Method, device and system for constructing a multi-level model system

Abstract

The present invention provides a method, device and system for constructing a multi-level model system, which belongs to the field of artificial intelligence technology. The multi-level model system includes: a cloud server, on which a cloud model is arranged; at least one edge node communicating with the cloud server, on which an edge model is arranged, and the edge node communicates with at least one device node; at least one device node, on which an end model is arranged; wherein the cloud model can update its own structure and model parameters according to the data of at least one of the edge models, the edge model can update its own structure and model parameters according to the data of the cloud model and the end model, and the end model can update its own structure and model parameters according to the data of the edge model. The technical solution of the present invention can effectively improve the adaptive ability and system robustness of the model on the distributed deployment nodes.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method, device and system for building a multi-level model system. Background Art

[0002] Device perception data is heterogeneous and changes over time, which leads to large differences in the inference accuracy and performance of the unified model trained with centralized data on different devices, which creates obstacles for the end-side deployment and operation of the model. For example, due to the different positions and lighting conditions of each smart camera, the image content and features they see are different, resulting in a decrease in the accuracy of the unified training model on the end. Therefore, on the device, even on the edge and in the cloud, it is necessary to continuously learn and update the model based on different business data.

[0003] Existing on-device learning methods mainly use device-side perception data for iterative learning. Due to the limitation of local device data, it is easy to cause problems such as difficulty in on-device model training. Distributed machine learning (DistributedML) methods, such as parameter servers, online learning, and federated learning, consider using distributed computing power and global data for model training, but the goal is to obtain a unified cloud model, and there are still data adaptation problems during the deployment process on the end side. Existing methods cannot meet the requirements of personalized and continuous learning of models on the end side.

[0004] Existing on-device learning and distributed machine learning methods have the following technical limitations:

[0005] Regarding the on-end learning method, after the pre-trained model is deployed to the terminal, as the data on the terminal accumulates, a new training data set is gradually formed on the terminal, and the model is retrained using the terminal computing power. Through iteration, the model can be continuously learned on the terminal. However, due to factors such as the deployment environment, there are hidden dangers in local data training, such as the small number of effective samples (low effective frame density of a camera at a certain angle), uneven distribution of sample categories (anomaly detection), etc. Relying solely on local data will cause overfitting or underfitting problems in model training. The lack of upper-level knowledge or interaction with multiple devices greatly limits the sustainability of on-end learning.

[0006] Regarding distributed machine learning methods, a distributed architecture (parameter server) was initially introduced to improve the performance of large-scale training tasks. The data and models were divided into different training nodes for distributed training, and then the aggregation nodes were used for task / resource scheduling, parameter aggregation, etc. to generate a unified cloud model. The advantage of distributed machine learning methods is that they coordinate multi-terminal data and computing power for global training. However, in the production deployment stage, fixed models are still deployed on the terminal in a "learn-than-deploy" manner, which essentially does not solve the personalized differences between devices and users, and the terminal models are not adaptive and personalized.

[0007] Therefore, existing technologies are rather polarized and cannot meet the requirements: 1) either lack consideration of global knowledge collaboration; 2) either lack adaptability to heterogeneous terminal data. Summary of the invention

[0008] The technical problem to be solved by the present invention is to provide a method, device and system for constructing a multi-level model system.

[0009] To solve the above technical problems, the embodiments of the present invention provide the following technical solutions:

[0010] In one aspect, a multi-level model system is provided, comprising:

[0011] A cloud server, wherein a cloud model is arranged on the cloud server;

[0012] at least one edge node in communication with the cloud server, the edge node having an edge model disposed thereon, the edge node in communication with at least one device node;

[0013] The at least one device node, wherein the device node is provided with an end model;

[0014] Among them, the cloud model can update its own structure and model parameters according to the data of at least one of the edge models, the edge model can update its own structure and model parameters according to the data of the cloud model and the end model, and the end model can update its own structure and model parameters according to the data of the edge model.

[0015] In some embodiments, the edge node includes:

[0016] Access edge nodes to communicate with the device nodes and core edge nodes respectively;

[0017] The core edge node communicates with the access edge node and the cloud server respectively.

[0018] In some embodiments, the end model is used to be updated collaboratively using a dynamically updated end database and edge knowledge of the edge model;

[0019] The edge model is used to be updated by using the end knowledge collaboration of the end model of at least one device node communicating with the edge node and the cloud knowledge collaboration of the cloud model of the cloud server communicating with the edge node;

[0020] The cloud model is used to collaboratively update using edge knowledge of an edge model of at least one edge node in communication with the cloud server.

[0021] In some embodiments, the cloud model is used to perform model compression based on computing power and business indicators of at least one edge node in communication with the cloud server;

[0022] The edge model is used to perform model compression according to the computing power and business indicators of at least one device node that communicates with the edge node.

[0023] In some embodiments, the device node includes: a vehicle-mounted device with a GPS module, and a smart camera.

[0024] In some embodiments, the device node is used to mark new data streams through automatic labeling to generate a dynamically updated end database.

[0025] An embodiment of the present invention further provides a method for constructing a multi-level model system, which is used to construct the multi-level model system as described above, comprising:

[0026] generating an initial cloud model based on the cloud data set;

[0027] Based on the computing resources of the edge nodes, the model is compressed, the initial cloud model is compressed into the initial edge model, and then deployed to the edge nodes;

[0028] According to the computing resources of the device node, the model is compressed, the initial edge model is compressed into the initial end model, and deployed to the device node.

[0029] In some embodiments, the method is applied to a device node, including:

[0030] After time t, the end training data set is formed by automatically labeling the perception data;

[0031] Request the latest edge model from the edge node, and based on the initial edge model and edge training data set, perform continuous learning of the edge model under the edge knowledge collaboration of the latest edge model to generate a new edge model;

[0032] Push the new end model to the edge node and perform on-end inference calculations based on the new end model.

[0033] In some embodiments, the method is applied to an edge node and includes:

[0034] Obtain the latest end model set and the latest end data set of multiple communicating device nodes, wherein the latest end data set includes the latest end training data sets of the multiple device nodes, and request the latest cloud model from the cloud server;

[0035] Based on the latest edge model and the latest aggregated edge data set, the edge model is continuously learned under the cloud knowledge collaboration of the latest cloud model and the end knowledge collaboration of the latest end model set to generate a new edge model;

[0036] Push the new edge model to the cloud server.

[0037] In some embodiments, the method is applied to a cloud server and includes:

[0038] Obtain the latest edge model set and the latest edge data set of multiple communicating edge nodes;

[0039] Based on the latest cloud model and the latest aggregated cloud dataset, the cloud model is continuously learned under the edge knowledge collaboration of the latest edge model set to generate a new cloud model.

[0040] An embodiment of the present invention further provides a device for constructing a multi-level model system, comprising:

[0041] A cloud model generation module, used to generate an initial cloud model based on a cloud data set;

[0042] The edge model generation module is used to compress the model according to the edge node computing resources, compress the initial cloud model into the initial edge model, and deploy it to the edge node;

[0043] The end model generation module is used to compress the model according to the computing resources of the device node, compress the initial edge model into the initial end model, and deploy it to the device node.

[0044] An embodiment of the present invention also provides a device for constructing a multi-level model system, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor; when the processor executes the program, the method for constructing a multi-level model system as described above is implemented.

[0045] An embodiment of the present invention further provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the method for constructing a multi-level model system as described above.

[0046] The embodiments of the present invention have the following beneficial effects:

[0047] In the above solution, the model is deployed on multi-level nodes, and then through continuous learning of new data and integration of multi-level knowledge collaboration mechanisms, the models on each layer and each node can be dynamically updated and knowledge collaboratively learned. The architecture can be flexibly scaled and data privacy is taken into consideration. It has a wide range of applications and is especially suitable for solving the multi-source heterogeneity and time-varying problems of distributed end-side data such as artificial intelligence, machine intelligence, and IoT. It can effectively improve the adaptability and system robustness of the model on the distributed deployment nodes. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 A schematic diagram of a multi-level model system according to an embodiment of the present invention;

[0049] Figure 2 A schematic diagram of a terminal model continuous learning mechanism based on knowledge collaboration in an embodiment of the present invention;

[0050] Figure 3 A schematic diagram of a continuous learning mechanism of an edge model based on knowledge collaboration according to an embodiment of the present invention;

[0051] Figure 4 A schematic diagram of a cloud model continuous learning mechanism based on knowledge collaboration according to an embodiment of the present invention;

[0052] Figure 5 A schematic diagram of an embodiment of the present invention performing multi-level model compression and deployment according to actual computing power and business;

[0053] Figure 6 A schematic diagram of a flow chart of a method for constructing a multi-level model system according to an embodiment of the present invention;

[0054] Figure 7 A schematic diagram of an embodiment of the present invention applied to vehicle network scenario to predict vehicle position;

[0055] Figure 8 This is a schematic diagram of an embodiment of the present invention applied to abnormal behavior detection in a smart elderly care scenario. DETAILED DESCRIPTION

[0056] In order to make the technical problems, technical solutions and advantages to be solved by the embodiments of the present invention more clear, a detailed description will be given below with reference to the accompanying drawings and specific embodiments.

[0057] An embodiment of the present invention provides a multi-level model system, comprising:

[0058] A cloud server, wherein a cloud model is arranged on the cloud server;

[0059] at least one edge node in communication with the cloud server, the edge node having an edge model disposed thereon, the edge node in communication with at least one device node;

[0060] The at least one device node, wherein the device node is provided with an end model;

[0061] Among them, the cloud model can update its own structure and model parameters according to the data of at least one of the edge models, the edge model can update its own structure and model parameters according to the data of the cloud model and the end model, and the end model can update its own structure and model parameters according to the data of the edge model.

[0062] In this embodiment, the model is deployed on multi-level nodes, and then through continuous learning of new data and integration of multi-level knowledge collaboration mechanisms, the models on each layer and each node can be dynamically updated and knowledge collaboratively learned. The architecture can be flexibly scaled and data privacy is taken into consideration. It has a wide range of applications and is particularly suitable for solving the multi-source heterogeneity and time-variability problems of distributed terminal data such as artificial intelligence, machine intelligence, and IoT. It can effectively improve the adaptability and system robustness of the model on the distributed deployment nodes.

[0063] In this embodiment, through the architecture design, the model is divided into cloud models (Cloud Model, represented by Cm t ), Edge Model (Edge Model, denoted as Em i t ) and the end model (Device Model, represented by Dm k t ), the cloud model, edge model, and end model can be various types of business models, such as target detection models and classification models. However, unlike existing deep learning models, the cloud model, edge model, and end model of this embodiment have a continuous learning mechanism and are continuous learning models. Continuously learning is a method of continuously learning new knowledge from newly acquired data after the model is deployed. This embodiment provides a multi-level continuous learning architecture based on knowledge collaboration. The multi-level is mainly reflected in the "end-edge-cloud" integrated architecture composed of device nodes, edge nodes, and cloud (i.e., cloud servers), such as Figure 1 As shown. Models at all levels are dynamically updated over time based on dynamic data. Considering the flexible scalability of the multi-level architecture for different application environments, the structure can be simplified to a two-level "end-edge" architecture, or expanded to "end-access-edge-core-edge-cloud", etc. The cloud server and edge node are designed to have a consistent structure in terms of core functions.

[0064] The continuous learning model in this embodiment not only obtains data from its own updated database, but also obtains data from the databases of continuous learning models on other devices. Through knowledge collaboration between all levels, knowledge is shared globally to achieve global knowledge collaboration. In related technologies, the learning model on a device node, such as a terminal, only uses its own data for training. The amount of data is not large enough, and the data categories are not rich. The trained model can only be applied to one device node but not to other device nodes, and lacks adaptability to heterogeneous terminal data. In the technical solution of the present application, in addition to using its own data for training, the device node can also update its own structure and model parameters according to the data of the edge model of the edge node, and can train a model suitable for multiple device nodes.

[0065] In some embodiments, the edge node includes:

[0066] Access edge nodes to communicate with the device nodes and core edge nodes respectively;

[0067] The core edge node communicates with the access edge node and the cloud server respectively.

[0068] In a specific embodiment, Figure 1 As shown, the multi-level model system includes a cloud (ie, a cloud server), an edge node, and a device node, where i is the index of the edge node and k is the index of the device node.

[0069] In some embodiments, the end model is used to be updated collaboratively using a dynamically updated end database and edge knowledge of the edge model;

[0070] The edge model is used to be updated by using the end knowledge collaboration of the end model of at least one device node communicating with the edge node and the cloud knowledge collaboration of the cloud model of the cloud server communicating with the edge node;

[0071] The cloud model is used to collaboratively update using edge knowledge of an edge model of at least one edge node in communication with the cloud server.

[0072] The continuous learning model consists of three parts: the continuous learning of the end model, the edge model, and the cloud model, which correspond to three learning mechanisms, such as Figure 2-Figure 4As shown in the figure, on-end continuous learning mainly relies on the continuous perception of accumulated new data and the collaboration of edge models, so that the end model not only retains existing knowledge, but also learns new knowledge through new data, thereby updating the end model and adapting local features in a small and precise manner. Edge continuous learning mainly relies on the new knowledge provided by multiple devices under its jurisdiction and the existing global knowledge of the upper-level cloud model to help the edge model have a more comprehensive vision than a single end model. On-cloud continuous learning is mainly based on end-edge continuous learning. By aggregating edge node knowledge to the cloud, the cloud model has global knowledge and even data, and the most powerful model can be obtained with the support of large-scale computing power.

[0073] Due to the limited computing resources of different device nodes and edge nodes, such as Figure 5 As shown in the figure, the cloud model and edge model need to be compressed according to the computing power and business indicators of the lower-level deployment nodes to meet the time requirements of on-end inference calculations and the training speed of end and edge nodes. Then, the compressed model is automatically deployed to the corresponding node.

[0074] In some embodiments, the cloud model is used to perform model compression based on computing power and business indicators of at least one edge node in communication with the cloud server;

[0075] The edge model is used to perform model compression according to the computing power and business indicators of at least one device node that communicates with the edge node.

[0076] In some embodiments, the device nodes have data perception and data storage capabilities. The device nodes include: vehicle-mounted devices with GPS modules and smart cameras. New data streams can be labeled through automatic annotation to generate training sets, laying a data foundation for continuous learning of on-end models.

[0077] In some embodiments, the device node is used to mark new data streams through automatic labeling to generate a dynamically updated end database.

[0078] In this embodiment, model compression makes efficient inference computing possible on the end, and business analysis and processing can be completed near the data end. On-end inference can reduce the network delay of the backhaul edge and cloud on the one hand, ensuring real-time inference, and on-end inference can effectively protect user privacy on the other hand.

[0079] The embodiment of the present invention also provides a method for constructing a multi-level model system, which is used to construct the multi-level model system as described above. Figure 6 As shown, including:

[0080] Step 101: Generate an initial cloud model based on the cloud data set;

[0081] Step 102: Perform model compression according to edge node computing resources, compress the initial cloud model into an initial edge model, and deploy it to the edge node;

[0082] Step 103: Perform model compression according to the computing resources of the device node, compress the initial edge model into an initial end model, and deploy it to the device node.

[0083] In some embodiments, the method is applied to a device node, comprising:

[0084] After time t, the end training data set is formed by automatically labeling the perception data;

[0085] Request the latest edge model from the edge node, and based on the initial edge model and edge training data set, perform continuous learning of the edge model under the edge knowledge collaboration of the latest edge model to generate a new edge model;

[0086] Push the new end model to the edge node and perform on-end inference calculations based on the new end model.

[0087] In some embodiments, the method is applied to an edge node, including:

[0088] Obtain the latest end model set and the latest end data set of multiple communicating device nodes, wherein the latest end data set includes the latest end training data sets of the multiple device nodes, and request the latest cloud model from the cloud server;

[0089] Based on the latest edge model and the latest aggregated edge data set, the edge model is continuously learned under the cloud knowledge collaboration of the latest cloud model and the end knowledge collaboration of the latest end model set to generate a new edge model;

[0090] Push the new edge model to the cloud server.

[0091] In some embodiments, the method is applied to a cloud server, including:

[0092] Obtain the latest edge model set and the latest edge data set of multiple communicating edge nodes;

[0093] Based on the latest cloud model and the latest aggregated cloud dataset, the cloud model is continuously learned under the edge knowledge collaboration of the latest edge model set to generate a new cloud model.

[0094] In a specific embodiment, the method for constructing a multi-level model system mainly includes 2 parts, 6 stages and 12 specific steps. The details are as follows:

[0095] 1. Model initialization

[0096] Phase 1: Cloud model initialization

[0097] 1. Cloud: Generate the initial cloud model Cm based on the cloud dataset (pre-collected dataset or open source dataset) 0 ; Phase 2: Edge model initialization

[0098] 2. Edge Node E i :New edge node E i Join, request the model from the cloud;

[0099] 3. Cloud: Based on edge nodes E i Computing resources, model compression, cloud model Cm 0 Compressed to the initial edge model Em i 0 , and deployed to edge node E i ;

[0100] Phase 3: End Model Initialization

[0101] 4. Device node D k : New device node D k Join, to the corresponding edge node E i Request model;

[0102] 5. Edge Node E i :According to device node D k Computing resources, model compression, edge model Em i 0 Compressed to the initial end model Dm k 0 , and deploy to device node D k ;

[0103] 6. Device node D k :Based on the end model Dm k 0 Start inference computing on the client.

[0104] 2. Multi-level Continuous Learning

[0105] Phase 4: Continuous learning of end models

[0106] 7. Device node D k :After time t, the end training data set Dd is formed by automatically labeling the perception data k t ;

[0107] 8. Device node D k :To edge node E i Request the latest model Em i 0 , based on the end model Dm k 0 And the end training data set Dd kt , in the edge model Em i 0 The end model is continuously learned under knowledge collaboration to generate a new end model Dm k t ;

[0108] 9. Device node D k :The newly generated end model Dm k t Push to edge node E i At the same time, based on the end model Dm k t Perform inference calculations on the client.

[0109] Phase 5: Continuous learning of edge models

[0110] 10.Edge Node E i : Get the latest end model collection Dms on multiple devices under your jurisdiction i t ={Dm1 t ,Dm2 t ,…,Dm k t} and the latest end data set Dds i t ={Dd1 t ,Dd2 t ,…,Dd k t}. Request the latest model Cm from the cloud 0 , based on the latest edge model Em i 0 and the latest aggregated edge dataset Ed i t = {Ed i 0 ,Dds i t}, in the cloud model Cm 0 , terminal model Dms i t Continuously learn edge models under knowledge collaboration to generate new edge models Em i t ;

[0111] 11.Edge Node E i :The newly generated end model Em i t Push to the cloud;

[0112] Phase 6: Continuous learning of cloud models

[0113] 12. Cloud: Get the latest edge model set Ems on multiple edge nodes under its jurisdictiont ={Em1 t ,Em2 t ,…,Em i t} and the latest edge data collection Eds t ={Ed1 t ,Ed2 t ,…,Ed i t}. Based on the latest cloud model Cm 0 and the latest aggregated cloud dataset Cd t ={Cd 0 ,Eds t}, in the latest edge model collection Ems t Continuously learn cloud models under knowledge collaboration to generate new cloud models Cd t .

[0114] In this embodiment, adaptive enhancement and accuracy optimization of the global model are achieved through multi-level continuous learning partial iteration. This embodiment also supports multiple variants: 1) The multi-level architecture is flexible and scalable, the levels can be increased or decreased, and the nodes can be expanded or reduced; 2) For data privacy protection, it can be degraded from the "model + data" knowledge collaboration mode to a collaboration mode where only the model is used as knowledge.

[0115] The technical solution of this embodiment can be applied to vehicle position prediction. In scenarios such as autonomous driving and assisted driving in connected vehicles, providing V2X (vehicle to everything) services to vehicles and users through vehicle position prediction is a very important capability. Figure 7 As shown in the figure, a solution is given to solve the business continuity problem of MEC (Mobile Edge Computing) in the Internet of Vehicles through the vehicle location prediction algorithm. In the solution, the location prediction algorithm is deployed on each vehicle. In the traditional method, all vehicles use an unchanging prediction model, but due to differences in vehicles and drivers' driving behavior habits, which change over time, the method used to build a multi-level model system is suitable for solving such problems. Through the model adaptation method based on multi-level continuous learning, the individual differences of each vehicle can be fully considered, and the data characteristics of each vehicle can be gradually adapted over time, and the accuracy of the vehicle-side location prediction algorithm will become higher and higher.

[0116] The technical solution of this embodiment can also be applied to abnormal behavior detection. In the smart community elderly care or nursing home environment, real-time detection of abnormal behaviors such as physical abnormalities or sudden dangers of the elderly plays an important role in emergency medical assistance and medical resource preparation. Through video surveillance in indoor or public areas, the behavioral characteristics of the elderly in different environments can be analyzed 24 hours a day. However, the abnormal behavior patterns of different monitoring environments and different individuals are very different and diverse. The method of multi-level continuous learning model adaptation can be applied to abnormal behavior detection, such as Figure 8 As shown, by building smart home camera nodes, home gateway / computing box nodes, and edge nodes in the homes of elderly people living alone, a two-level end-edge multi-node architecture is formed. Smart cameras and gateways / computing boxes can collaborate to complete data perception and automatic labeling, and deploy abnormal behavior detection end models on gateways / computing boxes with more sufficient computing resources to complete on-end reasoning, and use unmanned time to complete end model continuous learning. Considering privacy, only edge continuous learning is performed through the collaboration of end models and edge nodes. The abnormal behavior detection algorithms deployed in different families can gradually learn new abnormal patterns that are adapted to family members through this method, and the detection accuracy and recall rate will be improved.

[0117] The technical solution of this embodiment can be applied in multiple fields. In addition to being used in the field of Internet of Vehicles for vehicle position prediction, driver behavior analysis, automatic driving, etc., it can also be applied to the smart home / elderly care field for abnormal behavior analysis of smart cameras, personalized wake-up word learning of smart speakers, nursing robots, etc. It can also be applied to the field of smart agriculture for drone plant disease and pest identification, or to the field of smart factories for intelligent robot collaboration.

[0118] An embodiment of the present invention further provides a device for constructing a multi-level model system, comprising:

[0119] A cloud model generation module, used to generate an initial cloud model based on a cloud data set;

[0120] The edge model generation module is used to compress the model according to the edge node computing resources, compress the initial cloud model into the initial edge model, and deploy it to the edge node;

[0121] The end model generation module is used to compress the model according to the computing resources of the device node, compress the initial edge model into the initial end model, and deploy it to the device node.

[0122] In some embodiments, the device further comprises:

[0123] The data set formation module is used to form an end training data set by automatically annotating the perception data after time t;

[0124] The first learning module is used to request the latest edge model from the edge node, and based on the initial edge model and the edge training data set, continuously learn the edge model under the edge knowledge collaboration of the latest edge model to generate a new edge model;

[0125] The first push module is used to push the new end model to the edge node and perform on-end reasoning calculation based on the new end model.

[0126] In some embodiments, the device further comprises:

[0127] A first acquisition module is used to acquire the latest end model set and the latest end data set of multiple device nodes in communication, wherein the latest end data set includes the latest end training data sets of multiple device nodes, and request the latest cloud model from the cloud server;

[0128] The second learning module is used to continuously learn the edge model based on the latest edge model and the latest aggregated edge data set, under the cloud knowledge collaboration of the latest cloud model and the end knowledge collaboration of the latest end model set, to generate a new edge model;

[0129] The second push module is used to push the new edge model to the cloud server.

[0130] In some embodiments, the device further comprises:

[0131] A second acquisition module is used to acquire the latest edge model set and the latest edge data set of multiple edge nodes in communication;

[0132] The third learning module is used to perform continuous learning of cloud models based on the latest cloud model and the latest aggregated cloud dataset under the edge knowledge collaboration of the latest edge model set to generate a new cloud model.

[0133] An embodiment of the present invention also provides a device for constructing a multi-level model system, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor; when the processor executes the program, the method for constructing a multi-level model system as described above is implemented.

[0134] An embodiment of the present invention further provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the method for constructing a multi-level model system as described above.

[0135] Computer readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.

[0136] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A multi-level model system, characterized in that: include: A cloud server, wherein a cloud model is arranged on the cloud server; at least one edge node in communication with the cloud server, the edge node having an edge model disposed thereon, the edge node in communication with at least one device node; The at least one device node, wherein the device node is provided with an end model; Among them, the cloud model can update its own structure and model parameters according to the data of at least one of the edge models, the edge model can update its own structure and model parameters according to the data of the cloud model and the end model, and the end model can update its own structure and model parameters according to the data of the edge model.

2. The multi-level model system according to claim 1, characterized in that: The edge node includes: Access edge nodes to communicate with the device nodes and core edge nodes respectively; The core edge node communicates with the access edge node and the cloud server respectively.

3. The multi-level model system according to claim 1, characterized in that: The end model is used to be updated collaboratively using a dynamically updated end database and edge knowledge of the edge model; The edge model is used to be updated by using the end knowledge collaboration of the end model of at least one device node communicating with the edge node and the cloud knowledge collaboration of the cloud model of the cloud server communicating with the edge node; The cloud model is used to collaboratively update using edge knowledge of an edge model of at least one edge node in communication with the cloud server.

4. The multi-level model system according to claim 1, characterized in that: The cloud model is used to perform model compression according to the computing power and business indicators of at least one edge node communicating with the cloud server; The edge model is used to perform model compression according to the computing power and business indicators of at least one device node that communicates with the edge node.

5. The multi-level model system according to claim 1, characterized in that: The device nodes include: vehicle-mounted equipment with a GPS module and a smart camera.

6. The multi-level model system according to claim 1, characterized in that: The device node is used to mark the new data stream through automatic labeling to generate a dynamically updated end database.

7. A method for constructing a multi-level model system, characterized in that: Used to construct a multi-level model system as described in any one of claims 1 to 6, comprising: generating an initial cloud model based on the cloud data set; Based on the computing resources of the edge nodes, the model is compressed, the initial cloud model is compressed into the initial edge model, and then deployed to the edge nodes; According to the computing resources of the device node, the model is compressed, the initial edge model is compressed into the initial end model, and deployed to the device node.

8. The method for constructing a multi-level model system according to claim 7, characterized in that: Applies to device nodes, including: After time t, the end training data set is formed by automatically labeling the perception data; Request the latest edge model from the edge node, and based on the initial edge model and edge training data set, perform continuous learning of the edge model under the edge knowledge collaboration of the latest edge model to generate a new edge model; Push the new end model to the edge node and perform on-end inference calculations based on the new end model.

9. The method for constructing a multi-level model system according to claim 7, characterized in that: Applied to edge nodes, including: Obtain the latest end model set and the latest end data set of multiple communicating device nodes, wherein the latest end data set includes the latest end training data sets of the multiple device nodes, and request the latest cloud model from the cloud server; Based on the latest edge model and the latest aggregated edge data set, the edge model is continuously learned under the cloud knowledge collaboration of the latest cloud model and the end knowledge collaboration of the latest end model set to generate a new edge model; Push the new edge model to the cloud server.

10. The method for constructing a multi-level model system according to claim 7, characterized in that: Applied to cloud servers, including: Obtain the latest edge model set and the latest edge data set of multiple communicating edge nodes; Based on the latest cloud model and the latest aggregated cloud dataset, the cloud model is continuously learned under the edge knowledge collaboration of the latest edge model set to generate a new cloud model.

11. A device for constructing a multi-level model system, characterized in that: Used to construct a multi-level model system as described in any one of claims 1 to 6, comprising: A cloud model generation module, used to generate an initial cloud model based on a cloud data set; The edge model generation module is used to compress the model according to the edge node computing resources, compress the initial cloud model into the initial edge model, and deploy it to the edge node; The end model generation module is used to compress the model according to the computing resources of the device node, compress the initial edge model into the initial end model, and deploy it to the device node.

12. A device for constructing a multi-level model system, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor; characterized in that: When the processor executes the program, the method for constructing a multi-level model system as described in any one of claims 7 to 10 is implemented.

13. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps in the method for constructing a multi-level model system as described in any one of claims 7 to 10 are implemented.

Images