CN114936606A - Asynchronous decentralized model training method suitable for edge Internet of things agent device - Google Patents
Asynchronous decentralized model training method suitable for edge Internet of things agent device Download PDFInfo
- Publication number
- CN114936606A CN114936606A CN202210651051.5A CN202210651051A CN114936606A CN 114936606 A CN114936606 A CN 114936606A CN 202210651051 A CN202210651051 A CN 202210651051A CN 114936606 A CN114936606 A CN 114936606A
- Authority
- CN
- China
- Prior art keywords
- model
- edge internet
- things
- agent device
- edge
- 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
- 238000012549 training Methods 0.000 title claims abstract description 42
- 238000000034 method Methods 0.000 title claims abstract description 38
- 239000003795 chemical substances by application Substances 0.000 claims abstract description 87
- 230000004931 aggregating effect Effects 0.000 claims abstract description 7
- 230000007246 mechanism Effects 0.000 claims abstract description 6
- 230000006870 function Effects 0.000 claims description 26
- 238000004891 communication Methods 0.000 claims description 15
- 230000008569 process Effects 0.000 claims description 9
- 230000002776 aggregation Effects 0.000 claims description 6
- 238000004220 aggregation Methods 0.000 claims description 6
- 238000005070 sampling Methods 0.000 claims description 6
- 239000011159 matrix material Substances 0.000 claims description 5
- 238000002372 labelling Methods 0.000 claims description 4
- 238000011144 upstream manufacturing Methods 0.000 claims description 4
- 238000009795 derivation Methods 0.000 claims description 3
- 238000006116 polymerization reaction Methods 0.000 claims description 3
- 238000005516 engineering process Methods 0.000 abstract description 6
- 238000004422 calculation algorithm Methods 0.000 abstract description 4
- 238000004364 calculation method Methods 0.000 description 10
- 238000010801 machine learning Methods 0.000 description 10
- 238000001514 detection method Methods 0.000 description 9
- 238000007689 inspection Methods 0.000 description 7
- 238000007726 management method Methods 0.000 description 5
- 238000012545 processing Methods 0.000 description 4
- 238000002955 isolation Methods 0.000 description 3
- 239000002699 waste material Substances 0.000 description 3
- 238000010586 diagram Methods 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000012423 maintenance Methods 0.000 description 2
- 238000012544 monitoring process Methods 0.000 description 2
- 235000006629 Prosopis spicigera Nutrition 0.000 description 1
- 240000000037 Prosopis spicigera Species 0.000 description 1
- 230000002159 abnormal effect Effects 0.000 description 1
- 230000001133 acceleration Effects 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- QVFWZNCVPCJQOP-UHFFFAOYSA-N chloralodol Chemical compound CC(O)(C)CC(C)OC(O)C(Cl)(Cl)Cl QVFWZNCVPCJQOP-UHFFFAOYSA-N 0.000 description 1
- 230000008878 coupling Effects 0.000 description 1
- 238000010168 coupling process Methods 0.000 description 1
- 238000005859 coupling reaction Methods 0.000 description 1
- 238000013480 data collection Methods 0.000 description 1
- 238000013523 data management Methods 0.000 description 1
- 238000013499 data model Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000009977 dual effect Effects 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000010355 oscillation Effects 0.000 description 1
- 230000000737 periodic effect Effects 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 238000001228 spectrum Methods 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F21/00—Security arrangements for protecting computers, components thereof, programs or data against unauthorised activity
- G06F21/60—Protecting data
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F9/00—Arrangements for program control, e.g. control units
- G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
- G06F9/46—Multiprogramming arrangements
- G06F9/54—Interprogram communication
- G06F9/544—Buffers; Shared memory; Pipes
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y04—INFORMATION OR COMMUNICATION TECHNOLOGIES HAVING AN IMPACT ON OTHER TECHNOLOGY AREAS
- Y04S—SYSTEMS INTEGRATING TECHNOLOGIES RELATED TO POWER NETWORK OPERATION, COMMUNICATION OR INFORMATION TECHNOLOGIES FOR IMPROVING THE ELECTRICAL POWER GENERATION, TRANSMISSION, DISTRIBUTION, MANAGEMENT OR USAGE, i.e. SMART GRIDS
- Y04S10/00—Systems supporting electrical power generation, transmission or distribution
- Y04S10/50—Systems or methods supporting the power network operation or management, involving a certain degree of interaction with the load-side end user applications
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Software Systems (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Mathematical Physics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Computer Hardware Design (AREA)
- Bioethics (AREA)
- Computer Security & Cryptography (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Data Exchanges In Wide-Area Networks (AREA)
Abstract
The invention discloses an asynchronous decentralized model training method suitable for an edge Internet of things agent device, which comprises the following steps: each edge Internet of things agent device collects data from a user terminal and initializes the data; calculating a loss function value, performing back propagation according to the loss function value to obtain a gradient, and updating the model by using a random gradient descent formula; carrying out data privacy protection on the updated model by using a differential privacy mechanism; each edge Internet of things agent device sends the model to a neighbor and receives the neighbor model, and updates a local buffer pool when receiving the neighbor model; and aggregating the models in the local buffer pool to obtain a next iteration model. The training method disclosed by the invention reduces the bandwidth requirement through decentralization, improves the robustness of the algorithm, accelerates the training through an asynchronous mode, and improves the resource utilization rate; meanwhile, the protection of the local data privacy is realized by utilizing a differential privacy technology, and the privacy disclosure is effectively prevented.
Description
Technical Field
The invention belongs to the field of distributed machine learning, and particularly relates to an asynchronous decentralized model training method suitable for an edge Internet of things agent device.
Background
Distributed machine learning refers to a computing technique that distributes data across different machines for parallel large-scale machine learning training. The architecture based on the parameter server is a basic distributed model, each client performs gradient calculation according to local data of the client and then sends the gradient calculation to the server, and the server performs model aggregation and then sends the global model to each client to perform the next gradient calculation. This architecture breaks the bottleneck of stand-alone training and is able to process larger scale data, but with less robustness. First, this architecture is limited by the bandwidth and performance of the parameter servers, and when a parameter server is down or damaged, the entire training process will terminate until the server is functioning properly. Secondly, each client needs to wait for the completion of the training of other clients and can enter the next round of training after sending the model to the parameter server, so that the performance based on the parameter server architecture is limited by the worst performance of the client, and the waste of training resources is caused.
Decentralized machine learning can effectively solve the above problems. Under a decentralized architecture, all clients form a peer-to-peer network, and each client has own data and model. In each training round, each client uses its own data to perform gradient calculation, then shares the model with the connected neighbor clients, and calculates a local average model. Each client only needs to communicate with the connected neighbor clients and updates the model of the client, and the downtime or the error of any client only affects the training of a part of machines, so that the decentralized machine learning has higher robustness to the problems of communication bottleneck, edge equipment failure and the like.
The architecture of decentralized machine learning has great coupling with edge computing, and can be applied to various fields of edge computing. In the field of intelligent power grids, power loads intelligently predict power which needs to be used in multiple places, and power utilization data of each region are uploaded to local edge equipment to be cooperatively trained to obtain a global prediction model; for another example, in an intelligent inspection system, a machine learning model is needed to determine whether an inspected circuit or equipment has a fault, however, fault information of different scene areas is different, and data can be sourced from a large number of inspection equipment in different areas. Therefore, the inspection robot can be used for data acquisition and then uploaded to local edge equipment to cooperatively train a global model capable of covering various fault information. Edge computing essentially provides a decentralized service, and decentralized cooperative training has unprecedented advantages in the field of edge computing.
Although decentralized cooperative training algorithms have many advantages over the algorithms of the parameter server architecture, they still have certain disadvantages. As introduced above, each client needs to communicate with the neighbor's clients and share model data in each round of training. Therefore, any client needs to synchronously wait for the completion of the training of all the neighbor nodes and successfully receive the model to perform the next round of training, which causes resource waste to a certain extent. In addition, data privacy during communication needs to be further ensured.
Disclosure of Invention
In order to solve the technical problems, the invention provides an asynchronous decentralized model training method suitable for an edge Internet of things agent device, bandwidth requirements are reduced through decentralized and robustness of an algorithm is improved, and training is accelerated through an asynchronous mode and resource utilization rate is improved; meanwhile, the protection of the local data privacy is realized by utilizing a differential privacy technology.
In order to achieve the purpose, the technical scheme of the invention is as follows:
an asynchronous decentralized model training method suitable for a marginal Internet of things proxy device comprises the following steps:
(1) an initialization stage: each edge Internet of things agent device collects data from a user terminal, preprocesses the data to obtain a sample data set, and initializes a local model and a local buffer pool;
(2) and (3) updating the model: each edge Internet of things agent device calculates a loss function value of the sample data set, and then performs back propagation according to the loss function value to obtain a gradient; updating the model by using a random gradient descent formula;
(3) a privacy protection stage: each edge Internet of things agent device utilizes a differential privacy mechanism to introduce quantitative noise into the updated model for data privacy protection, and a noise-adding model is obtained;
(4) a model sharing stage: each edge Internet of things agent device sends the noise-added model to a neighbor edge Internet of things agent device, receives the model from the neighbor edge Internet of things agent device, and updates the model to a local buffer pool when receiving the neighbor model;
(5) a model polymerization stage: and (3) aggregating each edge Internet of things agent device by using the model in the local buffer pool to obtain a next iteration model, and returning to the step (2) until the maximum iteration round number is met and then outputting the model.
In the above scheme, the parameters initialized in step (1) include: the method comprises the following steps of calculating the total iteration number K, the learning rate gamma, the weight matrix W, the variance sigma of noise, the number n of edge Internet of things proxy devices and an initialization model of the edge Internet of things proxy device i
And initializing local buffer pool variables
Sample data set D i And a loss function, the loss function employing a crossoverAn entropy loss function F (·).
In the above scheme, the step (2) is specifically as follows:
(2.1) in the k-th iteration, the edge Internet of things agent i samples the data set D from the edge Internet of things agent i i In-between random sampling of a data sample
And is stored in the memory to be stored in the memory,
the characteristics of the sample are represented by,
representing a sample label, where d is the dimension of the sample feature, R d Representing a d-dimensional real number space;
(2.2) the edge Internet of things agent device i trains the local model by using an AI module, namely, samples to be obtained
Sample characteristics of
Model input to the k-th wheel
To obtain a predicted value y pred ；
(2.3) labeling the specimen
And predicted value y pred The input is input into a cross entropy loss function, namely a loss function value is obtained according to the following formula:
(2.4) obtaining a loss function from the back propagation derivation
About model
Gradient of (2)
Then, parameters are updated by using a random gradient descent formula to obtain an updated model
Where γ is the learning rate.
In the above scheme, the step (3) is specifically as follows:
(3.1) sampling a d-dimensional vector from the Gaussian distribution with d-dimensional variance as sigma to obtain a noise vector according to the initialized noise variance sigma
(3.2) noise vector to be sampled
Adding to updated model
To obtain a model of the noise-adding process
In the scheme, the step (4) is specifically as follows:
(4.1) edge agent of things i uses upstream communication interface and networkModel for network protocol to add its own noise
Sending the information to a neighbor edge Internet of things agent j;
(4.2) the edge IOT agent i uses the upstream communication interface and network protocol to receive the noise processing model from the neighbor edge IOT agent j
And model
If the buffer pool variable is saved in the buffer pool of the user, the updated buffer pool variable is stored in the buffer pool
In the above scheme, the step (5) is specifically as follows:
and aggregating the model data of the buffer pool by using the initialized weight matrix W, namely performing model aggregation on the edge Internet of things agent device i to obtain the model data of the next round:
wherein, W ij Elements representing ith row and jth column represent a model of an edge agent i receiving an edge agent j
The weight assigned.
Through the technical scheme, the asynchronous decentralized model training method suitable for the edge Internet of things agent device has the following beneficial effects:
(1) each edge Internet of things agent device initializes a buffer pool to receive model data from a neighbor edge Internet of things agent device, updates variables in the buffer pool when the neighbor data is successfully received, and finally uses the variables in the buffer pool to aggregate the models; by an asynchronous mode, any edge Internet of things agent device can directly use model data in the buffer pool for aggregation, and the training of other edge Internet of things agent devices is not required to be completed. This will greatly increase the speed of model training and solve the problem of resource waste caused by different equipment differences.
(2) The invention introduces a differential privacy mechanism, carries out noise processing on information transmitted by different edge Internet of things agent devices, and ensures privacy safety for data of each edge Internet of things agent device.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below.
FIG. 1 is a schematic diagram of a decentralized machine learning architecture according to the present disclosure;
FIG. 2 is a functional block diagram of an edge IOT agent according to the present disclosure;
FIG. 3 is a flowchart of an asynchronous decentralized model training method suitable for an edge Internet of things proxy device according to the present invention.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention.
The invention provides an asynchronous decentralized model training method suitable for an edge Internet of things agent device, wherein a decentralized machine learning architecture is shown in figure 1: and the user terminal of each region uploads the data to a local edge Internet of things agent device, and the edge Internet of things agent device processes the data to obtain a data set with a uniform format. And then, each edge Internet of things agent device performs machine learning model training, and shares the model trained by itself to other connected edge Internet of things agent devices in each iteration. And finally, each edge Internet of things agent device aggregates all collected models to perform the next iteration.
As shown in fig. 2, the edge internet of things proxy apparatus is an operation terminal for completing cooperative training, and the functional architecture of the edge internet of things proxy apparatus used in the embodiment of the present invention mainly includes the following parts:
(1) data acquisition module
The wireless communication device provides diversified wired and wireless interfaces, and is compatible with various wired communication protocols such as LAN, USB, RS485/232 and the like and wireless communication protocols such as RFID, Bluetooth, ZigBee, LoRa, NB-IoT and the like. Different wired communication and short-distance wireless communication technologies are adopted for different acquisition terminals to access the data of the terminals to the edge Internet of things agent device nearby, and unified acquisition of the data is achieved. Aiming at a wireless data acquisition scene, a wireless communication protocol frequency spectrum isolation technology is adopted, so that the mutual interference of different protocols is avoided, and the accuracy and the real-time performance of data acquisition are ensured. The northbound open interface supports multiple interface protocols such as MQTT, HTTP, COAP, XMPP, TCP/UDP and the like, and realizes unified data interaction with a platform layer Internet of things platform, a cloud platform and a full-service data center.
(2) Edge calculation module
The edge Internet of things agent device provides light-weight data caching capacity, format conversion is carried out on collected data according to a uniform service data model, and service-level analysis is carried out on the data by using an artificial intelligence technology. The device provides micro-service capability, can rapidly deploy customized application, and meets personalized requirements under different service scenes. The calculation result can be directly returned and transmitted to the terminal and the user, and the quick response of the service request is realized.
(3) AI function module
The edge Internet of things agent device hardware adopts a proprietary AI accelerating chip, deploys a general AI SDK, supports a mainstream deep learning reasoning framework and provides reasoning calculation accelerating capability for edge services. Especially for video monitoring services and high-bandwidth and low-delay service scenes, the local video coding and decoding hardware capability and AI calculation acceleration capability can effectively relieve the calculation load of a CPU (Central processing Unit) and meet the requirements of related services.
(4) Safety control module
In the aspect of data security, the edge Internet of things agent device provides a zero trust security access authentication mechanism, and performs multi-condition identity authentication on an access terminal by comprehensively analyzing fingerprint information such as an IP address, an MAC address, a communication protocol, a digital certificate and the like of equipment, so that the occurrence of equipment counterfeiting is avoided, and the identity credibility, safety and reliability of the access equipment are ensured. In the aspect of transmission safety, the edge Internet of things agent device establishes a virtual safe private network based on an encryption tunnel and an SD-WAN technology, so that a private safe isolation network is established on an unsafe wide area network, and meanwhile, different safe private networks can be established for different services, so that service isolation is achieved. In the aspect of equipment safety, unified maintenance management of the device and the access terminal is supported, wherein the unified maintenance management comprises login management, authority management, task management, data management, state monitoring, abnormal alarm and the like, and meanwhile periodic vulnerability scanning and automatic firmware upgrading of the device can be realized.
It should be noted that the existing edge internet of things agent apparatus can complete the model training method of the present invention, and is not limited to the above edge internet of things agent apparatus.
The invention discloses an asynchronous decentralized model training method suitable for an edge Internet of things agent device. Second, the method adds a certain amount of noise to the local model before model sharing to protect data privacy.
As shown in fig. 3, the method specifically includes the following steps:
(1) an initialization stage: each edge Internet of things agent device collects data from a user terminal, preprocesses the data to obtain a sample data set, and initializes a local model and a local buffer pool.
The method comprises the following specific steps:
(1.1) initializing the total number of iterations K, K usually being trained several times to find a suitable value.
(1.2) initialize the learning rate γ, which controls the convergence effect, usually through multiple training to find a suitable value. The early stage convergence of the larger learning rate is faster, but the oscillation convergence effect is poor easily in the vicinity of the optimal solution, the slower convergence speed is caused by the smaller learning rate, and the training time is very long.
(1.3) the number of the edge Internet of things agent devices is represented by a variable n, and the number of the n edge Internet of things agent devices is [1,2, …, n]. Initializing model for each edge Internet of things agent device i
And buffer pool variables
These values may be set to 0.
(1.4) initializing the local data set D i For each edge internet-of-things agent and loss function F (-) the data set is usually processed data containing features and labels, and the loss function can be a cross-entropy loss function.
(1.5) initializing a adjoint W with double random properties, and carrying out model aggregation by using the values of the adjoint by each edge Internet of things agent device. W ij And elements in the ith row and the jth column represent weights given by the edge internet of things agent i when receiving the model of the edge internet of things agent j.
(1.6) initializing the variance σ of the noise, wherein the larger the variance represents the more noise is added, the higher the degree of privacy protection is, but the larger the variance also causes the more noisy data can deviate from the original data, and the convergence speed can be reduced.
(2) And (3) updating the model: each edge Internet of things agent device calculates a loss function value of the sample data set by utilizing an AI function module thereof, and then performs reverse propagation according to the loss function value to obtain a gradient; and then updating the model by using a random gradient descent formula.
The method comprises the following specific steps:
(2.1) in the k-th iteration, the edge Internet of things agent i samples the data set D from the edge Internet of things agent i i In-between random sampling of a data sample
And is stored in the memory to be stored in the memory,
the characteristics of the sample are represented by,
representing a sample label, where d is the dimension of the sample feature, R d Representing a d-dimensional real number space;
(2.2) the edge Internet of things agent device i trains the local model by using an AI module, namely, samples to be obtained
Sample characteristics of
Model input to the k-th wheel
To obtain a predicted value y pred ；
(2.3) labeling the specimen
And predicted value y pred The input is input into a cross entropy loss function, namely a loss function value is obtained according to the following formula:
(2.4) obtaining a loss function from the back propagation derivation
About model
Gradient of (2)
Then, parameters are updated by using a random gradient descent formula to obtain an updated model
Where γ is the learning rate.
(3) A privacy protection stage: and each edge Internet of things agent device utilizes a differential privacy mechanism to introduce quantitative noise into the updated model for data privacy protection, so that a noise-adding model is obtained.
The method comprises the following specific steps:
(3.1) sampling a d-dimensional vector from a Gaussian distribution with d-dimensional variance sigma to obtain a noise vector according to the initialized noise variance sigma
(3.2) noise vector to be sampled
Adding to updated model
In the method, a model of noise-adding treatment is obtained
(4) A model sharing stage: each edge Internet of things agent device sends the noise-added model to a neighbor edge Internet of things agent device, receives the model from the neighbor edge Internet of things agent device, and updates the model to a local buffer pool when receiving the neighbor model.
The method comprises the following specific steps:
(4.1) edge Internet of things agent device i uses upstream communication interface and network protocol to process self-noise model
Sending the information to a neighbor edge Internet of things agent j;
(4.2) the edge Internet of things agent device i receives the noise-added model from the neighbor edge Internet of things agent device j by using an uplink communication interface and a network protocol
And model
If the buffer pool variable is saved in the buffer pool of the user, the updated buffer pool variable is stored in the buffer pool
(5) A model polymerization stage: and (3) aggregating each edge Internet of things agent device by using the model in the local buffer pool to obtain a next iteration model, and returning to the step (2) until the maximum iteration round number is met and then outputting the model.
The method comprises the following specific steps:
and aggregating the model data of the buffer pool by using the initialized weight matrix W, namely performing model aggregation on the edge Internet of things agent device i to obtain the model data of the next round:
wherein, W ij Elements representing ith row and jth column representing a model of an edge entity-association broker i receiving an edge entity-association broker j
The weight assigned. To ensure convergence, W needs to beEnsuring the dual random nature, i.e. the sum of any row and any column is 1. Wherein if i and j are not neighbors, W must be present ij ＝W ji ＝0。
The invention also discloses an asynchronous decentralized power equipment fault identification method suitable for the edge Internet of things agent device, which comprises the following steps:
first, data acquisition stage
1. The type of equipment and the type of fault to be detected are determined, depending on the actual scene requirements.
2. And (4) carrying out data collection according to the determined equipment, and putting the inspection robot to the area needing to be detected. Due to the fact that most scenes need to be covered, data need to be collected for the same power equipment under different weather conditions at different moments, not only pictures containing fault information but also normal pictures need to be collected.
3. A part of data is acquired from the Internet, the step is supplementary to the data acquisition of the previous step, and a fault detection model trained by more data sets has higher robustness.
4. And cutting all data pictures into uniform formats, and naming the pictures uniformly.
II, data processing stage:
1. the fault category, i.e. the number of sample tags and the specific sample tags, is first determined. Such as no device fault set to label No. 0, transformer fault set to label No. 1, line fault set to label No. 2, etc. This step depends on the actual scene requirements.
2. And marking the collected picture. This process can be accomplished using labeling tools such as labelme or labellimg. Specifically, the marking tool is used for marking the device to be detected in the picture, and a VOC data set format or a coco data set format is generated.
3. And segmenting the equipment fault data set into a training set, a verification set and a test set.
Thirdly, model training stage:
after sufficient data is collected, the data is uploaded to a local edge internet of things proxy device. Training a fault detection model in the edge Internet of things agent device according to the process shown in fig. 3 to obtain an output fault detection model. For target detection, a Fast-RCNN model or a Yolo model is generally used.
Fourthly, a model deployment stage:
and deploying environments required by the operation of the model on the inspection robot to be deployed, such as a python environment, a detection system, an early warning system and the like required by the operation. And then deploying the finally output fault detection model to each inspection robot in a docker or http service mode and the like.
Fifthly, fault detection stage:
the intelligent inspection robot acquires data of the power equipment and performs fault detection on the acquired picture data by using the fault detection model.
The previous description of the disclosed embodiments is provided to enable any person skilled in the art to make or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other embodiments without departing from the spirit or scope of the invention. Thus, the present invention is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims (6)
1. An asynchronous decentralized model training method suitable for a marginal Internet of things proxy device is characterized by comprising the following steps:
(1) an initialization stage: each edge Internet of things agent device collects data from a user terminal, preprocesses the data to obtain a sample data set, and initializes a local model and a local buffer pool;
(2) and (3) a model updating stage: each edge Internet of things agent device calculates a loss function value of the sample data set, and then performs reverse propagation according to the loss function value to obtain a gradient; updating the model by using a random gradient descent formula;
(3) a privacy protection stage: each edge Internet of things agent device utilizes a differential privacy mechanism to introduce quantitative noise into the updated model for data privacy protection, and a noise-adding model is obtained;
(4) a model sharing stage: each edge Internet of things agent device sends the noise-added model to a neighbor edge Internet of things agent device, receives the model from the neighbor edge Internet of things agent device, and updates the model to a local buffer pool when receiving the neighbor model;
(5) a model polymerization stage: and (3) aggregating each edge Internet of things agent device by using the model in the local buffer pool to obtain a next iteration model, and returning to the step (2) until the maximum iteration round number is met and then outputting the model.
2. The method for training the asynchronous decentralized model of the edge internet of things proxy device according to claim 1, wherein the parameters initialized in step (1) include: the method comprises the following steps of calculating the total iteration number K, the learning rate gamma, the weight matrix W, the variance sigma of noise, the number n of edge Internet of things proxy devices and an initialization model of the edge Internet of things proxy device i
And initializing local buffer pool variables
Sample data set D i And a loss function, the loss function employing a cross-entropy loss function F (·).
3. The method for training the asynchronous decentralized model suitable for the edge internet of things proxy device according to claim 1, wherein the step (2) is specifically as follows:
(2.1) in the k-th iteration, the edge Internet of things agent i samples the data set D from the edge Internet of things agent i i In-between random sampling of a data sample
And is stored in the internal memory of the computer,
the characteristics of the sample are represented by,
representing a sample label, where d is the dimension of the sample feature, R d Representing a d-dimensional real number space;
(2.2) the edge Internet of things agent device i trains the local model by using an AI module, namely, samples to be obtained
Sample characteristics of
Model input to the k-th wheel
To obtain a predicted value y pred ；
(2.3) labeling the specimen
And predicted value y pred The input is input into a cross entropy loss function, namely a loss function value is obtained according to the following formula:
(2.4) obtaining a loss function from the back propagation derivation
About model
Gradient of (2)
Then, parameters are updated by using a random gradient descent formula to obtain an updated model
Where γ is the learning rate.
4. The method for training the asynchronous decentralized model suitable for the edge internet of things proxy device according to claim 1, wherein the step (3) is specifically as follows:
(3.1) sampling a d-dimensional vector from a Gaussian distribution with d-dimensional variance sigma to obtain a noise vector according to the initialized noise variance sigma
(3.2) noise vector to be sampled
Adding to updated model
To obtain a model of the noise-adding process
5. The method for training the asynchronous decentralized model suitable for the edge agent device of the internet of things as claimed in claim 1, wherein the step (4) is specifically as follows:
(4.1) edge Internet of things agent device i uses upstream communication interface and network protocol to process self-noise model
Sending the information to a neighbor edge Internet of things agent j;
(4.2) the edge Internet of things agent device i receives the noise-added model from the neighbor edge Internet of things agent device j by using an uplink communication interface and a network protocol
And model
If the buffer pool variable is saved in the buffer pool of the user, the updated buffer pool variable is stored in the buffer pool
6. The method for training the asynchronous decentralized model suitable for the edge internet of things proxy device according to claim 1, wherein the step (5) is specifically as follows:
and aggregating the model data of the buffer pool by using the initialized weight matrix W, namely performing model aggregation on the edge Internet of things agent device i to obtain the model data of the next round:
wherein, W ij Elements representing ith row and jth column represent a model of an edge agent i receiving an edge agent j
The weight assigned.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210651051.5A CN114936606B (en) | 2022-06-10 | 2022-06-10 | Asynchronous decentralization model training method suitable for edge internet of things proxy device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210651051.5A CN114936606B (en) | 2022-06-10 | 2022-06-10 | Asynchronous decentralization model training method suitable for edge internet of things proxy device |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114936606A true CN114936606A (en) | 2022-08-23 |
| CN114936606B CN114936606B (en) | 2024-08-23 |
Family
ID=82866832
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210651051.5A Active CN114936606B (en) | 2022-06-10 | 2022-06-10 | Asynchronous decentralization model training method suitable for edge internet of things proxy device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114936606B (en) |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112818394A (en) * | 2021-01-29 | 2021-05-18 | 西安交通大学 | Self-adaptive asynchronous federal learning method with local privacy protection |
| CN114116198A (en) * | 2021-10-21 | 2022-03-01 | 西安电子科技大学 | Asynchronous federal learning method, system, equipment and terminal for mobile vehicle |
| US20220114475A1 (en) * | 2020-10-09 | 2022-04-14 | Rui Zhu | Methods and systems for decentralized federated learning |
| CN114363043A (en) * | 2021-12-30 | 2022-04-15 | 华东师范大学 | Asynchronous federated learning method based on verifiable aggregation and differential privacy in peer-to-peer network |
| CN114362940A (en) * | 2021-12-29 | 2022-04-15 | 华东师范大学 | Server-free asynchronous federated learning method for data privacy protection |
-
2022
- 2022-06-10 CN CN202210651051.5A patent/CN114936606B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20220114475A1 (en) * | 2020-10-09 | 2022-04-14 | Rui Zhu | Methods and systems for decentralized federated learning |
| CN112818394A (en) * | 2021-01-29 | 2021-05-18 | 西安交通大学 | Self-adaptive asynchronous federal learning method with local privacy protection |
| CN114116198A (en) * | 2021-10-21 | 2022-03-01 | 西安电子科技大学 | Asynchronous federal learning method, system, equipment and terminal for mobile vehicle |
| CN114362940A (en) * | 2021-12-29 | 2022-04-15 | 华东师范大学 | Server-free asynchronous federated learning method for data privacy protection |
| CN114363043A (en) * | 2021-12-30 | 2022-04-15 | 华东师范大学 | Asynchronous federated learning method based on verifiable aggregation and differential privacy in peer-to-peer network |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114936606B (en) | 2024-08-23 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112203282B (en) | 5G Internet of things intrusion detection method and system based on federal transfer learning | |
| Mohammed et al. | Budgeted online selection of candidate IoT clients to participate in federated learning | |
| Zainudin et al. | An efficient hybrid-dnn for ddos detection and classification in software-defined iiot networks | |
| CN107770263A (en) | A kind of internet-of-things terminal safety access method and system based on edge calculations | |
| EP4078899A1 (en) | Systems and methods for enhanced feedback for cascaded federated machine learning | |
| CN111510433A (en) | Internet of things malicious flow detection method based on fog computing platform | |
| CN107683597A (en) | Network behavior data collection and analysis for abnormality detection | |
| CN114330544A (en) | Method for establishing business flow abnormity detection model and abnormity detection method | |
| US20240106836A1 (en) | Learning of malicious behavior vocabulary and threat detection | |
| CN114710330A (en) | Anomaly detection method based on heterogeneous hierarchical federated learning | |
| CN116347492A (en) | 5G slice flow abnormality detection method, device, computer equipment and storage medium | |
| CN116319437A (en) | Network connectivity detection method and device | |
| Barsellotti et al. | Introducing data processing units (DPU) at the edge | |
| Ozer et al. | Offloading deep learning powered vision tasks from UAV to 5G edge server with denoising | |
| Abbasi et al. | FLITC: A novel federated learning-based method for IoT traffic classification | |
| Ageyev et al. | Traffic Monitoring and Abnormality Detection Methods for IoT | |
| CN112749403B (en) | Edge data encryption method suitable for edge Internet of things agent device | |
| Chen et al. | Federated meta-learning framework for few-shot fault diagnosis in industrial IoT | |
| WO2020192922A1 (en) | Intermediate network node and method performed therein for handling data of communication networks | |
| CN117729540A (en) | Perception equipment cloud edge safety control method based on unified edge computing framework | |
| CN114070775A (en) | Block chain network slice safety intelligent optimization method facing 5G intelligent network connection system | |
| CN114936606B (en) | Asynchronous decentralization model training method suitable for edge internet of things proxy device | |
| Chen et al. | A 5G Enabled Adaptive Computing Workflow for Greener Power Grid | |
| Malini et al. | An efficient deep learning mechanisms for IoT/Non-IoT devices classification and attack detection in SDN-enabled smart environment | |
| CN116155592A (en) | AMI network intrusion detection method based on DCGAN federal semi-supervised learning |
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 | ||
| CB03 | Change of inventor or designer information |
Inventor after: Wang Peng Inventor after: Zhang Liangxu Inventor after: Chen Shuzhen Inventor after: Yu Dongxiao Inventor after: Zou Yifei Inventor after: Du Chao Inventor before: Yu Dongxiao Inventor before: Zhang Liangxu Inventor before: Chen Shuzhen Inventor before: Zou Yifei Inventor before: Wang Peng Inventor before: Du Chao |
|
| CB03 | Change of inventor or designer information | ||
| TA01 | Transfer of patent application right |
Effective date of registration: 20221116 Address after: No.72 Binhai Road, Jimo District, Qingdao, Shandong 266200 Applicant after: SHANDONG University Applicant after: SHANGHAI STEP ELECTRIC Corp. Address before: No.72 Binhai Road, Jimo District, Qingdao, Shandong 266200 Applicant before: SHANDONG University |
|
| TA01 | Transfer of patent application right | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |