Coal mill core component fault prediction method based on vibration analysis

Abstract

The application discloses a coal mill core component fault prediction method based on vibration analysis and belongs to the technical field of rotary machinery state monitoring and predictive maintenance, which comprises collecting multi-point vibration signals to generate an optimized data set; performing variational modal decomposition and deep embedding to extract a high-level feature group; establishing a transfer learning model to generate a preliminary fault prediction distribution; performing model self-optimization through a reinforcement learning feedback loop to obtain refined prediction results; and performing multi-dimensional risk quantification and fault evolution path simulation to output a decision report. The application adopts a multi-level, self-evolution closed-loop intelligent prediction architecture, can perform online optimization of the model through a reinforcement learning agent based on the uncertainty of the prediction distribution, realizes rapid self-adaptation to a variable working condition environment, provides prospective maintenance decision support in combination with evolution path simulation, and significantly improves the capturing capability of early weak faults of the coal mill, the accuracy of cross-working condition prediction and the scientificity of final maintenance decisions.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of rotating machinery condition monitoring and predictive maintenance, and particularly relates to a coal mill core component fault prediction method based on vibration analysis. BACKGROUND

[0002] The coal mill is a key core equipment in the heavy industry fields such as thermal power generation, cement building materials and chemical industry, and the stability and reliability of its operation state are directly related to the safety and economic benefits of the entire production line. In order to ensure the safe and efficient operation of the coal mill, equipment condition monitoring and predictive maintenance based on vibration and other sensor data have become the main technical path in the industry. By collecting and analyzing the operation data of the core components of the coal mill, the real-time monitoring of the health state of the equipment can be realized to a certain extent, which is of great significance to ensure the continuity of industrial production.

[0003] However, the existing coal mill fault prediction methods still have universal technical defects. The traditional monitoring method based on fixed threshold can only realize alarm after the fault occurs, and it is difficult to effectively warn potential faults. Some diagnosis methods excessively rely on the personal experience of operation and maintenance personnel, and have the problems of strong subjectivity and insufficient reliability. The existing data-driven prediction models have limited ability to capture early weak fault features of the equipment, and have poor adaptability under variable working conditions, which is easy to produce false alarm or miss alarm. These problems restrict the implementation effect of the predictive maintenance strategy and affect the operation efficiency and safety of the equipment. SUMMARY

[0004] To solve the above problems, the present application provides a coal mill core component fault prediction method based on vibration analysis, which adopts a multi-level, self-evolving closed-loop intelligent prediction architecture. Based on the uncertainty of the prediction distribution, the model is optimized online by the reinforcement learning agent, which realizes the rapid self-adaptation to the variable working condition environment, and provides forward-looking maintenance decision support combined with the evolution path simulation, which significantly improves the ability to capture early weak faults of the coal mill, the accuracy of cross-condition prediction and the scientificity of the final maintenance decision.

[0005] The above object can be achieved by the following scheme:

[0006] The method for predicting faults of core components of a coal mill based on vibration analysis comprises collecting multi-point vibration signal data of core components of the coal mill, and performing sampling rate adjustment and noise pre-suppression to generate an optimized vibration data set; based on the optimized vibration data set, variational modal decomposition and deep embedding are performed to extract adaptive modal features and embedding vectors, and a high-level vibration feature group is generated; based on the high-level vibration feature group, a transfer learning integrated prediction model is established, and cross-condition fault transition probabilities are calculated to generate a preliminary fault prediction distribution; based on the preliminary fault prediction distribution, a reinforcement learning feedback iterative cycle is performed to optimize the transfer learning integrated prediction model, and a refined fault prediction result is generated; based on the refined fault prediction result, multi-dimensional risk quantification and evolution path simulation are performed, and a fault prediction report is output.

[0007] Optionally, the generating of the optimized vibration data set comprises: collecting multi-point vibration signal data of core components of the coal mill to generate an initial vibration signal group; based on the initial vibration signal group, adaptive sampling rate optimization is performed to adjust the sampling frequency to match vibration frequency spectrum changes, and an adjusted vibration signal group is generated; based on the adjusted vibration signal group, a noise pre-suppression cycle is performed, and iterative filtering operations are performed to remove noise interference, and a pre-suppression vibration data set is generated; based on the pre-suppression vibration data set, data standardization processing is performed to generate an optimized vibration data set.

[0008] Optionally, the generating of the high-level vibration feature group comprises: based on the optimized vibration data set, variational variational modal decomposition is performed to dynamically select the number of modes and decompose non-stationary vibration components, and a modal decomposition sequence group is generated; based on the modal decomposition sequence group, the attention mechanism is used to optimize the temporal and spatial relationship between modes, and an embedding vector set is generated; based on the embedding vector set, intelligent modal feature fusion and vector refinement are performed to generate a high-level vibration feature group.

[0009] Optionally, the method further comprises: based on the adjusted vibration signal group and the modal decomposition sequence group, time-space coupling graph modeling is performed to fuse the time-space relationship of the signal group and the dynamic characteristics of the modal sequence to generate a preliminary fusion index group; based on the preliminary fusion index group, multi-dimensional correlation analysis and noise residual compensation are performed to iteratively verify the time-space coupling strength and compensation parameters to generate a compensation fusion index group; based on the compensation fusion index group, dynamic threshold iterative optimization is performed to quantify vibration dynamics and refine index weights to generate a fusion vibration dynamic index.

[0010] Optionally, the generating the preliminary fault prediction distribution comprises: based on the high-level vibration feature group and the fusion vibration dynamic indicator, performing graph transfer learning by transferring cross-condition knowledge through domain adaptive graph structure, to generate a migration feature representation group; based on the migration feature representation group, constructing a transfer learning integrated prediction model, and quantifying a fault state probability path to generate a probability transition matrix; and based on the probability transition matrix, performing fault distribution refining and uncertainty quantification to obtain the preliminary fault prediction distribution.

[0011] Optionally, the obtaining the preliminary fault prediction distribution comprises: based on the probability transition matrix, performing sampling simulation to generate a random fault path sample group; based on the random fault path sample group, performing variational inference optimization to quantify uncertainty boundaries and refine distribution parameters to generate a refined distribution parameter group; and based on the refined distribution parameter group, fusing path samples and uncertainty indicators to obtain the preliminary fault prediction distribution.

[0012] Optionally, the generating the refined fault prediction result comprises: based on the preliminary fault prediction distribution, initializing the reinforcement learning feedback iterative cycle and constructing a multi-objective optimization function to generate an initial optimization parameter group; based on the initial optimization parameter group, performing dynamic reward adaptive adjustment, and iteratively updating a reward mechanism to optimize the transfer learning integrated prediction model to generate an intermediate optimization model; and based on the intermediate optimization model, performing convergence verification and parameter refining on the multi-objective optimization function to generate the refined fault prediction result.

[0013] Optionally, the generating the intermediate optimization model comprises: based on the initial optimization parameter group and the fusion vibration dynamic indicator, initializing reward evolution to generate an initial reward strategy group; based on the initial reward strategy group, performing reward evolution iteration and injecting the fusion vibration dynamic indicator to generate an evolved reward mechanism; and based on the evolved reward mechanism, iteratively updating parameters of the transfer learning integrated prediction model to generate the intermediate optimization model.

[0014] Optionally, the outputting the fault prediction report comprises: based on the refined fault prediction result and the intermediate optimization model, performing dynamic sensitivity multi-dimensional risk quantification, calculating risk indicator influence weights, and generating a multi-dimensional risk indicator group; based on the multi-dimensional risk indicator group, performing time sequence dependent evolution path simulation, predicting fault development trajectories and evaluating probability boundaries, and generating an evolution path group; and based on the evolution path group, performing risk indicator fusion and report visualization to generate the fault prediction report.

[0015] Optionally, the method further comprises: based on the probability transition matrix and the initial optimization parameter set, performing atlas multi-modal integration to generate a preliminary integrated indicator set; based on the preliminary integrated indicator set, performing multi-modal correlation refinement and uncertainty compensation to generate a refined integrated indicator set; and based on the refined integrated indicator set, performing closed-loop self-calibration feedback to the transfer learning integrated prediction model.

[0016] Compared with the prior art, the present application has the following advantages:

[0017] 1. The present application improves the data processing mode of "passive reception, shallow analysis" in traditional methods to "active capture, deep insight" system capability through the construction of "event-driven intelligent sampling" and "physical-data dual-domain deep feature extraction" mechanism. This mechanism ensures the fidelity of key fault transient information from the data source, and through adaptive decomposition and embedding technology, decouples the deep features highly related to the physical state from the complex signal, providing an unprecedented high signal-to-noise ratio decision basis for subsequent accurate prediction.

[0018] 2. The present application changes the fundamental limitation of "one-time training, static application" of traditional prediction models by constructing a multi-closed-loop self-evolution architecture of "transfer learning-reinforcement learning-top feedback". Not only can it perform real-time self-optimization and strategy adjustment based on internal reward signals through reinforcement learning, but also can use future prediction to guide current data acquisition strategy, realizing the transition from "passive prediction" to "active cognition and forward planning", and showing deep adaptive ability to complex working conditions and unknown modes.

[0019] Other features and advantages of the present application will be set forth in the following description, and in part will become apparent to those skilled in the art from the description, or can be learned by practice of the present application. The objects and other advantages of the present application can be achieved and obtained by the structures indicated in the description, claims and drawings. BRIEF DESCRIPTION OF DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0021] Figure 1 is a flowchart of the coal mill core component fault prediction method based on vibration analysis according to an embodiment of the present application.

[0022] Figure 2 is a variable parameter variational mode decomposition process curve according to an embodiment of the present application.

[0023] Figure 3 is a knowledge enhanced network diagram of cross-condition diagram migration learning of the embodiment of the application.

[0024] Figure 4 is a fault evolution path simulation and probability boundary schematic diagram of the embodiment of the application. DETAILED DESCRIPTION

[0025] To make the objects, technical solutions and advantages of the embodiments of the application clearer, the technical solutions in the embodiments of the application will be described below in a clear and complete manner with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are some but not all of the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the application.

[0026] With reference to Figure 1 , one embodiment of the application proposes a coal mill core component fault prediction method based on vibration analysis. A multi-level, self-evolution closed-loop intelligent prediction architecture is adopted. Based on the uncertainty of the prediction distribution, an online model optimization is performed by a reinforcement learning agent, the rapid self-adaptation to the variable working condition environment is realized, and the forward-looking maintenance decision support is provided in combination with the evolution path simulation, so that the early weak fault capturing capability of the coal mill, the accuracy of the cross-condition prediction and the scientificity of the final maintenance decision are significantly improved.

[0027] The method of the embodiment specifically includes:

[0028] Collecting multi-point vibration signal data of a coal mill core component, and performing sampling rate adjustment and noise pre-suppression to generate an optimized vibration data set;

[0029] Based on the optimized vibration data set, performing variational modal decomposition fusion deep embedding to extract adaptive modal features and embedding vectors, and generating a high-level vibration feature group;

[0030] Based on the high-level vibration feature group, establishing a transfer learning integrated prediction model, and calculating a cross-condition fault transition probability to generate a preliminary fault prediction distribution;

[0031] Based on the preliminary fault prediction distribution, performing a reinforcement learning feedback iterative cycle to optimize the transfer learning integrated prediction model, and generating a refined fault prediction result;

[0032] Based on the refined fault prediction result, performing multi-dimensional risk quantification and evolution path simulation, and outputting a fault prediction report.

[0033] A multi-level, self-evolution closed-loop intelligent prediction architecture is adopted, which can realize online model optimization by reinforcement learning agent based on the uncertainty of prediction distribution, realize rapid self-adaptation to variable working condition environment, and provide forward-looking maintenance decision support combined with evolutionary path simulation, significantly improve the capture ability of early weak faults of the coal mill, the accuracy of cross-condition prediction and the scientificity of final maintenance decision.

[0034] Optionally, the generating an optimized vibration dataset comprises:

[0035] Collecting multi-point vibration signal data of the core components of the coal mill to generate an initial vibration signal group;

[0036] Specifically, this step aims to obtain the most original device state information from the physical world. In key core components of the coal mill, such as the main bearing seat, the mill roller rocker arm, the reduction gearbox shell, etc., multiple high-frequency piezoelectric acceleration sensors are deployed. These sensors continuously collect vibration acceleration signals at each measurement point at a fixed basic sampling frequency that is several times higher than the highest analysis frequency, and these multi-channel signal data collected at the same time reference without any processing are collected to form an initial vibration signal group.

[0037] Based on the initial vibration signal group, adaptive sampling rate optimization is performed to adjust the sampling frequency to match the vibration frequency spectrum changes, and an adjusted vibration signal group is generated;

[0038] Specifically, this step aims to intelligently allocate collection resources to minimize data redundancy while ensuring information integrity. In this process, the initial vibration signal group is subjected to Short-Time Fourier Transform (STFT) to obtain a series of time-frequency spectra. For each frame of time-frequency spectrum, its spectral entropy is calculated to quantify the complexity of the signal within the time window. The calculation formula of spectral entropy can be:

[0039] ,

[0040] wherein, represents the time window, represents the th frequency component, is the total number of frequency components, is the number of frequency components in the time window , ​normalized power spectral density. The higher the spectral entropy value, the more abundant the frequency components of the signal. Subsequently, a feedback control logic is established, taking the calculated spectral entropy as the control input to dynamically adjust the sampling frequency of the next acquisition cycle. For example, when the spectral entropy exceeds a pre-set high complexity threshold, the sampling frequency is raised to the highest diagnostic level, thereby generating an adjusted vibration signal set.

[0041] Based on the adjusted vibration signal set, a noise pre-suppression cycle is performed to iteratively filter out noise interference, generating a pre-suppressed vibration dataset;

[0042] Specifically, this step aims to separate weak but critical fault feature signals from a strong noise background. This cycle adopts an iterative signal reconstruction and residual evaluation strategy. In each iteration, the signal is processed by an adaptive filter that adjusts its internal parameters based on the signal residual generated in the previous iteration. This process continues until the separation degree of signal components and noise components reaches the optimal, or the energy of the signal residual converges below a pre-set stable threshold, and the final signal is the pre-suppressed vibration dataset.

[0043] Based on the pre-suppressed vibration dataset, data standardization processing is performed to generate an optimized vibration dataset.

[0044] Specifically, this step aims to eliminate differences between different sensor channels and different physical dimensions, providing a unified numerical benchmark for subsequent model training. Exemplarily, the Z-score standardization method can be used. This method calculates the mean and standard deviation of all data points of each sensor channel, and then converts each data point value to the standard deviation multiple of its deviation from the mean, so that the data of all channels conform to the standard normal distribution with mean 0 and variance 1, finally generating the optimized vibration dataset.

[0045] Optionally, the generating a high-level vibration feature set comprises:

[0046] Based on the optimized vibration dataset, a variable-parameter variational modal decomposition is performed to dynamically select the number of modes and decompose non-stationary vibration components, generating a modal decomposition sequence set;

[0047] Specifically, this step aims to decompose the complex non-stationary signal containing various fault information coupled with operating noise in the optimized vibration data set into a series of simpler and more physically interpretable intrinsic mode functions. Unlike traditional variational mode decomposition, which requires manual pre-setting of the number of modes, the variational variational mode decomposition in this step adopts an adaptive optimization strategy. This strategy automatically and dynamically determines the optimal number of decomposition modes for each analysis time window by calculating indicators such as the center frequency, kurtosis, or energy of the signal, thereby more accurately decomposing vibration components with different characteristic frequencies related to different components such as bearings, gears, and rotors. The time series of multiple intrinsic mode functions generated ultimately constitute a modal decomposition sequence group, as shown in Figure 2 As shown in the figure, it shows how a complex original signal is adaptively decomposed into multiple simpler intrinsic mode functions containing different physical meanings by the variational variational mode decomposition method of the present application in the same coordinate system, intuitively reflecting the process of decoupling the core dynamic components from the complex coupled signal.

[0048] Based on the modal decomposition sequence group, the spatio-temporal relationship between modes is optimized using an attention mechanism to generate an embedding vector set;

[0049] Specifically, this step aims to deeply mine and quantify the complex internal relationships between different vibration modes. Since the evolution of device failures is often the result of the interaction of multiple physical components, there is a coupling relationship between different modal sequences in the time and space dimensions. In this step, an attention mechanism is introduced, which can automatically assign "attention weights" to different modes at different times, just like a human expert. For example, the mechanism can learn that when a certain high-frequency mode related to the bearing increases in energy, it should pay more "attention" to the changes in another low-frequency mode related to structural resonance. In this way, the original independent modal decomposition sequence group is transformed into a fixed-length low-dimensional vector that reflects the dynamic relationship between modes, i.e., the embedding vector set.

[0050] Based on the embedding vector set, intelligent modal feature fusion and vector refinement are performed to generate a high-level vibration feature group.

[0051] Specifically, this step aims to construct the final feature representation with the highest information density for driving the subsequent prediction model. In this process, first, the statistical features such as energy and entropy of each sequence in the modal decomposition sequence group are extracted to form a physical feature set. Then, the physical feature set is spliced and fused with the embedding vector set representing the abstract relationship between modes generated in the previous step. In order to further refine the information and eliminate redundancy, the spliced and fused vector is input into the encoding end of a deep autoencoder for vector refinement, and through nonlinear mapping, it is compressed into a more compact and more powerful final form, thereby generating a high-level vibration feature group.

[0052] Optionally, the method further comprises:

[0053] Based on the adjusted vibration signal group and the modal decomposition sequence group, a spatiotemporal coupling graph modeling is performed to fuse the spatiotemporal relationship of the signal group and the dynamic characteristics of the modal sequence, and a preliminary fusion index group is generated.

[0054] Specifically, this step aims to construct a parallel analysis path for directly extracting the dynamic and instability indicators of equipment operation from the original signals and decomposition modes. In this process, a spatiotemporal coupling graph is constructed, where the nodes of the graph represent each physical sensor, and the edge weights between the nodes are used to quantify the coupling strength between different sensor signals. The calculation of the edge weight fuses the spatial distance and the temporal synchronicity, and its calculation formula can be:

[0055] ,

[0056] wherein, is the physical distance between the sensors and , is a distance scaling factor, and the first term is used to represent the spatial coupling relationship. is the phase-locked value between the modal decomposition sequence groups corresponding to the two sensors within the time window , which is used to represent the closeness of the temporal coupling. The constructed spatiotemporal coupling graph is input into a graph neural network (GNN), and the spatiotemporal information of the adjacent nodes is aggregated through graph convolution operation. Finally, the embedding representation output by each node constitutes the preliminary fusion index group.

[0057] Based on the preliminary fusion index group, a multi-dimensional correlation analysis and noise residual compensation are performed, and the spatiotemporal coupling strength and the compensation parameters are iteratively verified to generate a compensation fusion index group.

[0058] Specifically, this step aims to refine and calibrate the preliminary graph model output. In this process, firstly, cross-correlation analysis is performed on each indicator within the preliminary fusion indicator group to assess the consistency of dynamic characteristics in different physical locations. At the same time, a noise residual compensation is performed. This compensation is obtained by calculating the difference between the total energy of the adjusted vibration signal group and the sum of the energy of all sequences in the modal decomposition sequence group, obtaining a residual energy that cannot be explained by any mode. This residual energy mainly corresponds to random noise or nonlinear coupling components. The preliminary fusion indicator group is modified using the residual energy, for example, when the residual energy is high, the confidence of the indicator is appropriately reduced, thereby generating a compensated fusion indicator group.

[0059] Based on the compensated fusion indicator group, dynamic threshold iterative optimization is performed to quantify the vibration dynamic change and refine the indicator weight, generating a fusion vibration dynamic indicator.

[0060] Specifically, this step aims to generate a final single indicator that can sensitively reflect the dynamic change of the equipment operating state. In this process, a statistical method based on a sliding window is used to continuously calculate the mean and standard deviation of the compensated fusion indicator group in the past period of time. Based on the statistical results, a dynamic and adaptive warning threshold is generated. The "iterative optimization" is reflected in the continuous sliding of the statistical window, so that the threshold can automatically adapt to the normal fluctuations of the equipment in different steady-state processes. Finally, the compensated fusion indicator group after normalization and smoothing processing is output as the fusion vibration dynamic indicator, which can be directly compared with the adaptive warning threshold to evaluate the current operating instability of the equipment.

[0061] Optionally, the generating a preliminary fault prediction distribution comprises:

[0062] Based on the high-level vibration feature group and the fusion vibration dynamic indicator, the domain adaptive graph structure is transferred to transfer cross-condition knowledge, the graph transfer learning is performed, and a transfer feature representation group is generated.

[0063] Specifically, this step aims to solve the core technical problem that the prediction performance of traditional models significantly decreases when the operating condition of the coal mill changes. In this process, first, two graph structures are constructed: one is a "source domain feature correlation graph" that contains universal fault mechanism knowledge, which is constructed based on a large amount of general vibration database; the other is a "target domain feature correlation graph" that is constructed based on the current coal mill real-time generated advanced vibration feature group and fusion vibration dynamic indicators. The core of graph transfer learning is to use the topology structure and knowledge of the "source domain feature correlation graph" as a priori information, and through a domain adaptation module to guide and enhance the construction of the "target domain feature correlation graph", especially to complete and strengthen the weak correlation caused by data sparseness in the target domain. After this cross-condition knowledge transfer, the node representation finally extracted from the enhanced target domain graph structure is the transfer feature representation group. The feature of this representation group is that it has robustness to different operating conditions, such as Figure 3 As shown in the figure, the "source domain" full connection structure with complete knowledge is represented by a gray dashed line superimposed on the same group of feature nodes, and the black solid line highlights the "target domain" sparse structure driven by local data, which shows the core idea of graph transfer learning using prior knowledge to enhance the target domain graph representation.

[0064] Based on the transfer feature representation group, a transfer learning integrated prediction model is constructed, and the fault state probability path is quantified to generate a probability transition matrix.

[0065] Specifically, this step aims to model the degradation process of the device as a multi-state random process with clear physical meaning. In this step, a set of implicit health states corresponding to the physical health degree of the core component is defined in advance, such as "healthy", "early degradation", "significant degradation", and "imminent failure". An integrated prediction model is constructed, such as a hidden Markov model or its variants, with these implicit health states as internal states and the transfer feature representation group as external observations. Through training, the model learns to infer the implicit health state that the device is most likely in from the observable transfer feature representation group, and more critically, quantifies the probability of transitioning from one implicit health state to another implicit health state in a unit of time. All state transition probabilities are organized into a matrix form, which is the probability transition matrix.

[0066] Based on the probability transition matrix, fault distribution refinement and uncertainty quantification are performed to obtain the preliminary fault prediction distribution.

[0067] In particular, this step aims to generate a dynamic, complete prediction distribution that can reflect the future likelihood and uncertainty from a static probability transition matrix. In this process, probabilistic programming techniques such as Monte Carlo sampling or variational inference are adopted. Based on the probability transition matrix, thousands or even tens of thousands of possible future fault evolution paths are simulated. Through statistical analysis of these massive simulation paths, not only the most likely fault development trajectory can be obtained, but more importantly, the uncertainty of the prediction can be quantified, for example, by calculating the dispersion range or variance of all simulation paths at each future time point. Finally, the complete probability information including the most likely evolution path, probability confidence interval, and uncertainty indicators are integrated to form the preliminary fault prediction distribution.

[0068] Optionally, the obtaining the preliminary fault prediction distribution comprises:

[0069] Based on the probability transition matrix, a sampling simulation is performed to generate a random fault path sample set;

[0070] In particular, this step aims to transform the static, one-step likelihood described by the probability transition matrix into a dynamic, multi-step, visualized set of future evolution trajectories. In this process, sampling techniques such as Markov Chain Monte Carlo (MCMC) are adopted. Starting from the current inferred device health state, according to the transition probabilities defined in the probability transition matrix, random sampling is performed to determine the health state at the next time. By independently repeating this process tens of thousands of times, a random fault path sample set containing a large number of possible future evolution trajectories is finally generated.

[0071] Based on the random fault path sample set, variational inference optimization is performed to quantify the uncertainty boundary and refine the distribution parameters to generate a refined distribution parameter set;

[0072] In particular, this step aims to refine a continuous, parameterized mathematical model that can describe the overall distribution of future states from the massive, discrete random path samples. In this process, variational inference (VI) techniques are adopted. The core idea of this technique is to approximate the true but difficult to directly calculate posterior distribution that is implied by the random fault path sample set with a simple, parameter-defined approximate probability distribution . This approximation process is achieved by maximizing an objective function called "Evidence Lower Bound (ELBO)", whose formula can be:

[0073] ,

[0074] wherein, represents the latent variables of the failure paths, represents the observed evidence, is an approximate distribution defined by a set of distribution parameters to be optimized, is a joint probability distribution. The parameters of are iteratively adjusted by an optimization algorithm to maximize the evidence lower bound, and the final set of optimal parameters, i.e., the refined distribution parameters, are obtained. The variance terms in this parameter set directly constitute the quantification of the uncertainty boundary of the prediction.

[0075] Based on the refined distribution parameter set, the path samples and the uncertainty indicators are fused to obtain a preliminary failure prediction distribution.

[0076] Specifically, this step aims to finally integrate and package the calculation results of the previous steps to form a structured and complete information output. In this process, the refined distribution parameter set is analyzed to extract the expected trajectory representing the "most likely" evolution path and the confidence interval representing the uncertainty boundary. The expected trajectory, confidence interval, and a representative set of random failure path samples are fused together to form a complete probability picture that contains both parametric statistical description and non-parametric sample examples. This picture is the preliminary failure prediction distribution.

[0077] Optionally, the generating the refined failure prediction result comprises:

[0078] Based on the preliminary failure prediction distribution, initializing the reinforcement learning feedback iterative loop and constructing a multi-objective optimization function to generate an initial optimization parameter set;

[0079] Specifically, this step aims to establish an intelligent feedback module with higher dimension and self-optimization capability. In this process, a reinforcement learning (RL) framework is constructed, where the state of the reinforcement learning agent is defined by the preliminary failure prediction distribution and its uncertainty indicators; the action space of the agent is defined as a set of adjustments to the hyperparameters or structural parameters of the transfer learning integrated prediction model. At the same time, a multi-objective optimization function is constructed, which contains at least two mutually balanced objectives: one is to maximize the accuracy of the prediction, and the other is to minimize the uncertainty of the prediction result. The initial optimization parameter set is the initial weight parameter of the policy network of the reinforcement learning agent.

[0080] Based on the initial optimization parameter set, performing dynamic reward adaptive adjustment to iteratively update the reward mechanism to optimize the transfer learning integrated prediction model, generating an intermediate optimization model;

[0081] Specifically, this step is the core of the reinforcement learning feedback iterative loop, aiming to enable the learning goal of the agent to dynamically adapt to the real-time physical state of the device. In this process, the reward function of the agent is designed as a dynamically weighted combined function, and the calculation of the reward value can be defined by the following formula:

[0082] ,

[0083] wherein, is the total reward value at the moment, is the "prediction refinement benefit" related to reducing the prediction uncertainty, is the "abnormal exploration benefit" related to improving the sensitivity of the model to mutation signals. The key is that the weight coefficient is a dynamic variable, and its value is determined by the real-time value of the fusion vibration dynamic indicator. For example, when the device runs smoothly, the indicator value is low, and increases, so that the agent focuses more on "refinement"; when the device runs unstably, the indicator value is high, and decreases, so that the agent focuses more on "change". Through this dynamic reward adaptive adjustment mechanism, the agent continuously generates intermediate optimization models in the iterative optimization process. Based on the intermediate optimization model, the multi-objective optimization function is verified and the parameters are refined to generate a refined fault prediction result.

[0084] Specifically, this step aims to ensure the stability of the optimization process and the reliability of the final result. After the reinforcement learning agent performs a preset number of iterative updates, it enters the convergence verification phase. In this phase, the change of the value of the multi-objective optimization function in continuous multiple iteration cycles is monitored, and when the fluctuation of the value is less than a preset convergence threshold, it is determined that the optimization process has converged. Then, from the multiple intermediate optimization models in the convergence phase, a model that performs best on the multi-objective optimization function is selected as the final refined model. Input the latest advanced vibration feature group into the final refined model, and the output is the refined fault prediction result.

[0085] Optionally, the generation of the intermediate optimization model comprises:

[0086] Based on the initial optimization parameter set and the fusion vibration dynamic indicator, initialize reward evolution to generate an initial reward strategy set;

[0087]

[0088] ​Specifically, this step aims to evolve the "reward mechanism" in reinforcement learning from a static, pre-defined function to a dynamic, optimizable agent. In this process, a "population" containing multiple potential reward strategies is initialized, i.e., the initial reward strategy group. Each reward strategy is an independent candidate that defines how to calculate the reward value based on the preliminary fault prediction distribution and uncertainty. For example, the population can include strategies that focus on punishing long-term risks, strategies that focus on rewarding short-term prediction stability, and multiple strategies with different tolerance to uncertainty. The fusion vibration dynamic indicator is used as a key input to evaluate the initial fitness of different strategies under the current equipment dynamics at this stage.

[0089] Based on the initial reward strategy group, perform reward evolution iterations and inject the fusion vibration dynamic indicator to generate an evolved reward mechanism;

[0090] Specifically, this step aims to select and incubate the optimal reward strategy from the initial population through the simulated "survival of the fittest" evolution process. In this process, an iterative evolution algorithm such as genetic algorithm or evolutionary strategy is executed. In each iteration, each reward strategy in the population is evaluated for its "fitness", which measures the performance it can achieve in guiding the reinforcement learning agent to optimize the transfer learning integrated prediction model. Here, the fusion vibration dynamic indicator is continuously "injected" into the fitness evaluation function as a dynamically changing environmental variable. This means that an excellent reward strategy not only performs well when the equipment is stable, but also guides the agent to make correct optimization when the equipment dynamicity indicator rises. Through multiple generations of genetic, crossover, and mutation operations, the reward strategy that survives and has the strongest adaptability is the evolved reward mechanism.

[0091] Based on the evolved reward mechanism, iteratively update the transfer learning integrated prediction model parameters to generate an intermediate optimization model.

[0092] Specifically, this step aims to apply the final result of the previous evolution process to perform the final and most efficient optimization of the main prediction model. In this process, the evolved reward mechanism is established as an immutable "gold standard" in the reinforcement learning feedback iteration loop. Using this well-honed reward mechanism, the network parameters or structural parameters of the transfer learning integrated prediction model are updated iteratively and finally targeted. Since the reward mechanism itself is optimal, the optimization process here converges faster, the results are more stable, and the final model obtained is the intermediate optimization model.

[0093] Optionally, the output fault prediction report includes:

[0094] Based on the refined failure prediction result and the intermediate optimization model, dynamic sensitivity multi-dimensional risk quantification is performed, risk indicator influence weights are calculated, and a multi-dimensional risk indicator group is generated;

[0095] Specifically, this step aims to deeply analyze and quantify the root driving factors that cause the current failure risk. In this process, a model-independent feature attribution method, such as the Shapley Additive Explanation (SHAP) algorithm, is used to perform dynamic sensitivity multi-dimensional risk quantification. This method considers each input feature as a "contributor" and calculates its marginal contribution value to the refined failure prediction result of the intermediate optimization model output. Through this method, the contribution degree of each input feature to the final prediction risk can be accurately calculated. Combining the risk indicator influence weights of all features with the overall risk probability, a multi-dimensional risk indicator group is generated.

[0096] Based on the multi-dimensional risk indicator group, time-dependent evolution path simulation is performed, the failure development trajectory is predicted and the probability boundary is evaluated, and an evolution path group is generated;

[0097] Specifically, this step aims to prospectively deduce the possible future failure development paths based on the current risk state and driving factors. Unlike the simulation based on static transition matrix, the time-dependent evolution path simulation here utilizes the time series memory ability of the intermediate optimization model itself. The current advanced vibration feature group is taken as the initial input, and through autoregressive prediction, the model's prediction output at one time step is taken as part of the input at the next time step, and multi-step prediction is iteratively performed. At the same time, random perturbations consistent with the preliminary failure prediction distribution are introduced in each step of prediction, and through thousands of independent simulations, a future development trajectory containing multiple possible and time-dependent relationships is finally generated, i.e., the evolution path group, as shown in Figure 4 The figure shows the future failure evolution paths simulated based on the current device state, which includes the most likely degradation trajectory calculated by the refined failure prediction result and the confidence interval composed of the uncertainty boundary, providing intuitive data support for prospective maintenance decisions.

[0098] Based on the evolution path group, risk indicator fusion and report visualization are performed, and a failure prediction report is generated.

[0099] Specifically, this step aims to present all the analysis results in a highly friendly and informative way to the operation personnel. In this process, the multi-dimensional risk indicator group and the evolution path group generated in the previous steps are deeply fused. The final generated failure prediction report is a visual, multi-level interactive report, which can include: a risk dashboard showing the overall failure probability, uncertainty boundary and risk level; a risk factor contribution view that clearly shows which specific feature indicators cause the risk to rise through a waterfall chart or contribution ranking; and a failure trajectory prediction view that shows the most likely future failure evolution path and probability boundary through a pie chart or trajectory bundle.

[0100] Optionally, the method further comprises:

[0101] Based on the probability transition matrix and the initial optimization parameter group, performing graph multi-modal integration to generate a preliminary integrated indicator group;

[0102] Specifically, this step aims to build a higher-dimensional self-calibration module for evaluating the consistency between the "model dynamics" and the "optimization strategy". In this process, the probability transition matrix is regarded as an adjacency matrix of a directed weighted graph describing the evolution law of the device health state space. At the same time, the reinforcement learning initial strategy defined by the initial optimization parameter group is regarded as a Markov decision process for decision-making on this state graph. Graph multi-modal integration is to analyze the consistency between the state transition path induced by the decision process and the internal physical transition law defined by the probability transition matrix. This consistency, or "harmony degree", can be quantified by a metric function, which can be:

[0103] ,

[0104] wherein, is the probability transition matrix, is the reinforcement learning strategy defined by the initial optimization parameter group the induced transition matrix of the generated strategy in the simulation environment, the Frobenius norm representing the difference between the two matrices. By comprehensively analyzing the "harmony degree" and the spectral properties of the graph, a preliminary integrated indicator group is generated.

[0105] Based on the preliminary integrated indicator group, multi-modal correlation refinement and uncertainty compensation are performed to generate a refined integrated indicator group;

[0106] Specifically, this step aims to improve the robustness and reliability of the preliminary integrated indicator set. During this process, an uncertainty compensation is performed. This compensation utilizes the uncertainty boundary of the probability transition matrix itself quantified in the variational inference optimization step. Specifically, if the "harmony" is low on a certain state transition path, but the transition probability corresponding to this path itself has very high uncertainty, the weight of this inharmony will be reduced. In this way, all preliminary integrated indicators are once weighted and corrected based on uncertainty, thereby avoiding incorrect calibration due to insufficient understanding of the model for some rare working conditions, and finally generating a refined integrated indicator set.

[0107] Based on the refined integrated indicator set, a closed-loop self-calibration feedback is performed to the transfer learning integrated prediction model.

[0108] Specifically, this step is the closed-loop link to achieve the entire top-level self-calibration. The refined integrated indicator set is finally converted into a calibration signal and fed back to the transfer learning integrated prediction model. The calibration signal directly acts on the regularization term or loss function of the transfer learning integrated prediction model. For example, when the refined integrated indicator set shows a low "harmony", it indicates that the optimization goal of reinforcement learning may conflict with the inherent cognition of the model, at which time the calibration signal will increase the regularization coefficient of the transfer learning integrated prediction model, forcing the model to learn more general and generalized features to avoid falling into a local optimal state that is physically unreasonable although it can be optimized by reinforcement learning. This closed-loop self-calibration feedback mechanism ensures the long-term stability of the entire prediction and the fidelity to the physical reality.

[0109] To verify the feasibility and advancement of the present application, it is applied to a 600 MW supercritical thermal power plant No. 3 MPS medium-speed coal mill. The mill has historically experienced unplanned shutdowns due to gradual wear and tear of core components and sudden coal blockage, severely affecting unit efficiency. This embodiment aims to apply the present application method to make forward-looking fault prediction on the health status of the core components of the mill.

[0110] At 10:42 on March 15, 2024, when a short transient impact was triggered by a fluctuation in coal supply, the change in vibration spectrum was identified and immediately triggered an event-driven high-fidelity sampling, instantaneously increasing the sampling rate of the relevant sensors to 20 kHz, successfully capturing an optimized vibration data set containing rich details. The data set was then subjected to variational variational modal decomposition and attention mechanism, identifying a weak high-frequency mode related to the mill roller impact and generating a high-level vibration feature set.

[0111] Meanwhile, the fusion vibration dynamic indicator keeps outputting, which starts to climb after the transient impact event, indicating that the overall stability of the equipment is declining. Combined with the advanced vibration feature group and the dynamic indicator, and using the source domain knowledge learned from other same type coal mills, a preliminary fault prediction distribution is generated. The distribution is quantified through the probability transition matrix and variational inference optimization, showing that the probability of the equipment entering the "early degradation state" within the next 200 hours is 75%, but with high uncertainty.

[0112] At this time, the reinforcement learning feedback iteration loop is activated. Due to the high fusion vibration dynamic indicator, the reward function is dynamically adjusted, making the optimization goal tilt towards "abnormal state exploration benefit", and finally generating a refined fault prediction result. The final fault prediction report is output at 14:00 on the same day, which not only gives the conclusion that "there is a high probability of early spalling of the grinding roller surface within 150 hours", but also quantifies the dynamic sensitivity, indicating that the risk dominant factor is the aforementioned weak high-frequency impact mode, and the evolution path simulation shows that the fault has an 80% probability of accelerating deterioration after 170 hours.

[0113] Most importantly, based on the deduced fault evolution path and probability transition matrix, a forward-looking information value map is constructed, and an acquisition strategy optimization instruction is generated and fed back to the data acquisition module. This instruction, in the subsequent operation, proactively improves the sampling resolution of the 2nd and 4th sensors most relevant to the grinding roller state. Finally, after about 145 hours, the power plant actively overhauls according to the prediction report, confirming that there is an early pitting defect on the grinding roller surface consistent with the predicted height, successfully avoiding an unplanned shutdown.

[0114] Table 1 Intelligent acquisition and optimization effect data table of coal mill vibration data

[0115]

[0116] Table 2 Comparison table of key indicators of coal mill fault prediction model

[0117]

[0118] Table 3 Verification table of coal mill fault prediction and actual maintenance results

[0119]

[0120] From the data recorded in the above Tables 1-3, it can be seen that the present application performs outstandingly in the examples. The data in Table 1 verifies the effectiveness of the adaptive sampling strategy in capturing key transient information. Table 2 shows the great advantages of the present application in key performance indicators such as prediction accuracy, lead time and false alarm rate through comparison with various benchmark prediction models. Table 3 records the whole process from the first generation of the prediction report to the final manual inspection verification, and the detailed data clearly proves the accuracy of the prediction results of the present application and the great engineering application value of the final decision report.

[0121] It should be noted that the above formulas can be translated into unitless standard values or parameters of the same dimension that can be superimposed by means of dimensional consistency principles and mathematical standardization methods (such as normalization processing, dimensionless parameter conversion or unit system unification), so as to eliminate the interference of different dimensions on the operation logic, so that the formulas have mathematical operation rationality and objective law adaptability while retaining the original data distribution characteristics. It is a conventional technical means, and will not be repeated here. The electrical connection between the above-mentioned units does not necessarily mean direct connection, and indirect connection can also be used as long as the purpose of the present application is achieved. The above-mentioned is only an exemplary embodiment of the present application, and cannot limit the scope of the present application.

[0122] That is, any equivalent changes and modifications made according to the teachings of the present application are still within the scope of the present application. Other embodiments of the present application will be readily apparent to those skilled in the art upon considering the specification and practice of the true principles disclosed. The present application is intended to cover any variations, uses or adaptive changes to the present application following the general principles of the present application and including commonly known or conventional technical means in the art not disclosed by the present application.

Claims

1. A method for predicting the failure of core components of a coal mill based on vibration analysis, characterized in that, The method includes: Multi-point vibration signal data of the core components of a coal mill are collected, and sampling rate adjustment and noise pre-suppression are performed to generate an optimized vibration dataset. The generation of the optimized vibration dataset includes: collecting multi-point vibration signal data of the core components of the coal mill to generate an initial vibration signal set; based on the initial vibration signal set, performing adaptive sampling rate optimization to adjust the sampling frequency to match the vibration spectrum changes, generating an adjusted vibration signal set; based on the adjusted vibration signal set, performing a noise pre-suppression loop, iteratively filtering to remove noise interference, generating a pre-suppressed vibration dataset; and based on the pre-suppressed vibration dataset, performing data standardization processing to generate an optimized vibration dataset. Based on the optimized vibration dataset, variable parametric variational mode decomposition is performed to dynamically select the number of modes and decompose non-stationary vibration components, generating a mode decomposition sequence group. Based on the mode decomposition sequence group, an attention mechanism is used to optimize the spatiotemporal relationship between modes, generating an embedding vector set. Based on the embedding vector set, intelligent modal feature fusion and vector refinement are performed to generate an advanced vibration feature group. The method further includes: based on the adjusted vibration signal group and the mode decomposition sequence group, spatiotemporal coupling graph modeling is performed to fuse the spatiotemporal relationship of the signal group and the dynamic characteristics of the modal sequence, generating a preliminary fusion index group. Based on the preliminary fusion index group, multi-dimensional correlation analysis and noise residual compensation are performed, iteratively verifying the spatiotemporal coupling strength and compensation parameters, generating a compensated fusion index group. Based on the compensated fusion index group, dynamic threshold iterative optimization is performed to quantify the dynamic changes of vibration and refine the index weights, generating a fused vibration dynamic index. Based on the advanced vibration feature set and the fused vibration dynamic index, graph transfer learning is performed through domain-adaptive graph structure transfer knowledge across operating conditions to generate a transfer feature representation set. Based on the transfer feature representation set, a transfer learning ensemble prediction model is constructed, and the probability paths of fault states are quantified to generate a probability transition matrix. Based on the probability transition matrix, sampling simulation is performed to generate a random fault path sample set. Based on the random fault path sample set, variational inference optimization is performed to quantify the uncertainty boundary and refine the distribution parameters to generate a refined distribution parameter set. Based on the refined distribution parameter set, the path samples and uncertainty index are fused to obtain a preliminary fault prediction distribution. Based on the preliminary fault prediction distribution, a reinforcement learning feedback iterative loop is executed to optimize the transfer learning ensemble prediction model and generate refined fault prediction results. Based on the refining fault prediction results, multi-dimensional risk quantification and evolution path simulation are performed, and a fault prediction report is output.

2. The method for predicting the failure of core components of a coal mill based on vibration analysis according to claim 1, characterized in that, The generated refining fault prediction results include: Based on the preliminary fault prediction distribution, the reinforcement learning feedback iteration loop is initialized, and a multi-objective optimization function is constructed to generate an initial set of optimization parameters. Based on the initial set of optimized parameters, dynamic reward adaptive adjustment is performed, and the reward mechanism is iteratively updated to optimize the transfer learning ensemble prediction model and generate an intermediate optimized model. Based on the intermediate optimization model, the multi-objective optimization function is converged and its parameters are refined to generate refined fault prediction results.

3. The method for predicting the failure of core components of a coal mill based on vibration analysis according to claim 2, characterized in that, The generation of the intermediate optimization model includes: Based on the initial optimized parameter set and the fused vibration dynamic index, initialize the reward evolution and generate the initial reward strategy set; Based on the initial reward strategy group, perform reward evolution iteration and inject fused vibration dynamic indicators to generate an evolutionary reward mechanism; Based on the evolutionary reward mechanism, the parameters of the transfer learning ensemble prediction model are iteratively updated to generate an intermediate optimized model.

4. The method for predicting the failure of core components of a coal mill based on vibration analysis according to claim 2, characterized in that, The output fault prediction report includes: Based on the refining failure prediction results and the intermediate optimization model, dynamic sensitivity multi-dimensional risk quantification is performed, the influence weight of risk indicators is calculated, and a multi-dimensional risk indicator group is generated. Based on the multi-dimensional risk indicator group, a time-dependent evolution path simulation is performed to predict the failure development trajectory and evaluate the probability boundary, thereby generating an evolution path group. Based on the aforementioned evolution path group, risk indicators are fused and the report is visualized to generate a fault prediction report.

5. The method for predicting the failure of core components of a coal mill based on vibration analysis according to claim 2, characterized in that, The method further includes: Based on the probability transition matrix and the initial optimization parameter set, perform map multimodal integration to generate a preliminary integration index set; Based on the preliminary integrated index set, multimodal correlation refinement and uncertainty compensation are performed to generate a refined integrated index set. Based on the refined integrated index set, a closed-loop self-calibration feedback is performed and fed back to the transfer learning integrated prediction model.

Images