Sensor-perception collaborative design method for fault diagnosis of monorail crane driving unit under cloud edge resource constraint

Abstract

This invention discloses a sensor-perception collaborative design method for fault diagnosis of the drive unit of a monorail crane under cloud-edge resource constraints, belonging to the field of monorail crane technology. First, taking the drive unit of the monorail crane as the object, the optimal monitoring position of the sensor is determined through finite element structure and temperature field analysis combined with simulation data quantitative evaluation. Then, based on this position, a conformal sensor precisely matching the curvature of the surface is designed and manufactured using direct-writing 3D printing technology. Next, a resource-constrained neural architecture search is introduced to adaptively design a lightweight diagnostic network model, balancing diagnostic accuracy and computational complexity. Finally, a federated learning distributed model evolution mechanism is constructed, achieving cross-node collaborative evolution of the model through local training at edge nodes, encrypted weight upload, cloud-based aggregation evaluation, and local fine-tuning at nodes. This invention can achieve accurate real-time fault diagnosis, improve model generalization ability, adapt to cloud-edge resource-constrained working conditions in coal mines, and ensure the safe and efficient operation of monorail cranes.

Description

Technical Field

[0001] This invention relates to the field of monorail technology, and more specifically to a sensing-perception collaborative design method for fault diagnosis of monorail drive unit under cloud-edge resource constraints. Background Technology

[0002] The monorail in coal mines is a key piece of auxiliary transportation equipment, responsible for hoisting and transporting materials, equipment, and personnel. Monorails operate over long distances in narrow tunnels, with frequent starts and stops and load fluctuations, making critical components such as the drive unit (motor, electric drive system, reducer, drive wheel / friction wheel and its bearings and cantilever support) prone to wear, loosening, overheating, and fatigue cracking. Once a malfunction occurs, it can lead to reduced traction, drive slippage, power interruption, or shutdown, affecting transportation efficiency and posing safety risks. Therefore, real-time status monitoring and fault diagnosis of the monorail drive unit are of great significance for ensuring safe and efficient coal mine production.

[0003] Traditional methods for diagnosing faults in monorail crane drive units typically rely on centralized data processing and large-scale diagnostic models, which often face the following problems in fault diagnosis:

[0004] 1. High difficulty in sensor integration: The installation space of monorail crane equipment is small, the surface is mostly cast or machined curved surface and there are oil stains, impacts and strong vibrations. Traditional adhesive strain gauges or temperature probes and external wiring are easy to fall off or wear, making it difficult to achieve high reliability and long life embedded integration of sensors in downhole equipment.

[0005] 2. Limited edge computing resources: Edge computing nodes in underground coal mine equipment are usually based on low-cost, low-power embedded processors (such as the ARM Cortex-M series), which have problems such as weak computing power, small storage space, and strict power consumption limitations, making it difficult to directly deploy complex diagnostic models.

[0006] 3. Lack of lightweight model adaptive design mechanism: Most existing models rely on manual design and fail to take edge hardware constraints (such as latency, memory, and energy consumption) as the direct target of model optimization, resulting in a significant performance degradation of the model when deployed in practice.

[0007] 4. Insufficient distributed intelligent collaboration: In the same mine, there are usually multiple monorail cranes operating in different roadways, and the data distribution of each accompanying node is significantly different. Existing methods lack efficient cross-node model collaborative evolution and knowledge sharing mechanisms, making it difficult to achieve continuous optimization and generalization of the model.

[0008] In summary, there is an urgent need to develop a fault diagnosis technology solution for monorail crane drive units that is suitable for cloud-edge resource-constrained scenarios in underground coal mines and can solve the above four major problems. This solution should organically combine highly reliable sensor integration, lightweight real-time edge diagnosis, and multi-node distributed intelligent collaboration to ensure accurate and real-time detection of monorail crane drive unit faults and continuous model optimization. Summary of the Invention

[0009] The purpose of this invention is to propose a sensor-perception collaborative design method for fault diagnosis of monorail crane drive unit under cloud-edge resource constraints, so as to realize accurate and real-time monitoring and diagnosis of monorail crane drive unit faults, improve the model generalization ability, and ensure safe and efficient production in coal mines.

[0010] The technical solution adopted in this invention is: a sensor-perception collaborative design method for fault diagnosis of a monorail crane drive unit under cloud-edge resource constraints, characterized by comprising the following steps:

[0011] Step S1: Taking the drive unit of the monorail as the object, candidate monitoring areas are obtained through finite element structure and temperature field analysis. After quantitative evaluation based on simulation data, the optimal monitoring position is determined as the best deployment point for the sensor.

[0012] Step S2: Based on the optimal sensor deployment point selected in Step S1, perform customized design and manufacturing of the sensor;

[0013] Step S3: Construct a lightweight diagnostic model that is adaptable to resources and operating conditions. Introduce a resource-constrained neural architecture search to automatically design a lightweight network. First, integrate classic lightweight network modules to build a module pool. Design a multi-objective reward function that combines diagnostic accuracy and computational complexity. Combine reinforcement learning algorithms and hard connection rules to complete the network architecture search. Then, use the Pareto optimal strategy to select models that are adapted to edge hardware. Finally, through adaptive design, dynamically adjust the cost of the diagnostic model or update parameters according to resources and operating conditions to achieve a balance between high accuracy and low complexity of the diagnostic model under different constraints.

[0014] Step S4: Deploy the designed diagnostic model to each monorail crane's onboard edge node. The node uses sensor data to complete local model training. The local model weight updates are encrypted and uploaded to the cloud. The cloud uses a federated averaging algorithm to aggregate the weights and generate a new generation of global model. The cloud performs performance evaluation on the global model and distributes the globally model with significantly improved performance to each edge node. If the performance does not meet the requirements, training is restarted. After receiving the new global model, the edge nodes complete personalized fine-tuning using a small amount of local data to generate a fault diagnosis model adapted to the underground working conditions.

[0015] As a further improvement of the present invention, in step S1, taking the monorail crane drive unit as the object, candidate monitoring areas are obtained through finite element structure and temperature field analysis, and the optimal monitoring position is determined after quantitative evaluation based on simulation data, which serves as the optimal deployment point for the sensor. Specifically:

[0016] Step S11: Establish a three-dimensional model of the key structure of the monorail crane drive unit, set material properties and perform mesh generation in ANSYS; then apply boundary condition loads to the structure to obtain the distribution of stress, displacement, strain energy density and safety factor.

[0017] Step S12: Perform temperature modeling on the key structures of the monorail crane drive unit, calculate the heat generation power and estimate the temperature, using the following formula:

[0018] ;

[0019] ;

[0020] ;

[0021] In the formula, This refers to the heating power; and These are the total energy power input to the heat source and the effective energy power output to the outside, respectively. The surface temperature of the heat source; The current ambient temperature; Heat flux density; The thickness of the heat source surface layer; To characterize the thermal conductivity of the material; The area of ​​the heat source;

[0022] Then, the temperature distribution along the heat conduction path is quantified, and the temperature at different locations is calculated along the main heat conduction path. The calculation formula is as follows:

[0023] ;

[0024] In the formula, Location along the heat conduction path Temperature at that location; Temperature at the reference point; Heat flow rate; For position The cross-sectional area perpendicular to the direction of heat flow; The coordinates represent the arc length of the heat conduction path;

[0025] Step S13: Use the entropy weight method to comprehensively select from multiple optimal deployment points obtained from structural analysis and temperature field analysis to determine the optimal deployment point of the sensor.

[0026] As a further improvement of the present invention, in step S13, the comprehensive selection of multiple optimal deployment points obtained from structural analysis and temperature field analysis using the entropy weight method specifically involves:

[0027] Step S131: Establish the original indicator matrix X and standardize the indicators:

[0028] ;

[0029] in, For the first The deployment point is at the Standardized values ​​for each evaluation indicator; Number of measurement points; For the number of indicators;

[0030] Step S132: Standardize the indicators to obtain a matrix. :

[0031] ;

[0032]

[0033] In the formula, For the first Deployment point at Standardized values ​​for each indicator;

[0034] Step S133: Calculate the entropy value of the index. and weight :

[0035] ; ;

[0036] ;

[0037] In the formula, ; Indicates the first The first indicator The proportion of indicator values ​​for each deployment point;

[0038] Step S134: Calculate the overall score for each deployment point:

[0039] ;

[0040] Based on the overall score The deployment points are sorted, and the higher the score, the more suitable they are for deployment.

[0041] As a further improvement of the present invention, in step S2, the customized design and manufacturing of the sensor based on the optimal sensor deployment point selected in step S1 specifically includes:

[0042] Step S21: Use a high-precision 3D scanner to scan the target surface of the downhole equipment from multiple angles to obtain complete surface point cloud data. Then, import the obtained point cloud data into MeshLab and use the iterative nearest point algorithm for registration and denoising to generate a surface model. ;

[0043] Step S22: Based on the surface model data obtained from the scan in S21, the substrate of the conformal surface sensor is designed as a flexible-rigid composite structure: the surface contact layer is made of flexible material and designed as a honeycomb to achieve zero-gap fit with the structural surface and provide mechanical buffer; the internal support layer is made of rigid material to provide structural support and maintain the stability of the sensing element.

[0044] Step S23: Plan the printing trajectory of the conformal sensor on the curved surface. The specific steps are as follows:

[0045] Step S231: Obtain the surface model through calculation. The centroid of the space The calculation formula is as follows: ;

[0046] In the formula, , , respectively surface model about , , The static moment of a plane; For curved surface models The area; For curved surface models Area elements;

[0047] Step S232: The curved surface model edge line towards the centroid Scale proportionally to obtain The set of curves is denoted as . Wire;

[0048] Step S233: Place each The line is evenly divided into A line segment is the set of its endpoints, and the set of its endpoints is denoted as a matrix. The formula is as follows:

[0049] ;

[0050] Step S234: Convert the matrix Transpose to obtain the transpose matrix The formula is as follows:

[0051] ;

[0052] Connect the points corresponding to the elements in the column vectors of the transpose matrix to generate... Wire, Line and Lines intertwine on the curved surface to form a grid;

[0053] Step S235: During printing, select the corresponding grid points according to the trajectory design scheme, connect them to obtain polylines, and smooth the polylines to obtain curves with continuous curvature and conformal to the surface contour. The final curve is obtained. This refers to the movement trajectory of the nozzle during the sensor printing process;

[0054] Step S24: Multi-material direct writing layer printing and curing.

[0055] Step S241, Surface pretreatment: Thoroughly clean the area to be monitored with anhydrous ethanol;

[0056] Step S242, Insulating substrate printing: In the pre-treated monitoring area, an insulating substrate is printed using direct-write molding 3D printing;

[0057] Step S243, Printing and curing of lower layer conductors: On the cured insulating substrate, using silver paste as the conductive material, three parallel lower layer conductors are printed through a 0.21 mm needle, an extrusion pressure of 0.25 MPa, and a printing speed of 5 mm / s.

[0058] Step S244, Insulating Strip Printing and Curing: Print three parallel epoxy resin insulating strips vertically above the lower conductor.

[0059] Step S245, Printing and curing of upper conductors: Print upper conductors with T-shaped connectors above the insulating strip, using the same printing parameters as the lower conductors;

[0060] Step S246, Sensor Unit Printing and Curing: Using carbon paste as the sensing material, the sensor unit array is printed one by one through a 0.21 mm needle at an extrusion pressure of 0.25 MPa and a printing speed of 5 mm / s. Each unit is approximately 0.35 mm long.

[0061] Step 247, Encapsulation Layer Printing and Curing: The encapsulation layer is printed with epoxy resin. The printing parameters and curing process are the same as in step S242. This is to protect the internal wires and sensing units and improve the mechanical strength and environmental adaptability of the overall structure.

[0062] As a further improvement of the present invention, in step S24, the multi-material direct writing layer printing and curing specifically includes:

[0063] Step S241, Surface pretreatment: Thoroughly clean the area to be monitored with anhydrous ethanol;

[0064] Step S242, Insulating substrate printing: In the pre-treated monitoring area, an insulating substrate is printed using direct-write molding 3D printing;

[0065] Step S243, Printing and curing of lower layer conductors: On the cured insulating substrate, using silver paste as the conductive material, three parallel lower layer conductors are printed through a 0.21 mm needle, an extrusion pressure of 0.25 MPa, and a printing speed of 5 mm / s.

[0066] Step S244, Insulating Strip Printing and Curing: Print three parallel epoxy resin insulating strips vertically above the lower conductor.

[0067] Step S245, Printing and curing of upper conductors: Print upper conductors with T-shaped connectors above the insulating strip, using the same printing parameters as the lower conductors;

[0068] Step S246, Sensor Unit Printing and Curing: Using carbon paste as the sensing material, the sensor unit array is printed one by one through a 0.21 mm needle at an extrusion pressure of 0.25 MPa and a printing speed of 5 mm / s. Each unit is approximately 0.35 mm long.

[0069] Step 247, Encapsulation Layer Printing and Curing: The encapsulation layer is printed with epoxy resin. The printing parameters and curing process are the same as in step S242. This is to protect the internal wires and sensing units and improve the mechanical strength and environmental adaptability of the overall structure.

[0070] As a further improvement of the present invention, in step S3, the construction of a lightweight diagnostic model that is adaptable to resources and operating conditions involves introducing a resource-constrained neural architecture search to automatically design a lightweight network. First, classic lightweight network modules are integrated to build a module pool. A multi-objective reward function that combines diagnostic accuracy and computational complexity is designed. Network architecture search is completed by combining reinforcement learning algorithms and hard-connection rules. Then, a Pareto optimal strategy is used to select models adapted to edge hardware. Finally, adaptive design is used to dynamically adjust the diagnostic model overhead or update parameters according to resources and operating conditions, achieving a balance between high accuracy and low complexity of the diagnostic model under different constraints. Specifically:

[0071] Step S31: Integrate the lightweight modules contained in the classic lightweight network, and construct a module pool in the action space that includes one-dimensional convolution, depthwise separable convolution, Ghost module, group convolution, attention mechanism, max pooling, pyramid pooling, fully connected layer, Softmax terminal state layer, and optional parameters for each module.

[0072] Step S32: Design a multi-objective reward function to form a comprehensive evaluation index to guide the direction of model optimization;

[0073] Step S33, Agent Search Strategy: The DQN reinforcement learning algorithm is adopted, and hard connection rules are integrated into the Ɛ-greedy strategy to form the search strategy of the network architecture. The Ɛ-greedy strategy is represented as follows:

[0074] ;

[0075] In the formula, λ represents the learning rate, which is set to 0.01; Indicates the execution of an action The instant reward received afterwards; The discount factor, representing the importance of future rewards, is set to 0.99; In the new state Below, the maximum effect of all possible actions value;

[0076] Step S34, Agent Search Process: The agent calculates the corresponding reward value based on the model structure existing in the current state space. The agent predicts and selects the next action based on the search strategy to obtain a higher cumulative reward, and so on, until the agent selects the termination layer, thus completing the design of the complete model; the actual performance of the designed model is then evaluated... The value is updated, and this guides the agent to construct higher-value modules in subsequent iterations. The model of value;

[0077] Step S35: Select a suitable diagnostic model for deployment at the edge layer using a Pareto optimal strategy. Choose a suitable set of diagnostic models from the constructed model space, based on the following criteria:

[0078] ;

[0079] In the formula, This indicates that the model Strictly superior to model ; and Representing the model respectively The diagnostic accuracy and floating-point calculations are then performed; from this set of Pareto optimal models, models that can be deployed on specific hardware resources are selected.

[0080] As a further improvement of the present invention, in step S32, the multi-objective reward function formula is as follows:

[0081] ;

[0082] In the formula, , These represent the total number of samples in the test set and the number of samples diagnosed with faults, respectively. Floating-point numbers for calculating the diagnostic model; This represents the maximum computing resource capacity for the edge layer. and These are the diagnostic accuracy indicators. and computational complexity metrics Weighting.

[0083] As a further improvement of the present invention, in step S4, the designed diagnostic model is first deployed to each monorail crane's onboard edge node, and the node completes local model training using sensor data; then, the local model weight update is encrypted and uploaded to the cloud, and the cloud uses a federated averaging algorithm to aggregate the weights to generate a new generation of global model; subsequently, the cloud performs performance evaluation on the global model, and distributes the significantly improved global model to each edge node; if the performance does not meet the requirements, training is restarted; after receiving the new global model, the edge node completes personalized fine-tuning using a small amount of local data to generate a fault diagnosis model adapted to the underground working conditions, specifically:

[0084] Step S41: Deploy the diagnostic model designed in step S3 to each monorail crane's onboard edge node. Each node uses the conformal sensor manufactured in step S2 to collect the monorail crane's drive unit's operating data. Then, the data processing module preprocesses and extracts features from the collected data to form a local database. Use local data Train the diagnostic model; once training is complete, obtain the updated diagnostic model. and the corresponding model weights ;

[0085] Step S42: Update and encrypt the locally trained model weights for each edge node and upload them to the cloud. The updated model weights are represented as follows:

[0086] ;

[0087] In the formula, These are the weights for the initial global model; For the first The model weights after local training at each edge node; For the first The weight change amount uploaded by each edge node;

[0088] Then, the cloud server initiates an aggregation algorithm to generate a new generation of global models. Aggregation is achieved using a federated average algorithm, as shown below:

[0089]

[0090] In the formula, These are the global model weights; Indicates the first One node; For the first The number of local data samples used by each edge node in this round of training; The sum of data samples used by all edge nodes participating in this round of aggregation;

[0091] Step S43: Analyze the global model generated by aggregation in the cloud. The model is trained and its performance is evaluated using metrics including accuracy, recall, and F1 score. Based on the evaluation results, a decision is made regarding whether to distribute the global model: if the new global model shows a significant performance improvement over the previous generation, the evolution is considered effective, and the new global model is repackaged and distributed to edge nodes for diagnostic model updates; if the performance improvement is not significant or deterioration occurs, the update is discarded, and the training process is restarted to further optimize the global model performance.

[0092] Step S44: Edge nodes receive and load the new global model. Then, a small amount of local data is used to fine-tune the model, generating a personalized model that is more suitable for fault diagnosis of downhole equipment.

[0093] Compared with the prior art, the present invention has the following technical advantages:

[0094] 1. This invention determines the optimal deployment point of the sensor by combining multiphysics simulation with the entropy weight method, taking into account both the acquisition of fault-sensitive information and the safety of the equipment structure, thereby improving the detectability of early faults. At the same time, it designs and manufactures a curved conformal sensor based on direct writing 3D printing technology, and adopts a flexible-rigid composite substrate structure and multi-material layered printing process to make the sensor fit with the complex curved surface of the monorail crane drive unit. This solves the problems of easy detachment and easy wear of wiring of traditional adhesive sensors, significantly reducing the frequency of sensor maintenance. It is suitable for harsh working conditions such as oil pollution, strong vibration and confined space in wells, and realizes the embedded integration of the sensing unit and the equipment structure, ensuring the continuity and accuracy of fault sensing data (strain, temperature, vibration) acquisition.

[0095] 2. This invention, constrained by edge node computing power, storage, and energy consumption budgets, introduces a resource-constrained neural architecture search. It constructs a lightweight module pool, designs a multi-objective reward function that integrates diagnostic accuracy and computational complexity, and combines reinforcement learning algorithms with hard-connection rules to achieve adaptive network architecture design. A Pareto optimal strategy is then used to select suitable models, achieving precise optimization of diagnostic model parameters and inference latency. This design allows the model to be directly deployed on low-cost embedded processors such as ARM Cortex-M and STM32F429, maintaining high diagnostic accuracy and low computational complexity under cloud-edge resource constraints. It enables real-time data processing, fault category identification, and health score output at the edge, meeting the engineering requirements for real-time fault diagnosis and early warning for monorail cranes and solving the problem that traditional large-scale diagnostic models cannot be deployed at the edge.

[0096] 3. This invention establishes a cloud-edge collaborative distributed model evolution mechanism based on federated learning. Each monorail crane's accompanying edge node completes model training using local sensor data. Only model weight updates, not the original data, are uploaded using AES-256 encryption. The cloud uses a federated averaging algorithm to aggregate and generate a global model. After evaluation based on multiple metrics including accuracy, recall, and F1 score, the optimized model is distributed to each node, where edge nodes then perform personalized fine-tuning using a small amount of local data. This mechanism, while ensuring the security of mine production data and reducing cloud-edge communication bandwidth consumption, achieves cross-node and cross-working-condition model knowledge sharing and periodic iterative evolution. It solves the problem of insufficient model generalization ability caused by significant differences in data distribution among multiple monorail cranes, allowing the diagnostic model to adapt to fault diagnosis of monorail crane drive units in different roadways and working conditions throughout the mine, thus continuously improving overall diagnostic performance.

[0097] 4. This invention organically combines highly reliable data acquisition in the embedded sensing layer, lightweight real-time diagnostics in the edge computing layer, and distributed model evolution in the federated collaboration layer, forming a full-link fault diagnosis system from physical layer hardware integration to algorithm layer model design and cloud-edge layer intelligent collaboration. It precisely solves four core problems in existing technologies: difficulty in sensor integration, poor edge resource adaptation, lack of adaptive mechanisms in model design, and insufficient distributed collaboration. This solution enables accurate and real-time detection of faults such as wear, loosening, overheating, and fatigue cracks in the drive unit of a monorail crane, effectively avoiding problems such as reduced traction, slippage, and power interruption caused by drive unit failures. It reduces equipment downtime risks, improves monorail crane transportation efficiency, and provides technical support for the safe, efficient, and intelligent operation and maintenance of auxiliary transportation equipment in coal mines, ensuring overall coal mine production safety.

[0098] 5. The direct-writing 3D printing technology used in this invention enables customized manufacturing of sensors, adapting to the curved surface features of different parts of a monorail crane without requiring additional dedicated mounting brackets, thus reducing sensor installation and modification costs. The lightweight diagnostic model is based on existing edge-embedded hardware design, eliminating the need to upgrade the computing hardware of downhole equipment and significantly reducing hardware investment costs. The federated learning mechanism reduces raw data transmission between the cloud and the edge, lowering bandwidth pressure and maintenance costs for downhole communication networks. Simultaneously, the long lifespan and low maintenance characteristics of the sensors, along with the autonomous evolution and generalization characteristics of the model, further reduce subsequent maintenance and debugging costs, improving fault diagnosis effectiveness while considering the economic efficiency and feasibility of engineering applications. Attached Figure Description

[0099] Figure 1 This is a flowchart of the sensing-perception collaborative design method for fault diagnosis of monorail gantry drive unit under cloud-edge resource constraints, as presented in this invention.

[0100] Figure 2 This is a flowchart illustrating the selection of the optimal sensor deployment point in the design method of this invention.

[0101] Figure 3 This is a flowchart illustrating the fabrication process of the conformal sensor for curved surfaces in the design method of this invention.

[0102] Figure 4 This is a flowchart illustrating the design of a lightweight diagnostic model in the design method of this invention.

[0103] Figure 5 This is a flowchart of the distributed model evolution based on federated learning in the design method of this invention. Detailed Implementation

[0104] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention are within the scope of protection of the present invention.

[0105] like Figure 1 As shown, the sensing-perception collaborative design method for fault diagnosis of monorail crane drive unit under resource constraints of the present invention specifically includes the following steps:

[0106] Step S1: Sensor deployment based on multiphysics simulation analysis.

[0107] When deploying sensors, it is necessary to determine the optimal monitoring locations for key components of the monorail crane drive unit. This invention focuses on the monorail crane drive unit (motor-reducer-drive wheel), using finite element structural analysis and temperature field analysis to obtain candidate monitoring areas. These areas are then quantitatively evaluated based on comprehensive simulation data to ultimately determine the optimal sensor deployment points. The sensors in this embodiment include strain sensors, temperature sensors, and vibration sensors, used to monitor the compressive stress, temperature, vibration information, and other data of key components in the drive unit.

[0108] Specifically, such as Figure 2 As shown, step S1 specifically includes:

[0109] Step S11: Finite element analysis of key components of downhole equipment.

[0110] First, a 3D model of the key structures of the monorail crane drive unit (such as the reducer housing, bearing housing, and drive wheel cantilever / support) was created using SolidWorks. Material properties were then set and meshed in ANSYS software (finite element analysis software). Boundary condition loads were then applied to the structure to obtain the distribution of stress, displacement, strain energy density, and safety factor. Under the premise of meeting safety factor requirements and not affecting structural strength, outer surface areas with high strain energy density, greater sensitivity to fault-induced load changes, and sufficient mounting area were preferentially selected as candidate deployment points, and strain simulation data for the corresponding locations were output.

[0111] Step S12: Temperature field analysis of downhole equipment.

[0112] Temperature modeling of a monorail structure requires calculating the equipment's heat output and estimating its temperature, using the following formulas:

[0113] ;

[0114] ;

[0115] ;

[0116] In the formula, This refers to the heating power; , These represent the total energy power input to the heat source and the effective energy power output to the external environment. Taking a drive motor as an example, these are respectively the total energy power input to the heat source and the effective energy power output to the external environment. , and Voltage and current, respectively. For mechanical efficiency; The surface temperature of the heat source; The current ambient temperature; Heat flux density; The thickness of the heat source surface layer; To characterize the thermal conductivity of the material; The area is the heat source area.

[0117] Furthermore, the temperature distribution along the heat conduction path is quantified, and the temperature at different locations is calculated along the main heat conduction path (such as component connection chains). The calculation formula is as follows:

[0118] ;

[0119] In the formula, Location along the heat conduction path Temperature at that location; Temperature at the reference point; Heat flow rate; For position The cross-sectional area perpendicular to the direction of heat flow; The coordinates represent the arc length of the heat conduction path.

[0120] Step S13: Optimal selection method for sensor deployment points.

[0121] The optimal sensor deployment point is determined by comprehensively selecting from multiple optimal deployment points obtained from structural and temperature field analyses using the entropy weight method. First, an original index matrix X is established, and the indexes are standardized.

[0122] ;

[0123] in, For the first The deployment point is at the Standardized values ​​for each evaluation indicator; Number of measurement points; This refers to the number of indicators.

[0124] Furthermore, the indicators are standardized to obtain a matrix. :

[0125] ;

[0126] ;

[0127] In the formula, For the first Deployment point at The standardized values ​​for each indicator; in structural analysis and temperature field analysis, positive indicators include: strain energy density, temperature gradient, and heat flux density, while negative indicators include: safety factor and heat dissipation efficiency.

[0128] Then, calculate the entropy value of the indicator. and weight :

[0129] ; ;

[0130] ;

[0131] In the formula, ; Indicates the first The first indicator The proportion of indicator values ​​for each deployment point.

[0132] Finally, calculate the overall score for each deployment point:

[0133] ;

[0134] Based on the overall score The deployment points are sorted, and the higher the score, the more suitable they are for deployment.

[0135] Step S2: Design and manufacturing of a single-track suspended curved surface conformal sensor system based on direct writing 3D printing technology.

[0136] Based on the optimal sensor deployment point selected by S1, the sensor is customized in design and manufacturing to ensure a perfect fit with the surface of the monorail drive unit. Addressing issues such as large curvature variations, oil contamination, strong vibrations, and difficulties in wiring and maintenance in areas like the monorail reducer housing, bearing housing, and drive wheel cantilever (bracket), the sensor structure is precisely matched to the actual curvature on the equipment surface by acquiring surface geometry data and using adaptive printing technology. This enables highly reliable, long-life, and embedded integrated sensors on the monorail.

[0137] Specifically, step S2 includes:

[0138] Step S21: 3D data acquisition and digital modeling of the curved surface.

[0139] like Figure 3 As shown, a high-precision 3D scanner, such as laser scanning or structured light scanning, is first used to scan the target surface of the downhole equipment from multiple angles to obtain complete surface point cloud data. Then, the acquired point cloud data is imported into MeshLab, and the iterative nearest-point algorithm is used for registration and denoising to generate a surface model. And save it as an STL file.

[0140] Step S22: Sensor substrate design based on mechanical conformality.

[0141] For applications where sensors are mounted on the outer surface of the key curved surface of a monorail crane drive unit, considering the compatibility between the sensor's bottom surface and the equipment's curved surface, as well as the impact of vibrations generated during operation on sensor performance, a sensor structure that meets practical requirements needs to be designed. Firstly, based on the curved surface model data obtained in step S21, the sensor substrate is designed as a flexible-rigid composite structure: the curved surface contact layer uses a flexible material and is designed in a honeycomb pattern to achieve zero-gap contact with the structural surface and provide mechanical buffering; the internal support layer uses a rigid material to provide structural support and maintain the stability of the sensing element.

[0142] Step S23: Planning the printing trajectory of the conformal sensor on the curved surface.

[0143] When fabricating sensors using direct-writing 3D printing technology, it is necessary to determine the printing trajectory. The steps are as follows:

[0144] First, the surface is obtained through calculation. The centroid of the space The calculation formula is as follows: ;

[0145] In the formula, , , respectively curved surfaces about , , The static moment of a plane; curved surface The area; curved surface Area elements.

[0146] Then, the curved surface edge line towards the centroid Scale proportionally to obtain The set of curves is denoted as . Wire.

[0147] Furthermore, each The line is evenly divided into A line segment is the set of its endpoints, and the set of its endpoints is denoted as a matrix. The formula is as follows:

[0148] ;

[0149] Then, the matrix Transpose to obtain the transpose matrix The formula is as follows:

[0150] ;

[0151] Connect the points corresponding to the elements in the column vectors of the transpose matrix to generate... Wire, Line and Lines interweave on the curved surface to form a grid. During printing, grid points (denoted as ) can be selected at corresponding positions according to the trajectory design scheme. ), and connect them to obtain a polyline (denoted as ). The polyline is then smoothed to obtain a curve with continuous curvature that conforms to the surface profile. Then the curve Translate the curve by 0.2 mm along the normal direction away from the surface to obtain the final curve. This refers to the movement trajectory of the nozzle during the sensor printing process.

[0152] Step S24: Multi-material direct writing layer printing and curing.

[0153] This step involves the physical fabrication of the sensor. Following the sensor structure, the substrate, insulating layer, sensing layer, and encapsulation layer are printed sequentially to achieve layered construction. The manufacturing steps are as follows:

[0154] 1. Surface pretreatment: Thoroughly clean the area to be monitored with anhydrous ethanol;

[0155] 2. Insulating Substrate Printing: An insulating substrate is printed using direct-write 3D printing on the pre-treated monitoring area. Printing parameters are set as follows: needle diameter 0.21 mm, printing spacing 0.1 mm, extrusion air pressure 0.125 MPa, and printing speed 10 mm / s. After printing, the substrate is allowed to cure naturally, forming a smooth and stable base layer.

[0156] 3. Printing and Curing of Lower Layer Conductors: On the cured insulating substrate, using silver paste as the conductive material, three parallel lower layer conductors are printed through a 0.21mm needle, at an extrusion pressure of 0.25 MPa, and a printing speed of 5 mm / s. The conductor spacing is adjustable according to the width of the monitoring area (generally 5–15 mm). After printing, the conductors are cured in a 75 °C oven for 30 minutes to achieve good conductivity and adhesion.

[0157] 4. Printing and Curing of Insulating Strips: Print three parallel epoxy resin insulating strips vertically above the lower conductor. Printing parameters are: air pressure 0.125 MPa, needle diameter 0.26 mm, printing speed 10 mm / s, and strip width 1 mm. After printing, dry at 71 °C for 2 hours to ensure complete curing of the insulation layer.

[0158] 5. Printing and Curing of Upper Conductors: Print upper conductors with T-joints above the insulating strip, using the same printing parameters as the lower conductors. After printing, cure in a 75 °C oven for 30 minutes to form a reliable upper and lower conductive connection.

[0159] 6. Sensing Unit Printing and Curing: Using carbon paste as the sensing material, an array of sensing units was printed one by one through a 0.21 mm needle at an extrusion pressure of 0.25 MPa and a printing speed of 5 mm / s. Each unit was approximately 0.35 mm in length. After printing, the units were cured at 80 °C for 2 hours to form a sensing layer with a stable electrical response.

[0160] 7. Encapsulation layer printing and curing: Finally, an encapsulation layer is printed with epoxy resin. The printing parameters and curing process are the same as in step 2. This is to protect the internal wires and sensing units and improve the mechanical strength and environmental adaptability of the overall structure.

[0161] Step S3: Adaptive design of a lightweight diagnostic network architecture.

[0162] like Figure 4 As shown, taking the fault diagnosis of a monorail crane drive unit as an example, edge nodes collect conformal strain / temperature signals and accelerometer vibration signals to construct multi-channel time-series samples (such as triaxial vibration + temperature + strain with a window length of T). Noise reduction, normalization, and feature extraction are performed at the edge, outputting fault categories and health scores for on-board early warning and maintenance decisions. However, due to limited cloud-edge resources, the computing power, storage, and energy consumption budget at the edge are tight, making it difficult to support real-time inference of large models. Therefore, a lightweight diagnostic model that can adapt to resources and operating conditions is required. To this end, a resource-constrained Neural Architecture Search (NAS) is introduced to automatically design a lightweight network, incorporating diagnostic accuracy, parameter quantity, FLOPs, inference latency, memory, and other overhead into the optimization. Through adaptive design, the diagnostic model overhead or updated parameters are dynamically adjusted according to resources and operating conditions to maintain high accuracy and low complexity under different constraints. Specifically, step S3 includes:

[0163] Step S31: Build a lightweight module pool.

[0164] It integrates lightweight modules contained in classic lightweight networks MobileNet, ShuffleNet, GhostNet, and GoogleNet, and constructs a module pool in the action space that includes one-dimensional convolution, depthwise separable convolution, Ghost module, grouped convolution, attention mechanism, max pooling, pyramid pooling, fully connected layer, Softmax terminal state layer, and optional parameters for each module.

[0165] Step S32: Design a multi-objective reward function.

[0166] A multi-objective reward function integrates the computational complexity and diagnostic accuracy of the diagnostic model to form a comprehensive evaluation index to guide the model's optimization direction. Reward Function The design is as follows:

[0167] ;

[0168] In the formula, , These represent the total number of samples in the test set and the number of samples diagnosed with faults, respectively. Floating-point numbers for calculating the diagnostic model; This represents the maximum computing resource capacity for the edge layer. and These are the diagnostic accuracy indicators. and computational complexity metrics Weighting.

[0169] Step S33: Agent search strategy.

[0170] The DQN (Deep Q-Network) reinforcement learning algorithm is adopted, and hard connection rules are integrated into the Ɛ-greedy policy to form the search strategy of the network architecture. The Ɛ-greedy policy is represented as follows:

[0171] ;

[0172] In the formula, λ represents the learning rate, which is set to 0.01; Indicates the execution of an action The instant reward received afterwards; The discount factor, representing the importance of future rewards, is set to 0.99; In the new state Below, the maximum effect of all possible actions value.

[0173] The hard constraint rules are as follows:

[0174] 1) The pyramid pooling layer is only connected to the optional parameter states of the fully connected layer;

[0175] 2) When two fully connected layers appear consecutively, the number of neurons in the upper layer must not be less than the number of neurons in the lower layer;

[0176] 3) Attention mechanism layers must not appear consecutively;

[0177] 4) During the connection process of each layer, the termination state layer can be directly connected at any position to end the layer connection process.

[0178] Step S34, Model Search Process.

[0179] First, the agent calculates the corresponding reward value based on the model structure existing in the current state space. Then, based on the search strategy, the agent predicts and selects the next action to obtain a higher cumulative reward, and so on, until the agent selects the termination layer, thus completing the design of the complete model. The actual performance of the designed model is then evaluated... The value is updated, and this guides the agent to construct higher-value modules in subsequent iterations. Value model.

[0180] Step S35: Model selection based on Pareto optimality.

[0181] A Pareto optimal strategy is adopted to select a suitable diagnostic model for deployment at the edge layer. First, a suitable set of diagnostic models is selected from the constructed model space, based on the following criteria:

[0182] ;

[0183] In the formula, This indicates that the model Strictly superior to model ; and Representing the model respectively The diagnostic accuracy and the calculation of floating-point numbers.

[0184] Then, from this set of Pareto optimal models, select the models that can be deployed on specific hardware resources.

[0185] Step S4: Distributed model evolution mechanism based on federated learning.

[0186] Step S41: Local model training. For example... Figure 5 As shown, the diagnostic model designed in S3 is first deployed to each edge node of the monorail crane. Each node uses the conformal sensor manufactured in S2 to collect the operating data (strain, temperature, and vibration information) of the monorail crane drive unit. Then, the STM32F429 microcontroller is used as the data processing module to preprocess and extract features from the collected data and form a local database. Furthermore, using local data Train the diagnostic model; once training is complete, obtain the updated diagnostic model. and the corresponding model weights .

[0187] Step S42: Model update upload and secure aggregation.

[0188] First, AES-256 (Advanced Encryption Standard) is used to encrypt the locally trained model weight updates at each edge node and then upload them to the cloud. The model weight updates are represented as follows:

[0189]

[0190] In the formula, These are the weights for the initial global model; For the first The model weights after local training at each edge node; For the first The weight change amount uploaded by each edge node;

[0191] Then, the cloud server initiates an aggregation algorithm to generate a new generation of global models. This invention uses a federated average algorithm to achieve aggregation, as shown below:

[0192]

[0193] In the formula, These are the global model weights; Indicates the first One node; For the first The number of local data samples used by each edge node in this round of training; The sum of data samples used by all edge nodes participating in this round of aggregation;

[0194] Step S43: Cloud-based model evaluation and redistribution.

[0195] First, the global model generated by aggregation is processed in the cloud. The model is trained and its performance is evaluated using metrics including accuracy, recall, and F1 score. Based on the evaluation results, a decision is made regarding whether to distribute the global model: if the new global model shows a significant performance improvement over the previous generation, the evolution is considered effective, and the new global model can be repackaged and distributed to each edge node for diagnostic model updates; if the performance improvement is not significant or deterioration occurs, the update is discarded, and the training process is restarted to further optimize the global model performance.

[0196] Step S44: Personalized fine-tuning of edge nodes.

[0197] Edge nodes receive and load new global models Then, a small amount of local data is used to fine-tune the model, generating a personalized model that is more suitable for fault diagnosis of downhole equipment.

[0198] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited thereto. Various changes that can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention are all within the protection scope of the claims of the present invention.

Claims

1. A sensor-perception collaborative design method for fault diagnosis of a monorail crane drive unit under cloud-edge resource constraints, characterized in that, Includes the following steps: Step S1: Taking the drive unit of the monorail as the object, candidate monitoring areas are obtained through finite element structure and temperature field analysis. After quantitative evaluation based on simulation data, the optimal monitoring position is determined as the best deployment point for the sensor. Step S2: Based on the optimal sensor deployment point selected in Step S1, perform customized design and manufacturing of the sensor; Step S3: Construct a lightweight diagnostic model that is adaptable to resources and operating conditions. Introduce a resource-constrained neural architecture search to automatically design a lightweight network. First, integrate classic lightweight network modules to build a module pool. Design a multi-objective reward function that combines diagnostic accuracy and computational complexity. Combine reinforcement learning algorithms and hard connection rules to complete the network architecture search. Then, use the Pareto optimal strategy to select models that are adapted to edge hardware. Finally, through adaptive design, dynamically adjust the cost of the diagnostic model or update parameters according to resources and operating conditions to achieve a balance between high accuracy and low complexity of the diagnostic model under different constraints. Step S4: Deploy the designed diagnostic model to each monorail crane's onboard edge node. The node uses sensor data to complete local model training. The local model weight update is encrypted and uploaded to the cloud. The cloud uses a federated averaging algorithm to aggregate the weights and generate a new generation of global model. The cloud performs performance evaluation on the global model and distributes the global model with significantly improved performance to each edge node. If the performance does not meet the requirements, training is restarted. After receiving the new global model, the edge node completes personalized fine-tuning using a small amount of local data to generate a fault diagnosis model adapted to downhole working conditions.

2. The sensing-perception collaborative design method for fault diagnosis of monorail gantry drive unit under cloud-edge resource constraints as described in claim 1, characterized in that, In step S1, taking the monorail crane drive unit as the object, candidate monitoring areas are obtained through finite element structural and temperature field analysis. The optimal monitoring location is then determined after quantitative evaluation using simulation data, serving as the best sensor deployment point. Specifically: Step S11: Establish a three-dimensional model of the key structure of the monorail crane drive unit, set material properties and perform mesh generation in ANSYS; then apply boundary condition loads to the structure to obtain the distribution of stress, displacement, strain energy density and safety factor. Step S12: Perform temperature modeling on the key structures of the monorail crane drive unit, calculate the heat generation power and estimate the temperature, using the following formula: ； ； ； In the formula, This refers to the heating power; and These are the total energy power input to the heat source and the effective energy power output to the outside, respectively. The surface temperature of the heat source; The current ambient temperature; Heat flux density; The thickness of the heat source surface layer; To characterize the thermal conductivity of the material; The area of ​​the heat source; Then, the temperature distribution along the heat conduction path is quantified, and the temperature at different locations is calculated along the main heat conduction path. The calculation formula is as follows: ； In the formula, Location along the heat conduction path Temperature at that location; Temperature at the reference point; Heat flow rate; For position The cross-sectional area perpendicular to the direction of heat flow; The coordinates represent the arc length of the heat conduction path; Step S13: Use the entropy weight method to comprehensively select from multiple optimal deployment points obtained from structural analysis and temperature field analysis to determine the optimal deployment point of the sensor.

3. The sensing-perception collaborative design method for fault diagnosis of monorail gantry drive unit under cloud-edge resource constraints as described in claim 2, characterized in that, In step S13, the entropy weight method is used to comprehensively select multiple optimal deployment points obtained from structural analysis and temperature field analysis, specifically as follows: Step S131: Establish the original indicator matrix X and standardize the indicators: ； in, For the first The deployment point is at the Standardized values ​​for each evaluation indicator; Number of measurement points; For the number of indicators; Step S132: Standardize the indicators to obtain a matrix. : ； ； In the formula, For the first Deployment point at Standardized values ​​for each indicator; Step S133: Calculate the entropy value of the index. and weight : ； ； ； In the formula, ; Indicates the first The first indicator The proportion of indicator values ​​for each deployment point; Step S134: Calculate the overall score for each deployment point: ； Based on the overall score The deployment points are sorted, and the higher the score, the more suitable they are for deployment.

4. The sensing-perception collaborative design method for fault diagnosis of monorail gantry drive unit under cloud-edge resource constraints as described in claim 1, characterized in that, In step S2, the customized design and manufacturing of the sensor based on the optimal sensor deployment point selected in step S1 specifically includes: Step S21: Use a high-precision 3D scanner to scan the target surface of the downhole equipment from multiple angles to obtain complete surface point cloud data. Then, import the obtained point cloud data into MeshLab and use the iterative nearest point algorithm for registration and denoising to generate a surface model. ; Step S22: Based on the surface model data obtained from the scan in S21, design the substrate of the conformal surface sensor as a flexible-rigid composite structure: the surface contact layer uses a flexible material and is designed as a honeycomb to achieve zero-gap fit with the structural surface and provide mechanical buffer; the internal support layer uses a rigid material to provide structural support and maintain the stability of the sensing element. Step S23: Plan the printing trajectory of the conformal sensor on the curved surface. The specific steps are as follows: Step S231: Obtain the surface model through calculation. The centroid of the space The calculation formula is as follows: ; In the formula, , , respectively surface model about , , The static moment of a plane; For curved surface models The area; For curved surface models Area elements; Step S232: The curved surface model edge line towards the centroid Scale proportionally to obtain The set of curves is denoted as . Wire; Step S233: Place each The line is evenly divided into A line segment is the set of its endpoints, and the set of its endpoints is denoted as a matrix. The formula is as follows: ； Step S234: Convert the matrix Transpose to obtain the transpose matrix The formula is as follows: ； Connect the points corresponding to the elements in the column vectors of the transpose matrix to generate... Wire, Line and Lines intertwine on the curved surface to form a grid; Step S235: During printing, select the corresponding grid points according to the trajectory design scheme, connect them to obtain polylines, and smooth the polylines to obtain curves with continuous curvature and conformal to the surface contour. The final curve is obtained. This refers to the movement trajectory of the nozzle during the sensor printing process; Step S24: Multi-material direct writing layer printing and curing; Step S241, Surface pretreatment: Thoroughly clean the area to be monitored with anhydrous ethanol; Step S242, Insulating substrate printing: In the pre-treated monitoring area, an insulating substrate is printed using direct-write molding 3D printing; Step S243, Printing and curing of lower layer conductors: On the cured insulating substrate, using silver paste as the conductive material, three parallel lower layer conductors are printed through a 0.21 mm needle, an extrusion pressure of 0.25 MPa, and a printing speed of 5 mm / s. Step S244, Insulating Strip Printing and Curing: Print three parallel epoxy resin insulating strips vertically above the lower conductor. Step S245, Printing and curing of upper conductors: Print upper conductors with T-shaped connectors above the insulating strip, using the same printing parameters as the lower conductors; Step S246, Sensor Unit Printing and Curing: Using carbon paste as the sensing material, the sensor unit array is printed one by one through a 0.21 mm needle at an extrusion pressure of 0.25 MPa and a printing speed of 5 mm / s. Each unit is approximately 0.35 mm long. Step 247, Encapsulation Layer Printing and Curing: The encapsulation layer is printed with epoxy resin. The printing parameters and curing process are the same as in step S242. This is to protect the internal wires and sensing units and improve the mechanical strength and environmental adaptability of the overall structure.

5. The sensing-perception collaborative design method for fault diagnosis of monorail crane drive unit under cloud-edge resource constraints as described in claim 1, characterized in that, In step S24, the multi-material direct writing layer printing and curing specifically includes: Step S241, Surface pretreatment: Thoroughly clean the area to be monitored with anhydrous ethanol; Step S242, Insulating substrate printing: In the pre-treated monitoring area, an insulating substrate is printed using direct-write molding 3D printing; Step S243, Printing and curing of lower layer conductors: On the cured insulating substrate, using silver paste as the conductive material, three parallel lower layer conductors are printed through a 0.21 mm needle, an extrusion pressure of 0.25 MPa, and a printing speed of 5 mm / s. Step S244, Insulating Strip Printing and Curing: Print three parallel epoxy resin insulating strips vertically above the lower conductor. Step S245, Printing and curing of upper conductors: Print upper conductors with T-shaped connectors above the insulating strip, using the same printing parameters as the lower conductors; Step S246, Sensor Unit Printing and Curing: Using carbon paste as the sensing material, the sensor unit array is printed one by one through a 0.21 mm needle at an extrusion pressure of 0.25 MPa and a printing speed of 5 mm / s. Each unit is approximately 0.35 mm long. Step 247, Encapsulation Layer Printing and Curing: The encapsulation layer is printed with epoxy resin. The printing parameters and curing process are the same as in step S242. This is to protect the internal wires and sensing units and improve the mechanical strength and environmental adaptability of the overall structure.

6. The sensing-perception collaborative design method for fault diagnosis of monorail gantry crane drive unit under cloud-edge resource constraints as described in claim 1, characterized in that, In step S3, the construction of a lightweight diagnostic model adaptable to resources and operating conditions involves introducing a resource-constrained neural architecture search to automatically design a lightweight network. First, classic lightweight network modules are integrated to build a module pool. A multi-objective reward function that balances diagnostic accuracy and computational complexity is designed. Network architecture search is completed using reinforcement learning algorithms and hard-connection rules. Then, a Pareto optimal strategy is used to select models adapted to edge hardware. Finally, adaptive design dynamically adjusts the diagnostic model overhead or updates parameters according to resources and operating conditions, achieving a balance between high accuracy and low complexity of the diagnostic model under different constraints. Specifically: Step S31: Integrate the lightweight modules contained in the classic lightweight network, and construct a module pool in the action space that includes one-dimensional convolution, depthwise separable convolution, Ghost module, group convolution, attention mechanism, max pooling, pyramid pooling, fully connected layer, Softmax terminal state layer, and optional parameters for each module. Step S32: Design a multi-objective reward function to form a comprehensive evaluation index to guide the direction of model optimization; Step S33, Agent Search Strategy: The DQN reinforcement learning algorithm is adopted, and hard connection rules are integrated into the Ɛ-greedy strategy to form the search strategy of the network architecture. The Ɛ-greedy strategy is represented as follows: ； In the formula, λ represents the learning rate, which is set to 0.01; Indicates the execution of an action The instant reward received afterwards; The discount factor, representing the importance of future rewards, is set to 0.99; In the new state Below, the maximum effect of all possible actions value; Step S34, Agent Search Process: The agent calculates the corresponding reward value based on the model structure existing in the current state space. The agent predicts and selects the next action based on the search strategy to obtain a higher cumulative reward, and so on, until the agent selects the termination layer, thus completing the design of the complete model; the actual performance of the designed model is then evaluated... The value is updated, and this guides the agent to construct higher-value modules in subsequent iterations. The model of value; Step S35: Select a suitable diagnostic model for deployment at the edge layer using a Pareto optimal strategy. Choose a suitable set of diagnostic models from the constructed model space, based on the following criteria: ； In the formula, This indicates that the model Strictly superior to model ; and Representing the model respectively The diagnostic accuracy and floating-point calculations are then performed; from this set of Pareto optimal models, models that can be deployed on specific hardware resources are selected.

7. The sensing-perception collaborative design method for fault diagnosis of monorail crane drive unit under cloud-edge resource constraints as described in claim 1, characterized in that, In step S32, the multi-objective reward function formula is as follows: ； In the formula, , These represent the total number of samples in the test set and the number of samples diagnosed with faults, respectively. Floating-point numbers for calculating the diagnostic model; This represents the maximum computing resource capacity for the edge layer. and These are the diagnostic accuracy indicators. and computational complexity metrics Weighting.

8. The sensing-perception collaborative design method for fault diagnosis of monorail gantry drive unit under cloud-edge resource constraints as described in claim 1, characterized in that, In step S4, the designed diagnostic model is first deployed to each monorail crane's onboard edge node, and the node uses sensor data to complete local model training; then the local model weight update is encrypted and uploaded to the cloud, and the cloud uses a federated averaging algorithm to aggregate the weights to generate a new generation of global model; subsequently, the cloud performs performance evaluation on the global model, and distributes the global model with significantly improved performance to each edge node, and if it does not meet the requirements, training is restarted. After receiving the new global model, the edge nodes perform personalized fine-tuning using a small amount of local data to generate a fault diagnosis model adapted to downhole working conditions, specifically: Step S41: Deploy the diagnostic model designed in step S3 to each monorail crane's onboard edge node. Each node uses the conformal sensor manufactured in step S2 to collect the monorail crane's drive unit's operating data. Then, the data processing module preprocesses and extracts features from the collected data to form a local database. Use local data Train the diagnostic model; once training is complete, obtain the updated diagnostic model. and the corresponding model weights ; Step S42: Update and encrypt the locally trained model weights for each edge node and upload them to the cloud. The updated model weights are represented as follows: ； In the formula, These are the weights for the initial global model; For the first The model weights after local training at each edge node; For the first The weight change amount uploaded by each edge node; Then, the cloud server initiates an aggregation algorithm to generate a new generation of global models. Aggregation is achieved using a federated average algorithm, as shown below: ； In the formula, These are the global model weights; Indicates the first One node; For the first The number of local data samples used by each edge node in this round of training; The sum of data samples used by all edge nodes participating in this round of aggregation; Step S43: Analyze the global model generated by aggregation in the cloud. The model is trained and its performance is evaluated using metrics including accuracy, recall, and F1 score. Based on the evaluation results, a decision is made regarding whether to distribute the global model: if the new global model shows a significant performance improvement over the previous generation, the evolution is considered effective, and the new global model is repackaged and distributed to edge nodes for diagnostic model updates; if the performance improvement is not significant or deterioration occurs, the update is discarded, and the training process is restarted to further optimize the global model performance. Step S44: Edge nodes receive and load the new global model. Then, a small amount of local data is used to fine-tune the model, generating a personalized model that is more suitable for fault diagnosis of downhole equipment.

Images