Cloud-edge-end collaborative predictive maintenance method based on DNN reasoning acceleration
By adopting a cloud-edge-device collaborative DNN predictive maintenance method in the oil and gas recovery and processing industry, and rationally allocating computing tasks, the problems of network congestion and excessive latency are solved, and efficient predictive maintenance of equipment is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHENYANG GOLDING NC & INTELLIGENCE TECH CO LTD
- Filing Date
- 2024-11-18
- Publication Date
- 2026-05-19
AI Technical Summary
In the oil and gas recovery and treatment industry, existing predictive maintenance methods for equipment rely on cloud computing, which leads to network congestion and excessive computational latency, resulting in performance bottlenecks in latency-sensitive and computationally intensive tasks.
A cloud-edge-device collaborative predictive maintenance method based on adaptive deep neural networks (DNN) is adopted. By rationally allocating computing tasks among the terminal, edge and cloud server, and optimizing resource scheduling with dynamic segmentation algorithm, inference latency is reduced and computing resource utilization is improved.
It significantly reduces inference latency, improves the real-time performance and efficiency of equipment maintenance, and can maintain high response speed and high inference accuracy in complex network environments.
Smart Images
Figure CN122069199A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of Industrial Internet of Things (IIoT), and particularly to predictive maintenance technology for the oil and gas recovery and processing industry.
[0002] With the rapid development of Industrial Internet of Things (IIoT) technology, predictive maintenance of equipment in the oil and gas recovery and processing industry has become a key technology for improving equipment reliability and reducing maintenance costs. However, existing technologies face problems such as network congestion and excessive computational latency when processing large-scale data, limiting their application in latency-sensitive and computationally intensive tasks. Currently, existing predictive maintenance methods are typically based on cloud computing, using deep neural network analysis to centrally process the massive amounts of data generated by equipment. However, with the increase in the number of devices and the improvement in the level of equipment intelligence, the amount of data generated by the devices has increased significantly. The reliance on centralized computing in the cloud faces problems of network congestion and excessive computational latency when processing large-scale data. This results in performance bottlenecks for traditional centralized cloud computing predictive maintenance methods when facing latency-sensitive and computationally intensive tasks. Summary of the Invention
[0003] To address the problems in existing technologies, this invention proposes a cloud-edge-device collaborative predictive maintenance method based on adaptive deep neural network (DNN) inference acceleration. This method significantly improves the real-time performance and efficiency of the maintenance process by optimizing the execution efficiency and accuracy of equipment health prediction tasks. Leveraging the collaborative work of cloud, edge, and terminal devices, and through the rational allocation of computing tasks and optimized resource scheduling, this invention significantly reduces inference latency and improves the utilization efficiency of computing resources. In particular, by using a dynamic segmentation algorithm to rationally divide the tasks of the deep neural network (DNN) model, the cross-device collaborative inference process is optimized. This fully utilizes the computing resources of cloud-edge-device devices for collaborative inference, reducing overall data transmission and computation latency, making the inference process more efficient, and thus enhancing the application effect of cloud-edge-device collaborative computing in the oil and gas recovery and processing industry.
[0004] The technical solution adopted in this invention is as follows: a cloud-edge-device collaborative predictive maintenance method based on DNN inference acceleration, comprising the following steps:
[0005] The terminal equipment collects data from the oil and gas processing equipment, performs data preprocessing to obtain equipment data, constructs task information, and compiles resource information of the terminal equipment. The task information and resource information are then sent to the edge server.
[0006] The edge server receives task and resource information from the terminal device and resource information from the cloud server, obtains the DNN model segmentation points through a dynamic segmentation algorithm, constructs a DNN collaborative inference scheme, and sends it to the terminal device and the cloud server.
[0007] Edge servers, terminal devices, and cloud servers perform DNN collaborative reasoning calculations according to the DNN collaborative reasoning scheme to obtain predictive maintenance prediction results;
[0008] Edge servers or cloud servers generate early warnings and maintenance information based on the prediction results, and issue maintenance instructions to terminal devices according to the pre-set maintenance plan, so that the terminal devices can perform maintenance operations according to the instructions.
[0009] The process involves data preprocessing to obtain device data, constructing task information, and statistically analyzing the resource information of the terminal device. The task information and resource information are then sent to the edge server, as detailed below:
[0010] The terminal device generates task information for predictive maintenance tasks based on the preprocessed device data; and packages the task information, the terminal device's current computing resources, and the current network resources and sends them to the edge server; the task information includes the amount of task data, the CPU clock cycles required for the task, and the task priority.
[0011] The edge server receives task and resource information from the terminal device and resource information from the cloud server, obtains DNN model segmentation points through a DNN segmentation algorithm, constructs a DNN collaborative inference scheme, and sends it to the terminal device and the cloud server, including the following steps:
[0012] The edge server receives task information from the terminal device, the terminal device's current computing and network resources, and the cloud server's current computing and network resources.
[0013] The edge server obtains the segmentation points of the CNN-LSTM model through the DNN segmentation algorithm based on the current computing resources and network resources of the three different computing devices, and divides the DNN model according to the segmentation points. Different parts of the divided CNN-LSTM model are allocated to the terminal device, cloud server and edge server respectively to perform corresponding operations to form a DNN collaborative inference scheme. The DNN collaborative inference scheme is then sent to the terminal device and cloud server.
[0014] The DNN segmentation algorithm includes the following steps:
[0015] a. Input parameters include task information for predictive maintenance tasks, current computing resources of each of the three computing devices, and current network resources;
[0016] b. Using the task information of the terminal device, as well as the current computing resources and current network resources of the terminal device, edge server, and cloud server as input, the execution latency of different layers in the CNN-LSTM model of the three different computing devices is predicted by the neural network model.
[0017] c. Following the chain order of the CNN-LSTM model, enumerate all possible combinations of split points and calculate the total inference latency for each combination;
[0018] d. By comparing the total inference latency of each combination, select the split point with the current minimum total inference latency, and iteratively update the corresponding split point and the minimum total inference latency;
[0019] e. Output the minimum total inference delay and the corresponding combination of split points.
[0020] The edge server, terminal device, and cloud server perform DNN collaborative inference calculations according to the DNN collaborative inference scheme to obtain predictive maintenance prediction results, including the following steps:
[0021] CNN-LSTM models are stored on edge servers, terminal devices, and cloud servers;
[0022] For each computing device—edge server, terminal device, and cloud server—as allocated in the DNN collaborative inference scheme, the predictive maintenance task is input into the CNN-LSTM model. Each computing device simultaneously executes its corresponding part of the operation according to the DNN collaborative inference scheme, and obtains the prediction result of the predictive maintenance task through the CNN-LSTM model.
[0023] The equipment data includes: pressure, concentration, temperature, humidity, and vibration.
[0024] A cloud-edge-device collaborative predictive maintenance system based on DNN inference acceleration includes:
[0025] Terminal equipment is used to collect data from oil and gas processing equipment, perform data preprocessing to obtain equipment data, construct task information, and collect resource information of terminal equipment, and send the task information and resource information to the edge server.
[0026] Edge servers are used to receive task and resource information from terminal devices and resource information from cloud servers. They obtain DNN model segmentation points through dynamic segmentation algorithms, construct DNN collaborative inference schemes, and send them to terminal devices and cloud servers. This enables edge servers, terminal devices, and cloud servers to perform DNN collaborative inference calculations according to the DNN collaborative inference schemes, and obtain predictive maintenance prediction results.
[0027] The cloud server is used to generate early warnings and maintenance information based on the prediction results, and to issue maintenance instructions to the terminal devices according to the pre-set maintenance plan, so that the terminal devices can perform maintenance operations according to the instructions.
[0028] The present invention has the following benefits and advantages:
[0029] 1. Improved timeliness: Through the collaborative work of cloud, edge, and device, latency is significantly reduced, especially when edge devices are under high load or network conditions are poor, it can still maintain a high response speed.
[0030] Resource optimization: By using an adaptive partitioning algorithm, computational tasks are allocated rationally, improving the utilization rate of computing resources and reducing energy consumption.
[0031] 2. Adaptable to complex network environments: The framework can work flexibly in different network environments, and can still maintain high inference accuracy and low latency under low bandwidth and high latency network conditions.
[0032] 3. Wide range of applications: This invention can be applied to fields such as oil and gas recovery and treatment, and is particularly suitable for large-scale equipment monitoring and predictive maintenance tasks, with strong industrial application value. Attached Figure Description
[0033] Figure 1 This is a diagram illustrating the overall architecture of the method of the present invention;
[0034] Figure 2 This is a flowchart of the method of the present invention;
[0035] Figure 3 This is a flowchart of the DNN model segmentation algorithm of the present invention. Detailed Implementation
[0036] The present invention will now be described in further detail with reference to the accompanying drawings.
[0037] This invention discloses a cloud-edge-device collaborative predictive maintenance method based on adaptive deep neural network inference acceleration, particularly suitable for predictive maintenance of equipment in the oil and gas recovery and processing industry within an industrial IoT environment. This method designs a cloud-edge-device collaborative predictive maintenance architecture, rationally allocating tasks to terminal devices, edge devices, and cloud servers, fully utilizing the computing power of the cloud, edge, and device ends. This method proposes an adaptive model segmentation algorithm for the cloud-edge-device environment, effectively realizing cross-device DNN collaborative inference, achieving efficient distributed inference of deep neural networks among cloud, edge, and device devices, and significantly accelerating the predictive maintenance inference process. To better describe the specific implementation of this invention, the following, combined with a practical application scenario, uses predictive maintenance in the oil and gas recovery and processing industry as an example to elaborate on the implementation steps, system configuration, and algorithm application of this invention.
[0038] A cloud-edge-device collaborative predictive maintenance method based on DNN inference acceleration includes the following steps:
[0039] 1) Establish a predictive maintenance framework that integrates cloud, edge, and device;
[0040] 2) The terminal equipment collects data from the oil and gas processing equipment;
[0041] 3) The terminal device performs data preprocessing on the collected data;
[0042] 4) The edge server receives messages from terminal devices and cloud servers, calculates the DNN model segmentation points using a dynamic segmentation algorithm based on the current system status information, and formulates a DNN collaborative inference scheme.
[0043] 5) Terminal devices, edge servers, and cloud servers perform DNN collaborative inference calculations according to the scheme to obtain prediction results;
[0044] 6) The edge server or cloud server generates corresponding early warning and maintenance information based on the prediction results, and issues maintenance instructions to the terminal device according to the pre-set maintenance plan. The terminal device performs maintenance operations according to the instructions.
[0045] The first step, establishing a cloud-edge-device collaborative predictive maintenance framework, specifically involves:
[0046] This invention constructs a cloud-edge-device collaborative predictive maintenance framework, comprising terminal devices, edge servers, and cloud servers. Terminal devices monitor equipment status in real time via sensors, performing data acquisition and preliminary processing. Edge servers perform preliminary inference calculations based on equipment status data and offload some computational tasks to the cloud. The cloud server handles complex computational tasks such as DNN model training and global data analysis. Terminal devices and edge servers communicate via 4G, 5G, or Wi-Fi for rapid data transmission and processing. Edge servers communicate with geographically distant terminal devices and cloud servers via fiber optic cables to ensure high-speed and stable data transmission. Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) models are deployed on three different computing devices to achieve real-time prediction of oil and gas processing equipment operation data. This framework significantly improves the real-time performance and efficiency of equipment maintenance through dynamic task allocation and collaborative inference.
[0047] The 2) terminal equipment performs data collection on the oil and gas processing equipment, specifically as follows:
[0048] The terminal equipment is deployed in the industrial field and communicates directly with the oil and gas processing equipment and its connected programmable logic controller (PLC) using the Modbus RTU protocol. The terminal equipment collects real-time operating data of the oil and gas processing equipment, including multi-physical field parameters such as pressure, concentration, temperature, humidity, and vibration, by running a multi-source heterogeneous data acquisition program.
[0049] The terminal device performs data preprocessing on the collected data, specifically as follows:
[0050] After collecting real-time operational data from the oil and gas processing equipment, the terminal device pushes the temporarily stored equipment data to the data preprocessing program. The data preprocessing program cleans, reduces, integrates, and extracts features from the equipment data to improve its usability and optimize the efficiency of subsequent inference.
[0051] The 4) edge server receives messages from terminal devices and cloud servers, and based on the current system status information, uses a dynamic segmentation algorithm to determine the segmentation points of the DNN model, and formulates a DNN collaborative inference scheme, including the following steps:
[0052] a. The terminal device generates basic information for predictive maintenance tasks based on the processed device data. It then packages the task information, the device's current computing power, and current network bandwidth resources and sends them to the edge server.
[0053] b. The cloud server packages its current computing power and network bandwidth resource information and sends it to the edge server;
[0054] c. The edge server receives messages from terminal devices and cloud servers, and based on current task information, device computing power, and network resource information, it predicts the execution latency of different devices in executing different layers of the CNN-LSTM model by running a neural network model.
[0055] d. Based on the predicted execution latency, the edge server runs a dynamic segmentation algorithm to calculate the segmentation points of the CNN-LSTM model, and divides the DNN model according to the segmentation points to form a DNN collaborative inference scheme.
[0056] In step 5), the terminal device, edge server, and cloud server perform DNN collaborative inference calculations according to the scheme to obtain prediction results, specifically as follows:
[0057] Terminal devices, edge servers, and cloud servers receive a DNN collaborative inference scheme and selectively activate the corresponding DNN blocks of the deployed CNN-LSTM model according to the scheme. The three different devices execute different parts of the complete CNN-LSTM model and transmit intermediate computation data via the network.
[0058] The 6) edge server or cloud server generates corresponding early warning and maintenance information based on the prediction results, and issues maintenance instructions to the terminal device according to the pre-set maintenance plan. The terminal device then performs maintenance operations according to the instructions. Specifically:
[0059] Predictive maintenance tasks are ultimately reasoned out on edge servers or cloud servers, yielding prediction results. These predictions include key information such as future pressure and concentration of the oil and gas processing equipment. Based on the prediction results and a pre-set maintenance plan, the edge server or cloud server issues predictive maintenance instructions and transmits these instructions to the terminal devices via network communication. The terminal devices then complete the corresponding maintenance tasks according to the instructions.
[0060] A dynamic segmentation algorithm is proposed to dynamically optimize the inference process of a DNN model. The main steps of the algorithm include:
[0061] a. Input parameters include the amount of task input data, inter-device bandwidth, and device computing resources;
[0062] b. Predict the inference latency of each layer when terminal devices, edge devices, and cloud servers execute CNN-LSTM models using neural network models;
[0063] c. Following the chained order of the CNN-LSTM model, calculate the total inference latency with the current position as the split point. The total inference latency of different devices is first predicted by the trained inference latency prediction model, and then the corresponding data transmission latency is added.
[0064] d. Dynamically calculate the split points of the DNN model and rationally allocate inference tasks to terminal devices, edge devices and cloud servers to minimize the total inference latency;
[0065] e. Optimize the allocation of computing resources based on the inference latency calculated at each split point, and dynamically schedule tasks among devices.
[0066] A cloud-edge-device collaborative predictive maintenance method, accelerated by adaptive DNN inference, is used in oil and gas recovery and processing equipment to monitor the equipment's health status and predict faults. For example... Figure 1 As shown, the system consists of three parts: terminal devices, edge servers, and cloud servers.
[0067] Terminal devices are deployed on oil and gas recovery and processing equipment, serving as the starting point for data acquisition. These devices primarily monitor the equipment's operational status (such as temperature, pressure, and vibration) through sensors. The terminal layer is responsible not only for real-time data acquisition but also for preliminary processing, including data cleaning, integration, and feature extraction. This processing helps reduce the amount of data transmitted and improves inference efficiency. Current terminal devices possess certain computing capabilities, enabling partial neural network inference. When a terminal device detects an abnormal trend, it can promptly trigger an alert and quickly upload key data to the edge or cloud. In this way, the terminal layer not only accelerates anomaly response speed but also effectively reduces the burden on the upstream layer.
[0068] The edge server is primarily responsible for controlling real-time DNN segmentation inference and local analysis. Located at the network edge, close to terminal devices, the edge layer possesses greater computing power. Therefore, an adaptive DNN collaborative inference module is deployed in this framework, responsible for receiving task information from edge devices and providing DNN model segmentation decisions. When the DNN segments to a single point, the edge layer, through partial inference from the DNN model, can quickly identify trends and potential faults in device status, thus providing preliminary analysis results for predictive maintenance. When the DNN segments to two points, the data is sent to the cloud for further collaborative inference. Furthermore, the edge layer can adaptively adjust the inference algorithm after the cloud model is updated to improve real-time performance and inference accuracy.
[0069] The cloud server acts as the core analysis center of the framework, primarily responsible for deep data analysis, model training, and global optimization. The cloud receives intermediate data from the edge layer during the inference process, utilizing its more powerful computing resources to execute subsequent inference and analysis tasks, generating device alerts and maintenance decisions. Simultaneously, based on long-term, massive data accumulation, the cloud layer iteratively trains the inference model using historical data, thereby generating global maintenance strategies and updating the model. By periodically updating the inference models at the edge and terminal layers, the cloud layer ensures the entire framework adapts to changes in device status and environment, achieving improved accuracy and response speed for predictive maintenance. Model updates completed in the cloud are transmitted to the edge and terminals, gradually enhancing the inference capabilities of edge devices, enabling them to perform predictive and maintenance tasks more independently.
[0070] The flowchart of the method of the present invention is as follows Figure 2 As shown, the process begins with the terminal device collecting data from the oil and gas processing equipment. This includes real-time pressure, concentration, and other operational status data, as well as multi-physical field data such as temperature, humidity, and vibration. The terminal device then performs basic data cleaning, conversion, and feature extraction preprocessing. Next, the terminal device pushes the task information to its connected edge server. The edge server calculates the DNN model segmentation points based on the current system status and formulates a DNN collaborative inference scheme. Following this, the terminal device, edge server, and cloud server perform DNN collaborative inference calculations according to the scheme. The final prediction result is generated on the edge server or cloud server. A PDM plan generation module and a maintenance instruction issuance module are deployed on the edge server or cloud server. These modules generate corresponding early warnings and maintenance information based on the prediction results and issue maintenance instructions to the terminal device according to the pre-set maintenance plan, ultimately achieving a closed-loop feedback for the entire predictive maintenance task process.
[0071] The flowchart of the dynamic segmentation algorithm using the DNN model in this invention is as follows: Figure 3 As shown, it mainly consists of the following steps:
[0072] (1) The edge server collects key data such as task information, network status and computing power of connected devices, and uses this information to train the DNN inference delay prediction model.
[0073] (2) The edge server uses the trained DNN inference latency prediction model to estimate the inference latency of the DNN layer on different computing devices in real time, providing a latency prediction basis for subsequent DNN model segmentation.
[0074] (3) Calculate data transmission latency: Calculate the data transmission latency based on the bandwidth and data volume between devices;
[0075] (4) Enumerate all possible combinations of split points, and calculate the total inference latency for each combination by traversing all possible combinations of split points between terminal devices, edge devices and cloud servers.
[0076] (5) By comparing the total inference latency of each combination, select the split point with the current minimum latency, and iteratively update the corresponding split point and the minimum total latency;
[0077] (6) The algorithm outputs the minimum total inference latency and the two corresponding split points in the current system state. These split points are used to determine the allocation of model tasks among different devices.
Claims
1. A cloud-edge-device collaborative predictive maintenance method based on DNN inference acceleration, characterized in that, Includes the following steps: The terminal equipment collects data from the oil and gas processing equipment, performs data preprocessing to obtain equipment data, constructs task information, and compiles resource information of the terminal equipment. The task information and resource information are then sent to the edge server. The edge server receives task and resource information from the terminal device and resource information from the cloud server. It obtains the DNN model segmentation points through a dynamic segmentation algorithm, constructs a DNN collaborative inference scheme, and sends it to the terminal device and the cloud server. Edge servers, terminal devices, and cloud servers perform DNN collaborative reasoning calculations according to the DNN collaborative reasoning scheme to obtain predictive maintenance prediction results; Edge servers or cloud servers generate early warnings and maintenance information based on the prediction results, and issue maintenance instructions to terminal devices according to the pre-set maintenance plan, so that the terminal devices can perform maintenance operations according to the instructions.
2. The cloud-edge-device collaborative predictive maintenance method based on DNN inference acceleration according to claim 1, characterized in that, The process involves data preprocessing to obtain device data, constructing task information, and statistically analyzing the resource information of the terminal device. The task information and resource information are then sent to the edge server, as detailed below: The terminal device generates task information for predictive maintenance tasks based on the preprocessed device data; and packages the task information, the terminal device's current computing resources, and the current network resources and sends them to the edge server; the task information includes the amount of task data, the CPU clock cycles required for the task, and the task priority.
3. The cloud-edge-device collaborative predictive maintenance method based on DNN inference acceleration according to claim 1, characterized in that, The edge server receives task and resource information from the terminal device and resource information from the cloud server, obtains DNN model segmentation points through a DNN segmentation algorithm, constructs a DNN collaborative inference scheme, and sends it to the terminal device and the cloud server, including the following steps: The edge server receives task information from the terminal device, the terminal device's current computing and network resources, and the cloud server's current computing and network resources. The edge server obtains the segmentation points of the CNN-LSTM model through the DNN segmentation algorithm based on the current computing resources and network resources of the three different computing devices, and divides the DNN model according to the segmentation points. Different parts of the divided CNN-LSTM model are allocated to the terminal device, cloud server and edge server respectively to perform corresponding operations to form a DNN collaborative inference scheme. The DNN collaborative inference scheme is then sent to the terminal device and cloud server.
4. The cloud-edge-device collaborative predictive maintenance method based on DNN inference acceleration according to claim 3, characterized in that, The DNN segmentation algorithm includes the following steps: a. Input parameters include task information for predictive maintenance tasks, current computing resources of each of the three computing devices, and current network resources; b. Using the task information of the terminal device, as well as the current computing resources and current network resources of the terminal device, edge server, and cloud server as input, the execution latency of different layers in the CNN-LSTM model of the three different computing devices is predicted by the neural network model. c. Following the chain order of the CNN-LSTM model, enumerate all possible combinations of split points and calculate the total inference latency for each combination; d. By comparing the total inference latency of each combination, select the split point with the current minimum total inference latency, and iteratively update the corresponding split point and the minimum total inference latency; e. Output the minimum total inference delay and the corresponding combination of split points.
5. The cloud-edge-device collaborative predictive maintenance method based on DNN inference acceleration according to claim 1, characterized in that, The edge server, terminal device, and cloud server perform DNN collaborative inference calculations according to the DNN collaborative inference scheme to obtain predictive maintenance prediction results, including the following steps: CNN-LSTM models are stored on edge servers, terminal devices, and cloud servers; For each computing device—edge server, terminal device, and cloud server—as allocated in the DNN collaborative inference scheme, the predictive maintenance task is input into the CNN-LSTM model. Each computing device simultaneously executes its corresponding part of the operation according to the DNN collaborative inference scheme, and obtains the prediction result of the predictive maintenance task through the CNN-LSTM model.
6. The cloud-edge-device collaborative predictive maintenance method based on DNN inference acceleration as described in claim 1, characterized in that, The equipment data includes: pressure, concentration, temperature, humidity, and vibration.
7. A cloud-edge-device collaborative predictive maintenance system based on DNN inference acceleration, characterized in that, include: Terminal equipment is used to collect data from oil and gas processing equipment, perform data preprocessing to obtain equipment data, construct task information, and collect resource information of terminal equipment, and send the task information and resource information to the edge server. Edge servers are used to receive task and resource information from terminal devices and resource information from cloud servers. They obtain DNN model segmentation points through dynamic segmentation algorithms, construct DNN collaborative inference schemes, and send them to terminal devices and cloud servers. This enables edge servers, terminal devices, and cloud servers to perform DNN collaborative inference calculations according to the DNN collaborative inference schemes, and obtain predictive maintenance prediction results. The cloud server is used to generate early warnings and maintenance information based on the prediction results, and to issue maintenance instructions to the terminal devices according to the pre-set maintenance plan, so that the terminal devices can perform maintenance operations according to the instructions.