A two-stage differentiated residual life prediction method of cold heading die

Abstract

The application provides a two-stage differentiated residual life prediction method of a cold upsetting die, comprising the following steps: step 1, periodically collecting relevant data in the running process of the cold upsetting die at preset interval of punch times; step 2, pre-processing and feature extraction, screening sensitive features to input a principal component analysis model to construct a health index; step 3, real-time monitoring of time series data of the health index to identify the inflection point of the die from the stable wear period to the accelerated wear period; and step 4, residual life prediction according to the health index. Through the dual judgment mechanism of the sliding window gradient change rate and the cumulative deviation, the identification from the stable wear period to the accelerated wear period is realized, the overall calculation amount is low, the wear condition change in the wear period can be followed in real time, and the prediction accuracy is improved.

Description

Technical Field

[0001] This invention relates to the field of cold heading die technology for fasteners, and particularly to a two-stage method for predicting the remaining life of cold heading dies. Background Technology

[0002] Cold heading dies are core equipment in the cold heading production of fasteners, and their wear condition directly determines the dimensional accuracy, surface quality, and production cost of the fasteners. When the cold heading die wears to a certain extent, fasteners may exhibit quality problems such as excessively large thread root diameter, insufficient head height, and flange face folding. In severe cases, sudden die breakage can lead to mass scrap and equipment downtime. Therefore, accurately predicting the remaining life of cold heading dies and replacing them as needed is of great significance for improving the economy and reliability of fastener production.

[0003] Currently, the die life prediction methods used by fastener manufacturers can be mainly divided into the following three categories: Category 1: Fixed-use-time replacement method. This method sets a fixed usage threshold for each type of mold based on historical experience, and forces the mold to be replaced once the threshold is reached. This method does not consider the impact of actual working conditions such as the hardness fluctuation of different batches of blanks, changes in lubrication conditions, and differences in the operating speed of cold heading machines, resulting in "early replacement wastes mold life, and late replacement leads to unstable product quality," and has extremely low prediction accuracy. The second type is the offline inspection method. This method involves periodically stopping the machine to remove the mold and using a profilometer or microscope to inspect the wear on the working surface of the mold. The remaining life is then estimated using empirical formulas. While this method estimates the remaining life by measuring the dimensional changes of the mold cavity, it requires stopping the machine to remove the mold, which seriously affects production efficiency. Furthermore, the mold may suddenly fail during the interval between two inspections, posing a safety hazard. The third category: Single-model online prediction method. With the development of industrial internet and artificial intelligence technologies, mold life prediction methods based on online monitoring data such as vibration signals and acoustic emission signals have emerged in recent years. For example, patent CN120974936A discloses a mold life assessment method based on machine learning. This method collects mold operation data, extracts multi-dimensional features, and constructs a single integrated model to output the remaining life prediction value.

[0004] However, the inventors of this application have discovered through long-term production practice and research that the aforementioned third type of method shares a common technical flaw: this type of method assumes that the wear rate of the mold remains constant throughout its entire lifecycle from commissioning to failure, i.e., it assumes that the degradation process is stable and linear. However, actual physical wear patterns show that the degradation process of cold heading dies exhibits significant two-stage characteristics: the first stage is the stable wear period, where the hardened layer on the die's working surface is intact, the wear rate is gradual, and health indicators decline slowly; the second stage is the accelerated wear period, where the hardened layer is worn through or microcracks initiate and propagate, the wear rate rises sharply, and health indicators rapidly approach the failure threshold. The single-model method uses the same set of model parameters throughout the entire lifecycle, leading to high-frequency noise interference during the stable wear period hindering the model's learning of the accelerated wear trend. Furthermore, it fails to quickly capture abrupt changes in the degradation rate during the accelerated wear period, ultimately resulting in significant prediction bias. Summary of the Invention

[0005] This invention addresses the aforementioned problems and aims to provide a two-stage remaining life prediction method for cold heading dies. By employing a dual judgment mechanism of sliding window gradient change rate and cumulative deviation, it achieves the identification of the transition from the stable wear stage to the accelerated wear stage. During the stable wear stage, an exponential smoothing model with extremely low computational overhead is used, while a deep learning model is activated only during the accelerated wear stage. The overall computational load is low, and it can follow the changes in wear conditions in real time, thereby improving prediction accuracy.

[0006] Specifically, the first aspect of the present invention provides a two-stage differentiated method for predicting the remaining life of a cold heading die, comprising: Step 1: At preset intervals, periodically collect vibration signals, cumulative stroke counts of the cold heading machine, and main slide speed during the operation of the cold heading die; The preset interval between strokes is 500 to 5000, which is selected based on the historical wear rate of the mold and the required prediction accuracy. Compared with the service life of more than 150,000 strokes of conventional molds, the amount of data collected is more controllable. If data is collected and calculated in real time, the amount of data is too large, and the data of most strokes is similar and not very valuable. The sampling frequency of the vibration signal is 20kHz to 30kHz, and the cumulative stroke count and main slide speed are read from the programmable logic controller of the cold heading machine through the OPC interface. Step 2: Preprocess and extract features from the collected raw data, select sensitive features to input into the principal component analysis model, and construct health indicators; Step 3: Monitor the time series data of health indicators in real time to identify the inflection point of the mold from the stable wear period to the accelerated wear period; Step 4: For test pieces whose health indicators are before the inflection point, use the first prediction model to predict the remaining lifespan; otherwise, use the second prediction model to predict the remaining lifespan.

[0007] Furthermore, step two includes: Step 2.1: Filter and denoise the acquired vibration signal, and extract time-domain and frequency-domain features; The acquired vibration signal was subjected to bandpass filtering for noise reduction, with a passband frequency range of 500Hz to 10kHz, and time-domain and frequency-domain features were extracted. Step 2.2: Use the main slider speed to perform a weighted correction on the cumulative stroke count and calculate the equivalent wear strokes; Step 2.3: Combine the extracted features with the equivalent wear impacts to form a multidimensional feature vector, input it into the principal component analysis model, and take the score of the first principal component as the original value of the health indicator; The five extracted vibration feature sequences are combined with the equivalent wear impact sequence to form a 6-dimensional feature vector. However, the equivalent wear impact does not participate in PCA dimensionality reduction and is only retained as an independent dimension to be used as the time axis in subsequent inflection point detection. After the vibration feature sequence is standardized, it is input into the principal component analysis model, and the score of the first principal component is taken as the original value of the health indicator. Step 2.4: Normalize the raw values ​​of the health indicators to obtain the final health indicators.

[0008] Furthermore, the time-domain and frequency-domain features include root mean square value, kurtosis index, margin index, energy proportion of the 2kHz-4kHz frequency band, and energy proportion of the 4kHz-8kHz frequency band.

[0009] The root mean square (RMS) value reflects the average energy of the signal. When the wear of the mold intensifies, the vibration amplitude increases, and the RMS value rises accordingly. The kurtosis index reflects the strength of the impact component in the signal and is very sensitive to early wear and micro-cracks; The margin index, also known as the gap factor, is the ratio of the signal peak value to the root mean square value. It is sensitive to changes in the peak value of the signal and reflects the relationship between the peak value and the overall amplitude level of the signal. When the mold wear intensifies, the peak value increases and the margin index increases. Energy percentage in the 2kHz-4kHz frequency band: reflects the mid-frequency wear characteristics of the mold, such as surface micro-cracks; Energy percentage in the 4kHz-8kHz frequency band: reflects the high-frequency impact characteristics of the mold, such as the peeling of cemented carbide particles.

[0010] The frequency band energy percentage is obtained by performing a fast Fourier transform on the signal to obtain the spectrum, and then calculating the percentage of energy in that frequency band relative to the total energy.

[0011] Furthermore, the equivalent wear strokes are calculated using the following formula: ; in: Equivalent wear strokes; This is the total number of impacts up to the current moment; The preset speed influence coefficient has a value range of 0.2-0.5. The real-time main slider speed during the i-th stroke; The preset standard speed is usually the rated operating speed of the cold heading machine.

[0012] Furthermore, step three includes: Step 3.1: Sort the sampling points of health indicators according to the equivalent wear impact to form a sparse time series. After smoothing the series, calculate the linear fitting slope within the sliding window. The sliding window contains 3 to 10 consecutive health indicator sampling points; Step 3.2: When the absolute value of the linear fitting slope continuously exceeds a preset first threshold and the duration is equivalent to 2 to 5 sliding windows, a suspected inflection point marker is triggered; The first set threshold is determined based on the statistical distribution of the slope of the health index during the stable wear period of similar molds, and is usually taken as 2-4 times the standard deviation of the absolute value of the average slope during the stable wear period. Step 3.3: After triggering the suspected inflection point marker, calculate the cumulative deviation between the actual value of the health indicator and the linear extrapolation value based on historical data of the stable wear period; Step 3.4: When the cumulative deviation exceeds the preset second threshold, confirm the inflection point and record the equivalent wear stroke position corresponding to the inflection point.

[0013] The second set threshold is set to 0.05-0.10. When the cumulative deviation exceeds the preset second set threshold, the inflection point is confirmed.

[0014] Furthermore, the preset first threshold is 2-4 times the standard deviation of the absolute value of the average slope of the health index data range within the first 2000 equivalent wear strokes after the same type of mold is put into use.

[0015] Furthermore, the linear extrapolation value based on historical data of stable wear period is obtained by taking health index data of 3-6 sampling points before the suspected inflection point trigger, fitting a linear trend line using the least squares method, and then substituting the equivalent wear impact of the current sampling point into the calculation.

[0016] Furthermore, the formula for calculating the cumulative deviation is: ; in: This is the cumulative deviation; This represents the number of sampling points from the time the suspected inflection point marker was triggered until the current time. The actual value of the health indicator at the j-th sampling point; It is the linear extrapolation value of the j-th sampling point.

[0017] For test pieces whose health indicators are before the inflection point, i.e., the current equivalent wear strokes are less than the inflection point equivalent wear strokes, the first prediction model is used to predict the remaining life; otherwise, i.e., the current equivalent wear strokes are greater than or equal to the inflection point equivalent wear strokes, the second prediction model is used to predict the remaining life.

[0018] Furthermore, the first prediction model is the Holt two-parameter exponential smoothing model.

[0019] The first prediction model adopts the Holt two-parameter exponential smoothing model, which includes two state variables: a horizontal component and a trend component. The horizontal smoothing parameter ranges from 0.1 to 0.3, and the trend smoothing parameter ranges from 0.03 to 0.1. The training method is to use the least squares method to fit the initial horizontal and initial trend values ​​in the initial stage of mold use. After that, for each new sampling point, the model parameters are updated by recursion formula. The values ​​of the two parameters are determined by minimizing the prediction error of the first 10 sampling points.

[0020] Furthermore, the second prediction model is a long short-term memory network.

[0021] The second prediction model is a Long Short-Term Memory (LSTM) network. This network includes an input layer, 1-2 LSTM hidden layers, an optional Dropout layer, and a fully connected output layer, where the LSTM hidden layers contain 16-64 LSTM units. The model undergoes supervised training using the first 5-10 sampling points after the inflection point as the initial training set, with mean squared error as the loss function and Adam as the optimizer. During prediction, a recursive prediction method is used, predicting health indicators point by point until a failure threshold is reached. When new sampling points are added, an incremental learning strategy is used to fine-tune the model parameters; when the prediction error continuously exceeds a set threshold, full retraining is triggered. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of this drawing or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this drawing. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0023] Figure 1 This is a flowchart of the steps of the present invention; The purpose, features, and advantages of this accompanying drawing will be further explained in conjunction with the embodiments and with reference to the accompanying drawing. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described and illustrated below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. All other embodiments obtained by those skilled in the art based on the embodiments provided by this invention without inventive effort are within the scope of protection of this invention.

[0025] Obviously, the accompanying drawings described below are merely some examples or embodiments of the present invention. Those skilled in the art can apply the present invention to other similar scenarios based on these drawings without any inventive effort. Furthermore, it is understood that although the efforts made in this development process may be complex and lengthy, for those skilled in the art related to the content disclosed in this invention, modifications to design, manufacturing, or production based on the technical content disclosed in this invention are merely conventional technical means and should not be construed as insufficient disclosure of the present invention.

[0026] Unless otherwise specified, all embodiments and optional embodiments of the present invention can be combined with each other to form new technical solutions.

[0027] Unless otherwise specified, all technical features and optional technical features of this invention can be combined to form new technical solutions.

[0028] Unless otherwise specified, all steps of the present invention may be performed sequentially or randomly, preferably sequentially. For example, the method includes steps (a) and (b), indicating that the method may include steps (a) and (b) performed sequentially, or it may include steps (b) and (a) performed sequentially. For example, the mention that the method may also include step (c) indicates that step (c) may be added to the method in any order. For example, the method may include steps (a), (b), and (c), or it may include steps (a), (c), and (b), or it may include steps (c), (a), and (b), etc.

[0029] Unless otherwise specified, the terms "comprising" and "including" as used in this invention can be open-ended or closed-ended. For example, "comprising" and "including" can mean that other components not listed may also be included, or that only the listed components may be included.

[0030] Unless otherwise specified, the term "or" is inclusive in this invention. For example, the phrase "A or B" means "A, B, or both A and B". More specifically, the condition "A or B" is satisfied by any of the following conditions: A is true (or exists) and B is false (or does not exist); A is false (or does not exist) and B is true (or exists); or both A and B are true (or exist).

[0031] To better understand the solutions of the embodiments of the present invention, some related terms and concepts that may be involved in the embodiments of the present invention will be introduced below: (1) Cold heading dies are forming tools installed on cold heading machines to shape metal wires or bars into fasteners such as bolts, nuts, and screws through upsetting, extrusion, etc. Cold heading dies typically include components such as punches, dies, and trimming dies, and their working surfaces will gradually wear down under repeated impact loads; (2) Equivalent wear strokes are the cumulative wear amount obtained by weighting the actual cumulative number of strokes according to the main slide speed. The basic idea is that the wear amount of a single stroke of the die is positively correlated with the impact speed—the higher the speed, the greater the impact energy, and the greater the wear amount of a single stroke. Therefore, when the cold heading machine operates at a speed higher than the rated speed, the contribution of that stroke to the wear is greater than one; when it is lower than the rated speed, the contribution is less than one. Equivalent wear strokes can more accurately reflect the cumulative wear amount of the die than the actual cumulative number of strokes.

[0032] In this embodiment, as Figure 1 As shown, a two-stage method for predicting the remaining life of a cold heading die includes: Step 1: At preset intervals, periodically collect vibration signals, cumulative stroke counts of the cold heading machine, and main slide speed during the operation of the cold heading die; Step 2: Preprocess and extract features from the collected raw data, select sensitive features to input into the principal component analysis model, and construct health indicators; Step 3: Monitor the time series data of health indicators in real time to identify the inflection point of the mold from the stable wear period to the accelerated wear period; Step 4: For test pieces whose health indicators are before the inflection point, use the first prediction model to predict the remaining lifespan; otherwise, use the second prediction model to predict the remaining lifespan.

[0033] In this embodiment, the preset interval between strokes is 1000 strokes. Data acquisition is triggered once every 1000 strokes. The duration of each acquisition is 0.2 seconds. The sampling frequency of the vibration signal is 25.6kHz. The current cumulative stroke count and the average main slider speed within the 1000 stroke cycle are read from the cold heading machine PLC through the OPC interface.

[0034] The single acquisition time is 0.2 seconds, the sampling frequency of the vibration signal is 25.6kHz, and 5120 vibration data points are acquired each time. Bandpass filtering is performed, with a passband range of 500Hz-10kHz. Then, the root mean square value, kurtosis index, margin index, energy proportion of the 2kHz-4kHz frequency band and the energy proportion of the 4kHz-8kHz frequency band are calculated.

[0035] Furthermore, the equivalent wear strokes are calculated using the following formula: ; in: Equivalent wear strokes; This is the total number of impacts up to the current moment; The preset speed influence coefficient; The real-time main slider speed during the i-th stroke; The preset standard speed is usually the rated operating speed of the cold heading machine.

[0036] In this embodiment, This represents the actual cumulative number of pulses corresponding to the current sampling point, and is an integer multiple of 1000; The rated speed of the cold heading machine is 120 strokes / minute; the speed influence coefficient is 0.35 after previous test calibration, and the cumulative value of equivalent wear strokes is calculated only once every 1000 strokes.

[0037] The five vibration characteristics of all sampling points were standardized and input into the principal component analysis model. The load vector of the first principal component was calculated as [0.42, 0.53, 0.30, 0.48, 0.44]. Since the PC1 score increases with the degree of wear, and the health index should be defined as 1 when it is brand new and 0 when it is failed, the negative of the PC1 score was taken as the original value of the health index. Then, the original value of the health index was mapped to the [0,1] interval through min-max normalization. The final health index is close to 1 for brand new molds and close to 0 for failed molds.

[0038] Furthermore, step three includes: Step 3.1: Sort the sampling points of health indicators according to the equivalent wear impact to form a sparse time series. After smoothing the series, calculate the linear fitting slope within the sliding window. The health index values ​​are sorted according to the equivalent wear impact. The sliding window contains 5 consecutive health index sampling points. The slope of the linear fit between the health index values ​​and the equivalent wear impact within the window is calculated. In this embodiment, the average slope of the first 10 windows is approximately −0.00018, and the standard deviation is 0.00005. Step 3.2: When the absolute value of the linear fitting slope continuously exceeds a preset first threshold and the duration is equivalent to two sliding windows, a suspected inflection point marker is triggered; The first set threshold is determined based on the statistical distribution of the slope of the health index during the stable wear period of similar molds, and is taken as 3 times the standard deviation of the absolute value of the average slope during the stable wear period; In this embodiment, the first set threshold is 3 times the standard deviation of the absolute value of the average slope, which is 0.00015. When the absolute value of the slope of two consecutive windows is greater than 0.00015, a suspected inflection point marker is triggered. In this embodiment, the suspected inflection point marker is first triggered at the 15th sampling point, which is approximately 152,000 equivalent wear impacts. Step 3.3: After triggering the suspected inflection point marker, calculate the cumulative deviation between the actual value of the health indicator and the linear extrapolation value based on historical data of the stable wear period; Take the health index data of the first 5 sampling points before the suspected inflection point, i.e., the first 5000 equivalent wear impacts, and fit a linear trend line using the least squares method: ; in: The health index data is given when the equivalent wear impact number is t. Equivalent wear strokes; The fitting result is =−0.00021, =1.02; Step 3.4: When the cumulative deviation exceeds the preset second threshold, confirm the inflection point and record the equivalent wear stroke position corresponding to the inflection point.

[0039] The second set threshold is set to 0.08. When the cumulative deviation exceeds the preset second set threshold, the inflection point is confirmed.

[0040] Then, starting from the suspected trigger point, the deviation between the actual health indicator data and the linear extrapolation value is calculated for each sampling point, and the cumulative deviation is calculated. When the cumulative deviation continuously exceeds the second set threshold of 0.08 for two sampling points, i.e., 2000 equivalent wear strokes, the inflection point is confirmed. In this embodiment, the inflection point is confirmed at the 16th sampling point, at approximately 157,000 equivalent wear strokes, and the equivalent wear stroke position corresponding to the inflection point is recorded as 157,000 equivalent wear strokes.

[0041] Furthermore, the formula for calculating the cumulative deviation is: ; in: This is the cumulative deviation; This represents the number of sampling points from the time the suspected inflection point marker was triggered until the current time. The actual value of the health indicator at the j-th sampling point; It is the linear extrapolation value of the j-th sampling point.

[0042] For test pieces whose health indicators are before the inflection point, i.e., the current equivalent wear strokes are less than the inflection point equivalent wear strokes, the first prediction model is used to predict the remaining life; otherwise, i.e., the current equivalent wear strokes are greater than or equal to the inflection point equivalent wear strokes, the second prediction model is used to predict the remaining life.

[0043] Furthermore, the first prediction model is the Holt two-parameter exponential smoothing model.

[0044] The first prediction model adopts the Holt two-parameter exponential smoothing model, which includes two state variables: a horizontal component and a trend component. The horizontal smoothing parameter ranges from 0.1 to 0.3, and the trend smoothing parameter ranges from 0.03 to 0.1. The training method is to use the least squares method to fit the initial horizontal and initial trend values ​​in the initial stage of mold use. After that, for each new sampling point, the model parameters are updated by recursion formula. The values ​​of the two parameters are determined by minimizing the prediction error of the first 10 sampling points.

[0045] In this embodiment, the health index data values ​​of the first 3 sampling points are taken, and a straight line is fitted using the least squares method. The intercept of the line is used as the initial level value, and the slope is used as the initial trend value. For each additional sampling point, i.e., every 1000 equivalent wear impacts, the model parameters are updated according to the recursive formula, as follows: ; ; in: The horizontal component of the i-th sampling point; For horizontal smoothing parameters; This is the linear extrapolation value of the i-th sampling point; This represents the horizontal component of the (i-1)th sampling point; This represents the trend component of the (i-1)th sampling point; The trend component of the i-th sampling point; For trend smoothing parameters; In this embodiment, the horizontal smoothing parameter is 0.15 and the trend smoothing parameter is 0.05.

[0046] After the 16th sampling point is processed, the system automatically switches from the first prediction model to the second prediction model. The switching operation is completed before the next sampling point, i.e., the 17th sampling.

[0047] Furthermore, the second prediction model is a long short-term memory network.

[0048] In this embodiment, the structure of the second prediction model is as follows: Input layer: receives a historical sequence of health indicators with a length of 5; LSTM layer: Single layer, containing 32 LSTM units. Each LSTM unit contains a forget gate, an input gate, an output gate, and a cell state.

[0049] Dropout layer: placed after the LSTM layer, with a dropout rate of 0.2%; Fully connected output layer: maps the hidden state of the last time step of the LSTM to a single output value, namely the predicted health metric value for the next sampling point; Activation function: LSTM internally uses the tanh activation function, and the output layer uses the linear activation function.

[0050] In this embodiment, the mold has run to the equivalent wear stroke count of 162,000, and the inflection point has been confirmed at 157,000 strokes. The second prediction model is used to predict the remaining life. The LSTM model state corresponding to the current sampling point is as follows: Input window: Take the actual values ​​of the health index of the last 5 sampling points, which are 0.62, 0.58, 0.53, 0.47, and 0.41 respectively. Input the input sequence [0.62, 0.58, 0.53, 0.47, 0.41] into the LSTM to obtain the predicted health index value of 0.36 for the next sampling point. Slide the sequence to [0.58, 0.53, 0.47, 0.41, 0.36] and input it again to obtain the second step prediction value of 0.30. Continuing the recursion: the predicted value in the third step is 0.24, the predicted value in the fourth step is 0.18, and the predicted value in the fifth step is 0.13. At this point, it is the first time that the value is lower than the failure threshold of 0.15. It takes a total of 5 steps to reach failure from the current sampling point, with each step corresponding to 1000 equivalent wear cycles. Therefore, the predicted remaining lifetime is 5000 equivalent wear cycles. At the same time, the MC Dropout method is used with a drop rate of 0.2 and forward propagation is repeated 100 times. The calculated 95% confidence interval is [4300 times, 5800 times]. After the actual mold continues to run for 5200 equivalent wear cycles, the health index drops to 0.14, with a deviation of only 200 times from the model's predicted value of 5000 times. The prediction accuracy is significantly better than the single exponential smoothing model.

[0051] It should be noted that the present invention is not limited to the above-described embodiments. The above embodiments are merely examples, and any embodiments that have the same structure and perform the same effects as the technical concept within the scope of the present invention are included within the scope of the present invention. Furthermore, various modifications that can be conceived by those skilled in the art to the embodiments, and other ways of constructing by combining some of the constituent elements of the embodiments, without departing from the spirit of the present invention, are also included within the scope of the present invention.

Claims

1. A two-stage method for predicting the remaining life of a cold heading die, characterized in that, include: Step 1: At preset intervals, periodically collect vibration signals, cumulative stroke counts of the cold heading machine, and main slide speed during the operation of the cold heading die; Step 2: Preprocess and extract features from the collected raw data, select sensitive features to input into the principal component analysis model, and construct health indicators; Step 3: Monitor the time series data of health indicators in real time to identify the inflection point of the mold from the stable wear period to the accelerated wear period; Step 4: For test pieces whose health indicators are before the inflection point, use the first prediction model to predict the remaining lifespan; otherwise, use the second prediction model to predict the remaining lifespan.

2. The method for predicting the remaining life of a cold heading die in two stages according to claim 1, characterized in that, Step two includes: Step 2.1: Filter and denoise the acquired vibration signal, and extract time-domain and frequency-domain features; Step 2.2: Use the main slider speed to perform a weighted correction on the cumulative stroke count and calculate the equivalent wear strokes; Step 2.3: Combine the extracted features with the equivalent wear impacts to form a multidimensional feature vector, input it into the principal component analysis model, and take the score of the first principal component as the original value of the health indicator; Step 2.4: Normalize the raw values ​​of the health indicators to obtain the final health indicators.

3. The method for predicting the remaining life of a cold heading die in two stages according to claim 2, characterized in that, The time-domain and frequency-domain features include root mean square value, kurtosis index, margin index, energy proportion of the 2kHz-4kHz frequency band, and energy proportion of the 4kHz-8kHz frequency band.

4. The method for predicting the remaining life of a cold heading die in two stages according to claim 2, characterized in that, The equivalent wear strokes are calculated using the following formula: ； in: Equivalent wear strokes; This is the total number of impacts up to the current moment; The preset speed influence coefficient has a value range of 0.2-0.

5. The real-time main slider speed during the i-th stroke; The preset standard speed is usually the rated operating speed of the cold heading machine.

5. The method for predicting the remaining life of a cold heading die in two stages according to claim 1, characterized in that, Step three includes: Step 3.1: Sort the sampling points of health indicators according to the equivalent wear impact to form a sparse time series. After smoothing the series, calculate the linear fitting slope within the sliding window. Step 3.2: When the absolute value of the linear fitting slope continuously exceeds a preset first threshold and the duration is equivalent to 2 to 5 sliding windows, a suspected inflection point marker is triggered; Step 3.3: After triggering the suspected inflection point marker, calculate the cumulative deviation between the actual value of the health indicator and the linear extrapolation value based on historical data of the stable wear period; Step 3.4: When the cumulative deviation exceeds the preset second threshold, confirm the inflection point and record the equivalent wear stroke position corresponding to the inflection point.

6. The method for predicting the remaining life of a cold heading die in two stages according to claim 5, characterized in that, The preset first threshold is 2-4 times the standard deviation of the absolute value of the average slope of the health index data range within the first 2000 equivalent wear strokes after the same type of mold is put into use.

7. The method for predicting the remaining life of a cold heading die in two stages according to claim 5, characterized in that, The linear extrapolation value based on historical data of stable wear period is obtained by taking the health index data of 3-6 sampling points before the suspected inflection point is triggered, fitting the linear trend line with the least squares method, and then substituting the equivalent wear impact of the current sampling point into the calculation.

8. The method for predicting the remaining life of a cold heading die in two stages according to claim 5, characterized in that, The formula for calculating the cumulative deviation is: ； in: This is the cumulative deviation; This represents the number of sampling points from the time the suspected inflection point marker was triggered until the current time. The actual value of the health indicator at the j-th sampling point; It is the linear extrapolation value of the j-th sampling point.

9. The method for predicting the remaining life of a cold heading die in two stages according to claim 1, characterized in that, The first prediction model is the Holt two-parameter exponential smoothing model.

10. The method for predicting the remaining life of a cold heading die in two stages according to claim 1, characterized in that, The second prediction model is a long short-term memory network.

Images