CN-121980969-A - CNN-LSTM-based high-temperature safety evaluation method for die steel

Abstract

The invention relates to a CNN-LSTM-based high-temperature safety evaluation method for die steel, which comprises the steps of firstly constructing a multi-dimensional parameter system covering mechanical property, thermal fatigue property, tissue stability, service environment and dynamic interaction, completing data preprocessing through abnormal value correction, dynamic sliding window standardization and feature interaction item construction, then adopting a deep learning architecture of CNN feature extraction-attention weight distribution-LSTM time sequence modeling-full connecting layer output to realize deep mining and time sequence evolution analysis of multi-parameter coupling features, and finally establishing a security level judgment standard with definite grades based on security risk values output by a model, thereby providing a quantitative and qualitative combined evaluation scheme for the high-temperature service safety of the die steel. According to the invention, the technical applicability is ensured through an incremental training and threshold calibration mechanism, the threshold can be calibrated for different die steel types, the model is updated along with service data accumulation, and the method is suitable for diversified high Wen Fuyi scenes.

Inventors

• WANG PEIJIAN
• ZHOU LAN
• WANG XIAO
• ZHANG FUBAO
• WU SHITIAN
• LI QIUXU
• LI HONGYU
• LI PENGFEI

Assignees

• 广东鸿图南通压铸有限公司
• 南通大学

Dates

Publication Date

20260505

Application Date

20260403

Claims (4)

1. 1. The method is characterized by comprising the collaborative design of multidimensional parameter system construction, data preprocessing, deep learning model modeling and safety grade judgment, and specifically comprises the following steps of: Firstly, multi-source data acquisition is carried out, and a multi-dimensional risk assessment index system covering mechanical properties, fatigue degradation, tissue stability and service environment is constructed according to the multi-source data acquisition; Then, entering a data processing link, and enhancing data characterization by introducing characteristic structures such as speed-carbide cooperative terms; on the basis, a CNN-LSTM-Attention deep learning fusion model is adopted to carry out core modeling, and finally the model integrates the characteristics through a full connection layer to output a security risk score; Based on the scores, the system carries out security level judgment, service life prediction and optimization strategy generation; In addition, in order to ensure the continuous effectiveness of the model, the system realizes online updating and self-adaptive optimization through an incremental learning and periodical threshold calibration mechanism.
2. 2. The method is characterized in that a CNN-LSTM-Attention deep learning fusion model takes 15-dimensional preprocessing characteristics as input, local correlation characteristics of 'mechanical-thermal-organization' parameters are extracted through a first convolution layer and a convolution formula, the local correlation characteristics are reduced in size through a first pooling layer, global characteristic fusion of cross-parameter categories is realized through a second convolution layer, after characteristic fusion, an Attention mechanism layer distributes weights to key failure characteristics through a score calculation and normalization process, an LSTM layer captures the time sequence evolution rule of parameters through a control mechanism, and finally high-dimensional time sequence characteristics are mapped into safety risk values through a full-connection layer and an output layer formula; wherein the input layer data preprocessing stage is performed in three steps: step1, correcting an abnormal value, namely correcting an out-of-range original monitoring value based on a physical reasonable boundary of a parameter, wherein a correction formula is as follows: ; wherein the first parameter is the original monitoring value, 、 Is a physically reasonable boundary for the parameter of item, and is within an industrially reasonable range, Is the corrected value; Step2 dynamic sliding Window normalization, with Sliding window updating statistics for each monitoring period to eliminate timing drift effects by Calculating a sliding average value and then according to Letter calculation sliding standard deviation: ; finally, dynamic standardization is realized: ; Wherein the method comprises the steps of For the current moment of monitoring, Is the first The value after the time-of-day correction, Is the first The sliding statistics of the moment, i.e. mean, standard deviation, As a value after the normalization to the value, Is the sliding window size; Step3, constructing a characteristic interaction item, capturing parameter coupling, constructing three types of interaction items aiming at a multi-parameter synergistic effect of high-temperature failure, and finally forming a 15-dimensional characteristic vector as a model input, wherein the three types of interaction items comprise 12 original parameters and three types of interaction items, and the construction formulas of the three types of interaction items are as follows: A. temperature-carbide coarsening synergy, multiplicative coupling: ; B. The hardness-crack propagation inversion term, division coupling, Avoiding denominator 0: ; C. load-thermal fatigue nonlinear terms, power-coupled, exponential fit by 100 sets of experiments: ; the 15-dimensional eigenvectors of the final input model are: ; Wherein the method comprises the steps of , , Three types of characteristic interaction items are respectively adopted, In order to be able to operate at a dynamic temperature, The coarsening rate of the carbide is increased, In order to achieve the high-temperature hardness, In order to accumulate the cyclic load, The total strain amplitude for the t-th cycle, i.e. the elastic and plastic strain amplitudes, For the temperature range of the t-th cycle, α, β are the fitting indices.
3. 3. The method for evaluating the high-temperature safety of the die steel based on the CNN-LSTM is characterized in that in the modeling stage of a CNN-LSTM hybrid deep learning model, the model adopts a four-module architecture of CNN feature extraction, attention weight distribution, LSTM time sequence modeling and full connection layer output, and the output of the former module is used as the input of the latter module to ensure that variables are consistent and physical significance is clear; The CNN feature extraction module is used for extracting CNN static features and capturing by adopting spatial correlation; A. layer 1 convolution local feature extraction: ; Wherein, the ( The number of convolution kernels), ( , Is the core size), , Represents a set of real numbers, Represented as A dimension real space, a He normal distribution initialization, The number of the cells is initialized to 0, In order to activate the function, Is the first Core number one Local features of the individual locations; B. layer 1 maximum pooling, feature dimension reduction: ; Wherein, the ( , For the step size), The local optimal characteristics are reserved after pooling; C. layer 2 convolution, global feature fusion: ; Wherein, the ( ), ( ), , , Is a cross-category global feature; attention weight distribution module, attention weight distribution, highlighting key features: A. attention score calculation: ; Wherein, the The Xavier normal distribution is initialized, , Is the first Importance scores for the nth feature of the kernel; B. Attention weight normalization: ; Wherein, the For normalized weights, the sum is 1, such as "crack propagation feature weight=0.2, dislocation density feature weight=0.05"; C. Attention enhancing features: ; Wherein, the Highlighting the contribution of the key failure feature as a weighted feature; the LSTM time sequence modeling module is used for modeling LSTM time sequence and capturing an evolution rule; A. An input gate controlling the current feature input: ; B. Forget door, control history state retention: ; C. cell status update, storing timing information: ; D. An output gate controlling cell state output: ; E. hidden layer output, time sequence feature compression: ; Wherein, the , The dimensions of the hidden layer, , , The dimensions of the attention feature, , As a function of the Sigmoid, Is the product of the elements and the product of the elements, The state is the previous time, and the time sequence is guaranteed to be consistent; The full-connection layer output module is used for entering the full-connection layer output module to calculate the security risk value, and the method comprises the following specific steps of firstly passing through ; Mapping high-dimensional features into one-dimensional raw scores, where The total time sequence step length is the total time of the current service, Is the full connection layer The weight, the He normal distribution is initialized, Is the first Time LSTM hidden layer The characteristics of the device are that, For the full connection layer bias term, Is one-dimensional original score, then passes Sigmoid normalization is carried out, the original score is mapped to the [0,1] interval, and finally the method is carried out by And (3) performing risk value scaling, and mapping the normalized score to a range of [0,100] to obtain a final high-temperature service safety risk value of the die steel.
4. 4. The method for evaluating high-temperature safety of CNN-LSTM based die steel as defined in claim 3, wherein the safety risk value is based on the safety level judgment stage (Range [0,100 ]) 5 security classes and explicit handling advice, class classification formula is ; The treatment suggestions corresponding to the grades are respectively that the performance of the die steel is stable under the extremely safe grade, the die steel can be continuously used for 1 time every 1000 working cycles, the die steel has no failure risk under the safe grade, 1 time every 500 working cycles is used for the sampling inspection, the die steel is slightly degraded under the critical safe grade, the die steel needs to be detected and made into a maintenance plan every 200 working cycles, the die steel has high failure risk under the unsafe grade, the repair feasibility needs to be immediately stopped for detection and evaluation, and the die steel is close to the failure threshold under the extremely unsafe grade, and is forbidden to be used and replaced.

Description

CNN-LSTM-based high-temperature safety evaluation method for die steel Technical Field The invention relates to a high-temperature service safety evaluation technology of die steel, belongs to the field of intersection of deep learning and material service performance analysis, and is suitable for a safety state monitoring and service life evaluation scene of the die steel under a high-temperature working condition. Background In the existing high-temperature service safety evaluation technology of die steel, part of schemes rely on single mechanical property parameters or simple linear models to carry out safety judgment, and failure mechanisms of thermal fatigue-tissue degradation are not fully considered. In addition, although multi-parameter analysis is introduced in the scheme, an effective characteristic interaction mechanism is not constructed, and the synergic effect among parameters is difficult to capture. Part of technologies adopt a traditional machine learning model for evaluation, and lack of capturing capability of a time sequence evolution rule leads to insufficient evaluation precision and timeliness, and cannot meet real-time and accurate safety evaluation requirements in industrial scenes. Patent number CN202410518925.9 is a method, system, equipment and medium for evaluating service safety of a metal structural member, and the method realizes indirect evaluation of service safety of the metal structural member through curve fitting of corrosion parameters and mechanical parameters, and has certain convenience in a conventional environment. The method aims at the high-temperature service scene of the die steel, has obvious limitations that firstly, the special thermal fatigue-tissue degradation failure mechanism of the high-temperature service of the die steel is not covered, only the single association between the focus corrosion and the mechanical property is realized, and the real safety state of the synergistic effect of multiple failure modes at high temperature can not be reflected; secondly, a system preprocessing flow for correcting abnormal values in industrial monitoring data and eliminating time sequence drift is lacking, the problem of parameter fluctuation caused by sensor errors in a high-temperature environment is difficult to solve, thirdly, a characteristic interaction item adapting to high-temperature failure characteristics is not constructed, the synergistic effect among parameters such as temperature-load, tissue-mechanics and the like cannot be captured, and finally, the precision and generalization capability of the method in the high-temperature service safety evaluation of the die steel cannot meet the requirements. Patent number CN202211700961.4 is entitled "a method and system for safety monitoring based on the hardness of a metal component", which implements safety monitoring of a metal component based on hardness parameters by establishing a mathematical correlation model of the hardness, yield strength and strain hardening index. The method is applied to high-temperature service safety evaluation of the die steel, and has the advantages that firstly, a linear correlation is built by a model according to a single hardness dimension, a failure mechanism of thermal fatigue-tissue degradation under the high temperature of the die steel is not considered, a failure process driven by multiple parameters together under a high-temperature environment cannot be covered, secondly, a data processing scheme of a system is not designed according to the specificity of monitoring data under an industrial scene (such as parameter drift and accidental errors under the high-temperature working condition), a characteristic interaction item adapting to high-temperature failure logic is not built, so that the characteristic quality of an input model is limited, thirdly, a traditional mathematical correlation model is adopted, the dynamic law of the evolution of parameters over time in the high-temperature service process of the die steel cannot be captured, and a attention mechanism is not introduced to highlight key failure characteristics, so that the evaluation result has limitation, and the suitability to a complex high-Wen Fuyi scene is insufficient. The patent number CN202411951215.1 is named as an internal threat detection method based on a CNN-LSTM algorithm, the internal threat detection precision is improved by fusion of CNN and LSTM in the network information safety field, but the core application scene and the high-temperature service safety evaluation of the die steel have essential differences, and the method has the specific limitations that firstly, the design is not carried out aiming at a 'thermal fatigue-tissue degradation' coupling failure mechanism under the high-temperature service scene of the die steel, the model architecture is not matched with the physical logic of high-temperature failure, secondly, the data preprocessing flow is only carried out aiming