Engineering machinery distributed edge computing system based on multi-heterogeneous teacher distillation method

Abstract

An engineering machinery distributed edge computing system based on a multi-heterogeneous teacher distillation method comprises a cloud training layer, an edge computing layer, a dynamic scheduling layer and a cross-layer asynchronous data processing module. The problems that in heterogeneous asynchronous data real-time processing of an engineering machinery group under the complex working condition that computing resources are limited, the parameter quantity of a large model is large, deployment is difficult, and the dynamic adaptability of edge node computing power and communication resources is poor are solved, efficient lightweight reasoning and cross-device computing load balancing of multi-modal data are achieved, and the method is suitable for engineering machinery group heterogeneous asynchronous data real-time processing. The parameter quantity and the calculation complexity are obviously reduced; a distributed deployment framework matched with the physical structure of the engineering machinery is constructed, and a lightweight reasoning module is deployed at a key part of equipment, so that near-source processing of multi-modal data is realized; task migration and load balancing scheduling are performed based on a real-time computing power state through a dynamic cooperation mechanism between edge nodes, and the problem of low resource utilization rate of heterogeneous nodes is solved. Therefore, a collaborative system with lightweight models, distributed deployment and flexible resources is formed, and the real-time decision-making capability and the system energy efficiency of the engineering machinery group under complex working conditions are integrally improved.

Description

Technical Field

[0001] The present invention relates to a technology in the field of construction machinery, and particularly to a distributed edge computing system for construction machinery based on a multi-heterogeneous teacher distillation method. Background Art

[0002] In the field of construction machinery intelligence, complex construction machinery generally uses multi-modal sensors for real-time working condition monitoring. In the prior art, deep learning models are usually deployed on the cloud or a central server to process heterogeneous time-series data such as hydraulic systems and boom stresses, resulting in a huge number of model parameters and difficulty in deploying on vehicle-mounted edge devices. Although a distributed edge computing network can decompose computing tasks to different nodes, there are spatio-temporal alignment differences in the asynchronous heterogeneous data generated by sensors at various parts of construction machinery, and there are significant differences in the computing power of edge nodes. Existing single model compression methods are difficult to adapt to different computing units, resulting in increased inference latency of local nodes and a decrease in the overall system energy efficiency ratio. Summary of the Invention

[0003] In view of the above deficiencies in the prior art, the present invention proposes a distributed edge computing system for construction machinery based on a multi-heterogeneous teacher distillation method. By integrating multi-source heterogeneous model compression and distributed edge collaborative computing technologies, it solves the problems of large model parameters and difficult deployment in the real-time processing of heterogeneous asynchronous data under complex working conditions with limited computing resources in a construction machinery group, and poor dynamic adaptability of the computing power and communication resources of edge nodes. It realizes efficient lightweight inference of multi-modal data and cross-device computing load balancing, significantly reducing the number of parameters and computing complexity; by constructing a distributed deployment framework matching the physical structure of construction machinery and deploying lightweight inference modules at key parts of the equipment, it realizes near-source processing of multi-modal data; through a dynamic cooperation mechanism between edge nodes, based on the real-time computing power state, task migration and load balancing scheduling are carried out to solve the problem of low resource utilization rate of heterogeneous nodes. Thus, a collaborative system with lightweight models, distributed deployment, and elastic resources is formed, overall improving the real-time decision-making ability and system energy efficiency of the construction machinery group under complex working conditions.

[0004] The present invention is realized through the following technical solutions:

[0005] The present invention relates to a construction machinery distributed edge computing system based on a multi - heterogeneous teacher distillation method, including: a cloud training layer, an edge computing layer, a dynamic scheduling layer, and a cross - layer asynchronous data processing module, where: The cloud training layer integrates visual, mechanical, and vibration heterogeneous data collected by multi - modal sensors of construction machinery, constructs a multi - teacher model cluster for collaborative training and knowledge distillation processing, generates a set of lightweight sub - model parameters adapted to edge nodes, and generates a student model with low memory occupancy through feature mapping distillation; The edge computing layer dynamically configures an inference model deployment plan according to the computing power characteristics of physical nodes, matches lightweight models at positions such as the embedded device at the end of the boom and the hydraulic control unit, and combines a mixed - precision inference engine to achieve millisecond - level real - time response; The dynamic scheduling layer calculates the load weights of GPU utilization, communication latency, and memory occupancy through fusion, dynamically plans the task migration path and resource allocation strategy, and combines a hybrid communication protocol to ensure job continuity; The cross - layer asynchronous data processing module uses a spatio - temporal alignment compensation algorithm to eliminate the clock deviation of multi - source sensors, and realizes millisecond - level synchronization of hook visual data and mechanical load signals through dynamic adjustment of a sliding window, supporting real - time decision - making of the crane lifting path.

[0006] The present invention also relates to a method for implementing construction machinery distributed edge computing based on the above - mentioned system, including:

[0007] Step 1, construct multi - modal feature partitioning and heterogeneous teacher models: Divide multi - modal features according to the construction machinery operation scenario, including heterogeneous data types such as visual perception data, mechanical parameters, and vibration signals, and train heterogeneous teacher models for different modal characteristics in the cloud respectively to form an expert model cluster;

[0008] Step 2, hierarchical knowledge distillation and model compression: Adopt a hierarchical knowledge distillation strategy to migrate the teacher model to the edge side;

[0009] The knowledge distillation strategy consists of global feature distillation and local feature distillation. Global feature distillation extracts common knowledge by aligning the output distributions of multi - teacher models, and local feature distillation generates a lightweight sub - model based on a dynamic adaptation module, where the dynamic adaptation module dynamically adjusts the feature migration intensity of the intermediate layer according to the computing power of the edge node;

[0010] Step 3, distributed deployment of lightweight models: Deploy lightweight models to core computing nodes according to the construction machinery physical network. For example, the visual inference module is deployed on the mobile perception unit, and the mechanical analysis module is deployed on the edge device on the power unit side, etc., and data interaction between nodes is realized through a priority - scheduling communication bus. The priority - scheduling communication bus allocates transmission bandwidth in real time according to the data type;

[0011] Step 4, Edge Node Dynamic Load Balancing: Based on the task migration strategy of device status awareness, the load weights of each node are monitored in real time. When the load of the sensing unit node exceeds the limit, part of the computing tasks are allocated to the idle computing nodes through the task migration strategy. The load weight is dynamically generated by comprehensively considering the GPU utilization rate and communication delay.

[0012] Step 5, Asynchronous Data Collaborative Inference: The problem of sensor clock deviation is solved through the spatio-temporal alignment compensation module, and the multi-source data stream synchronization is realized by combining the sliding window mechanism with an adaptively adjusted window length. The spatio-temporal alignment compensation module uses a linear interpolation algorithm to compensate for the sensor sampling time difference, and the sliding window mechanism automatically matches the timing alignment range according to the data stream frequency. Technical Effects

[0013] Through the construction of a multi-teacher model cluster for construction machinery through multi-modal feature collaborative partitioning, the present invention breaks through the limitation of the insufficient adaptability of traditional single-teacher distillation to heterogeneous data; designs a hierarchical dynamic distillation network, combines the global feature alignment and local adaptive migration mechanisms, and realizes the precise matching of the model compression ratio and the computing power of edge nodes through the dynamic adaptation module; based on the lightweight model deployment framework with physical topology constraints, uses the priority scheduling bus and spatio-temporal alignment compensation module to synchronously solve the spatio-temporal consistency problems of near-source data processing and asynchronous data collaborative inference; develops a dynamic task migration strategy driven by load weight prediction, and collaboratively determines the task allocation path through bandwidth awareness and computing power evaluation to solve the problem of systematic load imbalance caused by the computing power fluctuation of edge nodes, forming a closed-loop technology system of model lightweighting, deployment elasticity, and resource efficiency. The present invention significantly improves the modal adaptability of heterogeneous data of construction machinery through the construction of a multi-teacher model cluster, effectively overcomes the limitation of the representation ability of a single model for multi-source asynchronous data, and at the same time, the hierarchical dynamic distillation network maintains the consistency of inference accuracy among edge nodes while realizing model lightweighting, enabling the compressed model to precisely adapt to computing units with different computing power levels; combined with the deployment framework based on physical topology constraints, the consumption of the original data transmission bandwidth is reduced through near-source data processing, and the reliability of multi-modal data fusion is improved by means of the spatio-temporal consistency guarantee mechanism; further, the dynamic task migration strategy senses the node load and communication status in real time, realizes the efficient scheduling and elastic expansion of edge computing resources, thereby maintaining the overall energy efficiency balance of the system under complex working conditions. Finally, a technical closed-loop of collaborative optimization of model compression, distributed deployment, and resource scheduling is formed, significantly improving the response speed and decision-making accuracy of construction machinery groups in scenarios such as real-time monitoring and fault warning, and providing high-robustness edge computing support for the intelligent management of engineering equipment. Brief Description of the Drawings

[0014] Figure 1 It is a flowchart of the present invention;

[0015] Figure 2This is a schematic diagram of the main topology module structure of the system of the present invention. Detailed implementation manners

[0016] As Figure 1 and Figure 2 shown, this embodiment relates to a method for implementing distributed edge computing of construction machinery based on multi - heterogeneous teacher knowledge distillation, including:

[0017] Step 1, multi - modal feature partitioning and heterogeneous teacher model construction, specifically including:

[0018] 1.1. Multi - modal feature partitioning and data collection: According to the crane hoisting operation requirements, a visual perception module is deployed at the end of the boom to obtain hook positioning data, a mechanical parameter sensor is installed in the hydraulic system to collect oil pressure changes, and a vibration monitoring unit is configured at the slewing bearing part; each sensor performs spatio - temporal alignment pre - processing through an edge gateway to generate asynchronous data streams of vision, mechanics, and vibration with unified timestamps, and uploads them to the cloud feature partitioning engine.

[0019] 1.2. Multi - modal feature extraction and annotation: The cloud feature partitioning engine performs modal feature extraction on the received asynchronous data streams. The vision data extracts hook contour features and spatial coordinates, the mechanical data extracts pressure gradient change features, and the vibration data extracts frequency - domain energy distribution features, generates multi - modal feature vectors and adds modal labels.

[0020] 1.3. Cloud heterogeneous teacher model training: Based on the multi - modal feature vectors, a ResNet network as a vision analysis teacher model for image recognition, an LSTM network as a mechanical reasoning teacher model for analyzing time - series pressure features, and a 1D - CNN network as a vibration diagnosis teacher model for parsing frequency - domain features are respectively constructed on the cloud training platform to form an expert model cluster for the crane operation scenario.

[0021] 1.4. Model verification and knowledge base construction: Cross - verify the trained teacher models. The vision model verifies the error range of hook positioning accuracy, the mechanical model verifies the timeliness of pressure anomaly prediction, and the vibration model verifies the accuracy of health state classification, and stores the verified model weights and intermediate - layer features output to the cloud knowledge base.

[0022] 1.5. Establishment of multi - modal data - model mapping relationship: According to the physical logic of the crane operation scenario, bind the vision data stream to the vision analysis teacher model, associate the mechanical data stream with the mechanical reasoning teacher model, and correspond the vibration data stream to the vibration diagnosis teacher model to establish a dynamic mapping table between modal features and teacher models.

[0023] Step 2, adopt a hierarchical knowledge distillation strategy to migrate the teacher model to the edge side, specifically including:

[0024] 2.1. Construction of the global feature distillation strategy: Based on the output layer features of the cloud multi-teacher model cluster, align the common knowledge distributions of the vision, mechanics, and vibration models through the KL divergence loss function, and extract the cross-modal global features of the crane operation scenario;

[0025] 2.2. Dynamically transfer the local features. In view of the computing power differences of edge nodes, design a dynamic adaptation module to adaptively adjust the transfer intensity of the intermediate layer features of the teacher model according to the node hardware performance (such as the TOPS value of the in-vehicle GPU). Among them, the high transfer intensity (retaining 80% of the intermediate layer features) is adopted for the boom vision node, and the medium transfer intensity (retaining 60% of the features) is adopted for the chassis mechanics node;

[0026] 2.3. Generation of the lightweight student model: Integrate the global feature distillation results with the local dynamic transfer features to generate a lightweight model exclusive to the computing nodes of each physical part of the crane. The number of parameters of the vision inference model is significantly reduced to less than 1 / 4 of the original teacher model, and the number of parameters of the mechanics analysis model is reduced to less than 1 / 3 of the original model;

[0027] 2.4. Verification of the edge-side model: Deploy the lightweight model on the in-vehicle edge device, optimize the model adaptability through online incremental learning, and verify that the positioning accuracy error of the hook is significantly reduced compared with the original teacher model, and the detection delay of hydraulic pressure abnormality is significantly reduced compared with the cloud centralized computing, meeting the real-time operation requirements of the crane;

[0028] 2.5. Construction of the model-node mapping table: According to the physical topology structure of the crane, establish a binding relationship table between the lightweight model and the edge nodes, and define that the vision model is deployed on the boom end node, the mechanics model is deployed on the hydraulic control node, and the vibration model is deployed on the slewing bearing node.

[0029] Step 3. Distribute and deploy the generated lightweight model, specifically including:

[0030] 3.1. Physical topology matching deployment and communication protocol configuration: Based on the mechanical structure characteristics of the crane, deploy the lightweight model generated in Step 2 to the corresponding physical units, including deploying the vision inference model on the boom end embedded device, the mechanics analysis model on the hydraulic control unit node, and the vibration diagnosis model on the slewing bearing monitoring node; Synchronously establish a hierarchical communication bus, configure the node to define the three-level data transmission priority according to the flowchart communication protocol, and ensure the transmission of key tasks through dynamic bandwidth allocation;

[0031] 3.2 Real - time data stream processing and redundancy fault tolerance: Each edge node receives the real - time data stream of local sensors. The vision node processes the hook positioning image, the mechanics node analyzes the temporal change of hydraulic pressure, the vibration node analyzes the health status signal, and the inference results are synchronized to the central decision - making module through the bus. At the same time, a lightweight backup model is pre - loaded in the hydraulic control unit and the cab terminal, and it automatically switches when the main model is detected to be abnormal, ensuring the continuous operation of the system;

[0032] 3.3 Deployment verification and dynamic optimization: Verify the performance of the communication bus through pressure testing (corresponding to the verification branch in Figure 1 step 4), ensure that the highest - level data packet loss rate is lower than the domain standard value, and the bus utilization rate under multi - node concurrent tasks meets the preset threshold. Combine the flowchart - based dynamic load - balancing feedback mechanism to dynamically adjust the node resource allocation when the crane operation mode switches (such as waking up the chassis positioning node and loading the SLAM model under the condition of a transfer operation mode), and achieve continuous optimization of the system energy efficiency ratio.

[0033] Step 4 Edge node dynamic load balancing. Design a task migration strategy based on device - status awareness, specifically including:

[0034] 4.1 Real - time monitoring of node load status: Deploy load - monitoring agents at each edge - computing node of the crane to collect GPU utilization rate, memory occupancy rate, and communication delay metrics in real - time. The sampling period of the vision node is set to 100 ms, the sampling period of the mechanics node is set to 200 ms, and the sampling period of the vibration node is set to 500 ms. The status data is aggregated to the central scheduling module through the priority - scheduling bus;

[0035] 4.2 Calculation of dynamic load weights: The central scheduling module uses a weighted - fusion algorithm to generate the load weights of each node. The calculation formula is: load weight = 0.6×GPU utilization rate + 0.3×communication delay coefficient + 0.1×memory occupancy rate, where the communication delay coefficient is dynamically normalized according to the ratio of the bus transmission delay to the preset threshold. When the weight of the vision node exceeds the set threshold, task migration is triggered;

[0036] 4.3 Task migration decision - making and path planning: Based on the load - weight prediction model, predict the node overload risk. When the predicted weight of the boom vision node exceeds the threshold for multiple consecutive cycles, start the migration strategy. First, migrate the image pre - processing task to the idle chassis positioning node, including image down - sampling and ROI extraction. Second, select the migration path through the bandwidth - aware algorithm, and preferentially use the dedicated control bus to transmit the migration task context data;

[0037] 4.4 Seamless connection of migrated tasks: Dynamically load a lightweight model copy at the target node, use the checkpoint mechanism to save the intermediate state data of the source node, and achieve task switching through incremental synchronization to ensure that the interruption time of the hydraulic pressure analysis task meets the real - time requirements;

[0038] 4.5. Verification of load balancing effect: Tested under the combined working conditions of crane hoisting and transfer. When the load of the vision node reaches the peak, after migrating some preprocessing tasks, its weight decreases significantly, the utilization rate of the chassis node is increased to a reasonable range, and the overall inference delay of the system is significantly reduced.

[0039] Step 5. Asynchronous data collaborative inference. Solve the sensor clock deviation problem through the space-time alignment compensation module, specifically including:

[0040] 5.1. Multi-source data clock synchronization: Deploy the PTP precision clock protocol in the crane sensor network to synchronize the local clocks of vision, mechanics, and vibration sensors at the microsecond level. The residual clock deviation is processed by the space-time alignment compensation module;

[0041] 5.2. Implementation of space-time alignment compensation algorithm: For the hook vision coordinates and hydraulic pressure data that arrive asynchronously, use the linear interpolation algorithm to generate the estimated value of the pressure data at the image sampling moment.

[0042] 5.3. Dynamic adjustment of sliding window: According to the frequency difference between the vibration signal frequency and the vision data frequency, set the adaptive sliding window length. The vibration data within the window is aligned with the vision features in the frequency domain after FFT transformation;

[0043] 5.4. Multi-modal data fusion inference: Input the aligned vision positioning data, interpolated pressure data, and vibration spectrum features into the central decision-making module, and use the attention mechanism to weight and fuse multi-modal features to output the hoisting safety assessment result;

[0044] 5.5. Verification of collaborative inference effect: Tests show that space-time alignment compensation significantly improves the accuracy of multi-modal data fusion, the sliding window mechanism effectively reduces the processing delay of time-series data, and the detection response time of abnormal working conditions of the crane meets the requirements of real-time safety monitoring.

[0045] As Figure 2 shown, the present invention proposes a collaborative computing system for construction machinery equipment, which adopts a four-layer three-domain collaborative network, including a cloud training layer, an edge computing layer, a dynamic scheduling layer, and a cross-layer asynchronous data processing module. Each layer realizes deep collaboration through a special protocol stack for construction machinery. The cloud training layer constructs a multi-modal federated learning framework for heavy equipment such as cranes and concrete pumps. The edge computing layer deploys heterogeneous computing units that match the mechanical physical structure. The dynamic scheduling layer implements dynamic resource orchestration based on the digital twin model of the equipment working conditions. The asynchronous data processing module establishes a cross-layer channel to support real-time interaction of multi-domain data such as hydraulic systems and power units, forming a unique algorithm-computing power-data closed-loop optimization system for construction machinery. Taking the collaborative operation of a crane as an example, the system can realize real-time collaborative computing of functions such as hoisting parameter analysis, hydraulic state monitoring, and multi-crane collision avoidance decision-making.

[0046] The cloud training layer described above serves as the intelligent center of the system. It adopts a distributed learning mechanism enhanced by differential privacy, generates a lightweight knowledge model through multi-modal feature distillation technology, and significantly reduces the communication load of edge nodes by combining the self-developed adaptive gradient compression algorithm. Its version control module supports incremental hot updates of the model, ensuring the spatio-temporal consistency between algorithm iteration and hardware deployment, and solving the service interruption problem caused by model updates in traditional networks.

[0047] In the application in the construction machinery field, the cloud training layer described above is deeply adapted to the crane operation scenario: it strengthens the privacy protection mechanism of distributed learning through encryption of hydraulic system pressure characteristics, fuses multi-modal sensing data such as boom movement trajectory and load swing characteristics for feature distillation, and generates a lightweight safety warning model adapted to the embedded control unit. The gradient compression strategy optimized for the hoisting scenario significantly reduces the communication load, enables the model update cycle to be accurately matched with the crane transfer and maintenance window, and at the same time realizes version locking under high-risk working conditions based on the operation mode recognition algorithm, ensuring the continuity of hoisting operations and the stability of the system, and achieving seamless coordination between edge intelligence upgrade and construction machinery operation and maintenance processes.

[0048] The edge computing layer described above deploys a heterogeneous computing node cluster, designs a mixed-precision inference engine to prioritize time-sensitive tasks, dynamically allocates CPU / GPU / FPGA computing power resources through the hardware abstraction layer, executes local data filtering based on spatio-temporal constraints to greatly reduce the cloud transmission load, and combines an intelligent caching mechanism to enhance the autonomous decision-making ability of edge nodes, showing superior real-time response characteristics in complex environmental scenarios.

[0049] In the application in the construction machinery field, the edge computing layer described above is optimized in real time for crane hoisting operations: a mixed-precision inference engine is deployed at the embedded node at the end of the boom, and the FPGA computing power is dynamically allocated through the hardware abstraction layer to analyze lidar point cloud data, and the GPU resources are used to process the hydraulic pressure waveform, realizing millisecond-level fusion calculation of the load swing trajectory and the leg load-bearing state. Execute local data filtering based on the hoisting safety threshold, eliminate non-critical sensing information such as wind speed interference to reduce the cloud transmission load, and combine the edge intelligent caching mechanism of the slewing mechanism vibration characteristics to still maintain sub-second safety warning response in a strong electromagnetic interference environment, ensuring the continuity of hoisting operations under complex working conditions.

[0050] The dynamic scheduling layer described above constructs a digital twin scheduling system for end-edge-cloud collaboration, anticipates the system resource status through computing power profiling modeling and load prediction algorithms, adopts a multi-objective optimization strategy to generate a globally optimal task allocation plan, realizes elastic scaling and energy efficiency balance of three-level computing resources, and can significantly improve the task throughput and scheduling stability in high-concurrency scenarios through simulation verification.

[0051] After specific actual experiments, in the standardized test environment of cranes covering heavy-duty hoisting, multi-crane collaborative transfer, and sudden disturbance composite working conditions, the full process verification was carried out by deploying the system of the present invention. The results show that: compared with the traditional cloud centralized processing method, this embodiment has achieved significant improvements in core indicators such as model inference efficiency, multi-modal data fusion accuracy, and edge node resource utilization rate. Specifically, the inference latency of the edge lightweight model for hook visual positioning has been significantly reduced by about 60%, the cross-modal fusion error rate of the hydraulic system pressure time series characteristics and the boom vibration frequency domain energy distribution characteristics has decreased by more than 50% through the spatio-temporal alignment compensation module, and the dynamic load balancing efficiency of the slewing mechanism and the chassis power unit calculation nodes has been increased by more than 30%. Verified by standard benchmark test tools, the parameter order of the lightweight model has been compressed to the range of one-fifth to one-fourth of the original model scale, the spatio-temporal alignment compensation module has improved the time synchronization accuracy of multi-source sensor data to the millisecond level, and the dynamic task migration strategy has maintained the task switching interruption time within the range of hundreds of milliseconds during the peak load of the hydraulic system. Due to the core technology confidentiality agreement, the detailed experimental parameters and deployment configurations are not publicly available for the time being, but the test data fully prove that this embodiment can be widely applied to the intelligent transformation of complex construction machinery such as hydraulic machinery and lifting equipment, providing an edge computing solution with high real-time performance, strong robustness, and elastic expansion ability for the construction machinery field.

[0052] In the application in the construction machinery field, the dynamic scheduling layer deeply optimizes the collaborative operation of the crane fleet: constructs a digital twin of the crane based on multi-dimensional parameters such as hoisting moment, outrigger load-bearing status, and environmental wind speed, predicts the fluctuation of the computing power requirements of visual positioning and hydraulic diagnosis tasks through a load prediction algorithm, and dynamically optimizes the edge node resource allocation strategy. When a high-risk hoisting mode is detected, a multi-objective optimization algorithm is used to migrate non-core computing tasks to the cloud inference cluster, realizing elastic resource scheduling and millisecond-level response to core safety instructions, ensuring the safety and system stability of multi-crane collaborative hoisting operations in strong wind disturbance or complex space constraint scenarios, and ensuring the efficient execution of the full process of heavy component hoisting.

[0053] The asynchronous data processing module designs a cross-layer direct connection channel, constructs a circular data buffer based on a high-speed transmission protocol to eliminate hierarchical interaction latency, ensures the quality of service of critical tasks through data flow physical isolation technology, and integrates an anomaly detection mechanism to achieve dynamic fusing and self-healing recovery of computing tasks, forming a bottom-layer data path to support the reliable operation of the entire system.

[0054] In the application in the field of construction machinery, the asynchronous data processing module is customized for the multi-source data streams of cranes: a cross-layer direct connection channel for hoisting control instructions is established through corresponding protocols, enabling the emergency stop signal to bypass the conventional data processing link and directly reach the actuator, constructing a circular buffer queue for the hydraulic system pressure wave and the load swing data to eliminate the decision-making delay caused by the time sequence misalignment of multi-sensors. Based on physical isolation technology, the data stream of the safety protection system is separated from the conventional monitoring channel. When abnormal leg pressure or excessive wire rope swing is detected, non-core fault log analysis tasks are dynamically fused, triggering the self-healing mechanism to preferentially restore the torque balance calculation resources, and still maintaining the millisecond-level stable transmission of key control instructions in a strong vibration interference environment, forming a reliable data center to support the high-safety-level operation of cranes.

[0055] Through the design of hierarchical decoupling and cross-layer collaboration, compared with traditional edge computing networks, the present invention achieves breakthrough improvements in key indicators such as model communication efficiency, task response real-time performance, and resource scheduling flexibility, effectively solving technical problems such as federated learning data heterogeneity, edge node resource fragmentation, and contention for time-delay sensitive tasks, and providing a new computing network solution with high reliability, low latency, and easy expansion for the field of intelligent Internet of Things.

[0056] Compared with the technical defects of traditional construction machinery edge computing systems, such as difficulties in multi-modal data fusion, insufficient model generalization ability, and lagging real-time control response, the present invention achieves a breakthrough through the collaborative mechanism of multi-heterogeneous teacher distillation and dynamic resource scheduling. The prior art uses a single model to process multi-source heterogeneous sensing data, resulting in insufficient feature extraction and the risk of model overfitting. The present invention divides multi-modal features such as vision, mechanics, and vibration, etc., and constructs a dedicated teacher model cluster, significantly improving the feature representation ability under complex working conditions; the traditional static model deployment method is difficult to adapt to the computing power differences of edge nodes. The hierarchical knowledge distillation strategy of the present invention realizes the precise matching of the feature migration intensity and the computing unit ability through a dynamic adaptation module, greatly optimizing the communication efficiency while ensuring the model inference accuracy; aiming at the defect that the existing communication bus cannot distinguish data priorities, the three-level priority scheduling mechanism of this embodiment effectively guarantees the real-time transmission reliability of safety control instructions, while maintaining the collaborative computing stability among multiple nodes. These technical improvements are verified in typical construction machinery operation scenarios, proving that they can significantly improve the edge computing efficiency of intelligent equipment and provide a high-reliability and strong-real-time computing network support for the autonomous operation of heavy machinery.

[0057] The above specific implementation can be locally adjusted in different ways by those skilled in the art without departing from the principles and purposes of the present invention. The protection scope of the present invention is subject to the claims and is not limited by the above specific implementation, and all implementation solutions within its scope are subject to the constraints of the present invention.

Claims

1. A distributed edge computing system for construction machinery based on a multi-heterogeneous teacher distillation method, characterized in that: include: Cloud training layer, edge computing layer, dynamic scheduling layer and cross-layer asynchronous data processing module, among which: the cloud training layer integrates the visual, mechanical and vibration heterogeneous data collected by multi-modal sensors of engineering machinery, builds a multi-teacher model cluster for collaborative training and knowledge distillation processing, generates a lightweight sub-model parameter set adapted to the edge node, and generates a student model with low memory usage through feature mapping distillation; the edge computing layer dynamically configures the inference model deployment plan according to the computing power characteristics of the physical node, matches lightweight models in the embedded devices at the end of the boom and the hydraulic control unit, and combines the mixed precision inference engine to achieve millisecond-level real-time response; the dynamic scheduling layer dynamically plans the task migration path and resource allocation strategy by integrating the load weight calculation of GPU utilization, communication delay and memory usage, and combines the hybrid communication protocol to ensure the continuity of the operation; The cross-layer asynchronous data processing module adopts a spatiotemporal alignment compensation algorithm to eliminate the clock deviation of multi-source sensors, and achieves millisecond-level synchronization of the hook visual data and the mechanical load signal through dynamic adjustment of the sliding window, supporting real-time decision-making of the crane lifting path.

2. The distributed edge computing system for construction machinery based on the multi-heterogeneous teacher distillation method according to claim 1 is characterized in that: The cloud-based training layer serves as the intelligent core of the system and adopts a distributed learning mechanism enhanced by differential privacy. It generates lightweight knowledge models through multimodal feature distillation technology and significantly reduces the communication load of edge nodes by combining the independently developed adaptive gradient compression algorithm. Its version control module supports incremental hot updates of models, ensuring the spatiotemporal consistency of algorithm iteration and hardware deployment, and solving the problem of service interruptions caused by model updates in traditional networks.

3. The distributed edge computing system for construction machinery based on the multi-heterogeneous teacher distillation method according to claim 1 is characterized in that: The edge computing layer deploys a cluster of heterogeneous computing nodes, designs a mixed-precision inference engine to prioritize time-sensitive tasks, dynamically allocates CPU / GPU / FPGA computing resources through the hardware abstraction layer, performs localized data filtering based on time and space constraints to significantly reduce cloud transmission load, and combines an intelligent caching mechanism to enhance the autonomous decision-making capabilities of edge nodes, demonstrating superior real-time response characteristics in complex environmental scenarios.

4. The distributed edge computing system for construction machinery based on the multi-heterogeneous teacher distillation method according to claim 1 is characterized in that: The dynamic scheduling layer builds a digital twin scheduling system for end-edge-cloud collaboration, predicts the system resource status through computing power portrait modeling and load prediction algorithm, and adopts a multi-objective optimization strategy to generate a global optimal task allocation plan, thereby achieving elastic scaling and energy efficiency balance of three-level computing resources. Simulation verification shows that it can significantly improve task throughput and scheduling stability in high-concurrency scenarios.

5. A method for implementing distributed edge computing of construction machinery based on the system described in any one of claims 1 to 4, characterized in that: include: Step 1: Construct multimodal feature division and heterogeneous teacher model: divide multimodal features according to the operation scene of engineering machinery, including heterogeneous data types such as visual perception data, mechanical parameters and vibration signals, and train heterogeneous teacher models for different modal characteristics in the cloud to form an expert model cluster; Step 2: Hierarchical knowledge distillation and model compression: Use hierarchical knowledge distillation strategy to migrate the teacher model to the edge; The knowledge distillation strategy consists of global feature distillation and local feature distillation. The global feature distillation extracts common knowledge by aligning the output distribution of multiple teacher models. The local feature distillation generates a lightweight sub-model based on a dynamic adaptation module, where the dynamic adaptation module dynamically adjusts the migration strength of the middle layer features according to the computing power of the edge nodes. Step 3: Distributed deployment of lightweight models: Deploy lightweight models to core computing nodes based on the physical network of engineering machinery, such as visual reasoning modules deployed on mobile perception units, mechanical analysis modules deployed on edge devices on the power unit side, etc., and realize data interaction between nodes through a priority scheduling communication bus, which allocates transmission bandwidth in real time according to data type; Step 4: Dynamic load balancing of edge nodes: Based on the task migration strategy of device status perception, the load weight of each node is monitored in real time. When the load of the sensing unit node exceeds the limit, part of the computing tasks are allocated to the idle computing node through the task migration strategy. The load weight is dynamically generated by combining GPU utilization and communication delay. Step 5, asynchronous data collaborative reasoning: The problem of sensor clock deviation is solved by the spatiotemporal alignment compensation module, and the sliding window mechanism that adaptively adjusts the window length is combined to achieve multi-source data stream synchronization. The spatiotemporal alignment compensation module uses a linear interpolation algorithm to compensate for the sensor sampling time difference, and the sliding window mechanism automatically matches the timing alignment range according to the data stream frequency.

6. The method for implementing distributed edge computing of construction machinery according to claim 5 is characterized in that: The step 1 specifically includes: 1.

1. Multimodal feature segmentation and data collection: According to the requirements of crane hoisting operations, a visual perception module is deployed at the end of the boom to obtain hook positioning data, a mechanical parameter sensor is installed in the hydraulic system to collect oil pressure changes, and a vibration monitoring unit is configured at the slewing bearing. Each sensor is pre-processed for spatiotemporal alignment through the edge gateway to generate visual, mechanical, and vibration asynchronous data streams with unified timestamps, which are uploaded to the cloud feature segmentation engine. 1.

2. Multimodal feature extraction and labeling: The cloud feature segmentation engine extracts modal features from the received asynchronous data stream, extracts hook contour features and spatial coordinates from visual data, extracts pressure gradient change features from mechanical data, and extracts frequency domain energy distribution features from vibration data, generates multimodal feature vectors and adds modal labels; 1.

3. Cloud-based heterogeneous teacher model training: Based on multimodal feature vectors, a ResNet network as a visual analysis teacher model for image recognition, an LSTM network as a mechanical reasoning teacher model for analyzing time series pressure characteristics, and a 1D-CNN network as a vibration diagnosis teacher model for analyzing frequency domain characteristics are constructed on the cloud training platform to form an expert model cluster for crane operation scenarios; 1.

4. Model verification and knowledge base construction: Cross-validate the trained teacher model, verify the hook positioning accuracy error range with the visual model, verify the pressure anomaly prediction timeliness with the mechanical model, and verify the health status classification accuracy with the vibration model. The model weights and intermediate layer feature outputs that have passed the verification are stored in the cloud knowledge base. 1.

5. Establishment of multimodal data-model mapping relationship: According to the physical logic of the crane operation scene, the visual data stream is bound to the visual analysis teacher model, the mechanical data stream is associated with the mechanical reasoning teacher model, and the vibration data stream corresponds to the vibration diagnosis teacher model. A dynamic mapping table between modal features and teacher models is established.

7. The method for implementing distributed edge computing of construction machinery according to claim 5 is characterized in that: The step 2 specifically includes: 2.

1. Construction of global feature distillation strategy: Based on the output layer features of the multi-teacher model cluster on the cloud, the common knowledge distribution of the visual, mechanical, and vibration models is aligned through the KL divergence loss function to extract the cross-modal global features of the crane operation scene; 2.

2. Dynamically migrate local features. Aiming at the difference in computing power of edge nodes, a dynamic adaptation module is designed to adaptively adjust the migration intensity of the middle-layer features of the teacher model according to the hardware performance of the node. The boom vision node adopts high migration intensity, that is, retains 80% of the middle-layer features, and the chassis mechanics node adopts medium migration intensity, that is, retains 60% of the features; 2.

3. Lightweight student model generation: The global feature distillation results are integrated with the local dynamic migration features to generate lightweight models dedicated to computing nodes for each physical part of the crane. The number of parameters in the visual reasoning model is significantly reduced to less than 1 / 4 of the original teacher model, and the number of parameters in the mechanical analysis model is reduced to less than 1 / 3 of the original model. 2.

4. Edge-side model verification: Deploy lightweight models on vehicle-mounted edge devices, optimize model adaptability through online incremental learning, verify that the hook positioning accuracy error is significantly reduced compared to the original teacher model, and the hydraulic pressure anomaly detection delay is significantly reduced compared to cloud-based centralized computing, meeting the real-time operation requirements of cranes; 2.

5. Model and node mapping table construction: According to the physical topology of the crane, a binding relationship table between the lightweight model and the edge node is established, and the visual model for the boom end node deployment, the mechanical model for the hydraulic control node deployment, and the vibration model for the slewing bearing node deployment are defined.

8. The method for implementing distributed edge computing of construction machinery according to claim 5 is characterized in that: The step 3 specifically includes: 3.

1. Physical topology matching deployment and communication protocol configuration: Based on the mechanical structure characteristics of the crane, the lightweight model generated in step 2 is deployed to the corresponding physical unit, including the visual reasoning model deployed in the embedded device at the end of the boom, the mechanical analysis model deployed in the hydraulic control unit node, and the vibration diagnosis model deployed in the slewing bearing monitoring node; a hierarchical communication bus is established simultaneously, and the three-level data transmission priority is defined according to the flow chart communication protocol configuration node, and the transmission of key tasks is guaranteed through dynamic bandwidth allocation; 3.

2. Real-time data stream processing and redundant fault tolerance: Each edge node receives the real-time data stream of the local sensor, the visual node processes the hook positioning image, the mechanical node analyzes the time series change of the hydraulic pressure, and the vibration node analyzes the health status signal. The reasoning results are synchronized to the central decision module through the bus; at the same time, the backup lightweight model is preloaded in the hydraulic control unit and the cab terminal, and it automatically switches when the main model is detected to ensure the continuous operation of the system; 3.

3. Deployment verification and dynamic optimization: Verify the performance of the communication bus through stress testing to ensure that the highest-level data packet loss rate is lower than the standard value in the field, and that the bus utilization rate under multi-node concurrent tasks meets the preset threshold; combine the dynamic load balancing feedback mechanism of the flowchart to dynamically adjust the node resource allocation when the crane operation mode is switched to achieve continuous optimization of the system energy efficiency ratio.

9. The method for implementing distributed edge computing of construction machinery according to claim 5, characterized in that: The step 4 specifically includes: 4.

1. Real-time monitoring of node load status: Deploy load monitoring agents at each edge computing node of the crane to collect GPU utilization, memory occupancy and communication delay indicators in real time. The sampling period of the visual node is set to 100ms, the sampling period of the mechanical node is set to 200ms, and the sampling period of the vibration node is set to 500ms. The status data is aggregated to the central scheduling module through the priority scheduling bus; 4.

2. Dynamic load weight calculation: The central scheduling module uses a weighted fusion algorithm to generate the load weight of each node. The calculation formula is: load weight = 0.6×GPU utilization + 0.3×communication delay coefficient + 0.1×memory occupancy rate, where the communication delay coefficient is dynamically normalized according to the ratio of the bus transmission delay to the preset threshold. When the visual node weight exceeds the set threshold, task migration is triggered; 4.

3. Task migration decision and path planning: Based on the load weight prediction model, the node overload risk is predicted. When the prediction weight of the boom vision node exceeds the threshold for multiple cycles, the migration strategy is initiated. Specifically, the image preprocessing task is first migrated to the idle chassis positioning node, including image downsampling and ROI extraction; secondly, the migration path is selected through the bandwidth perception algorithm, and the dedicated control bus is used to transmit the migration task context data; 4.

4. Seamless connection of migration tasks: Dynamically load lightweight model copies on the target node, use the checkpoint mechanism to save the intermediate state data of the source node, and implement task switching through incremental synchronization to ensure that the interruption time of the hydraulic pressure analysis task meets the real-time requirements; 4.

5. Verification of load balancing effect: In the test under the crane lifting-transfer composite working condition, when the visual node load reaches the peak, its weight drops significantly after migrating some preprocessing tasks, the chassis node utilization rate is increased to a reasonable range, and the overall system reasoning delay is significantly reduced.

10. The method for implementing distributed edge computing of construction machinery according to claim 5, characterized in that: The step 5 specifically includes: 5.

1. Multi-source data clock synchronization: Deploy the PTP precision clock protocol in the crane sensor network to synchronize the local clocks of the visual, mechanical, and vibration sensors at the microsecond level. The residual clock deviation is processed by the spatiotemporal alignment compensation module. 5.

2. Implementation of the spatiotemporal alignment compensation algorithm: For the asynchronously arriving hook visual coordinates and hydraulic pressure data, a linear interpolation algorithm is used to generate an estimate of the pressure data at the image sampling time. 5.

3. Dynamic adjustment of sliding window: According to the difference between the frequency of vibration signal and visual data, the length of adaptive sliding window is set, and the vibration data in the window is aligned with the visual features in the frequency domain after FFT transformation; 5.

4. Multimodal data fusion reasoning: The aligned visual positioning data, interpolated pressure data and vibration spectrum features are input into the central decision module, and the attention mechanism is used to weightedly fuse the multimodal features to output the hoisting safety assessment results; 5.

5. Verification of collaborative reasoning effect: Tests show that spatiotemporal alignment compensation significantly improves the accuracy of multimodal data fusion, the sliding window mechanism effectively reduces the delay in time series data processing, and the abnormal condition detection response time of the crane meets the requirements of real-time safety monitoring.

Images