Deep-sea mining system fault diagnosis method and system based on digital twin and cloud edge collaboration

Abstract

The application provides a digital twin and cloud edge collaborative deep-sea mining system fault diagnosis method and system, and relates to the technical field of deep-sea mining system fault diagnosis. The method comprises the following steps: constructing a digital twin model and performing mechanism monitoring; edge side signal processing and component level intelligent diagnosis; cloud data aggregation; double-engine information fusion and decision making are performed in the cloud diagnosis center; uncertainty quantification and visual output are performed. The application constructs an intelligent diagnosis architecture based on digital twin and cloud edge collaboration, deeply integrates mechanism models and artificial intelligence, realizes early warning of system level chain faults, robust diagnosis of weak component faults in a strong noise environment, and provides high credibility diagnosis conclusions with quantitative confidence and decision support information, thereby significantly improving the safety, reliability and operation intelligent level of deep-sea mining operations.

Description

Fault Diagnosis Method and System for Deep-Sea Mining Systems Based on Digital Twin and Cloud-Edge Collaboration Technical Field

[0001] This invention relates to the field of fault diagnosis technology for deep-sea mining systems, specifically to a fault diagnosis method and system for deep-sea mining systems based on digital twins and cloud-edge collaboration. Background Technology

[0002] Deep-sea mining is a crucial pathway to obtaining strategic mineral resources such as polymetallic nodules and cobalt-rich crusts from the seabed. Its core equipment—the deep-sea mining system (typically including surface support vessels, hoisting pipes, relay stations, hoisting pumps, mining vehicles, and conveying hoses)—operates for extended periods in extreme marine environments characterized by high pressure, low temperatures, strong corrosion, and complex ocean current disturbances. The strong coupling of multiple physical fields within the system and the random and variable external environmental loads result in fault modes that are characterized by their frequent occurrence, concealment, cascading effects, and catastrophic consequences. Once a fault occurs, it can cause not only enormous economic losses but also potentially trigger serious marine environmental accidents. Therefore, developing efficient and reliable fault diagnosis technologies for deep-sea mining systems is of paramount importance for ensuring operational safety, improving mining efficiency, and reducing operation and maintenance costs.

[0003] Currently, research and practice on fault diagnosis for deep-sea mining systems are mainly developing in two directions: diagnostic methods based on physical mechanism models and intelligent diagnostic methods based on data-driven approaches.

[0004] Diagnostic methods based on physical mechanism models establish precise mathematical models of the system or key components (e.g., using the lumped mass method to establish a fluid-structure interaction dynamic model of the ore lifting pipe, or establishing a hydraulic model based on the pump's characteristic curve), and set theoretical thresholds or safety margins for key state parameters (such as stress, deflection, flow velocity, and efficiency). When the measured values ​​of the sensors continuously deviate from the theoretical values ​​or exceed the thresholds, a fault alarm is triggered. The advantage of this method is that the diagnostic process has clear physical meaning and interpretability. However, its limitations are significant: First, the deep-sea environment is complex and variable, and loads and boundary conditions are difficult to accurately obtain and simulate, leading to errors in the model itself; second, the preset fixed thresholds are difficult to adapt to dynamic operating conditions such as sea state fluctuations and time-varying degradation of equipment performance, easily resulting in false alarms or missed alarms; more importantly, this method usually analyzes single components or isolated physical quantities, lacking the system-level modeling and monitoring capabilities for the dynamic coupling and fault transmission chain between subsystems such as the "ore lifting pipe-pump-relay station-mining vehicle," making it difficult to predict chain system failures caused by interactions.

[0005] Data-driven intelligent diagnostic methods, particularly deep learning, provide powerful tools for automatically extracting fault features from massive, high-dimensional sensor data (such as vibration and acoustic emission signals). For example, convolutional neural networks (CNNs) can extract spatial local features from vibration signals, while gated recurrent units (GRUs) can capture the temporal dependencies of signals. Through training, the models can accurately classify fault types in rotating components such as bearings and gears. These methods have a strong ability to fit complex nonlinear relationships, reducing the reliance on precise mechanistic models to some extent. However, it faces unique challenges in deep-sea application scenarios: First, the signal-to-noise ratio of raw deep-sea signals is extremely low, and strong environmental noise (ocean currents, mechanical resonance) can easily overwhelm the characteristic components representing early weak faults, leading to a sharp decline in model performance; Second, most deep learning models are "black boxes," lacking transparency in their diagnostic decision-making process and without reliable quantification of output results, making them difficult for maintenance personnel to fully trust and use for critical decisions in high-risk operational scenarios; Finally, existing intelligent diagnostic models mostly focus on the state identification of single components, and the diagnostic conclusions are isolated "information fragments," failing to be effectively correlated and integrated with physical mechanism information reflecting the overall operating state of the system, thus forming information silos.

[0006] In summary, the current field of deep-sea mining system fault diagnosis faces the following pressing technical bottlenecks: lack of system-level coupled fault early warning, difficulty in extracting weak fault features under strong noise environments, contradiction between static diagnostic strategies and dynamic environmental adaptability, and insufficient decision-making credibility due to the separation of mechanism and data diagnostic information. Existing single-method systems are insufficient to independently address these challenges. Therefore, there is an urgent need for a next-generation systematic solution for deep-sea mining system fault diagnosis that integrates system mechanisms and data intelligence, balances global coupling with refined local perception, and possesses anti-interference and reliable decision-making capabilities. Summary of the Invention

[0007] This invention addresses the technical challenges of multi-level fault coupling, diagnostic lag, and high false alarm rates in existing deep-sea mining systems. It constructs an intelligent diagnostic architecture based on digital twins and cloud-edge collaboration, deeply integrating mechanistic models and artificial intelligence. This architecture enables proactive early warning of system-level cascading failures, robust diagnosis of minor component faults in high-noise environments, and provides highly reliable diagnostic conclusions with quantified confidence and decision support information. Consequently, it significantly improves the safety, reliability, and operational intelligence of deep-sea mining operations.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0009] A fault diagnosis method for a deep-sea mining system based on digital twin and cloud-edge collaboration, the method comprising: constructing a multi-physics coupled digital twin model of the deep-sea mining system, and performing mechanism monitoring based on the digital twin model; deploying an edge diagnostic unit near the key equipment of the deep-sea mining system, the edge diagnostic unit performing real-time filtering, feature extraction, and preliminary component-level fault diagnosis on raw signals collected by multiple sensors, and generating a depth feature vector; uploading the preliminary diagnosis results and depth feature vector generated by the edge diagnostic unit to a cloud diagnostic center; the cloud diagnostic center synchronously running the digital twin engine, receiving its system-level anomaly judgment and location information; performing dual-engine information fusion and decision-making in the cloud diagnostic center; and performing uncertainty quantification and visualization output based on the diagnosis results formed by the cloud diagnostic center; wherein, the edge diagnostic unit performs real-time filtering, feature extraction, and preliminary component-level fault diagnosis on raw signals collected by multiple sensors, and generates a depth feature vector, specifically including:

[0010] After completing adaptive noise suppression processing of the signal, the deep-sea mining system transmits the denoised signal to a lightweight deep learning model pre-deployed in the edge diagnostic unit. This deep learning model, based on an end-to-end architecture, uses a forward propagation mechanism to perform layer-by-layer feature analysis and computation on the input signal. The signal first passes through convolutional or recurrent layers to extract local time-frequency features, then through pooling layers to reduce data dimensionality and enhance feature robustness, and finally, a fully connected layer maps the feature space to fault categories. For a specific component, the deep learning model outputs a multi-dimensional fault classification probability vector, where each dimension corresponds to a preset fault type. The numerical value in the vector represents the confidence level that the signal belongs to that fault type. This result constitutes a preliminary diagnostic conclusion at the component level, providing a rapid reference for subsequent decisions. Simultaneously, to support more complex fault tracing and health status assessment, the system extracts a high-dimensional deep feature vector from the feature fusion layer of the deep learning model. This feature fusion layer retains detailed information from the lower-level features while incorporating high-level semantic features, enabling the encoding of fault representations. These features are used for refined classification of the current fault and are also passed as input to the digital twin system.

[0011] On the other hand, the present invention also provides a system applicable to the above-mentioned fault diagnosis method for deep-sea mining systems using digital twins and cloud-edge collaboration, the system comprising:

[0012] The physical sensing layer consists of a network of various sensors deployed on key components of the deep-sea mining system, used to collect raw data in real time.

[0013] The edge diagnostic layer consists of edge diagnostic units deployed in the near field of critical equipment. Each edge diagnostic unit is responsible for real-time filtering, feature extraction, and preliminary diagnosis of component-level faults in the raw signals collected by various sensors, and generates deep feature vectors.

[0014] The cloud-based intelligent layer, located on a surface support vessel or shore-based data center, has a cloud-based diagnostic server at its core. This server runs three main modules: a digital twin engine module for building and running a system-level multiphysics coupling model to achieve mechanism diagnosis; a fusion decision module for rule-based information fusion and arbitration of the diagnostic results from the edge layer and the mechanism diagnosis results from the digital twin; and an uncertainty quantification module for calculating the confidence level of the final diagnostic results.

[0015] Compared with the prior art, the beneficial effects of the present invention are:

[0016] 1. Advanced Early Warning Capability for System-Level Coupled Faults: Traditional methods often focus on single equipment or localized faults, while this invention achieves real-time monitoring and advanced early warning of coupled faults across the entire system by constructing a three-layer architecture of "physical perception - edge diagnosis - cloud intelligence." In deep-sea mining, the dynamic responses of mining machines, pump systems, and surface support systems are interconnected, and a single fault can trigger a chain reaction. This invention collects multi-source data (such as vibration, stress, and flow velocity) in real time at the edge layer, combined with joint analysis by a dual-engine approach (mechanism engine + data engine) in the cloud, enabling early identification of coupled fault risks and preventing the escalation of accidents.

[0017] 2. Robust Diagnosis of Weak Faults in High-Noise Environments: The deep-sea environment is characterized by high noise and dust, making it easy for traditional signal processing methods to miss weak faults. This invention proposes an edge diagnosis model combining "adaptive noise suppression + lightweight deep learning." Through real-time filtering and feature extraction of edge nodes, combined with cloud-based optimization, the model significantly improves the accuracy of weak fault identification.

[0018] 3. Real-time Performance and Computational Efficiency Optimization under Resource Constraints: Deep-sea mining systems have extremely high real-time requirements, but edge devices have limited computing resources. This invention adopts a cloud-edge collaborative strategy of "preliminary edge diagnosis + in-depth cloud analysis," where edge nodes only process key features (such as variance), while complex models run in the cloud, achieving dynamic allocation of computational load.

[0019] Other features and advantages of the embodiments of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0020] Figure 1 is a schematic diagram of the fault diagnosis system architecture for deep-sea mining systems provided by the present invention;

[0021] Figure 2 is a flowchart of the fault diagnosis method for deep-sea mining system provided by the present invention;

[0022] Figure 3 is a schematic diagram of the uncertainty quantification and visualization output provided by the present invention. Detailed Implementation

[0023] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings, so as to more clearly understand the purpose, features and advantages of this invention. It should be understood that the embodiments shown in the drawings are not intended to limit the scope of this invention, but are only for illustrating the essential spirit of the technical solutions of this invention. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this invention.

[0024] Unless the context requires otherwise, throughout the specification and claims, the word “comprising” and its variations, such as “including” and “having”, shall be understood to have an open, inclusive meaning, that is, to be interpreted as “including, but not limited to”.

[0025] Throughout this specification, references to "an embodiment" or "an embodiment" indicate that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment. Therefore, the appearance of "in an embodiment" or "an embodiment" in various places throughout the specification does not necessarily refer to the same embodiment. Furthermore, a particular feature, structure, or characteristic may be combined in any manner in one or more embodiments.

[0026] The singular forms “a” and “the” used in this specification and the appended claims include plural references unless otherwise expressly stated herein. It should be noted that the term “or” is generally used to mean “and / or” unless otherwise expressly stated herein.

[0027] In the following description, in order to clearly demonstrate the structure and working method of the present invention, a number of directional terms will be used. However, terms such as "front", "back", "left", "right", "outside", "inside", "outward", "inward", "up", and "down" should be understood as convenient terms and not as limiting terms.

[0028] The implementation details of the embodiments of the present invention will be described in detail below with reference to the accompanying drawings. The following content is only for the convenience of understanding the implementation details and is not necessary for implementing this solution.

[0029] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: a fault diagnosis method and system for deep-sea mining systems based on digital twins and cloud-edge collaboration, the core of which lies in building and running a fusion diagnosis system driven by "mechanism and data dual engines and cloud and edge collaborative computing".

[0030] Figure 1 shows a schematic diagram of the fault diagnosis system architecture for deep-sea mining systems provided by this invention. This fault diagnosis system adopts a three-layer collaborative architecture of "physical perception - edge diagnosis - cloud intelligence":

[0031] The physical sensing layer consists of a network of various sensors deployed on key components of the deep-sea mining system (such as ore lifting pipes, relay stations, lifting pumps, and mining vehicles) to collect multi-dimensional state data such as vibration, stress, displacement, flow velocity, and pressure in real time.

[0032] Edge diagnostic layer: Composed of edge diagnostic units deployed in the near field of critical equipment (such as relay station controllers and pump compartment control units). Each edge diagnostic unit has a built-in adaptive noise suppression module and a lightweight artificial intelligence diagnostic model, which is responsible for real-time filtering, feature extraction, and preliminary diagnosis of component-level faults of the raw signals from local sensors, and generates deep feature vectors.

[0033] Cloud-based intelligent layer: Located on surface support vessels or shore-based data centers, its core is a cloud-based diagnostic server. The server runs three core modules: a digital twin engine, used to build and run system-level multiphysics coupling models to achieve mechanism diagnosis; a fusion decision module, used to perform rule-based information fusion and arbitration of the diagnostic results from the edge layer and the mechanism diagnosis results from the digital twin; and an uncertainty quantification module, used to calculate the confidence level of the final diagnostic results.

[0034] Based on the above-mentioned deep-sea mining system fault diagnosis system, and referring to the flowchart of the deep-sea mining system fault diagnosis method shown in Figure 2, the implementation of the fault diagnosis method provided by the present invention is carried out cyclically according to the following steps:

[0035] Step S1: Construct a multi-physics coupled digital twin model of the deep-sea mining system, and conduct mechanism monitoring based on the digital twin model.

[0036] Based on the lumped mass method, fluid mechanics, and structural dynamics theories, a multi-physics coupled digital twin model is established, encompassing the mine lifting pipe, relay station, booster pump, and pipeline flow channels. This digital twin model receives boundary condition inputs from the physical system in real time (such as ship position, ocean current velocity profile, and pump speed commands), calculates the system's dynamic response, and outputs simulated values ​​of state variables such as stress, deflection, and flow velocity at various system locations. Simultaneously, the measured values ​​from sensors at corresponding locations are compared with the simulated values ​​in real time to generate a residual sequence. An adaptive algorithm, such as an exponentially weighted moving average control chart, is used to monitor this residual sequence. When the residual sequence continuously exceeds the dynamic control limits, the digital twin engine determines that a "system-level coupled dynamics anomaly" has occurred and outputs the anomaly type and the associated physical component region identifier.

[0037] Step S2: Deploy an edge diagnostic unit in the near field of key equipment in the deep-sea mining system. The edge diagnostic unit performs adaptive filtering based on real-time noise spectrum on locally acquired signals to suppress interference, and inputs the processed signals into a lightweight deep learning model to achieve real-time preliminary classification and deep feature extraction of component-level faults.

[0038] An edge diagnostic unit is deployed near the critical equipment in the deep-sea mining system. This unit continuously collects signals such as local vibration acceleration. These signals first enter the adaptive noise suppression module built into the edge diagnostic unit. This module analyzes the spectrum of the input signal in real time, identifies the dominant steady-state environmental noise component (such as the fundamental frequency noise of the pump group at a specific frequency), and dynamically generates the parameters of a notch filter or band-stop filter bank to filter the original signal, effectively suppressing background noise.

[0039] After adaptive noise suppression of the signal, the deep-sea mining system transmits the denoised signal to a lightweight deep learning model pre-deployed in the edge diagnostic unit. This deep learning model, based on an end-to-end architecture, employs a feedforward propagation mechanism to perform layer-by-layer feature analysis and computation on the input signal. The signal first passes through convolutional or recurrent layers to extract local time-frequency features, then pooling layers reduce data dimensionality and enhance feature robustness, and finally, a fully connected layer maps the feature space to fault categories. For a specific component (e.g., the bearing at the drive end of the No. 1 lifting pump), the deep learning model outputs a multi-dimensional fault classification probability vector. Each dimension corresponds to a preset fault type (e.g., inner ring damage, outer ring crack, cage loosening, etc.), and the numerical value in the vector represents the confidence level that the signal belongs to that fault type. This result constitutes a preliminary diagnostic conclusion at the component level, providing a rapid reference for subsequent decision-making. Meanwhile, to support more complex fault tracing and health status assessment, the system extracts a high-dimensional deep feature vector from the penultimate layer of the deep learning model (usually the feature fusion layer before the fully connected layer). This layer retains the detailed information of the lower-level features and integrates high-level semantic features, which can encode the abstract representation of the fault (such as the time-frequency distribution pattern of the impact component, periodic modulation features, etc.). These features can not only be used for the refined classification of the current fault, but also be passed as input to the subsequent digital twin system to realize the reuse of fault knowledge and the iterative optimization of the diagnostic model across devices and operating conditions.

[0040] Step S3: Upload the preliminary diagnostic results and deep feature vector generated by the edge diagnostic unit to the cloud diagnostic center; the cloud diagnostic center runs the digital twin engine synchronously and receives its system-level anomaly judgment and location information.

[0041] The edge diagnostic unit encapsulates the generated preliminary diagnostic results, deep feature vectors, device unique identifier IDs, and timestamps (UTC standard time) into standardized data packets according to a predefined communication protocol (such as MQTT or a custom binary protocol), and uploads them to the data receiving buffer of the cloud diagnostic center via an underwater wired or wireless communication link. Simultaneously, the cloud diagnostic center runs the digital twin engine, receiving its system-level anomaly judgment and location information.

[0042] Step S4: Perform dual-engine information fusion and decision-making in the cloud-based diagnostic center.

[0043] This step employs a rule-based and confidence-weighted fusion algorithm to comprehensively evaluate the output of the digital twin engine and the data engine outputs from multiple edge diagnostic units. When the diagnoses from the two engines are consistent, a high-confidence conclusion is generated; when the conclusions conflict, an inference arbitration mechanism based on a historical failure case library is triggered.

[0044] Specifically, the cloud-based diagnostic center processes two streams of information in parallel: one stream is the system-level anomaly alarms and location information output by the digital twin engine in step S1 (input to the mechanism engine); the other stream is the component-level preliminary diagnostic package from multiple edge diagnostic units in step S3 (input to the data engine). The dual-engine information fusion and decision-making are comprehensively judged according to the following rules:

[0045] Rule A (Consistent Evidence, High Confidence): If the mechanism engine reports an anomaly in a certain area (such as "connection between relay station and mining pipe"), and the edge diagnostic unit associated with that area simultaneously reports a corresponding component failure (such as "loose connecting flange bolts"), then the data engine conclusion is directly adopted and marked as high confidence.

[0046] Rule B (Conflict of Evidence, Arbitration Required): If the mechanism engine indicates that the system is normal, but multiple edge diagnostic units report serious faults; or if the root causes of the faults indicated by both are contradictory, then the case reasoning arbitration sub-process is triggered. This sub-process uses all current symptoms and operating conditions as query conditions, performs similarity matching in the historical fault case knowledge base, retrieves the most similar historical case, and uses its final handling solution as the priority reference for the current diagnosis;

[0047] Rule C (Single Evidence, Requires Verification): If only the data engine reports an isolated fault, while the mechanism engine does not respond, a medium-confidence diagnostic prompt with the "Review Recommended" label is generated.

[0048] Step S5: Uncertainty Quantification and Visualization Output. Using an uncertainty quantification module integrated into the cloud-based fusion decision module (e.g., employing the Monte Carlo Dropout method), the confidence index of the final diagnostic result is calculated and output. The system outputs the clear fault diagnosis conclusion along with its corresponding confidence level, providing a transparent and reliable basis for operational decisions.

[0049] As shown in Figure 3, any diagnostic conclusion formed through step S4 must be processed by the uncertainty quantification module. This uncertainty quantification module enables Monte Carlo Dropout (MCDropout) during the inference phase of the cloud-based deep learning model. Specifically, during deep learning model prediction, the Dropout layer remains active, and T (e.g., 50) random forward propagations are performed on the same set of input features to obtain T slightly different predicted probability distributions. Subsequently, the mean (as the final diagnostic category) and variance σ² (as a measure of uncertainty) of these T predictions are calculated. The system maps the variance to a 0-100% confidence score (CS); where the confidence score can be converted into a 0-100% quantification index (σ²_max is the maximum variance value in the data) using the normalization formula CS=1-(σ² / σ²_max), but is not limited to calculation using the normalization formula; this is only an example / selection of a preferred solution.

[0050] Finally, the visual output interface will receive a complete diagnostic report containing fault codes, fault descriptions, location information, occurrence time, and quantified confidence levels. As shown in Figure 3, the interface highlights the confidence level graphically (such as with progress bars and color coding) and provides links to relevant historical cases or supplementary testing suggestions for diagnoses with low to medium confidence levels, thus achieving a complete closed loop from data to credible decision-making.

[0051] This invention proposes a fault diagnosis method and system for deep-sea mining systems based on digital twins and cloud-edge collaboration. By constructing a three-layer architecture of "physical perception - edge diagnosis - cloud intelligence," it integrates a multi-physics coupled digital twin model with a lightweight deep learning model to achieve advanced early warning of system-level coupled faults and robust diagnosis of weak faults in noisy environments. Through adaptive noise suppression and real-time preliminary diagnosis at the edge layer, combined with information fusion and uncertainty quantification from a dual-engine cloud platform (mechanism engine + data engine), it solves the problems of lack of system-level fault early warning and difficulty in extracting weak features in traditional methods. Finally, it outputs a visualized diagnostic report with quantified confidence, significantly improving the safety, reliability, and intelligent operation and maintenance level of deep-sea mining operations.

[0052] Although the present invention has been described in detail with reference to the accompanying drawings and preferred embodiments, the invention is not limited thereto. Various equivalent modifications or substitutions can be made to the embodiments of the invention by those skilled in the art without departing from the spirit and essence of the invention. Such modifications or substitutions should all fall within the scope of the invention, or any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the invention should be covered within the protection scope of the invention. Therefore, the protection scope of the invention should be determined by the scope of the claims.

Claims

1. A fault diagnosis method for a deep-sea mining system using digital twin and cloud-edge collaboration, characterized in that, The method includes: constructing a multi-physics coupled digital twin model of a deep-sea mining system, and performing mechanism monitoring based on the digital twin model; deploying an edge diagnostic unit near the key equipment of the deep-sea mining system, which performs real-time filtering, feature extraction, and preliminary component-level fault diagnosis on raw signals collected by various sensors, and generates a depth feature vector; uploading the preliminary diagnosis results and depth feature vector generated by the edge diagnostic unit to a cloud diagnostic center; the cloud diagnostic center synchronously running the digital twin engine, receiving its system-level anomaly judgment and location information; performing dual-engine information fusion and decision-making in the cloud diagnostic center; and performing uncertainty quantification and visualization output based on the diagnosis results formed by the cloud diagnostic center; wherein, the edge diagnostic unit performs real-time filtering, feature extraction, and preliminary component-level fault diagnosis on raw signals collected by various sensors, and generates a depth feature vector, specifically including: after completing adaptive noise suppression processing of the signal, the deep-sea mining system transmits the noise-reduced signal to the pre-deployed edge diagnostic unit. In the lightweight deep learning model, the deep learning model is based on an end-to-end architecture design and uses a forward propagation mechanism to perform layer-by-layer feature analysis and calculation on the input signal. The signal first passes through convolutional or recurrent layers to extract local features in the time-frequency domain, then through pooling layers to reduce data dimensionality and enhance feature robustness, and finally completes the mapping from feature space to fault category in a fully connected layer. For a certain component, the deep learning model outputs a multi-dimensional fault classification probability vector, where each dimension corresponds to a preset fault type, and the value in the vector represents the confidence that the signal belongs to the fault type. This result constitutes a preliminary diagnostic conclusion at the component level, providing a rapid reference for subsequent decisions. At the same time, to support more complex fault tracing and health status assessment, the system extracts a high-dimensional deep feature vector from the feature fusion layer of the deep learning model. The feature fusion layer features retain the detailed information of the low-level features and fuse high-level semantic features, which can encode the representation of the fault. These features are used for the refined classification of the current fault and are also passed as input to the digital twin system.

2. The method according to claim 1, characterized in that, The construction of a multi-physics coupled digital twin model for the deep-sea mining system, and the mechanism monitoring based on this digital twin model, specifically includes: establishing a multi-physics coupled digital twin model covering the ore lifting pipe, relay station, booster pump, and pipeline flow channels based on the lumped mass method, fluid mechanics, and structural dynamics theory. The digital twin model receives boundary condition inputs from the physical system in real time, performs system dynamic response calculations, and outputs simulated values ​​of the state variables at each location of the system; simultaneously, it compares the measured values ​​of the sensors at the corresponding locations with the simulated values ​​in real time to generate a residual sequence; an adaptive algorithm using an exponentially weighted moving average control chart is used to monitor the residual sequence; when the residual sequence continuously exceeds the dynamic control limit, the digital twin engine determines that a "system-level coupled dynamics anomaly" has occurred and outputs the anomaly type and its associated physical component area identifier; wherein, the boundary conditions include the surface vessel position, ocean current velocity profile, and pump speed command.

3. The method according to claim 2, characterized in that, The deployment of an edge diagnostic unit near the key equipment in the deep-sea mining system involves real-time filtering of raw signals collected by various sensors. Specifically, this includes: deploying the edge diagnostic unit near the key equipment in the deep-sea mining system; the edge diagnostic unit continuously collecting signals; the signals first entering the adaptive noise suppression module built into the edge diagnostic unit; the adaptive noise suppression module analyzing the spectrum of the input signal in real time, identifying the dominant steady-state environmental noise component, and dynamically generating parameters for a notch filter or band-stop filter bank to filter the raw signal and suppress background noise.

4. The method according to claim 3, characterized in that, The process involves uploading the preliminary diagnostic results and depth feature vector generated by the edge diagnostic unit to the cloud diagnostic center. Simultaneously, the cloud diagnostic center runs the digital twin engine, receiving its system-level anomaly judgment and location information. Specifically, the edge diagnostic unit encapsulates the generated preliminary diagnostic results, depth feature vector, device unique identifier ID, and timestamp into standardized data packets according to a predefined communication protocol, and uploads them to the data receiving buffer of the cloud diagnostic center via an underwater wired or wireless communication link. Simultaneously, the cloud diagnostic center runs the digital twin engine, receiving its system-level anomaly judgment and location information.

5. The method according to claim 4, characterized in that, The process of performing dual-engine information fusion and decision-making in the cloud-based diagnostic center includes: using a rule-based and confidence-weighted fusion algorithm to comprehensively judge the output of the digital twin engine and the data engine output from multiple edge diagnostic units; generating a high-confidence conclusion when the two engines diagnose in agreement; and triggering an inference arbitration mechanism based on a historical fault case library when the conclusions conflict.

6. The method according to claim 5, characterized in that, The information fusion and decision-making of the two engines are comprehensively diagnosed according to the following rules: Rule A: If the digital twin engine reports an anomaly in a certain area, and the edge diagnostic unit associated with that area simultaneously reports a corresponding component failure, then the conclusion of the data engine is directly adopted and marked as high confidence; Rule B: If the digital twin engine shows that the system is normal, but multiple edge diagnostic units report serious failures; or the root causes of the failures indicated by the two are contradictory, then the case reasoning arbitration sub-process is triggered; this sub-process uses all current symptoms and operating conditions as query conditions, performs similarity matching in the historical failure case knowledge base, retrieves the most similar historical cases, and uses the final handling plan as the priority reference for the current diagnosis; Rule C: If only the data engine reports an isolated failure, and the digital twin engine does not respond, then a medium confidence diagnostic prompt with the label "recommend review" is generated.

7. The method according to claim 6, characterized in that, The uncertainty quantification and visualization output of the diagnostic results generated based on the cloud diagnostic center includes: calculating and outputting the confidence index of the final diagnostic result through the uncertainty quantification module integrated in the cloud fusion decision module; and outputting the fault diagnosis result and its corresponding confidence score together.

8. The method according to claim 7, characterized in that, The uncertainty quantification module integrated into the cloud-based fusion decision module calculates and outputs the confidence index of the final diagnostic result. Specifically, this includes: all diagnostic results generated by the cloud-based diagnostic center must be processed by the uncertainty quantification module; the uncertainty quantification module enables Monte Carlo Dropout during the inference phase of the cloud-based deep learning model; during deep learning model prediction, the Dropout layer remains active, and T random forward propagations are performed on the same set of input features to obtain T differentiated prediction probability distributions; subsequently, the mean and variance σ² of the T predictions are calculated, with the mean serving as the final diagnostic category and the variance as a measure of uncertainty; the variance is mapped to a 0-100% confidence score CS; where the confidence score is converted to a 0-100% quantification index using the normalization formula CS=1-(σ² / σ²_max), where σ²_max is the maximum variance value in the data.

9. A system applicable to the fault diagnosis method for a deep-sea mining system using digital twin and cloud-edge collaboration as described in any one of claims 1 to 8, characterized in that, The system comprises: a physical sensing layer, consisting of a network of multiple sensors deployed on key components of the deep-sea mining system for real-time acquisition of raw data; an edge diagnostic layer, composed of edge diagnostic units deployed near key equipment; each edge diagnostic unit is responsible for real-time filtering, feature extraction, and preliminary diagnosis of component-level faults in the raw signals acquired by multiple sensors, and generating depth feature vectors; and a cloud intelligence layer, located on a surface support vessel or shore-based data center, with a cloud diagnostic server at its core; the server runs three main modules: a digital twin engine module, used to build and run a system-level multi-physics coupling model to achieve mechanism diagnosis; a fusion decision module, used to perform rule-based information fusion and arbitration of the data diagnostic results from the edge layer and the mechanism diagnostic results from the digital twin; and an uncertainty quantification module, used to calculate the confidence level of the final diagnostic results.

Images