A mold service life evaluation method based on machine learning

Abstract

The application relates to the technical field of machine learning, and discloses a mold service life evaluation method based on machine learning, which comprises the following steps: collecting mold data and extracting features related to mold physical failure modes; performing adaptive decomposition, adjusting decomposition parameters, and obtaining signal components; constructing an integrated model to obtain a residual service life prediction value of the mold; analyzing feature contribution degrees, optimizing a model input feature set, and outputting maintenance decision suggestions based on analysis results; deploying the integrated model to an edge computing device, and adjusting an evaluation strategy according to degradation characteristics of the mold in different life cycles to realize real-time life evaluation and decision support. The application can significantly improve the accuracy and reliability of mold life prediction, enhance model transparency, and realize closed-loop management from prediction to maintenance decision.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and in particular to a method for evaluating the lifespan of molds based on machine learning. Background Technology

[0002] As a key process equipment in industrial production, the service life of molds directly affects production efficiency and product quality. In actual production, molds often gradually degrade due to various physical failure modes, including wear, fatigue, and crack propagation, eventually leading to failure. Traditional mold life assessment methods mostly rely on periodic maintenance, experience-based judgment, or predictions based on physical models. These methods often suffer from high subjectivity, low prediction accuracy, and inability to respond in real time.

[0003] With the development of smart manufacturing and industrial IoT technologies, it has become possible to collect mold operating status data in real time through sensors. However, due to the complexity of mold operating conditions, the heterogeneity of multi-source data, and the variety of failure mechanisms, how to extract effective features from massive monitoring data and establish accurate life prediction models remains a major challenge.

[0004] Existing methods often fail to fully integrate manufacturing process parameters, environmental data, and real-time monitoring signals, lacking adaptive decomposition and feature selection mechanisms for multimodal data, resulting in insufficient model generalization ability and poor interpretability. Furthermore, most existing lifespan prediction models remain at the cloud-based centralized processing stage, making it difficult to meet the real-time and reliability requirements of production sites. While edge computing offers advantages such as low latency and high responsiveness, how to lightweight and adapt complex machine learning models to edge devices, while simultaneously achieving dynamic adjustment of evaluation strategies at different stages of the lifecycle, remains a pressing technical challenge.

[0005] Therefore, this invention proposes a machine learning-based method for evaluating mold lifespan. Summary of the Invention

[0006] The purpose of this invention is to propose a machine learning-based method for evaluating mold lifespan in order to address the problems of low accuracy, poor real-time performance, and weak model interpretability in existing technologies.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: a machine learning-based method for evaluating mold lifespan, comprising the following steps:

[0008] Step S1: Collect mold manufacturing process parameters, working condition monitoring signals and workshop environmental data, and extract features related to the physical failure mode of the mold from the online working condition monitoring signals;

[0009] Step S2: Adaptively decompose the collected working condition monitoring signals, and adjust the decomposition parameters to obtain signal components with the goal of consistency between the decomposition results and the theoretical physical model.

[0010] Step S3: Construct an integrated model that combines deep learning and traditional machine learning, fuse features and signal components to obtain the predicted remaining lifespan of the mold.

[0011] Step S4: Analyze the feature contribution based on interpretable artificial intelligence technology, optimize the model input feature set according to the contribution, and output maintenance decision suggestions based on the analysis results;

[0012] Step S5 involves deploying the integrated model to an edge computing device and adjusting the evaluation strategy based on the degradation characteristics of the mold at different lifecycle stages to achieve real-time lifetime assessment and decision support.

[0013] Furthermore, step S1 also includes the following sub-steps:

[0014] S1-1, Obtain manufacturing process parameters from the manufacturing execution system, the manufacturing process parameters including mold material parameters, heat treatment process parameters and machining accuracy parameters;

[0015] S1-2, collects operating condition monitoring signals through a multi-source sensor network, the operating condition monitoring signals including vibration signals, acoustic emission signals and temperature signals;

[0016] S1-3, Collect workshop environmental data through workshop environmental sensors, including ambient temperature, humidity and dust concentration;

[0017] S1-4, calculate the kurtosis index and root mean square ratio from the vibration acceleration signal, as an early warning feature of the transition of wear from the steady period to the accelerated period;

[0018] S1-5, the energy of sudden events and the cumulative amount of ringing counts are statistically analyzed from the acoustic emission signal stream, which serves as the basis for determining the initiation and propagation of microcracks;

[0019] S1-6 calculates the maximum temperature rise rate and thermal cycling asymmetry from thermocouple temperature data as characteristic parameters for assessing the accumulation of thermal fatigue damage.

[0020] Step S1 further includes frequency domain feature extraction: performing Fourier transform on the vibration signal to extract the energy proportion of specific frequency bands corresponding to macroscopic wear and microscopic cracks; performing wavelet packet decomposition on the acoustic emission signal to extract the kurtosis factor of specific frequency bands corresponding to crack propagation, as a sensitive feature of crack propagation.

[0021] Furthermore, step S2 also includes the following sub-steps:

[0022] S2-1, Initialize the parameters of the fully integrated empirical mode decomposition adaptive noise algorithm, and set the initial range of noise standard deviation, the initial range of the overall mean number of iterations, and the maximum number of iterations;

[0023] S2-2 integrates the vibration signal input with the empirical mode decomposition adaptive noise algorithm to obtain multiple sets of intrinsic mode functions and a residual sequence.

[0024] S2-3, calculate the correlation coefficient between the decomposed residual sequence and the theoretical wear amount calculated based on the Achad wear model, and calculate the sample entropy of the intrinsic mode function components. The specific formula of the Achad wear model is as follows: Where dW is the wear depth and K is the wear factor. Here, is the normal pressure, u is the sliding velocity, and H is the mold hardness;

[0025] S2-4, Adjust the parameters of the fully integrated empirical mode decomposition adaptive noise algorithm by grid search until the correlation coefficient reaches the preset threshold and the sample entropy is minimized, and determine the decomposition parameters;

[0026] S2-5, Based on the correlation between the components and physical characteristics, eliminate intrinsic mode function components that are unrelated to mold degradation, and retain the effective intrinsic mode function components and residual sequences. The "unrelated to mold degradation" means that the correlation of the components is lower than a preset threshold.

[0027] The parameters of the Achad wear model in steps S2-3 are determined as follows: for different mold materials, multiple sets of standard wear test data are collected, and the material-specific wear factor is obtained by fitting using the least squares method; the normal pressure is obtained by finite element simulation of the mold to obtain the pressure distribution of key parts of the cavity, and the sliding speed is calculated based on the stroke and time of the stamping equipment.

[0028] Furthermore, step S3 also includes the following sub-steps:

[0029] S3-1, Construct a multi-layer architecture consisting of an encoder for processing long-term trend features, a recurrent neural network for processing short-term dynamic features, an attention fusion layer, and a meta-decision layer. The attention fusion layer includes a spatial attention sub-layer and a temporal attention sub-layer. The spatial attention sub-layer highlights the features of the cavity stress concentration area, and the temporal attention sub-layer strengthens the features of the early stage of failure.

[0030] S3-2, combine the extracted features and screened signal components with manufacturing process data and environmental data to form a feature vector, use the remaining number of working cycles as the sample label, and divide the training set and validation set according to a preset ratio;

[0031] S3-3, adopts a composite loss function that integrates prediction error and physical consistency penalty term, and uses an adaptive optimization algorithm for training. The training strategy includes an adaptive learning rate adjustment mechanism and an early stopping mechanism.

[0032] S3-4, the training set is input into the ensemble model for iterative training, the model hyperparameters are dynamically adjusted, and the performance is evaluated using the validation set until the model converges and the final model is saved. The model hyperparameters include the number of hidden layer nodes in the converter and the random deactivation probability of the gated recurrent unit.

[0033] The attention fusion layer in step S3-1 calculates dynamic weights according to the following formula: ,in, Let be the attention weight at time t. The preceding network outputs a feature vector, where v and R are learnable parameter matrices and b is a bias vector. The spatial attention sublayer generates a spatial weight map through convolution operations, assigning higher weights to features in stress concentration regions. The temporal attention sublayer assigns higher weights to features in the early stages of failure.

[0034] Furthermore, step S4 also includes the following sub-steps:

[0035] S4-1, Input the validation set into the trained model, use interpretable artificial intelligence technology to calculate the contribution of each input feature to the remaining useful life prediction value, generate a time-varying curve of feature contribution, and determine the contribution ratio of each key feature in different life cycles, wherein the different life cycles include the new model stage, the mid-term stage and the pre-failure stage.

[0036] S4-2, After completing the prediction of the remaining lifespan for the preset number of times, rank the features by contribution, retain the features of the first preset number, remove redundant features, and update the model input feature set.

[0037] S4-3, Establish a preset knowledge base of high contribution features and maintenance actions, and match and output corresponding maintenance decision suggestions from the preset knowledge base according to the type and value of high contribution features.

[0038] The interpretability AI technology in step S4-1 is the Shapley additive interpretation algorithm improved with physical prior constraints. The improvement includes: assigning high computational weights to features directly related to the mold failure mechanism and low computational weights to features with indirect influence; the contribution curve is dynamically updated using a sliding time window mechanism.

[0039] Furthermore, step S5 also includes the following sub-steps:

[0040] S5-1, the trained ensemble model, along with the dynamic feature selection function and maintenance decision mapping function implemented based on interpretability analysis, are jointly deployed to an edge computing device, which supports industrial Ethernet communication;

[0041] S5-2, in the new mold stage, calls manufacturing process data to calibrate the initial life baseline and reduce the new mold evaluation error. In the mid-term stage, it strengthens the trend analysis of the working condition monitoring signal to capture wear accumulation characteristics. In the early stage of failure, it starts the high-frequency sampling mode to enhance the ability to capture micro-crack characteristics.

[0042] S5-3, the edge computing device receives monitoring data in real time, automatically performs feature extraction, signal decomposition, model inference and decision output processes, and outputs the remaining service life prediction value and maintenance suggestions;

[0043] S5-4, periodically accumulated operational data, is used to adjust the model deployed on edge computing devices. During the adjustment, the underlying network parameters are fixed and the meta-decision layer parameters are optimized.

[0044] The beneficial effects of the technical solution provided by this invention include at least the following:

[0045] This invention significantly improves the accuracy and reliability of mold remaining service life prediction by integrating multi-source sensor data and manufacturing process parameters, combined with adaptive signal decomposition and physical model consistency optimization.

[0046] This invention constructs an integrated model that combines deep learning and traditional machine learning, and introduces an attention mechanism to enhance the extraction of key features. This effectively captures the degradation trend of the mold in different life cycles and enhances the model's temporal modeling and spatial perception capabilities.

[0047] This invention introduces interpretable artificial intelligence technology to dynamically analyze feature contribution and output maintenance decision suggestions, thereby improving model transparency and practicality and supporting on-site maintenance personnel to make accurate decisions.

[0048] This invention deploys the model to edge computing devices and implements an adaptive evaluation strategy for each lifecycle stage, enabling real-time, online lifecycle assessment and decision support, significantly improving the intelligence level and operational efficiency of the production system. Attached Figure Description

[0049] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention 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 the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 This is a flowchart provided for an embodiment of the present invention;

[0051] Figure 2 This is a schematic diagram of a method provided in an embodiment of the present invention. Detailed Implementation

[0052] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of a machine learning-based mold lifespan evaluation method proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0053] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0054] The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0055] The following description, in conjunction with the accompanying drawings, details a specific scheme for a machine learning-based mold lifespan assessment method provided by this invention.

[0056] Please see Figure 1 and Figure 2 The illustration shows a flowchart and a method diagram of a machine learning-based mold life assessment method according to an embodiment of the present invention, including the following steps:

[0057] Step S1: Collect mold manufacturing process parameters, working condition monitoring signals and workshop environmental data, and extract features related to the physical failure mode of the mold from the online working condition monitoring signals;

[0058] Step S1 further includes the following sub-steps:

[0059] S1-1, Obtain manufacturing process parameters from the manufacturing execution system. The manufacturing process parameters include mold material parameters, heat treatment process parameters, and machining accuracy parameters.

[0060] S1-2 collects operating condition monitoring signals through a multi-source sensor network. The operating condition monitoring signals include vibration signals, acoustic emission signals, and temperature signals.

[0061] S1-3 collects workshop environmental data through workshop environmental sensors. The workshop environmental data includes ambient temperature, humidity and dust concentration.

[0062] S1-4, calculate the kurtosis index and root mean square ratio from the vibration acceleration signal, as an early warning feature of the transition of wear from the steady period to the accelerated period;

[0063] S1-5, the energy of sudden events and the cumulative amount of ringing counts are statistically analyzed from the acoustic emission signal stream, which serves as the basis for determining the initiation and propagation of microcracks;

[0064] S1-6 calculates the maximum temperature rise rate and thermal cycling asymmetry from thermocouple temperature data as characteristic parameters for assessing the accumulation of thermal fatigue damage.

[0065] Furthermore, step S1 also includes frequency domain feature extraction: performing Fourier transform on the vibration signal to extract the specific frequency band energy proportion corresponding to macroscopic wear and microscopic cracks; performing wavelet packet decomposition on the acoustic emission signal to extract the specific frequency band kurtosis factor corresponding to crack propagation as a sensitive feature of crack propagation.

[0066] It should be noted that the data collection process involves installing a multi-source sensor network on the mold and the stamping machine. All sensor data is collected and initially summarized through an industrial IoT gateway deployed on-site, and then transmitted to the edge computing server in the workshop via industrial Ethernet. Manufacturing process parameters are automatically obtained from the enterprise's manufacturing execution system database through an API interface.

[0067] Manufacturing Execution System (MES): This is an information system used by enterprises to manage the production workshop level, recording the detailed processing history of each work order and each mold.

[0068] Mold material parameters: refer to the material grade and inherent properties used in the main parts of the mold body; for example, the hardness, yield strength and impact toughness of mold steel. These parameters determine the basic ability of the mold to resist deformation and failure.

[0069] Heat treatment process parameters: These refer to the key indicators of the heat treatment process used to obtain the above-mentioned material properties; for example: quenching temperature 1020°C, tempering temperature 520°C, and holding time. The heat treatment process directly affects the grain structure and final properties of the material and is crucial to avoiding early failure.

[0070] Machining accuracy parameters refer to the final machining quality of key parts of the mold cavity; for example, the dimensional tolerance of surface roughness Ra0.4μm and key fillet radius R8mm; rough surfaces or excessively small fillet radii are stress concentration points, which will significantly accelerate the initiation of fatigue cracks.

[0071] Multi-source sensor networks: refer to monitoring systems composed of various types of sensors, including:

[0072] Vibration acceleration sensor: An ICP type acceleration sensor is used and installed on the mold or press base to collect vibration signals generated during the stamping process. The sampling frequency is set to 10kHz to capture impact events.

[0073] Acoustic emission sensor: An acoustic emission sensor with a resonant frequency of 150kHz is used and is directly installed in the high-stress area near the mold cavity. The acoustic emission signal is extremely sensitive to microscopic deformation and crack propagation within the material.

[0074] Thermocouple: A K-type armored thermocouple is used, embedded 2mm below the surface of the mold cavity, to monitor the temperature change of the mold surface during each stamping cycle.

[0075] Workshop environment sensors: Deploy temperature and humidity sensors and laser dust sensors near the mold working area to monitor ambient temperature, relative humidity and PM10 dust concentration; humid and dusty environments can exacerbate mold corrosion and wear.

[0076] Kurtosis index: Kurtosis is very sensitive to the impact component in the signal. During the stable period of wear, the vibration signal distribution is close to normal and the kurtosis value is close to 3. When wear intensifies and enters the acceleration period, more transient impacts will be generated, resulting in a significant increase in the kurtosis value, thus providing an early warning.

[0077] Root mean square ratio: Calculates the ratio of the root mean square value of the signal within the current time window to the RMS value of a baseline stable period. If this ratio is consistently greater than 1 and shows an upward trend, it indicates that wear is intensifying and the overall energy level of the system is increasing.

[0078] Energy of sudden events: Acoustic emission signals are divided into continuous and sudden types. Sudden signals are generated by discrete events, such as the propagation of micro-cracks. Calculate the energy integral of these sudden events. The sudden increase in energy value is direct evidence of crack propagation.

[0079] Ringing count accumulation: The number of times a single burst event waveform exceeds a threshold is called the ringing count. The ringing counts of all events within a time period are accumulated to obtain the accumulation amount. The growth rate of this accumulation amount can reflect the severity and frequency of crack propagation.

[0080] Maximum temperature rise rate: Calculates the average rate at which the temperature rises from the lowest point to the highest point during each stamping cycle, in °C / s; an excessively high temperature rise rate will generate huge thermal stress, which is the main driving force for the initiation of thermal fatigue cracks.

[0081] Thermal cycling asymmetry: assessing the asymmetry of each temperature cycle; for example, calculating the ratio of heating time to cooling time. Asymmetric cycling can lead to different stress residence times at different temperatures, exacerbating creep or oxidation damage of certain mechanisms.

[0082] Fourier transform is applied to vibration signals: it converts time-domain vibration signals into frequency-domain signals. Macroscopic wear causes a general increase in vibration energy across the entire frequency band; while microscopic cracks produce specific high-frequency resonance components. Therefore, extracting the percentage of energy in the 5-8kHz high-frequency band relative to the total energy can serve as a characteristic of microscopic cracks.

[0083] Wavelet packet decomposition applied to acoustic emission signals: Wavelet packet decomposition provides a finer frequency band division than standard wavelet decomposition. Performing four-level wavelet packet decomposition on an acoustic emission signal yields 16 frequency bands. Crack propagation typically generates rich acoustic emission responses in a specific frequency band. Calculating the wavelet packet kurtosis factor of the signal in that specific frequency band indicates that the higher the kurtosis factor, the stronger the impact of that frequency band and the higher its correlation with crack propagation.

[0084] Step S2: Adaptively decompose the collected working condition monitoring signals, and adjust the decomposition parameters to obtain signal components with the goal of consistency between the decomposition results and the theoretical physical model.

[0085] Step S2 further includes the following sub-steps:

[0086] S2-1, Initialize the parameters of the fully integrated empirical mode decomposition adaptive noise algorithm, and set the initial range of noise standard deviation, the initial range of the overall mean number of iterations, and the maximum number of iterations;

[0087] S2-2 integrates the vibration signal input with the empirical mode decomposition adaptive noise algorithm to obtain multiple sets of intrinsic mode functions and a residual sequence.

[0088] S2-3, calculate the correlation coefficient between the decomposed residual sequence and the theoretical wear amount calculated based on the Achad wear model, and calculate the sample entropy of the intrinsic mode function components. The specific formula of the Achad wear model is as follows: Where dW is the wear depth and K is the wear factor. Here, is the normal pressure, u is the sliding velocity, and H is the mold hardness;

[0089] S2-4, Adjust the parameters of the fully integrated empirical mode decomposition adaptive noise algorithm by grid search until the correlation coefficient reaches the preset threshold and the sample entropy is minimized, and determine the decomposition parameters;

[0090] S2-5, based on the correlation between the components and physical characteristics, eliminate intrinsic mode function components that are unrelated to mold degradation, and retain the effective intrinsic mode function components and residual sequences. Unrelated to mold degradation means that the correlation of the components is lower than the preset threshold.

[0091] Furthermore, the parameters of the Achad wear model in steps S2-3 are determined in the following ways: for different mold materials, multiple sets of standard wear test data are collected, and the material-specific wear factor is obtained by fitting using the least squares method; the normal pressure is obtained by obtaining the pressure distribution of key parts of the cavity through mold finite element simulation, and the sliding speed is calculated based on the stroke and time of the stamping equipment.

[0092] It should be noted that the fully integrated empirical mode decomposition adaptive noise algorithm is an advanced signal decomposition method designed to solve the mode mixing problem. Its basic idea is to add paired white noise to the original signal multiple times, then perform empirical mode decomposition, and finally average the results of multiple decompositions, thereby effectively suppressing noise interference and obtaining purer and more stable intrinsic mode function components.

[0093] Noise standard deviation: refers to the intensity of the added white noise, usually set as a proportion of the original signal standard deviation, with an initial range of 0.1 to 0.3; this parameter controls the amount of noise added and affects the fineness of the decomposition.

[0094] Overall average number of repetitions: refers to the number of times the above "add noise and decompose" process is repeated; the more repetitions, the better the noise suppression effect, but the greater the computational load. The initial range is set to 50 to 150 times.

[0095] Maximum number of iterations: This refers to the maximum number of iterations allowed in the inner loop when decomposing each intrinsic mode function. It is used to ensure the convergence of the algorithm and is set to 1000 times.

[0096] Intrinsic mode functions: These are a series of components obtained after decomposition. Each component represents an intrinsic vibration mode in a specific frequency range of the original signal. These functions contain information from different scales of the signal, from high frequency to low frequency. High frequency components are usually related to noise or micro-variations, while low frequency components are related to trend changes.

[0097] Residual sequence: This is the trend term remaining after decomposition, representing the slowest changing component of the signal; in the context of mold wear, this residual sequence is crucial because it is often highly correlated with the cumulative amount of wear depth.

[0098] The Archad wear model is a classic wear theory model that describes wear as a process that is directly proportional to pressure and slip distance and inversely proportional to material hardness.

[0099] Wear depth: refers to the height of material loss on the surface of the mold due to wear. It is a direct indicator of the degree of wear and is measured in micrometers.

[0100] Wear factor: This is a comprehensive material coefficient that represents the material's ability to resist wear. It needs to be determined experimentally, and its value varies for different mold steels.

[0101] Normal pressure: refers to the pressure applied vertically to the die surface by the sheet metal during the stamping process, measured in megapascals (MPA), and describes the tightness of the contact.

[0102] Sliding speed: refers to the relative sliding speed between the sheet metal and the die surface during the stamping process. It is calculated from the stroke and cycle time of the stamping equipment and is measured in millimeters per second. It describes the friction process.

[0103] Mold hardness: refers to the hardness of the surface material of the mold, which is an indicator of the material's ability to resist plastic deformation, and is commonly expressed by Rockwell hardness.

[0104] Theoretical wear calculation: Based on each stamping stroke of the stamping equipment, a small theoretical wear depth increment is calculated using the above-mentioned Archad formula. By summing up the wear increments of all historical strokes, a theoretical wear curve that varies with time can be obtained.

[0105] Correlation coefficient: The Pearson correlation coefficient between the residual sequence obtained from the decomposition and the theoretical wear curve is calculated. The closer the coefficient is to 1, the higher the consistency between the decomposed trend term and the wear trend predicted by the physical model, proving that the decomposition is effective.

[0106] Sample entropy is a metric for measuring the complexity of a time series. The lower the entropy value, the more regular and smooth the series is; the higher the entropy value, the more chaotic the series is and the more noise it contains. It is desirable that the intrinsic mode functions decomposed are regular, so their sample entropy needs to be as small as possible.

[0107] Grid search method: a parameter optimization method. In this embodiment, within a pre-set initial range of noise standard deviation and overall average frequency, all possible parameter combinations are generated at a certain step size, such as a noise step size of 0.05 and an average frequency step size of 20.

[0108] Optimization objective: For each set of parameters, execute S2-2 and S2-3 to calculate the corresponding correlation coefficient and sample entropy. The optimal parameter combination is the one that minimizes the correlation coefficient exceeding a preset threshold (e.g., 0.85) and the average sample entropy of all intrinsic mode functions.

[0109] Selection criteria: Calculate the correlation coefficient between each intrinsic mode function component and the theoretical wear curve or the key features extracted in step S1.

[0110] Elimination rule: Set a correlation threshold, such as 0.5, and discard intrinsic mode function components with low correlation to wear that are below the threshold as irrelevant noise. Only retain the highly correlated intrinsic mode functions and the residual sequence representing the wear trend as input to the subsequent model.

[0111] Explanation of the method for determining the parameters of the Archad wear model:

[0112] Material-specific wear factor: It cannot be obtained directly from theory and must be calibrated experimentally. For the specific mold steel used in this invention, such as SKD11, multiple sliding wear tests need to be conducted on a standard wear testing machine under different pressures and speeds to measure the actual wear amount. Then, the least squares method is used to fit the experimental data with the Achad formula to inversely calculate the wear factor K value that best represents the wear characteristics of the material.

[0113] Normal pressure: The shape of the mold cavity is very complex, and the pressure distribution in various places is extremely uneven. By using finite element simulation technology, an accurate model of the mold and sheet metal is built in the computer to simulate the entire stamping process. The pressure distribution cloud map of the key parts in the cavity at each moment can be accurately calculated, thereby obtaining the accurate normal pressure value.

[0114] Sliding speed: Based on the slider stroke and time curve recorded by the sensors on the stamping equipment, the real-time speed change curve can be obtained by calculating its derivative, thereby determining the sliding speed.

[0115] Step S3: Construct an integrated model that combines deep learning and traditional machine learning, fuse features and signal components to obtain the predicted remaining lifespan of the mold.

[0116] Step S3 further includes the following sub-steps:

[0117] S3-1, construct a multi-layer architecture for processing long-term trend features, a recurrent neural network for processing short-term dynamic features, an attention fusion layer and a meta-decision layer. The attention fusion layer includes a spatial attention sub-layer and a temporal attention sub-layer. The spatial attention sub-layer highlights the features of the cavity stress concentration area, and the temporal attention sub-layer strengthens the features of the early stage of failure.

[0118] S3-2, combine the extracted features and screened signal components with manufacturing process data and environmental data to form a feature vector, use the remaining number of working cycles as the sample label, and divide the training set and validation set according to a preset ratio;

[0119] S3-3 employs a composite loss function that integrates prediction error and physical consistency penalty term, and uses an adaptive optimization algorithm for training. The training strategy includes an adaptive learning rate adjustment mechanism and an early stopping mechanism.

[0120] S3-4 involves inputting the training set into the ensemble model for iterative training, dynamically adjusting the model hyperparameters, and evaluating the performance using the validation set until the model converges and the final model is saved. The model hyperparameters include the number of hidden layer nodes in the converter and the random deactivation probability of the gated recurrent unit.

[0121] Furthermore, the attention fusion layer in step S3-1 calculates the dynamic weights according to the following formula: ,in, Let be the attention weight at time t. The preceding network outputs a feature vector, where v and R are learnable parameter matrices and b is a bias vector. The spatial attention sublayer generates a spatial weight map through convolution operations, assigning higher weights to features in stress concentration regions. The temporal attention sublayer assigns higher weights to features in the early stages of failure.

[0122] It should be noted that the encoder used to process long-term trend features refers to a model structure that can capture global information and long-term dependencies. It receives the residual sequence and manufacturing process parameters from step S2. These data are compressed into a highly abstract comprehensive state vector through multi-layer nonlinear transformation of the encoder. This vector summarizes the macro trend of the mold's "basic constitution" and "health status".

[0123] Recurrent neural networks (RNNs) for processing short-term dynamic features are network structures specifically designed for processing sequential data. They have a "memory" function and receive temporal features from steps S1 and S2, such as vibration kurtosis, acoustic emission energy, temperature change rate, and filtered intrinsic mode function components calculated for each stamping cycle or batch. The RNNs analyze these data step by step, keenly capturing subtle dynamic changes and short-term fluctuation patterns generated within each stamping cycle. These patterns are often early signs of failure.

[0124] Attention Fusion Layer: This is the core of the model's intelligent information filtering. It receives all information from the encoder and recurrent neural network and automatically determines which information is more important, thus assigning it higher weights.

[0125] Spatial Attention Sublayer: Its function is to identify which physical locations on the mold have more important features. The stress conditions of different parts of the mold cavity vary greatly. For example, fillets and draw beads are well-known stress concentration areas, which are most prone to wear and cracks. This sublayer analyzes features from different sensors installed in different locations through a convolution-like operation, generating a "spatial importance weight map". It automatically assigns higher weights to features from these high-risk areas, making the model pay more attention to the health status of these key parts.

[0126] The time attention sublayer identifies which time period features are more important for predicting the current lifespan. The feature changes of the mold in the early stage of failure are crucial for accurately predicting the remaining lifespan. This sublayer can automatically learn and strengthen the importance of features in the early stage of failure time window, while weakening the relatively unimportant information in the early stable stage, making the model more sensitive and accurate in judging the end of the lifespan.

[0127] Meta-decision layer: This is the "brain" and final output layer of the entire model. It receives all high-level features after weighted fusion through the attention mechanism and integrates them. In this embodiment, this layer can be composed of a series of fully connected layers, and finally outputs a continuous remaining service life prediction value, which can be in units of remaining stamping times or hours.

[0128] Feature vector: Combine all the data from the above sources—manufacturing process parameters, environmental data, time domain features, frequency domain features, and signal decomposition components—into a unified digital vector at each time sample point.

[0129] Sample label: The number of remaining working cycles corresponding to each feature vector, i.e. the number of remaining stamping times; this label needs to be calculated in reverse based on the historical full life cycle data of the mold, assuming that its final failure point is known.

[0130] Dataset partitioning: Divide all sample points in chronological order. For example, use the first 70% of the data as the training set for model learning; the middle 15% as the validation set for adjusting hyperparameters and preventing overfitting during training; and the last 15% as the test set for final model performance evaluation.

[0131] Composite loss function: The optimization objective of the model consists of two parts:

[0132] Prediction error: The mean square error between the model's predicted remaining useful life and the actual remaining useful life, which is the main optimization objective.

[0133] Physical consistency penalty: Introducing prior physical knowledge, for example, can constrain the predicted remaining useful life to be monotonically decreasing over time; if the model predicts a longer useful life than the previous time point, a penalty is imposed; this ensures that the model's predictions are not only accurate, but also consistent with physical common sense.

[0134] Adaptive optimization algorithm: This algorithm is an optimization algorithm that can automatically adjust the learning step size, such as the Adam algorithm. It uses a larger step size to converge quickly in the early stage of training and a smaller step size to fine-tune in the later stage, which is more efficient and stable than the traditional fixed step size algorithm.

[0135] Adaptive learning rate adjustment mechanism: A predefined rule, such as automatically halving the learning rate when the validation set error no longer decreases for several consecutive cycles, to help the model escape local optima and continue to converge toward a better point.

[0136] Early stopping mechanism: Training is stopped immediately when the validation set error stops decreasing and starts to rise continuously. This is a sign that the model is overfitting. The model is then backtracked to the state at the moment when the validation set error was the lowest, and that state is used as the final model.

[0137] Iterative training: The training set data is input into the model multiple times. The model calculates the loss based on the prediction results and updates millions of internal parameters through the backpropagation algorithm, gradually reducing the loss and improving the prediction accuracy.

[0138] Dynamically adjusting hyperparameters: Hyperparameters are parameters set before training begins that control the model structure and training process.

[0139] Number of hidden layer nodes in the converter: refers to the number of neurons in each layer of the encoder, such as 512. The more nodes there are, the stronger the model is, but it is also more prone to overfitting. The optimal number needs to be found by testing the performance on the validation set.

[0140] Random deactivation probability of gated recurrent units: a technique to prevent overfitting of recurrent neural networks. During training, a portion of the output of neurons is randomly "dropped", forcing the network to not over-rely on any single neuron, thereby learning more robust features. This dropout probability, for example, 0.2, is a key hyperparameter.

[0141] Model convergence: When both the training loss and validation loss remain stable at a low level and no longer change significantly, the model is considered to have been trained and the model parameters are saved.

[0142] Explanation of the attention weight formula:

[0143] The formula describes the computation process of the time attention sublayer, which is essentially a "scoring" mechanism: It is a "scoring function" that evaluates the feature vector output by the recurrent neural network at time t. The importance of this; the scoring process is accomplished through a learnable parameter matrix R and a bias vector b, followed by a nonlinear transformation using a hyperbolic tangent activation function; another learnable parameter vector v compresses the transformed result into a single score.

[0144] Finally, the scores of all time steps are converted into weights through the exponential normalization process in the formula. Weight The value is between 0 and 1, and the sum of the weights of all time steps is 1. The higher the score of a time step, the greater the proportion of its corresponding feature in the final fusion.

[0145] Step S4: Analyze the feature contribution based on interpretable artificial intelligence technology, optimize the model input feature set according to the contribution, and output maintenance decision suggestions based on the analysis results;

[0146] Step S4 further includes the following sub-steps:

[0147] S4-1: Input the validation set into the trained model, use interpretable artificial intelligence technology to calculate the contribution of each input feature to the remaining useful life prediction, generate a time-varying curve of feature contribution, and determine the contribution ratio of each key feature in different life cycles, including the new model stage, the mid-term stage and the pre-failure stage.

[0148] S4-2, After completing the prediction of the remaining lifespan for the preset number of times, rank the features by contribution, retain the features of the first preset number, remove redundant features, and update the model input feature set.

[0149] S4-3, Establish a preset knowledge base of high contribution features and maintenance actions, and match and output corresponding maintenance decision suggestions from the preset knowledge base according to the type and value of high contribution features.

[0150] Furthermore, the interpretability AI technology in step S4-1 is a Shapley additive interpretation algorithm improved with physical prior constraints. The improvements include: assigning high computational weights to features directly related to the mold failure mechanism and low computational weights to features with indirect influence; and dynamically updating the contribution curve using a sliding time window mechanism.

[0151] It should be noted that S4-1: Interpretable AI technology: This specifically refers to a Shapley additive interpretability algorithm improved with physical prior constraints. It is a method derived from game theory used to fairly allocate a prediction result among all input features.

[0152] Basic principle: The calculation process of this algorithm is likened to a "cooperative game": the "predicted remaining lifespan" is the total revenue of the game, and each input feature is a player participating in the game. The algorithm calculates the marginal contribution of each player in all possible player combinations and finally fairly determines the player's contribution to the total revenue.

[0153] Improved Physical Prior Constraints: Standard algorithms assume all features are equal. The key improvement of this invention lies in introducing physical prior knowledge as constraints. Based on domain knowledge, it is pre-defined which features are directly related to the mold failure mechanism and which have indirect influences. When calculating the contribution, the calculation process of directly related features is given higher weight, making their contribution more easily highlighted; while features with indirect influences are given lower weight. This makes the interpretation results not only mathematically fair but also more in line with physical laws.

[0154] Feature contribution: A value calculated by the above algorithm, which can be positive or negative. A positive contribution indicates that the appearance or increase of the feature makes the model predict a longer remaining lifetime; a negative contribution indicates that the appearance or increase of the feature makes the model predict a shorter remaining lifetime, that is, it indicates degradation.

[0155] Feature contribution time-varying curve: Connect the feature contributions calculated at each time point, such as every 100 stampings, to form a curve that changes over time; this curve clearly shows which features dominate the model's decision-making at different stages of the mold's life.

[0156] Key characteristics of different lifecycles:

[0157] New mold stage: Features with high contribution may be related to manufacturing process parameters and installation conditions, such as machining accuracy and hardness, which define the initial baseline of lifespan.

[0158] In the mid-term stage: the root mean square ratio of vibration signals, the thermal cycling asymmetry of temperature signals, and the contribution of features related to stable wear and thermal fatigue will continue to rise and become dominant.

[0159] Early stage of failure: The sudden energy of acoustic emission, the kurtosis of vibration signals and the characteristic contribution of sudden failure modes will increase sharply, becoming the strongest signal to predict the end of life.

[0160] Contribution ranking: After the system has been running for a period of time, for example, after 1000 predictions have been completed, the average contribution of all input features during this period is calculated and sorted from high to low.

[0161] Redundant features are removed: key features with the highest ranking, such as the top 20, are retained, and those features whose contribution is consistently close to zero are decisively removed. Redundant features may be noise or information that is irrelevant to the current degradation mode. This operation reduces the dimensionality of the model input, improves the subsequent computation speed and generalization ability, and reduces the risk of overfitting.

[0162] Update the model input feature set: Retrain or fine-tune the model in step S3 using the optimized new feature set to make it more streamlined and efficient.

[0163] High-contribution features and maintenance action preset knowledge base: This is a pre-built database based on expert experience that establishes a mapping relationship between "feature-problem-action".

[0164] Example Rule 1: If the feature with the highest contribution is the burst energy of acoustic emission, and its value exceeds the threshold A, then the conclusion is that the microcrack is actively propagating. It is recommended to perform fluorescence penetrant testing and arrange for the mold to be taken offline for maintenance within 50 stamping cycles.

[0165] Example Rule 2: If the feature contributing the most is vibration kurtosis and its value continues to rise, then the conclusion is that wear has entered an accelerated phase and spalling may occur. It is recommended to increase the lubrication frequency and re-evaluate after 200 stamping cycles.

[0166] Matching and Output: The system monitors the feature with the highest contribution and its value in real time, matches it with the rules in the knowledge base, and immediately outputs specific and actionable maintenance decision suggestions to the maintenance personnel on the operation interface, thereby directly transforming the data analysis results into productivity.

[0167] Explanation of the sliding time window mechanism: The contribution curve is dynamically updated using a sliding time window mechanism, which means that the system does not calculate the contribution of all historical data, but only focuses on data within the most recent period. For example, the most recent window contains 500 stamping cycles. The window slides forward as new data arrives, discarding old data and adding new data. This method ensures that the contribution analysis can dynamically and in real time reflect the current health status and dominant failure mode of the mold, rather than being averaged by outdated historical data, thus making the interpretation results always highly correlated with the current situation.

[0168] Step S5: Deploy the integrated model to the edge computing device and adjust the evaluation strategy according to the degradation characteristics of the mold at different life cycles to achieve real-time life assessment and decision support.

[0169] Step S5 further includes the following sub-steps:

[0170] S5-1 deploys the trained ensemble model, along with the dynamic feature selection and maintenance decision mapping functions implemented based on interpretability analysis, to an edge computing device that supports industrial Ethernet communication.

[0171] S5-2, in the new mold stage, calls manufacturing process data to calibrate the initial life baseline and reduce the new mold evaluation error. In the mid-term stage, it strengthens the trend analysis of the working condition monitoring signal to capture wear accumulation characteristics. In the early stage of failure, it starts the high-frequency sampling mode to enhance the ability to capture micro-crack characteristics.

[0172] S5-3, the edge computing device receives monitoring data in real time, automatically performs feature extraction, signal decomposition, model inference and decision output processes, and outputs the remaining service life prediction value and maintenance suggestions;

[0173] S5-4 uses periodically accumulated operational data to adjust the model deployed on edge computing devices, fixing the underlying network parameters and optimizing the meta-decision layer parameters during the adjustment process.

[0174] It should be noted that edge computing devices refer to industrial-grade high-performance computers deployed on-site in the workshop, close to stamping machines and molds. Unlike cloud servers, they are characterized by their robustness, low power consumption, rich industrial interfaces, and adaptability to the high temperature, dust, and vibration environments of the workshop.

[0175] Supports industrial Ethernet communication: This means that the device can directly exchange data at high speed with machine tool PLCs, sensor networks, and workshop manufacturing execution systems via industrial Ethernet protocols such as Profinet, EtherCAT, Modbus, and TCP / IP, ensuring the real-time performance and reliability of the data link.

[0176] Deployment details:

[0177] The trained ensemble model: Deploy the trained model parameters and structure files from step S3 onto the device.

[0178] Dynamic feature selection function: The feature selection logic formed in step S4 based on interpretability analysis, such as retaining only the top 20 important features, is encoded into a software function and deployed together.

[0179] Maintain the decision mapping function: Encode the "feature-problem-action" knowledge base and matching logic established in step S4 into software functions and deploy them together.

[0180] The edge system of this invention is not static, but possesses intelligent strategy adjustment capabilities:

[0181] New model stage: The new model is in its best performance state and the monitoring signal is stable. If the model trained on a large amount of degraded data is used directly for prediction, it may produce a large deviation.

[0182] Strategy: The system calls manufacturing process data to calibrate the initial lifespan baseline, reads the theoretical design lifespan of the mold, and uses the operating data of the first few hours of the new mold to fine-tune the initial output of the model, significantly reducing the evaluation error of the new mold and providing a reasonable starting point.

[0183] Mid-term stage: The mold enters a stable wear period, with slow but continuous degradation and a clear trend.

[0184] Strategy: Strengthen the trend analysis of operating condition monitoring signals. The system adopts an appropriate sampling frequency and pays more attention to the trend indicators of moving average lines and slope changes of characteristic values. It can keenly capture wear accumulation characteristics and achieve stable tracking of long-term slow degradation.

[0185] Early stage of failure: microcracks initiate and propagate, wear intensifies, and numerous transient and sudden events appear in the signal.

[0186] Strategy: Activate high-frequency sampling mode. When the system detects that the contribution of certain key features continues to increase and exceeds the threshold, it automatically triggers the signal acquisition card and sensors to enter high-speed sampling mode, such as increasing the sampling rate from 10kHz to 100kHz, to enhance the ability to capture high-frequency features of microcracks and capture key evidence for the final accurate end-of-life prediction.

[0187] Once the edge computing device is online, a fully automated closed-loop process begins:

[0188] Real-time reception: Continuously receive raw waveform data from sensors via industrial Ethernet.

[0189] Automatic execution: The device automatically executes step S1 (feature extraction), step S2 (signal decomposition), and step S3 (model inference) sequentially.

[0190] Decision output: The system instantly calculates and outputs the predicted remaining service life and maintenance recommendations, which can be communicated to on-site engineers via screen, indicator lights, or SMS.

[0191] Regularly accumulated operational data: The system stores new and valid data generated during daily operation, forming a continuously growing local dataset.

[0192] Adjusting the model: Every so often, such as once a month, the deployed model is fine-tuned and optimized using this new local dataset so that it can adapt to the specific working conditions of this equipment and this set of molds, making the model's prediction results more accurate.

[0193] The underlying network parameters refer to the parameters of the layers used to extract abstract features. These parameters are general feature extractors that have been trained on a large amount of data. Therefore, fixing them can prevent them from being "biased" when there is insufficient new data and maintain the stability of the model.

[0194] Meta-decision layer parameters: These refer to the parameters of the layer that ultimately integrates all information to predict lifespan, much like a "decision committee." Its decision-making rules can be optimized and fine-tuned based on new patterns revealed by new data. This approach can absorb new knowledge while avoiding overturning the model's existing capabilities, making it an efficient and reliable incremental learning strategy.

[0195] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for evaluating the service life of a mold based on machine learning, characterized in that, Includes the following steps: Step S1: Collect mold manufacturing process parameters, working condition monitoring signals and workshop environmental data, and extract features related to the physical failure mode of the mold from the online working condition monitoring signals; Step S2: Adaptively decompose the collected working condition monitoring signals, and adjust the decomposition parameters to obtain signal components with the goal of consistency between the decomposition results and the theoretical physical model. Step S3: Construct an integrated model that combines deep learning and traditional machine learning, fuse features and signal components to obtain the predicted remaining lifespan of the mold. Step S4: Analyze the feature contribution based on interpretable artificial intelligence technology, optimize the model input feature set according to the contribution, and output maintenance decision suggestions based on the analysis results; Step S5: Deploy the integrated model to the edge computing device and adjust the evaluation strategy according to the degradation characteristics of the mold at different life cycles to achieve real-time life assessment and decision support. Step S2 further includes the following sub-steps: S2-1, Initialize the parameters of the fully integrated empirical mode decomposition adaptive noise algorithm, and set the initial range of noise standard deviation, the initial range of the overall mean number of iterations, and the maximum number of iterations; S2-2 integrates the vibration signal input with the empirical mode decomposition adaptive noise algorithm to obtain multiple sets of intrinsic mode functions and a residual sequence. S2-3, calculate the correlation coefficient between the decomposed residual sequence and the theoretical wear amount calculated based on the Achad wear model, and calculate the sample entropy of the intrinsic mode function components. The specific formula of the Achad wear model is as follows: Where dW is the wear depth and K is the wear factor. Here, is the normal pressure, u is the sliding velocity, and H is the mold hardness; S2-4, Adjust the parameters of the fully integrated empirical mode decomposition adaptive noise algorithm by grid search until the correlation coefficient reaches the preset threshold and the sample entropy is minimized, and determine the decomposition parameters; S2-5, Based on the correlation between the components and physical characteristics, eliminate intrinsic mode function components that are unrelated to mold degradation, and retain the effective intrinsic mode function components and residual sequences. The "unrelated to mold degradation" means that the correlation of the components is lower than a preset threshold. Step S3 further includes the following sub-steps: S3-1, Construct a multi-layer architecture consisting of an encoder for processing long-term trend features, a recurrent neural network for processing short-term dynamic features, an attention fusion layer, and a meta-decision layer. The attention fusion layer includes a spatial attention sub-layer and a temporal attention sub-layer. The spatial attention sub-layer highlights the features of the cavity stress concentration area, and the temporal attention sub-layer strengthens the features of the early stage of failure. S3-2, combine the extracted features and screened signal components with manufacturing process parameters and workshop environment data to form a feature vector, and use the remaining number of working cycles as the sample label to divide the training set and validation set according to a preset ratio; S3-3, adopts a composite loss function that integrates prediction error and physical consistency penalty term, and uses an adaptive optimization algorithm for training. The training strategy includes an adaptive learning rate adjustment mechanism and an early stopping mechanism. S3-4, the training set is input into the ensemble model for iterative training, the model hyperparameters are dynamically adjusted, and the performance is evaluated using the validation set until the model converges and the final model is saved. The model hyperparameters include the number of hidden layer nodes in the converter and the random deactivation probability of the gated recurrent unit.

2. The machine learning-based mold life assessment method according to claim 1, characterized in that: Step S1 further includes the following sub-steps: S1-1, Obtain manufacturing process parameters from the manufacturing execution system, the manufacturing process parameters including mold material parameters, heat treatment process parameters and machining accuracy parameters; S1-2, collects operating condition monitoring signals through a multi-source sensor network, the operating condition monitoring signals including vibration signals, acoustic emission signals and temperature signals; S1-3, Collect workshop environmental data through workshop environmental sensors, including ambient temperature, humidity and dust concentration; S1-4, calculate the kurtosis index and root mean square ratio from the vibration acceleration signal, as early warning characteristics of the transition of wear from the steady period to the accelerated period; S1-5, the energy of sudden events and the cumulative amount of ringing counts are statistically analyzed from the acoustic emission signal stream, which serves as the basis for determining the initiation and propagation of microcracks; S1-6 calculates the maximum temperature rise rate and thermal cycling asymmetry from thermocouple temperature data as characteristic parameters for assessing the accumulation of thermal fatigue damage.

3. The machine learning-based mold life assessment method according to claim 2, characterized in that: Step S1 also includes frequency domain feature extraction: performing Fourier transform on the vibration signal to extract the energy proportion of specific frequency bands corresponding to macroscopic wear and microscopic cracks; performing wavelet packet decomposition on the acoustic emission signal to extract the kurtosis factor of specific frequency bands corresponding to crack propagation, as a sensitive feature of crack propagation.

4. The machine learning-based mold life assessment method according to claim 1, characterized in that: The parameters of the Achad wear model in steps S2-3 are determined as follows: for different mold materials, multiple sets of standard wear test data are collected, and the material-specific wear factor is obtained by fitting using the least squares method; the normal pressure is obtained by finite element simulation of the mold to obtain the pressure distribution of key parts of the cavity, and the sliding speed is calculated based on the stroke and time of the stamping equipment.

5. The machine learning-based mold life assessment method according to claim 1, characterized in that: The attention fusion layer in step S3-1 calculates dynamic weights according to the following formula: ,in, and The attention score is calculated as follows: t is the current time step, T is the total time step of the input sequence, and k is the loop index, iterating through all time steps from 1 to T. Let be the attention weight at time t. The preceding network outputs a feature vector, where v and R are learnable parameter matrices and b is a bias vector. The spatial attention sublayer generates a spatial weight map through convolution operations, assigning higher weights to features in stress concentration regions. The temporal attention sublayer assigns higher weights to features in the early stages of failure.

6. The machine learning-based mold life assessment method according to claim 1, characterized in that: Step S4 further includes the following sub-steps: S4-1, Input the validation set into the trained model, use interpretable artificial intelligence technology to calculate the contribution of each input feature to the remaining useful life prediction value, generate a time-varying curve of feature contribution, and determine the contribution ratio of each key feature in different life cycles, wherein the different life cycles include the new model stage, the mid-term stage and the pre-failure stage. S4-2, After completing the prediction of the remaining lifespan for the preset number of times, rank the features by contribution, retain the features of the first preset number, remove redundant features, and update the model input feature set. S4-3, Establish a preset knowledge base of high contribution features and maintenance actions, and match and output corresponding maintenance decision suggestions from the preset knowledge base according to the type and value of high contribution features.

7. The machine learning-based mold life assessment method according to claim 6, characterized in that: The interpretability AI technology in step S4-1 is the Shapley additive interpretation algorithm improved with physical prior constraints. The improvement includes: assigning high computational weights to features directly related to the mold failure mechanism and low computational weights to features that indirectly affect it; the time-varying curve of the feature contribution is dynamically updated using a sliding time window mechanism.

8. The machine learning-based mold life assessment method according to claim 1, characterized in that: Step S5 further includes the following sub-steps: S5-1, the trained ensemble model, along with the dynamic feature selection function and maintenance decision mapping function implemented based on interpretability analysis, are jointly deployed to an edge computing device, which supports industrial Ethernet communication; S5-2, in the new mold stage, calls manufacturing process data to calibrate the initial life baseline and reduce the new mold evaluation error. In the mid-term stage, it strengthens the trend analysis of the working condition monitoring signal to capture wear accumulation characteristics. In the early stage of failure, it starts the high-frequency sampling mode to enhance the ability to capture micro-crack characteristics. S5-3, the edge computing device receives monitoring data in real time, automatically performs feature extraction, signal decomposition, model inference and decision output processes, and outputs the remaining service life prediction value and maintenance suggestions; S5-4, periodically accumulated operational data, is used to adjust the model deployed on edge computing devices. During the adjustment, the underlying network parameters are fixed and the meta-decision layer parameters are optimized.

Images