Deep learning-based residual life prediction method for gear of cutting speed reducer of heading machine

Abstract

The invention discloses a heading machine cutting speed reducer gear remaining life prediction method based on deep learning, and relates to the life prediction related technical field, and the method comprises the steps: carrying out the multi-source heterogeneous sensing collection of a speed reducer gear, carrying out the interactive transmission to an operation and maintenance platform end according to an interface, executing the dual-order attention composition, and determining an attention state diagram; triggering a fuzzy discrimination threshold embedded in the platform, performing stage discrimination on the upper state and the attention state diagram, and activating a stage baseline relationship; and initializing the wear prediction unit, executing wear prediction and phase residual life calculation based on the attention state diagram, and determining a life prediction result. The technical problems that in the prior art, due to the fact that stage features of the gear degradation process are not depicted enough, dynamic degradation in the whole life cycle is difficult to adapt, and the precision and stability of the life prediction result are not enough are solved, and the technical effect of improving the prediction precision and stability of the remaining life of the heading machine cutting speed reducer gear is achieved.

Description

Technical Field

[0001] This application relates to the technical field of life prediction, and in particular to a method for predicting the remaining life of gears in a tunnel boring machine cutter reducer based on deep learning. Background Technology

[0002] As core equipment in tunneling engineering and mining, tunneling machines' critical components, such as the cutting reducer gears, operate under harsh conditions of high intensity, variable load, and strong impact for extended periods. This makes them prone to wear, fatigue, pitting, and even tooth breakage, leading to machine downtime and potentially major safety accidents. Therefore, accurate prediction of the remaining service life of the reducer gears is crucial. Traditional remaining service life prediction methods, which rely on establishing gear dynamics equations and wear evolution equations combined with operating parameters, have poor generalization ability in complex real-world environments. They struggle to dynamically adapt to individual differences and changes in operating conditions, failing to effectively uncover deep degradation characteristics and thus affecting the accuracy and stability of cutting reducer gear life prediction.

[0003] Current technologies suffer from insufficient characterization of the phased characteristics of gear degradation and difficulty in adapting to dynamic degradation throughout the entire life cycle, resulting in inadequate accuracy and stability of life prediction results. Summary of the Invention

[0004] This application provides a deep learning-based method for predicting the remaining life of gears in tunnel boring machine cutter reducers. This method addresses the technical problems in existing technologies, such as insufficient characterization of the stage-specific features of gear degradation and difficulty in adapting to dynamic degradation throughout the entire life cycle, which leads to insufficient accuracy and stability of life prediction results. This method achieves the technical effect of improving the accuracy and stability of remaining life prediction for tunnel boring machine cutter reducers.

[0005] This application provides a deep learning-based method for predicting the remaining life of a tunnel boring machine cutter reducer gear. The method includes: collecting multi-source heterogeneous sensor data on the reducer gear, transmitting the data to the operation and maintenance platform via interface interaction, performing a two-order attention graph construction to determine the attention state graph; triggering a fuzzy discrimination threshold embedded in the platform to perform stage discrimination between the upper-level state and the attention state graph and activate the stage baseline relationship, wherein the life cycle is divided into an initial linear decay relationship, a mid-term exponential growth relationship, and a late-term nonlinear failure relationship; initializing a wear prediction unit based on the stage baseline relationship, performing wear prediction and stage remaining life calculation based on the attention state graph, and determining the life prediction result, wherein the life prediction result is in the form of a ternary fuzzy number.

[0006] In a possible implementation, the deep learning-based method for predicting the remaining life of the gears in the tunnel boring machine cutter reducer further performs the following processing: constructing a first attention threshold with information focus and explicit relationship as the first-order attention objective; constructing a second attention threshold with implicit relationship and dynamic weight as the second-order attention objective; determining multi-source heterogeneous data and performing feature verification and homogeneous fitting reconstruction of the first and second attention thresholds to generate the attention state diagram.

[0007] In a possible implementation, the deep learning-based method for predicting the remaining life of the cutter gear of a tunneling machine further performs the following processing: the multi-source heterogeneous sensing type includes at least oil data, vibration signals, temperature signals, and acoustic emission signals; short-term acquisition based on the multi-source heterogeneous sensing type is performed with synchronization timestamp constraints to determine the multi-source heterogeneous data.

[0008] In a possible implementation, the deep learning-based method for predicting the remaining life of the cutting reducer gear of a tunneling machine further performs the following processing: adopting an adversarial training method to perform iterative training under a generation-decision architecture, wherein the generation architecture is used to generate the sample superordinate state and sample attention state map in multi-stage scenarios, and the determination architecture performs stage probability determination, and performs determination feedback training with the consistency test between the stage probability and the stage scenario; the determination architecture trained to convergence is used as the fuzzy discrimination threshold and embedded in the operation and maintenance platform.

[0009] In a possible implementation, the deep learning-based method for predicting the remaining life of the cutting reducer gear of a tunneling machine further performs the following processing: connecting to the operation and maintenance database and retrieving the upper-level status, wherein the upper-level status is the life prediction result of the upper-level time node; taking the upper-level status and the attention state diagram as input and the life cycle stage as output, performing a stage probability determination based on a fuzzy discrimination threshold and activating the stage baseline relationship.

[0010] In a possible implementation, the deep learning-based method for predicting the remaining life of the cutter reducer gear of a tunneling machine further performs the following processing: obtaining the specification information of the reducer gear and determining the constraints according to the installation scenario; using the specification information and constraints, retrieving operation records and organizing life curves to determine multiple life curves; for the multiple life curves, locating stage segmentation nodes based on the curve trend, performing linear fitting for the initial stage, exponential fitting for the middle stage, and nonlinear fitting for the later stage, and adding them to the stage baseline relationship.

[0011] In a possible implementation, the deep learning-based method for predicting the remaining life of the cutter gear of a tunneling machine further performs the following processing: the stage baseline relationship is activated as any one of a linear decay relationship, an exponential growth relationship, or a nonlinear failure relationship.

[0012] In a possible implementation, the deep learning-based method for predicting the remaining life of the cutting reducer gear of a tunneling machine further performs the following processing: for the attention state map, locate the fluctuation elements and perform uncertainty quantification to determine the confidence interval, wherein the confidence interval corresponds one-to-one with the fluctuation elements; using the confidence interval, determine the maximum value group, the median value group, and the minimum value group; based on the maximum value group, the median value group, and the minimum value group, perform three-thread prediction based on the attention state map to determine the ternary fuzzy number, wherein the ternary fuzzy number includes the predicted minimum value, the predicted median value, and the predicted maximum value.

[0013] In a possible implementation, the deep learning-based method for predicting the remaining life of the cutter gear of a tunneling machine further performs the following processing: identifying the fluctuation elements in the attention state diagram according to the maximum value group, performing wear prediction and stage remaining life calculation according to the initialized wear prediction unit, and determining the predicted minimum value.

[0014] In a possible implementation, the deep learning-based method for predicting the remaining life of the cutting reducer gear of a tunneling machine further performs the following processing: integrating the ternary fuzzy number as the life prediction result; adding the life prediction result into the operation and maintenance database, performing time-series prediction updates, and displaying it in a pop-up window on the operation and maintenance platform's display interface.

[0015] This application proposes a deep learning-based method for predicting the remaining life of gears in a tunnel boring machine (TBM) cutter reducer. The method involves multi-source heterogeneous sensing of the reducer gears, transmitting the data via interface interaction to an operation and maintenance platform for two-stage attention mapping to determine the attention state graph. This is followed by triggering a fuzzy discrimination threshold embedded in the platform to perform stage-based discrimination between the upper-level state and the attention state graph, activating the stage baseline relationship. The wear prediction unit is then initialized, and wear prediction and stage-based remaining life calculation based on the attention state graph are performed to determine the life prediction result. This method addresses the technical problems in existing technologies, such as insufficient characterization of the stage-based characteristics of gear degradation and difficulty in adapting to dynamic degradation throughout the entire life cycle, leading to insufficient accuracy and stability in life prediction results. Ultimately, this method improves the accuracy and stability of remaining life prediction for TBM cutter reducer gears. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments of the present invention will be briefly described below. Flowcharts are used in this application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed precisely in sequence. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from these processes.

[0017] Figure 1 A flowchart illustrating the method for predicting the remaining life of a tunneling machine cutting reducer gear based on deep learning, provided in an embodiment of this application. Detailed Implementation

[0018] The above description is merely an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, specific embodiments of this application are given below.

[0019] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description of this application will be provided in conjunction with the accompanying drawings. The described embodiments should not be considered as limitations on this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] In the following description, references to "some embodiments" describe a subset of all possible embodiments. However, it is understood that "some embodiments" can be the same or different subsets of all possible embodiments and can be combined with each other without conflict. The terms "first" and "second" are used merely to distinguish similar objects and do not represent a specific ordering of objects. The terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, product, or server that includes a series of steps is not necessarily limited to those steps explicitly listed, but may include other steps not explicitly listed or inherent to such processes, methods, products, or devices. 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 application belongs. The terminology used herein is for the purpose of describing embodiments of this application only.

[0021] This application provides a deep learning-based method for predicting the remaining life of gears in a tunnel boring machine cutter reducer, such as... Figure 1 As shown, the method includes: Step S100: Multi-source heterogeneous sensing data is collected from the reducer gears and transmitted to the operation and maintenance platform via interface interaction. Two-stage attention mapping is performed to determine the attention state diagram.

[0022] Preferably, the reducer gears are subjected to multi-source heterogeneous sensing acquisition, that is, multiple sensors are installed at key monitoring points of the reducer gears. These may include vibration acceleration sensors to collect vibration signals from the gearbox housing surface, acoustic emission sensors to collect high-frequency stress wave signals generated within the material due to fatigue and crack propagation, temperature sensors to collect temperature signals from bearings or oil sump, and online oil monitoring sensors to collect information such as lubricating oil viscosity, moisture content, and abrasive particle count. Then, the data is transmitted to the operation and maintenance platform via an interface, following a predefined communication protocol and data format, through wired or wireless networks for storage and processing. Finally, a two-order attention mapping is performed on the sensor data; specifically, for each… The system processes individual sensor data streams and uses an attention mechanism to calculate the correlation weights between different time points within each data stream. For example, it assigns higher attention weights to the moment when an impact event occurs in the vibration signal, while ignoring signals during the stable operation phase. This results in the output of a weighted feature vector for each sensor data stream. The weighted feature vectors of all sensors are then concatenated, and the attention mechanism is used again to calculate the interrelationships and importance between different sensor features. For instance, if the amplitude of a specific gear meshing frequency in the vibration spectrum increases while the number of ferromagnetic abrasive particles in the lubricating oil surges, indicating severe wear, it is assigned a very high weight. This results in the output of an attention state map, which integrates all sensor information and accurately quantifies their internal and inter-sensory importance.

[0023] Furthermore, step S100 also includes step S110, constructing a first attention threshold with information focus and explicit relationship as first-order attention targets; step S120, constructing a second attention threshold with implicit relationship and dynamic weight as second-order attention targets; and step S130, determining multi-source heterogeneous data and performing feature detection and homologous fitting reconstruction of the first and second attention thresholds to generate the attention state map.

[0024] Preferably, the information focus refers to a point or segment of significant change within a single data stream, such as a sudden impact pulse in a vibration signal. Explicit relationships refer to direct and easily observable correlations in a time series, such as the rise in vibration signal amplitude and the rise in temperature signal occurring almost synchronously in time. Then, based on a neural network, a first attention threshold is constructed using the information focus and explicit relationships as first-order attention targets. This threshold is used to scan the input single-source data and calculate weights for each time point based on the attention mechanism, marking the importance of the data at that time point.

[0025] Preferably, implicit relationships refer to nonlinear, complex coupling relationships between different sensor data that require in-depth analysis to discover. For example, the slow accumulation of energy of a specific frequency component in a vibration signal may lead to a change in the shape characteristics of a certain abrasive particles in the lubricating oil after a period of time. Dynamic weights refer to the dynamic changes in the importance of different sensors as the equipment degrades. For example, oil data may have the highest weight in the early stage of wear, vibration signals may have the highest weight in the period of severe wear, and temperature signals have the highest weight in the early stage of overheating failure. Based on neural networks, implicit relationships and dynamic weights are used as second-order attention targets to construct a second attention threshold, that is, to dynamically calculate the attention weights between the weighted feature vectors of different sensors after first-order processing through an attention mechanism. Then, multi-source heterogeneous data is identified, and feature verification is performed using a first attention threshold and a second attention threshold. Specifically, the first attention threshold processes the data from each sensor and verifies the time points assigned high weights, while the second attention threshold comprehensively processes the features from all sources and verifies the feature combinations assigned high weights. Next, homogeneous fitting and reconstruction are performed, which involves fusing and reducing the dimensionality of multi-source data guided by the same state element. All features after first-order and second-order attention weighting are fused and transformed through neural network layers to reconstruct an attention state map, thereby improving the accuracy of extracting degradation-related features from complex data.

[0026] Furthermore, step S100 also includes that the multi-source heterogeneous sensing type includes at least oil data, vibration signal, temperature signal, and acoustic emission signal; and performs short-term acquisition based on the multi-source heterogeneous sensing type with synchronous timestamp constraints to determine the multi-source heterogeneous data.

[0027] Preferably, the multi-source heterogeneous sensing type includes at least oil data, vibration signals, temperature signals, and acoustic emission signals. Oil data is used to monitor the chemical and physical state of the lubricating oil and wear products, including the number / type of abrasive particles, PQ index, viscosity, and water content. The number / type of abrasive particles directly reflects the severity and type of wear on the gear surface. The PQ index is used to quickly detect the total amount of ferromagnetic abrasive particles in the oil. Viscosity is a basic performance indicator of lubricating oil; excessively high or low viscosity will affect lubrication and accelerate wear. Water in the oil can lead to lubricating oil emulsification, additive failure, and component corrosion. Vibration signals are used to monitor the dynamic behavior of the gears, including time-domain and frequency-domain characteristics such as root mean square value and peak value. Temperature signals are used to monitor the thermal state, including gearbox temperature and bearing temperature. Acoustic emission signals are used to capture high-frequency stress waves and monitor the dynamic changes in the internal microstructure of materials. Then, short-term acquisition is performed with synchronous timestamp constraints, including cyclically acquiring a short segment of data of a fixed duration at a high frequency to generate multi-source heterogeneous data.

[0028] Step S200: Trigger the fuzzy discrimination threshold embedded in the platform to perform stage discrimination on the upper-level state and the attention state diagram and activate the stage baseline relationship. The life cycle is divided into an initial linear decay relationship, a mid-term exponential growth relationship and a late-term nonlinear failure relationship.

[0029] Step S200 further includes step S201, which adopts an adversarial training method to perform iterative training under the generation-decision architecture. The generation architecture is used to generate the sample upper-level state and sample attention state map in a multi-stage scenario. The determination architecture performs stage probability determination and performs determination feedback training by checking the consistency between the stage probability and the stage scenario. Step S202, the determination architecture trained to convergence is used as the fuzzy discrimination threshold and embedded in the operation and maintenance platform.

[0030] Preferably, an adversarial training approach is used to perform iterative training under the generator-determiner architecture. The generator architecture learns the distribution of real data and generates sample hyperstates and sample attention state maps for multi-stage scenarios. The decision architecture performs stage probability determination, i.e., determining whether the input sample is real or generated data. The generator generates sample hyperstates and sample attention state maps for a certain stage, and the state maps are passed to the discriminator for stage probability determination. Feedback is then given to the generator and the corresponding stage for verification, and further feedback is given to the discriminator. Iteration is performed in multi-stage scenarios to determine a convergent discriminator, which serves as a fuzzy discrimination threshold. Specifically, the generator outputs sample hyperstates and sample attention state maps using random noise vectors and specified stage labels to expand the training data. The sample hyperstates are fabricated. The historical lifetime prediction results are used, and the sample attention state map is a forged feature vector representing a specified stage. Then, the decision architecture is used to perform stage probability determination, outputting the probability of belonging to each stage. Then, the consistency between the stage probability and the stage scenario is tested for decision feedback training. For real data, the decision result of the decision architecture is highly consistent with the real stage label of the data. For generated data, the decision result of the decision architecture is highly consistent with the stage label specified at the time of generation. Then, the consistency metric is quantified into a loss function, and the parameters of the decision architecture are updated through backpropagation. When the adversarial training reaches equilibrium and training is completed, the final decision architecture is used as a fuzzy discrimination threshold and deployed on the operation and maintenance platform. In actual operation, the fuzzy discrimination threshold is called, the real-time calculated upper-level state and attention state map are input, and the current stage discrimination result is output.

[0031] Furthermore, step S200 also includes activating the stage baseline relationship as any one of a linear decay relationship, an exponential growth relationship, or a nonlinear failure relationship.

[0032] Preferably, upon receiving the latest attention state map, the platform's embedded fuzzy discrimination threshold is triggered, and the higher-level state and attention state map are input into it for stage discrimination. The attention state map represents the device's current instantaneous health state, while the higher-level state represents the historical trend, which is the lifespan prediction result of the previous time node, including the previously predicted remaining lifespan and its uncertainty. Then, a probability vector is output, and the stage with the highest probability is selected as the current stage discrimination result. Then, the stage baseline relationship is activated based on the stage discrimination result. The stage baseline relationship is a wear prediction mathematical relationship pre-established for each degradation stage. Specifically, the lifespan cycle is divided into an initial linear decay relationship, a mid-term exponential growth relationship, and a late-term nonlinear failure relationship. The initial linear decay relationship indicates that the remaining lifespan decreases uniformly over time during the initial failure period, with a low and relatively stable wear rate. The mid-term exponential growth relationship indicates that the rate of decrease in the remaining lifespan during the initial wear failure period is continuously accelerating, with pitting, microcracks, etc. appearing. The late-term nonlinear failure relationship indicates that damage accumulates to a critical point at the end of the wear failure period, performance degrades sharply, and complete failure may occur at any time, such as tooth breakage. The wear prediction relationship is intelligently switched according to changes in equipment status. It may be any of the following: linear decay relationship, exponential growth relationship, or nonlinear failure relationship. This allows for accurate prediction at each stage, greatly improving the accuracy and reliability of end-to-end prediction.

[0033] Furthermore, step S200 also includes step S210, connecting to the operation and maintenance database and retrieving the upper-level status, wherein the upper-level status is the lifetime prediction result of the upper-level time node; step S220, taking the upper-level status and attention state diagram as input and the lifetime cycle stage as output, performing stage probability determination based on fuzzy discrimination threshold, and activating the stage baseline relationship.

[0034] Preferably, the operation and maintenance database is used to store historical and real-time data of the equipment, recording sensor data, calculation features, and lifetime prediction results at each time point. Retrieving the upper-level status means obtaining the lifetime prediction results of the upper-level time node. For example, if a prediction is made every 10 minutes, the upper-level time node is the prediction time 10 minutes ago. The lifetime prediction result represents the ternary fuzzy number output of the previous cycle. Then, taking the upper-level status and attention state diagram as input and the lifetime cycle stage as output, the stage probability is determined using the fuzzy discrimination threshold. A probability distribution vector is predicted and output, representing the probability that the input data is judged as the initial, middle, and late stages, respectively. The one with the highest probability is selected as the current stage judgment result. Then, the stage baseline relationship is activated based on the current stage judgment result, and the wear prediction mathematical relationship of the corresponding stage is selected, which can be any one of linear decay relationship, exponential growth relationship, or nonlinear failure relationship. Finally, wear prediction and stage remaining lifetime calculation are performed to obtain the accurate lifetime prediction result at the current moment.

[0035] Furthermore, step S220 also includes step S221, obtaining the specification information of the reducer gear and determining the constraints according to the installation scenario; step S222, using the specification information and constraints, retrieving the operation records and organizing the life curves to determine multiple life curves; step S223, for the multiple life curves, locating the stage segmentation nodes by the curve trend, performing linear fitting for the initial stage, exponential fitting for the middle stage, and nonlinear fitting for the later stage, and adding the stage baseline relationship.

[0036] Preferably, the specifications of the reducer gears are obtained, including the gear's design and physical parameters, such as geometric parameters like module, number of teeth, tooth width, and pressure angle; material parameters like gear material type, heat treatment process, and surface hardness; and design parameters like rated power, rated torque, and design life. The installation scenario refers to the specific working environment and conditions of the gears, which may include equipment type (e.g., used in tunneling machine cutting sections, conveyors, hoists), load characteristics (e.g., heavy load, impact load, uniform load), and working environment (e.g., underground conditions, surface conditions). The reducer gear specifications and installation scenario are used as constraints for retrieving operational records, including filtering from the historical maintenance database to obtain full lifecycle historical data of all eligible gears of the same type from start to finish, including chronologically ordered sensor data for vibration, temperature, and oil. Next, the life curves are processed, which involves processing historical sensor data and calculating multiple health indicators for each time point. For example, the root mean square value of the vibration signal will increase as wear intensifies. Then, with time or number of operating cycles as the horizontal axis and the health indicator value as the vertical axis, a life curve from start to finish is plotted for the life data of each type of gear, and finally multiple life curves are determined.

[0037] Preferably, for multiple lifespan curves, the stage segmentation nodes are located based on the curve trend, dividing the curve into three stages: initial, intermediate, and late. Change point detection is used to determine the inflection point from the initial to the intermediate stage, and similarly, inflection points showing drastic fluctuations and a significant increase in variance are identified. This allows for the identification of two stage segmentation nodes for each lifespan curve, thus dividing the curve into three segments. Then, linear fitting is performed on the initial stage: data points from all lifespan curves in the initial stage are extracted, and a linear regression algorithm is used to fit and determine the optimal straight line, representing the linear decay relationship in the initial stage, where remaining lifespan decreases linearly over time. Exponential fitting is performed on the intermediate stage: data points from all lifespan curves in the intermediate stage are extracted, and a nonlinear least squares method is used to fit and determine the optimal exponential curve, representing the exponential growth relationship in the intermediate stage, where the rate of health deterioration accelerates. Nonlinear fitting is performed on the late stage: data points from all lifespan curves in the late stage are extracted, and a more complex stochastic process model or machine learning model is used for fitting to determine the nonlinear failure relationship in the late stage. Finally, the stage baseline relationship is added.

[0038] Step S300: Initialize the wear prediction unit based on the stage baseline relationship, perform wear prediction and stage remaining lifetime calculation based on the attention state diagram, and determine the lifetime prediction result, wherein the lifetime prediction result is in the form of a ternary fuzzy number.

[0039] Step S300 further includes step S310, for the attention state map, locating the fluctuation elements and performing uncertainty quantification to determine the confidence interval, wherein the confidence interval corresponds one-to-one with the fluctuation elements; step S320, using the confidence interval, determining the maximum value group, median value group, and minimum value group; step S330, based on the maximum value group, median value group, and minimum value group, performing three-thread prediction based on the attention state map to determine the ternary fuzzy number, wherein the ternary fuzzy number includes the predicted minimum value, the predicted median value, and the predicted maximum value.

[0040] Preferably, the wear prediction unit can predict the future evolution trajectory based on the current health indicator values. It contains a mapping function and initializes the wear prediction unit based on the stage baseline relationship. That is, the specific relationship parameters in the stage baseline relationship are loaded into the wear prediction unit, and it is initialized as a specific prediction model for the current gear type and the current working stage. For example, if the stage discrimination result is mid-term, the wear prediction unit is an exponential predictor. Then, wear prediction based on the attention state map is performed, that is, the attention state map is mapped to the current estimated value of the health indicator using the internal mapping function. Then, the remaining life of the stage is calculated. That is, the wear prediction unit substitutes the estimated value of the health indicator and the current time point into the stage baseline relationship to solve for the time required to reach the failure threshold, which is used as the estimated value of the remaining life.

[0041] Preferably, the uncertainty of lifetime prediction is quantified and the final lifetime prediction result is determined. Specifically, due to sensor noise, model error, and the randomness of future operating conditions, lifetime prediction is uncertain. A lifetime prediction range is generated through three-thread prediction. Then, the uncertainty of the fluctuation elements in the attention state diagram is quantified to generate confidence intervals, where each confidence interval corresponds one-to-one with a fluctuation element. Prediction is then performed using the minimum, median, and maximum sets of health indicator estimates, respectively. This includes substituting the maximum set of health indicator estimates into the stage baseline relationship to calculate the shortest remaining lifetime and determine the predicted minimum; substituting the median set of health indicator estimates into the stage baseline relationship to calculate the remaining lifetime and determine the predicted median; and substituting the minimum set of health indicator estimates into the stage baseline relationship to calculate the longest remaining lifetime and determine the predicted maximum. The predicted minimum, predicted median, and predicted maximum are combined to obtain a ternary fuzzy number, which serves as the final lifetime prediction result.

[0042] Furthermore, step S330 also includes identifying the fluctuation elements in the attention state diagram according to the maximum value group, performing wear prediction and stage remaining life calculation according to the initialized wear prediction unit, and determining the predicted minimum value.

[0043] Preferably, the maximum value group is the upper bound of the confidence interval obtained through uncertainty quantification, representing the combination of input conditions that may lead to the worst health status assessment. The fluctuation elements in the attention state diagram are identified based on the maximum value group, that is, the maximum value group is analyzed to identify the dimension in the attention state diagram that contributes the most to the formation of the most pessimistic assessment result, that is, to locate the most unreliable and most likely sensitive point that may lead to the deterioration of the assessment. Then, wear prediction is performed based on the initialized wear prediction unit, that is, the fluctuation elements are strengthened according to the identification results, and the strengthened attention state diagram that tends to the worst case is input into the mapping function of the wear prediction unit to output the pessimistic health index estimate. Then, the pessimistic health index estimate is substituted into the stage baseline relationship to calculate the stage remaining lifetime to obtain the shortest remaining lifetime prediction value, and finally the predicted minimum value is determined.

[0044] Furthermore, step S300 also includes step S340, integrating the ternary fuzzy number as the lifetime prediction result; step S350, adding the lifetime prediction result into the operation and maintenance database, performing time series prediction updates, and displaying it in a pop-up window on the operation and maintenance platform's display interface.

[0045] Preferably, the ternary fuzzy number is integrated as the lifespan prediction result, which may include the prediction timestamp, device identifier, current health stage, confidence level, and lifespan prediction ternary fuzzy number. The lifespan prediction result is then added to the operation and maintenance database and updated with time-series predictions. This means the lifespan prediction results are arranged according to timestamps, forming a historical prediction history of remaining lifespan changes over time. Operation and maintenance personnel can then use the historical prediction records of a specific device to draw a trend chart of remaining lifespan changes. They can also compare the prediction history with the actual failure time of the device to evaluate and optimize prediction performance. Finally, the lifespan prediction result is displayed in a pop-up window on the operation and maintenance platform, presenting the most essential decision-making information. For example, the alarm title could be: "Warning! Reducer gear X has insufficient remaining lifespan, estimated remaining lifespan 80-160 hours, most likely 120 hours. The device is currently in the middle of accelerated wear; it is recommended to arrange immediate inspection and replacement within 100 operating hours."

[0046] The specific embodiments described above do not constitute a limitation on the scope of protection of this application. Those skilled in the art should understand that various modifications, combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application. In some cases, the actions or steps described in this application can be performed in a different order than that shown in the embodiments and still achieve the desired results. Furthermore, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

Claims

1. A method for predicting the remaining life of gears in a tunnel boring machine cutter reducer based on deep learning, characterized in that, The method includes: Multi-source heterogeneous sensing data is collected from the reducer gears and transmitted to the operation and maintenance platform via interface interaction. A two-stage attention graph is constructed to determine the attention state graph. The platform's embedded fuzzy discrimination threshold is triggered to perform stage discrimination on the upper-level state and the attention state diagram and activate the stage baseline relationship. The life cycle is divided into an initial linear decay relationship, a mid-term exponential growth relationship and a late-term nonlinear failure relationship. The wear prediction unit is initialized based on the stage baseline relationship, and wear prediction and stage remaining lifetime calculation based on the attention state diagram are performed to determine the lifetime prediction result, wherein the lifetime prediction result is in the form of a ternary fuzzy number.

2. The method for predicting the remaining life of the cutting reducer gear of a tunneling machine based on deep learning as described in claim 1, characterized in that, Perform a two-order attention mapping to determine the attention state map, including: The first attention threshold is constructed by taking the information focus and explicit relationship as the first-order attention target; Using implicit relationships and dynamic weights as second-order attention objectives, a second attention threshold is constructed. The attention state map is generated by identifying multi-source heterogeneous data and performing feature verification and homology fitting reconstruction with the first and second attention thresholds.

3. The method for predicting the remaining life of the cutting reducer gear of a tunneling machine based on deep learning as described in claim 2, characterized in that, Multi-source heterogeneous sensing types include at least oil data, vibration signals, temperature signals, and acoustic emission signals; Using synchronization timestamp constraints, perform short-term data acquisition based on multi-source heterogeneous sensor types to determine multi-source heterogeneous data.

4. The method for predicting the remaining life of the cutting reducer gear of a tunneling machine based on deep learning as described in claim 1, characterized in that, Before triggering the platform's embedded fuzzy discrimination threshold, the construction of the fuzzy discrimination threshold includes: An adversarial training approach is adopted, and iterative training is performed under a generation-decision architecture. The generation architecture is used to generate sample superordinate states and sample attention state maps in multi-stage scenarios. The determination architecture performs stage probability determination and performs decision feedback training based on the consistency test between stage probability and stage scenario. The convergent decision architecture, trained to convergence, is used as the fuzzy discrimination threshold and is embedded in the operation and maintenance platform.

5. The method for predicting the remaining life of the cutting reducer gear of a tunneling machine based on deep learning as described in claim 4, characterized in that, Perform phase discrimination and activate phase baseline relationships, including: Connect to the operation and maintenance database and retrieve the upper-level status, wherein the upper-level status is the lifetime prediction result of the upper-level time node; Using the superordinate state and attention state diagram as input and the life cycle stage as output, a stage probability determination based on a fuzzy discrimination threshold is performed to activate the stage baseline relationship.

6. The method for predicting the remaining life of a tunneling machine cutting reducer gear based on deep learning as described in claim 5, characterized in that, Before activating the phase baseline relationship, the construction of the phase baseline relationship includes: Obtain the specifications of the reducer gears and determine the constraints based on the installation scenario; Based on the aforementioned specifications and constraints, operation records are retrieved and life curves are processed to determine multiple life curves. For the multiple lifespan curves, stage segmentation nodes are located based on the curve trends. Linear fitting is performed on the initial stage, exponential fitting on the middle stage, and nonlinear fitting on the later stage, and these are added to the stage baseline relationship.

7. The method for predicting the remaining life of the cutting reducer gear of a tunneling machine based on deep learning as described in claim 6, characterized in that, The stage baseline relationship is activated as any one of the following: linear decay relationship, exponential growth relationship, or nonlinear failure relationship.

8. The method for predicting the remaining life of the cutting reducer gear of a tunneling machine based on deep learning as described in claim 1, characterized in that, Perform wear prediction and stage remaining lifetime calculation based on attention state map, including: For the attention state diagram, the fluctuation elements are located and their uncertainty is quantified to determine the confidence interval, wherein the confidence interval corresponds one-to-one with the fluctuation elements; Based on the confidence interval, determine the maximum value group, median value group, and minimum value group; Based on the maximum, median, and minimum value groups, a three-thread prediction based on the attention state graph is performed to determine a ternary fuzzy number, wherein the ternary fuzzy number includes the predicted minimum, the predicted median, and the predicted maximum.

9. The method for predicting the remaining life of a tunnel boring machine cutting reducer gear based on deep learning as described in claim 8, characterized in that, Based on the maximum value group, the fluctuation elements in the attention state diagram are identified, and wear prediction and stage remaining life calculation are performed based on the initialized wear prediction unit to determine the predicted minimum value.

10. The method for predicting the remaining life of a tunnel boring machine cutting reducer gear based on deep learning as described in claim 8, characterized in that, After determining the ternary fuzzy number, the following is included: The ternary fuzzy numbers are integrated as the lifetime prediction result; The lifespan prediction results are added to the operation and maintenance database, and time-series predictions are updated. A pop-up window is then displayed on the operation and maintenance platform's interface.

Images