Digital drainage twinning large model end cloud collaborative deployment method
By constructing a three-tier deployment architecture and lightweight technology, the cloud dependency problem of the urban drainage twin system is solved, enabling edge autonomy and efficient model operation, adapting to low-power and low-latency scenarios, and improving the system's autonomous decision-making ability and operating efficiency.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHANGHAI ZHONGJING HANDING DIGITAL TECH CO LTD
- Filing Date
- 2026-02-09
- Publication Date
- 2026-05-12
AI Technical Summary
Existing urban drainage twin systems rely heavily on the cloud, resulting in delayed early warning responses when the network is interrupted, high power consumption at edge terminals, extended model inference time, and limited computing power scheduling, making it difficult to adapt to low-power, low-latency scenarios. Furthermore, the lack of lightweight deployment design hinders large-scale applications.
We construct a three-layer deployment architecture of terminal-edge-cloud, and compress hundreds of billions of parameters to the edge terminal through model lightweighting technology. This achieves a single terminal power consumption of ≤5W while maintaining an accuracy loss of ≤3%. We also optimize the scheduling of heterogeneous computing power to improve model running efficiency.
It achieves edge autonomy latency of ≤20ms in the offline state, model volume compression of 75%, accuracy loss of ≤3%, power consumption reduction of 60%, and running efficiency improvement of 40%.
Smart Images
Figure CN122019104A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of digital twin and lightweight artificial intelligence model technology, and in particular to a method for edge-cloud collaborative deployment of a large-scale digital drainage twin model. Background Technology
[0002] In the field of urban drainage twin system technology applications, cloud-based centralized deployment architecture is currently the mainstream implementation solution widely adopted in the industry. However, this architecture has three major technical defects in practical applications: First, the system is highly dependent on the cloud. When network transmission is interrupted, edge devices lose their autonomous decision-making capabilities, resulting in a flood warning response delay of more than 500ms, which seriously affects the timeliness of warnings and the efficiency of emergency response. Second, the large AI models used in drainage twin scenarios typically have hundreds of billions of parameters. If deployed directly on edge terminals, the power consumption of a single device will exceed 20W, and the model inference latency will exceed 100ms, making it unsuitable for special application scenarios with low power consumption and low latency, such as underground terminals and gate stations. Third, the existing architecture's computing power scheduling mode is relatively simple, and it has not carried out targeted allocation and optimization design for heterogeneous computing power resources of edge CPU, GPU, and NPU, resulting in low overall model operating efficiency and difficulty in fully utilizing the hardware computing power potential.
[0003] While existing technologies include AI-based solutions for building drainage unit twin systems and management systems using AI and 3D technologies, these solutions generally lack lightweight deployment designs adapted to edge devices, hindering their effective implementation on edge terminals. Furthermore, although HarmonyOS 6 possesses edge-cloud collaboration capabilities, it currently lacks a dedicated layered deployment standard and adaptation scheme for large-scale drainage twin models, thus failing to effectively address the balance challenge between model accuracy, terminal computing power, and operating power consumption.
[0004] In summary, existing urban drainage twin systems still have many technical shortcomings in terms of deployment architecture, model adaptation, computing power scheduling, and standards and specifications, which restrict their large-scale application and performance improvement in edge scenarios. There is an urgent need for a technical solution that can take into account autonomous decision-making capabilities, lightweight deployment requirements, and optimized scheduling of heterogeneous computing power. Summary of the Invention
[0005] To address the shortcomings of existing technologies, this invention proposes a digital drainage twin large-scale model edge-cloud collaborative deployment method, the purpose of which is: Construct a three-tier deployment architecture of terminal-edge-cloud to enable autonomous inference of edge nodes in the absence of network access, thereby reducing the system's dependence on the cloud. By using model lightweighting and quantization technology, large models with hundreds of billions of parameters can be compressed to a scale that edge terminals can handle, ensuring that the accuracy loss is ≤3%, while achieving edge inference latency ≤50ms and single terminal power consumption ≤5W. Optimize heterogeneous computing power scheduling strategies to enable collaborative work between CPU / GPU / NPU and improve model running efficiency.
[0006] This invention provides a method for edge-cloud collaborative deployment of a digital drainage twin large-scale model, comprising: Obtain the original model and perform lightweight processing to obtain a lightweight model; Deploy lightweight models using a three-tier architecture; Computing power is allocated according to the task; Conduct edge-cloud collaborative testing; and Model training and updates are performed based on data to complete model iteration.
[0007] In one embodiment of the present invention, obtaining the original model and performing lightweight processing to obtain a lightweight model includes: Obtain the original model; Calculate the contribution of each operator, prune redundant operators whose contribution is below the threshold, and fine-tune the pruned model. The original model after trimming is trained using knowledge distillation. Configure hybrid quantization parameters to perform quantization and obtain the quantization model; The quantization model is converted to a lightweight model in OM / HIM format.
[0008] In one embodiment of the present invention, the hybrid quantization parameters include INT4 weight quantization and INT8 activation value quantization. In one embodiment of the present invention, the process of obtaining the original model and performing lightweight processing must satisfy the accuracy loss constraint during the lightweight model acquisition process: .
[0009] In one embodiment of the present invention, the deployment of the lightweight model through a three-tier architecture includes: Deploy the HIM model at the edge terminal layer; Deploy the OM model at the edge computing layer; Deploy the full-precision original model at the cloud service layer.
[0010] In one embodiment of the present invention, the computing power scheduling based on the task includes mapping matrix operations, control logic, and attention mechanisms to NPU, CPU, and GPU computing power units respectively according to the task type.
[0011] In one embodiment of the present invention, the computing power scheduling efficiency in the computing power scheduling based on the task is: ; Where Pi is the power consumption of the i-th type of computing unit, and Ti is the running time of the i-th type of computing unit.
[0012] In one embodiment of the present invention, the end-to-cloud collaborative test includes starting the system and verifying the data synchronization and policy distribution functions under normal network conditions; The autonomous reasoning capability of edge nodes was verified by simulating a network outage scenario.
[0013] In one embodiment of the present invention, the edge-cloud collaborative testing must meet the edge-side inference latency constraint: .
[0014] The present invention has the following beneficial effects: (1) A three-layer deployment architecture of HarmonyOS edge cloud and cloud is proposed for drainage twin big model to solve the problems of high cloud dependence and network failure in traditional systems, and the edge autonomy latency is ≤20ms; (2) Innovative lightweight technology combining “quantization + distillation + trimming” achieves the dual goals of 75% model volume compression and ≤3% accuracy loss, adapting to the low power consumption requirements of edge terminals; (3) Design a CPU / GPU / NPU heterogeneous computing power scheduling algorithm. Compared with the traditional single computing power scheduling, the model running efficiency is improved by 40% and the energy consumption is reduced by 60%. Attached Figure Description
[0015] Figure 1 A flowchart of a digital drainage twin big model edge-cloud collaborative deployment method is shown in one embodiment of the present invention. Detailed Implementation
[0016] In the following description, the invention is described with reference to various embodiments. However, those skilled in the art will recognize that the embodiments may be practiced without one or more specific details or with other alternatives and / or additional methods, materials, or components. In other instances, well-known structures, materials, or operations are not shown or described in detail so as not to obscure the inventive points of the invention. Similarly, for illustrative purposes, specific quantities, materials, and configurations are set forth to provide a comprehensive understanding of embodiments of the invention. However, the invention is not limited to these specific details.
[0017] In this invention, the various embodiments are merely intended to illustrate the solutions of the invention and should not be construed as limiting.
[0018] In this specification, references to "an embodiment" or "this embodiment" mean that a particular feature, structure, or characteristic described in connection with that embodiment is included in at least one embodiment of the invention. The phrase "in one embodiment" appearing throughout this specification does not necessarily refer to the same embodiment in all instances.
[0019] Furthermore, the numbering of the steps in the methods of the present invention does not limit the execution order of the method steps. Unless otherwise specified, the method steps may be executed in different orders.
[0020] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.
[0021] Figure 1 A flowchart of a digital drainage twin big model edge-cloud collaborative deployment method is shown in one embodiment of the present invention.
[0022] In this embodiment, the present invention proposes a method for edge-cloud collaborative deployment of a digital drainage twin big model based on HarmonyOS 6.
[0023] In this embodiment, the hardware environment is as follows: Edge terminal layer: HarmonyOS 6 embedded terminal (equipped with NPU computing unit, computing power ≥2TOPS, power consumption ≤5W), well water level sensor, gate station control module; Edge computing layer: HarmonyOS 6 edge server (CPU ≥ 16 cores, GPU ≥ 8G VRAM, NPU computing power ≥ 20TOPS); Cloud service layer: cloud server cluster (CPU ≥ 64 cores, GPU ≥ 40G video memory, supports training of large models with hundreds of billions of parameters).
[0024] The software environment is as follows: (1) Operating system: HarmonyOS 6.0 and above (supports end-to-end cloud collaboration and heterogeneous computing power scheduling); (2) Model framework: MindSporeLite, TensorFlowLite (supports model quantization and conversion); (3) Communication protocols: MQTT, gRPC (supports low-latency data transmission between edge and cloud); (4) Deployment tool: HarmonyOS DevEcoStudio (supports OM / HIM model compilation and deployment).
[0025] In this embodiment, the method for collaborative deployment of the digital drainage twin big data model on the edge and cloud includes: S100: Model lightweighting processing. Obtain the target AI algorithm model, and sequentially perform operator pruning, knowledge distillation, and INT4 / INT8 quantization operations to output a lightweight model in OM / HIM format. Accuracy loss constraints must be met. .
[0026] The detailed steps are as follows: We extracted the AI twin management and control algorithm model and the pipeline water level siltation monitoring algorithm model, analyzed the model structure and parameter distribution, and used a standard test dataset (100,000+ drainage pipeline operation data) for pre-training and verification, recording the original model volume, inference accuracy, inference latency and other benchmark indicators. Operator pruning: Calculate the contribution score of each operator, prune redundant operators with a contribution score lower than 0.05, and perform 30 rounds of fine-tuning training on the pruned model to ensure that the model volume is compressed by ≥75% and the accuracy loss is ≤3%.
[0027] The operator contribution is calculated using the following formula: ; Where ∂Acc / ∂Oi is the partial derivative of the model accuracy with respect to the i-th operator, Comp Oi Let Comp be the computational complexity of the i-th operator. total This represents the total computational cost of the model and is used to filter redundant operators.
[0028] Using the original full-precision model as the teacher model and the quantized model as the student model, a combined loss function of "KL divergence soft-label loss + cross-entropy hard-label loss" is constructed. Distillation training is performed using a dataset of over 500,000 real-world samples (Epoch=100, BatchSize=32). Knowledge distillation is performed using the following formula, where the original model is the teacher model and the quantized model is the student model: ; Among them, α=0.7, β=0.3, KL(P T ,P S ) represents the KL divergence of the output probability distributions of the teacher model and the student model, CE(y,P) S ) represents the cross-entropy loss between the student model and the real labels, used to ensure the accuracy of the model after distillation.
[0029] An asymmetric quantization scheme is used to statistically analyze the range of weight values for each layer [min]. w ,max w The scaling factor S and zero point Z are calculated, and the 32-bit floating-point weights are converted into INT4 / INT8 integer weights, eliminating redundant parameters whose absolute values are less than the dynamic threshold. The quantization scaling factor is calculated using the following formula: ; Where b is the quantization bit width, and max w The maximum value of the weight is min. w This is the minimum weight value, used to map floating-point weights to the integer range.
[0030] The quantization zero point is calculated using the following formula: ; Here, round(・) is the rounding function, and Z is the zero offset of the quantized integer, which is used to reduce quantization mapping error.
[0031] Convert floating-point weights to integer weights using the following formula: ; Where w is the original 32-bit floating-point weight, w q The weights are quantized b-bit integers to achieve weight data compression.
[0032] The quantization model is converted to a lightweight model in OM / HIM format.
[0033] The S200 is deployed in a three-tier architecture, with the HIM model deployed at the edge terminal layer, the OM model deployed at the edge computing layer, and the full-precision model deployed at the cloud service layer, thus completing the configuration of functional modules at each layer.
[0034] A three-tier deployment architecture is constructed: edge terminal layer, edge computing layer, and cloud service layer. The edge terminal layer deploys a lightweight HIM model, responsible for real-time data acquisition, local inference, and alarm triggering, adapting to HarmonyOS 6 embedded devices. The edge computing layer deploys a hybrid precision OM model, responsible for region-level data fusion, twin simulation, and edge-cloud data synchronization, supporting autonomous operation even when the network is down. The cloud service layer deploys a full-precision large model, responsible for full-domain simulation, model training and updates, and global policy distribution. The three-tier deployment architecture process is as follows: Edge terminal layer deployment: Burn the HIM format lightweight model to the HarmonyOS 6 embedded device, configure the data acquisition interface and local inference engine, and set alarm trigger thresholds (such as triggering a level 1 alarm when the water level exceeds 80cm). Edge computing layer deployment: Deploy OM format mixed precision models on edge computing nodes, configure area-level data fusion rules and twin simulation parameters, and build a local cache database (caching model parameters and strategy data for the most recent 30 days). Deployment of cloud service layer: Deploy a full-precision large model on a cloud server cluster, build a model training platform and a global strategy decision module, and configure the end-to-cloud data synchronization interface and model update scheduling rules.
[0035] The S300 heterogeneous computing power scheduling configuration maps matrix operations, control logic, and attention mechanisms to NPU, CPU, and GPU computing units respectively, based on task type. Specific rules are as follows: Matrix operation tasks (such as model convolutional layer and fully connected layer calculations) are assigned to the NPU, and the NPU hardware acceleration interface is called. Control logic tasks (such as data acquisition scheduling, alarm command generation, and network disconnection switching control) are assigned to the CPU; Attention mechanism tasks (such as whole-basin data correlation analysis and twin scene feature matching) are assigned to GPUs.
[0036] The computing power scheduling efficiency is: ; Where Pi is the power consumption of the i-th type of computing unit, and Ti is the running time of the i-th type of computing unit.
[0037] S400 and edge-cloud collaborative testing: The system was started to verify data synchronization and policy distribution functions under normal network conditions. Simulated network outage scenarios were used to verify the autonomous inference capabilities of edge nodes, including: Start the system and verify the data synchronization and policy distribution functions under normal network conditions; The autonomous reasoning capability of edge nodes was verified by simulating a network outage scenario.
[0038] In one embodiment of the present invention, the edge-cloud collaborative testing must meet the edge-side inference latency constraint: .
[0039] S500, Model Iteration and Optimization: The cloud performs model training and updates based on data uploaded from the edge, and distributes the data to the edge nodes to complete the model iteration.
[0040] In this embodiment, the technical effectiveness is verified through actual testing. The experimental environment is a simulated urban drainage pipe network scenario, and the test data is as follows: (1) Model lightweighting effect: The original model with hundreds of billions of parameters was 20GB in size. After combination optimization, it was compressed to 5GB, with a compression rate of 75% and an accuracy loss rate of 2.1%, which meets the requirement of ≤3%. (2) Edge-side operation indicators: The edge terminal inference latency is 38ms and the power consumption of a single device is 3.2W, both of which are better than the design targets (≤50ms, ≤5W). (3) Edge autonomy capability: The autonomous inference latency of edge nodes in the offline state is 18ms, which is much lower than the 500ms or more of the traditional system; (4) Computing power scheduling efficiency: Compared with traditional single CPU scheduling, heterogeneous computing power scheduling improves the model running efficiency by 42% and reduces energy consumption by 61%.
[0041] Although various embodiments of the invention have been described above, it should be understood that they are presented by way of example only and not as limitations. It will be apparent to those skilled in the art that various combinations, modifications, and alterations can be made without departing from the spirit and scope of the invention. Therefore, the breadth and scope of the invention disclosed herein should not be limited by the exemplary embodiments disclosed above, but should be defined solely by the appended claims and their equivalents.
Claims
1. A method for collaborative deployment of a digital drainage twin large-scale model via edge-cloud, characterized in that, include: Obtain the original model and perform lightweight processing to obtain a lightweight model; Deploy lightweight models using a three-tier architecture; Computing power is allocated according to the task; Conduct edge-cloud collaborative testing; as well as Model training and updates are performed based on data to complete model iteration.
2. The method according to claim 1, characterized in that, The process of obtaining the original model and performing lightweight processing to obtain a lightweight model includes: Obtain the original model; Calculate the contribution of each operator, prune redundant operators whose contribution is below the threshold, and fine-tune the pruned model. The original model after trimming is trained using knowledge distillation. Configure hybrid quantization parameters to perform quantization and obtain the quantization model; The quantization model is converted to a lightweight model in OM / HIM format.
3. The method according to claim 2, characterized in that, The hybrid quantization parameters include INT4 weight quantization and INT8 activation value quantization.
4. The method according to claim 1, characterized in that, The process of obtaining the original model and performing lightweight processing must satisfy accuracy loss constraints during the lightweight model acquisition process. 。 5. The method according to claim 1, characterized in that, The deployment of the lightweight model via a three-tier architecture includes: Deploy the HIM model at the edge terminal layer; Deploy the OM model at the edge computing layer; Deploy the full-precision original model at the cloud service layer.
6. The method according to claim 1, characterized in that, The task-based computing power scheduling includes mapping matrix operations, control logic, and attention mechanisms to NPU, CPU, and GPU computing power units, respectively, based on the task type.
7. The method according to claim 1, characterized in that, The computing power scheduling efficiency in the task-based computing power scheduling is: ； Where Pi is the power consumption of the i-th type of computing unit, and Ti is the running time of the i-th type of computing unit.
8. The method according to claim 1, characterized in that, The edge-cloud collaboration test includes starting the system and verifying the data synchronization and policy distribution functions under normal network conditions. The autonomous reasoning capability of edge nodes was verified by simulating a network outage scenario.
9. The method according to claim 1, characterized in that, The edge-cloud collaborative testing must meet the edge-side inference latency constraint: 。