Lightweight helicopter fault diagnosis method and device based on cloud-side cooperation, and medium

Abstract

According to the lightweight helicopter fault diagnosis method and device based on cloud edge cooperation and the medium, a dynamic cooperative distributed intelligent diagnosis system is constructed between the cloud end and the edge end, so that efficient lightweight reasoning of the edge end and centralized optimization training of the cloud end are realized; therefore, an aviation health management platform with self-learning, self-adaption and online updating capabilities is constructed. The whole system adopts a cloud-edge-end three-level architecture, wherein the end side is responsible for data acquisition and preprocessing; the edge side undertakes real-time diagnosis and lightweight model reasoning; and the cloud is responsible for global training, model scheduling and strategy optimization. Cyclic interaction of model parameters, task instructions and diagnosis results is achieved between the cloud and the edge through a secure communication link, and a closed-loop intelligent updating mechanism is formed. According to the method, the real-time performance, the computing power efficiency, the model generalization ability and the system adaptability of a fault diagnosis system are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of helicopter fault diagnosis technology, and specifically to a lightweight helicopter fault diagnosis method, device and storage medium based on cloud-edge collaboration. Background Technology

[0002] In existing technologies, fault diagnosis for aviation equipment, especially helicopters, has undergone a phased evolution from threshold-based rule-based discrimination and frequency-domain feature-driven vibration diagnosis to deep learning-driven end-to-end pattern recognition. In the traditional stage, diagnostic logic was typically deployed at ground stations or backend servers, with data acquisition and storage only performed during flight, followed by offline analysis after flight. While this approach has no deployment constraints in terms of computational resources and algorithm complexity, it cannot meet real-time requirements, especially in the face of sudden faults, in-process degradation, and time-sensitive safety threats. Furthermore, the data closure delay caused by offline operation leads to significant lags in health management and maintenance decisions, making it difficult to support rapid response under support conditions.

[0003] To overcome the inherent latency problem of offline analysis, some studies have attempted to move diagnostic functions to the edge, such as deploying them on airborne terminals, avionics integrated processing platforms, or lightweight embedded computing units for online identification. However, the operating characteristics of helicopters are significantly different from those of fixed-wing aircraft. Their power transmission chain, rotor system, airframe aeroelastic load, and vibration coupling result in highly non-stationary signal spectrum characteristics and strong noise background. Running large models directly on the edge is difficult due to constraints in terms of computing power, power consumption, and storage. Excessive model compression sacrifices accuracy and robustness, especially under extreme conditions, complex interference, or rare fault samples, where the identification capability of lightweight edge models is significantly reduced, leading to insufficient reliability of online diagnostics and failing to achieve airworthiness-grade stability. Furthermore, upgrading and maintaining edge deployments is costly, infrequently adjustable, and difficult to guarantee version consistency, limiting their lifecycle evolution and adaptation.

[0004] like Figure 1As shown, to compensate for the shortcomings of insufficient edge computing power and the limitations of model evolution and manageability, a collaborative paradigm has emerged in academia and industry in recent years. This paradigm utilizes cloud resources for centralized training, unified management, and online updates of diagnostic models, and then quantizes and tailors them into lightweight inference subnetworks that can run on mobile devices. This cloud-edge model attempts to decouple training and inference, achieving high-precision model evolution through the cloud and low-latency closed-loop execution through the edge, thus ensuring that the diagnostic process balances real-time performance and maintainability. However, this direction is still in the experimental exploration stage, and there are several unresolved issues in aviation scenarios, especially helicopter scenarios: First, the communication link between the cloud and the flight end has inherent constraints such as intermittent and limited bandwidth, and the inability to guarantee continuous connection, which makes it impossible to achieve deterministic and on-demand synchronization of model updates, and may result in lagging or even inconsistent end-side models; Second, the cloud-based large model pruning and distillation strategies are mostly based on general operating conditions, which do not fully cover the operating condition disturbance modes and early characteristics of minor faults of rotorcraft. In the pruning process, key sensitivities are easily sacrificed, causing the combination of cloud-based generation and edge-based push to lose its ability to cover extreme safety events; Third, the cloud-edge collaborative control plane and data plane still lack formal provable mechanisms for airworthiness and safety audits, making it difficult for such systems to be implemented in real helicopter service environments.

[0005] Against this backdrop, the industry has attempted to introduce lightweight techniques such as model quantization, pruning, and knowledge distillation into fault diagnosis algorithms to achieve runnable inference in edge-side computing environments. However, these lightweight methods are still in the exploratory stage in helicopter scenarios, and their applicability and reliability have not yet been verified at the engineering level. Taking quantization as an example, general int8 and mixed-precision quantization schemes can significantly reduce model size and computational load in fields such as natural image recognition. However, when applied to vibration and multi-source sensor data with strong noise, strong non-stationarity, and weak early fault characteristics in helicopters, quantization errors will be directly superimposed on the already low signal-to-noise ratio representation, further compressing the discrimination margin of early signs and leading to a significant decrease in detection sensitivity and robustness. In other words, quantization technology cannot simply reuse general parameter strategies in such high-safety scenarios; its accuracy loss is unacceptable, and a mature scenario-based methodology is still lacking.

[0006] Similarly, while model pruning can reduce redundant parameters at both structured and unstructured levels, making the model deployable at the edge, existing pruning strategies mostly originate from natural data distribution or steady-state industrial equipment diagnostic problems. Their sensitive layer retention and pruning criteria do not constrain the time-varying nature of rotor system vibration modes, which can easily lead to the mispruning of critical discrimination channels, resulting in missed detections or misjudgments under extreme operating conditions and rare faults. In addition, the stability and interpretability of pruned models lack verifiable means, and they cannot pass formal audits or demonstrations in airworthiness environments. Therefore, they remain in the laboratory or pilot research stage and are difficult to directly incorporate into airborne operations.

[0007] Knowledge distillation is a common technique in general machine learning used to guide lightweight student models with complex teacher models to achieve near-performance. However, distillation assumes that the output or intermediate representation of the teacher model itself has sufficient reliability and stability. Large models for helicopter fault diagnosis are often trained on limited samples, and their predictions are uncertain under abnormally scarce operating conditions and combined perturbation scenarios. This causes the distillation process to synchronously transfer potential biases from the teacher model to the student model, "solidifying" the distortion and thus generating uncontrollable systematic bias risks in unsupervised real-world flight environments. Furthermore, existing distillation strategies mostly optimize logit or feature distribution for classification tasks, while helicopter diagnosis often involves multimodal time-series signal fusion. Cross-layer and cross-modal distillation lacks a stable and effective framework, and empirical accumulation is insufficient to support engineering implementation.

[0008] As aviation equipment develops towards intelligence and networking, helicopters, as complex mechatronic systems, rely heavily on fault diagnosis technology for flight safety and maintenance. Current helicopter fault diagnosis largely depends on the integrated application of airborne sensor data acquisition, signal processing, and model recognition technologies. However, despite significant progress in aviation health management in recent years using artificial intelligence, big data, and cloud computing, existing helicopter fault diagnosis systems still have many shortcomings. These deficiencies severely restrict the system's real-time performance, accuracy, and reliability in practical applications, especially in environments with multi-scenario collaboration, limited edge computing resources, and highly dynamic flight missions. The following analysis examines the shortcomings of existing diagnostic methods: 1) Insufficient real-time performance Existing fault diagnosis systems rely excessively on centralized cloud computing, making it difficult to meet real-time requirements. Traditional cloud-based aviation health management systems typically upload flight data to ground servers or cloud platforms via communication links for unified analysis and processing. While this architecture can utilize the powerful computing resources of the cloud for complex algorithm training and model inference, it cannot achieve real-time status monitoring and fault early warning during flight due to transmission latency, bandwidth limitations, and network interruptions. When helicopters are in remote areas, complex weather environments, or low-altitude communication blind spots, the cloud computing model becomes almost ineffective, failing to support online diagnosis and emergency response for critical components. Therefore, existing technologies have significant shortcomings in terms of real-time performance and mission continuity. At the edge, insufficient computing power and unbalanced system load lead to low diagnostic efficiency. Although some intelligent maintenance terminals based on edge computing have emerged in recent years, offloading some data processing tasks to airborne equipment or near-field nodes, the limited size, power consumption, and computing power of airborne computing units make it difficult to support complex deep learning model operation. Furthermore, edge algorithms have deficiencies in training data scale, computational accuracy, and model update mechanisms, resulting in instability and decreased accuracy of diagnostic results. Especially in multi-task flight scenarios, edge devices have difficulty dynamically allocating resources according to task priorities, resulting in "computing bottlenecks" and "data silos," which seriously affect the overall system performance.

[0009] 2) Research on the optimization and lightweighting of multimodal model algorithms is lacking. Existing helicopter fault diagnosis systems suffer from shortcomings in data fusion and model generalization capabilities. Helicopters have complex structures, tight coupling between systems, and diverse fault types with complex evolution paths. Current mainstream diagnostic algorithms are mostly based on single sensor signals or single-source data modeling, making it difficult to fully utilize multi-source heterogeneous data, such as the inherent correlations between vibration signals, temperature, current, voltage, acoustic characteristics, and flight state parameters. This results in diagnostic models being sensitive to environmental changes, having poor generalization ability, and experiencing significant performance degradation across different aircraft types, missions, or load conditions. Furthermore, the lack of unified data standards and high-quality labeled samples makes it difficult for existing algorithms to perform continuous learning and adaptive optimization, leading to gradual model degradation over long-term operation.

[0010] In cloud-based centralized processing architectures, a large amount of sensitive flight data needs to be transmitted via public networks or dedicated links. Attacks on these communication links or data leaks directly threaten flight safety and information security. Existing systems mostly employ static encryption or access control methods, lacking dynamic trust mechanisms and end-to-end security guarantees, making them ill-equipped to handle complex network attack scenarios. Furthermore, the centralized processing model of cloud platforms carries the risk of single points of failure; if a server crashes or the network malfunctions, the entire diagnostic system will be paralyzed, unable to provide continuous support for flight missions. In addition, the conflict between system lightweighting and energy consumption control is becoming increasingly prominent. Airborne diagnostic systems are subject to strict weight, size, and energy consumption constraints in their design, while existing intelligent algorithms often have high computational complexity and large storage requirements, making them difficult to run efficiently on embedded platforms. Some research has attempted to achieve algorithm lightweighting through model pruning, quantization, or distillation techniques, but a unified standard and reliable framework have not yet been established, and model compression often comes at the cost of accuracy, affecting the stability and reliability of fault identification. Therefore, how to achieve high-precision fault diagnosis under limited computing power and energy consumption conditions remains a bottleneck that current technologies struggle to overcome.

[0011] 4) Low efficiency of cloud-edge collaborative scheduling Existing technologies lack effective coordination and scheduling strategies for cloud-edge collaboration mechanisms. Although the concept of cloud-edge collaboration has been applied to some extent in the industrial internet field, its practice in the aviation field is still in its early stages. Current systems are mostly based on simple interactions of data uploading and result distribution, lacking refined task division and dynamic scheduling mechanisms. An efficient collaborative inference link has not been formed between edge nodes and the cloud, resulting in high data transfer latency, unstable task synchronization, and an inability to achieve a dynamic optimal balance between computing resources, communication bandwidth, and energy consumption. Furthermore, the lack of a unified collaboration framework leads to poor compatibility between different manufacturers and different models, making it difficult to achieve system-level cross-platform collaboration and unified management. Summary of the Invention

[0012] The present invention proposes a lightweight helicopter fault diagnosis method, device and storage medium based on cloud-edge collaboration, which can at least solve one of the technical problems in the background art.

[0013] To achieve the above objectives, the present invention adopts the following technical solution: A lightweight helicopter fault diagnosis method based on cloud-edge collaboration involves performing the following steps via computer equipment. S1. In the initial stage of system deployment, the basic architecture and task scheduling framework for a global diagnostic model are established in the cloud. S2. Deploy a lightweight neural network model on edge computing nodes based on S1; S3. After completing the edge model pruning and lightweight optimization, a knowledge distillation algorithm combining physical equation constraints is proposed. It not only uses the high-precision cloud model, i.e. the teacher model, to transfer high-level semantic knowledge to the edge lightweight model, i.e. the student model, but also introduces physical priors such as helicopter dynamics equations, structural vibration equations and energy conservation constraints to achieve the integration of data-driven and mechanism model learning. S4. The cloud scheduling algorithm is used to realize the core hub of the "cloud-edge" collaborative intelligent diagnostic system. It not only undertakes the resource management and task allocation of multiple nodes, but also is responsible for the dynamic scheduling of model updates, computing power coordination, network traffic control and collaborative feedback. S5. After completing the lightweight optimization of the edge model, physical constraint distillation, and cloud intelligent scheduling, a cloud-edge collaborative loop and online model update mechanism is proposed to build the system's self-learning and continuous evolution capabilities. Through this mechanism, cloud and edge nodes form a dynamic closed loop of data, model, and feedback, enabling the system to have intelligent characteristics of self-awareness, self-adjustment, and self-evolution in long-term operation, thereby achieving continuous improvement in diagnostic accuracy and dynamic optimization of system performance.

[0014] In another aspect, the present invention also discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method described above.

[0015] In another aspect, the present invention also discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method described above.

[0016] As can be seen from the above technical solution, the lightweight helicopter fault diagnosis method based on cloud-edge collaboration of the present invention aims to overcome the problems of insufficient real-time performance, limited computing power, poor model generalization, low scheduling efficiency, and lagging system updates in the existing technology. By constructing a dynamically collaborative distributed intelligent diagnostic system between the cloud and the edge, efficient lightweight inference at the edge and centralized optimization training in the cloud are achieved, thereby constructing an aviation health management platform with self-learning, self-adaptation, and online update capabilities.

[0017] The system adopts a three-tier architecture of "cloud-edge-device": the device side is responsible for data collection and preprocessing; the edge side is responsible for real-time diagnosis and lightweight model inference; and the cloud side is responsible for global training, model scheduling, and policy optimization. The cloud and edge sides communicate cyclically through a secure communication link to achieve the interaction of model parameters, task instructions, and diagnostic results, forming a closed-loop intelligent update mechanism.

[0018] The core technological innovations of this invention are reflected in three aspects: ① edge model pruning and reinforcement learning collaborative optimization algorithm; ② knowledge distillation mechanism based on physical equation constraints; ③ cloud task scheduling and cloud-edge collaborative update strategy.

[0019] Specifically, the beneficial effects of the present invention are as follows: The lightweight helicopter fault diagnosis system and method based on cloud-edge collaboration of the present invention significantly improves the real-time performance, computing efficiency, model generalization ability and system adaptability of the fault diagnosis system through three core technological innovations: edge model pruning and reinforcement learning collaborative optimization algorithm, knowledge distillation mechanism based on physical equation constraints, and cloud task scheduling and cloud-edge collaborative update strategy.

[0020] Starting with the technical solution and comparing it with the shortcomings of existing technologies, the advantages and effects of each innovation are explained in detail below: 1) Edge model pruning and reinforcement learning co-optimization algorithm Traditional helicopter fault diagnosis systems rely on centralized cloud-based models. Limited by the computing power, storage, and energy consumption constraints of the airborne platform, complete deep learning models cannot be directly deployed at the edge, resulting in insufficient real-time diagnostic capabilities. Existing model lightweighting methods (such as manual pruning and rule-based compression) lack adaptive mechanisms, easily leading to accuracy loss or low compression efficiency. This invention proposes a multi-stage pruning optimization algorithm based on reinforcement learning (RL). By dynamically evaluating the pruning feasibility of the model structure, automating policy search, and optimizing hierarchical dependencies, it achieves efficient lightweighting of the edge model. The cloud-based analysis module quantifies the contribution and redundancy of each layer's parameters by calculating channel importance indicators (Grad-CAM gradient weighting), redundancy indicators (inter-channel correlation coefficients), and structural dependency matrices, generating a pruning candidate list.

[0021] The RL agent dynamically adjusts the pruning ratio (e.g., retaining, deleting, or proportionally pruning) to avoid over-compression or accuracy degradation caused by fixed-threshold pruning in traditional methods. For example, experiments show that this method reduces the number of model parameters by 72% and improves inference speed by 3.5 times while maintaining 98% diagnostic accuracy. The RL agent adopts an Actor-Critic architecture with policy gradient, using a reward function to comprehensively balance accuracy loss, computational latency, and energy consumption. The reward mechanism guides the agent to prioritize pruning channels that contribute little to the output but have high energy consumption, achieving an optimal trade-off between computational power and accuracy. Compared with traditional pruning methods (such as L1 regularization), this method reduces diagnostic error by 41% and energy consumption by 28% at the same compression rate.

[0022] 2) Knowledge distillation mechanism based on physical equation constraints Purely data-driven knowledge distillation (such as soft label transfer based on KL divergence) lacks physical mechanism constraints, leading to "physical distortion" in student models when facing nonlinear coupled faults or rare operating conditions, resulting in insufficient interpretability. Existing methods struggle to balance lightweight design with diagnostic stability. This invention proposes Physically Constrained Knowledge Distillation (P-KD), which incorporates physical priors such as helicopter dynamics equations and the law of energy conservation into the distillation process. Through multi-layer feature alignment and joint optimization, it enables student models to achieve diagnostic capabilities similar to teacher models with low computational complexity.

[0023] The distillation process simultaneously transfers the output layer probability distribution, intermediate layer feature representations, and gradient directions, achieving structural mapping through KL divergence and gradient consistency loss. Compared to traditional distillation methods (which only align the output layer), this method maintains 96% feature similarity on a lightweight model (reducing the number of parameters by 85%), and improves the F1 score for fault identification by 14%.

[0024] 3) Cloud-based task scheduling and cloud-edge collaborative update strategies Traditional scheduling strategies rely on static priorities or fixed rules, making it difficult to adapt to the dynamic environment of helicopter swarms (such as fluctuations in computing power and changes in network latency), resulting in high task execution latency and low resource utilization. Existing system update mechanisms are lagging and unable to respond promptly to new failure modes. This invention proposes a cloud-based scheduling framework of "intelligent perception—adaptive decision-making—dynamic collaboration—closed-loop optimization," which combines task semantic relevance and energy consumption awareness mechanisms to achieve optimal global resource allocation and continuous model evolution. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the cloud-edge collaboration process in existing technologies; Figure 2 This is the flowchart of the cloud-edge collaborative loopback and online model update mechanism in an embodiment of the present invention; Figure 3 This is the edge-side model pruning and reinforcement learning optimization process in an embodiment of the present invention; Figure 4 This is the overall flow of the knowledge distillation algorithm combined with physical equation constraints in an embodiment of the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0027] This invention proposes a lightweight helicopter fault diagnosis system and method based on cloud-edge collaboration, aiming to overcome the problems of insufficient real-time performance, limited computing power, poor model generalization, low scheduling efficiency, and lagging system updates in existing technologies. By constructing a dynamically collaborative distributed intelligent diagnostic system between the cloud and the edge, efficient lightweight inference at the edge and centralized optimized training in the cloud are achieved, thereby building an aviation health management platform with self-learning, self-adaptation, and online update capabilities.

[0028] The system adopts a three-tier architecture of "cloud-edge-device": the device side is responsible for data collection and preprocessing; the edge side is responsible for real-time diagnosis and lightweight model inference; and the cloud side is responsible for global training, model scheduling, and policy optimization. The cloud and edge sides communicate cyclically through a secure communication link to achieve the interaction of model parameters, task instructions, and diagnostic results, forming a closed-loop intelligent update mechanism.

[0029] The core technological innovations of this invention are reflected in three aspects: ① a collaborative optimization algorithm for edge model pruning and reinforcement learning; ② a knowledge distillation mechanism based on physical equation constraints; and ③ a cloud task scheduling and cloud-edge collaborative update strategy. The specific process is as follows: Figure 2 As shown, the following is a detailed explanation: Step 1: System Initialization and Data Collaborative Modeling In the initial deployment of this invention's system, a global diagnostic model infrastructure and task scheduling framework are established in the cloud. The cloud model employs a Transformer-based fault diagnosis model to extract features from multi-source sensor data and achieve intelligent identification of multiple fault modes. Input data includes, but is not limited to, vibration signals, rotational speed, temperature, current, voltage, sound waves, and flight status parameters.

[0030] Edge sensor nodes acquire raw data through a high-speed acquisition module. The data then undergoes feature preprocessing via an edge gateway, including signal denoising, normalization, wavelet packet decomposition, temporal spectrum extraction, and feature dimensionality reduction. After being standardized into standard feature vectors by a feature mapping module, the data is fed into the edge computing unit for real-time inference.

[0031] The cloud-based system establishes multi-domain data mapping relationships based on different device models, task scenarios, and environmental parameters, and utilizes incremental learning technology to fuse historical and real-time data for modeling. Through a time-window-based sliding sampling strategy, new data is incorporated into the online training set, dynamically adjusting global model parameters to achieve rapid adaptation to different operating conditions.

[0032] During this phase, the system simultaneously establishes a cloud-edge communication topology table and a computing resource mapping table. The topology table identifies the network bandwidth, transmission latency, and communication reliability of each helicopter node; the resource mapping table records the CPU and GPU load rates, storage capacity, and available energy consumption limits of each edge computing unit, providing a foundation for subsequent intelligent scheduling and model deployment.

[0033] Step 2: Edge-side model pruning and reinforcement learning optimization To achieve real-time and efficient fault diagnosis at the edge, this invention deploys a lightweight neural network model on edge computing nodes. Due to limitations in computing power, storage, and energy consumption on airborne platforms, a complete deep learning model cannot be run directly; therefore, structural optimization is needed to reduce model size. This invention proposes a reinforcement learning-driven multi-stage pruning optimization algorithm that significantly reduces model complexity while maintaining diagnostic accuracy, thereby achieving high efficiency and intelligence in edge inference. The process mainly includes four steps, as follows: Figure 3 As shown.

[0034] Step 21: Model Structure Analysis and Pruning Capability Assessment Before performing pruning operations, the system first analyzes the full model structure in the cloud. The model includes multiple convolutional layers, batch normalization layers, fully connected layers, and activation layers, each containing a large number of parameters and feature channels. The cloud analysis module evaluates the model's pruning potential by calculating the following metrics: Channel importance indicators The contribution of each channel to the final output is measured by gradient-weighted average (Grad-CAM) and parameter sensitivity analysis.

[0035] Redundancy index : Calculate the correlation coefficient and mutual information between adjacent channels, and identify channels with high redundancy as potential pruning targets.

[0036] Structural dependency matrix Used to analyze the degree of dependency between features in different layers, preventing pruning from disrupting the information transmission chain.

[0037] Based on this, the cloud creates a pruning candidate list and priority index table for each layer, and distributes this structured metadata to the edge to provide the initial state for pruning decisions for the reinforcement learning agent.

[0038] Step 22: Reinforcement learning modeling of pruning strategies To achieve automated structure search and optimal pruning decisions, this invention designs a reinforcement learning (RL) intelligent agent for policy learning and iterative optimization in complex pruning spaces. First, each state vector in the state space is defined. It consists of the current layer parameter quantity, number of channels, feature importance index, computational cost (FLOPs), and energy consumption weight.

[0039] ; Next, define the action space. This includes operations such as "keep", "delete", and "crop proportionally". The cropping ratio for each layer can be selected as discrete values ​​such as {0.1, 0.2, 0.3, 0.4, 0.5}.

[0040] With this setup, the RL agent can dynamically adjust the pruning intensity of different layers during the search process to achieve globally optimal allocation. To effectively reward model learning, a reward function based on comprehensive evaluation is designed, taking into account both accuracy loss and reward function. Calculation delay Energy consumption Three factors ; in, and This is a balancing factor. This reward mechanism incentivizes agents to minimize computation while maintaining diagnostic performance.

[0041] The training mechanism employs an Actor-Critic architecture based on policy gradients. Actors generate pruning decisions, and Critics evaluate the merits of these decisions based on performance feedback. Through iterative iteration, the agent gradually converges to the optimal pruning strategy under different constraints, achieving autonomous structural exploration.

[0042] Step 23: Channel-level pruning and local retraining After the reinforcement learning agent determines the optimal pruning ratio for each layer, the edge nodes enter the actual pruning stage. First, channel importance is calculated; through gradient sensitivity analysis, the importance coefficient of each convolutional channel is calculated. : ; in, For loss function, Let be the weight of the ij-th channel. N is the total number of channels.

[0043] Next, channel filtering and deletion are performed. Based on the proportion of reinforcement learning output, the least important channels are removed from each layer, and the parameters of the BN layer and the input dimensions of adjacent layers are adjusted simultaneously to ensure structural continuity. After pruning, the model is locally retrained at the edge nodes, using a cached dataset to correct for the accuracy drop caused by structural changes. Retraining employs a mini-batch incremental learning approach, with the learning rate adaptively decreasing to prevent overfitting.

[0044] Step 24: Hierarchical pruning and cross-level dependency optimization After channel-level pruning, this invention further optimizes the hierarchical structure. The stage uses a cross-layer analysis module to identify functionally redundant layers (repeated convolutional layers or low-contribution fully connected layers) in the model and performs hierarchical pruning in conjunction with reinforcement learning strategies. Feature Map Similarity (FMS) and mutual information metrics are used to identify layer structures with highly similar outputs. ; in, and The feature outputs of layers i and j are respectively. If the FMS exceeds the threshold (0.95), they are considered as mergeable layers. and These are the activated feature values. Similar layers are merged or deleted through parameter remapping and residual compensation. The residual compensation term is used to maintain the stability of the feature flow, ensuring that the pruned model still possesses complete semantic delivery capabilities. If the model accuracy decreases by more than a preset threshold of 3% after pruning, the system automatically triggers a rollback mechanism to restore the previous optimal structure and adjust the pruning ratio, achieving safe redundancy control.

[0045] Step 3: Distillation algorithm based on knowledge of physical equation constraints After completing edge model pruning and lightweight optimization, this invention proposes a Physics-Constrained Knowledge Distillation (P-KD) algorithm to further improve diagnostic accuracy, robustness, and interpretability. This method not only utilizes a high-precision cloud model (teacher model) to transfer high-level semantic knowledge to the lightweight edge model (student model), but also introduces physical priors such as helicopter dynamics equations, structural vibration equations, and energy conservation constraints, achieving a fusion of data-driven and mechanistic model learning. Through this distillation process, the student model can achieve diagnostic capabilities similar to or even more stable than the teacher model with lower computational complexity, and possess interpretable responses to changes in physical states.

[0046] The physical confinement distillation process of this invention consists of the following five main steps, as follows: Figure 4 As shown: Step 31: Teacher Model Establishment and Multi-Source Feature Extraction In a cloud environment, a high-precision teacher network is first constructed. This model is based on a deep temporal convolutional CNN-Transformer hybrid network, used to comprehensively process multi-source heterogeneous signals, including: Vibration acceleration signals (main rotor, tail rotor, drive shaft, reducer); Temperature and current signals (motors and servo systems); Flight attitude and operating parameters (heading angle, pitch angle, load changes, etc.); Acoustic emission and pressure signals, etc.

[0047] The teacher model achieves cross-domain feature alignment and weighting through a multi-modal fusion block. Each signal is processed by a feature extraction unit to form a time-frequency fusion feature matrix, which is then aggregated by a self-attention module to obtain a global feature representation. Trained on a large amount of flight data, the teacher model exhibits high fault identification accuracy and the ability to capture nonlinear coupling features. Its intermediate layer features, output probability distribution, and gradient response map serve as the knowledge source for distillation.

[0048] Step 32: Student Model Design and Input Mapping Structure The Student Network is deployed on edge nodes, and its structure consists of a lightweight network obtained after pruning and optimization in step two. To ensure knowledge compatibility with the teacher model, the student model retains key hierarchical structures (such as kernel size, channel arrangement, and normalization strategies) but reduces its depth and width to adapt to limited computing power. The student model uses the same standardized feature vectors as the teacher model, but an input mapping layer is added at the front end of the network to adaptively align data from different sampling frequencies and sensor sources. Step 33: Modeling Physical Equations and Defining Constraint Terms To overcome the physical distortion problem caused by "pure statistical transfer" in traditional knowledge distillation, this invention introduces helicopter dynamics into the distillation process to construct a physical consistency constraint term.

[0049] 1) Establishment of dynamic model Simplified dynamic equations based on helicopter transmission system: ; in, Let M be the structural vibration displacement vector, and C, K be the mass, damping, and stiffness matrices, respectively. The external load or fault excitation is used. The model obtains key parameters through finite element modeling (FEM) or experiments to describe the vibration response characteristics of the system under different operating conditions.

[0050] 2) Energy conservation and stability constraints: Introducing the energy conservation equation. ; in, To input power, To output power, The energy constraint loss is defined as the energy loss by comparing the difference between the predicted and measured energy spectra. ; in, To predict the energy spectrum. The measured energy spectrum is used. Considering both dynamic equilibrium and energy conservation constraints, a physical consistency loss term is established: ; in, This represents the theoretical response mapping calculated based on the dynamic equations. This constraint ensures that the student model output not only shares similarities with the teacher model but also satisfies physical consistency.

[0051] Step 34: Knowledge Transfer and Multi-Layer Feature Alignment In the distillation process, this invention employs a multi-layer alignment mechanism to achieve knowledge transfer and structural mapping. The knowledge transfer between the teacher and student models includes three types of information: (1) probability distribution knowledge of the output layer; (2) feature representation knowledge of the intermediate layer; and (3) the gradient direction and geometric relationship of the feature space. KL divergence is implemented on the output of each layer, and gradient consistency is used to make the student and teacher networks align.

[0052] Step 35: Joint Optimization and Training Strategy for the Distillation Process Combining the aforementioned loss terms, this invention constructs the overall objective function. ; ; ; in, The standard classification loss for the student model, , , and These are adjustable weighting coefficients. This represents the KL divergence for soft labels. For intermediate layer alignment loss, For physical constraint loss, This is a gradient constraint. The probability score is the softmax score for classification. For linear projection mapping functions, Let be the input features of the i-th layer. The optimization objective is to achieve a balance between physical consistency and structural robustness of the model while ensuring classification accuracy.

[0053] Step 4: Cloud-based scheduling algorithm and intelligent task allocation mechanism The cloud scheduling algorithm of this invention is the core hub for realizing a cloud-edge collaborative intelligent diagnostic system. It not only manages resources and allocates tasks across multiple nodes, but also handles dynamic scheduling for model updates, computing power coordination, network traffic control, and collaborative feedback. Due to the complex operating environment and varied flight missions of helicopter swarms, the computing power, bandwidth, and power consumption of edge nodes change dynamically over time. Therefore, a single static scheduling strategy is insufficient to meet performance requirements in the long term. The cloud scheduling algorithm proposed in this invention adopts a full-process control framework of "intelligent perception—adaptive decision-making—dynamic collaboration—closed-loop optimization," aiming to achieve the overall goals of optimal global resource utilization, minimum task execution latency, and most stable diagnostic performance.

[0054] First, in the initial stage of system operation, the cloud establishes a real-time status awareness mechanism through communication links with all edge nodes, continuously collecting key operational metrics from each node. These metrics include current CPU and GPU load rates, memory usage, network bandwidth, link latency, signal quality, remaining battery power, model version information, and task queue length. The cloud performs multi-dimensional aggregation and standardization processing on this information to form node status vectors, and builds a resource profile database based on these vectors. This database can be regarded as a global "health map" of the system, providing data support for subsequent task scheduling and model allocation. At the same time, the cloud uses a sliding time window mechanism to monitor the trend of node status changes, identify potential performance fluctuations and network bottlenecks, and provide early warning basis for the early adjustment of scheduling strategies.

[0055] Having grasped the overall state, the cloud-based scheduling module enters the decision-making phase. Traditional scheduling methods, often based on fixed rules or static priority ranking, struggle to adapt to task complexity and environmental uncertainty. This invention introduces a dynamic scheduling algorithm framework based on intelligent policy optimization, treating the scheduling process as a multi-objective decision problem. The system needs to strike a balance between diagnostic task latency, communication overhead, energy consumption constraints, model accuracy, and node stability. To achieve this goal, the cloud adopts a multi-layered decision structure: the upper layer is the global policy planning layer, responsible for determining the overall task distribution pattern and model version selection; the lower layer is the local scheduling execution layer, responsible for specifically allocating tasks, determining the execution order, and configuring parameters within each node. Through this hierarchical design, the system can balance global optimization and local adaptability.

[0056] The core of the global planning layer lies in the strategy generation mechanism. The cloud-based scheduling system trains a scheduling strategy model using historical task execution records, node performance curves, and network topology. This model employs a self-learning structure, continuously optimizing its parameters during operation, making the strategy selection process increasingly intelligent. For example, when a node's computing power decreases or network latency increases, the system can automatically identify the node's "abnormal load" state and proactively migrate some diagnostic computation tasks to adjacent, higher-performing nodes or the cloud for execution. Simultaneously, for high-priority or safety-critical flight missions, the cloud allocates redundant diagnostic nodes for dual-path or multi-path verification to ensure the reliability and timeliness of the results.

[0057] During task allocation, the cloud-based scheduling algorithm not only considers computing power and latency factors but also comprehensively incorporates the concept of "task semantic relevance." Different flight missions or fault types have varying dependencies on data features and model versions. The cloud maps tasks based on the matching degree between task feature vectors and the model capabilities of each node. For example, if an edge node has historically performed well in identifying transmission system faults, the system tends to prioritize allocating similar tasks to that node to maximize model experience utilization. This semantically relevant scheduling mechanism breaks through the traditional "computing power priority" allocation logic, making resource allocation more intelligent and context-aware.

[0058] During task execution, the cloud monitors the task progress and result feedback of each node in real time through an asynchronous control channel. The scheduling module has a dynamic adjustment mechanism. Once an execution anomaly is detected (such as task delay timeout, low result confidence, or model drift), the system will immediately trigger an adaptive scheduling process. Adaptive scheduling does not rely on manual intervention but is automatically executed through predefined adjustment rules and strategy models. For example, when the confidence of the fault diagnosis result of an edge node is lower than the threshold, the cloud will automatically forward the corresponding sample to a high-performance node or the cloud main model for re-examination; if a communication interruption is detected, the task will be temporarily stored and will continue to synchronize after the link is restored, ensuring that the task is not interrupted or lost.

[0059] The cloud scheduling algorithm of this invention also introduces an energy consumption awareness and bandwidth adaptive control mechanism. During each task allocation, the system comprehensively evaluates the data transmission volume, computational requirements, and node energy status of the task, and achieves optimal scheduling within available communication resources through a dynamic bandwidth allocation algorithm. When the flight mission is in a phase with good communication conditions, the system tends to distribute complex tasks to edge nodes to fully utilize local computing power; while when the link is unstable or energy consumption is limited, the task is moved to the cloud or a nearby node for execution to reduce energy consumption and transmission overhead. Through this dynamic adjustment mechanism, the cloud can maintain the overall energy efficiency balance and mission continuity of the system at different operational stages.

[0060] At the algorithm implementation level, the cloud-based scheduling system adopts a modular architecture design, consisting of a state awareness module, a decision engine module, an execution management module, and a feedback optimization module. The state awareness module is responsible for collecting system information in real time and updating resource status; the decision engine module generates the optimal task allocation scheme based on the strategy model; the execution management module is responsible for task distribution, node synchronization, and execution monitoring; and the feedback optimization module updates the strategy model based on actual operating results, achieving continuous learning and adaptive optimization. The modules communicate efficiently through message queues and shared memory mechanisms, ensuring that the system completes the scheduling loop within millisecond-level response time.

[0061] Step 5: Cloud-edge collaborative loopback and online model update mechanism After completing the lightweight optimization of the edge model, physical constraint distillation, and cloud-based intelligent scheduling, this invention further proposes a cloud-edge collaborative loop and online model update mechanism to build the system's self-learning and continuous evolution capabilities. Through this mechanism, cloud and edge nodes form a dynamic closed loop of data, models, and feedback, enabling the system to possess intelligent characteristics of self-awareness, self-adjustment, and self-evolution during long-term operation, thereby achieving continuous improvement in diagnostic accuracy and dynamic optimization of system performance.

[0062] The design philosophy of this mechanism is to drive model updates with data flow, promote system optimization with model updates, and feed back the optimization results into the data flow. The cloud, as the global control and knowledge center, is responsible for aggregating operational data from multiple nodes and performing centralized training and evaluation; edge nodes, as frontline intelligent agents, are responsible for real-time inference, local learning, and feedback reporting. The two continuously interact through a highly reliable communication link, forming a closed-loop process of "collection—diagnosis—feedback—relearning—redeployment".

[0063] After receiving feedback data from multiple edge nodes, the cloud platform enters the global analysis phase. The system compares the edge prediction results with the cloud-based main model prediction results using a consistency detection module, calculating a consistency index. When the consistency index falls below a threshold, it indicates that the model's performance has degraded under the current operating conditions or that the environment has changed, and the cloud platform initiates an online update process. Simultaneously, the cloud platform also uses clustering and anomaly detection algorithms to identify new fault samples or rare patterns, adding them to the knowledge base to enrich the training set.

[0064] During the update process, the cloud employs a combination of incremental learning and transfer learning to iteratively train the model. Incremental learning absorbs new data without forgetting old knowledge, while transfer learning enhances the model's adaptability to new tasks through parameter sharing and feature mapping. To prevent catastrophic forgetting, the system introduces a resilient weight retention mechanism, limiting changes in key parameters and thus ensuring the stability of the original diagnostic capabilities. After the updated model has been validated and achieved stable convergence, it is marked as a candidate version by the cloud and enters the deployment phase.

[0065] In view of the embodiments of the present invention, the relevant alternative solutions of the present invention are as follows: The "Lightweight Helicopter Fault Diagnosis System and Method Based on Cloud-Edge Collaboration" proposed in this invention, while possessing unique innovations in its architecture, algorithm framework, and online update mechanism, still offers several functionally equivalent or similarly effective alternatives for different engineering needs and operating environments from a technical implementation perspective. These alternatives are consistent with the overall concept of this invention and all represent equivalent improvements that can be made by those skilled in the art based on the fundamental concept of this invention.

[0066] At the system architecture level, this invention adopts a three-tiered collaborative structure of "cloud-edge-device" to achieve distributed intelligent diagnostics. However, this structure can be flexibly adjusted according to the task scale and network conditions. In environments with limited resources or unstable communication, a two-tiered "edge-device" structure can be used instead, pushing some cloud functions down to edge servers. In this case, edge nodes not only undertake real-time inference tasks but can also locally perform model fine-tuning and policy updates, thus retaining a certain degree of self-learning capability even when disconnected from the cloud. Conversely, for large-scale clusters or complex operation and maintenance systems, it can also be expanded into a multi-tiered "cloud-edge-edge-device" structure, deploying regional cloud nodes in different geographical locations to build a distributed scheduling network, thereby improving overall computing elasticity and disaster recovery capabilities. Although these alternative structures have different levels, they still maintain the information loop and intelligent optimization logic of cloud-edge collaboration.

[0067] This invention employs a reinforcement learning-driven multi-stage pruning algorithm to achieve adaptive compression and efficient inference of edge-end models. Alternatively, model lightweighting can also be achieved through other algorithms, such as: (1) The Neural Architecture Search (NAS) method is used to automatically search for the optimal model structure under given computing power and latency constraints; (2) Use model quantization technology to convert floating-point parameters into low-bit integers, thereby reducing computational overhead and memory usage; (3) Employ weight sharing and module reconstruction techniques to merge or share parameters of repeated convolutional modules in order to further improve operating efficiency; (4) Using the knowledge distillation direct compression method, a lightweight network is generated by joint distillation of a multi-teacher model without the need for reinforcement learning.

[0068] These methods differ in their lightweight principles, but they can achieve similar results to reinforcement learning pruning algorithms in terms of performance and objectives, and can also reduce edge computing pressure while ensuring diagnostic accuracy.

[0069] The core innovation of this knowledge distillation algorithm, constrained by physical equations, is used to fuse mechanistic models with deep network knowledge. For this part, the physical constraint module can adopt different forms such as energy conservation constraints, dynamic equilibrium constraints, or fluid dynamics equation constraints, with the specific choice flexibly determined according to the characteristics of helicopter components. For systems where it is inconvenient to establish a clear physical model, a data-driven pseudo-physical modeling method can also be used. This method establishes an "empirical physical constraint" function by fitting patterns from historical signals, thereby achieving approximate physical consistency. The teacher model for knowledge distillation can also adopt different architectures, such as CNN-BiLSTM, Transformer, or graph neural network (GNN) based structures, to adapt to different types of multi-source signal data. If higher energy consumption is required in specific application scenarios, distillation can be achieved between the teacher model and the student model through feature pruning or channel alignment networks, thereby reducing the amount of feature transfer. Although these alternative algorithms differ in form, their core idea remains the same: to utilize the knowledge transfer and physical prior fusion of high-precision models to improve the stability and interpretability of lightweight models.

[0070] Regarding cloud scheduling algorithms, this invention employs a self-learning-driven dynamic policy scheduling framework to achieve multi-node collaboration and optimal allocation of task resources. For different system scales and communication conditions, the following alternatives can be adopted: In small-scale or localized network scenarios, the scheduling strategy can be simplified to a heuristic algorithm based on priority queues, such as Min-Latency First or Max-Utilization strategies; in large-scale distributed systems, a hybrid scheduling mechanism combining reinforcement learning and game theory can be introduced to achieve autonomous competition and load balancing among nodes; and for military or emergency flight scenarios with high reliability requirements, the scheduling algorithm can also adopt a hierarchical disaster recovery mechanism based on fault-tolerant replication and task mirroring to ensure that tasks can automatically take over execution even if any node fails. These alternatives differ in algorithmic form and implementation path, but all can operate effectively within the cloud perception-decision-execution framework proposed in this invention, thereby achieving equivalent scheduling optimization effects.

[0071] In summary, the core innovative technologies of the embodiments of the present invention are as follows: 1) Multi-stage model pruning algorithm driven by reinforcement learning In terms of computing power utilization and energy consumption control, this invention significantly improves edge inference efficiency by introducing a multi-stage model pruning algorithm driven by reinforcement learning. Traditional pruning methods rely on manual experience to set pruning ratios, which can easily lead to loss of model accuracy or waste of computing resources. This invention, however, utilizes an intelligent agent to automatically search for the optimal pruning strategy under multi-dimensional constraints, significantly reducing model complexity and computational load while maintaining accuracy. This optimization method based on adaptive structure search enables the diagnostic model to run efficiently on low-power computing platforms, successfully overcoming the technical obstacle of limited airborne hardware resources.

[0072] 2) Knowledge distillation algorithm combining physical equation constraints Regarding model accuracy and physical consistency, this invention constructs an innovative diagnostic framework that integrates data-driven approaches and mechanistic modeling by combining a knowledge distillation algorithm constrained by physical equations. Traditional knowledge distillation often uses the probability distribution output by the teacher model as the learning objective, neglecting the inherent physical laws of the system, leading to model failure under unseen conditions. This invention, however, introduces helicopter dynamics equations, vibration energy conservation constraints, and structural dynamic characteristic constraints into the distillation process, ensuring that the student model, while learning the characteristic distribution of the teacher model, is constrained by real physical mechanisms. This strategy ensures that the model output not only meets data consistency requirements but also conforms to physical interpretability requirements, fundamentally improving the model's generalization performance and stability.

[0073] 3) Dynamic perception and strategy learning In terms of task scheduling and system resource coordination, the cloud-based intelligent scheduling algorithm proposed in this invention significantly improves the resource utilization efficiency of multi-node systems through dynamic perception and policy learning. Existing technologies often employ static allocation or manual rule scheduling, which cannot cope with complex situations such as fluctuations in node computing power, changes in network latency, and uneven task load. The scheduling algorithm of this invention is cloud-centric, collecting node operating status, communication bandwidth, and energy consumption data in real time, and dynamically generating task allocation schemes through a self-learning model. The system can adaptively match and diagnose tasks based on task semantic relevance and node performance, achieving three-dimensional optimization of "task-resource-model". This mechanism effectively avoids task backlog and resource idleness.

[0074] 4) Cloud-edge collaborative loop mechanism Regarding model updates and system self-learning, this invention addresses the issues of delayed model updates and frequent manual intervention in traditional systems by constructing a cloud-edge collaborative loop mechanism. In existing technologies, model optimization typically relies on manual data uploads and centralized retraining, resulting in long cycles and slow responses. This invention establishes an automatic feedback channel between the cloud and the edge, where edge nodes report diagnostic results and performance metrics in real time during operation. The cloud automatically determines whether the model has degraded based on consistency checks and triggers an incremental update process. The cloud utilizes transfer learning and elastic weight retention algorithms for rapid retraining, and the updated model is then distributed to edge nodes in a tiered manner. The entire process is fully automated, reducing the model self-learning cycle from several weeks to several hours, significantly enhancing the system's long-term stability and adaptability.

[0075] In another aspect, the present invention also discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method described above.

[0076] In another aspect, the present invention also discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method described above.

[0077] It is understood that the systems, devices, and storage media provided in the embodiments of the present invention correspond to the methods provided in the embodiments of the present invention, and the explanations, examples, and beneficial effects of the relevant content can be referred to the corresponding parts of the above methods.

[0078] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid state disk (SSD)).

[0079] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0080] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0081] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A cloud-edge collaboration-based lightweight helicopter fault diagnosis method, characterized in that, The following steps are performed by a computer device: S1, in the initial stage of system deployment, a global diagnostic model infrastructure and task scheduling framework is established through the cloud; S2, based on S1, a lightweight neural network model is deployed on the edge computing node; S3, after completing edge model pruning and lightweight optimization, a knowledge distillation algorithm combined with physical equation constraints is proposed, which not only transfers high-level semantic knowledge from the cloud high-precision model (teacher model) to the edge lightweight model (student model), but also introduces physical priors such as helicopter dynamics equation, structural vibration equation and energy conservation constraint to realize the fusion learning of data-driven and mechanism model; S4, cloud scheduling algorithm is used to realize the core hub of "cloud-edge" collaborative intelligent diagnosis system, which not only undertakes resource management and task allocation of multiple nodes, but also is responsible for model updating, computing power coordination, network traffic control and dynamic scheduling of collaborative feedback; S5, after completing the lightweight optimization of edge model, physical constraint distillation and cloud intelligent scheduling, a cloud-edge collaborative loop and online model updating mechanism is proposed to build the self-learning and continuous evolution ability of the system; through this mechanism, the cloud and edge nodes form a dynamic closed loop of data, model and feedback, enabling the system to have intelligent characteristics of self-perception, self-adjustment and self-evolution in long-term operation, thereby realizing continuous improvement of diagnostic accuracy and dynamic optimization of system performance.

2. The cloud-edge collaboration based lightweight helicopter fault diagnosis method according to claim 1, characterized in that, S1 specifically includes: The cloud model uses a Transformer-based fault diagnosis large model to extract features from multi-source sensor data and realize intelligent identification of multiple fault modes; The input data includes vibration signals, rotation speed, temperature, current, voltage, sound waves and flight state parameters; The end-side sensor node acquires raw data through a high-speed acquisition module, and performs feature preprocessing through an edge gateway, including signal denoising, normalization, wavelet packet decomposition, time-frequency spectrum extraction and feature dimension reduction. After being mapped to a standard feature vector by a feature mapping module, it is transmitted to an edge computing unit for real-time inference; The cloud establishes a multi-domain data mapping relationship according to different aircraft types, task scenarios and environmental parameters, and uses incremental learning technology to fuse historical data and real-time data for modeling; Through a sliding sampling strategy based on a time window, new data is included in the online training set, and the global model parameters are dynamically corrected to realize rapid adaptation to different working conditions; At this stage, the system also establishes a cloud-edge communication topology table and a computing resource mapping table; the topology table is used to identify the network bandwidth, transmission delay and communication reliability of each helicopter node; the resource mapping table records the CPU, GPU load rate, storage capacity and available energy consumption upper limit of each edge computing unit, providing a basis for subsequent intelligent scheduling and model deployment.

3. The cloud-edge collaboration based lightweight helicopter fault diagnosis method according to claim 2, characterized in that, S2 specifically includes: Step 21, model structure analysis and pruning feasibility evaluation Before performing pruning operations, the cloud full model structure is first analyzed; the model includes multiple convolutional layers, batch normalization layers, fully connected layers and activation layers, each containing a large number of parameters and feature channels; The cloud analysis module evaluates the pruning feasibility of the model by calculating the following indicators: Channel importance indicator : measure the contribution degree of each channel to the final output through gradient weighted average (Grad-CAM) and parameter sensitivity analysis; redundancy index : statistics of correlation coefficient and mutual information between adjacent channels, high redundancy channel is potential pruning object; structure dependency matrix : for analyzing the dependency degree of different inter-layer features, preventing pruning from destroying information transmission chain On this basis, the cloud establishes a pruning candidate list and a priority index table for each layer and issues the structural metadata to the edge to provide the initial state of the pruning decision for the reinforcement learning agent; Step 22, reinforcement learning modeling of pruning strategy An intelligent agent based on reinforcement learning (RL) is designed to learn and optimize the pruning strategy in a complex pruning space. Firstly, the state space is defined as a vector The state vector is composed of the current layer parameter size, channel number, feature importance index, computational cost (FLOPs), and energy consumption weight. ； After that, define the action space, action space including "reserved", "delete" and "trim" these operations; each layer of the cutting ratio selected discrete value; A reward function based on comprehensive evaluation is designed, which comprehensively considers the factors of accuracy loss , calculation delay and energy consumption ​ ； wherein and is a balancing factor; Step 23, channel-level pruning and local retraining After the reinforcement learning agent determines the optimal pruning ratio of each layer, the edge node enters the actual pruning stage; first, the channel importance is calculated, and through gradient sensitivity analysis, the importance coefficient of each convolution channel is calculated : ； wherein, is a loss function, is a weight for the ijth channel; N is the total number of channels; Then, channel screening and deletion are performed, the least important channels are removed from each layer according to the proportion output by the reinforcement learning, and the BN layer parameters and the input dimension of the adjacent layer are adjusted synchronously to ensure the structural continuity; The pruned model is locally retrained on the edge node, the cache dataset is used to correct the accuracy drop caused by the structural change, the small batch incremental learning method is used for retraining, and the learning rate is adaptively reduced to prevent overfitting; Step 24, hierarchical pruning and cross-layer dependency optimization After completing the channel-level pruning, hierarchical structure optimization is further performed.

4. The cloud-edge collaboration based lightweight helicopter fault diagnosis method according to claim 3, characterized in that: In step S22, the policy gradient-based Actor-Critic architecture is also used in the training mechanism, the Actor generates pruning decisions, and the Critic evaluates the pros and cons of the decisions according to performance feedback; through repeated iterations, the agent gradually converges to the optimal pruning strategy under different constraints, realizing autonomous structure exploration.

5. The cloud-edge collaboration based lightweight helicopter fault diagnosis method according to claim 4, characterized in that, Step S24 specifically includes: Through the cross-layer analysis module, the functional redundant layers, i.e., repeated convolutional layers or low-contribution fully connected layers, are identified in the model, and the hierarchical pruning is performed in combination with the reinforcement learning strategy; Feature map similarity (FMS) and mutual information metrics are used to identify layer structures with high similarity ； wherein, and are the characteristic outputs of i and j layers, respectively, and if the FMS exceeds a threshold value, the layers are considered to be mergable layers; and are the activated feature values, respectively; similar layers are merged or deleted through parameter remapping and residual compensation; if the model accuracy decreases by more than 3% after pruning, the system automatically triggers a rollback mechanism to restore the last optimal structure and adjust the pruning ratio, achieving safe redundancy control.

6. The cloud-edge collaboration based lightweight helicopter fault diagnosis method according to claim 5, characterized in that, Step S3 specifically includes: Step 31, teacher model establishment and multi-source feature extraction In the cloud environment, a high-precision teacher model (Teacher Network) is constructed; the model is based on a deep time convolutional CNN-Transformer hybrid network and is used to comprehensively process multi-source heterogeneous signals; The teacher model realizes cross-domain feature alignment and weighting through a multi-modal fusion block (Multi-modal Fusion Block); after each signal is processed by a feature extraction unit, a time-frequency fusion feature matrix is formed, and then a global feature representation is obtained through a self-attention mechanism module (Self-Attention); the teacher model is trained with a large amount of flight data and has high fault recognition accuracy and the ability to capture nonlinear coupling features; the intermediate layer features, output probability distribution, and gradient response graph will serve as the distillation knowledge source; Step 32, student model design and input mapping structure The student model (Student Network) is deployed on the edge node, and its structure is composed of the lightweight network obtained after pruning optimization; the student model input uses the same standardized feature vector as the teacher model, but an input mapping layer (Input Mapping Layer) is added at the front end of the network to adaptively align data from different sampling frequencies and different sensor sources; Step 33, physical equation modeling and constraint term definition The helicopter dynamics mechanism is introduced into the distillation process, and a physically consistent constraint term is constructed. 1) Kinetic model establishment Based on the simplified dynamic equation of the helicopter transmission system: ； wherein, is the structural vibration displacement vector, M, C, K are the mass, damping and stiffness matrices, respectively, is the external load or fault excitation; the model is obtained by finite element modeling (FEM) or experimental identification of key parameters for describing the vibration response characteristics of the system under different working conditions; 2) Energy conservation and stability constraints, introduce the energy conservation equation ； where, is the input power, is the output power, is the loss; by comparing the difference between the predicted energy spectrum and the measured energy spectrum, the energy constraint loss is defined: ； wherein, is a predicted energy spectrum; is a measured energy spectrum; a physically consistent loss term is established by comprehensively considering two types of constraints of kinetic balance and energy conservation: ； wherein, represents a theoretical response map calculated based on a kinetic equation; Step 34, knowledge transfer and multi-layer feature alignment In the distillation process, the multi-layer alignment mechanism is used to realize knowledge transfer and structure mapping; The knowledge transfer between the teacher and student models includes three types of information: (1) the probability distribution knowledge of the output layer; (2) the feature representation knowledge of the intermediate layer; (3) the gradient direction and the geometric relationship in the feature space; The KL divergence is implemented on each layer output, and the gradient consistency is used to make the student and teacher network direction consistent; Step 35, joint optimization and training strategy of distillation process Combine the loss terms mentioned above to construct the total objective function ； ； ； where, is the standard classification loss for the student model, , , and are tunable weight coefficients; is the soft label KL divergence; is the intermediate layer alignment loss, is the physical constraint loss, is the gradient constraint; is the probability score of the classification softmax; is the linear projection mapping function, is the input feature of the i-th layer.

7. The cloud-edge collaboration based lightweight helicopter fault diagnosis method according to claim 6, characterized in that, Step S4 specifically includes: In the early stage of system operation, the cloud establishes a real-time state perception mechanism through the communication link with all edge nodes, and continuously collects key operating indicators of each node; These indicators include current CPU and GPU load rate, memory usage, network bandwidth, link latency, signal quality, remaining power, model version information, and task queue length; The cloud aggregates and standardizes these information in multiple dimensions to form a node state vector, and builds a resource portrait database based on this; At the same time, the cloud uses a sliding time window mechanism to monitor the trend of node state changes, identify potential performance fluctuations and network bottlenecks, and provide early warning basis for advance adjustment of scheduling strategy; Based on the global state, the cloud scheduling module enters the decision-making stage; A dynamic scheduling algorithm framework based on intelligent strategy optimization is introduced, and the scheduling process is regarded as a multi-objective decision problem; The cloud adopts a multi-layer decision structure: the upper layer is the global strategy planning layer, which is responsible for determining the overall task distribution mode and model version selection; The lower layer is the local scheduling execution layer, which is responsible for specifically allocating tasks, determining the execution order and parameter configuration within each node range; The cloud scheduling system trains a set of scheduling strategy model based on historical task execution records, node performance curves and network topology; This model uses a self-learning structure to continuously optimize its parameters during operation, making the strategy selection process intelligent; At the same time, for high-priority or safety-critical flight tasks, the cloud will allocate redundant diagnostic nodes for double or multiple verification to ensure the reliability and timeliness of the results; In the task allocation process, the cloud scheduling algorithm not only considers the computing power and delay factors, but also comprehensively introduces the concept of "task semantic relevance"; Different flight tasks or fault types have different dependencies on data features and model versions, and the cloud maps tasks according to the matching degree of task feature vectors and model capabilities of each node; In the task execution stage, the cloud monitors the task progress and result feedback of each node in real time through the asynchronous control channel; The scheduling module has a dynamic adjustment mechanism, and once it detects node execution abnormalities, the system will immediately trigger the adaptive scheduling process; The cloud scheduling algorithm also introduces an energy consumption perception and bandwidth adaptive control mechanism; the system comprehensively evaluates the data transmission volume, computing demand and node energy state of the task at each task allocation, and realizes optimal scheduling in the available communication resources through a bandwidth dynamic allocation algorithm; when the flight task is in a good communication condition stage, the system tends to assign complex tasks to edge nodes to fully utilize local computing power; when the link is unstable or energy consumption is limited, the task is moved up to the cloud or adjacent nodes for execution to reduce energy consumption and transmission overhead; At the algorithm implementation level, the cloud scheduling system adopts a modular architecture design, consisting of a state perception module, a decision engine module, an execution management module and a feedback optimization module; the state perception module is responsible for real-time collection of system information and updating of resource state; the decision engine module generates an optimal task allocation scheme based on a strategy model; the execution management module is responsible for task assignment, node synchronization and execution monitoring; the feedback optimization module updates the strategy model based on actual operation results to realize continuous learning and adaptive optimization; the modules communicate efficiently through a message queue and shared memory mechanism to ensure that the system completes the scheduling closed loop within a millisecond-level response time.

8. The cloud-edge collaboration based lightweight helicopter fault diagnosis method according to claim 7, characterized in that, The step S5 specifically comprises: The cloud is responsible for gathering multi-node operation data and performing centralized training and evaluation as the global control and knowledge center; the edge node is responsible for real-time inference, local learning and feedback reporting as the front-line intelligent agent; the two continuously interact through a high-reliability communication link to form a closed-loop process of "collection-diagnosis-feedback-relearning-redeployment"; After receiving feedback data from multiple edge nodes, the cloud enters the global analysis stage; the system compares the differences between edge prediction results and cloud master model prediction results through a consistency detection module to calculate a consistency degree index; when the consistency degree is below a threshold, it indicates that the model performance has degraded or the environment has changed, and the cloud will start the online update process; at the same time, the cloud also identifies new fault samples or rare patterns through clustering and anomaly detection algorithms and adds them to the knowledge base to enrich the training set; During the update process, the cloud iteratively trains the model by combining incremental learning and transfer learning; incremental learning is used to absorb new data without forgetting old knowledge, and transfer learning improves the adaptability of new tasks through parameter sharing and feature mapping; to prevent catastrophic forgetting, the system introduces an elastic weight maintenance mechanism to limit the change of key parameters, thereby ensuring the stability of the original diagnostic capability; after the updated model is verified to achieve stable convergence, the cloud marks it as a candidate version and enters the deployment stage.

9. A computer readable storage medium storing a computer program, characterized in that: The computer program, when executed by a processor, causes the processor to perform the steps of the method of any one of claims 1 to 8.

10. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that: The computer program, when executed by the processor, causes the processor to perform the steps of the method of any one of claims 1 to 8.

Images