CN111126618A - Multi-source heterogeneous system-based federal learning method and device - Google Patents
Multi-source heterogeneous system-based federal learning method and device Download PDFInfo
- Publication number
- CN111126618A CN111126618A CN201911236454.8A CN201911236454A CN111126618A CN 111126618 A CN111126618 A CN 111126618A CN 201911236454 A CN201911236454 A CN 201911236454A CN 111126618 A CN111126618 A CN 111126618A
- Authority
- CN
- China
- Prior art keywords
- model
- node
- nodes
- isomorphic
- network
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Software Systems (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Medical Informatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Artificial Intelligence (AREA)
- Mobile Radio Communication Systems (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention discloses a federal learning method and a federal learning device based on a multi-source heterogeneous system, wherein the multi-source heterogeneous system comprises a first system and a second system, the first system and the second system are in a heterogeneous relationship, and a plurality of isomorphic nodes of the first system and a plurality of isomorphic nodes of the second system respectively acquire data with different dimensions for a same object; the method comprises the following steps: the method comprises the steps of obtaining model parameters of a first system sub-model obtained through federal learning of a local model of each isomorphic node of a first system, obtaining model parameters of a second system sub-model obtained through federal learning of a local model of each isomorphic node of a second system, and obtaining a federal model of the multi-source heterogeneous system through federal learning of the model parameters of the first system sub-model and the model parameters of the second system sub-model. The technical scheme is used for reducing the processing complexity of the multi-source heterogeneous network during federal learning and improving the learning efficiency of the whole model.
Description
Technical Field
The embodiment of the invention relates to the field of financial technology (Fintech), in particular to a federal learning method and a federal learning device based on a multi-source heterogeneous system.
Background
With the development of computer technology, more and more technologies are applied in the financial field, the traditional financial industry is gradually changing to financial science and technology, and machine learning technology is no exception, but higher requirements are also provided for the machine learning technology due to the requirements of security and real-time performance of the financial and payment industries.
Federal learning refers to a method of machine learning by federating different participants. When the federal learning network is a multi-source heterogeneous network, due to the fact that the performance of each network node is different, the problem that the processing of the federal learning process is complex and the efficiency is low exists. For example, the satellite and the mobile phone form a multi-source heterogeneous network for federal learning, the data transmitted by the satellite as a model participant is data corresponding to the satellite, the data transmitted by the mobile phone as a model participant is data corresponding to the mobile phone, the formats of the data transmitted by the satellite and the mobile phone are different, format conversion is required to be performed firstly when the satellite and the mobile phone perform data interaction, the processing capacity of the satellite and the processing capacity of the mobile phone are greatly different, the model effects of the two local training models are different, and the rates of the training models are also different. The method is characterized in that different types of network nodes are used as participants to conduct federal learning, and problems of format conversion, different processing capabilities and the like are involved in the learning process, so that the multi-source heterogeneous network is complex in processing and low in efficiency when federal learning is conducted.
Disclosure of Invention
The embodiment of the invention provides a federal learning method and a federal learning device based on a multi-source heterogeneous system, which are used for reducing the processing complexity of a multi-source heterogeneous network during federal learning and improving the learning efficiency of an overall model.
The embodiment of the invention provides a federal learning method based on a multi-source heterogeneous system, wherein the multi-source heterogeneous system comprises a first system with a plurality of homogeneous nodes and a second system with a plurality of homogeneous nodes; the first system and the second system are in a heterogeneous relationship; the plurality of isomorphic nodes of the first system and the plurality of isomorphic nodes of the second system respectively acquire data with different dimensions for the same object;
the method comprises the following steps:
obtaining model parameters of a first system sub-model obtained by performing federal learning on local models of isomorphic nodes of the first system, wherein the local models of the isomorphic nodes of the first system are obtained by training the isomorphic nodes of the first system according to acquired data of the local models;
obtaining model parameters of a second system sub-model obtained by performing federal learning on local models of all isomorphic nodes of the second system; the local model of each isomorphic node of the second system is obtained by training each isomorphic node of the second system according to the acquired data of the local model;
and carrying out federal learning through the model parameters of the first system submodel and the model parameters of the second system submodel to obtain a federal model of the multi-source heterogeneous system.
In the technical scheme, in the federal learning process of the multi-source heterogeneous system, the isomorphic nodes of the same type form a first system, the isomorphic nodes of the other same type form a second system, federal learning is carried out inside the first system to obtain a submodel of the first system, federal learning is carried out inside the second system to obtain a submodel of the second system, and for each system, because the network nodes inside the system are the isomorphic nodes of the same type, the problem of different performances among the network nodes does not exist, namely the problem of overlarge model effect or training speed difference of local training models without format conversion or with different network nodes is solved. Furthermore, the first system and the second system are used as network nodes for federal learning, so that data interaction among different types of network nodes can be greatly reduced, a large amount of data processing caused by different node types in the system is avoided, the processing complexity of the multisource heterogeneous network during federal learning is reduced, and the learning efficiency of the whole model is improved. In addition, when the first system and the second system are used as network nodes for federal learning, model parameters are transmitted between the first system and the second system, and compared with training data, the data volume transmitted between the first system and the second system is small, so that the data processing complexity of the system is further reduced.
Optionally, the method further includes:
dividing a plurality of network nodes according to the node types; the node types of the plurality of network nodes comprise a first type and a second type;
determining a network node of which the node type belongs to the first type as a homogeneous node in the first system;
determining a network node of which the node type belongs to the second type as a homogeneous node in the second system;
the node type of the network node is determined according to the performance of the network node, and the performance of the network node at least comprises one or a combination of the following: processing power, storage power, data transmission format.
In the above technical solution, when performing federated learning, each node may be divided into the first system and the second system according to the performance of each node constituting federated learning, so that the node in the first system is an isomorphic node, the node in the second system is also an isomorphic node, and the first system and the second system are in a heterogeneous relationship. Optionally, the multi-source heterogeneous system is established based on a delay tolerant network;
the method further comprises the following steps:
sending a first model parameter update to a first system through a communication link of a first delay tolerant network of the first system, and if the sending fails, caching the first model parameter update to a cache layer of the first delay tolerant network; after communication link connection of the first delay tolerant network is detected, the first model updating parameters in the cache layer of the first delay tolerant network are sent to the first system; and/or
Sending a second model parameter update to a second system through a communication link of a second delay tolerant network of the second system, and if the sending fails, caching the second model parameter update to a cache layer of the second delay tolerant network; and after communication link connection of the second delay tolerant network is detected, sending the second model updating parameters in the cache layer of the second delay tolerant network to the second system.
In the technical scheme, the multi-source heterogeneous system is established based on the delay tolerant network, so that the data packets which are failed to be transmitted due to network interruption are effectively cached, and the complete transmission of data is guaranteed.
Optionally, the method further includes:
acquiring first acquisition data of a plurality of isomorphic nodes in the first system aiming at the same object;
acquiring second acquisition data of a plurality of isomorphic nodes in the second system aiming at the same object;
and inputting the first collected data and the second collected data of the same object into the federal model to obtain a predicted value of the same object.
In the technical scheme, the first acquisition data are acquired by a plurality of isomorphic nodes of the first system, the second acquisition data are acquired by a plurality of isomorphic nodes of the second system, and the first system and the second system are in a heterogeneous relationship, so that the first acquisition data and the second acquisition data of the same object acquired by the first system and the second system are of different dimensions, and the first acquisition data and the second acquisition data are input into the federal model together for prediction, so that the prediction accuracy of the model on the object can be improved.
Optionally, the plurality of isomorphic nodes of the first system are a plurality of satellite nodes in a satellite system; the second system comprises a plurality of ground nodes in at least one ground system; the ground systems in the at least one ground system are in a heterogeneous relationship.
In the technical scheme, the multi-source heterogeneous system comprises a satellite system and a ground system, the satellite system can acquire large-range data of the same object, the ground system can acquire small-range data of the same object, the small-range data and the small-range data are combined to determine relatively comprehensive data of the same object, and therefore the acquired model can further improve the accuracy of prediction of the object.
Optionally, the ground system is an internet of things system, a mobile phone system or an unmanned aerial vehicle system.
In the technical scheme, the ground system is an internet of things system or a mobile phone system or an unmanned aerial vehicle system, the data dimensionalities of the same object acquired by different ground systems are different, the comprehensiveness of the acquired data of the same object is further improved, and therefore the prediction accuracy of the object can be further improved through the acquired model.
Correspondingly, the embodiment of the invention also provides a federal learning device based on the multi-source heterogeneous system, wherein the multi-source heterogeneous system comprises a first system with a plurality of homogeneous nodes and a second system with a plurality of homogeneous nodes; the first system and the second system are in a heterogeneous relationship; the plurality of isomorphic nodes of the first system and the plurality of isomorphic nodes of the second system respectively acquire data with different dimensions for the same object;
the device comprises:
the acquiring unit is used for acquiring model parameters of a first system sub-model obtained by federate learning through local models of isomorphic nodes of the first system, wherein the local models of the isomorphic nodes of the first system are obtained by training the isomorphic nodes of the first system according to acquired data of the local models;
the obtaining unit is further configured to obtain model parameters of a second system sub-model obtained through federal learning of local models of each isomorphic node of the second system; the local model of each isomorphic node of the second system is obtained by training each isomorphic node of the second system according to the acquired data of the local model;
and the processing unit is used for carrying out federal learning through the model parameters of the first system submodel and the model parameters of the second system submodel to obtain a federal model of the multi-source heterogeneous system.
Optionally, the processing unit is further configured to:
dividing a plurality of network nodes according to the node types; the node types of the plurality of network nodes comprise a first type and a second type;
determining a network node of which the node type belongs to the first type as a homogeneous node in the first system;
determining a network node of which the node type belongs to the second type as a homogeneous node in the second system;
the node type of the network node is determined according to the performance of the network node, and the performance of the network node at least comprises one or a combination of the following: processing power, storage power, data transmission format.
Optionally, the multi-source heterogeneous system is established based on a delay tolerant network;
the processing unit is further to:
sending a first model parameter update to a first system through a communication link of a first delay tolerant network of the first system, and if the sending fails, caching the first model parameter update to a cache layer of the first delay tolerant network; after communication link connection of the first delay tolerant network is detected, the first model updating parameters in the cache layer of the first delay tolerant network are sent to the first system; and/or
Sending a second model parameter update to a second system through a communication link of a second delay tolerant network of the second system, and if the sending fails, caching the second model parameter update to a cache layer of the second delay tolerant network; and after communication link connection of the second delay tolerant network is detected, sending the second model updating parameters in the cache layer of the second delay tolerant network to the second system.
Optionally, the obtaining unit is further configured to:
acquiring first acquisition data of a plurality of isomorphic nodes in the first system aiming at the same object;
acquiring second acquisition data of a plurality of isomorphic nodes in the second system aiming at the same object;
the processing unit is further to:
and inputting the first collected data and the second collected data of the same object into the federal model to obtain a predicted value of the same object.
Optionally, the plurality of isomorphic nodes of the first system are a plurality of satellite nodes in a satellite system; the second system comprises a plurality of ground nodes in at least one ground system; the ground systems in the at least one ground system are in a heterogeneous relationship.
Optionally, the ground system is an internet of things system, a mobile phone system or an unmanned aerial vehicle system.
Correspondingly, an embodiment of the present invention further provides a computing device, including:
a memory for storing program instructions;
and the processor is used for calling the program instruction stored in the memory and executing the federated learning method based on the multi-source heterogeneous system according to the obtained program.
Correspondingly, the embodiment of the invention also provides a computer-readable non-volatile storage medium, which comprises computer-readable instructions, and when the computer reads and executes the computer-readable instructions, the computer is enabled to execute the above federated learning method based on the multi-source heterogeneous system.
Drawings
In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings needed to be used in the description of the embodiments will be briefly introduced below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings based on these drawings without creative efforts.
Fig. 1 is a schematic diagram of a system architecture according to an embodiment of the present invention;
FIG. 2 is a diagram of another system architecture according to an embodiment of the present invention;
fig. 3 is a schematic flow chart of a federated learning method provided in the embodiment of the present invention;
fig. 4 is a schematic structural diagram of a bang learning device according to an embodiment of the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention clearer, the present invention will be described in further detail with reference to the accompanying drawings, and it is apparent that the described embodiments are only a part of the embodiments of the present invention, not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
The embodiment of the invention provides a system architecture applicable to a federated learning method based on a multi-source heterogeneous system, which can be the multi-source heterogeneous system, wherein the multi-source heterogeneous system comprises a first system with a plurality of homogeneous nodes and a second system with a plurality of homogeneous nodes, the first system and the second system are in a heterogeneous relationship, and the plurality of homogeneous nodes of the first system and the plurality of homogeneous nodes of the second system respectively acquire data of different dimensions for the same object.
The data transmission formats of the first system and the second system are different, and the data transmission formats of the plurality of nodes in the first system are the same, so that the plurality of nodes in the first system are in a homogeneous relationship.
Fig. 1 exemplarily illustrates a multi-source heterogeneous system, wherein the multi-source heterogeneous system includes a first system and a second system, wherein the first system may be a satellite system, and the satellite system is composed of a plurality of satellite nodes; the second system may include at least one ground system, and the ground system may be an internet of things system or a mobile phone system or an unmanned aerial vehicle system or the like, specifically, the internet of things system is composed of a plurality of internet of things nodes, the mobile phone system is composed of a plurality of mobile phone nodes, and the unmanned aerial vehicle system is composed of a plurality of unmanned aerial vehicle nodes. The ground systems in the at least one ground system are in a heterogeneous relationship.
And each network node respectively performs federal learning in the system where the network node is positioned, and determines a federal learning submodel. Taking a satellite system as an example, federal learning is carried out among satellite nodes, so that satellite submodels are determined. Furthermore, a satellite system, an internet of things system, a mobile phone system and an unmanned aerial vehicle system are combined to conduct federal learning, and a federal model of the multi-source heterogeneous system is determined.
Furthermore, a delay tolerant network can be adopted for communication among the satellite nodes and between the satellite nodes and the ground node, a cache layer (Bundle layer) is added in the delay tolerant network, data packets which are failed to be transmitted due to network interruption are effectively cached, and the data packets are continuously transmitted after a link is connected.
In the embodiment of the present invention, another multi-source heterogeneous system is further provided, which may also be called space-Temporal Federal Learning (STFL), as shown in FIG. 2. And the network nodes in each system locally perform modeling calculation and send the generated model parameters to the coordinator of the system, so that the submodel corresponding to the system is established, and each system sends the model parameters of the trained submodel to the coordinator of the whole federal learning system, so that the federal model between the systems is trained. The processing capacity, the storage capacity, the data transmission format and other performances of different types of network nodes are greatly different, so that the same type of nodes form a Sub-federated network (Sub-federated network), the Sub-model construction is completed by utilizing the horizontal federated learning, the output of the Sub-model participates in the top-level federated learning after being encrypted, the data interaction among subsystems can be reduced by the mode, the frequent data transmission and conversion among the subsystems are avoided, and the rapid and effective establishment of the federated model of the multi-source heterogeneous system is realized.
Based on the above description, fig. 3 exemplarily shows a flow of a federated learning method based on a multi-source heterogeneous system according to an embodiment of the present invention, where the flow may be executed by a federated learning apparatus based on a multi-source heterogeneous system.
As shown in fig. 3, the process specifically includes:
The local model of each isomorphic node of the first system is obtained by training each isomorphic node of the first system according to the acquired data of the local model. And the local model of each isomorphic node is subjected to federal learning to obtain a first system submodel.
Optionally, the plurality of isomorphic nodes of the first system are a plurality of satellite nodes in the satellite system, and each satellite node performs local training according to the acquired data, which is equivalent to performing in-orbit calculation on each satellite node, so as to solve the problem of timeliness of data transmission between satellites. Local models are trained by all satellite nodes, and then the local models are combined to train the models, so that the limitation of data acquisition of each satellite (the limitation of an observation visual angle and a window at the same moment) can be effectively overcome. In the process of establishing the federal learning model between satellites, satellite data is not transmitted, and only encryption parameters after local model training are transmitted, so that the data transmission quantity is reduced, and data sharing between the satellites is ensured.
Each satellite node can acquire data with the same dimensionality and can also acquire data with different dimensionalities. In one implementation mode, each satellite node collects data of different dimensions of the same observation area so as to improve the richness of information of the observation area, and by taking a crop estimated production scene in a certain area as an example, a large-range crop observation is firstly obtained through a low-resolution satellite, and the total area and the macro growth are calculated; and then, specific calculation is carried out on certain local observation growth vigor, water, fertilizer and disasters of the crops in the area through a high-resolution satellite. And acquiring an image picture of a radar wave band by cooperating with the SAR satellite under the condition of cloud to perform local estimation. Each type of satellite is locally modeled by using the acquired information, and then a satellite sub-model is established by using an interplanetary network to cooperate with a result (model parameter) of local processing.
When the satellite system establishes a satellite sub-model (a first system sub-model), a sink node may be set, and the sink node is configured to establish the satellite sub-model according to a sink result, where the sink node may be a satellite node in the satellite system or a ground receiving station.
And 302, obtaining model parameters of the second system submodel obtained by performing federal learning on the local model of each isomorphic node of the second system.
And the local model of each isomorphic node of the second system is obtained by training each isomorphic node of the second system according to the acquired data of the local model.
The ground system can be an internet of things system, a mobile phone system or an unmanned aerial vehicle system, for example, when the ground system is the internet of things system, the temperature and humidity information can be obtained and an internet of things submodel can be established; when the ground system is a mobile phone system, a pest and disease damage picture photographed by a mobile phone can be obtained and a mobile phone sub-model is established; when the ground system is an unmanned aerial vehicle system, the crop growth picture photographed by the unmanned aerial vehicle can be acquired, and the unmanned aerial vehicle sub-model is established.
And 303, performing federal learning through the model parameters of the first system submodel and the model parameters of the second system submodel to obtain a federal model of the multi-source heterogeneous system.
And performing federal learning between the systems through the model parameters of the first system submodel and the model parameters of the second system submodel, and determining a federal model of the multi-source heterogeneous system.
In the embodiment of the invention, the sink node can be a ground receiving station, not only can receive the model parameters of the satellite submodel determined by the satellite system, but also can receive the model parameters of the internet of things submodel determined by the internet of things system, the model parameters of the mobile phone system submodel determined by the mobile phone system and the model parameters of the unmanned aerial vehicle system submodel determined by the unmanned aerial vehicle system, and further performs the federal model training of the multi-source heterogeneous system according to the model parameters sent by each system.
In addition, the multi-source heterogeneous system can be established based on a delay tolerant network, and can be established among systems or established in the system. In the specific model training, sending a first model parameter update to a first system through a communication link of a first delay tolerant network of the first system, and if the sending fails, caching the first model parameter update to a cache layer of the first delay tolerant network; after communication link connection of the first delay tolerant network is detected, a first model updating parameter in a cache layer of the first delay tolerant network is sent to a first system; and/or sending a second model parameter update to the second system through a communication link of a second delay tolerant network of the second system, and if the sending fails, caching the second model parameter update to a cache layer of the second delay tolerant network; and after communication link connection of the second delay tolerant network is detected, sending a second model updating parameter in a cache layer of the second delay tolerant network to the second system.
In one implementation, the first system may be a satellite system, and due to the unstable connection of the satellite network, the link related to the satellite data transmission may be established based on a delay tolerant network, such as a delay tolerant network between a satellite and a satellite, a delay tolerant network between a satellite and a ground receiver, and the like. In the specific model training, model parameter updating can be sent to the satellite system through a communication link of the delay tolerant network, if the sending fails, the model parameter updating is cached to a cache layer of the delay tolerant network, and after the communication link is detected to be communicated, the model updating parameters in the cache layer are sent to the satellite system.
After the federal model training is completed, prediction can be performed according to the trained federal model, specifically, first collected data of a plurality of isomorphic nodes in a first system for the same object is obtained, second collected data of a plurality of isomorphic nodes in a second system for the same object is obtained, and the first collected data and the second collected data of the same object are input into the federal model to obtain a predicted value of the same object.
In the embodiment of the present invention, the plurality of homogeneous nodes in the first system and the plurality of homogeneous nodes in the second system may be pre-divided, may be manually divided by a worker according to the performance of each of the plurality of network nodes in the system, or may be automatically divided by the system according to the performance of each of the plurality of network nodes. When the system is automatically divided, the type of each network node may be determined according to the performance of each network node, wherein the performance of the network node may include at least one or a combination of the following: the processing capability, the storage capability and the data transmission format may determine the types of the plurality of network nodes as a first type and a second type, that is, the types of the plurality of network nodes include the first type and the second type, and further may divide the plurality of network nodes according to the node types, determine the network node of which the node type belongs to the first type as a homogeneous node of a first system, and determine the network node of which the node type belongs to the second type as a homogeneous node of a second system.
In order to better explain the embodiment of the present invention, the federal learning procedure based on the multi-source heterogeneous system will be described in a specific example.
Example one:
analyzing the rural economic development index using the STFL and spatio-temporal data is a very typical and socially valuable case. First, the accessibility of rural roads is one of the important indicators reflecting the economic development thereof. And constructing a satellite federal learning network through high-resolution series remote sensing satellite networking. And modeling a road in each satellite node by the remote sensing image to extract a sub-model. Each sub-model is located in different high-score satellites, and encryption parameters are interacted through the delay tolerant network, so that each sub-model is cooperated. Secondly, social diversity is also an important index reflecting the economic activity of the region. The mobile phones of all farmers build federal study through a telecommunication network. Each mobile phone node constructs a local model, and model parameters are exchanged among the mobile phone nodes through a telecommunication network, so that a social diversity result of the region is obtained. And collecting crop information by the nodes of the Internet of things, and constructing an estimated yield model locally. Federal learning is established among nodes of the Internet of things through near-field radio, and each node is trained through an interactive encryption parameter collaborative model. And finally, exchanging training parameters through federal learning by the sub-models constructed by the different types of nodes so as to construct and complete the rural economic development index model. It can be seen that the model is constructed based on the STFL, so that the heterogeneous data mining performance can be guaranteed, and the data privacy safety can be protected.
Example two: federal learning can solve the problem of information differences between different rivet points observed by the same satellite. For example, the tolerance of different features to atmospheric pollution. And establishing a local model for each satellite according to data of different landforms captured at the same time period to analyze the correlation between the current landform and the atmospheric pollution, and then obtaining overall information through a sink node to construct a correlation model between the landform and the atmospheric pollution. And an atmospheric pollution source analysis model can be further constructed by combining a ground Internet of things system and a mobile phone system.
In the above embodiment, in the federal learning process of the multi-source heterogeneous system, the isomorphic nodes of the same type form a first system, and the isomorphic nodes of the other same type form a second system, federal learning is performed inside the first system to obtain a sub-model of the first system, and federal learning is performed inside the second system to obtain a sub-model of the second system. Furthermore, the first system and the second system are used as network nodes for federal learning, so that data interaction among different types of network nodes can be greatly reduced, a large amount of data processing caused by different node types in the system is avoided, the processing complexity of the multisource heterogeneous network during federal learning is reduced, and the learning efficiency of the whole model is improved. In addition, when the first system and the second system are used as network nodes for federal learning, model parameters are transmitted between the first system and the second system, and compared with training data, the data volume transmitted between the first system and the second system is small, so that the data processing complexity of the system is further reduced.
Based on the same inventive concept, fig. 4 exemplarily shows a structure of a federated learning apparatus based on a multi-source heterogeneous system provided in an embodiment of the present invention, where the multi-source heterogeneous system includes a first system having a plurality of homogeneous nodes and a second system having a plurality of homogeneous nodes; the first system and the second system are in a heterogeneous relationship; the plurality of isomorphic nodes of the first system and the plurality of isomorphic nodes of the second system respectively acquire data with different dimensions for the same object; the apparatus may perform a process of a federated learning method based on a multi-source heterogeneous system.
The device comprises:
an obtaining unit 401, configured to obtain a model parameter of a first system sub-model obtained through federate learning of a local model of each isomorphic node of the first system, where the local model of each isomorphic node of the first system is obtained through training of each isomorphic node of the first system according to acquired data of the local model;
the obtaining unit 401 is further configured to obtain model parameters of a second system sub-model obtained through federal learning of a local model of each isomorphic node of the second system; the local model of each isomorphic node of the second system is obtained by training each isomorphic node of the second system according to the acquired data of the local model;
and the processing unit 402 is configured to perform federal learning on the model parameters of the first system sub-model and the model parameters of the second system sub-model to obtain a federal model of the multi-source heterogeneous system.
Optionally, the processing unit 402 is further configured to:
dividing a plurality of network nodes according to the node types; the node types of the plurality of network nodes comprise a first type and a second type;
determining a network node of which the node type belongs to the first type as a homogeneous node in the first system;
determining a network node of which the node type belongs to the second type as a homogeneous node in the second system;
the node type of the network node is determined according to the performance of the network node, and the performance of the network node at least comprises one or a combination of the following: processing power, storage power, data transmission format.
Optionally, the multi-source heterogeneous system is established based on a delay tolerant network;
the processing unit 402 is further configured to:
sending a first model parameter update to a first system through a communication link of a first delay tolerant network of the first system, and if the sending fails, caching the first model parameter update to a cache layer of the first delay tolerant network; after communication link connection of the first delay tolerant network is detected, the first model updating parameters in the cache layer of the first delay tolerant network are sent to the first system; and/or
Sending a second model parameter update to a second system through a communication link of a second delay tolerant network of the second system, and if the sending fails, caching the second model parameter update to a cache layer of the second delay tolerant network; and after communication link connection of the second delay tolerant network is detected, sending the second model updating parameters in the cache layer of the second delay tolerant network to the second system.
Optionally, the obtaining unit 401 is further configured to:
acquiring first acquisition data of a plurality of isomorphic nodes in the first system aiming at the same object;
acquiring second acquisition data of a plurality of isomorphic nodes in the second system aiming at the same object;
the processing unit 402 is further configured to:
and inputting the first collected data and the second collected data of the same object into the federal model to obtain a predicted value of the same object.
Optionally, the plurality of isomorphic nodes of the first system are a plurality of satellite nodes in a satellite system; the second system comprises a plurality of ground nodes in at least one ground system; the ground systems in the at least one ground system are in a heterogeneous relationship.
Optionally, the ground system is an internet of things system, a mobile phone system or an unmanned aerial vehicle system.
Based on the same inventive concept, an embodiment of the present invention further provides a computing device, including:
a memory for storing program instructions;
and the processor is used for calling the program instruction stored in the memory and executing the federated learning method based on the multi-source heterogeneous system according to the obtained program.
Based on the same inventive concept, the embodiment of the invention also provides a computer-readable non-volatile storage medium, which comprises computer-readable instructions, and when the computer reads and executes the computer-readable instructions, the computer is enabled to execute the federal learning method based on the multi-source heterogeneous system.
The present invention is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each flow and/or block of the flow diagrams and/or block diagrams, and combinations of flows and/or blocks in the flow diagrams and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
While preferred embodiments of the present invention have been described, additional variations and modifications in those embodiments may occur to those skilled in the art once they learn of the basic inventive concepts. Therefore, it is intended that the appended claims be interpreted as including preferred embodiments and all such alterations and modifications as fall within the scope of the invention.
It will be apparent to those skilled in the art that various changes and modifications may be made in the present invention without departing from the spirit and scope of the invention. Thus, if such modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include such modifications and variations.
Claims (10)
1. A federal learning method based on a multi-source heterogeneous system is characterized in that the multi-source heterogeneous system comprises a first system with a plurality of homogeneous nodes and a second system with a plurality of homogeneous nodes; the first system and the second system are in a heterogeneous relationship; the plurality of isomorphic nodes of the first system and the plurality of isomorphic nodes of the second system respectively acquire data with different dimensions for the same object;
the method comprises the following steps:
obtaining model parameters of a first system sub-model obtained by performing federal learning on a local model of each isomorphic node of the first system; the local model of each isomorphic node of the first system is obtained by training each isomorphic node of the first system according to the acquired data of the local model;
obtaining model parameters of a second system sub-model obtained by performing federal learning on local models of all isomorphic nodes of the second system; the local model of each isomorphic node of the second system is obtained by training each isomorphic node of the second system according to the acquired data of the local model;
and carrying out federal learning through the model parameters of the first system submodel and the model parameters of the second system submodel to obtain a federal model of the multi-source heterogeneous system.
2. The method of claim 1, wherein the method further comprises:
dividing a plurality of network nodes according to the node types; the node types of the plurality of network nodes comprise a first type and a second type;
determining a network node of which the node type belongs to the first type as a homogeneous node in the first system;
determining a network node of which the node type belongs to the second type as a homogeneous node in the second system;
the node type of the network node is determined according to the performance of the network node, and the performance of the network node at least comprises one or a combination of the following: processing power, storage power, data transmission format.
3. The method of claim 1, wherein the multi-source heterogeneous system is established based on a delay-tolerant network;
the method further comprises the following steps:
sending a first model parameter update to a first system through a communication link of a first delay tolerant network of the first system, and if the sending fails, caching the first model parameter update to a cache layer of the first delay tolerant network; after communication link connection of the first delay tolerant network is detected, the first model updating parameters in the cache layer of the first delay tolerant network are sent to the first system; and/or
Sending a second model parameter update to a second system through a communication link of a second delay tolerant network of the second system, and if the sending fails, caching the second model parameter update to a cache layer of the second delay tolerant network; and after communication link connection of the second delay tolerant network is detected, sending the second model updating parameters in the cache layer of the second delay tolerant network to the second system.
4. The method of claim 1, wherein the method further comprises:
acquiring first acquisition data of a plurality of isomorphic nodes in the first system aiming at the same object;
acquiring second acquisition data of a plurality of isomorphic nodes in the second system aiming at the same object;
and inputting the first collected data and the second collected data of the same object into the federal model to obtain a predicted value of the same object.
5. The method of any one of claims 1 to 4, wherein the plurality of homogeneous nodes of the first system are a plurality of satellite nodes in a satellite system; the second system comprises a plurality of ground nodes in at least one ground system; the ground systems in the at least one ground system are in a heterogeneous relationship.
6. The method of claim 5, wherein the ground system is an Internet of things system or a cell phone system or a drone system.
7. The federal learning device based on the multi-source heterogeneous system is characterized in that the multi-source heterogeneous system comprises a first system with a plurality of homogeneous nodes and a second system with a plurality of homogeneous nodes; the first system and the second system are in a heterogeneous relationship; the plurality of isomorphic nodes of the first system and the plurality of isomorphic nodes of the second system respectively acquire data with different dimensions for the same object;
the device comprises:
the acquiring unit is used for acquiring model parameters of a first system sub-model obtained by federate learning through local models of isomorphic nodes of the first system, wherein the local models of the isomorphic nodes of the first system are obtained by training the isomorphic nodes of the first system according to acquired data of the local models;
the obtaining unit is further configured to obtain model parameters of a second system sub-model obtained through federal learning of local models of each isomorphic node of the second system; the local model of each isomorphic node of the second system is obtained by training each isomorphic node of the second system according to the acquired data of the local model;
and the processing unit is used for carrying out federal learning through the model parameters of the first system submodel and the model parameters of the second system submodel to obtain a federal model of the multi-source heterogeneous system.
8. The apparatus as recited in claim 7, said processing unit to further:
dividing a plurality of network nodes according to the node types; the node types of the plurality of network nodes comprise a first type and a second type;
determining a network node of which the node type belongs to the first type as a homogeneous node in the first system;
determining a network node of which the node type belongs to the second type as a homogeneous node in the second system;
the node type of the network node is determined according to the performance of the network node, and the performance of the network node at least comprises one or a combination of the following: processing power, storage power, data transmission format.
9. A computing device, comprising:
a memory for storing program instructions;
a processor for calling program instructions stored in said memory to execute the method of any one of claims 1 to 6 in accordance with the obtained program.
10. A computer-readable non-transitory storage medium including computer-readable instructions which, when read and executed by a computer, cause the computer to perform the method of any one of claims 1 to 6.
Priority Applications (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911236454.8A CN111126618B (en) | 2019-12-05 | 2019-12-05 | Federal learning method and device based on multi-source heterogeneous system |
| PCT/CN2020/112976 WO2021109647A1 (en) | 2019-12-05 | 2020-09-02 | Federated learning method and apparatus based on multi-source heterogeneous system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911236454.8A CN111126618B (en) | 2019-12-05 | 2019-12-05 | Federal learning method and device based on multi-source heterogeneous system |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN111126618A true CN111126618A (en) | 2020-05-08 |
| CN111126618B CN111126618B (en) | 2023-08-04 |
Family
ID=70496183
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201911236454.8A Active CN111126618B (en) | 2019-12-05 | 2019-12-05 | Federal learning method and device based on multi-source heterogeneous system |
Country Status (2)
| Country | Link |
|---|---|
| CN (1) | CN111126618B (en) |
| WO (1) | WO2021109647A1 (en) |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112087766A (en) * | 2020-08-24 | 2020-12-15 | 浙江大学 | Unmanned system heterogeneous network communication channel access method and device |
| CN112286752A (en) * | 2020-09-29 | 2021-01-29 | 深圳致星科技有限公司 | Algorithm verification method and system for federated learning heterogeneous processing system |
| WO2021109647A1 (en) * | 2019-12-05 | 2021-06-10 | 深圳前海微众银行股份有限公司 | Federated learning method and apparatus based on multi-source heterogeneous system |
| CN113160021A (en) * | 2021-03-18 | 2021-07-23 | 天津中科物联科技有限公司 | Safe production early warning system based on multi-source heterogeneous data federal learning |
| CN113743677A (en) * | 2021-09-16 | 2021-12-03 | 成都数融科技有限公司 | Personal credit evaluation model training method and evaluation method based on federal learning |
| CN114090983A (en) * | 2022-01-24 | 2022-02-25 | 亿景智联(北京)科技有限公司 | Heterogeneous federated learning platform communication method and device |
| CN116362352A (en) * | 2023-06-01 | 2023-06-30 | 广州思迈特软件有限公司 | Model automatic updating method, system, medium and terminal based on machine learning |
Families Citing this family (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114221957A (en) * | 2021-11-30 | 2022-03-22 | 中国电子科技网络信息安全有限公司 | Country management system |
| CN114385376B (en) * | 2021-12-09 | 2024-05-31 | 北京理工大学 | Client selection method for federal learning of lower edge side of heterogeneous data |
| CN114611718B (en) * | 2022-02-16 | 2024-10-29 | 爱动超越人工智能科技(苏州)有限责任公司 | Federal learning method and system for heterogeneous data |
| CN117592555B (en) * | 2023-11-28 | 2024-05-10 | 中国医学科学院北京协和医院 | Federal learning method and system for multi-source heterogeneous medical data |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1841415A (en) * | 2005-03-29 | 2006-10-04 | 上海宝信软件股份有限公司 | Data transmission method employing data soft bus for application system of metallurgic enterprise |
| CN102238089A (en) * | 2011-07-01 | 2011-11-09 | 中兴通讯股份有限公司 | Service interaction method, device and system |
| US20120023062A1 (en) * | 2010-01-15 | 2012-01-26 | Telcordia Technologies, Inc. | Robust information fusion methods for decision making for multisource data |
| CN104932895A (en) * | 2015-06-26 | 2015-09-23 | 南京邮电大学 | Middleware based on SOA (Service-Oriented Architecture) and information publishing method thereof |
| CN107579845A (en) * | 2017-08-20 | 2018-01-12 | 西南电子技术研究所（中国电子科技集团公司第十研究所） | Spatial information web services architectural framework |
| CN110266771A (en) * | 2019-05-30 | 2019-09-20 | 天津神兔未来科技有限公司 | Distributed intelligence node and distributed swarm intelligence system dispositions method |
Family Cites Families (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20040128541A1 (en) * | 2002-12-31 | 2004-07-01 | Iinternational Business Machines Corporation | Local architecture for federated heterogeneous system |
| CN110377648A (en) * | 2018-04-11 | 2019-10-25 | 西安邮电大学 | A kind of multi-source heterogeneous Data Analysis Platform towards intelligence manufacture |
| US11423254B2 (en) * | 2019-03-28 | 2022-08-23 | Intel Corporation | Technologies for distributing iterative computations in heterogeneous computing environments |
| CN111126618B (en) * | 2019-12-05 | 2023-08-04 | 深圳前海微众银行股份有限公司 | Federal learning method and device based on multi-source heterogeneous system |
-
2019
- 2019-12-05 CN CN201911236454.8A patent/CN111126618B/en active Active
-
2020
- 2020-09-02 WO PCT/CN2020/112976 patent/WO2021109647A1/en active Application Filing
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1841415A (en) * | 2005-03-29 | 2006-10-04 | 上海宝信软件股份有限公司 | Data transmission method employing data soft bus for application system of metallurgic enterprise |
| US20120023062A1 (en) * | 2010-01-15 | 2012-01-26 | Telcordia Technologies, Inc. | Robust information fusion methods for decision making for multisource data |
| CN102238089A (en) * | 2011-07-01 | 2011-11-09 | 中兴通讯股份有限公司 | Service interaction method, device and system |
| CN104932895A (en) * | 2015-06-26 | 2015-09-23 | 南京邮电大学 | Middleware based on SOA (Service-Oriented Architecture) and information publishing method thereof |
| CN107579845A (en) * | 2017-08-20 | 2018-01-12 | 西南电子技术研究所（中国电子科技集团公司第十研究所） | Spatial information web services architectural framework |
| CN110266771A (en) * | 2019-05-30 | 2019-09-20 | 天津神兔未来科技有限公司 | Distributed intelligence node and distributed swarm intelligence system dispositions method |
Non-Patent Citations (1)
| Title |
|---|
| DAVE CONWAY-JONES ETC: "Demonstration of Federated Learning in aResource-Constrained Networked Environment", pages 484 - 486 * |
Cited By (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2021109647A1 (en) * | 2019-12-05 | 2021-06-10 | 深圳前海微众银行股份有限公司 | Federated learning method and apparatus based on multi-source heterogeneous system |
| CN112087766A (en) * | 2020-08-24 | 2020-12-15 | 浙江大学 | Unmanned system heterogeneous network communication channel access method and device |
| CN112087766B (en) * | 2020-08-24 | 2022-03-25 | 浙江大学 | Unmanned system heterogeneous network communication channel access method and device |
| CN112286752A (en) * | 2020-09-29 | 2021-01-29 | 深圳致星科技有限公司 | Algorithm verification method and system for federated learning heterogeneous processing system |
| CN113160021A (en) * | 2021-03-18 | 2021-07-23 | 天津中科物联科技有限公司 | Safe production early warning system based on multi-source heterogeneous data federal learning |
| CN113160021B (en) * | 2021-03-18 | 2022-04-12 | 天津中科物联科技有限公司 | Safe production early warning system based on multi-source heterogeneous data federal learning |
| CN113743677A (en) * | 2021-09-16 | 2021-12-03 | 成都数融科技有限公司 | Personal credit evaluation model training method and evaluation method based on federal learning |
| CN113743677B (en) * | 2021-09-16 | 2023-06-30 | 成都数融科技有限公司 | Personal credit evaluation model training method and evaluation method based on federal learning |
| CN114090983A (en) * | 2022-01-24 | 2022-02-25 | 亿景智联(北京)科技有限公司 | Heterogeneous federated learning platform communication method and device |
| CN116362352A (en) * | 2023-06-01 | 2023-06-30 | 广州思迈特软件有限公司 | Model automatic updating method, system, medium and terminal based on machine learning |
Also Published As
| Publication number | Publication date |
|---|---|
| WO2021109647A1 (en) | 2021-06-10 |
| CN111126618B (en) | 2023-08-04 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111126618B (en) | Federal learning method and device based on multi-source heterogeneous system | |
| US20200242591A1 (en) | Network Performance Testing with Blockchain | |
| Ahmad et al. | An efficient divide-and-conquer approach for big data analytics in machine-to-machine communication | |
| CN107682068B (en) | Task-driven reconfigurable spatial information network resource management architecture and method | |
| CN114611705A (en) | Data processing method, training method for machine learning, and related device and equipment | |
| Hayashi et al. | A study on the variety and size of input data for radio propagation prediction using a deep neural network | |
| CN101819661A (en) | Method for analyzing and evaluating decision supporting capability of complex system | |
| CN115249073A (en) | Method and device for federated learning | |
| CN115580891B (en) | Flow prediction model training method, prediction method and device based on federal learning | |
| CN110087041A (en) | Video data processing and transmission method and system based on the base station 5G | |
| CN115277789A (en) | Safety protection system and method for cascade hydropower station | |
| CN116665040A (en) | Building change detection method based on attention mechanism and with multi-scale input and output | |
| CN114943284A (en) | Data processing system and method of behavior prediction model | |
| CN111818146A (en) | SOA cloud computing intelligent agricultural data processing method and system | |
| Liu et al. | HPL-ViT: A unified perception framework for heterogeneous parallel LiDARs in V2V | |
| CN109062677A (en) | Unmanned aerial vehicle system calculation migration method | |
| Curiel-Ramirez et al. | A simulation approach of the internet of intelligent vehicles for closed routes in urban environments | |
| CN110022541A (en) | A kind of sparse acquisition of WSN crop growth environment information and transmission method based on NB-IoT and FPGA | |
| CN115665805A (en) | Point cloud analysis task-oriented edge computing resource scheduling method and system | |
| CN110400373B (en) | Video information processing method and server | |
| CN113780371A (en) | Insulator state edge recognition method based on edge calculation and deep learning | |
| CN116600361B (en) | Unmanned aerial vehicle networking configuration method, unmanned aerial vehicle networking configuration equipment and readable storage medium | |
| Qu et al. | SiCP: Simultaneous Individual and Cooperative Perception for 3D Object Detection in Connected and Automated Vehicles | |
| CN118157999B (en) | Model training method, device, terminal and storage medium | |
| Xu et al. | Deep Reinforcement Learning for Image-Based Multi-Agent Coverage Path Planning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |