Deep learning-based intelligent predictive maintenance system and method for mine equipment

Abstract

This invention discloses a deep learning-based intelligent predictive maintenance system and method for mining equipment, belonging to the field of industrial maintenance technology. The system includes a data acquisition and synchronization module, a physical preprocessing module, a feature extraction module, a physical modeling module, an uncertainty quantification module, an intelligent early warning module, a human-machine collaboration module, and an online optimization module, connected sequentially. The method is executed through the above system. This invention achieves a deep integration of data-driven approaches and physical mechanisms, solving the problem of superficial integration of physical constraints in existing methods. Simultaneously, through multi-source data synchronization and simulation enhancement, multi-scale and graph structure feature extraction, uncertainty quantification and risk assessment, human-machine knowledge interaction, and online model optimization, it improves the accuracy and reliability of fault prediction in critical degradation stages and extrapolated operating conditions, and provides a scientific basis for maintenance decisions based on risk quantification.

Description

Technical Field

[0001] This invention relates to the field of industrial maintenance technology, specifically to an intelligent predictive maintenance system and method for mining equipment based on deep learning. Background Technology

[0002] In the process of industrial intelligentization, predictive maintenance has become an important means to ensure the safe and stable operation of critical equipment, especially in high-risk industries where its value is even more prominent. The mining industry has a wide variety of equipment and operates in harsh environments. Large equipment such as excavators, crushers, and hoisting systems are subjected to high-intensity loads for extended periods. Equipment failures not only cause huge economic losses but can also threaten personnel safety. In recent years, with the rapid development of deep learning technology, data-driven intelligent predictive maintenance systems have gradually been applied. These systems typically collect operational data by deploying sensors such as vibration, temperature, and pressure sensors on the equipment and use neural network models to analyze the equipment's health status, thereby predicting the timing of failures and providing a basis for maintenance planning.

[0003] Currently, these deep learning-based predictive maintenance systems exhibit a fundamental limitation in practical engineering applications: the integration of physical mechanisms and data-driven models often remains superficial. Most systems employ a purely data-driven, end-to-end approach, relying on neural networks to learn failure modes directly from sensor data. Even when some studies attempt to introduce physical knowledge as constraints, they typically only add penalty terms related to physical laws during model training, failing to deeply embed the physical mechanisms describing the intrinsic processes of equipment failure into the model's structural design. The failure mechanisms of large mining equipment are essentially dynamic processes involving the coupling of multiple physics fields. For example, bearing degradation involves a cyclical feedback loop of frictional heating, temperature rise leading to material property evolution, thermal stress inducing deformation, and deformation further exacerbating wear. However, existing models, for the sake of reducing computational complexity, generally oversimplify these complex interactions, either treating them as independent or modeling them as simple unidirectional causal relationships. The direct consequence of this simplification is the loss of key nonlinear feedback effects and state abrupt change characteristics in the physical evolution process, leading to significant deviations in the model's predictive performance and decreased reliability when the equipment enters a rapid degradation phase. Summary of the Invention

[0004] This invention proposes a deep learning-based intelligent predictive maintenance system and method for mining equipment. Its objectives are: 1. To achieve deep integration of physical mechanisms and data-driven models, embedding the physical laws describing the inherent dynamic process of equipment failure deep into the model's structural design, rather than merely using them as shallow constraints during the training phase. 2. To accurately capture and model the multi-physics coupling, nonlinear feedback effects, and abrupt state changes involved in the failure process of mining equipment, thereby improving the model's predictive reliability and accuracy during critical stages such as rapid equipment degradation.

[0005] The technical solution of this invention is as follows:

[0006] A deep learning-based intelligent predictive maintenance system for mining equipment includes, in sequence, a data acquisition and synchronization module, a physical preprocessing module, a feature extraction module, a physical modeling module, an uncertainty quantification module, an intelligent early warning module, a human-machine collaboration module, and an online optimization module; wherein:

[0007] The data acquisition and synchronization module obtains operational data from vibration, temperature, strain, and acoustic emission sensors and performs timestamp alignment processing.

[0008] The physical preprocessing module detects outliers based on the physical constraint boundaries of the equipment, generates virtual fault samples using simulation, and reconstructs the internal physical field distribution.

[0009] The feature extraction module uses a multi-branch convolutional neural network to extract features at different time scales and constructs a graph neural network to infer the associated states of components.

[0010] The physical modeling module design encodes conservation laws and constitutive relations into a differentiable network layer in the physical constraint layer, and adopts a dual-path structure to fuse analytical models and neural networks.

[0011] The uncertainty quantification module assesses cognitive uncertainty through variational inference and establishes an out-of-distribution detector and a multi-layer propagation mechanism.

[0012] The intelligent early warning module comprehensively calculates the risk index and establishes a four-level early warning mechanism, and visualizes the equipment status through a digital twin interface;

[0013] The human-machine collaboration module provides a knowledge input interface and an interpretable reasoning path, and establishes a conversational inquiry mechanism;

[0014] The online optimization module monitors performance metrics in real time, uses incremental learning to update the model, and integrates cross-device fault cases to achieve knowledge reuse.

[0015] As a further improvement to the deep learning-based intelligent predictive maintenance system for mining equipment, the data acquisition and synchronization module includes a multi-source data acquisition unit and a time-series synchronization processing unit.

[0016] The multi-source data acquisition unit collects operating data from the equipment's vibration sensors, temperature sensors, strain sensors, acoustic emission sensors, and oil analysis sensors. It also obtains load values, speed values, and pressure values ​​from the control system. The unit performs multi-rate acquisition for different sampling frequencies and adds timestamps. Through signal amplitude checks, sampling rate analysis, and continuity judgment, it detects sensor drift, faults, or missing data, and generates quality markers and alarm information.

[0017] The timing synchronization processing unit receives multi-source data streams with timestamps and quality label information, establishes a unified time reference, downsamples high-frequency data or interpolates low-frequency data, maps the measured values ​​to a unified time node, identifies and isolates abnormal data, generates a timing-synchronized multi-source fusion dataset, and outputs it to the physical preprocessing module.

[0018] As a further improvement to the deep learning-based intelligent predictive maintenance system for mining equipment, the physical preprocessing module includes a constraint detection unit and a simulation enhancement unit.

[0019] The constraint detection unit receives the multi-source fusion dataset output by the data acquisition synchronization module, detects abnormal values ​​based on the temperature change rate threshold, vibration amplitude limit, stress-strain relationship range, and material performance parameter boundary, distinguishes between real extreme working conditions and sensor errors and marks them, and solves the heat conduction equation and elasticity equation based on sparse measurement point data, geometric structure model and material property parameters to invert the internal temperature field and stress field distribution, and outputs the detected measured data and reconstructed physical field data.

[0020] The simulation enhancement unit receives the measured data and physical field data, calls the finite element model and multibody dynamics model, changes the load spectrum, boundary conditions, material properties and environmental condition parameters to generate virtual sensor response data, simulates the occurrence and evolution of faults for scarce fault samples to generate virtual fault data, mixes them to form an enhanced dataset, normalizes it according to physical dimensions and coupling relationships, considers causal relationships and coupling strength weights, and outputs it to the feature extraction module.

[0021] As a further improvement to the deep learning-based intelligent predictive maintenance system for mining equipment, the feature extraction module includes a multi-scale extraction unit and a graph reasoning unit.

[0022] The multi-scale extraction unit receives the normalized enhanced dataset output by the physical preprocessing module, and uses a multi-branch convolutional neural network to process the time series data in parallel. The short-time branch, medium-time branch, and long-time branch use convolutional kernels of increasing size to extract transient impact features, periodic temperature change features, and long-term fatigue accumulation features, respectively. Each branch independently completes convolution and pooling operations and outputs corresponding feature vectors. The three types of feature vectors are concatenated in the feature dimension to form a multi-scale fused feature vector.

[0023] The graph inference unit receives the multi-scale fused feature vector, constructs a graph neural network model to represent the device topology, with nodes corresponding to components and edges representing physical connections and energy transfer paths. The fused feature vector is used as the initial feature input for the nodes, and the feature information is propagated through graph convolution operations to infer the health status of adjacent components. The graph feature representation is output, and an adaptive attention mechanism is introduced to calculate weight coefficients based on degradation stage parameters. Features at different scales are weighted and fused with the graph features, and then output to the physical modeling module.

[0024] As a further improvement to the deep learning-based intelligent predictive maintenance system for mining equipment, the physical modeling module includes a physical coding unit and a residual supervision unit.

[0025] The physical coding unit receives the final feature representation output by the feature extraction module, designs a physical constraint layer to encode the law of energy conservation, material constitutive relations and crack propagation laws into a differentiable network layer, constructs a dual-path parallel structure, the physical calculation path runs a simplified analytical model to calculate the state evolution of the computing device, and the neural network calculation path uses a multilayer perceptron to learn complex nonlinear relationships, introduces a gating fusion mechanism to calculate the fusion weight coefficients according to the input feature distribution characteristics, and weights the outputs of the two paths to form a fusion state representation;

[0026] The residual supervision unit receives the fusion state representation, introduces physical residual connections in each hidden layer of the neural network, calculates the deviation between the current layer output and the prediction of the physical constraint layer as a residual term and adds it to the loss function, adopts a course learning strategy, sets a low weight coefficient to fit the data in the early stage of training, gradually increases the weight coefficient in the middle stage to guide physical consistency convergence, fixes a high weight coefficient in the later stage, and outputs the trained physical embedded prediction model to the uncertainty quantification module.

[0027] As a further improvement to the deep learning-based intelligent predictive maintenance system for mining equipment, the uncertainty quantification module includes an uncertainty assessment unit and a confidence management unit.

[0028] The uncertainty assessment unit receives the physical embedded prediction model output by the physical modeling module, uses variational inference to convert the network weight parameters into a probability distribution, performs multiple forward propagation calculations using Monte Carlo sampling to assess the statistical dispersion of the prediction results and evaluate cognitive uncertainty, adds random perturbations to the input data for testing, evaluates the change in the data augmentation output as random uncertainty, constructs an out-of-distribution detector to calculate the Mahalanobis distance between the current input and the training manifold, marks it as an out-of-distribution sample when it exceeds a preset threshold and increases the uncertainty assessment value, and outputs the uncertainty assessment result.

[0029] The confidence management unit receives the uncertainty assessment results and multi-level output data, establishes a multi-layer uncertainty propagation mechanism, calculates local uncertainty values ​​in the feature extraction layer, state assessment layer, and life prediction layer, tracks the cumulative and amplification effects of each layer through the error propagation formula to generate a complete probability distribution, sets dynamic confidence thresholds for bearing failure, gear failure, structural fatigue failure, and lubrication failure based on historical data and the severity of consequences, compares and generates high, medium, and low confidence level prediction intervals, and outputs quantitative prediction results to the intelligent early warning module.

[0030] As a further improvement to the deep learning-based intelligent predictive maintenance system for mining equipment, the intelligent early warning module includes a risk assessment unit and a visualization unit.

[0031] The risk assessment unit receives the quantitative prediction results output by the uncertainty quantification module, extracts the failure probability data and remaining usage time data, combines the failure consequence severity database and maintenance resource availability data, integrates failure probability, consequence severity, time urgency and resource constraint weight coefficients through a weighted summation method to calculate a comprehensive risk index, classifies the risk index into blue monitoring level, yellow attention level, orange warning level and red alarm level according to the range of risk index values, calls the cost-benefit optimization model to solve for the optimal maintenance time window and maintenance strategy scheme, and outputs decision information;

[0032] The visualization unit receives the decision information and quantitative prediction results, constructs a digital twin 3D model, maps health status data to corresponding components and identifies levels through color coding, highlights degraded areas and labels the type, overlays temperature and stress field distributions with pseudo-color cloud maps, generates time series curves to display historical and predicted trends and adds confidence interval error bands, displays warning levels with icons and text, displays risk indices on dashboards, marks maintenance times on the timeline, displays maintenance plan costs and benefits, outputs to terminal devices and synchronizes with the human-machine collaboration module.

[0033] As a further improvement to the deep learning-based intelligent predictive maintenance system for mining equipment, the human-machine collaboration module includes a knowledge interaction unit and an interpretable decision-making unit.

[0034] The knowledge interaction unit provides a knowledge input interface to receive domain knowledge input by engineers, including fault mode descriptions, physical law expressions, material performance parameters, and operation and maintenance experience rules. It uses natural language processing technology for semantic parsing and entity recognition, extracts key physical parameters, constraints, and causal relationships, and transforms them into physical constraint expressions to update the physical modeling module or into data annotation information to mark the training dataset. It establishes a conversational inquiry mechanism to receive engineers' questions, retrieves reasoning process records and historical case data to generate response information, and outputs knowledge update results and dialogue records.

[0035] The interpretable decision unit receives decision information and model reasoning process data from the intelligent early warning module, generates an interpretable reasoning path, including input feature importance ranking, intermediate layer feature activation mode, and output result contribution decomposition, extracts key physical quantity time series data to display the evolution process and marks abrupt change points, identifies fatigue crack propagation, wear accumulation, thermal stress damage, and corrosion degradation as the dominant degradation mechanisms, retrieves similar cases from the historical failure case database and calculates similarity values ​​for ranking and display, generates multiple maintenance strategy alternatives for high-risk and high-uncertainty scenarios, and outputs decision support information to the visualization unit.

[0036] As a further improvement to the deep learning-based intelligent predictive maintenance system for mining equipment, the online optimization module includes a performance monitoring unit and an audit migration unit.

[0037] The performance monitoring unit tracks prediction accuracy, false alarm rate, and false negative rate in real time. It compares the actual fault conditions with the prediction results to calculate the performance index values. It uses statistical process control methods to establish a control chart and set upper and lower control limits. When the performance index exceeds the control limits, it is judged as performance degradation and triggers model update. It uses elastic weight consolidation technology or memory playback technology to update the model parameters online. It implements an active learning strategy to identify high-value samples for priority labeling and incremental training. It establishes a model version management mechanism to record update information. It implements an A / B testing mechanism to compare the performance of the old and new versions and replace the model. It outputs the updated model parameters and performance monitoring report.

[0038] The audit migration unit periodically audits the physical constraint layer's compliance, extracts predicted physical quantity values ​​to verify the degree of conformity with the law of conservation of energy, material constitutive relations, and crack propagation laws, and triggers the calibration process when the calculated physical residual deviation value exceeds a preset threshold. It also periodically audits the calibration accuracy of the uncertainty quantification module, collects predicted probability distribution and actual fault frequency data to calculate the deviation value of the calibration curve and adjusts the parameters, integrates cross-equipment and cross-mine fault case data to construct a cross-domain database, uses transfer learning technology to pre-train in the source domain and then fine-tunes the parameters in the target domain, aligns the feature distribution through a domain adaptive method to achieve knowledge reuse, and outputs audit reports and transfer learning model parameters to the performance monitoring unit.

[0039] This invention also provides a deep learning-based intelligent predictive maintenance method for mining equipment, characterized in that the method is based on the deep learning-based intelligent predictive maintenance system for mining equipment as described in claim 1, and includes the following steps:

[0040] S1. The data acquisition synchronization module acquires equipment operation data from vibration sensors, temperature sensors, strain sensors and acoustic emission sensors, and performs timestamp alignment processing on data with different sampling rates to form a multi-source fusion dataset; the physical preprocessing module detects outliers based on physical constraint boundaries and inverts the internal physical field distribution, calls the simulation model to generate virtual fault samples, mixes them with the measured data and normalizes them to form an enhanced dataset;

[0041] S2. The feature extraction module uses a multi-branch convolutional neural network to extract features at different time scales, constructs a graph neural network model to infer the component association state, and fuses multi-scale features and graph features through an adaptive attention mechanism.

[0042] S3. The physical modeling module designs a physical constraint layer to encode physical laws into a differentiable network layer, constructs a dual-path structure of physical calculation path and neural network path and merges them through a gating mechanism, introduces physical residual connections in each hidden layer and trains them using a course learning strategy.

[0043] S4. The uncertainty quantification module uses variational inference to assess cognitive uncertainty, identifies abnormal working conditions through an out-of-distribution detector, and establishes a multi-level uncertainty propagation mechanism to generate multi-level confidence prediction intervals.

[0044] S5. The intelligent early warning module integrates fault probability, severity of consequences, time urgency, and resource constraints to calculate a comprehensive risk index, classifies early warning levels, and solves for the optimal maintenance strategy. The equipment status and maintenance suggestions are visualized through a digital twin model. The human-machine collaboration module provides a knowledge input interface and a conversational query mechanism to realize human-machine collaboration.

[0045] S6. The online optimization module monitors the model's performance indicators in real time and triggers online updates when performance degrades. It also periodically audits the accuracy of physical consistency and uncertainty calibration and achieves cross-device knowledge reuse through transfer learning technology.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] 1. This invention fundamentally improves the superficiality of existing predictive maintenance systems in terms of physics and data fusion by deeply embedding physical mechanisms into the model structure design. Specifically, the physical modeling module designs a differentiable physical constraint layer, encoding key laws such as energy conservation, material constitutive modeling, and crack propagation into network layers, and adaptively fusing them with the neural network path through a gating mechanism. Simultaneously, physical residual connections and a curriculum learning strategy are introduced during training to ensure that the intermediate states from the feature layer to the output layer are consistently constrained by physical laws. This deeply embedded modeling approach enables the system to accurately capture the complex dynamic characteristics of multi-physics coupling, nonlinear feedback, and state abrupt changes during mining equipment failure. This significantly improves the reliability and accuracy of predictions when equipment enters critical stages such as rapid degradation, effectively overcoming the prediction bias caused by the loss of physical mechanisms in traditional data-driven models under extrapolation conditions.

[0048] 2. The system proposed in this invention constructs a complete technical chain from data source to maintenance decision-making through multi-module collaboration. The data acquisition and synchronization module not only achieves accurate time alignment of multi-source heterogeneous data, but also effectively distinguishes between real extreme working conditions and sensor anomalies through a sensor health monitoring mechanism, ensuring data quality from the source. The physical preprocessing module identifies outliers based on physical constraints and generates virtual fault samples using simulation technology, maintaining the physical rationality of the data while expanding scarce samples. The feature extraction module adopts a combination of multi-branch convolutional neural networks and graph neural networks to achieve parallel extraction of features at multiple time scales and reasoning of topological relationships between equipment components. These measures together provide a high-quality, high-information input foundation for subsequent deep modeling, solving engineering challenges such as data inconsistency, scarce samples, and complex component relationships in mining scenarios.

[0049] 3. This invention further provides a reliable confidence assessment of the prediction results through an uncertainty quantification module. This module not only separates and quantifies model cognitive uncertainty and data stochastic uncertainty, but also establishes an out-of-distribution detection mechanism and a multi-layer uncertainty propagation mechanism, enabling it to identify abnormal operating conditions and track the cumulative effect of uncertainty within the model. Combined with the risk index calculated by the intelligent early warning module and the optimized maintenance strategy, the system outputs no longer a single deterministic value, but a probabilistic prediction with confidence intervals and tiered early warning suggestions. This solution enables maintenance decisions to shift from experience-based judgment to scientific planning based on risk quantification, more rationally balancing maintenance costs and equipment availability while ensuring safety, and reducing unnecessary preventative downtime.

[0050] 4. The human-machine collaboration module of this invention deeply integrates expert experience with the model reasoning process, enhancing the interpretability and usability of the system. Through a natural language processing interface, the domain knowledge input by engineers can be transformed into physical constraints or data annotations usable by the model, enabling continuous injection of expert experience. Simultaneously, the system can generate interpretable reasoning paths, showcasing key feature contributions, degradation mechanisms, and similar historical cases, and responding to questions in a dialogue format. This not only enhances maintenance personnel's understanding and trust in the prediction results but also enables the system to adapt to the specific equipment characteristics and maintenance habits of different sites, becoming an interactive and evolving auxiliary decision-making platform.

[0051] 5. The online optimization module of this invention ensures long-term system performance stability and efficient knowledge reuse. By monitoring performance indicators in real time and employing incremental learning techniques to prevent catastrophic forgetting, the model can adapt to changes in operating conditions and new failure modes, avoiding performance degradation over time. A periodic auditing mechanism maintains the physical consistency of the model and the accuracy of uncertainty calibration. Through transfer learning and domain adaptation techniques, the system can integrate failure cases across devices and scenarios, enabling rapid knowledge transfer and reuse, significantly shortening the deployment and debugging cycle of new equipment, and improving the overall efficiency of engineering promotion of the solution. Attached Figure Description

[0052] Figure 1 This is an architecture diagram of a deep learning-based intelligent predictive maintenance system for mining equipment.

[0053] Figure 2 This is a flowchart of a deep learning-based intelligent predictive maintenance method for mining equipment. Detailed Implementation

[0054] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0055] like Figure 1 A deep learning-based intelligent predictive maintenance system for mining equipment includes a data acquisition and synchronization module, a physical preprocessing module, a feature extraction module, a physical modeling module, an uncertainty quantification module, an intelligent early warning module, a human-machine collaboration module, and an online optimization module connected in sequence.

[0056] The data acquisition and synchronization module obtains operational data from vibration, temperature, strain, and acoustic emission sensors and performs timestamp alignment processing.

[0057] The physical preprocessing module detects outliers based on the physical constraint boundaries of the equipment, generates virtual fault samples using simulation, and reconstructs the internal physical field distribution.

[0058] The feature extraction module uses a multi-branch convolutional neural network to extract features at different time scales and constructs a graph neural network to infer the associated states of components.

[0059] The physical modeling module designs a physical constraint layer that encodes conservation laws and constitutive relations into a differentiable network layer, and adopts a dual-path structure to fuse analytical models and neural networks.

[0060] The uncertainty quantification module assesses cognitive uncertainty through variational inference and establishes an out-of-distribution detector and a multi-layer propagation mechanism.

[0061] The intelligent early warning module comprehensively calculates the risk index and establishes a four-level early warning mechanism, and visualizes the equipment status through a digital twin interface.

[0062] The human-machine collaboration module provides a knowledge input interface and an interpretable reasoning path, and establishes a conversational inquiry mechanism.

[0063] The online optimization module monitors performance metrics in real time, uses incremental learning to update the model, and integrates cross-device fault cases to achieve knowledge reuse.

[0064] In this embodiment, the data acquisition synchronization module obtains operational data from vibration sensors, temperature sensors, strain sensors, acoustic emission sensors, and oil analysis sensors. It performs timestamp alignment processing on data from different sampling frequencies, solving the problem of accurately matching multi-source heterogeneous data in the time dimension. Traditional data acquisition systems often ignore the health status of the sensors themselves, leading to abnormal data being mixed into the analysis process. The sensor health monitoring mechanism established in this module can detect sensor drift, malfunctions, or data loss and generate quality markers, avoiding subsequent misjudgments caused by sensor anomalies. The time-aligned multi-source dataset output by this module, with quality markers, lays the data foundation for subsequent physical field reconstruction and multivariate coupling analysis, improving the overall system's data reliability.

[0065] The physical preprocessing module performs outlier detection based on the physical constraint boundaries of the equipment, distinguishing between normal responses under extreme operating conditions and abnormal data generated by sensor malfunctions, thus avoiding misclassification of real extreme conditions as noise and its rejection. Obtaining fault samples for mining equipment in actual operation is difficult and costly. This module utilizes finite element models and multibody dynamics simulations to generate virtual fault samples, effectively expanding the number of samples for scarce fault types in the training dataset. Through physical field reconstruction technology, this module inverts the complete temperature and stress field distributions inside the equipment from limited measurement point data, expanding local information to a global state. After normalizing different physical quantities according to their dimensions and coupling relationships, the enhanced dataset output by this module conforms to physical laws and has sufficient sample diversity, reducing the dependence of deep learning models on a large amount of actual fault data.

[0066] The feature extraction module employs a multi-branch convolutional neural network. The short-time branch uses small-sized convolutional kernels to capture transient features such as impact, the medium-time branch uses medium-sized convolutional kernels to extract periodic features such as temperature, and the long-time branch uses large-sized convolutional kernels to identify long-term cumulative features such as fatigue. This avoids the limitation of a single time window being unable to capture multi-scale features simultaneously. Mining equipment components have complex mechanical connections and energy transfer relationships. The graph neural network constructed in this module encodes the equipment topology as relationships between nodes and edges. Through graph convolution operations, it enables the inference of the health status of adjacent components from local sensor signals, allowing for early identification of cascading failure risks caused by single-point failures. The introduced adaptive attention mechanism dynamically adjusts the weights of features at different time scales according to the current degradation stage of the equipment. This allows the model to focus on transient features in the early stages of degradation and on long-term cumulative features in the later stages, improving the specificity of feature representation.

[0067] The physical modeling module's physical constraint layer expresses the law of energy conservation, material constitutive relations, and crack propagation laws as a differentiable network layer, enabling physical knowledge to participate in the gradient backpropagation process of the neural network. In the dual-path structure employed in this module, the physical path runs a simplified analytical model based on physical laws, while the neural network path learns complex nonlinear relationships that the analytical model cannot describe. A gating fusion mechanism dynamically adjusts the contribution ratio of the two paths based on the characteristics of the input features. The introduced physical residual connection calculates the deviation between the predicted features and the physical constraint features at each layer of the network, using this deviation as a loss term to supervise the training process and ensure that the intermediate layer feature representations conform to physical evolution laws. The adopted curriculum learning strategy uses lower physical constraint weights in the early stages of training to allow the model to quickly fit the data, and gradually increases the physical constraint weights in the later stages of training to guide the model towards physical consistency. This module improves the model's ability to predict extrapolated scenarios beyond the training data distribution, reducing the need for large samples in purely data-driven methods.

[0068] The uncertainty quantification module uses variational inference to convert deterministic network weights into a probability distribution. It generates multiple predictions through Monte Carlo sampling and calculates their dispersion to assess cognitive uncertainty, which reflects the degree of uncertainty in the model parameters. Data augmentation is performed by adding random perturbations to the input data that conform to noise statistics. This module evaluates the magnitude of output change caused by the input perturbations as random uncertainty, reflecting the noise level of the data itself. An out-of-distribution detector calculates the Mahalanobis distance between the current input data and the training data manifold. When the distance exceeds a threshold, it is identified as an out-of-distribution sample, and the uncertainty assessment value is increased, preventing the model from making overly confident erroneous predictions under unfamiliar conditions. A multi-layer uncertainty propagation mechanism tracks the accumulation and amplification effects of uncertainty between the feature extraction layer, state assessment layer, and lifetime prediction layer. Dynamic confidence thresholds are set for different fault types based on historical data and the severity of consequences. The complete probability distribution output by this module provides risk quantification information for maintenance decisions, enabling decision-makers to choose conservative or aggressive maintenance plans based on their risk tolerance.

[0069] The intelligent early warning module comprehensively considers fault probability, severity of consequences, remaining usage time, and maintenance resource availability to calculate a comprehensive risk index, realizing the transformation from a single technical indicator to comprehensive management decision-making. The module establishes a four-level early warning mechanism, classifying equipment status into blue monitoring, yellow attention, orange warning, and red alarm levels based on the risk index. Different levels correspond to different inspection frequencies and response measures, avoiding overload of early warning information and inefficient allocation of maintenance resources. The introduced cost-benefit optimization model aims to minimize maintenance costs and maximize equipment availability, using safety risk thresholds and resource constraints as limitations to solve for the optimal maintenance time window and maintenance strategy, reducing unnecessary preventative downtime while ensuring safety. A digital twin 3D model visualizes the equipment's health status, degradation location identification, internal temperature and stress field distribution, performance parameter trends, and confidence intervals of prediction results, providing maintenance personnel with intuitive status perception and decision-making basis, improving decision-making efficiency and accuracy.

[0070] The human-machine collaboration module provides a knowledge input interface that allows engineers to input knowledge in areas such as equipment failure modes, physical laws, material parameters, and maintenance experience. This knowledge is then transformed into physical constraints or data annotations through semantic parsing and entity recognition using natural language processing technology, realizing the conversion of expert experience into a system knowledge base. The module's output interpretable reasoning path includes input feature importance ranking, the temporal evolution of key physical quantities, identification of dominant degradation mechanisms, and comparison with similar historical cases. This transforms the deep learning model's prediction process from a black box into a traceable and transparent process, enhancing maintenance personnel's understanding and trust in the system's prediction results. The established dialogic inquiry mechanism allows engineers to question the prediction results, and the system retrieves reasoning process records and historical case data to generate explanatory responses, promoting knowledge exchange between humans and machines. For decision-making scenarios with high comprehensive risk indices and significant uncertainty, this module generates multiple maintenance strategy alternatives and displays the cost estimates and risk reduction levels of each option, preserving the final judgment of humans in key decisions. This module improves the overall decision-making level of the maintenance team and accelerates the learning process of newly hired engineers for equipment condition assessment methods.

[0071] The online optimization module performs real-time statistical analysis of prediction accuracy, false positive rate, and false negative rate. A performance index control chart is established using statistical process control methods. When performance indicators exceed control limits, the model update process is automatically triggered, ensuring long-term system stability. This module employs elastic weight consolidation technology, adding a regularization term to the loss function to constrain the variation of important parameters, or uses memory replay technology to mix and retain historical key samples during new data training. This effectively avoids catastrophic forgetting during incremental learning, enabling the model to adapt to new conditions without losing its ability to identify learned fault modes. The implemented active learning strategy calculates the information entropy or prediction uncertainty of unlabeled samples, identifying high-value samples that contribute most to model improvement and prioritizing their submission for manual annotation, improving the efficiency of annotation resources. A model version management mechanism records the parameter configuration and performance indicators for each update. The implemented A / B testing mechanism ensures the safety and traceability of model updates by deploying new and old versions of the model in parallel and comparing their performance. Regular audits of the physical constraint layer's satisfaction and the calibration accuracy of uncertainty quantification maintain the model's physical consistency and probabilistic prediction reliability. By integrating fault case data across equipment and mines, and using transfer learning technology to pre-train models on source domain data and then fine-tuning parameters on target domain data, fault diagnosis knowledge can be reused across different equipment and environments, shortening the commissioning cycle of new equipment.

[0072] By working together with the data acquisition synchronization module and the physical preprocessing module, the system completed tasks such as accurate time alignment of multi-source heterogeneous data, sensor health status monitoring, outlier identification based on physical constraints, and virtual fault sample generation. This solved long-standing problems in engineering applications, such as inconsistent sampling frequency of monitoring data for mining equipment, confusing judgment of sensor faults, and scarcity of actual fault samples, and provided a high-quality and large-scale training dataset for subsequent analysis.

[0073] By working in concert with the feature extraction module and the physical modeling module, the system has completed tasks such as parallel feature extraction at multiple time scales, equipment topology modeling, networked encoding of physical laws, and fusion modeling driven by data and mechanism. This solves the problems of traditional methods, such as single feature representation, neglect of component relationships, and weak extrapolation ability due to lack of physical constraints. The system can maintain high prediction accuracy under normal operating conditions, boundary conditions, and unknown operating conditions, and significantly reduces the false alarm rate and false negative rate.

[0074] By working in tandem with the uncertainty quantification module and the intelligent early warning module, the system has accomplished tasks such as multi-source uncertainty quantification, out-of-distribution sample identification, comprehensive risk index calculation, dynamic classification of early warning levels, and cost-effectiveness optimization of maintenance strategies. This solves the problems of traditional methods, which only provide deterministic prediction values, cannot quantify the reliability of predictions, and are difficult to balance safety and economy. It transforms maintenance decisions from experience-based judgments to scientific decisions based on probability distribution and risk quantification, reducing unnecessary preventative downtime and improving the overall utilization rate of equipment while ensuring equipment safety.

[0075] Through the collaborative work of the human-machine collaboration module and the online optimization module, the system has completed tasks such as expert knowledge input transformation, interpretable presentation of the prediction process, continuous monitoring of model performance, incremental learning to prevent forgetting and updating, and cross-domain fault knowledge transfer. This has solved problems such as low trust level caused by the black box nature of traditional deep learning methods, the inability of static models to adapt to changes in working conditions, and the need for new equipment to be trained from scratch. The system has been transformed from an uninterpretable prediction tool into an understandable, interactive, and evolvable auxiliary decision-making platform, shortening the deployment cycle of new equipment.

[0076] Through the collaborative work of the above eight modules, this solution realizes the transformation of mine equipment maintenance from reactive, post-failure response maintenance to proactive, pre-failure predictive maintenance; from relying on the experience and judgment of individual engineers to intelligent decision-making based on data analysis and physical model reasoning; and from independently formulating maintenance plans for single pieces of equipment to system optimization that comprehensively considers the resource constraints of multiple equipment and production plans. It solves to the greatest extent the technical problems of insufficient integration of physical mechanisms, oversimplification of multi-physics coupling relationships, and low predictive reliability when data is scarce and operating conditions change.

[0077] like Figure 1The data acquisition and synchronization module includes a multi-source data acquisition unit and a time-series synchronization processing unit.

[0078] The multi-source data acquisition unit collects operating data from the equipment's vibration sensors, temperature sensors, strain sensors, acoustic emission sensors, and oil analysis sensors. It also obtains load, speed, and pressure values ​​from the control system. The unit performs multi-rate acquisition for different sampling frequencies and adds timestamps. Through signal amplitude checks, sampling rate analysis, and continuity judgment, it detects sensor drift, faults, or missing data, and generates quality markers and alarm information.

[0079] The timing synchronization processing unit receives multi-source data streams with timestamps and quality label information, establishes a unified time reference, downsamples high-frequency data or interpolates low-frequency data, maps the measured values ​​to a unified time node, identifies and isolates abnormal data, generates a timing-synchronized multi-source fusion dataset, and outputs it to the physical preprocessing module.

[0080] In this embodiment, the sampling frequency of different sensors in the mining equipment is determined according to the physical characteristics of the monitored object. Vibration and acoustic emission sensors use a higher sampling frequency to capture rapidly changing dynamic signals, while temperature and oil sensors use a lower sampling frequency to track slowly changing thermal and chemical states. The sampling frequency of the strain sensor is selected according to the load change rate.

[0081] The multi-source data acquisition unit employs a multi-rate acquisition method to address the aforementioned sampling frequency differences. Each sensor operates independently according to a preset frequency, and each acquired data point is appended with a microsecond-level timestamp. Traditional acquisition systems typically assume that sensors will continuously operate normally. However, in actual operation, sensors may experience measurement drift or data interruption due to loose installation, aging cables, circuit faults, etc. If these abnormal data are directly included in the analysis process, they will interfere with subsequent judgments. The multi-source data acquisition unit establishes three detection mechanisms to identify sensor anomalies: checking the signal amplitude to determine if the measured value exceeds the physically reasonable range, analyzing the sampling rate to determine if the data arrival interval deviates from the normal cycle, and checking for missing segments in the data sequence based on continuity. When any of the above anomalies are detected, the multi-source data acquisition unit adds a quality marker to the corresponding data point and sends an alarm message to the monitoring system to prompt maintenance personnel to check the sensor status.

[0082] The timing synchronization processing unit receives timestamped data streams and quality marker information from the multi-source data acquisition unit and establishes a unified time coordinate based on the device's main control system clock. Because the sampling frequencies of each sensor differ, the distribution density of their data points on the time axis varies; directly performing multivariate analysis may result in mismatched time points.

[0083] The time-series synchronization processing unit employs downsampling methods for high-frequency data, calculating statistics such as mean and variance within a sliding time window, or using wavelet transform to extract energy features from different frequency bands, converting the dense raw data into a feature sequence. For low-frequency data, interpolation methods are used, selecting linear interpolation, spline interpolation, or prediction interpolation based on the variation law of physical quantities. After the above processing, the time-series synchronization processing unit uniformly maps the measurement values ​​of each sensor to time nodes at the same interval, with the time interval set according to the needs of subsequent analysis.

[0084] For anomalous data with quality labels, the time-series synchronization processing unit handles them according to the type of anomalous data: data with anomalous amplitude but continuous temporal sequence retains the label for subsequent weight adjustment during analysis; timestamp offsets caused by sampling rate anomalous data are corrected by referencing adjacent normal data; and data missing caused by continuity anomalies are marked as missing intervals. In the dataset output by the time-series synchronization processing unit, multiple physical quantity values ​​at the same time node correspond to the actual operating status of the equipment at that moment, and the temporal correlation between physical quantities is accurately maintained.

[0085] In this system, the data acquisition synchronization module, through the cooperation of a multi-source data acquisition unit and a time-series synchronization processing unit, achieves end-to-end quality control from the data source to the pre-processed output. The multi-source data acquisition unit is responsible for sensor status monitoring and anomaly identification, while the time-series synchronization processing unit is responsible for multi-rate data time alignment and anomaly classification. Compared with existing technologies, the data acquisition synchronization module in this system avoids the long-term accumulation of abnormal data by monitoring sensor status in real time during the acquisition phase through an integrated quality control method. It replaces the coarse pairing method with timestamp-based alignment processing, enabling the response of equipment operating condition changes to multiple physical quantities to be fully reflected. The quality marking mechanism records the source and type of anomalies, reducing manual intervention.

[0086] like Figure 1 The physical preprocessing module includes a constraint detection unit and a simulation enhancement unit.

[0087] The constraint detection unit receives a multi-source fusion dataset, detects outliers based on temperature change rate threshold, vibration amplitude limit, stress-strain relationship range, and material property parameter boundaries, distinguishes between real extreme working conditions and sensor errors and marks them, and solves the heat conduction equation and elasticity equation based on sparse measurement point data, geometric structure model and material property parameters to invert the internal temperature field and stress field distribution, and outputs the measured data and reconstructed physical field data after detection.

[0088] The simulation enhancement unit receives measured data and physical field data, calls the finite element model and multibody dynamics model, changes parameters such as load spectrum, boundary conditions, material properties and environmental conditions to generate virtual sensor response data, simulates the occurrence and evolution of faults for scarce fault samples to generate virtual fault data, mixes them to form an enhanced dataset, normalizes it according to physical dimensions and coupling relationships, considers causal relationships and coupling strength weights, and outputs it to the feature extraction module.

[0089] In this embodiment: The constraint detection unit receives the multi-source fusion dataset output by the data acquisition synchronization module and performs outlier detection and internal physical field inversion based on the physical constraint boundaries of the device. It determines the upper limit of the temperature change rate based on material heat capacity and heat dissipation conditions, the vibration amplitude limit based on structural strength and load conditions, and the stress-strain range based on material constitutive relations. Judgments are made using these constraint boundaries based on physical laws. When a measured value exceeds the physical constraint boundary, it is marked as a sensor error. Values ​​within the boundary, even if they deviate from the statistical mean, are retained as true data, thus distinguishing between the actual extreme operating condition response and sensor errors. Simultaneously, based on sparse measurement point data from the surface or reachable locations, combined with the geometric structure model and material property parameters, the constraint detection unit solves the heat conduction equation to invert the internal temperature field distribution and solves the elasticity equation to invert the internal stress field distribution, extending the monitoring information from local measurement points to the entire device domain, thereby providing a complete state description for subsequent analysis.

[0090] The simulation enhancement unit calls upon pre-established finite element models and multibody dynamics models, and generates virtual sensor response data under different working conditions by changing input parameters such as load spectrum parameters, boundary conditions, material property parameters, and environmental conditions. For rare fault types such as bearing pitting, gear cracking, structural fatigue, and lubrication failure, corresponding fault mechanisms are introduced into the simulation model. Virtual sensor response data during the fault occurrence and evolution process are generated through simulation calculations, and the measured data and virtual data are mixed to form an enhanced dataset.

[0091] Meanwhile, considering the differences in the dimensions and numerical ranges of different physical quantities in the dataset, the simulation enhancement unit normalizes the data according to the physical dimensions and coupling relationships, taking into account the causal relationships and coupling strength weights between physical quantities. For example, load changes cause vibration and stress changes, and temperature increases affect material stiffness and thus vibration characteristics. The normalization process maintains the relative strength of these relationships, thereby providing a sufficient number of diverse and physically consistent samples for training deep learning models.

[0092] Existing preprocessing methods rely on statistical features for anomaly detection, which is prone to misjudgment when processing extreme operating condition data. They also lack means to obtain internal physical quantities by relying on measured data. Oversampling or data augmentation techniques used to address the scarcity of fault samples are essentially statistical transformations of existing samples, lacking physical diversity. The physical preprocessing module in this system, through the cooperation of a constraint detection unit and a simulation enhancement unit, introduces physical constraints and simulation techniques into the preprocessing stage. The constraint detection unit retains real extreme operating condition data by judging physical constraint boundaries and obtains internal physical field information through physical field inversion. The simulation enhancement unit generates virtual samples with physical diversity by changing simulation parameters and covers scarce fault modes through fault mechanism simulation. Therefore, the dataset output by the physical preprocessing module outperforms existing technologies in terms of accuracy, completeness, and diversity, providing high-quality training input for the feature extraction module and strengthening the physical consistency and generalization ability of the model from the data preparation stage.

[0093] like Figure 1 The feature extraction module includes a multi-scale extraction unit and a graph reasoning unit.

[0094] The multi-scale extraction unit receives the normalized augmented dataset and uses a multi-branch convolutional neural network to process the time series data in parallel. The short-time branch, medium-time branch, and long-time branch use convolutional kernels of increasing size to extract transient impact features, periodic temperature change features, and long-term fatigue accumulation features, respectively. Each branch independently completes convolution and pooling operations and outputs the corresponding feature vector. The three types of feature vectors are concatenated in the feature dimension to form a multi-scale fusion feature vector.

[0095] The graph inference unit receives multi-scale fused feature vectors, constructs a graph neural network model to represent the device topology, with nodes corresponding to components and edges representing physical connections and energy transfer paths. The fused feature vectors are used as the initial feature inputs for the nodes, and feature information is propagated through graph convolution operations to infer the health status of adjacent components. The output is a graph feature representation. An adaptive attention mechanism is introduced to calculate weight coefficients based on degradation stage parameters, weighting features at different scales and fusing them with graph features, and then outputting to the physical modeling module.

[0096] In this embodiment: The multi-scale extraction unit receives the normalized enhanced dataset and uses a multi-branch convolutional neural network to process the time series data in parallel, completing the synchronous extraction of features at different time scales. The multi-scale extraction unit sets up three parallel processing channels: short-time branch, medium-time branch, and long-time branch, each using convolutional kernels of increasing size to extract features from the time series data. The short-time branch uses small-sized convolutional kernels to extract transient impact features from high-frequency vibration signals; the medium-time branch uses medium-sized convolutional kernels to extract periodic temperature change features related to the equipment's operating cycle; and the long-time branch uses large-sized convolutional kernels to extract fatigue accumulation features reflecting the long-term performance trend of the equipment.

[0097] Each branch independently performs convolution and pooling operations, outputting its corresponding feature vector. These vectors are then concatenated along the feature dimension to form a multi-scale fused feature vector. This allows the feature representation to simultaneously include degradation information across three time scales: transient, periodic, and cumulative. This multi-branch parallel extraction method solves the problem that a single time window cannot adequately capture multi-scale features. The transient signals captured by the short window, the periodic changes captured by the medium window, and the cumulative trends captured by the long window are all fully preserved in the fused features, providing a more comprehensive state description for subsequent models.

[0098] The graph inference unit receives multi-scale fused feature vectors and constructs a graph neural network model to represent the device's topology, completing the modeling of inter-component relationships and the inference of adjacent component states. The graph inference unit models the device as a graph structure, where nodes correspond to the various components of the device, and edges represent the physical connections and energy transfer paths between components. The multi-scale fused feature vectors are used as the initial feature inputs to the nodes. Through graph convolution operations, node features propagate along the edges to adjacent nodes. The feature update of each node not only depends on its own measurement data but also incorporates feature information from neighboring nodes, thus achieving feature propagation and state inference based on topological relationships.

[0099] The graph inference unit introduces an adaptive attention mechanism, calculating weight coefficients based on parameters of the equipment's degradation stage. During normal operation, long-term features are given higher weight to monitor cumulative degradation, while transient features are given higher weight to capture abnormal signals during the fault initiation stage. After weighting features at different scales and fusing them with graph features, the output is a graph feature representation that reflects the current state of the equipment and the relationships between components. This graph-based inference method utilizes the physical connections and energy transfer relationships between components in the equipment. Correlation effects such as vibrations from bearing failures being transmitted to the gearbox via the shaft system, and heat from gear wear being transferred to adjacent bearings via lubricating oil, are reflected in the graph features.

[0100] Existing feature extraction methods typically employ fixed time windows or single-scale convolutional neural networks for feature learning, making it difficult to simultaneously capture degradation features across multiple time scales. They also process sensor data from each component independently without considering the device topology and physical connections between components. This application's feature extraction module integrates multi-time scale feature learning and device topology modeling into the feature extraction process through the collaboration of a multi-scale extraction unit and a graph inference unit. The multi-scale extraction unit achieves simultaneous extraction and fusion of transient, periodic, and cumulative features through multi-branch parallel processing. The graph inference unit uses graph neural network modeling to achieve feature propagation based on the device topology and inference of adjacent component states. An adaptive attention mechanism dynamically adjusts the weights of features at different scales according to the degradation stage. This results in the feature representation output by the feature extraction module outperforming existing technologies in terms of time scale coverage, component relationship representation, and degradation stage adaptability. It provides the physical modeling module with more comprehensive and physically meaningful feature inputs, improving the model's ability to represent complex degradation processes at the feature level.

[0101] like Figure 1 The physical modeling module includes a physical coding unit and a residual supervision unit.

[0102] The physical coding unit receives the final feature representation, designs a physical constraint layer to encode the law of energy conservation, material constitutive relations and crack propagation laws into a differentiable network layer, constructs a dual-path parallel structure, the physical calculation path runs a simplified analytical model to calculate the state evolution of the computing device, and the neural network calculation path uses a multilayer perceptron to learn complex nonlinear relationships, introduces a gating fusion mechanism to calculate the fusion weight coefficients based on the distribution characteristics of the input features, and weights the outputs of the two paths to form a fused state representation.

[0103] The residual supervision unit receives the fusion state representation, introduces physical residual connections in each hidden layer of the neural network, calculates the deviation between the current layer output and the prediction of the physical constraint layer as a residual term and adds it to the loss function. A course learning strategy is adopted, setting a low weight coefficient to fit the data in the early stage of training, gradually increasing the weight coefficient in the middle stage to guide physical consistency convergence, and fixing a higher weight coefficient in the later stage. The trained physical embedded prediction model is then output to the uncertainty quantification module.

[0104] In this embodiment: The physical encoding unit receives the final feature representation output by the feature extraction module, and designs a physical constraint layer to encode the law of energy conservation, material constitutive relations, and crack propagation laws into a differentiable network layer, constructing a dual-path parallel structure of physical computation path and neural network computation path. The physical computation path runs a simplified analytical model based on physical laws to calculate the state evolution of the computing device, while the neural network computation path uses a multilayer perceptron to learn complex nonlinear relationships in the data. The physical encoding unit introduces a gating fusion mechanism, calculates fusion weight coefficients based on the distribution characteristics of the input features, and weights and sums the outputs of the two paths to form a fused state representation.

[0105] This dual-path parallel structure allows physical knowledge to directly participate in the forward propagation and gradient backpropagation processes of the network, strengthening the constraint effect of physical laws on model prediction. The gating fusion mechanism dynamically adjusts the weights of the two paths based on the distribution characteristics of the input features. When the input features are within the coverage of the training data, the weights of the neural network path are higher; when the input features deviate from the training distribution, the weights of the physical computation path are higher, thereby improving the prediction reliability of the model in small sample scenarios and extrapolation conditions.

[0106] The residual supervision unit receives the fused state representation output by the physical coding unit and introduces physical residual connections into each hidden layer of the neural network, completing the propagation and supervision of physical constraints in the deep structure of the network. The residual supervision unit calculates the deviation between the current layer's output and the prediction of the physical constraint layer as a residual term, which is then added to the loss function. This ensures that the parameter updates of each layer are not only guided by the gradient of the final prediction error but also constrained by the physical consistency deviation of intermediate layers. The residual supervision unit employs a curriculum learning strategy to control the strength of physical constraints. In the early stages of training, lower weight coefficients are set to allow the model to preferentially fit the training data. In the mid-stage, the weight coefficients are gradually increased to guide the model towards physical consistency. In the later stages, higher weight coefficients are fixed to ensure that the model output conforms to physical laws.

[0107] This training strategy, which adjusts the strength of physical constraints in stages, avoids the problem of model convergence difficulties caused by excessively strong physical constraints in the early stages, while ensuring that the model follows physical laws in the later stages of training. Physical residual connections propagate physical constraint information in the deep structure of the network. Compared with applying physical constraints only to the output layer, the physical meaning of the features in each layer of the network is enhanced, the interpretability of the model and its generalization ability under unknown conditions are improved, and the output physical embedded prediction model provides a prediction foundation for the uncertainty quantification module that combines data fitting ability and physical consistency.

[0108] Existing technologies primarily employ purely data-driven methods for predictive modeling. Even when physical constraints are introduced, they are often treated as auxiliary terms in the loss function rather than deeply embedded in the network architecture. This results in models performing well under conditions covered by training data, but struggling to effectively infer physical principles when data is scarce or conditions change. The physical modeling module in this application deeply integrates physical knowledge into the network structure and training process through the cooperation of a physical encoding unit and a residual supervision unit. The physical encoding unit achieves adaptive fusion of physical computation and data-driven approaches through a dual-path parallel structure and a gating fusion mechanism, allowing physical principles to directly participate in forward propagation and gradient calculation. The residual supervision unit, through physical residual connections and a curriculum learning strategy, enables the propagation of physical constraints in the deep network structure and dynamic adjustment of their intensity during training. Consequently, the predictive model output by the physical modeling module outperforms existing technologies in terms of physical consistency, small-sample generalization ability, and reliability of extrapolated conditions. The model maintains high fitting accuracy within the training data coverage range and maintains reasonable prediction results through physical constraints under scenarios of scarce data and changing conditions, providing a more reliable prediction foundation for the uncertainty quantification module and improving the system's adaptability to complex conditions from the modeling method perspective.

[0109] The physical constraint layer in the physical modeling module is implemented using a differentiable physical computation structure. This structure embeds the physical laws involved in the device's operation into the forward propagation process of the neural network in the form of computational units. Specifically, based on pre-established laws of energy conservation, material constitutive relations, and crack propagation, the physical constraint layer discretizes the corresponding mathematical model and constructs it into computational units composed of basic operators. This allows the model to participate in model training as a network layer and support gradient backpropagation.

[0110] The physical constraint layer receives feature representations from the feature extraction module or intermediate layers of the neural network, calculates the device state evolution results under physical constraints based on the features, and outputs the corresponding physical prediction results. These physical prediction results, as the output of the physical calculation path, are used together with the prediction results obtained by the neural network calculation path using a data-driven approach for subsequent fusion processing.

[0111] In the physical encoding unit, a dual-path structure is employed, with the physical computation path and the neural network computation path running in parallel, to process the input features. The physical computation path runs a simplified analytical model based on the physical constraint layer to physically model the device state; the neural network computation path performs nonlinear mapping on the input features through a multilayer perceptron to learn the complex relationships implicit in the data. The physical encoding unit sets up a gating fusion mechanism, automatically calculating fusion weight coefficients based on the distribution characteristics of the input features, and weighting and fusing the outputs of the two paths according to the fusion weight coefficients to form a fused state representation.

[0112] The gating fusion mechanism can dynamically adjust the contribution ratio of the two paths according to the distribution area of ​​the input features. When the input features are within the coverage of the training data, the weight of the neural network calculation path is increased; when the input features deviate from the distribution of the training data or are in a region with scarce samples, the weight of the physical calculation path is increased, so that the model can obtain more reasonable prediction results under different working conditions.

[0113] The residual supervision unit receives the fused state representation and introduces physical residual connections in multiple hidden layers of the neural network. The residual supervision unit calculates the deviation between the network output of each hidden layer and the physical prediction result obtained by the physical constraint layer based on the same input features, and uses this deviation as a physical residual term in model training. This physical residual term is superimposed on the overall loss function of the model, so that the network parameter update process is simultaneously constrained by both prediction error and physical consistency deviation.

[0114] During training, the residual supervision unit adopts a course learning strategy to dynamically adjust the weight of the physical residual term in the loss function. In the early stage of training, a lower weight is set to prioritize fitting the training data. As the training process progresses, the weight of the physical residual is gradually increased to guide the model to converge toward physical consistency. In the later stage of training, a higher weight of the physical residual is maintained to ensure that the model output conforms to the constraints of physical laws.

[0115] like Figure 1 The uncertainty quantification module includes an uncertainty assessment unit and a confidence management unit.

[0116] The uncertainty assessment unit receives the physically embedded prediction model, uses variational inference to convert the network weight parameters into a probability distribution, performs multiple forward propagation calculations through Monte Carlo sampling to assess the statistical dispersion of the prediction results and evaluate cognitive uncertainty, adds random perturbations to the input data for testing, evaluates the magnitude of the output change as random uncertainty, constructs an out-of-distribution detector to calculate the Mahalanobis distance between the current input and the training manifold, marks samples exceeding a preset threshold as out-of-distribution samples and increases the uncertainty assessment value, and outputs the uncertainty assessment result.

[0117] The confidence management unit receives uncertainty assessment results and multi-level output data, establishes a multi-layer uncertainty propagation mechanism, calculates local uncertainty values ​​in the feature extraction layer, state assessment layer, and life prediction layer, and generates a complete probability distribution by tracking the cumulative and amplification effects of each layer through the error propagation formula. For bearing failure, gear failure, structural fatigue failure, and lubrication failure, dynamic confidence thresholds are set based on historical data and the severity of consequences. The unit compares and generates prediction intervals for high, medium, and low confidence levels, and outputs quantitative prediction results to the intelligent early warning module.

[0118] In this embodiment: The uncertainty assessment unit receives the physical embedded prediction model output by the physical modeling module, uses variational inference to convert the network weight parameters into a probability distribution, and calculates the statistical dispersion of the prediction results through multiple forward propagations using Monte Carlo sampling to assess cognitive uncertainty. When random perturbations are added to the input data for testing, the magnitude of the output change is used as the random uncertainty. Cognitive uncertainty reflects the model's own lack of knowledge; it is higher in regions with scarce training samples. Random uncertainty reflects inherent noise and measurement errors in the data. The uncertainty assessment unit constructs an out-of-distribution detector to calculate the Mahalanobis distance between the current input and the training data manifold. When the distance exceeds a preset threshold, it is marked as an out-of-distribution sample, and the uncertainty assessment value is increased. This enables the identification of abnormal operating conditions beyond the training range, avoiding the model from outputting overconfident and erroneous predictions under unknown conditions.

[0119] The training manifold described in the uncertainty assessment unit is obtained through statistical modeling of the distribution characteristics of training samples in the feature space during the model training phase. Specifically, after the physical embedded prediction model is trained, the training data is input into the selected feature layer or state evaluation layer in the model to obtain the corresponding feature representations. Statistical analysis is then performed on these feature representations to form an overall characterization of the distribution range and concentration of training samples in the feature space, thus serving as the representation basis for the training manifold.

[0120] During the prediction process, the uncertainty assessment unit maps the current input data to a feature space consistent with that of the training phase and evaluates the degree of deviation between the current input and the training data distribution based on the statistical properties of the training manifold. When the deviation exceeds a preset threshold, the current input sample is determined to be outside the coverage of the training data and is marked as an out-of-distribution sample. Simultaneously, the uncertainty assessment result is increased accordingly to reflect the reduced predictive reliability of the model under this input condition. This approach enables effective identification of unknown or abnormal operating conditions, preventing the model from outputting unreasonably high-confidence predictions when exceeding the training range.

[0121] The confidence management unit receives uncertainty assessment results and multi-level model output data, establishes a multi-layer uncertainty propagation mechanism, calculates local uncertainty values ​​at the feature extraction layer, state assessment layer, and life prediction layer, and generates a complete probability distribution by tracking the cumulative and amplified effects of uncertainty at each layer through the error propagation formula. For different failure types such as bearing failure, gear failure, structural fatigue failure, and lubrication failure, the confidence management unit sets dynamic confidence thresholds based on historical data statistics and the severity of consequences, generating three confidence level prediction intervals (high, medium, and low) by comparing the uncertainty levels of the prediction results. A narrower high-confidence prediction interval indicates reliable prediction results, while a wider low-confidence prediction interval suggests significant uncertainty requiring careful decision-making. This allows decision-makers to select appropriate prediction intervals for planning based on risk tolerance and maintenance resource constraints.

[0122] The confidence management unit incorporates a multi-layered uncertainty propagation mechanism to describe the transmission and accumulation of uncertainty across different levels of the model. Specifically, the confidence management unit quantifies the uncertainty of the model output at the feature extraction layer, state assessment layer, and lifetime prediction layer, and uses the uncertainty results from each layer as input for subsequent layer uncertainty calculations. During uncertainty propagation, the impact of the uncertainty from the previous layer on the output of the current layer is evaluated based on the error propagation principle, and this is combined with the uncertainty introduced by the current layer itself to form a layer-by-layer accumulated uncertainty description. This approach comprehensively reflects the propagation path and cumulative effect of uncertainty within the model.

[0123] To address the confidence management needs of different fault types, the confidence management unit sets dynamic confidence thresholds based on the statistical characteristics of historical operating data and the severity of the corresponding fault consequences. Specifically, for each type of fault, the distribution characteristics of prediction error or lifespan deviation are first determined based on historical samples, forming a corresponding confidence interval. Then, the confidence interval is adjusted based on the impact of this type of fault on equipment safety and operational reliability. For fault types with high consequences, the sensitivity of low-confidence judgments is increased to strengthen risk warnings; for fault types with relatively low consequences, the confidence interval is appropriately widened to balance prediction stability and maintenance costs. Through this method, multi-level confidence prediction intervals matching the risk level of different fault types are generated.

[0124] By comprehensively implementing out-of-distribution sample identification, uncertainty propagation, and dynamic confidence threshold setting, the uncertainty quantification module in this embodiment can provide an uncertainty description corresponding to the prediction reliability while outputting the prediction result. This gives the prediction result a clear confidence level indication and risk reference value, providing a more comprehensive and actionable decision-making basis for the subsequent intelligent early warning module.

[0125] Existing prediction systems typically output deterministic numerical results without providing confidence information, making it difficult for decision-makers to assess the reliability of the predictions. They also lack quantitative risk assessment criteria when facing high-risk decisions and lack mechanisms to identify abnormal operating conditions beyond the training data range. This application's uncertainty quantification module, through the cooperation of an uncertainty assessment unit and a confidence management unit, introduces uncertainty quantification and confidence management into the prediction system. The uncertainty assessment unit separates and quantifies cognitive uncertainty and stochastic uncertainty through variational inference and Monte Carlo sampling, and identifies abnormal operating conditions beyond the training range through out-of-distribution detection. The confidence management unit tracks the accumulation and amplification effects of uncertainty at each layer of the network through a multi-layer uncertainty propagation mechanism, and generates multi-level prediction intervals for different fault types through dynamic confidence thresholds. Therefore, the prediction results output by the uncertainty quantification module outperform existing technologies in terms of confidence assessment, risk identification, and decision support. The confidence information accompanying the prediction results enables decision-makers to adjust maintenance strategies based on the level of reliability. Out-of-distribution detection reduces the risk of misjudgment by the model under unknown operating conditions. Multi-level prediction intervals provide quantitative basis for decision-making based on different risk preferences and provide more complete prediction information for the intelligent early warning module. This enhances the practical value of the system in terms of the interpretability and operability of the prediction results.

[0126] like Figure 1 The intelligent early warning module includes a risk assessment unit and a visualization unit.

[0127] The risk assessment unit receives the quantitative prediction results, extracts the failure probability data and remaining usage time data, combines the failure consequence severity database and maintenance resource availability data, integrates the failure probability, consequence severity, time urgency and resource constraint weight coefficients through a weighted summation method to calculate the comprehensive risk index, classifies the risk index into blue monitoring level, yellow attention level, orange warning level and red alarm level according to the range of risk index values, calls the cost-benefit optimization model to solve for the optimal maintenance time window and maintenance strategy scheme, and outputs decision information;

[0128] The visualization unit receives decision-making information and quantitative prediction results, constructs a digital twin 3D model, maps health status data to corresponding components and identifies levels through color coding, highlights degraded areas and labels the type, overlays temperature and stress field distributions with pseudo-color cloud maps, generates time series curves to display historical and predicted trends and adds confidence interval error bands, displays warning levels with icons and text, displays risk indices on dashboards, marks maintenance times on the timeline, displays maintenance plan costs and benefits, and outputs to terminal devices.

[0129] In this embodiment: The risk assessment unit receives the quantitative prediction results output by the uncertainty quantification module, extracts fault probability data and remaining usage time data, and combines them with the fault consequence severity database and maintenance resource availability data. It then calculates a comprehensive risk index by integrating the weighted coefficients of fault probability, consequence severity, time urgency, and resource constraints using a weighted summation method. Based on the risk index value range, the risk assessment unit classifies the system into four warning levels: blue monitoring level, yellow attention level, orange warning level, and red alert level. It then calls a cost-benefit optimization model to solve for the optimal maintenance time window and maintenance strategy, comprehensively considering factors such as maintenance costs, downtime losses, fault risks, and resource constraints. This achieves the transformation from prediction results to maintenance decisions, avoiding problems such as inappropriate maintenance timing or unreasonable resource allocation that may result from relying solely on experience.

[0130] The visualization unit receives decision-making information and quantitative prediction results from the risk assessment unit, constructs a digital twin 3D model, maps health status data to corresponding components, identifies levels through color coding, highlights degraded areas and labels fault types, and overlays temperature and stress field distributions onto a pseudo-color cloud map, presenting physical field information that cannot be directly observed inside the equipment in an intuitive visual form. The visualization unit generates time-series curves to display historical data and predicted trends, adding confidence interval error bands. It displays warning levels in both icon and text format, shows risk indices in dashboard format, marks recommended maintenance times on the timeline, and displays the costs and expected benefits of maintenance plans, thus providing comprehensive information support for decision-makers and presenting complex predictive data and decision recommendations in an easy-to-understand and operational manner.

[0131] Existing early warning systems typically trigger alarms based on fixed thresholds, failing to comprehensively consider multi-dimensional factors such as fault probability, severity of consequences, time urgency, and resource constraints. This leads to frequent false alarms or missed alarms, making it difficult for decision-makers to directly formulate maintenance plans based on alarm information. Information presentation is often limited to simple numerical tables or two-dimensional graphs, lacking intuitive three-dimensional visualization and physical field distribution displays. The intelligent early warning module of this application integrates comprehensive risk assessment and intuitive information presentation into the early warning decision-making process through the cooperation of a risk assessment unit and a visualization unit. The risk assessment unit calculates a comprehensive risk index through multi-factor weighted summation and classifies it into four early warning levels. It solves for the optimal maintenance time window and strategy scheme through a cost-benefit optimization model. The visualization unit intuitively displays equipment status and physical field distribution through a digital twin 3D model and pseudo-color cloud map, shows predictive trends and uncertainties through time series graphs and confidence intervals, and displays early warning levels, risk indices, and maintenance recommendations through icons, dashboards, and timelines. Therefore, the intelligent early warning module outperforms existing technologies in three aspects: comprehensive risk assessment, maintenance decision optimization, and information visualization. Multi-dimensional risk assessment reduces false alarm and false alarm rates, maintenance decision optimization improves resource allocation efficiency, and intuitive visualization shortens the information interpretation time for decision-makers. It enhances the system's support capability for maintenance management in terms of the scientific nature and operability of early warning decisions.

[0132] like Figure 1 The human-machine collaboration module includes a knowledge interaction unit and an explainable decision-making unit.

[0133] The knowledge interaction unit provides a knowledge input interface to receive domain knowledge input from engineers, including fault mode descriptions, physical law expressions, material performance parameters, and operation and maintenance experience rules. It uses natural language processing technology for semantic parsing and entity recognition, extracts key physical parameters, constraints, and causal relationships, and transforms them into physical constraint expressions to update the physical modeling module or into data annotation information to label the training dataset. It establishes a conversational inquiry mechanism to receive questions from engineers, retrieves reasoning process records and historical case data to generate response information, and outputs knowledge update results and dialogue records.

[0134] The interpretable decision unit receives decision information and model inference process data, generates interpretable inference paths, including input feature importance ranking, intermediate layer feature activation modes, and output result contribution decomposition, extracts key physical quantity time series data to display the evolution process and marks abrupt change points, identifies dominant degradation mechanisms such as fatigue crack propagation, wear accumulation, thermal stress damage, and corrosion deterioration, retrieves similar cases from the historical failure case database and calculates similarity values ​​for ranking and display, generates multiple maintenance strategy alternatives for high-risk and high-uncertainty scenarios, and outputs decision support information to the visualization unit.

[0135] In this embodiment: The knowledge interaction unit provides a knowledge input interface to receive domain knowledge input from engineers, including fault mode descriptions, physical law expressions, material performance parameters, and operation and maintenance experience rules. Natural language processing technology is used to perform semantic parsing and entity recognition on the input content, extracting key physical parameters, constraints, and causal relationships. The extracted results are then converted into physical constraint expressions to update the physical modeling module, or into data annotation information and labeled into the training dataset. The knowledge interaction unit establishes a dialogic inquiry mechanism to receive questions from engineers regarding the model's reasoning results. It retrieves relevant reasoning process records and historical case data to generate targeted response information. Engineers can understand the specific basis of the model's predictions through inquiries, and can input corrective knowledge when they find discrepancies between the model's reasoning and the actual situation. The knowledge interaction unit completes the conversion of engineers' experience knowledge into a usable form for the model and the interactive verification of the model's reasoning logic. This allows the model to continuously receive knowledge supplementation and logical verification from human experts during operation, thereby improving the model's adaptability to specific field conditions and equipment characteristics.

[0136] During the knowledge transformation process, the knowledge interaction unit verifies the consistency between the extracted results and the existing model structure and physical modeling rules to ensure that the resulting constraint information is logically reasonable and feasible in the model. Empirical descriptions that cannot be directly used for model constraint updates are retained in the form of rules or explanatory information for the interpretation or manual verification of subsequent model inference results, thereby avoiding the direct participation of ambiguous or incomplete information in model calculations.

[0137] The explainable decision unit receives decision information and model inference process data, generates an explainable inference path, including input feature importance ranking, intermediate layer feature activation patterns, and output contribution decomposition. It extracts time-series data of key physical quantities to display the evolution process and marks abrupt change points, identifying dominant degradation mechanisms such as fatigue crack propagation, wear accumulation, thermal stress damage, and corrosion degradation. The explainable decision unit retrieves similar cases from a historical failure case database, calculates similarity values, and ranks them for display. For high-risk, high-uncertainty scenarios, it generates multiple maintenance strategy alternatives, enabling decision-makers to understand the basis of model predictions. Through feature importance ranking, it understands the impact of monitoring parameters on predictions; through degradation mechanism identification, it understands the main failure modes currently faced by the equipment; and through similar case retrieval, it verifies decisions based on historical experience.

[0138] When generating interpretable inference paths, the interpretable decision unit analyzes the formation process of the model output results. By evaluating the degree to which the model output responds to changes in input features, it determines the relative roles of different input features in the prediction process and decomposes the contribution of the prediction results accordingly. This decomposition result is used to illustrate the degree of influence of each monitoring parameter and intermediate state variable on the prediction results, thus providing a clear basis for the prediction conclusions.

[0139] The interpretable decision unit (IDU) combines information on changes in the intermediate state layers of the model to extract the time-series evolution characteristics of key physical quantities, analyze their trends, and mark the corresponding key moments when significant changes occur. This helps to explain the correlation between changes in equipment state and prediction results. In this way, the IDU transforms the model's internal reasoning process into analytical information that engineers can understand, enabling prediction results to provide clear and traceable explanatory evidence while offering conclusions.

[0140] Existing prediction systems often employ end-to-end black-box models, lacking interpretability in their reasoning processes. Decision-makers struggle to understand the basis for predictions, and accuracy is difficult to assess when results contradict empirical judgments. Furthermore, engineers' accumulated domain knowledge and operational experience are difficult to effectively integrate into the model. This application's human-machine collaboration module, through the cooperation of a knowledge interaction unit and an interpretable decision-making unit, integrates human expert knowledge and makes the decision-making process transparent into the system's operation. The knowledge interaction unit uses natural language processing to transform engineer-input domain knowledge into usable physical constraints and data annotations for the model. A conversational inquiry mechanism supports engineers in questioning and verifying the model's reasoning logic. The interpretable decision-making unit reveals the model's reasoning path through feature importance analysis and contribution decomposition, identifies and explains the basis for predictions through key physical quantity evolution and degradation mechanisms, and supports decision verification and scheme comparison through similar case retrieval and multi-solution generation. Thus, the human-machine collaboration module outperforms existing technologies in knowledge fusion, reasoning transparency, and decision credibility.

[0141] like Figure 1 The online optimization module includes a performance monitoring unit and an audit migration unit.

[0142] The performance monitoring unit tracks prediction accuracy, false alarm rate, and false negative rate in real time. It compares actual fault conditions with prediction results to calculate performance index values. It uses statistical process control methods to establish control charts and set upper and lower control limits. When the performance index exceeds the control limits, it is judged as performance degradation and triggers model updates. It uses elastic weight consolidation technology or memory playback technology to update model parameters online. It implements an active learning strategy to identify high-value samples for priority labeling and incremental training. It establishes a model version management mechanism to record update information. It implements an A / B testing mechanism to compare the performance of the old and new versions and then replaces the model. It outputs updated model parameters and performance monitoring reports.

[0143] The audit migration unit periodically audits the physical constraint layer's compliance, extracts predicted physical quantity values ​​to verify the degree of conformity with the law of conservation of energy, material constitutive relations, and crack propagation laws, and triggers the calibration process when the calculated physical residual deviation value exceeds a preset threshold. It also periodically audits the calibration accuracy of the uncertainty quantification module, collects predicted probability distribution and actual fault frequency data to calculate the deviation value of the calibration curve and adjust parameters, integrates cross-equipment and cross-mine fault case data to construct a cross-domain database, uses transfer learning technology to pre-train in the source domain and then fine-tunes parameters in the target domain, aligns feature distributions through domain adaptation methods to achieve knowledge reuse, and outputs audit reports and transfer learning model parameters to the performance monitoring unit.

[0144] In this embodiment, the performance monitoring unit tracks performance indicators such as prediction accuracy, false alarm rate, and false negative rate in real time. It compares actual fault conditions with predicted results to calculate indicator values, establishes a control chart using statistical process control methods, and sets upper and lower control limits. When performance indicators exceed these limits, performance degradation is identified, triggering a model update process. During model updates, elastic weight consolidation or memory playback techniques are used to update model parameters online. An active learning strategy is implemented to identify high-value samples with large amounts of information, prioritizing manual annotation and incremental training. A model version management mechanism records the time, reason, and performance changes of each update. An A / B testing mechanism compares the performance of old and new versions in a real-world operating environment to determine whether to replace the model. The performance monitoring unit completes the tasks of continuous monitoring and adaptive updating of model performance. When equipment operating conditions change or new fault modes not covered during the training phase occur, the model can absorb information from the new data through the online learning mechanism, avoiding performance degradation due to operating condition drift over time.

[0145] During model updates, the performance monitoring unit employs a restricted update strategy for adjusting model parameters to minimize disruption to existing knowledge. Specifically, by identifying parameters that played a major role in prediction results during historical training, the update magnitude of these parameters is limited. Simultaneously, historical sample information is incorporated into the update process, allowing the model to absorb new data features while retaining its memory of existing operating conditions. For newly collected data samples, the performance monitoring unit, considering the model's current prediction state, prioritizes samples that significantly improve the model's discriminative ability for manual annotation and incremental training, thereby achieving effective performance improvement with limited annotation costs. After the model update is completed, relevant parameter changes and performance comparison results are recorded and used for version management to support subsequent backtracking and effectiveness evaluation.

[0146] The audit migration unit periodically audits the fulfillment of physical constraint layers, extracts predicted physical quantities from the model output to verify whether they conform to physical constraints such as the law of conservation of energy, material constitutive relations, and crack propagation laws, calculates physical residual deviation values, and triggers a calibration process to adjust physical constraint layer parameters when the deviation value exceeds a preset threshold. It also periodically audits the calibration accuracy of the uncertainty quantification module, collecting predicted probability distribution data and actual fault frequency data over a period of time to calculate calibration curves, evaluate the degree of agreement between predicted probabilities and actual frequencies, and adjust relevant parameters. The audit migration unit integrates fault case data from different equipment and mines to construct a cross-domain database. Using transfer learning technology, it pre-trains the model in the data-rich source domain and then fine-tunes the parameters in the data-scarce target domain. Through domain adaptation methods, it aligns the feature distribution differences between different domains, completing periodic verification of model physical consistency and cross-device, cross-scenario knowledge reuse. This allows model knowledge trained on data-rich devices to be transferred and applied to new devices with scarce data.

[0147] The audit migration unit periodically verifies the key constraints and output characteristics of the model during model operation. Specifically, it analyzes the predicted physical quantities output by the model to verify whether they meet the preset physical constraints. When deviations exceed the allowable range, it triggers a calibration process for the relevant physical constraint parameters to restore the model's adherence to physical laws. Simultaneously, it performs statistical audits on the probabilistic predictions output by the model, assessing the rationality of the uncertainty quantification results by comparing the predicted probabilities with the frequency of actual events, and adjusting relevant parameters as necessary.

[0148] In cross-device or cross-scenario applications, the audit migration unit integrates operational data and fault cases from different sources to form a reference dataset that can be used for model migration. By pre-training the model in scenarios with sufficient data, and then adjusting the parameters using the limited data of the target scenario, the model can adapt to the characteristics of new devices or new operating conditions while maintaining general knowledge. This process avoids the data and time costs of completely retraining the model in new scenarios, while ensuring the consistency and stability of the model in different application environments.

[0149] Existing prediction systems typically operate with fixed model parameters after deployment. When equipment operating conditions change or new failure modes emerge, model performance gradually degrades, requiring manual collection of new data for offline retraining. Furthermore, there is a lack of regular auditing mechanisms for the physical consistency and uncertainty quantification accuracy of the model. When deploying new equipment, data must be accumulated from scratch to train a dedicated model. This application's online optimization module integrates continuous model optimization and cross-domain knowledge reuse into the system operation process through the cooperation of a performance monitoring unit and an audit migration unit. The performance monitoring unit monitors model performance in real time using statistical process control methods and triggers updates when performance degradation is detected. It achieves efficient online incremental updates of parameters through elastic weight consolidation and active learning strategies, and ensures the traceability and controllability of model updates through version management and A / B testing mechanisms. The audit migration unit ensures the physical consistency and probabilistic prediction accuracy of the model by regularly auditing physical constraints and uncertainty calibration, and achieves knowledge transfer between different devices and scenarios through transfer learning and domain adaptation techniques. Therefore, the online optimization module outperforms existing technologies in terms of model adaptability, physical consistency maintenance, and cross-domain knowledge reuse. The model's performance automatically adjusts to changes in operating conditions, maintaining the accuracy of predictions. The regular auditing mechanism ensures the reliability of the model's output. Transfer learning technology shortens the deployment cycle of new equipment and reduces data requirements. From the perspective of continuous system improvement capabilities, it enhances the long-term stability of the prediction system and the efficiency of engineering promotion.

[0150] This application also includes a deep learning-based intelligent predictive maintenance method for mining equipment, such as... Figure 2 The specific steps are as follows:

[0151] S1. The data acquisition synchronization module acquires equipment operation data from vibration sensors, temperature sensors, strain sensors and acoustic emission sensors, and performs timestamp alignment processing on data with different sampling rates to form a multi-source fusion dataset; the physical preprocessing module detects outliers based on physical constraint boundaries and inverts the internal physical field distribution, calls the simulation model to generate virtual fault samples, mixes them with the measured data and normalizes them to form an enhanced dataset.

[0152] S2. The feature extraction module uses a multi-branch convolutional neural network to extract features at different time scales, constructs a graph neural network model to infer the associated state of components, and fuses multi-scale features and graph features through an adaptive attention mechanism.

[0153] S3. The physical modeling module designs a physical constraint layer to encode physical laws into a differentiable network layer, constructs a dual-path structure of physical calculation path and neural network path and merges them through a gating mechanism, introduces physical residual connections in each hidden layer and trains them using a course learning strategy.

[0154] S4. The uncertainty quantification module uses variational inference to assess cognitive uncertainty, identifies abnormal operating conditions through an externally distributed detector, and establishes a multi-level uncertainty propagation mechanism to generate multi-level confidence prediction intervals.

[0155] S5. The intelligent early warning module integrates the failure probability, severity of consequences, time urgency, and resource constraints to calculate a comprehensive risk index, classifies early warning levels, and solves the optimal maintenance strategy. The equipment status and maintenance suggestions are visualized through a digital twin model. The human-machine collaboration module provides a knowledge input interface and a conversational query mechanism to realize human-machine collaboration.

[0156] S6. The online optimization module monitors the model's performance indicators in real time and triggers online updates when performance degrades. It also periodically audits the accuracy of physical consistency and uncertainty calibration and achieves cross-device knowledge reuse through transfer learning technology.

[0157] The contents not described in detail in this specification are existing technologies known to those skilled in the art.

[0158] It should be noted that, as will be apparent to those skilled in the art, the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics thereof. The scope of the present invention is defined by the claims rather than the foregoing description.

Claims

1. A deep learning-based intelligent predictive maintenance system for mining equipment, characterized in that: It includes, in sequence, a data acquisition and synchronization module, a physical preprocessing module, a feature extraction module, a physical modeling module, an uncertainty quantification module, an intelligent early warning module, a human-machine collaboration module, and an online optimization module; among which: The data acquisition and synchronization module obtains operational data from vibration, temperature, strain, and acoustic emission sensors and performs timestamp alignment processing. The physical preprocessing module detects outliers based on the physical constraint boundaries of the equipment, generates virtual fault samples using simulation, and reconstructs the internal physical field distribution. The feature extraction module uses a multi-branch convolutional neural network to extract features at different time scales and constructs a graph neural network to infer the associated states of components. The physical modeling module design encodes conservation laws and constitutive relations into a differentiable network layer in the physical constraint layer, and adopts a dual-path structure to fuse analytical models and neural networks. The uncertainty quantification module assesses cognitive uncertainty through variational inference and establishes an out-of-distribution detector and a multi-layer propagation mechanism. The intelligent early warning module comprehensively calculates the risk index and establishes a four-level early warning mechanism, and visualizes the equipment status through a digital twin interface; The human-machine collaboration module provides a knowledge input interface and an interpretable reasoning path, and establishes a conversational inquiry mechanism; The online optimization module monitors performance metrics in real time, uses incremental learning to update the model, and integrates cross-device fault cases to achieve knowledge reuse.

2. The intelligent predictive maintenance system for mining equipment based on deep learning as described in claim 1, characterized in that, The data acquisition and synchronization module includes a multi-source data acquisition unit and a time-series synchronization processing unit. The multi-source data acquisition unit collects operating data from the equipment's vibration sensors, temperature sensors, strain sensors, acoustic emission sensors, and oil analysis sensors. It also obtains load values, speed values, and pressure values ​​from the control system. The unit performs multi-rate acquisition for different sampling frequencies and adds timestamps. Through signal amplitude checks, sampling rate analysis, and continuity judgment, it detects sensor drift, faults, or missing data, and generates quality markers and alarm information. The timing synchronization processing unit receives multi-source data streams with timestamps and quality label information, establishes a unified time reference, downsamples high-frequency data or interpolates low-frequency data, maps the measured values ​​to a unified time node, identifies and isolates abnormal data, generates a timing-synchronized multi-source fusion dataset, and outputs it to the physical preprocessing module.

3. The intelligent predictive maintenance system for mining equipment based on deep learning as described in claim 1, characterized in that, The physical preprocessing module includes a constraint detection unit and a simulation enhancement unit; The constraint detection unit receives the multi-source fusion dataset output by the data acquisition synchronization module, detects abnormal values ​​based on the temperature change rate threshold, vibration amplitude limit, stress-strain relationship range, and material performance parameter boundary, distinguishes between real extreme working conditions and sensor errors and marks them, and solves the heat conduction equation and elasticity equation based on sparse measurement point data, geometric structure model and material property parameters to invert the internal temperature field and stress field distribution, and outputs the detected measured data and reconstructed physical field data. The simulation enhancement unit receives the measured data and physical field data, calls the finite element model and multibody dynamics model, changes the load spectrum, boundary conditions, material properties and environmental condition parameters to generate virtual sensor response data, simulates the occurrence and evolution of faults for scarce fault samples to generate virtual fault data, mixes them to form an enhanced dataset, normalizes it according to physical dimensions and coupling relationships, considers causal relationships and coupling strength weights, and outputs it to the feature extraction module.

4. The intelligent predictive maintenance system for mining equipment based on deep learning as described in claim 1, characterized in that, The feature extraction module includes a multi-scale extraction unit and a graph reasoning unit; The multi-scale extraction unit receives the normalized enhanced dataset output by the physical preprocessing module, and uses a multi-branch convolutional neural network to process the time series data in parallel. The short-time branch, medium-time branch, and long-time branch use convolutional kernels of increasing size to extract transient impact features, periodic temperature change features, and long-term fatigue accumulation features, respectively. Each branch independently completes convolution and pooling operations and outputs corresponding feature vectors. The three types of feature vectors are concatenated in the feature dimension to form a multi-scale fused feature vector. The graph inference unit receives the multi-scale fused feature vector, constructs a graph neural network model to represent the device topology, where nodes correspond to components, edges represent physical connections and energy transfer paths, and the fused feature vector is used as the initial feature input for each node. It then propagates feature information through graph convolution operations to infer the health status of adjacent components, outputting a graph feature representation. The adaptive attention mechanism calculates weight coefficients based on degradation stage parameters, weights features at different scales and fuses them with the graph features, and outputs them to the physical modeling module.

5. The intelligent predictive maintenance system for mining equipment based on deep learning as described in claim 1, characterized in that, The physical modeling module includes a physical coding unit and a residual supervision unit; The physical coding unit receives the final feature representation output by the feature extraction module, designs a physical constraint layer to encode the law of energy conservation, material constitutive relations and crack propagation laws into a differentiable network layer, constructs a dual-path parallel structure, the physical calculation path runs a simplified analytical model to calculate the state evolution of the computing device, and the neural network calculation path uses a multilayer perceptron to learn complex nonlinear relationships, introduces a gating fusion mechanism to calculate the fusion weight coefficients according to the input feature distribution characteristics, and weights the outputs of the two paths to form a fusion state representation; The residual supervision unit receives the fusion state representation, introduces physical residual connections in each hidden layer of the neural network, calculates the deviation between the current layer output and the prediction of the physical constraint layer as a residual term and adds it to the loss function, adopts a course learning strategy, sets a low weight coefficient to fit the data in the early stage of training, gradually increases the weight coefficient in the middle stage to guide physical consistency convergence, fixes a high weight coefficient in the later stage, and outputs the trained physical embedded prediction model to the uncertainty quantification module.

6. The intelligent predictive maintenance system for mining equipment based on deep learning as described in claim 1, characterized in that, The uncertainty quantification module includes an uncertainty assessment unit and a confidence management unit; The uncertainty assessment unit receives the physical embedded prediction model output by the physical modeling module, uses variational inference to convert the network weight parameters into a probability distribution, performs multiple forward propagation calculations using Monte Carlo sampling to assess the statistical dispersion of the prediction results and evaluate cognitive uncertainty, adds random perturbations to the input data for testing, evaluates the change in the data augmentation output as random uncertainty, constructs an out-of-distribution detector to calculate the Mahalanobis distance between the current input and the training manifold, marks it as an out-of-distribution sample when it exceeds a preset threshold and increases the uncertainty assessment value, and outputs the uncertainty assessment result. The confidence management unit receives the uncertainty assessment results and multi-level output data, establishes a multi-layer uncertainty propagation mechanism, calculates local uncertainty values ​​in the feature extraction layer, state assessment layer, and life prediction layer, tracks the cumulative and amplification effects of each layer through the error propagation formula to generate a complete probability distribution, sets dynamic confidence thresholds for bearing failure, gear failure, structural fatigue failure, and lubrication failure based on historical data and the severity of consequences, compares and generates high, medium, and low confidence level prediction intervals, and outputs quantitative prediction results to the intelligent early warning module.

7. The intelligent predictive maintenance system for mining equipment based on deep learning as described in claim 1, characterized in that, The intelligent early warning module includes a risk assessment unit and a visualization unit; The risk assessment unit receives the quantitative prediction results output by the uncertainty quantification module, extracts the failure probability data and remaining usage time data, combines the failure consequence severity database and maintenance resource availability data, integrates failure probability, consequence severity, time urgency and resource constraint weight coefficients through a weighted summation method to calculate a comprehensive risk index, classifies the risk index into blue monitoring level, yellow attention level, orange warning level and red alarm level according to the range of risk index values, calls the cost-benefit optimization model to solve for the optimal maintenance time window and maintenance strategy scheme, and outputs decision information; The visualization unit receives the decision information and quantitative prediction results, constructs a digital twin 3D model, maps health status data to corresponding components and identifies levels through color coding, highlights degraded areas and labels the type, overlays temperature and stress field distributions with pseudo-color cloud maps, generates time series curves to display historical and predicted trends and adds confidence interval error bands, displays warning levels with icons and text, displays risk indices on dashboards, marks maintenance times on the timeline, displays maintenance plan costs and benefits, outputs to terminal devices and synchronizes with the human-machine collaboration module.

8. The intelligent predictive maintenance system for mining equipment based on deep learning as described in claim 1, characterized in that, The human-machine collaboration module includes a knowledge interaction unit and an explainable decision-making unit; The knowledge interaction unit provides a knowledge input interface to receive domain knowledge input by engineers, including fault mode descriptions, physical law expressions, material performance parameters, and operation and maintenance experience rules. It uses natural language processing technology for semantic parsing and entity recognition, extracts key physical parameters, constraints, and causal relationships, and transforms them into physical constraint expressions to update the physical modeling module or into data annotation information to mark the training dataset. It establishes a conversational inquiry mechanism to receive engineers' questions, retrieves reasoning process records and historical case data to generate response information, and outputs knowledge update results and dialogue records. The interpretable decision unit receives decision information and model reasoning process data from the intelligent early warning module, generates an interpretable reasoning path, including input feature importance ranking, intermediate layer feature activation mode, and output result contribution decomposition, extracts key physical quantity time series data to display the evolution process and marks abrupt change points, identifies fatigue crack propagation, wear accumulation, thermal stress damage, and corrosion degradation as the dominant degradation mechanisms, retrieves similar cases from the historical failure case database and calculates similarity values ​​for ranking and display, generates multiple maintenance strategy alternatives for high-risk and high-uncertainty scenarios, and outputs decision support information to the visualization unit.

9. The intelligent predictive maintenance system for mining equipment based on deep learning as described in claim 1, characterized in that, The online optimization module includes a performance monitoring unit and an audit migration unit; The performance monitoring unit tracks prediction accuracy, false alarm rate, and false negative rate in real time. It compares the actual fault conditions with the prediction results to calculate the performance index values. It uses statistical process control methods to establish a control chart and set upper and lower control limits. When the performance index exceeds the control limits, it is judged as performance degradation and triggers model update. It uses elastic weight consolidation technology or memory playback technology to update the model parameters online. It implements an active learning strategy to identify high-value samples for priority labeling and incremental training. It establishes a model version management mechanism to record update information. It implements an A / B testing mechanism to compare the performance of the old and new versions and replace the model. It outputs the updated model parameters and performance monitoring report. The audit migration unit periodically audits the physical constraint layer's compliance, extracts predicted physical quantity values ​​to verify the degree of conformity with the law of conservation of energy, material constitutive relations, and crack propagation laws, and triggers the calibration process when the calculated physical residual deviation value exceeds a preset threshold. It also periodically audits the calibration accuracy of the uncertainty quantification module, collects predicted probability distribution and actual fault frequency data to calculate the deviation value of the calibration curve and adjusts the parameters, integrates cross-equipment and cross-mine fault case data to construct a cross-domain database, uses transfer learning technology to pre-train in the source domain and then fine-tunes the parameters in the target domain, aligns the feature distribution through a domain adaptive method to achieve knowledge reuse, and outputs audit reports and transfer learning model parameters to the performance monitoring unit.

10. A method for intelligent predictive maintenance of mining equipment based on deep learning, characterized in that, This method, based on the deep learning-based intelligent predictive maintenance system for mining equipment as described in claim 1, includes the following steps: S1. The data acquisition synchronization module acquires equipment operation data from vibration sensors, temperature sensors, strain sensors and acoustic emission sensors, and performs timestamp alignment processing on data with different sampling rates to form a multi-source fusion dataset; the physical preprocessing module detects outliers based on physical constraint boundaries and inverts the internal physical field distribution, calls the simulation model to generate virtual fault samples, mixes them with the measured data and normalizes them to form an enhanced dataset; S2. The feature extraction module uses a multi-branch convolutional neural network to extract features at different time scales, constructs a graph neural network model to infer the component association state, and fuses multi-scale features and graph features through an adaptive attention mechanism. S3. The physical modeling module designs a physical constraint layer to encode physical laws into a differentiable network layer, constructs a dual-path structure of physical calculation path and neural network path and merges them through a gating mechanism, introduces physical residual connections in each hidden layer and trains them using a course learning strategy. S4. The uncertainty quantification module uses variational inference to assess cognitive uncertainty, identifies abnormal working conditions through an out-of-distribution detector, and establishes a multi-level uncertainty propagation mechanism to generate multi-level confidence prediction intervals. S5. The intelligent early warning module integrates fault probability, severity of consequences, time urgency, and resource constraints to calculate a comprehensive risk index, classifies early warning levels, and solves for the optimal maintenance strategy. The equipment status and maintenance suggestions are visualized through a digital twin model. The human-machine collaboration module provides a knowledge input interface and a conversational query mechanism to realize human-machine collaboration. S6. The online optimization module monitors the model's performance indicators in real time and triggers online updates when performance degrades. It also periodically audits the accuracy of physical consistency and uncertainty calibration and achieves cross-device knowledge reuse through transfer learning technology.

Images