An AGV residual life prediction method, system, device and storage medium considering time-varying uncertainty

Abstract

The application relates to the technical field of automatic guided vehicle (AGV) life prediction, and specifically discloses an AGV residual life prediction method, system, device and storage medium considering time-varying uncertainty, which comprises the following steps: acquiring multi-source heterogeneous sensor data of the AGV; extracting shared and private features of the data by using a Dify model to obtain a comprehensive degradation index; introducing a nonlinear diffusion term function in a Wiener process to construct a Wiener process degradation model considering time-varying uncertainty, defining the life and residual service life of a component by a first passage time (FPT) concept, and deducing a residual service life probability density function under the FPT concept; estimating model deterministic parameters, and completing parameter updating in combination with a Bayesian algorithm and a current time degradation state; inputting the comprehensive degradation index into the degradation model after parameter updating, and outputting a residual service life prediction result according to the residual service life probability density function; and the application can improve the adaptability and robustness of the model to the life prediction task under complex working conditions.

Description

Technical Field

[0001] This invention relates to the field of predictive maintenance and health management technology for industrial control equipment, and in particular to a method, system, device, and storage medium for predicting the remaining life of AGVs that takes into account time-varying uncertainties. Background Technology

[0002] With the rapid development of intelligent manufacturing and smart logistics, Automated Guided Vehicles (AGVs) have been widely used in warehousing management, flexible manufacturing, and unmanned delivery, becoming one of the key pieces of equipment for improving automation levels and transportation efficiency. AGVs typically possess the ability to operate continuously around the clock, placing higher demands on their reliability and stability. In AGV systems, bearings, as key fundamental components in drive motors, steering systems, and rolling parts, directly affect the stability and efficiency of the entire vehicle. Because AGVs often operate in complex industrial environments, such as high humidity, high dust, and high load conditions, bearings are more prone to fatigue damage, poor lubrication, and other degradation phenomena, which can lead to abnormal vehicle operation or even shutdown. As a core component of the critical transmission system of AGVs, bearings bear the responsibility of supporting equipment movement and transmitting loads; their health directly affects the safety and efficiency of AGV operation. Once a bearing fails, it can easily cause the entire machine to shut down, resulting in system malfunction and economic losses.

[0003] Currently, statistical modeling methods are widely used in Remaining Useful Life (RUL) prediction, especially modeling frameworks based on stochastic processes, such as Wiener processes and Gamma processes. These frameworks can effectively characterize the randomness and uncertainty in degradation paths and support dynamic reasoning about parameter updates and lifetime distributions. Although existing research has made significant progress in improving the modeling capabilities of Wiener processes, these methods mainly focus on modeling specific stochastic factors and do not consider the parameter uncertainties that change over time during degradation, i.e., time-varying uncertainties. This modeling deficiency may lead to increased bias in RUL estimates in real-world engineering applications, reducing the model's adaptability and generalization ability. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method for predicting the remaining life of AGVs that considers time-varying uncertainties. By introducing the Dify model, the method automatically learns the relationship between multi-source sensor data and time series. At the same time, a nonlinear diffusion modeling mechanism is introduced into the Wiener process modeling to characterize the dynamic changes of the disturbance term during the actual degradation process, thereby improving the model's adaptability and robustness to life prediction tasks under complex working conditions.

[0005] On the one hand, a method for predicting the remaining life of AGVs that considers time-varying uncertainties is provided, including: Acquire multi-source heterogeneous sensor data from AGV; The Dify model is used to extract shared and private features from multi-source heterogeneous sensor data to obtain a comprehensive degradation index; A Wiener process degradation model considering time-varying uncertainties is constructed. Specifically, a nonlinear diffusion term function is introduced into the Wiener process to construct a Wiener process model considering time-varying uncertainties. For the first time, the lifetime and remaining useful life of the component are defined using the concept of time. Based on the drift term function and the diffusion term function, the probability density function of the remaining useful life under the concept of FPT is derived. The deterministic parameters of the model are estimated using the maximum likelihood estimation algorithm, and the parameters are updated by combining the Bayesian algorithm with the degradation state at the current moment. The degradation model, after being updated with the comprehensive degradation index as input parameters, is used to obtain the predicted remaining useful life based on the probability density function of the remaining useful life.

[0006] Preferably, the Dify model adopts a dual-branch spatiotemporal encoder structure, with the first branch being a sensor-private feature encoding branch and the second branch being a cross-sensor shared feature fusion branch.

[0007] Preferably, the first branch extracts private features from multi-source heterogeneous sensor data to obtain private spatiotemporal features corresponding to each sensor; the second branch maps the private spatiotemporal features of different sensors into query vectors, key vectors and value vectors, calculates the dynamic correlation weights between different channels through cross-sensor attention, and performs weighted fusion to obtain shared features characterizing the common degradation trend of bearings.

[0008] Preferably, a nonlinear diffusion term function is introduced into the Wiener process to construct a Wiener process model that considers time-varying uncertainties, specifically as follows:

[0009] in, It is a diffusion term function; This represents the degradation state of the component at monitoring time t. Let be the initial degenerate state of the component. Without loss of generality, assume... ; This is the drift term in the component degradation process, used to describe the overall fluctuation trend during the degradation process, where... The drift coefficient, For the drift term function, .

[0010] Preferably, the deterministic parameters of the model are estimated using the maximum likelihood estimation algorithm, and the parameters are updated by combining the Bayesian algorithm with the current degradation state. Specifically, the deterministic parameters in the model are estimated by maximizing the likelihood function based on the observed degradation data. The random parameters are updated by posterior inference based on the prior distribution of the random parameters obtained by the maximum likelihood estimation algorithm and the current degradation state.

[0011] Preferably, the method of maximizing the likelihood function based on observed degradation data to estimate deterministic parameters in the model involves: first, determining the drift timescale parameter using the least squares method, which controls the rate of change and curvature of the deterministic degradation trend function; second, determining the diffusion timescale parameter and the relative proportion parameter between individual differences and process randomness by maximizing the profile likelihood function based on the drift timescale parameter; third, determining the unknown parameters based on the diffusion timescale parameter and the relative proportion parameter.

[0012] Preferably, the remaining useful life probability density function expression under the FPT concept is as follows: .

[0013] On the other hand, an AGV remaining life prediction system considering time-varying uncertainties is provided, including: The data acquisition module acquires multi-source heterogeneous sensor data from the AGV; The feature extraction module uses the Dify model to extract shared and private features from multi-source heterogeneous sensor data to obtain a comprehensive degradation index. The degradation model construction module constructs a Wiener process degradation model considering time-varying uncertainties. Specifically, it introduces a nonlinear diffusion term function into the Wiener process to construct a Wiener process model considering time-varying uncertainties, thus defining the lifetime and remaining useful life of the component for the first time using the concept of time. Based on the drift term function and the diffusion term function, it derives the probability density function of the remaining useful life under the concept of FPT. It uses the maximum likelihood estimation algorithm to estimate the deterministic parameters of the model, and combines the Bayesian algorithm with the degradation state at the current moment to complete the parameter update. The prediction module takes the comprehensive degradation index as input to the updated degradation model and obtains the predicted remaining useful life based on the probability density function of the remaining useful life.

[0014] Furthermore, an electronic device is also provided, including: A memory for non-transitory storage of computer-readable instructions; and a processor for executing the computer-readable instructions, wherein the computer-readable instructions, when executed by the processor, perform the method described in the first aspect above.

[0015] In another aspect, a storage medium is also provided for non-transitory storage of computer-readable instructions, wherein when the non-transitory computer-readable instructions are executed by a computer, the method described in the first aspect is performed.

[0016] The above technical solution has the following advantages or beneficial effects: This invention introduces the Dify model to automatically learn the relationship between multi-source sensor data and time series, and introduces a nonlinear diffusion modeling mechanism in Wiener process modeling to characterize the dynamic changes of the disturbance term in the real degradation process, thereby improving the model's adaptability and robustness to lifetime prediction tasks under complex working conditions.

[0017] This invention employs a data-driven intelligent fusion model to perform channel-independent encoding and cross-sensor attention fusion of horizontal and vertical vibration signals. This allows for dynamic adjustment of the contribution of information from different sensors, reducing the impact of redundant information and local noise on the characterization of the degradation state. By setting drift time scale and diffusion time scale separately in the Wiener process, the intensity of random fluctuations can change with the running time, thereby more accurately describing the time-varying uncertainty of the AGV bearing degradation process under complex working conditions.

[0018] This invention is based on the first-ever construction of the remaining useful life probability density function (PDF) over time, which can simultaneously output the lifetime point estimate, probability distribution, and confidence interval, providing a basis for risk quantification and predictive maintenance decisions. It employs a two-step parameter estimation method combining least squares estimation and profile likelihood estimation to reduce the computational complexity of solving multiple parameters jointly, and uses Bayesian updates to achieve online correction of model parameters and lifetime prediction results. Attached Figure Description

[0019] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0020] Figure 1 This is a flowchart of the method in Example 1; Figure 2 This is a diagram of the experimental setup for Example 1; Figure 3 This is the time-domain data and root mean square error feature map of the bearing 1_3 bearing dataset from Example 1. Figure 3 (a) in the figure represents the vertical time domain data of bearing 1_3. Figure 3 (b) in the figure represents the horizontal time-domain data of bearing 1_3. Figure 3 (c) in the figure represents the root mean square error value of the vertical bearing 1_3. Figure 3In the figure, (d) represents the root mean square error value of bearing 1_3. Figure 4 This is the time-domain data and root mean square error feature map of the bearing 1_5 bearing dataset from Example 1. Figure 4 (a) in the figure represents the vertical time domain data of bearing 1_5. Figure 4 (b) in the figure represents the horizontal time-domain data of bearing 1_5. Figure 4 (c) in the figure represents the root mean square error value of the vertical bearing 1_5. Figure 4 In the figure, (d) represents the root mean square error value of bearing 1_5. Figure 5 This is a fusion data diagram of bearings 1 and 3 in Example 1; Figure 6 This is a fusion data diagram of bearing 1_5 in Example 1; Figure 7 The graph shows the probability density function of the remaining service life of bearings 1 and 3 in Example 1 at different monitoring times. Figure 8 This is a probability density function graph of the remaining service life of bearing 1_5 in Example 1 at different monitoring times. Detailed Implementation

[0021] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, 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.

[0022] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments of the invention. The terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0023] In this embodiment of the invention, "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Furthermore, in the description of this invention, "multiple" refers to two or more.

[0024] Furthermore, to facilitate a clear description of the technical solutions of the embodiments of the present invention, the terms "first" and "second" are used in the embodiments of the present invention to distinguish identical or similar items with essentially the same function and effect. Those skilled in the art will understand that the terms "first" and "second" do not limit the quantity or execution order, and the terms "first" and "second" are not necessarily different.

[0025] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0026] Example 1 Since the health of bearings is directly related to the safety and efficiency of AGV operation, accurately predicting the remaining useful life (RUL) of bearings and implementing predictive maintenance before failure becomes a prerequisite for ensuring the efficient operation of AGVs and extending equipment life.

[0027] Currently, prediction methods for rolling bearing range under constant wear (RUL) mainly include data-driven methods, physical model methods, and hybrid methods. In recent years, with the continuous advancement of sensor technology and industrial data acquisition methods, data-driven methods have become a research hotspot for RUL prediction. These methods construct RUL prediction models by mining the relationship between historical operating data and degradation characteristics, offering advantages such as flexible modeling, no need for complex physical prior information, and strong adaptability. Among them, deep learning methods, relying on their powerful nonlinear modeling capabilities, can automatically extract degradation features from high-dimensional, multi-source data, exhibiting excellent feature learning and prediction capabilities, especially in multi-sensor data fusion scenarios. However, deep learning models typically rely on a large number of training samples and suffer from insufficient interpretability and limited generalization ability, making them difficult to adapt to individual equipment differences. To improve the interpretability and reliability of models, statistical modeling methods are also widely used in RUL prediction, especially modeling frameworks based on stochastic processes, such as Wiener processes and Gamma processes, which can effectively characterize the randomness and uncertainty in degradation paths and support dynamic inference of parameter updates and lifetime distributions. Statistical models offer significant advantages in handling small sample data and modeling degradation mechanisms, making them suitable for online condition assessment and lifetime prediction of individual devices. However, single statistical modeling often struggles to fully utilize the deep feature information contained in multidimensional sensor data. Therefore, methods integrating deep learning and statistical modeling have gained increasing attention, aiming to combine the data representation capabilities of deep models with the interpretability and uncertainty modeling capabilities of statistical models to achieve accurate modeling of device degradation states and dynamic, reliable prediction of near-term downtime (RUL).

[0028] While existing research has made significant progress in improving the modeling capabilities of Wiener processes, current methods primarily focus on modeling specific stochastic factors and do not consider the time-varying uncertainties of parameters during degradation, i.e., time-varying uncertainties. This modeling deficiency may lead to increased bias in RUL estimates in real-world engineering applications, reducing the model's adaptability and generalization ability.

[0029] To improve prediction accuracy, increasing research focuses on effectively fusing multi-dimensional degradation information. Compared to traditional lifetime prediction methods relying on single sensors or feature signals, multi-source data fusion methods can more comprehensively capture the evolution of equipment health status, significantly improving the accuracy and robustness of Remaining Usage (RUL) prediction. Furthermore, multi-dimensional sensor data fusion is an important method for improving the accuracy of RUL prediction. Although methods such as weighted averaging, information contribution rate allocation, and multi-index evaluation have achieved effective fusion of multi-sensor data and improved lifetime prediction accuracy to some extent, these methods mostly rely on manually set weighting mechanisms or evaluation indicators, exhibiting subjectivity and limitations, and failing to fully explore the nonlinear relationships and deep features between multi-source data. In addition, traditional weighted fusion methods struggle to achieve robust dynamic modeling when faced with noise, inconsistencies, or dynamic feature evolution among sensor data. Therefore, to improve the intelligence and adaptability of multi-dimensional degradation information fusion, it is necessary to introduce deep learning methods with strong feature learning and high-dimensional information abstraction capabilities to automatically capture the dependencies between multi-source data and time-series data, achieving data-driven, efficient fusion modeling.

[0030] In summary, constructing a multi-sensor data fusion method can help to more comprehensively capture equipment degradation characteristics and improve the reliability and practicality of remaining life prediction. However, existing research still has the following limitations: (1) Multi-sensor data fusion methods determine the fusion weight by simply weighting the average or the information contribution rate, without considering the interaction between sensor data, and it is difficult to fully reflect the dynamic contribution of each sensor in the degradation process; (2) Most studies based on Wiener process model are based on the assumption of linear Brownian motion, ignoring the characteristic that local fluctuations under actual working conditions may change nonlinearly with the working environment.

[0031] See Figure 1 The flowchart illustrates a method for predicting the remaining life of an AGV that takes into account time-varying uncertainties according to an embodiment of the present disclosure: S1: Acquire multi-source heterogeneous sensor data of AGV.

[0032] Horizontal and vertical acceleration vibration sensors are installed on the bearing housing to collect multi-source heterogeneous sensor data in real time. The multi-source heterogeneous sensor data includes horizontal and vertical vibration acceleration data installed in different directions on the key bearing housing of the AGV, as well as the corresponding acquisition timestamps and monitoring cycle numbers.

[0033] S2: Use the Dify model to extract the shared and private features of multi-source heterogeneous sensor data to obtain a comprehensive degradation index.

[0034] In multi-source sensor monitoring scenarios for AGV bearings, traditional weighted fusion methods struggle to dynamically capture the nonlinear correlations and temporal dependencies between sensors. To address this, this invention proposes a data-driven intelligent fusion model (Dify model) to achieve adaptive fusion and degradation feature representation learning of multi-source heterogeneous sensor data.

[0035] The core of the Dify model is to extract shared and private features from multi-source heterogeneous sensor data through a deep temporal encoder, and dynamically learn the correlation weights between sensor data based on an attention mechanism, ultimately generating a one-dimensional comprehensive degradation index as input to the subsequent Wiener process degradation model.

[0036] The Dify model has two branches: the first branch is for encoding sensor-specific features, and the second branch is for fusing shared features across sensors.

[0037] For horizontal and vertical vibration signals, the private features of multi-source heterogeneous sensor data are extracted through the first branch to obtain the private spatiotemporal features corresponding to each sensor. The first branch is equipped with a one-dimensional convolutional network and a gated recurrent unit with independent parameters. The one-dimensional convolutional network is used to extract local degradation features such as impact and periodic fluctuations, and the gated recurrent unit is used to extract the time dependence between different monitoring periods, thereby obtaining the private spatiotemporal features corresponding to each sensor.

[0038] The second branch is used to map the private spatiotemporal features of different sensors into query vectors, key vectors, and value vectors. It calculates the dynamic correlation weights between different channels through cross-sensor attention and performs weighted fusion to obtain shared features that characterize the common degradation trend of bearings.

[0039] The Dify model employs a dual-branch spatiotemporal encoder structure, including a sensor-specific feature encoding branch and a cross-sensor shared feature fusion branch. The sensor-specific feature encoding branch uses 1D convolutional layers (Conv1D) and gated recurrent units (GRUs) to extract local features and temporal dependencies. (1) in, Represented as the first Within the monitoring period, the first Vibration sequences of a sensor after preprocessing; Represented as the first The sensor at the first Private spatiotemporal feature vectors corresponding to each monitoring cycle; ,in, This refers to the number of sensor channels; ,in, Total number of monitoring cycles; This represents a one-dimensional convolutional feature extraction network corresponding to different sensors; This indicates the gated loop unit corresponding to different sensors.

[0040] Map the private features of each sensor to query vectors. Key vector Sum value vector :

[0041]

[0042]

[0043] in, This indicates the sensor channel index that participates in cross-sensor feature fusion.

[0044] Next, the Softmax function is used to process the query vector. With key vector Normalize the dot product result to obtain the dynamic correlation weight: (5) Based on the dynamic association weights, the value vector By weighting, the shared fusion characteristics are obtained. : (6) (7) Finally, the fusion features A one-dimensional comprehensive degradation index is obtained by using linear mapping and the Sigmoid activation function: (8) Furthermore, to meet the requirements of monotonicity and trend in degradation modeling, a regularization term is added to the loss function: (9) in, The coefficient representing the monotonicity constraint of the degradation trend; This represents the smoothness constraint coefficient of the degradation curve; and These represent two adjacent monitoring times; This represents the second difference of the comprehensive degradation index sequence; This indicates the degree of local fluctuation in the degradation curve.

[0045] The total loss function is the sum of the mean squared error (MSE) and the regularization term: (10) It should be noted that the one-dimensional comprehensive degradation index output by the Dify model is defined as the degradation state observation value of the Wiener process degradation model at discrete monitoring times. This observation value corresponds to the Wiener process degradation model. At discrete monitoring time The observed values ​​of the degradation state, This is used for the subsequent derivation and prediction of the probability density function of RUL.

[0046] S3: Constructing a Wiener process degradation model considering time-varying uncertainties: Introducing a nonlinear diffusion term function into the Wiener process to construct a Wiener process model considering time-varying uncertainties; defining the lifetime and remaining useful life of components for the first time using the concept of time; deriving the probability density function of remaining useful life under the concept of FPT based on the drift term function and the diffusion term function; estimating the deterministic parameters of the model using the maximum likelihood estimation algorithm, and updating the parameters by combining the Bayesian algorithm with the degradation state at the current moment.

[0047] S3.1 introduces a nonlinear diffusion term function into the Wiener process to construct a Wiener process model that considers time-varying uncertainties, thereby defining the life and remaining useful life of a component for the first time through the concept of time.

[0048] Common Wiener process models: (11) in, This represents the degradation state of the component at monitoring time t. Let be the initial degenerate state of the component. Without loss of generality, assume... . This is the drift term in the component degradation process, used to describe the overall fluctuation trend during the degradation process, where... The drift coefficient, For the drift term function, when When, formula (11) is a linear Wiener process model, when When the function is nonlinear, formula (11) is a nonlinear Wiener process model. The diffusion term describes the instantaneous fluctuations during the degradation process, where... The diffusion coefficient is... This is standard Brownian motion.

[0049] In traditional Wiener process models, the diffusion term is typically set as standard Brownian motion, representing a fixed level of uncertainty in the degradation process. However, in real-world engineering scenarios, such as AGV bearings operating under varying loads and complex environments, their degradation rate and uncertainty often exhibit nonlinear characteristics as time or degradation level changes.

[0050] To more realistically describe the uncertain evolution of the degradation process, a nonlinear diffusion term function is introduced. Specifically, a nonlinear function is used in the diffusion term function, as shown in the following formula: (12) in, This is the diffusion term function.

[0051] Existing Wiener process models are all limiting cases of this model: 1) When and Then, formula (11) will become a linear Wiener process model: (13) 2) When At that time, formula (11) will become a nonlinear Wiener process model with the same time scale: (14) 3) When Then, formula (11) will become a nonlinear Wiener process model: (15) 4) When When the time is uncertain, formula (11) will become the Wiener process model that takes into account time uncertainty.

[0052] In degradation modeling, component failure is typically defined as the moment when a condition monitoring index first exceeds a certain set threshold. Therefore, this invention uses the first-pass time as the core concept for defining lifetime, as shown in the following formula: (16) in: To pre-set thresholds, The lifespan of the component; This represents the initial degradation state of the bearing.

[0053] For the remaining service life of a component, it is defined as the current monitoring time of the equipment. until the time of failure The time difference, i.e. Based on the initial definition of the remaining service life of a component using the concept of time, the remaining service life of the component is the time from the current state monitoring moment to the component failure moment, as shown in the following formula: (17) in, For at any time The remaining useful life under certain conditions is a random variable. This represents one possible value of the random variable representing the remaining useful life. Indicates future time The state of degradation; This indicates the degradation status obtained at the current monitoring moment.

[0054] S3.2 derives the probability density function of remaining useful life under the FPT concept.

[0055] Because the degradation model constructed in this invention contains nonlinear drift terms and nonlinear time-transformed Brownian motion, its degradation process lacks an analytical state distribution function, which further makes it difficult to explicitly solve the probability density function of RUL.

[0056] Therefore, this invention will employ a time scale transformation method to simplify the modeling and RUL prediction process.

[0057] First, variable substitution was performed on formula (11) to extract the nonlinear time transformation from Brownian motion and convert it into the standard Brownian motion form. Specifically, it is assumed that... ,So Formula (11) can be transformed into: (18) For ease of calculation, formula (17) is transformed as follows: (19) in, .

[0058] For formula (18), the problem is transformed into a time-based problem involving the variable s for the first time. The FPT formulas for lifetime and remaining useful life are as follows: (20) (twenty one) Lemma 1: For the model In other words, under the FPT concept, reaching the set threshold The approximate analytical expression for the probability density function can be expressed as: (twenty two) For the first pass time of the degradation process in formula (18), considering it as a process driven by standard Brownian motion, based on Lemma 1, we can obtain an approximate analytical expression for the probability density function of the lifetime under the FPT concept in formula (18), as follows: (twenty three) Will Substituting back into formula (22), we get: (twenty four) Formula (23) can be further transformed into: (25) To further demonstrate the flexibility of the proposed model, several typical forms of the drift term function and the diffusion term function are considered. By substituting these specific functions into equation (24), the probability density function expression for lifetime under the FPT concept can be derived, which is equivalent to the device remaining lifetime distribution.

[0059] (1) When both the drift term function and the diffusion term function are power functions, i.e. At that time, the first Wiener process model was obtained: (26) Substituting into formula (24), we can obtain the approximate analytical expression of the probability density function of the first lifetime as follows: (27) When a and b are 1, formula (25) is a common linear Wiener process model, and formula (26) follows an inverse Gaussian distribution, proving that the derivation of the proposed model is correct; when b is 1, formula (25) is a common nonlinear Wiener process model.

[0060] (2) When both the drift term function and the diffusion term function are exponential functions, i.e. The second Wiener process model is obtained: (28) The approximate analytical expression for the probability density function of the second lifetime, based on the second Wiener process model, is as follows: (29) (3) When the drift term function is a power function and the diffusion term function is an exponential function, i.e. The third Wiener process model is obtained: (30) The approximate analytical expression for the probability density function of the third lifetime, based on the second Wiener process model, is as follows: (31) In practical industrial applications, components of the same type operating under the same conditions often exhibit different degradation trajectories due to inherent manufacturing, material, and assembly differences. Therefore, this invention considers the drift coefficient as a variable following a normal distribution, i.e. .

[0061] When the drift coefficient is a single variable, according to Bayes' law of total probability and Lemma 2, the approximate analytical expression of the probability density function of the lifetime of the proposed model considering heterogeneity can be obtained as follows: (32) Lemma 2: For ease of calculation, we can obtain the following conclusion: (33) In practical prediction and health management, the goal is not merely to estimate the lifespan of a component, but to assess its remaining useful life (RUL) based on its current degradation state. Having obtained the distribution of lifespan, it is necessary to derive the expression for RUL under the concept of FPT (Factor-Preference Time). Similarly, the probability density function expression for RUL under the FPT concept is as follows: (34) S3.3 uses the maximum likelihood estimation algorithm to estimate the deterministic parameters of the model, and combines the Bayesian algorithm with the current degenerate state to complete the parameter update.

[0062] Accurate estimation of unknown parameters in the model is crucial for the reliability of the prediction model and the accuracy of the remaining useful life prediction. In the proposed time-varying Wiener process model framework, the parameters that need to be estimated include the drift coefficient, the diffusion coefficient, and parameters related to the time transformation function. These parameters control the stochastic process model and directly affect the remaining useful life prediction results.

[0063] To this end, this invention employs two complementary estimation methods: maximum likelihood estimation and Bayesian algorithm. First, the deterministic parameters in the model are estimated by maximizing the likelihood function based on observed degradation data. Then, based on the prior distribution of the random parameters obtained from the maximum likelihood estimation algorithm, and combined with the current degradation state, the random parameters are updated through posterior inference.

[0064] Specifically, deterministic parameters in the model are estimated by maximizing the likelihood function based on observed degradation data. First, the drift timescale parameter is determined using the least squares method; this parameter controls the rate of change and curvature of the deterministic degradation trend function. Based on the drift timescale parameter, the diffusion timescale parameter and the relative proportion parameter between individual differences and process randomness are determined by maximizing the profile likelihood function. The diffusion timescale parameter controls the cumulative rate of the random fluctuation timescale function over time. Finally, the unknown parameters are determined based on the diffusion timescale parameter and the relative proportion parameter.

[0065] Assuming the degradation process of a single component occurs at different monitoring times... Observations were conducted, and the corresponding degradation status monitoring values ​​were recorded as follows: ,in, Let be the transpose of the matrix, and n be the number of observations performed on that cell. For computational convenience, define... , Based on the properties of independent Brownian motion increments and the normal distribution of increments, the degradation state monitoring values ​​of the component can be obtained. It follows a multivariate normal distribution. Ignoring heterogeneity, its mean and variance are as follows: (35) in: .

[0066] When considering the uncertainty of the drift coefficient, the mean and variance of the degradation state of the component are as follows: (36) in: .

[0067] Therefore, based on the observed degenerative state To estimate unknown parameters By maximizing The likelihood function is used to estimate it, and its expression is as follows: (37) To simplify the calculation process, the likelihood function is taken in logarithmic form, which is then transformed into a linear form, as shown in the following formula: (38) in, The drift timescale parameter is used to control the deterministic degradation trend function. The rate of change and curvature; The diffusion timescale parameter is used to control the diffusion timescale function. The cumulative rate over time; Random drift coefficient The mean value reflects the average degradation rate of different objects; The variance of the drift coefficient reflects the degradation differences between different individual bearings; The variance of the diffusion term reflects the intensity of random fluctuations during the degradation process.

[0068] Because the formula contains It is difficult to estimate these five unknown parameters by directly maximizing the log-likelihood function. Therefore, we transform formula (37), extract the diffusion term coefficient, and define... Therefore, we can obtain the following formula: (39) Formula (38) for unknown parameters and Differentiation yields: (40) Setting formula (39) to 0, we can obtain the following formula: (41) Substituting equation (40) back into equation (38), we can obtain the contour likelihood optimization function. As shown in the following formula: (42) Since there are three unknown parameters in the profile likelihood function, direct optimization would incur a large computational burden. Therefore, a two-step estimation method is adopted. Specifically, parameter 'a' determines the deterministic trend of the degradation trajectory. First, the drift time scale parameter is determined using the least squares method. This parameter controls the deterministic degradation trend function. The rate of change and curvature. Then, based on the drift timescale parameter, the diffusion timescale parameter and the relative proportion parameter between individual variability and process randomness are determined by maximizing the profile likelihood function. The diffusion timescale parameter is used to control the random fluctuation timescale function. The cumulative rate over time. Finally, substituting the diffusion timescale parameter and the relative proportion parameter back into equation (40), we can obtain the parameter estimation results. .

[0069] S4: Input the comprehensive degradation index into the updated degradation model and obtain the predicted remaining useful life based on the remaining useful life probability density function.

[0070] S4.1 in Each monitoring cycle utilizes the Dify model to obtain the current overall degradation status: .

[0071] S4.2 Determine whether the current degradation state has reached the failure threshold. ; like If the bearing fails, its remaining service life is set to 0. like If so, then lifespan prediction will continue.

[0072] S4.3 combines the offline estimated model parameters with the degraded observation sequence up to the current time, and updates the posterior distribution of the drift parameters using a Bayesian method: (43) in, This represents the set of model parameters.

[0073] S4.4 Substituting the current degradation state, monitoring time, failure threshold, drift timescale function, diffusion timescale function, and updated parameters into the remaining lifetime probability density function, we obtain: (44) S4.5 normalizes the remaining useful life probability density and calculates the lifetime point estimate: (45) S4.6 calculates the lifetime prediction interval at the corresponding confidence level based on the cumulative distribution function of the probability density, and outputs the remaining lifetime point estimate, probability density curve and confidence interval.

[0074] S4.7 Once new monitoring data is obtained, repeat the above process to achieve online updates of the remaining service life.

[0075] The following is the experimental procedure: Due to the complex operating environment of actual automated guided vehicles (AGVs), fault data is scarce and difficult to obtain, especially data covering the complete degradation process from healthy state to failure. Therefore, to verify the effectiveness of the proposed remaining life prediction method, bearings, a typical key component in AGVs, were selected as the research object. As an important rotating component in the AGV drive system, the health status of bearings directly affects the operational reliability and safety of the AGV. Furthermore, their degradation process exhibits typical randomness and observability, making them suitable as a subject for life prediction. To overcome the difficulty in obtaining full-life degradation data in actual AGVs, the publicly available Xi'an Jiaotong University bearing full-life degradation dataset was used as the experimental data source. Rolling bearings on an accelerated degradation test platform were collected using vibration acceleration sensors. This dataset records the complete process of bearings from initial healthy state to failure, possessing completeness, stability, and high resolution, and is widely used in remaining life prediction research.

[0076] The SJTU-SY dataset contains three different operating condition datasets, each operating under constant speed and load. In the experiment, horizontal and vertical acceleration vibration sensors were arranged on the bearing housing. The experimental setup is as follows: Figure 2 As shown in the figure. Vibration signals were acquired every minute for 1.28 seconds, with a sampling frequency of 25.6 kHz. The vertical vibration signals from the bearing datasets 1_3 and 1_5 under operating condition 1 were used for the experiment. The 1_3 dataset ran for 158 minutes, while the 1_5 dataset ran for 52 minutes. The time-domain plot of the vibration signal from the 1_3 dataset is shown in the figure. Figure 3 As shown (length 10), Figure 3 (a) in the figure represents the vertical time domain data of bearing 1_3. Figure 3 (b) in the figure represents the horizontal time-domain data of bearing 1_3. Figure 3 (c) in the figure represents the root mean square error value of the vertical bearing 1_3. Figure 3 In the figure, (d) represents the root mean square error value of bearing 1_3; the time domain plot of the vibration signal of bearing 1_5 dataset is shown below. Figure 4 As shown, Figure 4 (a) in the figure represents the vertical time domain data of bearing 1_5. Figure 4 (b) in the figure represents the horizontal time-domain data of bearing 1_5. Figure 4 (c) in the figure represents the root mean square error value of the vertical bearing 1_5. Figure 4 In the figure, (d) represents the root mean square error of bearing 1_5.

[0077] Figure 5 and Figure 6 The figures show the trend curves of the comprehensive degradation index constructed by fusing multi-source sensor data using the Dify model on the Bearing 1_3 and Bearing 1_5 datasets, reflecting the overall evolution law of the bearing degradation process. The horizontal axis represents the monitoring period, indicating the discrete sequence of equipment operating time; the vertical axis represents the comprehensive degradation index, indicating the fused comprehensive degradation index. The green curve represents the fusion result, i.e., the comprehensive degradation index formed after modeling, compressing, and reconstructing multiple single sensor information based on the Dify model. In both datasets, the fused degradation state curves show good monotonicity and stability, especially as the bearing enters the later stage of failure, where the comprehensive degradation index rises sharply. In summary, the application of the Dify model in degradation modeling demonstrates excellent feature extraction and fusion capabilities, effectively suppressing redundancy and noise between sensors, enhancing the characterization performance of the degradation state index, providing a stable and physically meaningful input basis for subsequent remaining service life prediction, and significantly improving the accuracy and robustness of the prediction.

[0078] Furthermore, this invention selects the power function model as the basic form for degradation path modeling, mainly based on its good adaptability in characterizing the slow degradation process of equipment. The power function model has the advantages of simple structure and strong parameter interpretability, and can effectively capture the nonlinear trend of degradation rate changing with time, making it particularly suitable for modeling the degradation behavior of mechanical components such as bearings during long-term service. Figure 7 and Figure 8 The figure illustrates the probability density function (PDF) of the remaining useful life (RUL) of bearings 1_3 and 1_5 at different monitoring times. The horizontal axis represents the predicted RUL, the vertical axis represents the current monitoring time, and the vertical column represents the probability density of the RUL at the current monitoring time. The figure shows that as the monitoring cycle progresses (i.e., the equipment's continuous operating time increases), the PDF curve gradually shifts to the left and converges, consistent with the bearings approaching failure. This indicates that the model can dynamically update its RUL prediction. Simultaneously, the estimated value gradually approaches the actual RUL, especially in the near-failure stage, where the model prediction is more accurate, indicating higher reliability in later prediction stages. Furthermore, the PDF curve is relatively broad in the early monitoring stages, reflecting significant prediction uncertainty in the initial phase; however, it becomes sharper before near failure, indicating that the model gradually accumulates more degradation information, making the prediction more focused and reliable. Overall, the results in the figure demonstrate that the proposed method can effectively provide RUL probability distribution predictions at different monitoring times, exhibiting good time-varying dynamic modeling capabilities and uncertainty characterization capabilities, validating the applicability and effectiveness of the proposed model in RUL prediction.

[0079] To more clearly demonstrate the advantages of the proposed method, the performance of common linear Wiener process models and nonlinear Wiener process models in remaining lifetime prediction under single-sensor data is compared, as shown in Table 1. The proposed method is denoted by M1, the linear Wiener process model for modeling horizontal data is denoted by M2, and the nonlinear power-function Wiener process model for modeling horizontal data is denoted by M3. The differences in model prediction accuracy are systematically evaluated using root mean square error (RMSE) and mean absolute error (MAE) as evaluation metrics, as shown in the following formula: (46) (47) Table 1 Comparison of prediction results from different methods

[0080] Table 1 compares the prediction performance of different methods on the Bearing1_3 and Bearing1_5 datasets, using root mean square error (RMSE) and mean absolute error (MAE) as evaluation metrics. The results clearly show that the M1 method has significantly lower RMSE and MAE than the M2 and M3 methods on both datasets. For example, in the Bearing1_3 dataset, the RMSE and MAE of M1 are 3.3818 and 3.2553, respectively, far lower than M2's 106.9082 and 95.4031, and M3's 46.4187 and 51.4187. Similarly, in the Bearing1_5 dataset, the RMSE and MAE of M1 are 5.1613 and 5.0434, respectively, also significantly better than M2 and M3. This result demonstrates that the proposed modeling method not only possesses higher prediction accuracy but also stronger robustness and generalization ability. By appropriately introducing a nonlinear modeling framework and optimization strategy, the M1 method can more effectively characterize the dynamic characteristics of bearing degradation and improve the accuracy of remaining service life prediction.

[0081] This invention first uses the Dify model to map multidimensional sensor data into 1D data to characterize the true degradation state of bearings during the degradation process. Then, a nonlinear Brownian process is introduced into the common nonlinear Wiener process model to capture the degradation process of bearings under complex operating environments. By using a time-scale variation method, the proposed model is approximated as a common Wiener process model. The PDF of remaining service life under the concept of time is derived for the first time and extended to consider the variability between units. Finally, the effectiveness of the proposed method is verified using a publicly available dataset of rolling bearings.

[0082] Example 2 This embodiment provides an AGV remaining life prediction system that considers time-varying uncertainties, including: The data acquisition module acquires multi-source heterogeneous sensor data from the AGV; The feature extraction module uses the Dify model to extract shared and private features from multi-source heterogeneous sensor data to obtain a comprehensive degradation index. The degradation model construction module constructs a Wiener process degradation model considering time-varying uncertainties. Specifically, it introduces a nonlinear diffusion term function into the Wiener process to construct a Wiener process model considering time-varying uncertainties; it defines the lifetime and remaining useful life of a component for the first time using the concept of time; based on the drift term function and the diffusion term function, it derives the probability density function of the remaining useful life under the concept of FPT; it uses the maximum likelihood estimation algorithm to estimate the deterministic parameters of the model, and combines the Bayesian algorithm with the degradation state at the current moment to complete the parameter update. The prediction module takes the comprehensive degradation index as input to the updated degradation model and obtains the predicted remaining useful life based on the probability density function of the remaining useful life.

[0083] Example 3 This embodiment also provides an electronic device, including: one or more processors, one or more memories, and one or more computer programs; wherein, the processor is connected to the memory, and the one or more computer programs are stored in the memory. When the electronic device is running, the processor executes the one or more computer programs stored in the memory to cause the electronic device to perform the method described in Embodiment 1.

[0084] Example 4 This embodiment also provides a computer-readable storage medium for storing computer instructions, which, when executed by a processor, complete the method described in Embodiment 1.

[0085] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting the remaining life of an AGV considering time-varying uncertainties, characterized in that, include: Acquire multi-source heterogeneous sensor data from AGV; The Dify model is used to extract shared and private features from multi-source heterogeneous sensor data to obtain a comprehensive degradation index; A Wiener process degradation model considering time-varying uncertainties is constructed. Specifically, a nonlinear diffusion term function is introduced into the Wiener process to construct a Wiener process model considering time-varying uncertainties. For the first time, the lifetime and remaining useful life of the component are defined using the concept of time. Based on the drift term function and the diffusion term function, the probability density function of the remaining useful life under the concept of FPT is derived. The deterministic parameters of the model are estimated using the maximum likelihood estimation algorithm, and the parameters are updated by combining the Bayesian algorithm with the degradation state at the current moment. The degradation model, after being updated with the comprehensive degradation index as input parameters, is used to obtain the predicted remaining useful life based on the probability density function of the remaining useful life.

2. The method for predicting the remaining life of an AGV considering time-varying uncertainties according to claim 1, characterized in that, The Dify model adopts a dual-branch spatiotemporal encoder structure. The first branch is a sensor-private feature encoding branch, and the second branch is a cross-sensor shared feature fusion branch.

3. The method for predicting the remaining life of an AGV considering time-varying uncertainties according to claim 2, characterized in that, The first branch extracts private features from multi-source heterogeneous sensor data to obtain private spatiotemporal features corresponding to each sensor; the second branch maps the private spatiotemporal features of different sensors into query vectors, key vectors and value vectors, calculates the dynamic correlation weights between different channels through cross-sensor attention, and performs weighted fusion to obtain shared features characterizing the common degradation trend of bearings.

4. The method for predicting the remaining life of an AGV considering time-varying uncertainties according to claim 1, characterized in that, A nonlinear diffusion term function is introduced into the Wiener process to construct a Wiener process model that considers time-varying uncertainties, specifically as follows: in, It is a diffusion term function; This represents the degradation state of the component at monitoring time t. Let be the initial degenerate state of the component. Without loss of generality, assume... ; This is the drift term in the component degradation process, used to describe the overall fluctuation trend during the degradation process, where... The drift coefficient, For the drift term function, .

5. The method for predicting the remaining life of an AGV considering time-varying uncertainties according to claim 1, characterized in that, The maximum likelihood estimation algorithm is used to estimate the deterministic parameters of the model, and the parameters are updated by combining the Bayesian algorithm with the current degradation state. Specifically, the deterministic parameters in the model are estimated by maximizing the likelihood function based on the observed degradation data. The random parameters are updated by posterior inference based on the prior distribution of the random parameters obtained by the maximum likelihood estimation algorithm and the current degradation state.

6. The method for predicting the remaining life of an AGV considering time-varying uncertainties according to claim 5, characterized in that, The method estimates deterministic parameters in the model by maximizing the likelihood function based on observed degradation data. Specifically, the method involves: first, determining the drift timescale parameter using the least squares method, which controls the rate of change and curvature of the deterministic degradation trend function; second, determining the diffusion timescale parameter and the relative proportion parameter between individual differences and process randomness by maximizing the profile likelihood function based on the drift timescale parameter; third, determining the unknown parameters based on the diffusion timescale parameter and the relative proportion parameter.

7. The method for predicting the remaining life of an AGV considering time-varying uncertainties according to claim 1, characterized in that, The remaining useful life probability density function expression under the FPT concept is as follows: 。 8. An AGV remaining life prediction system considering time-varying uncertainties, characterized in that, include: The data acquisition module acquires multi-source heterogeneous sensor data from the AGV; The feature extraction module uses the Dify model to extract shared and private features from multi-source heterogeneous sensor data to obtain a comprehensive degradation index. The degradation model construction module constructs a Wiener process degradation model considering time-varying uncertainties. Specifically, it introduces a nonlinear diffusion term function into the Wiener process to construct a Wiener process model considering time-varying uncertainties; it defines the lifetime and remaining useful life of a component for the first time using the concept of time; based on the drift term function and the diffusion term function, it derives the probability density function of the remaining useful life under the concept of FPT; it uses the maximum likelihood estimation algorithm to estimate the deterministic parameters of the model, and combines the Bayesian algorithm with the degradation state at the current moment to complete the parameter update. The prediction module takes the comprehensive degradation index as input to the updated degradation model and obtains the predicted remaining useful life based on the probability density function of the remaining useful life.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the method according to any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, it implements the steps of the method described in any one of claims 1-7.

Images