Die edge wear life prediction method and system based on punch number

Abstract

The application relates to the technical field of intelligent manufacturing, and discloses a die edge wear life prediction method and system based on stamping times. The method comprises the following steps: collecting continuous pressure signals in a stamping process and intercepting a single stamping tonnage waveform sequence, and constructing a two-dimensional tonnage evolution matrix; performing manifold clustering analysis on the two-dimensional tonnage evolution matrix to determine a current wear stage label and a corresponding standard tonnage growth gradient threshold value; inputting the two-dimensional tonnage evolution matrix and the wear stage label into a preset deep time sequence prediction model to obtain a basic residual life prediction value; calculating a real-time tonnage growth gradient of the two-dimensional tonnage evolution matrix and combining the standard tonnage growth gradient threshold value to generate a degradation rate deviation factor; and weighting and correcting the basic residual life prediction value by using the degradation rate deviation factor to obtain a final edge wear life evaluation result. The method can realize high-precision die dynamic life prediction.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing technology, and in particular to a method and system for predicting the wear life of die cutting edges based on the number of stamping cycles. Background Technology

[0002] Currently, with the increasing precision requirements of modern manufacturing for stamped parts, the wear condition of the cutting edge of stamping dies, as a core forming equipment, directly affects the forming quality and production efficiency of products. Accurately predicting the wear life of the die cutting edge is of great significance for formulating reasonable maintenance plans and preventing batch scrap and unexpected downtime caused by cutting edge failure.

[0003] In existing technologies, die life management primarily relies on empirical statistics or simple counters, i.e., setting a fixed number of stamping cycles as the maintenance period. Some improved solutions integrate a data acquisition and control module to monitor the maximum tonnage value during the stamping process in real time. When the monitored peak pressure exceeds a preset safety threshold, the system determines that the die wear is excessive and issues an alarm signal. However, this approach has several drawbacks. First, relying solely on a fixed threshold or a single peak characteristic ignores the morphological evolution of the tonnage waveform throughout the stamping stroke, making it difficult to capture subtle mechanical changes caused by edge wear. Second, existing data acquisition and control logic often processes single signals in isolation, lacking the ability to perform time-series analysis and manifold feature mining of signal evolution trends throughout the entire life cycle. This prevents dynamic adjustment of the remaining life prediction value based on the die's current specific wear stage (e.g., stable or severe wear) and its real-time degradation rate, resulting in assessment results that often lag behind the actual wear state and fail to meet the demands of high-precision automated production.

[0004] Existing technologies suffer from low prediction accuracy and a lack of adaptability. Summary of the Invention

[0005] This invention provides a method and system for predicting the wear life of die cutting edges based on the number of stamping cycles, in order to solve the technical problems of low prediction accuracy and lack of adaptability in the prior art.

[0006] In a first aspect, to solve the above-mentioned technical problems, the present invention provides a method for predicting the wear life of a die cutting edge based on the number of stamping cycles, comprising: By collecting continuous pressure signals through a tonnage sensor installed on the stamping die, and using a sliding window algorithm to extract the pressure data of a single stamping stroke, a single stamping tonnage waveform sequence is obtained. The waveform sequence of single stamping tonnage is spliced ​​and stacked in the time dimension according to the increasing order of the number of stampings to construct a two-dimensional tonnage evolution matrix with the single sampling point as the horizontal axis and the cumulative number of stampings as the vertical axis. Manifold clustering analysis is performed on the two-dimensional tonnage evolution matrix to identify the current wear state of the mold, determine the corresponding wear stage label and the standard tonnage growth gradient threshold under the corresponding stage; The two-dimensional tonnage evolution matrix and the wear stage label are used as joint features input to a pre-set deep time series prediction model. By extracting spatiotemporal evolution features and performing regression calculations, the basic remaining life prediction value of the mold is output. Calculate the rate of change of peak tonnage within the most recent preset number of times in the two-dimensional tonnage evolution matrix to obtain the real-time tonnage growth gradient, calculate the ratio between the real-time tonnage growth gradient and the standard tonnage growth gradient threshold, and generate a degradation rate deviation factor. The baseline remaining life prediction value is weighted and corrected using the degradation rate deviation factor to obtain the final edge wear life assessment result.

[0007] Preferably, the step of acquiring continuous pressure signals through a tonnage sensor installed on the stamping die, and using a sliding window algorithm to extract pressure data for a single stamping stroke to obtain a single stamping tonnage waveform sequence includes: The raw voltage signal of the tonnage sensor is acquired at a preset high-frequency sampling rate and synchronously triggered based on the crank angle signal or slider displacement signal of the stamping equipment. Set effective pressure start threshold and end threshold, slide and scan on the original voltage signal according to the preset time window length, identify the segment in the original voltage signal where the pressure value is continuously higher than the start threshold, and extract it as the original single waveform; The original single waveform is resampled to map it into a feature vector of fixed length. An anomaly detection tree is constructed based on an isolated forest. The path length of the feature vector in the tree structure is calculated to generate an anomaly score. If the anomaly score is lower than a preset outlier threshold, the corresponding feature vector is retained and included in the single stamping tonnage waveform sequence.

[0008] Preferably, the step of splicing and stacking the single stamping tonnage waveform sequence in ascending order of stamping count along the time dimension to construct a two-dimensional tonnage evolution matrix with the single sampling point as the horizontal axis and the cumulative number of stamping counts as the vertical axis includes: Construct an initially empty two-dimensional matrix container, setting the number of columns of the two-dimensional matrix container to be equal to the fixed length of a single waveform in the single stamping tonnage waveform sequence, and the number of rows to be equal to the preset historical observation window size; According to the timestamp sequence of the number of stampings, the amplitude data in the waveform sequence of the single stamping tonnage is filled into the two-dimensional matrix container row by row to form the original data matrix; The original data matrix is ​​subjected to global max-min normalization to map the pressure amplitude within the matrix to the interval between 0 and 1, thus obtaining the two-dimensional tonnage evolution matrix.

[0009] Preferably, the step of performing manifold clustering analysis on the two-dimensional tonnage evolution matrix to identify the current wear state of the mold, and determining the corresponding wear stage label and the standard tonnage growth gradient threshold under the corresponding stage includes: The high-dimensional two-dimensional tonnage evolution matrix is ​​mapped to a low-dimensional feature space using t-SNE nonlinear dimensionality reduction mapping to obtain the dimensionality-reduced sample point set. The sample point set was clustered using the DBSCAN density clustering algorithm. Based on the time distribution order of the clusters, the clusters were sequentially labeled as the initial break-in stage, the stable wear stage, and the severe wear stage, which served as the wear stage labels. For each marked wear stage, the average slope of the tonnage peak value changing with the number of stampings in the historical samples of the corresponding stage is extracted, and the average slope is defined as the standard tonnage growth gradient threshold under the corresponding stage.

[0010] Preferably, the step of inputting the two-dimensional tonnage evolution matrix and the wear stage label as joint features into a pre-set deep temporal prediction model, and outputting the basic remaining life prediction value of the mold by extracting spatiotemporal evolution features for regression calculation, includes: Construct a hybrid deep learning model that includes convolutional neural network layers and long short-term memory network layers; The morphological features of a single waveform in the two-dimensional tonnage evolution matrix are extracted using the convolutional neural network layer, and the time degradation features of the waveform as the number of stampings evolves are captured using the long short-term memory network layer. The wear stage labels are vectorized using one-hot encoding and then spliced ​​and fused with the time degradation features in a fully connected layer. Linear regression is performed on the output layer of the hybrid deep learning model to directly output the basic remaining life prediction value in units of remaining stamping counts.

[0011] Preferably, the step of calculating the rate of change of peak tonnage within the most recent preset number of times in the two-dimensional tonnage evolution matrix to obtain the real-time tonnage growth gradient, and calculating the ratio between the real-time tonnage growth gradient and the standard tonnage growth gradient threshold to generate a degradation rate deviation factor includes: Extract the row vectors corresponding to the most recent N stampings in the two-dimensional tonnage evolution matrix, and obtain the maximum value of each row vector to form a peak sequence; The peak sequence is linearly fitted using the least squares method, and the slope of the fitted line is calculated. The slope is then determined as the real-time tonnage growth gradient. The degradation rate deviation factor is obtained by dividing the real-time tonnage growth gradient by the standard tonnage growth gradient threshold corresponding to the current wear stage label.

[0012] Preferably, the step of using the degradation rate deviation factor to weight and correct the basic remaining life prediction value to obtain the final edge wear life assessment result includes: The Kalman filter algorithm is used to smooth and track the multiple continuously generated degradation rate deviation factors to obtain the smoothing deviation factor. Construct an inverse correction formula, multiply the predicted base remaining life value by a preset standard weighting coefficient, and then divide by the smoothing deviation factor to obtain the correction value; The corrected values ​​are processed by moving average filtering to output the final edge wear life assessment result.

[0013] Secondly, the present invention provides a die edge wear life prediction system based on the number of stamping cycles, comprising: The graph construction module is used to splice and stack the single stamping tonnage waveform sequence in the time dimension according to the increasing order of the number of stampings, and construct a two-dimensional tonnage evolution matrix with the single sampling point as the horizontal axis and the cumulative number of stampings as the vertical axis. The working condition identification module is used to perform manifold clustering analysis on the two-dimensional tonnage evolution matrix, identify the current wear state of the mold, and determine the corresponding wear stage label and the standard tonnage growth gradient threshold under the corresponding stage. The basic prediction module is used to input the two-dimensional tonnage evolution matrix and the wear stage label as joint features into a preset deep time series prediction model, and output the basic remaining life prediction value of the mold by extracting spatiotemporal evolution features for regression calculation. The deviation calculation module is used to calculate the rate of change of the peak tonnage within the most recent preset number of times in the two-dimensional tonnage evolution matrix, obtain the real-time tonnage growth gradient, calculate the ratio between the real-time tonnage growth gradient and the standard tonnage growth gradient threshold, and generate the degradation rate deviation factor. The correction evaluation module is used to perform weighted correction on the basic remaining life prediction value using the degradation rate deviation factor to obtain the final edge wear life evaluation result.

[0014] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention constructs a two-dimensional tonnage evolution matrix by stacking discrete single stamping waveforms along the time dimension, and extracts spatiotemporal evolution features using a hybrid model containing convolutional neural networks and recurrent neural networks (such as Long Short-Term Memory networks LSTM), thereby achieving in-depth capture of the degradation law of the entire life cycle of the mold. Existing technologies often only focus on a single tonnage peak or an isolated single waveform, ignoring the subtle gradual changes in waveform morphology as the number of stampings increases. This invention integrates "transient waveform morphology" and "long-term evolution trend" into the same data structure, effectively overcoming the technical difficulties of one-sided feature extraction and low prediction accuracy in traditional methods.

[0015] (2) This invention automatically identifies the current wear stage of the mold (initial break-in, stable wear, severe wear) by introducing manifold clustering analysis (t-SNE and DBSCAN), and determines the standard tonnage growth gradient threshold unique to each stage, thus realizing adaptive benchmark setting based on physical state. Traditional methods usually adopt a uniform linear assumption or fixed threshold throughout the entire life cycle, which is difficult to adapt to the nonlinear characteristics of the wear process. This method dynamically matches the evaluation criteria according to the physical characteristics of different stages, significantly improving the scientificity and interpretability of the wear state judgment.

[0016] (3) This invention generates a degradation rate deviation factor by calculating the ratio of the real-time tonnage growth gradient to the standard gradient, and combines Kalman filtering to dynamically weight and correct the basic remaining life prediction value, thus establishing a closed-loop prediction mechanism of "historical big data-driven + real-time operating condition feedback". In response to the problem of sudden wear acceleration caused by material hardness fluctuations or lubrication failure in actual production, this mechanism can quickly detect abnormal deviations in the real-time degradation rate and automatically adjust the prediction results, effectively solving the problem of prediction lag caused by the inability of static models to respond to dynamic operating condition changes, and ensuring the real-time performance and robustness of life assessment. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the method for predicting the wear life of a die cutting edge based on the number of stamping cycles provided in the first embodiment of the present invention; Figure 2 This is a schematic diagram of the die edge wear life prediction system based on the number of stamping cycles provided in the second embodiment of the present invention. Detailed Implementation

[0018] 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.

[0019] Reference Figure 1 The first embodiment of the present invention provides a method for predicting the wear life of a die cutting edge based on the number of stamping cycles, including the following steps: S11: Continuous pressure signals are collected by a tonnage sensor installed on the stamping die, and the pressure data of a single stamping stroke is extracted by a sliding window algorithm to obtain a single stamping tonnage waveform sequence. S12, the single stamping tonnage waveform sequence is spliced ​​and stacked in the time dimension according to the increasing order of the number of stampings to construct a two-dimensional tonnage evolution matrix with the single sampling point as the horizontal axis and the cumulative number of stampings as the vertical axis; S13, perform manifold clustering analysis on the two-dimensional tonnage evolution matrix to identify the current wear state of the mold, and determine the corresponding wear stage label and the standard tonnage growth gradient threshold under the corresponding stage. S14, the two-dimensional tonnage evolution matrix and the wear stage label are used as joint features input to the preset deep time series prediction model. By extracting spatiotemporal evolution features, regression calculation is performed to output the basic remaining life prediction value of the mold. S15, calculate the rate of change of peak tonnage within the most recent preset number of times in the two-dimensional tonnage evolution matrix to obtain the real-time tonnage growth gradient, calculate the ratio between the real-time tonnage growth gradient and the standard tonnage growth gradient threshold, and generate the degradation rate deviation factor. S16, the basic remaining life prediction value is weighted and corrected using the degradation rate deviation factor to obtain the final edge wear life assessment result.

[0020] In step S11, continuous pressure signals are acquired by a tonnage sensor installed on the stamping die, and the pressure data of a single stamping stroke is extracted using a sliding window algorithm to obtain a single stamping tonnage waveform sequence, including: The raw voltage signal of the tonnage sensor is acquired at a preset high-frequency sampling rate and synchronously triggered based on the crank angle signal or slider displacement signal of the stamping equipment. Set effective pressure start threshold and end threshold, slide and scan on the original voltage signal according to the preset time window length, identify the segment in the original voltage signal where the pressure value is continuously higher than the start threshold, and extract it as the original single waveform; The original single waveform is resampled to map it into a feature vector of fixed length. An anomaly detection tree is constructed based on an isolated forest. The path length of the feature vector in the tree structure is calculated to generate an anomaly score. If the anomaly score is lower than a preset outlier threshold, the corresponding feature vector is retained and included in the single stamping tonnage waveform sequence.

[0021] It should be noted that setting effective pressure start and stop thresholds is achieved using baseline noise statistical analysis technology. This operation first collects the background noise signal of the press under no-load operation and calculates the average amplitude of the background noise signal. with standard deviation Based on the 3-Sigma criterion in statistics, select... The calculated results are used as the pressure start and end thresholds. This setting ensures that the cutting action is triggered only by effective forming pressure, effectively shielding against mechanical vibration and environmental electromagnetic noise interference.

[0022] It is worth noting that the predetermined time window length is determined based on statistical analysis of the temporal distribution characteristics of historical standard production data. The system pre-acquires a large number of historical single-press waveforms covering different process parameters, identifies the effective duration of each waveform, and constructs a duration distribution histogram. The duration value corresponding to the 99th percentile of this distribution is selected and multiplied by a predetermined safety redundancy coefficient (e.g., 1.2) as the predetermined time window length. This setting aims to ensure that even with fluctuations in process parameters or adjustments in equipment operating speed (such as reduced production speed), the sliding window can still completely cover the entire physical process from the cutting edge contacting the sheet metal to complete separation, preventing the loss of key features due to window truncation.

[0023] It is worth noting that the preset safety redundancy coefficient is used to ensure that the time window covers the stamping stroke of the equipment operating at the lowest permissible speed. It is determined based on the dynamic boundary of the equipment. The actual possible lowest speed of the equipment is the difference between 1 and the maximum speed drop rate, multiplied by the rated speed of the equipment. At the same time, according to the process card, the ratio of the effective speed to the rated speed under extreme slow conditions is approximately the difference between 1 and the maximum speed drop rate, multiplied by the ratio of the lowest permissible production speed to the rated speed. Since the stamping stroke time is inversely proportional to the speed, under extreme slow conditions, the time required for the same stroke will be extended to the reciprocal of the ratio of the effective speed to the rated speed at the rated speed. This reciprocal is the required safety redundancy coefficient, which is used to expand the reference time window to ensure full coverage of all operating conditions.

[0024] It should be noted that the resampling of the original single waveform is implemented using a cubic spline interpolation algorithm. Since minute fluctuations in the stamping speed result in a variable number of data points for the acquired original waveform, this operation first constructs a piecewise cubic polynomial function on the time axis of the original waveform and applies a constraint condition ensuring the continuity of the first and second derivatives at the nodes to guarantee waveform smoothness. Subsequently, a fixed number (e.g., 1024) of sampling points are selected at equal intervals within the normalized time interval, and the corresponding pressure amplitude is calculated. This process maps the original signals with varying physical durations into feature vectors of consistent dimension.

[0025] It should be noted that the construction of anomaly detection trees and the calculation of anomaly scores based on isolated forests are implemented using unsupervised anomaly detection techniques. The system inputs the feature vectors into the isolated forest algorithm, which constructs multiple binary trees (iTrees) by randomly selecting features and split points. For each input feature vector, the system calculates the average path length required to reach a leaf node from the root node. The physical logic is that anomalous waveforms (such as those caused by double-material or card-like patterns) typically have significantly different morphological characteristics, making them easier to isolate, and their path length is significantly shorter than that of normal waveforms. The system converts this path length into a value between 0 and 1 using an exponential mapping, which is the anomaly score.

[0026] It is worth noting that the preset high-frequency sampling rate is determined based on the spectral analysis of the transient impact signal from die stamping. By performing a Fast Fourier Transform (FFT) on historical high-precision pressure data, an effective cutoff frequency covering 99% of the signal energy is determined. Based on the Nyquist sampling theorem and engineering redundancy requirements, 10 to 20 times the effective cutoff frequency is selected as the preset high-frequency sampling rate (e.g., 20kHz) to fully capture the microsecond-level high-frequency oscillation characteristics of the cutting edge as it enters the material.

[0027] It is worth noting that the preset outlier threshold is determined based on statistical analysis of the distribution of abnormal scores in historical normal production batch data. The system calculates the abnormal scores of normal data throughout the entire lifecycle and constructs a histogram or probability density function of the score distribution. The 99th percentile of this distribution is selected as the preset outlier threshold (e.g., 0.65) to accurately remove abnormal wear-related outlier samples caused by sensor malfunctions or foreign object intrusion, ensuring the purity of the input to the subsequent lifespan prediction model.

[0028] For example, the tonnage sensor acquires signals at a sampling rate of 20 kHz. The initial pressure threshold is set to 0.5 tons. The system captures a raw single waveform containing 2050 data points. After cubic spline interpolation resampling, the waveform is converted into a feature vector of length 1024 points. Isolation forest calculations determine the outlier score of this vector to be 0.35. Since 0.35 is lower than the preset outlier threshold of 0.65, the system determines that the waveform is valid and stores it in the single stamping tonnage waveform sequence.

[0029] In step S12, the waveform sequence of single stamping tonnage is spliced ​​and stacked in ascending order of the number of stampings along the time dimension to construct a two-dimensional tonnage evolution matrix with the single sampling point as the horizontal axis and the cumulative number of stampings as the vertical axis, including: Construct an initially empty two-dimensional matrix container, setting the number of columns of the two-dimensional matrix container to be equal to the fixed length of a single waveform in the single stamping tonnage waveform sequence, and the number of rows to be equal to the preset historical observation window size; According to the timestamp sequence of the number of stampings, the amplitude data in the waveform sequence of the single stamping tonnage is filled into the two-dimensional matrix container row by row to form the original data matrix; The original data matrix is ​​subjected to global max-min normalization to map the pressure amplitude within the matrix to the interval between 0 and 1, thus obtaining the two-dimensional tonnage evolution matrix.

[0030] It should be noted that the construction of the two-dimensional matrix container and the filling of data are implemented using a queue data structure or sliding memory block based on the First-In-First-Out (FIFO) principle. This operation arranges the time-discrete, one-dimensional single-stroke waveform in an ordered spatial order. Physically, this is equivalent to stacking the transient snapshots of each stroke vertically along the time axis, thereby forming a spatiotemporal evolution map (i.e., the original data matrix) that can simultaneously reflect the "mechanical characteristics within a single stroke (horizontal axis)" and the "performance degradation trend throughout the entire life cycle (vertical axis)". In this matrix, each row represents a specific stamping event, while the subtle texture changes between rows contain the cumulative information of cutting edge wear.

[0031] It should be noted that the global maximum-minimum normalization process on the original data matrix is ​​implemented using a linear mapping transformation technique. This operation first traverses the entire original data matrix, searching for the global maximum and minimum pressure amplitudes. Then, for each element in the matrix, the difference between that element's value and the global minimum pressure amplitude is calculated, and this difference is divided by the difference between the global maximum and minimum pressure amplitudes (i.e., the range). This calculation process uniformly maps all pressure data to a closed interval between 0 and 1. This step eliminates the dimensional differences in absolute pressure values ​​between different tonnages and mold models, allowing subsequent deep learning models to focus on the distortion and evolution of waveform geometry, rather than being limited by the magnitude of the absolute values.

[0032] It is worth noting that the predetermined historical observation window size is determined based on a correlation analysis of the die wear rate and the predictive response timeliness. The system analyzes historical wear curves to calculate the minimum average number of stamping cycles required for a significant transition in wear state (e.g., from a stable wear period to a severe wear period). One-tenth of this interval is selected as the predetermined historical observation window size (e.g., the most recent 500 stamping cycles) to ensure that the input matrix contains sufficient historical evolution context information while maintaining a rapid response capability to changes in the current wear state, avoiding information lag due to an excessively large window.

[0033] For example, the length of a single stamping waveform is fixed at 1024 sampling points. The preset historical observation window size is 500 times. The system constructs a... The system reads the waveform data of the last 500 stamping operations and fills them into rows 1 to 500 of the matrix in chronological order. Assume the global maximum value in the matrix is ​​100 tons and the minimum value is 0 tons. For a point in the matrix with an element value of 50 tons, after normalization, (50-0) / (100-0) = 0.5. The final generated two-dimensional tonnage evolution matrix is ​​a value distributed between 0 and 1. A data block can be viewed as a single-channel grayscale image input to a subsequent model.

[0034] In step S13, manifold clustering analysis is performed on the two-dimensional tonnage evolution matrix to identify the current wear state of the mold, determine the corresponding wear stage label and the standard tonnage growth gradient threshold under the corresponding stage, including: The high-dimensional two-dimensional tonnage evolution matrix is ​​mapped to a low-dimensional feature space using t-SNE nonlinear dimensionality reduction mapping to obtain the dimensionality-reduced sample point set. The sample point set was clustered using the DBSCAN density clustering algorithm. Based on the time distribution order of the clusters, the clusters were sequentially labeled as the initial break-in stage, the stable wear stage, and the severe wear stage, which served as the wear stage labels. For each marked wear stage, the average slope of the tonnage peak value changing with the number of stampings in the historical samples of the corresponding stage is extracted, and the average slope is defined as the standard tonnage growth gradient threshold under the corresponding stage.

[0035] It should be noted that the t-SNE (t-Distributed Stochastic Neighbor Embedding) nonlinear dimensionality reduction mapping is implemented based on manifold learning theory. This operation treats each row vector in the two-dimensional tonnage evolution matrix as a data object in a high-dimensional space. The system first calculates the Euclidean distance between each pair of data points in the high-dimensional space and transforms it into a conditional probability under a Gaussian distribution to characterize the similarity between points; then, it constructs the t-distribution probability of the corresponding points in the low-dimensional feature space (usually a two-dimensional plane); finally, it uses the gradient descent algorithm to minimize the Kullback-Leibler divergence between the high-dimensional and low-dimensional distributions. This process preserves the local structural features of the high-dimensional waveform data, causing stamping waveforms with similar wear patterns to cluster together in the low-dimensional space, forming the dimensionality-reduced sample point set.

[0036] It should be noted that the DBSCAN (Density-Based Spatial Clustering of Applications with Noise) density clustering algorithm is implemented based on core objects and density reachability logic. This operation sets a preset neighborhood radius and minimum neighbor count. For any point in the sample point set, if the number of points within the neighborhood radius is greater than the minimum neighbor count, it is marked as a core point, and all points density-reachable from it are grouped into the same cluster. The DBSCAN algorithm can automatically discover clusters of arbitrary shapes and mark sparse points not belonging to any cluster as noise (corresponding to abnormal stamping). After clustering, the system reads the stamping count and timestamp of all sample points in each cluster and calculates the time centroid of each cluster. Based on the chronological order of the time centroids, the earliest appearing cluster is marked as the initial break-in stage, the middle clusters as the stable wear stage, and the latest clusters as the severe wear stage. The system determines which cluster's geometric boundary the latest input data point falls into, thus determining the wear stage label to which the mold currently belongs.

[0037] It is worth noting that the determination of the preset neighborhood radius and the minimum number of neighboring points is based on statistical analysis of the K-distance distribution characteristics. Regarding the minimum number of neighboring points, given that the feature space after t-SNE dimensionality reduction is two-dimensional, 4 is selected as the minimum number of neighboring points according to the dimensionality constraint criterion in statistics, to ensure that the formation of clusters has minimal statistical support. For the neighborhood radius, the elbow detection method of the K-distance graph is used. This method first calculates the Euclidean distance between each point in the sample set and its fourth nearest neighbor; secondly, it plots a distribution curve by arranging all distance values ​​in ascending order; finally, it selects the distance value corresponding to the inflection point of the curve slope abruptly as the preset neighborhood radius. This setting objectively defines the physical boundary between high-density regions and sparse noise regions.

[0038] It is worth noting that the determination of the standard tonnage growth gradient threshold is based on piecewise linear regression analysis of historical lifecycle data of the same type of mold. The system pre-collects data on a large number of historical failed molds and divides their lifecycle into three stages using the aforementioned clustering method. For each stage (e.g., the stable wear stage), the peak tonnage and cumulative stamping count of all sample points within that stage are extracted, and a univariate linear regression model is constructed with the cumulative stamping count as the independent variable and the peak tonnage as the dependent variable. The slope parameter of this model is solved using the least squares method, and the slope parameter of all historical samples is statistically averaged. The calculated average slope value is defined as the standard tonnage growth gradient threshold for that stage. This threshold objectively quantifies the normal wear rate benchmark of the mold at different lifecycle stages.

[0039] For example, the system receives the two-dimensional tonnage evolution matrix output in step S12, which contains waveform data from the most recent 500 stampings. After t-SNE dimensionality reduction, the feature coordinates of the latest 500th stamping on the low-dimensional two-dimensional plane are (2.15, 3.40). The system calls the DBSCAN clustering model built based on 100,000 historical stamping data. This model has formed three clusters arranged in chronological order: cluster A (initial break-in, center coordinates (-5.2, 1.1)), cluster B (stable wear, center coordinates (2.1, 3.5)), and cluster C (severe wear, center coordinates (8.6, -4.2)). The system calculates and finds that the current coordinates (2.15, 3.40) are within the density range of cluster B, therefore the current state of the mold is identified as the stable wear stage. Subsequently, the system queried the historical statistics database for this period and read that the standard tonnage growth gradient threshold for this period was 0.0025 tons / time (that is, for each additional stamping, the peak tonnage increases naturally by an average of 0.0025 tons).

[0040] In step S14, the two-dimensional tonnage evolution matrix and the wear stage label are used as joint features input to a preset deep temporal prediction model. Regression calculations are performed by extracting spatiotemporal evolution features to output the basic remaining life prediction value of the mold, including: Construct a hybrid deep learning model that includes convolutional neural network layers and long short-term memory network layers; The morphological features of a single waveform in the two-dimensional tonnage evolution matrix are extracted using the convolutional neural network layer, and the time degradation features of the waveform as the number of stampings evolves are captured using the long short-term memory network layer. The wear stage labels are vectorized using one-hot encoding and then spliced ​​and fused with the time degradation features in a fully connected layer. Linear regression is performed on the output layer of the hybrid deep learning model to directly output the basic remaining life prediction value in units of remaining stamping counts.

[0041] It should be noted that the construction of the hybrid deep learning model and the extraction of spatiotemporal evolution features are achieved using a hybrid model architecture that includes convolutional neural networks and recurrent neural networks, such as Long Short-Term Memory (LSTM) networks. Specifically, the model first treats the input two-dimensional tonnage evolution matrix as a time series sample set, where each row of the matrix corresponds to a time step. The system uses a one-dimensional convolutional layer (Conv1D) to independently extract features from each row (i.e., a single stamping waveform), capturing local morphological key points in the waveform (such as contact peaks, shearing plateaus, and unloading troughs) through sliding convolution kernels, compressing the original high-dimensional waveform data into a low-dimensional morphological feature vector. Subsequently, the morphological feature vector sequence arranged in chronological order is input into the LSTM network layer. The LSTM unit performs recursive calculations in the sequence dimension through gating mechanisms (forget gate, input gate, output gate), remembering the long-term gradual change trend of waveform features as the cumulative number of stampings increases, and finally outputting a deep hidden state vector containing historical evolution information, i.e., the temporal degradation feature.

[0042] It should be noted that the concatenation and fusion of the wear stage labels is achieved through a feature vector concatenation operation. The system first converts the discrete text labels (such as stable wear stages) output in step S13 into sparse vectors (e.g., vectors with only one element being 1 and the rest being 0) using one-hot encoding. Then, before the fully connected layer of the model, this sparse vector is concatenated with the deep hidden state vector output from the LSTM layer along the channel dimension to construct a joint feature vector that integrates prior information about the current lifecycle stage with dynamic information about the historical waveform evolution trend.

[0043] It is worth noting that the construction and training of the pre-built deep time series prediction model were completed offline before system deployment, based on historical full lifecycle data. The construction process is as follows: First, a large number of complete stamping tonnage records of the same type of mold from the installation of new cutting edges to scrapping are exported from the enterprise's historical database. Second, each record is back-annotated to calculate the actual remaining lifespan at each time point (i.e., total lifespan minus the current cumulative number of stampings). Then, a network architecture is constructed, consisting of three cascaded one-dimensional convolutional layers (Conv1D) and two long short-term memory (LSTM) layers. The number of convolutional kernels is set to 64, the kernel size is set to 3, the stride is set to 1, and the rectified linear unit (ReLU) is used as the activation function. The number of hidden units in the LSTM layer is set to 128. The initial learning rate is set to 0.001, the batch size is set to 32, the mean squared error (MSE) is selected as the loss function, and the Adam optimizer is used to iteratively update the model weights. Finally, when the prediction error of the model on the independent validation set is less than the preset convergence threshold, the model parameters are saved as a pre-set model.

[0044] It is worth noting that the preset convergence threshold is determined based on statistical analysis of the time granularity of production maintenance scheduling. This process first analyzes the smallest scheduling unit of the workshop production plan (typically a standard work shift, corresponding to approximately 2000 stamping operations); then, 10% (i.e., 200 operations) of this smallest scheduling unit value is selected as the preset convergence threshold. This setting aims to ensure that the model's average prediction deviation is significantly less than the length of a production shift, thereby guaranteeing that maintenance decisions are accurately implemented within specific work shifts and avoiding planning chaos caused by prediction errors spanning multiple shifts.

[0045] For example, the input two-dimensional tonnage evolution matrix contains historical data from the most recent 500 stampings, with each stamping containing 1024 sampling points. The wear stage is labeled as the stable wear stage. In feature extraction, the convolutional layer in the model compresses each row of waveform data into a short morphological feature vector, generating a feature map containing 500 time steps. In temporal modeling, a long short-term memory network layer processes this time series, outputting a deep hidden state vector that can characterize the long-term degradation trend. The system converts the stable wear stage label into a corresponding one-hot encoded vector. Subsequently, the system concatenates this one-hot encoded vector with the aforementioned deep hidden state vector to form a joint feature vector. A fully connected layer maps this joint feature vector to a scalar value. The model finally outputs the value "12500", indicating that the predicted basic remaining life of the die is 12500 stampings.

[0046] In step S15, the rate of change of peak tonnage within the most recent preset number of times in the two-dimensional tonnage evolution matrix is ​​calculated to obtain the real-time tonnage growth gradient. The ratio between the real-time tonnage growth gradient and the standard tonnage growth gradient threshold is calculated to generate a degradation rate deviation factor, including: Extract the row vectors corresponding to the most recent N stampings in the two-dimensional tonnage evolution matrix, and obtain the maximum value of each row vector to form a peak sequence; The peak sequence is linearly fitted using the least squares method, and the slope of the fitted line is calculated. The slope is then determined as the real-time tonnage growth gradient. The degradation rate deviation factor is obtained by dividing the real-time tonnage growth gradient by the standard tonnage growth gradient threshold corresponding to the current wear stage label.

[0047] It should be noted that the linear fitting of the peak sequence using the least squares method is based on the criterion of minimizing the sum of squared residuals. This operation first constructs a univariate linear regression model with the number of stampings as the independent variable and the peak tonnage as the dependent variable. Subsequently, the system mathematically solves for an optimal slope parameter that minimizes the sum of the squares of the vertical distances between each actual observation point in the peak sequence and the corresponding point on the fitted line. This calculated slope parameter quantifies the wear acceleration trend of the current mold within a very short time window and is defined as the real-time tonnage growth gradient.

[0048] It is worth noting that the value N for the most recent N stampings is determined based on autocorrelation analysis of short-term random fluctuation noise in the tonnage signal. The system calculates the autocorrelation function of the historical tonnage peak sequence to determine the lag step (i.e., decorrelation time) required for the correlation of the noise signal to decay to zero. Five times this lag step is selected as the preset value N (e.g., N=50) to ensure that the selected window length is sufficient to smooth out the random process fluctuations of a single stamping through fitting, thereby accurately reflecting the true short-term wear trend.

[0049] For example, the system extracts data from the most recent 50 stampings (i.e., N=50) from the two-dimensional tonnage evolution matrix. The peak tonnage of these 50 stampings forms a sequence that increases slightly with the number of stampings. After least squares fitting, the rising slope of this sequence is 0.012 tons / stamp (i.e., the real-time tonnage growth gradient). According to the identification result of step S13, the current die is in a stable wear stage, and the standard tonnage growth gradient threshold for this stage is 0.010 tons / stamp. The system performs a division operation. The final generated degradation rate deviation factor is 1.2, indicating that the current die wear rate is 20% faster than the standard rate for this stage.

[0050] In step S16, the basic remaining life prediction value is weighted and corrected using the degradation rate deviation factor to obtain the final edge wear life assessment result, including: The Kalman filter algorithm is used to smooth and track the multiple continuously generated degradation rate deviation factors to obtain the smoothing deviation factor. Construct an inverse correction formula, multiply the predicted base remaining life value by a preset standard weighting coefficient, and then divide by the smoothing deviation factor to obtain the correction value; The corrected values ​​are processed by moving average filtering to output the final edge wear life assessment result.

[0051] It should be noted that the smoothing tracking using the Kalman filter algorithm is achieved through state-space estimation. This operation treats the true deviation factor as an unobservable system state and the noisy deviation factor calculated in step S15 as an observed value. The system establishes state equations and observation equations, and through recursive calculations in two stages—prediction and update—dynamically adjusts the weights of the predicted and observed values ​​(i.e., Kalman gain) using the covariance matrix. This process effectively filters out the degradation rate calculation jitter caused by fluctuations in a single stamping process (such as uneven lubrication or slight changes in material thickness), estimating the optimal smoothing deviation factor from the noisy observation sequence.

[0052] It should be noted that the inverse correction formula is constructed based on the physical assumption that "wear rate is inversely proportional to remaining life". The specific calculation logic is as follows: the baseline remaining life prediction value output in step S14 is multiplied by a preset standard weighting coefficient to obtain the baseline life; subsequently, this baseline life is divided by the smoothing deviation factor. The physical meaning of this calculation logic is that when the smoothing deviation factor is greater than 1, it indicates that the current actual wear rate is higher than the standard level of the same historical period, therefore the remaining life expectation needs to be revised downwards; conversely, if it is less than 1, it is revised upwards.

[0053] It should be noted that the moving average filtering process applied to the corrected values ​​is implemented using a sliding window averaging technique. The system maintains a fixed-length (e.g., 5) first-in-first-out queue. This fixed length is determined based on a trade-off analysis between result stability and display lag, aiming to ensure that while effectively filtering out high-frequency calculation jitter, the signal delay introduced by the smoothing process is controlled within an operator-acceptable visual perception threshold (e.g., 0.5 seconds). This queue stores the most recently calculated corrected values. Before each output, the system calculates the arithmetic mean of all values ​​in the queue as the final edge wear life assessment result. This step further smooths the output curve, preventing frequent jumps in the prediction results on the display terminal and improving the user experience.

[0054] It is worth noting that the determination of the visual perception threshold is based on statistical experimental analysis of the audiovisual synchronization characteristics in a human-computer interaction environment. This operation first establishes a multimodal delay... The testing platform introduces a time delay variable with controllable step size between the physical impact sound of the stamping equipment (auditory stimulus) and the numerical display on the monitoring screen (visual stimulus). Subsequently, a random sample group meeting preset physiological criteria is selected. The preset physiological criteria are determined with reference to the international ergonomics standard (ISO 9241) regarding the audiovisual ability requirements of users of visual display terminals. Subjects are required to have corrected visual acuity of 1.0 or above in both eyes, and a pure-tone hearing threshold below 25 dB within the commonly used frequency range, to eliminate the interference of sensory defects on the experimental data.

[0055] The determination of the sample size is based on the Central Limit Theorem in statistics, which states that when the size of an independent and identically distributed random sample reaches 30 or more, the sampling distribution of its sample mean will approximately follow a normal distribution, thus ensuring that the experimental data meets the preconditions for parametric statistical testing and has sufficient statistical power. Next, a synchronicity discrimination task is performed using the constant stimulus method, recording the probability distribution of the sample group's judgment of audio-visual synchronization under different delay conditions. Finally, a psychophysical curve is constructed using the sigmoid logistic function to calculate the delay time corresponding to the decrease in recognition probability to a preset critical value, which is then determined as the visual perception threshold. This setting eliminates individual subjective differences through standardized psychophysical experiments, ensuring the scientific validity and universality of the system response delay index setting.

[0056] It is worth noting that the determination of the standard weight coefficients is based on statistical calibration of the model's average systematic bias on the validation set. During the offline training phase, the system calculates the average prediction error rate of the base model on the independent validation dataset. For example, if statistics show that the model tends to overestimate lifespan by an average of 10%, a calibration factor is constructed based on this error rate, for example, 1 divided by 1.1, and this factor is set as the standard weight coefficients to eliminate the model's inherent systematic bias.

[0057] For example, the base remaining lifetime output by the deep time-series prediction model is 10,000 cycles. Currently in a stable wear phase, the standard tonnage growth gradient threshold is 0.010. The most recently calculated real-time tonnage growth gradient is 0.012, generating a degradation rate bias factor of 1.2 (indicating a wear rate 20% faster than the standard). After Kalman filtering to eliminate transient fluctuations, a smoothing bias factor of 1.18 is obtained. The standard weighting coefficient is assumed to be 1.0 (no systematic bias). The system performs an inverse proportional correction calculation. This value is stored in a moving average queue. After averaging the results from previous calculations, the final evaluation result output to the user is 8470.

[0058] In summary, this invention constructs a two-dimensional tonnage evolution matrix containing the single waveform morphology and the evolution trend of the entire life cycle by collecting high-frequency pressure data during the stamping process and performing sliding window interception. This effectively overcomes the technical difficulties of fragmented and weak temporal correlation of traditional discrete monitoring data. By combining t-SNE dimensionality reduction and DBSCAN density clustering algorithm, the current physical wear state of the die is automatically identified and an adaptive growth gradient benchmark is matched, eliminating the blindness and misjudgment risk of setting static thresholds based on experience. A hybrid model of convolutional neural network and long short-term memory network is introduced to accurately capture the weak waveform distortion and long-term degradation characteristics caused by edge wear, solving the problem of difficulty in modeling nonlinear wear laws. Furthermore, the deviation factor generated by the real-time tonnage growth gradient is used, and the basic prediction value is dynamically closed-loop corrected by Kalman filtering, breaking through the bottleneck that static models cannot respond to sudden changes in working conditions, and realizing high robustness and high accuracy real-time prediction of die edge wear life.

[0059] 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.

[0060] Reference Figure 2 The second embodiment of the present invention provides a die cutting edge wear life prediction system based on the number of stamping cycles, comprising: The data acquisition and cleaning module is used to acquire continuous pressure signals through a tonnage sensor set on the stamping die, and to use a sliding window algorithm to extract the pressure data of a single stamping stroke to obtain a single stamping tonnage waveform sequence. The graph construction module is used to splice and stack the single stamping tonnage waveform sequence in the time dimension according to the increasing order of the number of stampings, and construct a two-dimensional tonnage evolution matrix with the single sampling point as the horizontal axis and the cumulative number of stampings as the vertical axis. The working condition identification module is used to perform manifold clustering analysis on the two-dimensional tonnage evolution matrix, identify the current wear state of the mold, and determine the corresponding wear stage label and the standard tonnage growth gradient threshold under the corresponding stage. The basic prediction module is used to input the two-dimensional tonnage evolution matrix and the wear stage label as joint features into a preset deep time series prediction model, and output the basic remaining life prediction value of the mold by extracting spatiotemporal evolution features for regression calculation. The deviation calculation module is used to calculate the rate of change of the peak tonnage within the most recent preset number of times in the two-dimensional tonnage evolution matrix, obtain the real-time tonnage growth gradient, calculate the ratio between the real-time tonnage growth gradient and the standard tonnage growth gradient threshold, and generate the degradation rate deviation factor. The correction evaluation module is used to perform weighted correction on the basic remaining life prediction value using the degradation rate deviation factor to obtain the final edge wear life evaluation result.

[0061] It should be noted that the die edge wear life prediction system based on the number of stamping cycles provided in this embodiment of the invention is used to execute all the process steps of the die edge wear life prediction method based on the number of stamping cycles described in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0062] 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.

[0063] 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 wear life of a die cutting edge based on the number of stamping cycles, characterized in that, include: By collecting continuous pressure signals through a tonnage sensor installed on the stamping die, and using a sliding window algorithm to extract the pressure data of a single stamping stroke, a single stamping tonnage waveform sequence is obtained. The waveform sequence of single stamping tonnage is spliced ​​and stacked in the time dimension according to the increasing order of the number of stampings to construct a two-dimensional tonnage evolution matrix with the single sampling point as the horizontal axis and the cumulative number of stampings as the vertical axis. Manifold clustering analysis is performed on the two-dimensional tonnage evolution matrix to identify the current wear state of the mold, determine the corresponding wear stage label and the standard tonnage growth gradient threshold under the corresponding stage, including: using t-SNE nonlinear dimensionality reduction mapping to map the high-dimensional two-dimensional tonnage evolution matrix to a low-dimensional feature space to obtain a dimensionality-reduced sample point set; using the DBSCAN density clustering algorithm to cluster the sample point set, and according to the temporal distribution order of the clusters, the clusters are sequentially labeled as the initial running-in stage, the stable wear stage, and the severe wear stage, as the wear stage label; for each labeled wear stage, the average slope of the tonnage peak value changing with the number of stampings in the historical samples of the corresponding stage is extracted, and the average slope is defined as the standard tonnage growth gradient threshold under the corresponding stage; The two-dimensional tonnage evolution matrix and the wear stage label are used as joint features input to a pre-set deep time series prediction model. By extracting spatiotemporal evolution features and performing regression calculations, the basic remaining life prediction value of the mold is output. Calculate the rate of change of peak tonnage within the most recent preset number of times in the two-dimensional tonnage evolution matrix to obtain the real-time tonnage growth gradient, calculate the ratio between the real-time tonnage growth gradient and the standard tonnage growth gradient threshold, and generate a degradation rate deviation factor. The baseline remaining life prediction value is weighted and corrected using the degradation rate deviation factor to obtain the final edge wear life assessment result.

2. The method for predicting the wear life of a die cutting edge based on the number of stamping cycles according to claim 1, characterized in that, The process involves acquiring continuous pressure signals through a tonnage sensor mounted on the stamping die, and using a sliding window algorithm to extract pressure data from a single stamping stroke to obtain a single stamping tonnage waveform sequence, including: The raw voltage signal of the tonnage sensor is acquired at a preset high-frequency sampling rate and synchronously triggered based on the crank angle signal or slider displacement signal of the stamping equipment. Set effective pressure start threshold and end threshold, slide and scan on the original voltage signal according to the preset time window length, identify the segment in the original voltage signal where the pressure value is continuously higher than the start threshold, and extract it as the original single waveform; The original single waveform is resampled to map it into a feature vector of fixed length. An anomaly detection tree is constructed based on an isolated forest. The path length of the feature vector in the tree structure is calculated to generate an anomaly score. If the anomaly score is lower than a preset outlier threshold, the corresponding feature vector is retained and included in the single stamping tonnage waveform sequence.

3. The method for predicting the wear life of a die cutting edge based on the number of stamping cycles according to claim 1, characterized in that, The step of splicing and stacking the single-stroke tonnage waveform sequence in ascending order of the number of strokes along the time dimension to construct a two-dimensional tonnage evolution matrix with the single sampling point as the horizontal axis and the cumulative number of strokes as the vertical axis includes: Construct an initially empty two-dimensional matrix container, setting the number of columns of the two-dimensional matrix container to be equal to the fixed length of a single waveform in the single stamping tonnage waveform sequence, and the number of rows to be equal to the preset historical observation window size; According to the timestamp sequence of the number of stampings, the amplitude data in the waveform sequence of the single stamping tonnage is filled into the two-dimensional matrix container row by row to form the original data matrix; The original data matrix is ​​subjected to global max-min normalization to map the pressure amplitude within the matrix to the interval between 0 and 1, thus obtaining the two-dimensional tonnage evolution matrix.

4. The method for predicting the wear life of a die cutting edge based on the number of stamping cycles according to claim 1, characterized in that, The deep time-series prediction model, which uses the two-dimensional tonnage evolution matrix and the wear stage label as joint features as input, extracts spatiotemporal evolution features for regression calculation and outputs the basic remaining life prediction value of the mold, including: Construct a hybrid deep learning model that includes convolutional neural network layers and long short-term memory network layers; The morphological features of a single waveform in the two-dimensional tonnage evolution matrix are extracted using the convolutional neural network layer, and the time degradation features of the waveform as the number of stampings evolves are captured using the long short-term memory network layer. The wear stage labels are vectorized using one-hot encoding and then spliced ​​and fused with the time degradation features in a fully connected layer. Linear regression is performed on the output layer of the hybrid deep learning model to directly output the basic remaining life prediction value in units of remaining stamping counts.

5. The method for predicting the wear life of a die cutting edge based on the number of stamping cycles according to claim 1, characterized in that, The calculation of the tonnage peak change rate within the most recent preset number of times in the two-dimensional tonnage evolution matrix yields the real-time tonnage growth gradient. The ratio between the real-time tonnage growth gradient and the standard tonnage growth gradient threshold is then calculated to generate a degradation rate deviation factor, including: Extract the row vectors corresponding to the most recent N stampings in the two-dimensional tonnage evolution matrix, and obtain the maximum value of each row vector to form a peak sequence; The peak sequence is linearly fitted using the least squares method, and the slope of the fitted line is calculated. The slope is then determined as the real-time tonnage growth gradient. The degradation rate deviation factor is obtained by dividing the real-time tonnage growth gradient by the standard tonnage growth gradient threshold corresponding to the current wear stage label.

6. The method for predicting the wear life of a die cutting edge based on the number of stamping cycles according to claim 1, characterized in that, The step of using the degradation rate deviation factor to weight and correct the predicted basic remaining life value to obtain the final edge wear life assessment result includes: The Kalman filter algorithm is used to smooth and track the multiple continuously generated degradation rate deviation factors to obtain the smoothing deviation factor. Construct an inverse correction formula, multiply the predicted base remaining life value by a preset standard weighting coefficient, and then divide by the smoothing deviation factor to obtain the correction value; The corrected values ​​are processed by moving average filtering to output the final edge wear life assessment result.

7. A die cutting edge wear life prediction system based on stamping cycles, characterized in that, include: The data acquisition and cleaning module is used to acquire continuous pressure signals through a tonnage sensor set on the stamping die, and to use a sliding window algorithm to extract the pressure data of a single stamping stroke to obtain a single stamping tonnage waveform sequence. The graph construction module is used to splice and stack the single stamping tonnage waveform sequence in the time dimension according to the increasing order of the number of stampings, and construct a two-dimensional tonnage evolution matrix with the single sampling point as the horizontal axis and the cumulative number of stampings as the vertical axis. The working condition identification module is used to perform manifold clustering analysis on the two-dimensional tonnage evolution matrix, identify the current wear state of the mold, and determine the corresponding wear stage label and the standard tonnage growth gradient threshold under the corresponding stage. This includes: using t-SNE nonlinear dimensionality reduction mapping to map the high-dimensional two-dimensional tonnage evolution matrix to a low-dimensional feature space, obtaining a dimensionality-reduced sample point set; using the DBSCAN density clustering algorithm to cluster the sample point set, and according to the temporal distribution order of the clusters, labeling the clusters sequentially as the initial break-in stage, stable wear stage, and severe wear stage, as the wear stage label; for each labeled wear stage, extracting the average slope of the tonnage peak value changing with the number of stamping cycles in the historical samples of the corresponding stage, and defining the average slope as the standard tonnage growth gradient threshold under the corresponding stage. The basic prediction module is used to input the two-dimensional tonnage evolution matrix and the wear stage label as joint features into a preset deep time series prediction model, and output the basic remaining life prediction value of the mold by extracting spatiotemporal evolution features for regression calculation. The deviation calculation module is used to calculate the rate of change of the peak tonnage within the most recent preset number of times in the two-dimensional tonnage evolution matrix, obtain the real-time tonnage growth gradient, calculate the ratio between the real-time tonnage growth gradient and the standard tonnage growth gradient threshold, and generate the degradation rate deviation factor. The correction evaluation module is used to perform weighted correction on the basic remaining life prediction value using the degradation rate deviation factor to obtain the final edge wear life evaluation result.

Images