrounding normal area, forming a clear high-brightness band in the map.

[0020] Specifically, the next step is to extract the high-frequency components from the mechanical vibration signal and determine whether they exceed the limits. First, a 5th-order Butterworth low-pass filter is used to remove high-frequency noise above 5kHz. Then, a fast Fourier transform is used to convert the time-domain signal into a frequency-domain signal, defining 1kHz-5kHz as the high-frequency component range (corresponding to the vibration response caused by coating peeling and substrate microcrack propagation). The sum of signal amplitudes within this frequency band is calculated and compared with a preset amplitude threshold (based on the statistical analysis of high-frequency vibration amplitudes under normal operating conditions, taking the average value plus 3 times the standard deviation) to determine whether it exceeds the limit. If the high-frequency component amplitude exceeds the limit, an adaptive dynamic threshold coefficient is obtained through weighted calculation. This coefficient comprehensively considers the degree of high-frequency amplitude exceeding the limit and the duration of exceeding the limit. The former is the ratio of the difference between the current high-frequency amplitude and the threshold, and the latter is the ratio of the duration of exceeding the limit to the reference duration (0.1s). Both are assigned weights of 0.6 and 0.4 respectively, and the weighted sum is used to obtain the coefficient (with a value range of 1.0-2.0). If it does not exceed the limit, the coefficient is set to 1.0 by default.

[0021] It should be noted that this coefficient is subsequently applied to perform double-threshold hysteresis segmentation on the gradient magnitude map. First, an initial threshold is obtained using the Otsu algorithm, and then the low and high thresholds are adjusted according to the coefficient. During segmentation, pixels with gradient magnitudes greater than the high threshold are marked as confirmed defect points, those less than the low threshold are marked as background points, and those in between are considered suspected defect points. For suspected defect points, it is determined whether they have an 8-neighborhood connectivity relationship with confirmed defect points; if so, they are determined as defect points, otherwise as background points. After segmentation, a 3×3 circular structuring element is used for dilatation-erosion morphological processing to fill internal holes in the defects and remove edge burrs, resulting in a smooth, complete binary contour map of surface defects.

[0022] It should be noted that when constructing the local degradation index matrix and generating the coating degradation feature distribution map, each connected defect region in the contour map is traversed to extract three core feature parameters: defect area, perimeter, and shape complexity. The area is converted to the actual area using the number of pixels (a single pixel is calibrated to 10μm×10μm), the perimeter is the sum of the Euclidean distances between the defect boundary pixels, and the shape complexity reflects the degree of irregularity of the defect shape. Based on these three parameters, a quantization score is assigned, and a degradation index (range 0-1, with larger values ​​indicating more severe degradation) is obtained by weighting the values ​​at 0.5, 0.3, and 0.2. Specifically, each connected defect region in the contour map is traversed, and three core feature parameters are extracted: defect area, perimeter, and shape complexity. The shape complexity is the square of the perimeter divided by 4π times the defect area, which is the reciprocal of the roundness. Each feature parameter is normalized and scored, and the relative severity of a certain feature value of the defect in all defects detected in the current test is calculated. For example, the score is obtained by using the proportion of the feature value to the maximum value of the feature among all defects. Then, the scores are weighted and summed with weights of 0.5, 0.3, and 0.2 to obtain the degradation index of each defect region.

[0023] When constructing the local degradation index matrix, the critical surfaces of the mold are divided into 100×100 grid cells (each cell is 1mm×1mm in actual size). The degradation index of each cell is assigned according to the following rules: 0 for defect-free areas, the degradation index of a single defective area is taken directly, and the maximum value is taken for multiple defective areas. Finally, the Jet color mapping table is used to perform pseudo-color mapping on the matrix, corresponding the degradation index 0-1 to a gradient of blue (low degradation) to red (high degradation), forming a coating degradation feature distribution map.

[0024] In step S12, a multidimensional wear feature vector is extracted based on the coating degradation feature distribution map. This multidimensional wear feature vector is then input into a preset coating wear model to obtain a coating integrity quantification score, including: Morphological gradient calculations were performed on the coating degradation feature distribution map to extract the coordinates of thickness variation regions and peeling points; A multidimensional wear feature vector is constructed based on the average grayscale attenuation rate of the thickness variation region, the proportion of the total area of ​​the variation region, the number of peeling points, the average cluster density of the peeling points, and the maximum cluster radius. The multidimensional wear feature vector is input into a preset coating wear model, and a support vector machine is used to map it to a high-dimensional space with a radial basis kernel function. The optimal classification hyperplane is then solved to obtain the classification decision value. The Euclidean distance between the classification decision value and the preset center of the intact coating sample is calculated, and the coating integrity quantification score is obtained through normalization mapping.

[0025] In one implementation, this embodiment achieves accurate quantitative assessment of coating integrity through multi-step feature extraction and intelligent model analysis, providing core data support for subsequent substrate fatigue analysis and failure risk assessment. For the morphological gradient operation of the coating degradation feature distribution map, this embodiment adopts a standardized morphological processing flow to highlight areas of coating thickness variation and accurately locate the coordinates of the peeling points. First, the coating degradation feature distribution map is preprocessed using a 3×3 rectangular structuring element for opening operations to remove tiny noise spots in the image while preserving the overall outline of the defect area, avoiding the impact of noise interference on subsequent gradient extraction. Then, morphological gradient operations are performed. This operation, by subtracting the result of the erosion operation from the image's dilation operation, effectively highlights areas in the image where grayscale values ​​change drastically. These areas correspond precisely to transition zones where the coating thickness is significantly thinned or peeled off. In this embodiment, both dilation and erosion operations use 3×3 circular structuring elements. Circular structuring elements can act uniformly in all directions of the image, avoiding edge distortion problems that may occur with rectangular structuring elements. After morphological gradient calculation, the gradient value of the normal coating area is usually maintained between 8 and 18, while the gradient value of the transition zone with drastic thickness changes will exceed 45, forming obvious high gradient response areas. These high gradient connected areas are identified as thickness change areas.

[0026] It should be noted that in the peeling point coordinate extraction stage, this embodiment combines the morphological gradient calculation results with a local maximum suppression algorithm to achieve accurate localization. First, a gradient magnitude threshold of 60 is set, and pixels with gradient magnitudes greater than this threshold are selected as a candidate set of suspected peeling points. This threshold is determined by analyzing a large amount of historical defect data and can effectively filter out false gradient response points. Then, local maximum suppression processing is performed on the pixels in the candidate set, retaining only the pixels with the largest gradient magnitude within an 8-neighborhood as the final peeling point coordinates, ensuring that each peeling point is marked only once and avoiding duplicate counting.

[0027] In another implementation, when constructing the multidimensional wear feature vector, this embodiment selects five key feature parameters to comprehensively characterize the wear state of the coating. The first feature parameter is the average grayscale attenuation rate of the thickness variation region, which is obtained by calculating the ratio of the difference between the grayscale values ​​of all pixels in the thickness variation region and the average grayscale value of the normal coating region, reflecting the overall degree of coating thickness reduction. The second feature parameter is the total area ratio of the variation region, that is, the ratio of the total number of pixels in all thickness variation regions to the total number of pixels in the image, used to measure the coverage of the coating wear region. The third feature parameter is the number of peeling points, which is directly used as the total number of extracted peeling point coordinates, intuitively reflecting the severity of coating peeling. The fourth feature parameter is the average cluster density of peeling points, which is calculated by dividing the total number of peeling points by the actual area of ​​the thickness variation region (derived from the image pixel size calibration; in this embodiment, the actual size of a single pixel is 10μm×10μm), reflecting the density of peeling points. The fifth feature parameter is the maximum cluster radius. The K-means clustering algorithm is used to cluster the coordinates of the peeling points, calculating the minimum circumscribed circle radius of each cluster. The maximum value is taken as the maximum cluster radius, reflecting the concentrated area of ​​the peeling region. For example, in a certain test, this radius was 4.2 mm, indicating a large area of ​​concentrated peeling. Combining these five feature parameters in sequence forms a five-dimensional wear feature vector under the current operating condition. This vector comprehensively covers key information such as the intensity, range, and density of coating wear.

[0028] It should be noted that when inputting the multidimensional wear feature vector into the preset coating wear model, this embodiment uses the support vector machine algorithm to construct a classification model, and uses the radial basis function (RBF) kernel function to achieve high-dimensional space mapping and solve for the optimal classification hyperplane. The training process of the coating wear model adopts a standardized machine learning process. First, a training dataset containing a large number of labeled samples is constructed. The samples cover four typical states: intact coating, light wear, moderate wear, and severe peeling. The sample numbers of each class account for 25%, 30%, 30%, and 15%, respectively, to ensure a balanced sample distribution and avoid model bias. During training, the multidimensional wear feature vector is used as input, and the corresponding wear state label is used as output. The input features are mapped to a high-dimensional feature space through the RBF kernel function. The optimal classification hyperplane that maximizes the interval between samples of different classes is solved in the high-dimensional space. The RBF kernel function has good nonlinear fitting ability and can effectively handle the complex nonlinear relationship between coating wear features and wear state. At the same time, the model overfitting is avoided by adjusting the regularization parameter. After the model is trained, the multidimensional wear feature vector obtained from the current detection is input into the model to obtain the signed classification decision value of the vector in the classification hyperplane. The sign of this value reflects the wear category to which the current coating state belongs, and the absolute value of the value reflects the distance from the classification hyperplane. For example, a classification decision value of -1.87 obtained in a certain detection indicates that the current state has crossed the boundary between intact and light wear, belonging to light wear or above.

[0029] It should be noted that, in calculating the coating integrity quantification score, this embodiment calculates the Euclidean distance between the classification decision value and the preset center of intact coating samples, and obtains the final score through normalization mapping. First, the preset center of intact coating samples is determined, which is obtained by calculating the mean of the classification decision values ​​of all intact coating samples. This distance is then normalized and mapped, with a mapping range of 0-100 points, where 100 points represents a brand-new, intact coating state, and 0 points represents a completely failed coating state. The mapping formula is implemented through a linear transformation to ensure that the distance and score have an inverse linear relationship.

[0030] Specifically, the multidimensional feature extraction and support vector machine evaluation model constructed in this embodiment has strong anti-interference ability and working condition adaptability, and can effectively resist noise interference caused by environmental factors such as uneven lighting and slight vibration in the stamping workshop.

[0031] In step S13, if the coating integrity quantization score is lower than a preset quantization score threshold, high-frequency vibration components are filtered and extracted from the mechanical vibration signal data, and spectral analysis is performed on the high-frequency vibration components to obtain a substrate fatigue vibration spectrum, including: If the coating integrity quantization score is lower than the preset quantization score threshold, then synchronously acquired mechanical vibration signal data is obtained. The cutoff frequency is determined based on the coating integrity quantization score, and the mechanical vibration signal data is filtered by a high-pass filter to extract the high-frequency vibration component. Spectral analysis is performed on the high-frequency vibration components to convert the time-domain signal into a frequency-domain signal, thereby obtaining vibration energy distribution data; Extract fatigue characteristic frequency points with high amplitude intensity from the vibration energy distribution data, and construct a fatigue vibration spectrum diagram of the matrix based on the fatigue characteristic frequency points.

[0032] In one implementation, this embodiment constructs a precise vibration signal analysis process based on coating state feedback. By dynamically adjusting filtering parameters and spectral feature extraction strategies, it achieves efficient capture and visualization of high-frequency vibration signals related to substrate fatigue, providing reliable spectral data support for subsequent microcrack size estimation. Setting a preset quantization score threshold is a crucial prerequisite for triggering this step's analysis. This threshold is determined through extensive statistical analysis of historical failure data and verification in engineering practice. In this embodiment, the preset quantization score threshold is set to 50 points (out of 100). When the coating integrity quantization score output in step S12 is lower than this threshold, it indicates that the coating has experienced significant wear or peeling, its protective effect on the substrate is significantly weakened, and the substrate may suffer fatigue damage due to stress concentration. At this point, in-depth analysis of the vibration signal needs to be initiated.

[0033] It should be noted that in the synchronous acquisition of mechanical vibration signal data, this embodiment employs a timestamp synchronization mechanism to ensure data consistency. The system reads mechanical vibration signal data whose timestamp deviation from the coating degradation characteristic distribution map acquisition timetamp does not exceed 1ms. The data reading duration is set to 10 minutes, which is determined by analyzing the continuous characteristics of vibration signals caused by fatigue crack propagation in the substrate. This duration can fully cover the vibration response of at least 3-5 stamping cycles, ensuring the capture of stable fatigue vibration characteristics. The vibration signal data originates from the accelerometer deployed in step S11. The data format is a 16-bit digital signal, containing key information such as the amplitude and sampling time of the time-domain waveform. It is imported into the spectrum analysis unit through a high-speed data transmission interface, with the transmission delay controlled within 20ms to ensure data real-time performance.

[0034] Specifically, when determining the cutoff frequency based on the coating integrity quantification score, this embodiment employs a dynamic adaptation strategy to ensure a positive correlation between the cutoff frequency and the degree of coating wear. The base value for the cutoff frequency is set to 1000Hz. When the coating integrity quantification score is between 40 and 50, the cutoff frequency is set to 1500Hz; when the score is between 30 and 40, the cutoff frequency is set to 2000Hz; and when the score is below 30, the cutoff frequency is set to 2500Hz. The more severe the coating wear, the more intense the direct contact between the substrate and the workpiece, resulting in higher fatigue vibration frequencies, requiring higher cutoff frequencies to capture the relevant high-frequency components. For example, when the coating integrity quantification score is 41.3, the system sets the cutoff frequency to 2000Hz to ensure coverage of high-frequency vibration signals related to coating peeling and substrate microcrack propagation.

[0035] In another implementation, this embodiment uses a Chebyshev Type I high-pass filter. The high-pass filter design and filtering process employ a standardized signal processing workflow, effectively preserving high-frequency signals above the cutoff frequency while significantly suppressing low-frequency noise. The filter order is set to 8th, with a passband ripple not exceeding 1dB and a stopband attenuation greater than 80dB, ensuring the stability and reliability of the filtering effect. During the filtering process, the collected mechanical vibration signal data is input into the high-pass filter. The filter filters the signal according to the preset cutoff frequency, eliminating low-frequency background vibrations generated during normal equipment operation (such as motor rotation, frame resonance, etc.), and retaining only high-frequency vibration components above the cutoff frequency. These high-frequency components mainly correspond to the vibration responses generated by damage processes such as impact contact between the coating peeling area and the substrate, and the initiation and propagation of fatigue cracks in the substrate.

[0036] Specifically, the spectral analysis described in this invention includes, but is not limited to, frequency domain analysis methods such as Fast Fourier Transform (FFT), Short-Time Fourier Transform (SFT), and Wavelet Transform, used to extract spectral features related to matrix fatigue from time-domain vibration signals. When performing Fast Fourier Transform (FFT) on high-frequency vibration components, this embodiment employs optimized transform parameters to ensure the accuracy and efficiency of the frequency domain analysis. First, the high-frequency vibration components are preprocessed using a Hanning window function to reduce spectral leakage. The length of the window function is matched to the number of signal sampling points to ensure signal integrity. Then, the number of transform points for the FFT is set to 1024. The FFT converts the time-varying high-frequency vibration signal in the time domain into vibration energy distribution data in the frequency domain. This data, with frequency as the abscissa and energy amplitude as the ordinate, visually presents the distribution of vibration energy at different frequencies.

[0037] In another implementation, the extraction of fatigue characteristic frequency points employs a strategy combining amplitude ranking and threshold screening. First, all frequency points in the vibration energy distribution data are sorted from highest to lowest energy amplitude, and the top five frequencies with the highest amplitudes are selected as candidate characteristic frequencies. Then, an amplitude intensity threshold is set, which is three times the average amplitude intensity of all frequency points. Frequency points with amplitude intensities exceeding this threshold are selected as the final fatigue characteristic frequencies. This screening strategy effectively eliminates false high-amplitude frequency points generated by random noise, ensuring that the extracted fatigue characteristic frequencies are directly related to the fatigue damage of the matrix. For example, in one analysis, the amplitude intensity at 2500Hz was 3.7, and the amplitude intensity at 3200Hz was 2.9, both exceeding three times the average amplitude intensity, and were therefore identified as fatigue characteristic frequencies.

[0038] It should be noted that when constructing the fatigue vibration spectrum diagram of the matrix, this embodiment adopts a professional spectrum visualization format, with frequency (Hz) as the horizontal axis and energy amplitude (after normalization) as the vertical axis. The horizontal axis ranges from 0 to 5000 Hz, and the vertical axis ranges from 0 to 5, facilitating intuitive observation of the energy amplitude differences at each frequency point. In the spectrum diagram, the extracted fatigue characteristic frequency points are specially marked (e.g., highlighted with red dots), and their corresponding frequency values ​​and amplitude intensities are labeled, enabling operators to quickly locate key characteristic frequency points. Simultaneously, a vibration energy distribution reference curve under normal operating conditions is added to the spectrum diagram. By comparing the current spectrum with the reference curve, the changes in the fatigue state of the matrix are intuitively reflected.

[0039] In step S14, the peak frequency offset and amplitude attenuation value are calculated based on the fatigue vibration spectrum of the matrix. The peak frequency offset and amplitude attenuation value are then classified and retrieved based on a preset support vector machine model to obtain an estimated microcrack size, including: Analyze the energy distribution of the fatigue vibration spectrum of the matrix to determine the frequency point of maximum energy; The difference between the maximum energy frequency point and the reference frequency point under normal operating conditions of the equipment is calculated to obtain the peak frequency offset; Based on the spectral position corresponding to the peak frequency offset, calculate the amplitude attenuation value of the corresponding spectral position compared to the normal state; The peak frequency offset and the amplitude attenuation value are combined to construct the feature vector to be judged, and the vector is input into a preset support vector machine model to determine its position in the classification hyperplane. The crack level label is identified based on the location, and a preset database is retrieved based on the crack level label to obtain the estimated value of the microcrack size.

[0040] In one implementation, this embodiment employs a multi-dimensional energy screening strategy to ensure accurate positioning. First, the frequency domain data of the matrix fatigue vibration spectrum is gridded, dividing the frequency axis into several frequency bands at 10Hz intervals. The total energy within each frequency band is calculated and sorted, initially screening out the top 5 candidate frequency bands with the highest total energy. Subsequently, a refined search is performed within each candidate frequency band, using a quadratic interpolation algorithm to improve frequency resolution and accurately locate the frequency point with the highest energy amplitude as the maximum energy frequency point. This method effectively avoids peak shifts caused by discrete spectrum sampling, ensuring that the identification error of the maximum energy frequency point is less than 5Hz. For example, in a certain detection, after initial screening, 2700-2900Hz was identified as a high-energy candidate frequency band. Through a quadratic interpolation algorithm, 2800Hz was ultimately located as the maximum energy frequency point. The energy amplitude at this frequency point is significantly higher than in other areas and is highly correlated with the vibration response caused by the propagation of microcracks in the matrix.

[0041] It should be noted that the reference frequency point is set using a dynamic calibration mechanism to ensure it accurately reflects the vibration characteristics under normal equipment operation. Before the mold is first put into use, reference data is collected for 200 production cycles. During the collection process, the mold is in brand new condition (intact coating, undamaged substrate). The vibration spectrum of each cycle is recorded, and the maximum energy frequency point is extracted. The average value of these frequency points is calculated as the initial reference frequency point. During subsequent service, the system automatically triggers a reference frequency calibration every 1000 production cycles, updating the reference frequency point after removing abnormal data to ensure that the reference value can adapt to the minor vibration changes caused by normal equipment aging. The peak frequency offset is directly calculated by the difference between the current maximum energy frequency point and the reference frequency point. For example, when the current maximum energy frequency point is 2800Hz, the peak frequency offset is 300Hz. This value directly reflects the impact of substrate damage on vibration frequency characteristics. A larger offset usually means a larger microcrack size or a faster propagation rate.

[0042] In another implementation, the amplitude attenuation value is calculated by focusing on the spectral position corresponding to the peak frequency offset, and quantification is achieved by comparing it with the amplitude under normal conditions. First, the amplitude intensity of the frequency point corresponding to the peak frequency offset (i.e., the current maximum energy frequency point) in the current spectrum is extracted and recorded as the current amplitude. Then, the normal state spectrum consistent with the current operating conditions (such as stamping materials, process parameters, etc.) is retrieved from the reference spectrum database, and the amplitude intensity of the same frequency point is extracted and recorded as the reference amplitude. The amplitude attenuation value is calculated by dividing the difference between the reference amplitude and the current amplitude by the reference amplitude. This value reflects the degree of energy loss of the vibration signal at a specific frequency point and is directly related to the decrease in structural stiffness caused by microcracks.

[0043] It should be noted that the construction of the feature vector to be judged adopts a standardized feature combination method, combining the peak frequency offset and amplitude attenuation value in sequence into a two-dimensional feature vector. To eliminate the influence of dimensional differences on model classification, the two feature parameters are normalized before combination. The peak frequency offset is normalized using maximum-minimum normalization, mapping it to the [0,1] interval. The maximum value of normalization is referenced from the maximum offset of 500Hz in historical data. The amplitude attenuation value is directly retained in percentage form and also mapped to the [0,1] interval.

[0044] It should be noted that the preset support vector machine model adopts a binary classification extended multi-class architecture, specifically optimized for microcrack level recognition. The model's training dataset comes from a large number of mold experiments with artificially pre-fabricated cracks. By pre-fabricating microcracks of different sizes (0.1mm, 0.3mm, 0.5mm, 0.8mm, 1.0mm, 1.5mm) on the mold substrate, vibration spectrum data corresponding to each crack size were collected. Peak frequency shift and amplitude attenuation values ​​were extracted to construct training samples, with a total of over 5000 samples covering different working conditions and crack locations. During training, a radial basis function kernel function was used to achieve feature space mapping, and the penalty parameter C and kernel function parameter σ were optimized using cross-validation. Finally, it was determined that C=10 and σ=0.1 resulted in the optimal model classification accuracy. After inputting the feature vector to be discriminated into the trained model, the model calculates the distance and relative position between the feature vector and the classification hyperplane, and outputs the corresponding classification result.

[0045] Specifically, the crack grade labels are defined using a five-level classification standard, each corresponding to a different range of microcrack sizes: Level 1 (no crack), Level 2 (microcrack size 0.1-0.3 mm), Level 3 (microcrack size 0.3-0.8 mm), Level 4 (microcrack size 0.8-1.2 mm), and Level 5 (microcrack size greater than 1.2 mm). The pre-built database combines experimental data with theoretical calculations to ensure the accuracy of the microcrack size range. The database stores the microcrack size range and typical value corresponding to each crack grade label. The size range is determined by statistically analyzing crack size measurements from a large number of experimental samples, and the typical value is the median of the range. For example, the microcrack size range corresponding to the Level 3 crack label is 0.3-0.8 mm, with a typical value of 0.55 mm; the size range corresponding to the Level 4 crack label is 0.8-1.2 mm, with a typical value of 1.0 mm. Once the system identifies a crack level label, it automatically retrieves the corresponding size range and typical value from the database as an estimate of the microcrack size. Operators can select either the typical value or the size range as a reference based on actual needs. For example, when a level 3 crack label is identified, the estimated microcrack size is 0.3-0.8 mm (typical value 0.55 mm). Subsequent non-destructive testing verified that the error between this estimate and the actual crack size is less than 0.1 mm, meeting the accuracy requirements for engineering applications.

[0046] In step S15, if the estimated microcrack size exceeds a preset critical size, the coating integrity quantification score and the estimated microcrack size are normalized, and a weighted calculation is performed to obtain the coupling failure risk index, including: If the estimated size of the microcrack exceeds the preset critical size, the hyperspectral scanning program on the coating surface is started to obtain hyperspectral scanning data of the coating surface; The coating integrity quantification score is obtained by analyzing the spectral reflectance characteristics of the hyperspectral scanning data. The estimated microcrack size and the quantification score of coating integrity are normalized to unify the dimensions, and the objective weighting factors of the estimated microcrack size and the quantification score of coating integrity are calculated. The coupling failure risk index is obtained by weighted averaging the normalized values ​​and the corresponding objective weighting factors.

[0047] In one implementation, this embodiment sets the preset critical size to 1.0 mm. When the estimated microcrack size output in step S14 exceeds this value, it indicates that the microcracks in the substrate have entered a rapid propagation stage. Relying solely on the previous coating integrity assessment is insufficient to fully reflect the overall mold failure risk, necessitating the initiation of a hyperspectral scanning program to collect refined data on the coating surface. The initiation and data acquisition of the hyperspectral scanning program on the coating surface employ a high-precision, non-contact measurement method. This embodiment uses a pushbroom hyperspectral imager with a spectral range covering 400-2500 nm (including visible and near-infrared bands), a spectral resolution of 5 nm, and a spatial resolution of 0.1 mm / pixel, capable of capturing subtle changes in the spectral reflectance characteristics of the coating surface caused by wear, oxidation, and peeling. The hyperspectral imager is mounted above the mold using a custom bracket, maintaining a fixed distance of 300 mm from the coating surface. The scanning range covers the thickness variation areas and concentrated peeling points on the key surfaces of the mold, ensuring accurate correspondence between the scanned data and the previously detected defect areas. During the scanning process, a diffuse reflection standard plate (99% reflectivity) is used for real-time calibration to eliminate the influence of light intensity fluctuations on spectral data. The scanning speed is set to 0.5 mm / s, and the data volume of a single scanned image is 1024×1024 pixels. The data is transmitted to the analysis unit through optical fiber, and the transmission delay is controlled within 30ms to ensure the timeliness and integrity of the data.

[0048] It should be noted that the preset critical size is determined based on the fatigue limit and engineering safety standards of the mold matrix material (Cr12MoV mold steel). Tensile fatigue tests were conducted on the matrix material to obtain the crack propagation threshold under alternating loads. Combined with the maximum stress level during actual mold service (the maximum stress in the rounded corner area of ​​the mold during stamping is approximately 800-1000 MPa), statistical analysis determined 1.0 mm as the critical node for rapid microcrack propagation. When the microcrack size exceeds 1.0 mm, the crack propagation rate increases significantly (approximately 3-5 times that below 1.0 mm), therefore, it is set as the critical size.

[0049] It should be noted that this embodiment employs a combination of spectral feature extraction and pattern recognition when analyzing the coating integrity quantification score based on hyperspectral scanning data. First, the hyperspectral data is preprocessed, using Savitzky-Golay filtering to eliminate spectral noise and multivariate scattering correction (MSC) to eliminate spectral interference caused by differences in coating surface roughness. Then, key spectral feature parameters of the coating surface are extracted, including reflectance in characteristic bands, spectral absorption peak area, and the rate of change of reflectance slope. These parameters effectively reflect the coating thickness, density, and surface oxidation degree. The extracted spectral feature parameters are input into a preset coating wear model, which is trained on a large number of coating samples with different damage states and outputs a quantification score between 0 and 1 (the closer the score is to 1, the better the coating integrity).

[0050] It is worth noting that the coating integrity quantification score obtained based on surface image data in step S12 is a rapid and efficient assessment of the coating status within a normal monitoring cycle. When step S14 determines that the microcrack size exceeds the preset critical size (e.g., 1.0 mm), it indicates that the mold has entered a high-risk state. To obtain more accurate coating degradation information to support key risk decisions, the system triggers a hyperspectral scanning program as a one-time, higher-precision supplementary detection. The coating integrity quantification score obtained through hyperspectral data analysis, with its evaluation dimensions such as chemical composition and micro-oxidation state, complements the score obtained based on grayscale image morphology analysis, together forming a comprehensive description of the coating status. In the subsequent calculation of the coupled failure risk index, the quantification score obtained from hyperspectral data analysis is used preferentially or weighted and fused to improve the accuracy of the risk index calculation. This design achieves a balance between monitoring accuracy and efficiency, that is, using efficient methods under normal conditions and using high-precision methods for verification when the risk escalates.

[0051] It should be noted that normalizing the estimated microcrack size and the coating integrity quantification score is to eliminate the difference in their dimensions and ensure the rationality of the weighted fusion. This embodiment uses the maximum-minimum normalization method to map the two indicators to the [0,1] interval. For the estimated microcrack size, the maximum normalized value is set to 2.0 mm, which is set with reference to the preset critical size (1.0 mm) and historical extreme data. This value is approximately twice the critical size, reserving sufficient numerical space for severe crack propagation. The minimum value is 0 mm. The normalization formula is that the normalized microcrack size equals the estimated microcrack size divided by the maximum value. The objective weighting factor is calculated using the entropy weighting method. First, historical detection data of the same model mold under various wear and crack conditions are collected to construct a decision matrix. The rows of the matrix correspond to a single detection sample, and the columns correspond to the two indicators: the normalized microcrack size and the coating integrity quantification score. Then, the information entropy of each indicator is calculated. The smaller the information entropy, the greater the dispersion of the indicator and the higher its contribution to risk assessment. Finally, the weighting factor is calculated based on the information entropy. The coupling failure risk index is obtained by weighted averaging the normalized values ​​and objective weighting factors.

[0052] In step S16, the coupling failure risk index is matched with the historical risk index sequence stored in the historical production cycle data to screen out similar historical decline trajectories. Linear regression fitting is performed based on these similar historical decline trajectories to obtain a lifetime prediction model. The remaining lifetime prediction cycles are calculated based on the lifetime prediction model and a preset critical failure threshold, including: The coupling failure risk index is numerically matched with the evolution sequence of the historical risk index in the historical production cycle data to filter out similar historical decline trajectory data. Extract the similar historical decline trajectory data to construct a linear regression training set, use the linear regression algorithm to fit the training set to obtain regression model parameters, and construct a life prediction model based on the regression model parameters; The theoretical total cycle life value is calculated by substituting the preset critical failure threshold into the life prediction model. The difference between the theoretical total cycle life value and the current cumulative number of production cycles completed is calculated to obtain the number of remaining cycles for the life prediction.

[0053] In one implementation, this embodiment employs a standardized data cleaning and structured storage strategy. Historical production cycle data originates from inspection records and operation logs during the mold's past service life, including the coupling failure risk index corresponding to each inspection, the cumulative number of production cycles, the stamping process parameters at that time (such as stamping material, pressure value, speed, etc.), and environmental parameters (such as workshop temperature, humidity, etc.). First, the raw data is cleaned to remove outliers caused by equipment failures or abnormal data transmission (such as sudden changes in the risk index, mismatches between cycle counts and timestamps, etc.), and a small amount of missing data is filled in using linear interpolation. Subsequently, the cleaned data is categorized and stored according to mold number and process type, constructing a structured historical database that supports rapid retrieval based on multi-dimensional conditions. For example, it can filter out historical data subsets with the same mold model and consistent process parameters as the current one, ensuring the relevance of subsequent trajectory matching.

[0054] It should be noted that the numerical matching of the coupled failure risk index and the historical risk index evolution sequence adopts a dynamic similarity calculation method. The core is to select the historical trajectory that is closest to the current mold decline trend. First, the current coupled failure risk index and the risk index sequence of the last 5 detections are extracted to form the current decline trend segment. Then, the complete historical risk index evolution sequence under the same mold model and similar process parameters is retrieved from the historical database (each sequence contains all risk indices from the mold's commissioning to its retirement and the corresponding cumulative production cycle number). The similarity calculation adopts the Dynamic Time Warping (DTW) algorithm, which can effectively handle the sequence matching problem of different lengths and time scales. By calculating the distance between the current trend segment and each sub-segment in the historical sequence, the smaller the distance, the higher the similarity. A similarity threshold of 0.85 (distance value less than 0.15) is set, and historical sequences with a distance less than this threshold are selected as similar historical decline trajectory data.

[0055] It should be noted that when constructing the linear regression training set, feature extraction and normalization are performed on each similar historical decline trajectory data. Using the cumulative number of production cycles in each trajectory as the independent variable (x) and the corresponding coupling failure risk index as the dependent variable (y), all data points of the risk index from the current similar starting point to the critical failure threshold are extracted from the trajectory to form the corresponding sample pair (x_i, y_i). All sample pairs of similar trajectories are aggregated to form the linear regression training set, ensuring that the training set contains a sufficient number of samples (at least 500 sample pairs) to guarantee the fitting accuracy of the regression model.

[0056] Specifically, when fitting the training set using a linear regression algorithm, the parameters (slope k and intercept b) of the linear equation y=kx+b are solved using the least squares method, where k reflects the rate at which the risk index increases with the number of iterations, and b is the initial offset. During the fitting process, cross-validation is used to evaluate the model performance. The training set is divided into a training subset and a validation subset in a 7:3 ratio. Through multiple iterations of training and parameter optimization, the model's coefficient of determination (R²) on the validation subset is ensured to be no less than 0.92, indicating that the model can fit the linear relationship between the risk index and the number of iterations well.

[0057] In another implementation, the preset critical failure threshold is determined based on the mold's safe service requirements and engineering practice standards. In this embodiment, it is set to 0.8, meaning that when the coupled failure risk index reaches 0.8, the mold's failure risk has reached an unacceptable level, requiring it to be taken out of service and repaired or replaced. Substituting this critical failure threshold (y=0.8) into the fitted linear equation and solving for the corresponding independent variable x yields the theoretical total cycle life value. The remaining cycle life prediction is calculated by the difference between the theoretical total cycle life value and the current cumulative production cycle count. The current cumulative production cycle count is directly read from the mold's operation log to ensure the accuracy of the value.

[0058] In step S17, if the remaining lifespan prediction cycle count is less than the preset safe lifespan cycle count, then the maintenance timing is determined based on the remaining lifespan prediction cycle count, a maintenance timing alarm signal is generated and output, all calculated data is recorded, and the final decision basis is obtained, including: If the remaining lifespan prediction cycle number is less than the preset safe cycle number, a high-risk status indicator is generated, and the maintenance timing cutoff point is determined in conjunction with the production plan and equipment operation schedule. A maintenance timing alarm signal is constructed based on the high-risk status indicator and the maintenance timing cutoff point; In response to the maintenance timing alarm signal, the predicted data, risk identifiers, and calculated data of maintenance cut-off points are extracted and structured log entries are assembled. The structured log entries are written into the system log storage queue to generate the basis for the final decision.

[0059] In one implementation, the preset safety cycle count in this embodiment is set to 200 production cycles. This value is derived from statistical analysis of a large number of historical maintenance cases, ensuring sufficient time is allocated to complete maintenance preparations after an early warning. When the predicted remaining lifespan cycles output in step S16 are less than 200, it indicates that the mold has entered a high-risk service stage, and the maintenance early warning process must be initiated immediately. For example, in a certain inspection, the predicted remaining lifespan cycles are 180, which is less than the safety threshold of 200, and the system automatically generates a high-risk status indicator.

[0060] Specifically, the high-risk status identification adopts a standardized hierarchical coding rule, including three core dimensions: risk level, remaining cycle count warning range, and urgency level. The risk level is divided into three levels: "high-risk," "medium-risk," and "low-risk." A remaining cycle count of less than 100 cycles is marked as high-risk, 100-150 cycles as medium-risk, and 150-200 cycles as low-risk. The urgency level is correspondingly divided into three levels: immediate shutdown for maintenance, priority maintenance, and planned maintenance. For example, a high-risk status identification corresponding to 180 remaining cycles is "Risk Level: Low-risk; Warning Range: 150-200 cycles; Urgency Level: Planned Maintenance."

[0061] It's important to note that determining the maintenance cutoff point requires balancing equipment safety and production efficiency, employing a dual strategy of production plan adaptation and remaining lifespan buffer. First, the system establishes communication with the production line's MES system via an industrial Ethernet interface to obtain production schedule data for the next 30 days, filtering out non-critical production task periods or production gaps as candidate maintenance windows. Then, based on the predicted remaining cycles, a buffer margin is calculated. The maintenance cutoff point is set as the current cumulative production cycle count plus the remaining cycle count minus the buffer margin, with the buffer margin set at 30 production cycles to ensure maintenance is completed before the cutoff point, avoiding production interruptions due to sudden failures. If the cycle count corresponding to a candidate maintenance window differs from the calculated cutoff point, the system automatically selects the cycle count closest to the cutoff point and falling within a non-critical production period as the final maintenance cutoff point.

[0062] It should be noted that the construction of maintenance timing alarm signals adopts a multi-channel synchronous push mechanism to ensure that relevant personnel receive early warning information in a timely manner. The alarm signal includes four core information modules: basic information (equipment number, mold model, current cycle count), risk information (high-risk status indicator, remaining lifespan cycle count, maintenance timing cutoff point), data basis (coupled failure risk index, estimated microcrack size, coating integrity quantification score), and handling suggestions (maintenance type, required spare parts, estimated maintenance duration). The assembly of structured log entries aims to achieve full-process data traceability, ensuring that maintenance decisions are reviewable and verifiable. The system automatically extracts all key calculation data from steps S11 to S16 and constructs structured logs in the format of "data type-calculation result-timestamp-related parameters". Log entries contain core information including image acquisition data (resolution, shooting time), vibration signal data (sampling frequency, high-frequency component amplitude), feature extraction data (multidimensional wear feature vector, peak frequency offset), model calculation data (coating integrity quantification score, microcrack size estimate, coupling failure risk index, remaining life prediction cycles), and early warning decision data (high-risk status identifier, maintenance timing cutoff point).

[0063] In another implementation, this embodiment employs a time-series storage and multiple backup strategies for the system log storage queue to ensure data security and rapid retrieval. Log entries are written to an industrial-grade database in timestamp order, using a dual storage mechanism of local storage and cloud backup. Local storage retains log data from the most recent year, while cloud backup permanently stores all log data to prevent data loss. The database supports multi-dimensional retrieval by device number, time range, data type, etc., facilitating the equipment management team to trace the mold failure process, verify the accuracy of prediction models, and optimize maintenance strategies. For example, after a mold has been repaired, the log entries for that mold can be retrieved to compare various data indicators before and after repair, evaluating the repair effectiveness. When there is a deviation between the predicted results and the actual lifespan, the model parameters and threshold settings can be optimized by analyzing changes in key parameters in the log data. The final decision-making basis is generated by combining data aggregation and expert rules, transforming scattered log data into actionable maintenance decision recommendations. The system first aggregates structured log entries, calculates the changing trends of key indicators (such as the growth rate of the coupling failure risk index and the propagation speed of microcrack size), and compares them with similar indicators in historical maintenance cases. Then, it calls a pre-set expert rule base to generate targeted maintenance decision suggestions based on factors such as risk level, maintenance timing, and production plan.

[0064] It is worth noting that the preset thresholds involved in this invention, such as the preset quantitative score threshold of 50 points and the number of safe cycles of 200 times, can be determined by statistical analysis of the score and remaining cycles of the last safe inspection before failure in a large amount of historical service data of the same type of mold, such as taking the average value or the lower limit of a certain confidence interval; the preset critical size of 1.0 mm can be selected as the critical point before the crack enters the accelerated propagation stage, based on the fatigue crack propagation rate curve of the matrix material and combined with the mold design safety factor.

[0065] In summary, this invention discloses a method for predicting the fatigue life of the coating and substrate on the surface of a drawing die. This method includes simultaneously acquiring surface image data and mechanical vibration signal data of the die using an industrial camera and an accelerometer, and retrieving historical production cycle data containing historical risk index evolution sequences and corresponding cumulative production cycle counts. The surface image data is processed through convolution operations and dynamic threshold segmentation to generate a coating degradation feature distribution map. A multidimensional wear feature vector is then extracted, and a coating integrity quantification score is obtained using a support vector machine model. When this score is below a preset threshold, the vibration signal is subjected to high-frequency filtering and Fourier transform. A fatigue vibration spectrum of the substrate is constructed, and the peak frequency shift and amplitude attenuation value are calculated. The estimated value of microcrack size is obtained by combining support vector machine classification and database retrieval. If the microcrack size exceeds the critical size, hyperspectral scanning is initiated to supplement the coating status. After normalizing the two types of indicators, the coupled failure risk index is obtained by weighting them using the entropy weight method. The index is matched with historical data for similar trajectories, and the remaining number of life prediction cycles is obtained by linear regression fitting. When the remaining number of cycles is less than the number of safe cycles, an alarm signal containing the risk level and maintenance timing is generated. The structured log is assembled and stored to form a traceable final decision basis.

[0066] It is worth noting that this invention is based on a digital production management platform for stamping dies. This platform is a comprehensive system integrating multiple technologies, covering the entire process of digital production management from design to manufacturing to inspection, forming a closed-loop system from process simulation, production scheduling, machining control to quality inspection. The design and planning stage includes a die surface correction system, which corrects the die surface using simulation and measured data to guide the final die surface finishing. The manufacturing stage includes an intelligent production scheduling and dispatching system, which improves the dynamic adaptability and execution accuracy of the scheduling plan through dynamic scheduling, order insertion response, and production scheduling optimization; and an adaptive machining system, which improves the accuracy and efficiency of die remachining through machine measurement, error compensation, and adaptive machining. The inspection stage includes a die cutting edge wear life prediction system, which achieves high-precision cutting edge wear life prediction. The prediction results affect the equipment health in the production scheduling system, facilitating preventative maintenance scheduling; and a drawing die critical surface life prediction system, which predicts the remaining life by monitoring the health status of the die body, providing support for maintenance plans.

[0067] Reference Figure 2 The second embodiment of the present invention provides a fatigue life prediction system for the coating and substrate of a drawing die, comprising: The data acquisition and preprocessing module is used to acquire surface image data and mechanical vibration signal data of the mold, obtain historical production cycle data of the mold, and perform image processing on the surface image data to obtain a coating degradation feature distribution map. The coating integrity assessment module is used to extract a multi-dimensional wear feature vector based on the coating degradation feature distribution map, and input the multi-dimensional wear feature vector into the training preset coating wear model to obtain a coating integrity quantification score. The substrate fatigue spectrum analysis module is used to filter and extract high-frequency vibration components from the mechanical vibration signal data if the coating integrity quantization score is lower than a preset quantization score threshold, and perform spectrum analysis on the high-frequency vibration components to obtain a substrate fatigue vibration spectrum diagram. The microcrack size estimation module is used to calculate the peak frequency offset and amplitude attenuation value based on the fatigue vibration spectrum of the matrix, and to classify and retrieve the peak frequency offset and amplitude attenuation value based on a preset support vector machine model to obtain the microcrack size estimate. The coupling failure risk calculation module is used to normalize the coating integrity quantification score and the microcrack size estimate if the estimated value of the microcrack size exceeds the preset critical size, and then calculate the coupling failure risk index by weighting. The remaining life prediction module is used to match the coupled failure risk index with the historical risk index sequence stored in the historical production cycle data, filter out similar historical decline trajectories, perform linear regression fitting based on the similar historical decline trajectories to obtain the life prediction model, and calculate the remaining life prediction cycle number based on the life prediction model and the preset critical failure threshold. The alarm and decision module is used to determine the maintenance timing based on the predicted remaining lifespan if the predicted remaining lifespan is less than the preset safe lifespan, generate and output a maintenance timing alarm signal, record all calculation data, and obtain the final decision basis.

[0068] It should be noted that the fatigue life prediction system for the coating on the surface of the drawing die and the substrate provided in this embodiment of the invention is used to execute all the process steps of the fatigue life prediction method for the coating on the surface of the drawing die and the substrate in the above embodiment. The working principle and beneficial effect of the two are one-to-one, so they will not be described again.

[0069] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0070] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for predicting the fatigue life of a draw die profile coating and substrate, characterized by, include: Collect surface image data and mechanical vibration signal data of the mold, obtain historical production cycle data of the mold, and perform image processing on the surface image data to obtain a coating degradation feature distribution map; A multidimensional wear feature vector is extracted based on the coating degradation feature distribution map. The multidimensional wear feature vector is then input into a preset coating wear model to obtain a coating integrity quantification score. If the coating integrity quantization score is lower than the preset quantization score threshold, then high-frequency vibration components are filtered and extracted from the mechanical vibration signal data, and the high-frequency vibration components are subjected to spectral analysis to obtain the substrate fatigue vibration spectrum. The peak frequency offset and amplitude attenuation value are calculated based on the fatigue vibration spectrum of the matrix. The peak frequency offset and amplitude attenuation value are then classified and retrieved based on a preset support vector machine model to obtain the estimated value of the microcrack size. If the estimated microcrack size exceeds the preset critical size, the coating integrity quantification score and the estimated microcrack size are normalized and weighted to obtain the coupling failure risk index. The coupling failure risk index is matched with the historical risk index sequence stored in the historical production cycle data to screen out similar historical decline trajectories. Linear regression fitting is performed based on the similar historical decline trajectories to obtain the life prediction model. The remaining life prediction cycle number is calculated according to the life prediction model and the preset critical failure threshold. If the remaining lifespan prediction cycle number is less than the preset safe lifespan cycle number, then the maintenance timing is determined based on the remaining lifespan prediction cycle number, a maintenance timing alarm signal is generated and output, all calculation data is recorded, and the final decision basis is obtained.

2. The method of predicting the life of a draw die profile coating and substrate fatigue according to claim 1, characterized by, The process involves acquiring surface image data and mechanical vibration signal data of the mold, obtaining historical production cycle data of the mold, and performing image processing on the surface image data to obtain a coating degradation feature distribution map, including: Acquire surface image data of the mold, mechanical vibration signal data, and historical production cycle data containing historical risk index evolution sequences and corresponding cumulative production cycle counts; The surface image data is convolved to extract horizontal and vertical gradients and synthesize a gradient magnitude map. Extract the high-frequency components from the mechanical vibration signal data. If the amplitude of the high-frequency components exceeds a preset amplitude threshold, an adaptive dynamic threshold coefficient is obtained by weighted calculation. The adaptive dynamic threshold coefficient is applied to perform double threshold hysteresis segmentation on the gradient magnitude map to obtain a surface defect profile map; Based on the feature quantification and scoring of the area, perimeter, and shape complexity of each defect region in the surface defect profile map, a local degradation index matrix is ​​constructed, and the local degradation index matrix is ​​used to obtain a coating degradation feature distribution map through pseudo-color mapping.

3. The method of predicting the life of a draw die profile coating and substrate fatigue according to claim 1, characterized by, The step of extracting a multidimensional wear feature vector based on the coating degradation feature distribution map, and inputting the multidimensional wear feature vector into a preset coating wear model to obtain a coating integrity quantification score includes: Morphological gradient calculations were performed on the coating degradation feature distribution map to extract the coordinates of thickness variation regions and peeling points; A multidimensional wear feature vector is constructed based on the average grayscale attenuation rate of the thickness variation region, the proportion of the total area of ​​the variation region, the number of peeling points, the average cluster density of the peeling points, and the maximum cluster radius. The multidimensional wear feature vector is input into a preset coating wear model, and a support vector machine is used to map it to a high-dimensional space with a radial basis kernel function. The optimal classification hyperplane is then solved to obtain the classification decision value. The Euclidean distance between the classification decision value and the preset center of the intact coating sample is calculated, and the coating integrity quantification score is obtained through normalization mapping.

4. The method for predicting the fatigue life of the coating and substrate on the surface of a drawing die according to claim 1, characterized in that, If the coating integrity quantization score is lower than a preset quantization score threshold, then high-frequency vibration components are filtered and extracted from the mechanical vibration signal data, and spectral analysis is performed on the high-frequency vibration components to obtain a substrate fatigue vibration spectrum, including: If the coating integrity quantization score is lower than the preset quantization score threshold, then synchronously acquired mechanical vibration signal data is obtained. The cutoff frequency is determined based on the coating integrity quantization score, and the mechanical vibration signal data is filtered by a high-pass filter to extract the high-frequency vibration component. Spectral analysis is performed on the high-frequency vibration components to convert the time-domain signal into a frequency-domain signal, thereby obtaining vibration energy distribution data; Extract fatigue characteristic frequency points with high amplitude intensity from the vibration energy distribution data, and construct a fatigue vibration spectrum diagram of the matrix based on the fatigue characteristic frequency points.

5. The method for predicting the fatigue life of the coating and substrate on the surface of a drawing die according to claim 1, characterized in that, The peak frequency offset and amplitude attenuation value are calculated based on the fatigue vibration spectrum of the matrix. Then, based on a preset support vector machine model, the peak frequency offset and amplitude attenuation value are classified and retrieved to obtain an estimated microcrack size, including: Analyze the energy distribution of the fatigue vibration spectrum of the matrix to determine the frequency point of maximum energy; The difference between the maximum energy frequency point and the reference frequency point under normal operating conditions of the equipment is calculated to obtain the peak frequency offset; Based on the spectral position corresponding to the peak frequency offset, calculate the amplitude attenuation value of the corresponding spectral position compared to the normal state; The peak frequency offset and the amplitude attenuation value are combined to construct the feature vector to be judged, and the vector is input into a preset support vector machine model to determine its position in the classification hyperplane. The crack level label is identified based on the location, and a preset database is retrieved based on the crack level label to obtain the estimated value of the microcrack size.

6. The method for predicting the fatigue life of the coating and substrate on the surface of a drawing die according to claim 1, characterized in that, If the estimated microcrack size exceeds a preset critical size, the coating integrity quantification score and the estimated microcrack size are normalized, and a weighted calculation is performed to obtain a coupling failure risk index, including: If the estimated size of the microcrack exceeds the preset critical size, the hyperspectral scanning program on the coating surface is started to obtain hyperspectral scanning data of the coating surface; The coating integrity quantification score is obtained by analyzing the spectral reflectance characteristics of the hyperspectral scanning data. The estimated microcrack size and the quantification score of coating integrity are normalized to unify the dimensions, and the objective weighting factors of the estimated microcrack size and the quantification score of coating integrity are calculated. The coupling failure risk index is obtained by weighted averaging the normalized values ​​and the corresponding objective weighting factors.

7. The method for predicting the fatigue life of the coating and substrate on the surface of a drawing die according to claim 1, characterized in that, The process involves matching the coupled failure risk index with the historical risk index sequence stored in the historical production cycle data to screen out similar historical decline trajectories, performing linear regression fitting based on these similar historical decline trajectories to obtain a lifespan prediction model, and calculating the remaining lifespan cycle number based on the lifespan prediction model and a preset critical failure threshold, including: The coupling failure risk index is numerically matched with the evolution sequence of the historical risk index in the historical production cycle data to filter out similar historical decline trajectory data. Extract the similar historical decline trajectory data to construct a linear regression training set, use the linear regression algorithm to fit the training set to obtain regression model parameters, and construct a life prediction model based on the regression model parameters; The theoretical total cycle life value is calculated by substituting the preset critical failure threshold into the life prediction model. The difference between the theoretical total cycle life value and the current cumulative number of production cycles completed is calculated to obtain the number of remaining cycles for the life prediction.

8. The method for predicting the fatigue life of the coating and substrate on the surface of a drawing die according to claim 1, characterized in that, If the remaining lifespan predicted is less than the preset safe lifespan, then a maintenance opportunity is determined based on the remaining lifespan predicted, a maintenance opportunity alarm signal is generated and output, all calculation data is recorded, and the final decision basis is obtained, including: If the remaining lifespan prediction cycle number is less than the preset safe cycle number, a high-risk status indicator is generated, and the maintenance timing cutoff point is determined in conjunction with the production plan and equipment operation schedule. A maintenance timing alarm signal is constructed based on the high-risk status indicator and the maintenance timing cutoff point; In response to the maintenance timing alarm signal, the predicted data, risk identifiers, and calculated data of maintenance cut-off points are extracted and structured log entries are assembled. The structured log entries are written into the system log storage queue to generate the basis for the final decision.

9. A fatigue life prediction system for the coating and substrate of a drawing die, characterized in that, include: The data acquisition and preprocessing module is used to acquire surface image data and mechanical vibration signal data of the mold, obtain historical production cycle data of the mold, and perform image processing on the surface image data to obtain a coating degradation feature distribution map. The coating integrity assessment module is used to extract a multi-dimensional wear feature vector based on the coating degradation feature distribution map, and input the multi-dimensional wear feature vector into the training preset coating wear model to obtain a coating integrity quantification score. The substrate fatigue spectrum analysis module is used to filter and extract high-frequency vibration components from the mechanical vibration signal data if the coating integrity quantization score is lower than a preset quantization score threshold, and perform spectrum analysis on the high-frequency vibration components to obtain a substrate fatigue vibration spectrum diagram. The microcrack size estimation module is used to calculate the peak frequency offset and amplitude attenuation value based on the fatigue vibration spectrum of the matrix, and to classify and retrieve the peak frequency offset and amplitude attenuation value based on a preset support vector machine model to obtain the microcrack size estimate. The coupling failure risk calculation module is used to normalize the coating integrity quantification score and the microcrack size estimate if the estimated value of the microcrack size exceeds the preset critical size, and then calculate the coupling failure risk index by weighting. The remaining life prediction module is used to match the coupled failure risk index with the historical risk index sequence stored in the historical production cycle data, filter out similar historical decline trajectories, perform linear regression fitting based on the similar historical decline trajectories to obtain the life prediction model, and calculate the remaining life prediction cycle number based on the life prediction model and the preset critical failure threshold. The alarm and decision module is used to determine the maintenance timing based on the predicted remaining lifespan if the predicted remaining lifespan is less than the preset safe lifespan, generate and output a maintenance timing alarm signal, record all calculation data, and obtain the final decision basis.

Images