CN114580087B - Method, device and system for predicting federal remaining service life of shipborne equipment - Google Patents
Method, device and system for predicting federal remaining service life of shipborne equipment Download PDFInfo
- Publication number
- CN114580087B CN114580087B CN202210483195.4A CN202210483195A CN114580087B CN 114580087 B CN114580087 B CN 114580087B CN 202210483195 A CN202210483195 A CN 202210483195A CN 114580087 B CN114580087 B CN 114580087B
- Authority
- CN
- China
- Prior art keywords
- data
- service life
- training
- gradient
- model
- 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.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/10—Geometric CAD
- G06F30/15—Vehicle, aircraft or watercraft design
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
-
- 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
- G06N3/044—Recurrent networks, e.g. Hopfield 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
- G06N3/04—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
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
-
- 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
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2119/00—Details relating to the type or aim of the analysis or the optimisation
- G06F2119/04—Ageing analysis or optimisation against ageing
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Software Systems (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Data Mining & Analysis (AREA)
- Computational Linguistics (AREA)
- Molecular Biology (AREA)
- General Health & Medical Sciences (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Geometry (AREA)
- Computer Hardware Design (AREA)
- Medical Informatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Automation & Control Theory (AREA)
- Aviation & Aerospace Engineering (AREA)
- Computational Mathematics (AREA)
- Mathematical Analysis (AREA)
- Mathematical Optimization (AREA)
- Pure & Applied Mathematics (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention provides a method, a device and a system for predicting the residual service life of a ship-borne device, belongs to the technical field of machine learning, and solves the problems that the residual service life of the ship-borne device cannot be accurately predicted due to small data sample amount and multi-user sample dispersion, and the method comprises the following steps: constructing a residual service life prediction model based on machine learning; during federal training, the technical center server receives N groups of encryption gradients of all participants and decrypts the encryption gradients by using a reserved private key to obtain the gradients; clustering and screening the gradient obtained by decryption, and updating the parameters of the model by using the screened gradient; and predicting the residual service life of the ship-borne equipment in the running state in real time and early warning potential equipment faults by using the trained residual service life prediction model based on machine learning. The method can provide real-time residual service life prediction and intelligent early warning for potential equipment faults for the operating shipborne equipment.
Description
Technical Field
The invention belongs to the technical field of machine learning, and particularly relates to a method, a device and a system for predicting the residual federal service life of shipborne equipment.
Background
The statements in this section merely provide background information related to the present disclosure and may not necessarily constitute prior art.
With the development of artificial intelligence technology, more and more industries actively promote the integration with intelligent technology, and the traditional industries are intelligently upgraded and reformed. The residual service life prediction technology is used for training a model by constructing a machine learning model and taking equipment operation data as input, and finally enabling the model to predict the residual service life of the equipment. However, in the field of the remaining service life prediction technology of the onboard equipment, the machine learning technology is limited by insufficient data volume.
Shipboard equipment produced by manufacturers is installed on civil ships, warships and submarines and is handed to different users for use. However, as a technologically intensive ultra-large electromechanical product, the number of newly manufactured ships is very limited every year; and due to the confidentiality requirement and privacy protection, equipment manufacturers cannot collect data resources of different users at one place, further cannot develop effective data-driven residual service life prediction models, and cannot provide advanced intelligent prediction technical services for the users.
Therefore, the operation data of the shipborne equipment has the defects of small sample amount reserved by a single user and dispersion of samples of multiple users, a data island is formed, and accurate prediction of the residual service life of the shipborne equipment cannot be realized.
In addition, different users perform homomorphic real-time state monitoring on the same type of equipment under the technical support of manufacturers, the obtained state monitoring data are independently and identically distributed, but are influenced by potential factors such as equipment installation deviation, electronic device precision difference, environmental difference and the like, data held by different users do not strictly obey the same distribution, and the data distribution has different degrees of deviation, so that in the federal training process, if a technical center server performs the same processing on all received gradients, the model performance is reduced.
Disclosure of Invention
In order to overcome the defects of the prior art, the invention provides a method for predicting the residual service life of the federal of the shipborne equipment, which can realize accurate prediction of the residual service life of the shipborne equipment.
In order to achieve the above object, one or more embodiments of the present invention provide the following technical solutions:
in a first aspect, a method for predicting the residual federal service life of shipborne equipment is disclosed, which comprises the following steps:
establishing a residual service life prediction model based on machine learning by using historical monitoring data and real-time data of shipborne equipment and pre-training;
carrying out federal training on the built residual service life prediction model by using a federal training mode, and carrying out joint training by using a multi-user scattered data set during training;
and predicting the residual service life of the ship-borne equipment in the running state in real time and early warning potential equipment faults by using the trained residual service life prediction model based on machine learning.
As a further technical scheme, a residual service life prediction model based on machine learning is jointly trained by using a federal learning mechanism, and the method comprises the following specific steps:
sending the model training parameters to all federal training participants;
each participant processes data of self-contained shipborne equipment, performs multi-round training on the model and calculates loss, the model automatically calculates gradient by using a back propagation algorithm, and then performs algorithm encryption on the gradient by using a reserved public key;
all user terminals send encryption gradients to a technical center server;
the technical center server receives the N groups of encryption gradients of all the participants and decrypts the encryption gradients by using the preserved private key to obtain the gradients;
clustering the gradient obtained by decryption and screening to solve the problem of non-independent and same distribution of the multi-party scattered data, updating the local model by using the screened gradient, sending the updated local model to users in the cluster, and repeatedly finishing model updating of all the clusters;
and judging whether the updated model is tested by the user with own data and reaches the expected prediction performance, if not, repeatedly carrying out the federal training, and if so, stopping the training.
As a further technical scheme, the method screens the gradient obtained by decryption, and specifically comprises the following steps:
calculate each cluster package for the gradient set of all usersMean gradient of the included gradient samples; calculating the gradient samples contained in other clusters in the participating cluster D i Weight W when updating parameter i,j ；
The screening mechanism is W i,h >0.5 Cluster participating Cluster D i Otherwise, let W be updated i,h And =0, namely, not participating in the secondary round parameter update.
As a further technical scheme, the built machine learning-based residual service life prediction model specifically comprises the following steps:
a first channel in which time-dependent information of input data is learned using an LSTM to obtain time characteristic information;
the second channel is used for extracting the characteristics of the multi-sensor data to obtain important spatial information;
calculating the condition vectors of the acquired time characteristic information and the space important information by using the limiting conditions, and supervising the parameter updating of the characteristic learning network by using the condition vectors;
splicing the time characteristic information and the space important information obtained by the two channels in the last step with respective limiting conditions respectively, and taking the spliced time characteristic information and the space important information as the input of two-dimensional convolution networks respectively to carry out deeper characteristic extraction;
fine-grained fusion is carried out on the characteristics through a characteristic fusion network, so that the degradation process is better described;
and the multilayer full-connection layer is used for processing the fusion characteristics to obtain a predicted value of the residual service life.
As a further technical solution, before constructing the machine learning-based remaining service life prediction model, the method further comprises: and acquiring state monitoring data, namely monitoring the operation period of the shipborne equipment by utilizing various sensors to acquire the state data of the shipborne equipment, wherein the monitoring mode and the sensor arrangement are established and installed by an equipment manufacturer.
As a further technical scheme, the acquired state monitoring data is subjected to data preprocessing, including data normalization, signal denoising and data segmentation.
As a further technical solution, the data segmentation is to segment data in the model training data set, and segment the data into data segments of a set length according to a segmentation stride.
As a further technical solution, the data in the model training data set specifically is:
the data collected at time t consists of the outputs of the F sensors and is recorded as
；
Represents the output of the f-th sensor at time t to obtain the sensor data of the m-th equipment, and is recorded as
。
The state monitoring starts from the start of the device, denoted T =0, and the device reaches the end of the failure threshold, denoted T = T, i.e. the total life of the device is T, T m Representing the total life of the mth device.
In a second aspect, a device for predicting the residual federal service life of a ship-borne device is disclosed, which comprises: a technology center server, the technology center server comprising:
a model building module configured to: constructing a residual service life prediction model based on machine learning by utilizing historical monitoring data and real-time data of shipborne equipment;
a training module configured to: performing joint training on the built residual service life prediction model based on machine learning by using a federal learning mechanism, and performing joint training by using N dispersed data sets self-owned by each user during training;
a prediction module configured to: and predicting the residual service life of the ship-borne equipment in the running state in real time and early warning potential equipment faults by using the trained residual service life prediction model based on machine learning.
In a third aspect, a system for predicting the residual federal service life of a ship-borne device is disclosed, which comprises:
a technical center server, a plurality of user terminals and a transmission link;
the user terminals are respectively communicated with the technical center server through transmission links;
the technology center server includes:
a model building module configured to: constructing a residual service life prediction model based on machine learning by utilizing historical monitoring data and real-time data of shipborne equipment; the data is stored in the technical center;
a training module configured to: performing joint training on the built residual service life prediction model based on machine learning by using a federal learning mechanism, and performing joint training by using N dispersed data sets self-owned by each user during training;
a prediction module configured to: and predicting the residual service life of the ship-borne equipment in the running state in real time and early warning potential equipment faults by using the trained residual service life prediction model based on machine learning.
By applying the system, the residual service life of the shipborne equipment can be accurately predicted, maintenance personnel can predict the potential fault information of the equipment in advance according to the prediction result, scientific equipment scheduling arrangement and health management are carried out, and a maintenance plan of the equipment is made in advance, so that potential safety hazards caused by system faults are avoided, and the navigation reliability of ships and naval vessels is improved.
The above one or more technical solutions have the following beneficial effects:
the trained prediction model is provided for a user to use, real-time residual service life prediction and intelligent early warning of potential equipment faults are provided for operating shipborne equipment, maintenance plans of the equipment are made in advance by maintenance personnel, meanwhile, the user feeds back the use conditions of the prediction system to manufacturers, and the designer is helped to optimize and upgrade the system.
The machine learning model constructed by the invention is subjected to combined training on the model by means of a dispersed data isolated island under a federal learning mechanism without collecting dispersed data resources, so that the model training with few samples is realized, an accurate model is obtained, and the accurate prediction on the residual service life of shipborne equipment is realized.
The invention provides a method for screening different source gradients, which reduces negative influence of data distribution deviation on model training, improves the prediction performance of a model on the premise of protecting user privacy and guaranteeing data safety, and implements intelligent residual service life prediction technical service based on data driving on shipborne equipment.
Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention.
Drawings
The accompanying drawings, which are incorporated in and constitute a part of this specification, are included to provide a further understanding of the invention, and are incorporated in and constitute a part of this specification, illustrate exemplary embodiments of the invention and together with the description serve to explain the invention and not to limit the invention.
FIG. 1 is a schematic structural diagram of a federal residual service life prediction system in accordance with an embodiment of the present invention;
FIG. 2 is a diagram of a vibration monitoring sensor arrangement for an on-board ventilator;
FIG. 3 is a diagram of a model structure for predicting remaining useful life according to an embodiment of the present invention;
FIG. 4 illustrates a feature fusion network according to an embodiment of the present invention;
FIG. 5 is a flow chart of federal residual useful life prediction system training in accordance with an embodiment of the present invention;
fig. 6 is a federal remaining useful life prediction system device according to an embodiment of the present invention.
Detailed Description
It is to be understood that the following detailed description is exemplary and is intended to provide further explanation of the invention as claimed. Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.
It is noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of exemplary embodiments according to the invention.
The embodiments and features of the embodiments of the present invention may be combined with each other without conflict.
Example one
The embodiment discloses a method for predicting the residual federal service life of shipborne equipment, as shown in fig. 1:
on one hand, equipment manufacturers provide technical support, and the content comprises a residual service life prediction model designed by the manufacturer technical center based on machine learning by utilizing self historical monitoring data and real-time data, a design model federal training mode, data specification formulation, a data screening and processing method and the like.
On the other hand, different equipment users firstly judge whether own data meet data specifications set by manufacturers, and users meeting the specifications screen and preprocess the data according to manufacturer instructions under the condition that local data are not moved, and then train the prediction model together according to a federal training mode set by manufacturers.
And finally, the technical center server provides the trained prediction model for the user to use, provides real-time residual service life prediction and intelligent early warning of potential equipment faults for the running shipborne equipment, helps maintenance personnel to make a maintenance plan of the equipment in advance, and simultaneously, the user feeds back the service condition of the prediction system to a manufacturer, and helps designers to optimize and upgrade the system.
In order to realize the technical scheme, the method specifically comprises the following steps:
the method comprises the following steps: obtaining effective data resources;
the state monitoring means that during the operation of the equipment, various sensors are used for monitoring the equipment to obtain equipment state data, and a monitoring mode and sensor arrangement are generally established and installed by equipment manufacturers and delivered to users together with the equipment. Common monitoring modes include vibration monitoring, oil analysis, infrared imaging, ultrasonic monitoring and the like, and monitoring of environmental indexes such as temperature, humidity, wind speed and the like.
Various mathematical statistical indexes of the original signal output by the sensor can also be used as data resources of model training, such as monotonicity and concave-convex property of the signal, and indexes of frequency spectrum characteristics, peak value, peak-peak value, vibration intensity and the like of the vibration signal. In this embodiment, an acceleration sensor is used to perform vibration monitoring on a ship-borne ventilator, and a monitoring method and a data source are described, but not limited to this example, and a sensor arrangement point is shown in fig. 2.
Step two: and (2) carrying out data preprocessing on the original signal, wherein the data preprocessing comprises (1) data normalization, (2) signal denoising and (3) data segmentation.
(1) Data normalization
The data was normalized by Min-max according to the following formula:
in the formula x min And x max Respectively representing the minimum and maximum values of the original signal sequence, so that the output sequence of each group of sensors is linearly transformed to [0, 1%]Decimal within the interval.
(2) Signal de-noising
The method comprises the following steps that a sensor is used for monitoring shipborne equipment, in the processes of signal acquisition, transmission and processing, under the influence of factors such as environmental interference, sensor resolution, power frequency interference and the like, a large amount of noise exists in an original sensing signal, and data are subjected to denoising processing according to the following formula:
in the formula x i Denotes the ith element in the sequence, P (x) i ) Represents x i Denoised values. In this embodiment, data denoising with a sliding window l =5 is adopted.
(3) Data partitioning
Model training data set X training Collected by monitoring the state of M devices throughout their life cycle. The state monitoring starts from the start of the device, denoted T =0, and the device reaches the end of the failure threshold, denoted T = T, i.e. the total life of the device is T, T m Representing the total life of the mth device. the data collected at time t consists of the outputs of the F sensors, and is recorded as:
x f,t the output of the f-th sensor at the time t is represented, and then the data collected from the start of the operation of the m-th equipment to the full life cycle of the fault is obtained and recorded as:
in addition, the total service life T of different monitoring units is different under the influence of factors such as initial wear degree of equipment, manufacturing errors and randomness in the operation process, and the input of a residual service life prediction model is a vector with fixed dimension.
Therefore, it is necessary to divide the data into segments of length T m X of (2) m Dividing the data into data segments with the length L to obtain the data segments with the fixed shape of F x L:
the data segments are used as input of a model, wherein i =0,1 m 。
Step three: the condition-constrained multi-channel feature fusion residual service life prediction model is shown in FIG. 3:
a first stage;
and (4) enabling input data to enter a channel I, learning time dependence information of the input data by using the LSTM, and further inputting the characteristic information to an attention module to dynamically give a larger weight to more key characteristics to obtain time characteristic information.
And (3) the input data enters a channel II, the association degree of the sensor signals at different positions and the equipment degradation degree can be dynamically changed at different stages of equipment degradation, and a one-dimensional convolution neural network channel is adopted to extract the characteristics of the multi-sensor data to obtain important spatial information.
And a second stage:
the degradation process of the onboard equipment is irreversible from the delivery operation until the failure occurs, thus deriving the condition limit 1:
the degree of equipment failure is monotonically increasing and the degradation curve should be monotonically changing. From healthy operation to operation in the initial failure mode, not only the failure degree monotonically increases, but also the speed of the failure degree change increases, yielding the constraint 2:
the degradation curve has a concave-convex nature.
The constraint 1 is calculated by the following formula:
the constraint 2 is calculated by the following formula:
in the formula x i Denotes the ith element in the sequence, F (x) i ) Representing characteristic curves for all
If f (x) i ) Are all positive numbers, limiting the characteristic curve F to a monotonically increasing function, if F (x) i ) Are all negative numbers, and limit the characteristic curve F to be a monotone decreasing function; and g (x) i ) The positive and negative characteristics of (2) restrict the unevenness of the characteristic curve F.
The constraint condition vectors f, g of the feature information obtained in the first stage are calculated by using the formulas (3) and (4), and conditional screening is performed.
Where Fi denotes the ith feature vector, V denotes the number of feature vectors, T denotes the length of the feature vector, when N is Fi Conditional constraint vector f representing Fi i Number of negative numbers involved, monotonic decreasing characteristic of the limiting characteristic curve, N Fi Denotes f i The number of positive numbers involved limits the monotonic increasing nature of the characteristic curve.
Similarly, the optimum second constraint condition is selected according to equation (6)
。
The extracted features are supervised by the two conditions, and the monotonicity and the concave-convex property of the feature variables are limited, so that the model can learn useful features in data more efficiently, and the model can be converged quickly.
In the third stage, the limiting conditions calculated in the second stage and the feature information extracted in the first stage are spliced to be used as the input of a two-dimensional convolutional neural network for deeper feature extraction, and a first channel obtains a deep feature C 1 Channel two gets C 2 Compared with the common modes of feature splicing, feature adding and the like, the method can carry out deeper fusion on the features of different channels, thereby better describing the degradation process.
The converged network is shown in FIG. 4, first, C 1 ，C 2 Transforming into characteristic diagram with same size k m by deformation network
Then will be
Transposing to size m x k
And finally
And
calculating a fusion feature vector H by the formula (7) in which
Respectively represent
Is/are as follows
An element of a location.
k and m respectively represent
K is the width and m is the length. That is to say that
All of which are k m.
The fourth stage, inputting the fusion characteristic H into the multilayer full-connection layer to obtain a predicted value of the residual service life, and calculating the mean square error of the predicted value and the actual value according to a formula (8) as a loss function, wherein r j To enter the true remaining life label for the sequence,
n represents the number of input samples for the remaining life prediction of the corresponding sequence.
Step four: and performing joint training on the constructed residual service life prediction by using a federal learning mechanism. The federal remaining useful life prediction system training process proposed in this embodiment is shown in fig. 5:
the federal training preparation work is completed first: and the manufacturer technical center sends the established data processing specification and the built residual service life model to the users participating in the federal training, and the N users are assumed to participate in the federal training of the model. Thus completing the work of the second step and the third step. And the technical center server performs pre-training by using local data to obtain model initial parameters.
In fig. 5, 1 is to send updated model parameters to the user, 2 is to process data, train the local model and calculate the gradient in multiple rounds, 3 is to send encryption gradient, 4 is to perform security aggregation and gradient screening, and update parameters.
The federal training procedure is as follows:
1. the technical center model training parameters are sent to all federal training participants;
2. each participant processes data of own data, and performs multiple rounds of training on the model to apply a formula (8) to calculate loss, the model automatically calculates gradient g by using a back propagation algorithm, and performs algorithm encryption on the gradient g by using a reserved public key { n/e }, wherein the public key and the private key are calculated by an RSA algorithm and held by the participant and a technical center respectively, and the ciphertext is calculated by a formula (9);
3. all user terminals send encryption gradient G to the technical center;
4. and the technical center server receives the N groups of encryption gradients G of all the participants and decrypts the encryption gradients G by using the preserved private key { N/d }. In addition, in order to deal with the problem of non-independent and same distribution of scattered data, the decryption gradient is further screened.
The gradient of all users is clustered by utilizing a K-medoids algorithm, the number of clustering clusters is K, the setting of a K value is related to the distribution difference of scattered data, and the smaller the difference is, the smaller the K value is; one round of the federally trained gradient screening procedure is as follows:
gradient screening process
for each round of gradient screening t =1,2, … do
The K-medoids clustering algorithm willDDivided into K clusters
，
for each cluster D i， i=1，2，...Kdo
Averaging the gradients within each cluster
：
Cluster D i Centroids with other cluster centroids O h Expressed as d i,h ，h=1,2,...K；
Calculating the sum of distances
；
end for
for each cluster D i， i=1，2，...Kdo
for each cluster D j， j=1，2，...Kdo
Cluster calculation D j Participating cluster D j Weight W when updating parameter i,j ：
Screening Cluster D j Whether or not to participate in cluster D i Updating parameters:
IfW i,j >0.5, then D j participating in Cluster D i Updating parameters;
else D j not participating in the Cluster D i Updating of parameters, W i,j =0；
end for
Participation cluster D i Weighted gradient of parameter update
：
end for
Gradient set for all users
，g n Representing the gradient of the nth user; partitioning D into K clusters by clustering
The centroid corresponding to each cluster is respectively
Each cluster containing a number of samples of
The total number of samples for all clusters should be equal to N, i.e.
；
Calculating the average gradient of the gradient samples contained in each cluster according to the formula (11) to obtain
；
For the ith cluster D i Containing the number of samples m i The centroid is O i Mean of gradient samples within a cluster of
(ii) a Let the h-th cluster D h Center of mass O h And O i Has a Euclidean distance of d i,h Then there is the calculation cluster D of the formula (12) i Total distance S of centroid from all other clusters i ；
The weighting and screening mechanisms are performed by mathematical statistics of Euclidean distances, cluster D j Containing gradient samples in cluster D i Weight value W of the updated parameter i,j Calculated by the formula (13), the formula 13 shows that the cluster D is j And cluster D i Of Euclidean distance d i,h The larger the difference between the data distributions representing the two clusters, the cluster D j Containing gradient samples in cluster D i The smaller the weight value when updating the parameters of (1).
When W is i,j >0.5 Cluster participating Cluster D i Updating the parameters of (1); when W is i,j <At 0.5, it represents a cluster D j And cluster D i The Euclidean distance of (D) is too large, that is, the distribution difference of data samples is large, if cluster D is j Continue to participate in cluster D i The parameter update of (2) adversely affects the update of the parameter due to a large data distribution difference, and therefore, W is set to i,j =0, i.e. cluster D j Does not participate in the secondary round of parameter updating, thereby completing parameter weighting and screening.
Is represented by the formula (13) Calculate all clusters in cluster D i The weight value of (2) when updating the parameter.
Because of d i,i =0 can yield W i,i= 1；
Then cluster D i The weighted gradient of (c) is calculated by equation (14):
updating the local model by using the screened gradient, sending the updated local model to users contained in the cluster, and repeatedly finishing model updating of K clusters; and (4) testing whether the updated model achieves the expected predicted performance by the user by using own data, returning to the step one if the updated model does not achieve the expected predicted performance, repeatedly carrying out the federal training, and stopping the training if the updated model achieves the expected predicted performance.
Example two
The object of this embodiment is to provide a federal remaining service life prediction device of shipborne equipment, including: a technology center server, the technology center server comprising:
a model building module configured to: the technical center utilizes the historical monitoring data and the real-time data of the shipborne equipment stored by the technical center to construct a residual service life prediction model based on machine learning; and enabling a designer to build a deep learning prediction model through a model building module.
A training module configured to: performing joint training on the built residual service life prediction model based on machine learning by using a federal learning mechanism, and performing joint training by using N dispersed data sets self-owned by each user during training;
the training module utilizes the processed historical data to carry out federal training on the model, and after the training is finished, the model has reliable residual service life prediction performance.
A prediction module configured to: and predicting the residual service life of the ship-borne equipment in the running state in real time and early warning potential equipment faults by using the trained residual service life prediction model based on machine learning.
The prediction module is used for inputting equipment state data obtained through real-time monitoring into the residual service life prediction model by a user to predict the residual service life in real time.
EXAMPLE III
Referring to fig. 6, the purpose of the present embodiment is to provide a system apparatus for predicting remaining useful life of federal nation, which includes a technology center server, a plurality of user terminals and a transmission link.
The user terminals are respectively communicated with the technical center server through transmission links;
the computing module of the technical center server is used for processing computer instructions and completing tasks such as data processing, machine learning model modeling, model training and the like;
the storage module is used for storing data, computer instructions, model codes and the like;
the display module is used for human-computer interface interaction;
the decryption module is used for processing the encrypted data;
the screening module is used for screening the received data; the data acquisition module is used for aggregating the monitoring data;
the transmission link is used for completing information transmission between the technical center and the user.
The user terminal computing module is used for processing computer instructions and completing tasks such as data processing, model training and the like; the storage module is used for storing data, computer instructions, model codes and the like; the display module is used for tasks such as human-computer interaction, data visualization and the like; the encryption module is used for encrypting data; the data acquisition module is used for aggregating the monitoring data.
For the specific data processing procedure of the user terminal and the technical center server, reference is made in detail to the data processing method in the first embodiment.
Those skilled in the art will appreciate that the modules or steps of the present invention described above can be implemented using general purpose computer means, or alternatively, they can be implemented using program code that is executable by computing means, such that they are stored in memory means for execution by the computing means, or they are separately fabricated into individual integrated circuit modules, or multiple modules or steps of them are fabricated into a single integrated circuit module. The present invention is not limited to any specific combination of hardware and software.
Although the embodiments of the present invention have been described with reference to the accompanying drawings, it is not intended to limit the scope of the present invention, and it should be understood by those skilled in the art that various modifications and variations can be made without inventive efforts by those skilled in the art based on the technical solution of the present invention.
Claims (8)
1. A method for predicting the residual federal service life of shipborne equipment is characterized by comprising the following steps:
establishing a residual service life prediction model based on machine learning by using historical monitoring data and real-time data of shipborne equipment and pre-training;
carrying out federal training on the built residual service life prediction model by using a federal training mode, and carrying out joint training by using a multi-user scattered data set during training;
predicting the residual service life of the ship-mounted equipment in the running state in real time and early warning potential equipment faults by using a trained residual service life prediction model based on machine learning;
the method comprises the following steps of performing joint training on a built residual service life prediction model based on machine learning by using a federal learning mechanism, and specifically comprises the following steps:
sending the model training parameters to all federal training participants;
each participant processes data of self-contained shipborne equipment, performs multi-round training on the model and calculates loss, the model automatically calculates gradient by using a back propagation algorithm, and then performs algorithm encryption on the gradient by using a reserved public key;
all user terminals send encryption gradients to a technical center server;
the technical center server receives the N groups of encryption gradients of all the participants and decrypts the encryption gradients by using the preserved private key to obtain the gradients;
clustering the gradient obtained by decryption and screening to solve the problem of non-independent and same distribution of the multi-party scattered data, updating the local model by using the screened gradient, sending the updated local model to users in the cluster, and repeatedly finishing model updating of all the clusters;
judging whether the updated model tested by the user with own data reaches the expected prediction performance, if not, repeatedly carrying out federal training, and if so, stopping training;
and screening the gradient obtained by decryption, specifically:
calculating the average gradient of the gradient samples contained in each cluster aiming at the gradient set of all users; calculating the gradient samples contained in other clusters in the participating clustersD i Weight value when updating parameterW i,j The formula is as follows:
in the above formula, the first and second carbon atoms are,S i into a clusterD i The total distance from the centroid of all other clusters,d i,h is as followshAn individual clusterD h Center of massO h AndO i the Euclidean distance of (a) is,kto be the number of clusters to be clustered,jthe value range of (a) is [1,k]；
the screening mechanism isW i,j >0.5 cluster participation clusterD i Otherwise, the parameters ofW i,j And =0, namely, not participating in the secondary round parameter update.
2. The method for predicting the federal remaining service life of a shipborne device as claimed in claim 1, wherein the built machine learning-based remaining service life prediction model specifically comprises:
a first channel in which time-dependent information of input data is learned using an LSTM to obtain time characteristic information;
the second channel is used for extracting the characteristics of the multi-sensor data to obtain important spatial information;
calculating a condition vector of the acquired time characteristic information and the space important information by using the limiting condition, and supervising the parameter updating of the characteristic learning network by using the condition vector;
splicing the time characteristic information and the space important information obtained by the two channels in the last step with respective limiting conditions respectively, and taking the spliced time characteristic information and the space important information as the input of two-dimensional convolution networks respectively to carry out deeper characteristic extraction;
fine-grained fusion is carried out on the characteristics through a characteristic fusion network, so that the degradation process is better described;
and the multilayer full-connection layer is used for processing the fusion characteristics to obtain a predicted value of the residual service life.
3. The method for predicting the federal remaining useful life of a ship-borne device as claimed in claim 1, wherein the data in the model training data set is specifically:
tthe data collected at any moment is composed ofFComposition of output of individual sensors, note
；
To representtAt the first momentfThe output of the sensor, thereby obtainingmSensor data of individual devices, note
；
Condition monitoring begins with equipment startup, as indicatedt=0, the device reaches the end of the failure threshold, noted ast=TI.e. total life of the device isT，T m Is shown asmTotal life of the stage equipment.
4. The method for forecasting the federal remaining service life of a ship-borne device as claimed in claim 1, wherein before the step of constructing the machine learning based remaining service life forecasting model, the method further comprises: and acquiring state monitoring data, namely monitoring the operation period of the shipborne equipment by utilizing various sensors to acquire the state data of the shipborne equipment, wherein the monitoring mode and the sensor arrangement are established and installed by an equipment manufacturer.
5. The method for predicting the federal residual service life of a ship-borne device as claimed in claim 1, wherein the data preprocessing is performed on the acquired state monitoring data, and the data preprocessing includes data normalization, signal denoising and data segmentation.
6. The method as claimed in claim 5, wherein the data segmentation is implemented by segmenting data in a model training data set, and dividing the data into data segments of a set length according to a segmentation step.
7. A device for predicting the residual service life of the federal of shipborne equipment is characterized by comprising: a technology center server, the technology center server comprising:
a model building module configured to: the technical center utilizes the historical monitoring data and the real-time data of the self-contained shipborne equipment to construct a residual service life prediction model based on machine learning;
a training module configured to: performing joint training on the built residual service life prediction model based on machine learning by using a federal learning mechanism, and performing joint training by using N dispersed data sets self-owned by each user during training;
a prediction module configured to: predicting the residual service life of the ship-mounted equipment in the running state in real time and early warning potential equipment faults by using a trained residual service life prediction model based on machine learning;
the method comprises the following steps of performing joint training on a built residual service life prediction model based on machine learning by using a federal learning mechanism, and specifically comprises the following steps:
sending the model training parameters to all federal training participants;
each participant processes data of self-contained shipborne equipment, performs multi-round training on the model and calculates loss, the model automatically calculates gradient by using a back propagation algorithm, and then performs algorithm encryption on the gradient by using a reserved public key;
all user terminals send encryption gradients to a technical center server;
the technical center server receives the N groups of encryption gradients of all the participants and decrypts the encryption gradients by using the preserved private key to obtain the gradients;
clustering the gradient obtained by decryption and screening to solve the problem of non-independent and same distribution of the multi-party scattered data, updating the local model by using the screened gradient, sending the updated local model to users in the cluster, and repeatedly finishing model updating of all the clusters;
judging whether the updated model tested by the user with own data reaches the expected prediction performance, if not, repeatedly carrying out federal training, and if so, stopping training;
and screening the gradient obtained by decryption, specifically:
calculating the average gradient of the gradient samples contained in each cluster aiming at the gradient set of all users; calculating the gradient samples contained in other clusters in the participating clustersD i Weight value when updating parameterW i,j The formula is as follows:
in the above formula, the first and second carbon atoms are,S i into a clusterD i The total distance from the centroid of all other clusters,d i,h is as followshAn individual clusterD h Center of massO h AndO i the Euclidean distance of (a) is,kto be the number of the clusters to be clustered,jthe value range of (a) is [1,k]；
the screening mechanism isW i,j >0.5 cluster participation clusterD i Otherwise, the parameters ofW i,j And =0, namely, not participating in the secondary round parameter update.
8. A system for predicting the residual federal service life of shipborne equipment is characterized by comprising the following components:
a technical center server, a plurality of user terminals and a transmission link;
the user terminals are respectively communicated with the technical center server through transmission links;
the technology center server includes:
a model building module configured to: constructing a residual service life prediction model based on machine learning by utilizing historical monitoring data and real-time data of shipborne equipment;
a training module configured to: performing joint training on the built residual service life prediction model based on machine learning by using a federal learning mechanism, and performing joint training by using N dispersed data sets self-owned by each user during training;
a prediction module configured to: predicting the residual service life of the ship-mounted equipment in the running state in real time and early warning potential equipment faults by using a trained residual service life prediction model based on machine learning;
the method comprises the following steps of performing joint training on a built residual service life prediction model based on machine learning by using a federal learning mechanism, and specifically comprises the following steps:
sending the model training parameters to all federal training participants;
each participant processes data of self-contained shipborne equipment, performs multi-round training on the model and calculates loss, the model automatically calculates gradient by using a back propagation algorithm, and then performs algorithm encryption on the gradient by using a reserved public key;
all user terminals send encryption gradients to a technical center server;
the technical center server receives the N groups of encryption gradients of all the participants and decrypts the encryption gradients by using the preserved private key to obtain the gradients;
clustering the gradient obtained by decryption and screening to solve the problem of non-independent and same distribution of the multi-party scattered data, updating the local model by using the screened gradient, sending the updated local model to users in the cluster, and repeatedly finishing model updating of all the clusters;
judging whether the updated model tested by the user with own data reaches the expected prediction performance, if not, repeatedly carrying out federal training, and if so, stopping training;
and screening the gradient obtained by decryption, specifically:
calculating the average gradient of the gradient samples contained in each cluster aiming at the gradient set of all users; calculating the gradient samples contained in other clusters in the participating clustersD i Weight value when updating parameterW i,j The formula is as follows:
in the above formula, the first and second carbon atoms are,S i into a clusterD i The total distance from the centroid of all other clusters,d i,h is as followshAn individual clusterD h Center of massO h AndO i the Euclidean distance of (a) is,kto be the number of clusters to be clustered,jthe value range of (a) is [1,k]；
the screening mechanism isW i,j >0.5 cluster participation clusterD i Otherwise, the parameters ofW i,j And =0, namely, not participating in the secondary round parameter update.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210483195.4A CN114580087B (en) | 2022-05-06 | 2022-05-06 | Method, device and system for predicting federal remaining service life of shipborne equipment |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210483195.4A CN114580087B (en) | 2022-05-06 | 2022-05-06 | Method, device and system for predicting federal remaining service life of shipborne equipment |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114580087A CN114580087A (en) | 2022-06-03 |
| CN114580087B true CN114580087B (en) | 2022-08-02 |
Family
ID=81784641
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210483195.4A Active CN114580087B (en) | 2022-05-06 | 2022-05-06 | Method, device and system for predicting federal remaining service life of shipborne equipment |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114580087B (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115876475B (en) * | 2023-02-06 | 2023-06-09 | 山东大学 | Fault diagnosis method, system, equipment and storage medium |
| CN118114056A (en) * | 2024-04-26 | 2024-05-31 | 北京东方通科技股份有限公司 | Uninterrupted system state evaluation method and system |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111274737A (en) * | 2020-02-25 | 2020-06-12 | 山东大学 | Method and system for predicting remaining service life of mechanical equipment |
Family Cites Families (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR20190103088A (en) * | 2019-08-15 | 2019-09-04 | 엘지전자 주식회사 | Method and apparatus for recognizing a business card using federated learning |
| CN112288097B (en) * | 2020-10-29 | 2024-04-02 | 平安科技（深圳）有限公司 | Federal learning data processing method, federal learning data processing device, computer equipment and storage medium |
| CN112712182B (en) * | 2021-03-29 | 2021-06-01 | 腾讯科技（深圳）有限公司 | Model training method and device based on federal learning and storage medium |
| CN113487083A (en) * | 2021-07-06 | 2021-10-08 | 新智数字科技有限公司 | Method and device for predicting residual service life of equipment, computer equipment and computer-readable storage medium |
| CN113536670B (en) * | 2021-07-09 | 2024-01-12 | 北京航空航天大学 | Federal migration learning method for predicting residual life of group product |
| CN114139441A (en) * | 2021-11-04 | 2022-03-04 | 中国华能集团清洁能源技术研究院有限公司 | HT-based method and system for determining residual service life of fan bearing |
| CN113962988B (en) * | 2021-12-08 | 2024-04-09 | 东南大学 | Power inspection image anomaly detection method and system based on federal learning |
| CN114330740A (en) * | 2021-12-17 | 2022-04-12 | 青岛鹏海软件有限公司 | Manufacturing equipment fault monitoring model training system based on federal learning |
-
2022
- 2022-05-06 CN CN202210483195.4A patent/CN114580087B/en active Active
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111274737A (en) * | 2020-02-25 | 2020-06-12 | 山东大学 | Method and system for predicting remaining service life of mechanical equipment |
Non-Patent Citations (2)
| Title |
|---|
| 基于LSTM的风机滚动轴承剩余寿命预测研究;薛嫣等;《工业控制计算机》;20200625(第06期);全文 * |
| 基于长短期记忆网络的齿轮剩余寿命预测研究;石慧等;《太原科技大学学报》;20201013(第05期);全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114580087A (en) | 2022-06-03 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN114580087B (en) | Method, device and system for predicting federal remaining service life of shipborne equipment | |
| CN107786369B (en) | Power communication network security situation perception and prediction method based on IRT (intelligent resilient test) hierarchical analysis and LSTM (local Scale TM) | |
| Liao et al. | Enhanced restricted Boltzmann machine with prognosability regularization for prognostics and health assessment | |
| US10592725B2 (en) | Neural network systems | |
| CN112763214B (en) | Rolling bearing fault diagnosis method based on multi-label zero-sample learning | |
| DE112021004261T5 (en) | DUALMODAL RELATIONSHIP NETWORKS FOR AUDIOVISUAL EVENT LOCALIZATION | |
| CN111539515A (en) | Complex equipment maintenance decision method based on fault prediction | |
| Henriksson et al. | Towards structured evaluation of deep neural network supervisors | |
| CN102208028A (en) | Fault predicting and diagnosing method suitable for dynamic complex system | |
| US11137322B2 (en) | Diagnosing method of engine condition and diagnostic modeling method thereof | |
| WO2020148992A1 (en) | Model generation device, model generation method, model generation program, model generation system, inspection system, and monitoring system | |
| CN112529234A (en) | Surface water quality prediction method based on deep learning | |
| CN111695594A (en) | Image category identification method and device, computer equipment and medium | |
| CN116340796B (en) | Time sequence data analysis method, device, equipment and storage medium | |
| CN111680407B (en) | Satellite health assessment method based on Gaussian mixture model | |
| CN114913923A (en) | Cell type identification method aiming at open sequencing data of single cell chromatin | |
| CN115114128A (en) | Satellite health state evaluation system and evaluation method | |
| CN114500335B (en) | SDN network flow control method based on fuzzy C-means and mixed kernel least square support vector machine | |
| CN115859777A (en) | Method for predicting service life of product system in multiple fault modes | |
| CN113344470B (en) | Intelligent power supply system management method based on block chain | |
| CN112528554A (en) | Data fusion method and system suitable for multi-launch multi-source rocket test data | |
| CN115618743B (en) | State evaluation method and state evaluation system of sighting telescope system | |
| CN115375038A (en) | Aircraft engine failure mode identification and service life prediction method | |
| Luo et al. | A novel method for remaining useful life prediction of roller bearings involving the discrepancy and similarity of degradation trajectories | |
| CN114154415A (en) | Equipment life prediction method and device |
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 |