Coal mill core component fault prediction method based on vibration analysis

Abstract

The invention discloses a coal mill core component fault prediction method based on vibration analysis, and belongs to the technical field of rotating machinery state monitoring and predictive maintenance, and the method comprises the steps: collecting multi-point vibration signals, and generating an optimization data set; variational mode decomposition and deep embedding are carried out to extract an advanced feature group; establishing a transfer learning model to generate preliminary fault prediction distribution; performing model self-optimization through reinforcement learning feedback circulation to obtain a refining prediction result; and multi-dimensional risk quantification and fault evolution path simulation are carried out to output a decision report. According to the method, a multi-level self-evolution closed-loop intelligent prediction architecture is adopted, online model optimization can be carried out through a reinforcement learning agent based on the uncertainty of prediction distribution, rapid self-adaption to a variable working condition environment is achieved, prospective maintenance decision support is provided in combination with evolution path simulation, and the method has the advantages of being high in adaptability and high in reliability. And the capability of capturing early weak faults of the coal mill, the accuracy of cross-working-condition prediction and the scientificity of final maintenance decision are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of rotating machinery condition monitoring and predictive maintenance technology, and in particular to a method for predicting the failure of core components of a coal mill based on vibration analysis. Background Technology

[0002] Coal mills are key equipment in heavy industries such as thermal power generation, cement and building materials, and chemicals. Their operational stability and reliability directly affect the safety and economic efficiency of the entire production line. To ensure the safe and efficient operation of coal mills, equipment condition monitoring and predictive maintenance based on vibration and other sensor data have become the main technical approach in the industry. By collecting and analyzing the operating data of the core components of the coal mill, real-time monitoring of the equipment's health status can be achieved to a certain extent, which is of great significance for ensuring the continuity of industrial production.

[0003] However, existing methods for predicting coal mill failures still suffer from common technical shortcomings. Traditional monitoring methods based on fixed thresholds can only provide alarms after a failure has occurred, making it difficult to effectively warn of potential failures. Some diagnostic methods rely excessively on the personal experience of maintenance personnel, resulting in high subjectivity and insufficient reliability. Existing data-driven prediction models have limited ability to capture early, subtle fault characteristics and are not adaptable to varying operating conditions, easily leading to false alarms or missed alarms. These problems restrict the effectiveness of predictive maintenance strategies and affect the operational efficiency and safety of the equipment. Summary of the Invention

[0004] To address the aforementioned issues, this invention provides a fault prediction method for core components of a coal mill based on vibration analysis. It employs a multi-level, self-evolving closed-loop intelligent prediction architecture, which, based on the uncertainty of the prediction distribution, optimizes the model online through a reinforcement learning agent. This enables rapid adaptation to changing operating conditions and, combined with evolutionary path simulation, provides forward-looking maintenance decision support. This significantly improves the ability to detect early, minor faults in the coal mill, the accuracy of cross-condition prediction, and the scientific rigor of the final maintenance decisions.

[0005] The above objectives can be achieved through the following approach:

[0006] A method for predicting the failure of core components of a coal mill based on vibration analysis includes: collecting multi-point vibration signal data of the core components of the coal mill, and performing sampling rate adjustment and noise pre-suppression to generate an optimized vibration dataset; based on the optimized vibration dataset, performing variational mode decomposition and deep embedding to extract adaptive mode features and embedding vectors to generate a high-level vibration feature set; based on the high-level vibration feature set, establishing a transfer learning ensemble prediction model and calculating the cross-operating condition failure transfer probability to generate a preliminary failure prediction distribution; based on the preliminary failure prediction distribution, performing reinforcement learning feedback iterative loops to optimize the transfer learning ensemble prediction model and generate refining failure prediction results; based on the refining failure prediction results, performing multi-dimensional risk quantification and evolution path simulation, and outputting a failure prediction report.

[0007] Optionally, generating 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 to generate an adjusted vibration signal set; based on the adjusted vibration signal set, performing a noise pre-suppression loop and iterative filtering to remove noise interference to generate a pre-suppressed vibration dataset; and based on the pre-suppressed vibration dataset, performing data standardization processing to generate an optimized vibration dataset.

[0008] Optionally, generating the advanced vibration feature set includes: performing variable parametric variational mode decomposition based on the optimized vibration dataset, dynamically selecting the number of modes and decomposing non-stationary vibration components to generate a mode decomposition sequence set; optimizing the spatiotemporal relationship between modes using an attention mechanism based on the mode decomposition sequence set to generate an embedding vector set; and performing intelligent mode feature fusion and vector refinement based on the embedding vector set to generate the advanced vibration feature set.

[0009] Optionally, the method further includes: performing spatiotemporal coupling graph modeling based on the adjusted vibration signal group and the modal decomposition sequence group, fusing the spatiotemporal relationship of the signal group and the dynamic characteristics of the modal sequence to generate a preliminary fusion index group; performing multi-dimensional correlation analysis and noise residual compensation based on the preliminary fusion index group, iteratively verifying the spatiotemporal coupling strength and compensation parameters to generate a compensated fusion index group; and performing dynamic threshold iterative optimization based on the compensated fusion index group to quantify the dynamic changes of vibration and refine the index weights to generate a fused vibration dynamic index.

[0010] Optionally, generating the preliminary fault prediction distribution includes: based on the advanced vibration feature set and the fused vibration dynamic index, performing graph transfer learning through domain-adaptive graph structure transfer cross-condition knowledge to generate a transfer feature representation set; based on the transfer feature representation set, constructing a transfer learning integrated prediction model and quantifying the fault state probability path to generate a probability transfer matrix; and based on the probability transfer matrix, refining the fault distribution and quantifying uncertainty to obtain the preliminary fault prediction distribution.

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

[0012] Optionally, generating refined fault prediction results includes: initializing the reinforcement learning feedback iteration loop based on the preliminary fault prediction distribution, constructing a multi-objective optimization function, and generating an initial set of optimization parameters; performing dynamic reward adaptive adjustment based on the initial set of optimization parameters, iteratively updating the reward mechanism to optimize the transfer learning ensemble prediction model, and generating an intermediate optimization model; and performing convergence verification and parameter refinement on the multi-objective optimization function based on the intermediate optimization model to generate refined fault prediction results.

[0013] Optionally, the generation of the intermediate optimization model includes: initializing reward evolution based on the initial optimization parameter set and the fused vibration dynamic index to generate an initial reward strategy set; performing reward evolution iteration based on the initial reward strategy set and injecting the fused vibration dynamic index to generate an evolutionary reward mechanism; and iteratively updating the parameters of the transfer learning ensemble prediction model based on the evolutionary reward mechanism to generate an intermediate optimization model.

[0014] Optionally, the output fault prediction report includes: based on the refined fault prediction results and the intermediate optimization model, performing dynamic sensitivity multi-dimensional risk quantification, calculating the influence weight of risk indicators, and generating a multi-dimensional risk indicator group; based on the multi-dimensional risk indicator group, performing time-dependent evolution path simulation, predicting the fault development trajectory and evaluating the probability boundary, and generating an evolution path group; based on the evolution path group, performing risk indicator fusion and report visualization, and generating a fault prediction report.

[0015] Optionally, the method further includes: performing graph multimodal integration based on the probability transition matrix and the initial optimization parameter set to generate a preliminary integrated index set; performing multimodal association refinement and uncertainty compensation based on the preliminary integrated index set to generate a refined integrated index set; and performing closed-loop self-calibration feedback based on the refined integrated index set to feed back to the transfer learning integrated prediction model.

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

[0017] 1. This invention elevates the traditional data processing model of "passive reception and shallow analysis" to a system capability of "active capture and deep insight" by constructing a collaborative mechanism of "event-driven intelligent sampling" and "physical-data dual-domain deep feature extraction". This mechanism ensures the fidelity of key fault transient information from the data source and, through adaptive decomposition and embedding technology, decouples deep features highly correlated with physical states from complex signals, providing an unprecedented high signal-to-noise ratio decision-making foundation for subsequent accurate prediction;

[0018] 2. This invention, by constructing a multi-loop self-evolving architecture of "transfer learning-reinforcement learning-top-level feedback," fundamentally changes the limitations of traditional prediction models that are "trained once and applied statically." It not only enables real-time self-optimization and strategy adjustment based on intrinsic reward signals through reinforcement learning, but also uses future predictions to guide current data collection strategies, achieving a leap from "passive prediction" to "proactive cognition and forward-looking planning," demonstrating a deep adaptive capability to complex working conditions and unknown patterns.

[0019] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating the fault prediction method for core components of a coal mill based on vibration analysis, according to an embodiment of the present invention.

[0022] Figure 2 This is a graph showing the variable parameter variational mode decomposition process according to an embodiment of the present invention.

[0023] Figure 3 This is a knowledge enhancement network graph for cross-condition graph transfer learning according to an embodiment of the present invention.

[0024] Figure 4 This is a schematic diagram of the fault evolution path simulation and probability boundary in an embodiment of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] Reference Figure 1 One embodiment of the present invention proposes a fault prediction method for core components of a coal mill based on vibration analysis. It adopts a multi-level, self-evolving closed-loop intelligent prediction architecture, which can optimize the model online through reinforcement learning agents based on the uncertainty of the prediction distribution, achieve rapid adaptation to changing working conditions, and provide forward-looking maintenance decision support by combining evolution path simulation. This significantly improves the ability to capture early minor faults of the coal mill, the accuracy of cross-working condition prediction, and the scientific nature of the final maintenance decision.

[0027] The method described in this embodiment specifically includes:

[0028] Multi-point vibration signal data of the core components of the coal mill are collected, and sampling rate adjustment and noise pre-suppression are performed to generate an optimized vibration dataset;

[0029] Based on the optimized vibration dataset, variational mode decomposition and deep embedding are performed to extract adaptive mode features and embedding vectors, generating a high-level vibration feature set.

[0030] Based on the aforementioned advanced vibration feature set, a transfer learning integrated prediction model is established, and the cross-condition fault transfer probability is calculated to generate a preliminary fault prediction distribution.

[0031] 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.

[0032] Based on the refining fault prediction results, multi-dimensional risk quantification and evolution path simulation are performed, and a fault prediction report is output.

[0033] A multi-level, self-evolving closed-loop intelligent prediction architecture is adopted, which can optimize the model online through reinforcement learning agents based on the uncertainty of the prediction distribution, achieve rapid adaptation to changing operating conditions, and provide forward-looking maintenance decision support by combining evolution path simulation. This significantly improves the ability to detect early minor faults in coal mills, the accuracy of cross-operating condition prediction, and the scientific nature of final maintenance decisions.

[0034] Optionally, the generation of the optimized vibration dataset includes:

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

[0036] Specifically, this step aims to obtain the most raw equipment status information from the physical world. Multiple high-frequency piezoelectric accelerometers are deployed in key components of the coal mill, such as the main bearing housing, grinding roller rocker arm, and gearbox housing. These sensors continuously collect vibration acceleration signals from each measuring point at a fixed base sampling frequency several times higher than the highest analysis frequency. These unprocessed, multi-channel signal data, collected under the same time reference, are then aggregated to form an initial vibration signal set.

[0037] Based on the initial vibration signal set, adaptive sampling rate optimization is performed to adjust the sampling frequency to match the changes in the vibration spectrum and generate an adjusted vibration signal set.

[0038] Specifically, this step aims to intelligently allocate acquisition resources to minimize data redundancy while ensuring information integrity. In this process, a windowed short-time Fourier transform (STFT) is performed on the initial vibration signal group to obtain a series of time spectra. For each frame's time spectrum, its spectral entropy is calculated to quantify the complexity of the signal within that time window. (Spectral entropy) The calculation formula can be:

[0039] ,

[0040] in, Indicates a time window. Indicates the first One frequency component, This represents the total number of frequency components. In the time window Internal and frequency components The normalized power spectral density is calculated. A higher spectral entropy value indicates a richer frequency composition of the signal. Subsequently, a feedback control logic is established, using the calculated spectral entropy as a control input to dynamically adjust the sampling frequency of the next acquisition cycle. For example, when the spectral entropy exceeds a preset high complexity threshold, the sampling frequency is increased to the highest diagnostic level, thereby generating an adjusted vibration signal set.

[0041] Based on the adjusted vibration signal group, a noise pre-suppression loop is executed, and iterative filtering operations are performed to remove noise interference and generate a pre-suppressed vibration dataset.

[0042] Specifically, this step aims to separate weak but crucial fault characteristic signals from a strong noise background. The loop employs an iterative signal reconstruction and residual evaluation strategy. In each iteration, the signal is processed by an adaptive filter whose internal parameters are adjusted based on the signal residuals generated in the previous iteration. This process continues until the separation between the signal and noise components reaches its optimal level, or the energy of the signal residuals converges below a preset stability threshold. The resulting signal is the pre-suppressed vibration dataset.

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

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

[0045] Optionally, the generation of the advanced vibration feature set includes:

[0046] Based on the optimized vibration dataset, variable parametric variational mode decomposition is performed, the number of modes is dynamically selected and non-stationary vibration components are decomposed, and a mode decomposition sequence group is generated.

[0047] Specifically, this step aims to decompose the complex non-stationary signal coupled with various fault information and operating noise contained in the optimized vibration dataset into a series of simpler and more physically interpretable intrinsic mode functions (EMFs). Unlike traditional variational mode decomposition (VMD), which requires manually setting the number of modes in advance, this step employs 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 signal's center frequency, kurtosis, or energy. This allows for more accurate decomposition of vibration components with different characteristic frequencies related to different parts such as bearings, gears, and rotors. The resulting time series of multiple EMFs together constitute a mode decomposition sequence set, such as... Figure 2 As shown in the figure, under the same coordinate system, a complex original signal is adaptively decomposed into multiple simpler eigenmode functions with different physical meanings by the variable parameter variational mode decomposition method of the present invention, intuitively demonstrating the process of decoupling the core dynamic components from complex coupled signals.

[0048] Based on the modality decomposition sequence set, the spatiotemporal relationship between modalities is optimized using an attention mechanism to generate an embedding vector set.

[0049] Specifically, this step aims to deeply explore and quantify the complex intrinsic relationships between different vibration modes. Since equipment failure evolution is often the result of the interaction of multiple physical components, there are coupling relationships between different modal sequences in both time and space dimensions. In this step, an attention mechanism is introduced, which, like a human expert, automatically assigns "attention weights" to different modes at different times. For example, this mechanism can learn that when the energy of a specific high-frequency mode related to a bearing increases, more attention should be paid to changes in another low-frequency mode related to structural resonance. In this way, the original, independent modal decomposition sequence set is transformed into a fixed-length low-dimensional vector that reflects the dynamic correlation between modes—that is, an embedded vector set.

[0050] Based on the embedded vector set, intelligent modal feature fusion and vector refinement are performed to generate advanced vibration feature groups.

[0051] Specifically, this step aims to construct the final, information-dense feature representation to drive subsequent prediction models. First, statistical features such as energy and entropy are extracted from each sequence in the modal decomposition sequence set to form a physical feature set. Then, this physical feature set is concatenated and fused with the set of embedding vectors generated in the previous step, representing abstract relationships between modes. To further refine information and eliminate redundancy, the concatenated and fused vectors are input into the encoder of a deep autoencoder for vector refinement. Through nonlinear mapping, they are compressed into a more compact and representative final form, thereby generating a high-level vibration feature set.

[0052] Optionally, the method further includes:

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

[0054] Specifically, this step aims to construct a parallel analysis path to directly extract dynamic and instability indicators of device operation from the raw signal and decomposed modes. In this process, a spatiotemporal coupling graph is constructed, where the nodes represent each physical sensor, and the edge weights between nodes are used to quantify the coupling strength between signals from different sensors. These edge weights... The calculation integrates spatial distance and temporal synchronization, and its formula can be:

[0055] ,

[0056] in, It is a sensor and The physical distance between them It is a distance scaling factor, and the first term is used to characterize spatial coupling relationships. In the time window The phase-locked value between the modal decomposition sequence groups corresponding to the two sensors is used to characterize the tightness of temporal coupling. The constructed spatiotemporal coupling graph is input into a graph neural network (GNN), and the spatiotemporal information of neighboring nodes is aggregated through graph convolution operations. Finally, the embedding representation output by each node constitutes the preliminary fusion index group.

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

[0058] Specifically, this step aims to refine and calibrate the initial graphical model output. In this process, cross-correlation analysis is first performed on the various indices within the initial fusion index group to assess the consistency of dynamic characteristics at different physical locations. Simultaneously, noise residual compensation is performed. This compensation calculates the difference between the total energy of the adjusted vibration signal group and the sum of the energies of all sequences in the mode decomposition sequence group, resulting in a residual energy that cannot be explained by any mode. This residual energy primarily corresponds to random noise or nonlinear coupling components. This residual energy is then used to correct the initial fusion index group; for example, when the residual energy is high, the confidence level of the indices is appropriately reduced, thereby generating a compensated fusion index group.

[0059] Based on the aforementioned compensation fusion index group, dynamic threshold iterative optimization is performed to quantify the dynamic changes in vibration and refine the index weights, thereby generating a fused vibration dynamic index.

[0060] Specifically, this step aims to generate a single, final indicator that sensitively reflects the dynamic changes in the equipment's operating status. In this process, a sliding window-based statistical method is used to continuously calculate the mean and standard deviation of the compensation fusion indicator group over a past period. Based on this statistical result, a dynamically changing adaptive warning threshold is generated. "Iterative optimization" is reflected in the continuous sliding of this statistical window, allowing the threshold to automatically adapt to normal fluctuations in the equipment under different steady-state processes. Finally, the normalized and smoothed compensation fusion indicator group is output as a fused vibration dynamic indicator, which can be directly compared with the adaptive warning threshold to assess the current operational instability of the equipment.

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

[0062] Based on the advanced vibration feature set and the fused vibration dynamic index, graph transfer learning is performed through domain-adaptive graph structure transfer cross-condition knowledge to generate a transfer feature representation set.

[0063] Specifically, this step aims to address the core technical challenge of traditional models predicting significant performance degradation when coal mill operating conditions change. In this process, two graph structures are first constructed: one is a "source domain feature association graph" built upon a massive general vibration database, containing universal fault mechanism knowledge; the other is a "target domain feature association graph" built upon real-time generated advanced vibration feature sets and fused vibration dynamic indicators from the current coal mill. The core of graph transfer learning lies in using the topological structure and knowledge of the "source domain feature association graph" as prior information, guiding and enhancing the construction of the "target domain feature association graph" through a domain adaptation module, particularly by completing and strengthening weak associations in the target domain caused by data sparsity. After this cross-condition knowledge transfer, the node representations extracted from the enhanced target domain graph structure are the transfer feature representation set. This representation set is characterized by robustness to different operating conditions, such as… Figure 3 As shown in the figure, the fully connected structure of the knowledge-complete "source domain" is represented by gray dashed lines superimposed on the same set of feature nodes, while the sparse structure of the "target domain" driven by local data is highlighted by black solid lines. This demonstrates the core idea of ​​graph transfer learning to enhance the graph representation of the target domain by utilizing prior knowledge.

[0064] Based on the aforementioned set of transfer feature representations, a transfer learning ensemble prediction model is constructed, and the probability path of the fault state is quantified to generate a probability transition matrix.

[0065] Specifically, this step aims to model the device degradation process as a multi-state stochastic process with explicit physical meaning. In this step, a set of implicit health states corresponding to the physical health of core components is predefined, such as "healthy," "early degradation," "significant degradation," and "imminent failure." An ensemble predictive model, such as a Hidden Markov Model or a variant, is constructed, with these implicit health states as internal states and a set of transition feature representations as external observations. Through training, this model learns to infer the most likely implicit health state of the device from the observable set of transition feature representations, and, more importantly, quantifies the probability of transitioning from one implicit health state to another per unit time. The transition probabilities between all states are organized into a matrix form, namely the probability transition matrix.

[0066] Based on the probability transition matrix, the fault distribution is refined and the uncertainty is quantified to obtain a preliminary fault prediction distribution.

[0067] Specifically, this step aims to generate a dynamic and complete predictive distribution that reflects future probabilities and uncertainties from a static probability transition matrix. This process employs probabilistic programming techniques such as Monte Carlo sampling or variational inference. Based on the probability transition matrix, thousands or even tens of thousands of possible future fault evolution paths are simulated. Statistical analysis of these massive simulated paths not only yields the most probable fault development trajectory but, more importantly, quantifies the uncertainty of the prediction, for example, by calculating the dispersion or variance of all simulated paths at various future time points. Finally, the complete probabilistic information, including the most probable evolution path, probability confidence intervals, and uncertainty indicators, is integrated to form a preliminary fault prediction distribution.

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

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

[0070] Specifically, this step aims to transform the static, one-step possibilities described by the probability transition matrix into a dynamic, multi-step, and visualized set of future evolutionary trajectories. In this process, sampling techniques such as Markov Chain Monte Carlo (MCMC) are employed. Starting from the currently inferred device health state, random sampling is performed based on the transition probabilities defined in the probability transition matrix to determine the health state at the next moment. By independently repeating this process tens of thousands of times, a random fault path sample set containing a large number of possible future evolutionary trajectories is ultimately 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, generating a refined distribution parameter set.

[0072] Specifically, this step aims to extract a continuous, parameterized mathematical model from a massive amount of discrete random path samples that can describe the overall distribution of future states. In this process, variational inference (VI) is employed. The core idea of ​​this technique is to use a simple approximate probability distribution defined by a set of parameters. To approximate the true but difficult-to-compute posterior distribution implicit in the sample set of random fault paths. This approximation process is achieved by maximizing an objective function called the "Evidence Lower Bound (ELBO)," which can be expressed as:

[0073] ,

[0074] in, Potential variables representing fault paths, Representing observed evidence, It is an approximate distribution defined by a set of distribution parameters to be optimized. It is a joint probability distribution. The algorithm is iteratively adjusted through optimization. The parameters are optimized to maximize the lower bound of evidence, and the resulting optimal set of parameters is the refined distribution parameter set. The variance term in this parameter set directly constitutes the quantification of the uncertainty boundary of the prediction.

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

[0076] Specifically, this step aims to integrate and encapsulate the calculation results from the previous steps, forming a structured and comprehensive output. During this process, the refined distribution parameter set is analyzed to extract the expected trajectory representing the "most likely" evolutionary path, and the confidence interval representing the uncertainty boundary. This expected trajectory and confidence interval are then fused with a sample set of representative random failure paths to form a complete probabilistic picture that includes both parametric statistical descriptions and non-parametric sample examples; this picture constitutes the preliminary failure prediction distribution.

[0077] Optionally, the generation of refining fault prediction results includes:

[0078] 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.

[0079] Specifically, this step aims to establish a higher-dimensional, self-optimizing intelligent feedback module. In this process, a reinforcement learning (RL) framework is constructed, where the state of the reinforcement learning agent is defined by the initial fault prediction distribution and its uncertainty index; the agent's action space is defined as a set of hyperparameters or structural parameters adjusted for the transfer learning ensemble prediction model. Simultaneously, a multi-objective optimization function is constructed, containing at least two mutually constraining objectives: maximizing prediction accuracy and minimizing prediction uncertainty. The initial set of optimization parameters constitutes the initial weight parameters of the reinforcement learning agent's policy network.

[0080] 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.

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

[0082] ,

[0083] in, Is Total reward value at each moment This refers to the "forecast refining benefit" associated with reducing forecast uncertainty. This is the "anomaly detection benefit" related to improving the model's sensitivity to mutation signals. The key lies in the weighting coefficients. It is a dynamic variable, and its value is determined by the real-time value of the integrated vibration dynamic index. For example, when the equipment is operating smoothly, the value of this index is lower. As the value increases, the agent becomes more focused on "refinement"; when the equipment is unstable, this value is higher. This reduction allows the agent to focus more on "seeking change." Through this dynamic reward adaptive adjustment mechanism, the agent continuously generates intermediate optimized models during the iterative optimization process.

[0084] Based on the intermediate optimization model, the multi-objective optimization function is converged and its parameters are refined to generate refined fault prediction results.

[0085] 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 iterations, it enters the convergence verification phase. In this phase, the value of the multi-objective optimization function is monitored over several consecutive iterations. When the fluctuation of its value is less than a preset convergence threshold, the optimization process is considered to have converged. Subsequently, from the multiple intermediate optimization models in the convergence phase, the model with the best overall performance on the multi-objective optimization function is selected as the final refined model. The latest advanced vibration feature set is input into this final refined model, and its output is the refined fault prediction result.

[0086] Optionally, the generation of the intermediate optimization model includes:

[0087] Based on the initial optimized parameter set and the fused vibration dynamic index, initialize the reward evolution and generate the initial reward strategy set;

[0088] Specifically, this step aims to evolve the "reward mechanism" itself in reinforcement learning from a static, pre-defined function into a dynamic, optimizable agent. In this process, a "population" containing multiple potential reward strategies—the initial reward strategy set—is initialized. Each reward strategy is an independent candidate, defining how the reward value is calculated based on the initial fault prediction distribution and uncertainty. For example, the population may include strategies that focus on penalizing long-term risk, strategies that focus on rewarding short-term prediction stability, and multiple strategies with different tolerances to uncertainty. The fused vibration dynamics index is used at this stage as a key input to the environmental state to evaluate the initial fitness of different strategies under the current device dynamics.

[0089] Based on the initial reward strategy group, perform reward evolution iteration and inject fused vibration dynamic indicators to generate an evolutionary reward mechanism;

[0090] Specifically, this step aims to select and cultivate the optimal reward strategy from the initial population by simulating an evolutionary process of "survival of the fittest." In this process, an iterative evolutionary algorithm, such as a genetic algorithm or evolutionary strategy, is executed. In each iteration, the "fitness" of each reward strategy in the population is evaluated, which measures its performance in guiding a reinforcement learning agent to optimize a transfer learning ensemble prediction model. Here, fused vibrational dynamics metrics are continuously "injected" into the fitness evaluation function as a dynamically changing environmental variable. This means that an excellent reward strategy should not only perform well when the equipment is stable but also guide the agent to make correct optimizations when the equipment's dynamics metrics increase. Through multiple generations of genetic, crossover, and mutation operations, the reward strategy that ultimately survives and has the strongest fitness is the evolutionary reward mechanism.

[0091] Based on the evolutionary reward mechanism, the parameters of the transfer learning ensemble prediction model are iteratively updated to generate an intermediate optimized model.

[0092] Specifically, this step aims to apply the final results of the preceding evolutionary process to perform the final, most efficient optimization of the master prediction model. In this process, the evolutionary reward mechanism is established as the immutable "gold standard" in the reinforcement learning feedback iterative loop. Using this well-tested reward mechanism, the network parameters or structural parameters of the transfer learning ensemble prediction model are ultimately updated in a targeted manner. Since the reward mechanism itself is already optimal, the optimization process here converges faster and yields more stable results; the final model obtained is the intermediate optimized model.

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

[0094] 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.

[0095] Specifically, this step aims to deeply analyze and quantify the fundamental driving factors leading to the current failure risk. In this process, a model-independent feature attribution method, such as the Shapley Additive Explanations (SHAP) algorithm, is employed to perform dynamic sensitivity multidimensional risk quantification. This method treats each input feature as a "contributor" and calculates its marginal contribution to the refined failure prediction result output by the intermediate optimization model. This method allows for the precise calculation of the contribution of each input feature to the final predicted risk. Combining the risk index influence weights of all features with the overall risk probability generates a multidimensional risk index set.

[0096] 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.

[0097] Specifically, this step aims to prospectively extrapolate possible future failure development paths based on the current risk status and driving factors. Unlike simulations based on static transition matrices, this time-dependent evolution path simulation utilizes the time-series memory capability of the intermediate optimization model itself. Using the current high-level vibration feature set as initial input, the model iteratively performs multi-step predictions by using the predicted output of one time step as part of the input for the next time step through autoregression. Simultaneously, random perturbations conforming to the initial failure prediction distribution are introduced into each prediction step. Through thousands of independent simulations, a set of future development trajectories containing multiple possibilities and conforming to time-series dependencies is ultimately generated—this is the evolution path set, such as... Figure 4 As shown in the figure, the future failure evolution path simulated based on the current equipment state includes the most likely degradation trajectory calculated from the refined failure prediction results, as well as the confidence interval formed by the uncertainty boundary, providing intuitive data support for forward-looking maintenance decisions.

[0098] Based on the aforementioned evolution path group, risk indicators are fused and the report is visualized to generate a fault prediction report.

[0099] Specifically, this step aims to present all analysis results in a highly user-friendly and information-rich manner for operations and maintenance personnel. During this process, the multi-dimensional risk indicator groups and evolution path groups generated in previous steps are deeply integrated. The final generated fault prediction report is a visualized, multi-level, interactive report that may include: a risk dashboard displaying the overall fault probability, uncertainty boundaries, and risk level; a view showing the contribution of risk factors, sorted by waterfall chart or contribution, clearly illustrating which specific characteristic indicators are currently causing increased risk; and a fault trajectory prediction view, displaying the most likely future fault evolution paths and probability boundaries, using pie charts or trajectory bundles.

[0100] Optionally, the method further includes:

[0101] Based on the probability transition matrix and the initial optimization parameter set, perform map multimodal integration to generate a preliminary integration index set;

[0102] Specifically, this step aims to construct a higher-dimensional self-calibration module for evaluating the consistency between "model dynamics" and "optimization strategy." In this process, the probability transition matrix is ​​viewed as an adjacency matrix of a directed weighted graph describing the evolution of the device's health state space. Simultaneously, the reinforcement learning initial strategy defined by the initial set of optimization parameters is considered a Markov decision process making decisions on this state graph. Graph multimodal integration analyzes the consistency between the state transition path induced by this decision process and the intrinsic physical transition laws defined by the probability transition matrix. This consistency, or "harmony," can be quantified by a metric function, which can be:

[0103] ,

[0104] in, It is a probability transition matrix. The reinforcement learning strategy is defined by the initial set of optimized parameters. The policy-induced transition matrix generated during execution in the simulation environment The Frobenius norm represents the difference between two matrices. A preliminary integrated index set is generated through a comprehensive analysis of this "harmony" and the spectral properties of the graph.

[0105] Based on the preliminary integrated index set, multimodal correlation refinement and uncertainty compensation are performed to generate a refined integrated index set.

[0106] Specifically, this step aims to improve the robustness and reliability of the initial integrated index 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 corresponding transition probability itself has extremely high uncertainty, then the weight of that disharmony is reduced. In this way, all initial integrated indices undergo an uncertainty-based weighted correction, thereby avoiding incorrect calibrations due to insufficient model awareness of certain rare operating conditions, ultimately generating a refined integrated index set.

[0107] 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.

[0108] Specifically, this step is the closed-loop link in achieving the entire top-level self-calibration. The refined ensemble index set is ultimately transformed into a calibration signal and fed back to the transfer learning ensemble prediction model. This calibration signal directly affects the regularization term or loss function of the transfer learning ensemble prediction model. For example, when the refined ensemble index set shows a low "harmony," it indicates that the optimization objective of reinforcement learning may conflict with the model's intrinsic cognition. In this case, the calibration signal will increase the regularization coefficient of the transfer learning ensemble prediction model, forcing the model to learn more general and universal features, thus avoiding it from getting trapped in a locally optimal state that, while optimized by reinforcement learning, is physically unreasonable. This closed-loop self-calibration feedback mechanism ensures the long-term stability of the entire prediction and its fidelity to physical reality.

[0109] To verify the feasibility and advancement of this invention, it was applied to the No. 3 MPS medium-speed coal mill in a 600MW supercritical thermal power plant. This coal mill has historically experienced unplanned shutdowns due to progressive wear of core components and sudden coal blockages, severely impacting unit efficiency. This embodiment aims to use the method of this invention to proactively predict the health status of the coal mill's core components.

[0110] At 10:42 AM on March 15, 2024, when a fluctuation in coal feed caused a brief transient impact, the vibration spectrum change was identified, and event-driven high-fidelity sampling was immediately triggered, instantly increasing the sampling rate of the relevant sensors to 20kHz, successfully capturing an optimized vibration dataset containing rich details. Subsequently, parametric variational mode decomposition and attention mechanisms were performed on the dataset, identifying a weak high-frequency mode related to the grinding roller impact and generating a high-level vibration feature set.

[0111] Meanwhile, a fusion vibration dynamic index was continuously output. This index began to rise steadily after the transient impact event, indicating a declining trend in the overall operational stability of the equipment. Combining the advanced vibration feature set and this dynamic index, and utilizing source domain knowledge learned from other similar coal mills, a preliminary fault prediction distribution was generated. This distribution, quantized through probability transition matrix and variational inference optimization, shows a 75% probability that the equipment will enter an "early degradation state" within the next 200 hours, but with high uncertainty.

[0112] At this point, the reinforcement learning feedback iterative loop was activated. Due to the high fusion vibration dynamic index, the reward function was dynamically adjusted, tilting the optimization objective towards "abnormal state exploration gains," ultimately generating refined fault prediction results. The final fault prediction report was output at 14:00 that afternoon. This report not only concluded that "early spalling of the grinding roller surface has a high probability of occurring within 150 hours," but also, through dynamic sensitivity quantification, indicated that the dominant risk factor is the aforementioned weak high-frequency impact mode, and its evolution path simulation showed 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 was constructed, and a data acquisition strategy optimization instruction was generated and fed back to the data acquisition module. In subsequent operation, this instruction proactively improved the sampling resolution of sensors 2 and 4, which were most relevant to the grinding roller's condition. Ultimately, approximately 145 hours later, the power plant conducted proactive maintenance based on this prediction report, confirming early pitting defects on the grinding roller surface that were highly consistent with the prediction, successfully averting an unplanned downtime.

[0114] Table 1. Data Table of Intelligent Acquisition and Optimization Effects of Coal Mill Vibration Data

[0115]

[0116] Table 2 Comparison 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] As can be seen from the data recorded in Tables 1-3 above, the present invention performs exceptionally well in the embodiments. Table 1 verifies the effectiveness of the adaptive sampling strategy in capturing key transient information. Table 2, through comparison with various benchmark prediction models, demonstrates the significant advantages of the present invention in key performance indicators such as prediction accuracy, lead time, and false alarm rate. Table 3 records the entire process from the initial generation of the prediction report to the final manual verification; detailed data clearly demonstrates the accuracy of the prediction results and the significant engineering application value of the final decision report.

[0121] It should be noted that the formulas described above, through the principle of dimensional consistency and mathematical standardization methods (such as normalization, dimensionless parameter conversion, or unit system unification), can translate physical quantities with different properties into unitless standard values ​​or parameters that can be superimposed in the same dimension. This eliminates the interference of different dimensions on the computational logic, allowing the formulas to retain the original data distribution characteristics while possessing mathematical rationality and adaptability to objective laws. These are conventional technical methods and will not be elaborated further. The electrical connections between the various units described above do not necessarily represent direct or indirect connections; any indirect connection method is applicable to the embodiments of this invention as long as it achieves the purpose of this invention. The above descriptions are merely exemplary embodiments of this invention and should not be construed as limiting the scope of this invention.

[0122] All equivalent changes and modifications made in accordance with the teachings of this invention are still within the scope of this invention. Those skilled in the art will readily conceive of other embodiments of this invention upon considering the specification and the disclosure of practical truth. This application is intended to cover any variations, uses, or adaptations of this invention that follow the general principles of this invention and include common knowledge or conventional techniques in the art not described herein.

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 the coal mill are collected, and sampling rate adjustment and noise pre-suppression are performed to generate an optimized vibration dataset; Based on the optimized vibration dataset, variational mode decomposition and deep embedding are performed to extract adaptive mode features and embedding vectors, generating a high-level vibration feature set. Based on the aforementioned advanced vibration feature set, a transfer learning integrated prediction model is established, and the cross-condition fault transfer probability is calculated to generate 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 optimized vibration dataset includes: Collect 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, adaptive sampling rate optimization is performed to adjust the sampling frequency to match the changes in the vibration spectrum and generate an adjusted vibration signal set. Based on the adjusted vibration signal group, a noise pre-suppression loop is executed, and iterative filtering operations are performed to remove noise interference and generate a pre-suppressed vibration dataset. Based on the pre-suppressed vibration dataset, data standardization is performed to generate an optimized vibration dataset.

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 generated high-level vibration feature set includes: Based on the optimized vibration dataset, variable parametric variational mode decomposition is performed, the number of modes is dynamically selected and non-stationary vibration components are decomposed, and a mode decomposition sequence group is generated. Based on the modality decomposition sequence set, the spatiotemporal relationship between modalities is optimized using an attention mechanism to generate an embedding vector set. Based on the embedded vector set, intelligent modal feature fusion and vector refinement are performed to generate advanced vibration feature groups.

4. The method for predicting the failure of core components of a coal mill based on vibration analysis according to claim 3, characterized in that, The method further includes: Based on the adjusted vibration signal group and the modal decomposition sequence group, spatiotemporal coupling diagram modeling is performed to fuse the spatiotemporal 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 set, multi-dimensional correlation analysis and noise residual compensation are performed, the spatiotemporal coupling strength and compensation parameters are iteratively verified, and a compensation fusion index set is generated. Based on the aforementioned compensation fusion index group, dynamic threshold iterative optimization is performed to quantify the dynamic changes in vibration and refine the index weights, thereby generating a fused vibration dynamic index.

5. The method for predicting the failure of core components of a coal mill based on vibration analysis according to claim 4, characterized in that, The generation of the preliminary fault prediction distribution includes: Based on the advanced vibration feature set and the fused vibration dynamic index, graph transfer learning is performed through domain-adaptive graph structure transfer cross-condition knowledge to generate a transfer feature representation set. Based on the aforementioned set of transfer feature representations, a transfer learning ensemble prediction model is constructed, and the probability path of the fault state is quantified to generate a probability transition matrix. Based on the probability transition matrix, the fault distribution is refined and the uncertainty is quantified to obtain a preliminary fault prediction distribution.

6. The method for predicting the failure of core components of a coal mill based on vibration analysis according to claim 5, characterized in that, The obtained preliminary fault prediction distribution includes: Based on the probability transition matrix, a sampling simulation is performed to generate a random fault path sample group; Based on the random fault path sample set, variational inference optimization is performed to quantify the uncertainty boundary and refine the distribution parameters, generating a refined distribution parameter set. Based on the refined distribution parameter set, the path samples and uncertainty indicators are fused to obtain the preliminary fault prediction distribution.

7. The method for predicting the failure of core components of a coal mill based on vibration analysis according to claim 6, 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.

8. The method for predicting the failure of core components of a coal mill based on vibration analysis according to claim 7, 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.

9. The method for predicting the failure of core components of a coal mill based on vibration analysis according to claim 7, 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.

10. The method for predicting the failure of core components of a coal mill based on vibration analysis according to claim 7, 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