Search

CN-117035055-B - Equipment RUL prediction method and system based on BFL and semantics

CN117035055BCN 117035055 BCN117035055 BCN 117035055BCN-117035055-B

Abstract

The invention discloses a method and a system for predicting RUL of equipment based on BFL and semantics, which comprises the following steps of 1, carrying out identity verification through an ID verifier after FL participants receive task release information from a working node, 2, carrying out local model training on data acquired by a sensor by the FL participants through a CNN-LSTM method, 3, judging whether the local model training reaches the maximum iteration number or a root mean square error threshold according to an aggregation strategy, if so, ending the round training, 4, judging whether to participate in the federal learning training according to semantic reasoning, and determining whether to send the local model to a blockchain network by the FL participants, and 5, carrying out weighted aggregation on the local model by the working node according to the federal gray wolf weighting strategy, and returning the updated global model to the FL participants. The invention realizes information security sharing among the edge nodes and improves the operation efficiency of the whole Internet of things system.

Inventors

  • GENG DAOQU
  • WANG SHOUZHENG

Assignees

  • 重庆邮电大学

Dates

Publication Date
20260512
Application Date
20230718

Claims (10)

  1. 1. The RUL prediction method for the equipment based on BFL and semantics is characterized by comprising the following steps: After receiving task release information from a working node of a blockchain network, FL participants perform local model training on data acquired by a sensor, wherein the data acquired by the sensor is operation data generated in the working process of equipment; after the local model training is finished, judging whether to finish the iterative process according to the iteration times or the root mean square error threshold value; If the iteration process is determined to be ended, the FL participant sends weight data after the local model training to a blockchain network; When the working node of the blockchain network receives weight data sent by all FL participants, weighting and aggregating the weight data, sending an updated global model back to the FL participants for convergence detection, if the model has reached convergence, the participants quit federal learning, if the model has not reached convergence, starting a new round of federal learning training from local model training by using the global model, wherein the weight data is subjected to differential privacy treatment and then sent to the blockchain network for weighting and aggregating, and the weighting and aggregating is based on a consensus algorithm and weighting and aggregating the weight data according to a federal gray wolf algorithm and an aggregation policy, specifically, a model aggregation formula is as follows: Wherein, the 、 Representing global weights obtained by a federal averaging algorithm and by a federal wolf algorithm, respectively; Represent the first Wheel set Data volume of individual FL participants; indicating the number of FL participants; Representing a scoring function; the weight value corresponding to the maximum scoring function is obtained; The final calculated global weight value is obtained by a federal average algorithm and a federal gray wolf algorithm, wherein the scoring function is selected from the scoring functions in PHM2008 data challenge race, and is defined as follows: Wherein, the Is the difference between the RUL predicted value and the true value; Calculating a score for the final; The number of data for the test set; Is a natural constant; And stopping training when the training times reach the preset maximum training times or all FL participants exit from federal learning through convergence detection.
  2. 2. The method for predicting the RUL of the device based on BFL and semantics as claimed in claim 1, wherein FL participants perform local model training on the data collected by the sensor by adopting a CNN-LSTM (convolutional neural network-long short-term memory neural network) method, comprising the steps of: the FL participants collect data of the sensor network, preprocess the data and put the data into a local database; the FL participants adopt a CNN-LSTM method based on an attention mechanism to carry out model training on the data in a local database; the method based on the attention mechanism comprises the following specific steps of: Step 1, preprocessing the input original data to obtain Individual feature vectors As input to the CNN-LSTM network structure based on the attention mechanism; Step 2, obtaining local space state characteristics contained in the CNN (Convolutional Neural Networks, convolutional neural network) layer; Step 3, extracting time characteristic information by utilizing an LSTM (Long Short Term Memory, long-term memory) layer; Step 4, inputting the characteristic information extracted by the CNN-LSTM into a self-attention mechanism for weighting and outputting; step 5, converting the output of the attention mechanism into one-dimensional data through a tiling layer, and outputting a residual service life prediction result in a mode of many to one through 2 full-connection layers; the evaluation indexes of the life prediction result are a scoring function and root mean square error (Root Mean Square Error, RMSE); The RMSE calculation formula is: wherein: predicting the residual service life; is a true value; Represent the first Performing round iteration; Representing the total number of iterations.
  3. 3. The method for predicting the RUL of the equipment based on the BFL and the semantics as claimed in claim 1, wherein before the FL participants send the weight data trained by the local model to the blockchain network, the weight data is subjected to differential privacy processing, specifically, the weight data to be transmitted to the blockchain network is subjected to noise adding processing by using a differential privacy algorithm, and the formula of the differential privacy algorithm is as follows: Wherein, the Indicating that the output result is Probability of (2); And A dataset representing a hamming distance of 1; Representing a differential privacy algorithm; representing an arbitrary set of outputs; Is a natural constant; Representing a minimum value for measuring the effect of privacy budget; Is a relaxed term that indicates how far the differential privacy is acceptable to be unsatisfied.
  4. 4. The method for predicting the RUL of the device based on BFL and semantics of claim 1, wherein before the FL participant performs the local model training on the data collected by the sensor, the method further comprises the FL participant performing ID verification through a blockchain network, wherein the ID verification is implemented through a precompiled ID verifier smart contract, and the method comprises the steps of: When the FL participant performs identity registration, the intelligent contract automatically puts the identity ID of the FL participant into an identity ID pool; After receiving the task release information from the working node, the FL participant provides an identity ID of the FL participant to the ID verifier for verification; If the identity ID is matched with the registration ID in the ID pool, agreeing to upload data to the blockchain network by the FL participant; if the identity ID is not matched with the registration ID in the ID pool or is not verified after overtime, the verification is not confirmed to pass, and the participant is not agreed to upload data to the blockchain network.
  5. 5. The BFL and semantic based device RUL prediction method according to claim 2, wherein the weighted aggregation of the weight data is performed by the working node based on a consensus algorithm and according to a federal gray wolf algorithm and an aggregation policy, and sent back to the FL participants for convergence detection, comprising the steps of: After receiving the weight data of all FL participants, the working node aggregates the weight data based on a consensus algorithm according to a federal gray wolf algorithm and an aggregation strategy to obtain a new initial model, and broadcasts the new initial model to the FL participants; The FL participants receive the initial model sent by the working node and then carry out model convergence check, if the model has reached convergence, the participants exit federal learning, and if the model has not reached convergence, a new round of federal learning is carried out from local model training by using the global model.
  6. 6. The method for predicting the RUL of the device based on BFL and semantics of claim 5, wherein the federal wolf algorithm is based on an original federal average algorithm FedAvg, and the federal wolf algorithm FedGWO is provided to perform optimization calculation on weight data of a local model, and specifically comprises: the objective function is a scoring function in the PHM2008 data challenge, and the position is selected Bringing the objective function into three solutions with the maximum scoring function value The formula of the weight data calculation mode of the local model is as follows: In the formula, 、 、 Respectively represent the current population 、 、 Is a position vector of (2); A position vector representing a gray wolf; 、 、 Respectively representing the distances between the current candidate gray wolves and the optimal three wolves; the current iteration number; representing hadamard product operations; And Is a synergistic coefficient vector when When the wolves are found in various areas, the hunting objects are found as dispersed as possible among the wolves At this time, the wolf will intensively search for hunting in a certain or some area, Is that Random values in between, represent the random weight of the current position of the wolf on the effect of the prey, Indicating that the influence weight is large, otherwise, indicating that the influence weight is small; 、 、 And 、 、 Respectively represent the current population 、 、 Is a co-coefficient vector of (a) And Is a random vector, and is in the whole iterative process Linearly decreasing from 2 to 0; And Is that Random vectors between; representing the final position of the current candidate wolves in the wolf group; is the aggregate weight obtained by the wolf algorithm.
  7. 7. The method of claim 1, further comprising generating a fixed length 16-ary string from final model and data contribution information of the FL participant and aggregated result information of the working node by a hash algorithm after the FL participant exits federal learning, and storing the 16-ary string on a blockchain.
  8. 8. The method for predicting the RUL of the device based on BFL and semantics of claim 1, wherein before the FL participant sends the weight data after the local model training to the blockchain network, further comprising determining whether the FL participant participates in the federal learning training and decides whether to send the local model to the blockchain network according to semantic reasoning after the iterative process is finished, comprising the steps of: creating ontology classes and relationships; Creating an ontology attribute, wherein the ontology attribute comprises a data attribute and an object attribute; writing an inference rule, and formulating a mapping rule from a node resource state to a next execution task; acquiring data from a real world scene; creating an instance, namely determining a specific class of the acquired data according to the ontology on the basis of creating the ontology, and instantiating; defining instance attributes, including adding object attributes and data attributes to the instance according to the generated data; The method comprises the steps of realizing reasoning, namely using a reasoning machine to reason the nodes, obtaining the current storage, calculation and other resource states of the nodes and whether the accuracy of the model meets the minimum threshold requirement, and obtaining the next state of the nodes for training, reasoning or model aggregation.
  9. 9. The method for predicting the RUL of the equipment based on BFL and semantics as claimed in claim 1, wherein the working node works on a decentralised blockchain, the blockchain network calculates the contribution degree of the working node according to a reward and punishment intelligent contract as the basis of the selection of the working node, the working node and FL participants keep communication and model parameter data exchange, the FL participants can view historical transaction data through the working node, check bits are added in the original data frame format of the blockchain to realize the management of the lifecycle of the blockchain, and the method comprises the following specific steps: The block chain data format is added with check bits on the original basis, and when the check bit is 1, the current block chain model data is considered to have problems and is broadcasted to other working nodes; Adopting a voting mechanism, when 2/3 working nodes consider that the data has problems, calling a pre-compiled intelligent contract to initiate transaction to a pre-compiled contract with a fixed address of 0x1005, and reading and writing the state of the contract; And determining whether to stop the current FL training according to the contract state, and performing intelligent contract freezing, thawing and revocation.
  10. 10. A BFL and semantic based device RUL prediction system, characterized in that a BFL and semantic based device RUL prediction method according to any of claims 1 to 9 is used, comprising: The FL participant module is used as an edge server, is internally provided with sensors of various components and environments and is used for collecting data of a sensor network and realizing the processing of bottom data by edges, the data preprocessing module is used for preprocessing the collected data, the local database is used for storing the preprocessed data, the CNN-LSTM equipment residual service life prediction model training module is used for carrying out simulation training on the data stored in the local database, and the ID generation module is used for providing an ID of the FL participant module for an ID verifier of a blockchain network so as to verify the identity; The working node is used for working on the decentralized block chain, and always keeping data communication with the FL participant module, initiating a federal training task to the FL participant module, and carrying out model aggregation through a consensus algorithm and an intelligent contract; The block chain network module comprises a consensus algorithm, an intelligent contract, a full node, an identity ID pool and an ID verifier; the block chain network module receives the data transmitted by the FL participant module, and the working node performs weighted aggregation on the weight data based on a consensus algorithm according to a Federal gray wolf algorithm and an aggregation strategy so as to realize Federal learning; the ID verifier is realized through an intelligent contract of a blockchain, and when the identity ID sent by the FL participant module is received, the intelligent contract can automatically trigger and compare the identity ID in the identity ID pool to verify; The FL participant module judges whether training conditions, including power resources, communication resources, storage resources, calculation resources and models, are met or not through each node of the semantic reasoning module, and uploads model information to the blockchain network if the training conditions are met.

Description

Equipment RUL prediction method and system based on BFL and semantics Technical Field The invention belongs to the field of the Internet of things, relates to research on a residual service life prediction technology of equipment based on BFL and semantics, and belongs to the field of combination of artificial intelligence, edge calculation, privacy protection and semantic technology. Background The statements in this section merely provide background information related to the present disclosure and may constitute prior art. In carrying out the present invention, the inventors have found that at least the following problems exist in the prior art. In the modern industrial development process, large-scale high-end key electromechanical equipment presents a development trend of complexity, automation and centralization, and is often in a continuous running state with high load and variable working conditions. To ensure proper operation and ease of maintenance of the machine, the design of reliability for the initial stage of the machine, as well as on-line monitoring and health management after the machine is put into service, must be enhanced. In the age background of big data, the industry internet in the united states, industry 4.0 in germany, and the like are all driving the formation of health assessment framework construction based on data driving and the development of information management systems, so that fault Prediction and Health Management (PHM) have been developed. PHM makes the maintenance activity transition from passive post maintenance, periodic maintenance, optionally maintenance, etc. to active predictive maintenance. The core of the predictive maintenance is that the residual service life (REMAINING USEFUL LIFE, RUL) is predicted, so that production tasks can be more purposefully arranged, a reasonable maintenance plan can be formulated, and the maximization of production efficiency is realized. Existing methods for predicting the residual service life of equipment are mainly divided into a model-based method, a data-driven-based method and a hybrid-based method. Model-based methods require some knowledge of the mechanism of the device in order to build an appropriate model to predict the remaining useful life of the device. However, for complex working conditions in a real environment, due to the complex structure of mechanical equipment, stronger coupling interference is generated between sub-components, and the influence of internal and external nonlinear factors such as damping, variable stiffness, variable external load and the like in the running process is overcome, so that an accurate physical model is difficult to build and apply to monitoring and prediction, and prediction accuracy cannot be ensured. Hybrid-based approaches combine the advantages of model-based and data-based to improve prediction accuracy, but it is also difficult to build accurate models. Instead of using laws of physics to estimate the degradation process, the data-driven method based on historical failure data analysis may be used to implement the prediction of RUL. In recent years, with the development of sensor technology and signal processing technology, operation data generated during operation of equipment can be better collected and processed, and therefore, the method has become the mainstream of prediction methods. Training of the device residual life prediction model using deep learning methods, however, may result in a significant amount of private data, such as plant production data, user usage data, model parameter data, and the like. These data are collected in large amounts and stored in corporate databases, and users cannot delete or control the use of these data. Meanwhile, the production data may contain a large amount of sensitive information, such as capacity, work efficiency and the like. Uploading these data to the cloud creates a significant security risk. And the data collected by the sensor has great heterogeneity, and due to the influence of factors such as environment, the data loss caused by packet loss can be generated, and the non-independent and uniform distribution of the data on each device can seriously influence the learning rate of the deep learning algorithm. On the other hand, information security is put into important protection, so that industry researchers often only can mine and analyze related industry data, and the problem of data islanding is caused. Imperfections in the data set can result in poor results in models that are trained by researchers, which significantly limit the effectiveness of deep learning due to the privacy of the data. In response to the above problems, federal learning (FEDERATED LEARNING, FL) has been proposed to solve these problems to some extent. Federal learning can be built on edge devices, and training of a prediction model is achieved through cooperative work of multiple edge devices. The traditional federal learning needs a trusted