CN116502696A - Rolling bearing residual life prediction method based on federal learning and model pruning - Google Patents
Rolling bearing residual life prediction method based on federal learning and model pruning Download PDFInfo
- Publication number
- CN116502696A CN116502696A CN202310405004.7A CN202310405004A CN116502696A CN 116502696 A CN116502696 A CN 116502696A CN 202310405004 A CN202310405004 A CN 202310405004A CN 116502696 A CN116502696 A CN 116502696A
- Authority
- CN
- China
- Prior art keywords
- model
- client
- pruning
- rolling bearing
- prediction
- 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.)
- Pending
Links
- 238000013138 pruning Methods 0.000 title claims abstract description 54
- 238000000034 method Methods 0.000 title claims abstract description 53
- 238000005096 rolling process Methods 0.000 title claims abstract description 46
- 238000000605 extraction Methods 0.000 claims abstract description 30
- 238000012549 training Methods 0.000 claims abstract description 30
- 238000007781 pre-processing Methods 0.000 claims abstract description 11
- 238000004891 communication Methods 0.000 claims description 12
- 230000006870 function Effects 0.000 claims description 12
- 230000004913 activation Effects 0.000 claims description 9
- 238000004364 calculation method Methods 0.000 claims description 6
- 238000012935 Averaging Methods 0.000 claims description 4
- 230000000694 effects Effects 0.000 abstract description 2
- 238000012360 testing method Methods 0.000 description 10
- 239000000306 component Substances 0.000 description 4
- 238000006731 degradation reaction Methods 0.000 description 3
- 238000011156 evaluation Methods 0.000 description 3
- 238000005070 sampling Methods 0.000 description 3
- 230000008901 benefit Effects 0.000 description 2
- 230000015556 catabolic process Effects 0.000 description 2
- 238000010276 construction Methods 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 238000011478 gradient descent method Methods 0.000 description 2
- 238000012544 monitoring process Methods 0.000 description 2
- 238000010606 normalization Methods 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 238000000638 solvent extraction Methods 0.000 description 2
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 1
- 238000005299 abrasion Methods 0.000 description 1
- 230000001133 acceleration Effects 0.000 description 1
- 230000004931 aggregating effect Effects 0.000 description 1
- 238000013528 artificial neural network Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 239000008358 core component Substances 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 230000001737 promoting effect Effects 0.000 description 1
Classifications
-
- 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
- G06N3/082—Learning methods modifying the architecture, e.g. adding, deleting or silencing nodes or connections
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01M—TESTING STATIC OR DYNAMIC BALANCE OF MACHINES OR STRUCTURES; TESTING OF STRUCTURES OR APPARATUS, NOT OTHERWISE PROVIDED FOR
- G01M13/00—Testing of machine parts
- G01M13/04—Bearings
- G01M13/045—Acoustic or vibration analysis
-
- 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
- G06N3/098—Distributed learning, e.g. federated learning
Landscapes
- Physics & Mathematics (AREA)
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Acoustics & Sound (AREA)
- Testing Of Devices, Machine Parts, Or Other Structures Thereof (AREA)
Abstract
The invention discloses a rolling bearing residual life prediction method based on federal learning and model pruning, which comprises the following steps: collecting vibration signals of the rolling bearing in the whole life cycle; dividing, storing and preprocessing the acquired samples to obtain sequence data which can be used as model input; constructing a prediction model comprising a multi-scale feature expansion module, a depth feature extraction module and a prediction module; in the federal learning framework, a central server and a plurality of clients are used for cooperatively training a prediction model, and a pruning strategy is matched, so that the model structure becomes light; and predicting the residual service life of the rolling bearing by using the finally obtained lightweight model. According to the invention, the prediction model is built for the distributed data which are not shared by each other through federal learning, so that the safety of the data of each client is greatly improved, and the prediction effect and the reasoning speed of the model are effectively improved by matching with pruning operation.
Description
Technical Field
The disclosure belongs to the technical field of bearing residual life prediction, and particularly relates to a rolling bearing residual life prediction method based on federal learning and model pruning.
Background
The rolling bearing is a core component of a rotary machine, is degraded due to abrasion during use, and is one of the most likely parts to fail. Monitoring the running state of the rolling bearing and developing a practical and effective residual life prediction research is important to improving the stability of a mechanical system.
With the rapid development of deep learning and sensor technology, a rolling bearing residual life prediction method based on a deep neural network has achieved remarkable results. Due to the strong nonlinear mapping capability, the deep network can fit the highly complex bearing degradation process under the support of massive monitoring data so as to realize life prediction. However, existing correlation models and training data have the following problems: 1) In order to extract deep degradation characteristics, the number of layers of the existing model is generally large, the parameter amount is large, and the reasoning speed is low; 2) The rolling bearing has the advantages that the whole life data acquisition is difficult, samples are fewer, the data island phenomenon exists, and in the traditional method, the privacy protection of a plurality of data possession ends is not considered in the centralized method of fusing data and then performing model training.
The above information disclosed in the background section is only for enhancement of understanding of the background of the invention and therefore may contain information that does not form the prior art that is already known in the country to a person of ordinary skill in the art.
Disclosure of Invention
Aiming at the defects in the prior art, the invention aims to provide a rolling bearing residual life prediction method based on federal learning and model pruning. According to the method, a life prediction model is built for the distributed rolling bearing data which are not shared with each other under federal learning, and a pruning strategy is matched, so that the model is in a lightweight structure.
In order to achieve the above object, the present disclosure provides the following technical solutions:
a rolling bearing residual life prediction method based on federal learning and model pruning comprises the following steps:
step 1, collecting vibration signals of the rolling bearing in the whole life cycle as samples;
step 2, taking part of the collected samples as a training set, taking a single sample as a basic unit, storing the samples of the training set to a plurality of clients, and preprocessing the samples of the training set of each client to obtain sequence data which is input as a model;
step 3, constructing a global model for predicting the residual life of the rolling bearing, wherein the global model comprises a multi-scale feature expansion module, a depth feature extraction module and a prediction module, the multi-scale feature expansion module comprises a full-connection layer FC1, a multi-scale convolution layer MC2 and a multi-scale convolution layer MC3, the depth feature extraction module comprises a convolution layer C1, a convolution layer C2, a convolution layer C3 and a convolution layer C4, and the prediction module comprises a full-connection layer FC2 and a regression layer;
step 4, in the federal learning framework, the central server and a plurality of clients cooperatively train a global model, and training the global model into a lightweight prediction model based on pruning strategies;
and 5, predicting the residual life of the rolling bearings of the same type based on the lightweight global model.
In the rolling bearing residual life prediction method based on federal learning and model pruning, preprocessing comprises the steps of performing fast Fourier transform on a vibration signal to obtain a frequency domain amplitude signal, performing normalization processing on the frequency domain amplitude, and dividing the obtained sequence into a plurality of short sequences serving as sequence data based on 1024 data points.
In the method for predicting the residual life of the rolling bearing based on federal learning and model pruning, the step 4 comprises the following steps:
step 4-1) setting the maximum communication round number N in federal learning round Characteristic number epsilon of pruning and initial prediction model beta 0 ；
Step 4-2) the central server will initially pre-provisionMeasurement model beta 0 Issuing to all clients;
step 4-3), in each round of communication iteration, each client performs parameter update on the received model, and uploads the updated model to a central server, and the central server aggregates the collected models to form a global model;
step 4-4), the central server transmits the global model to each client, each client evaluates the global model and marks a client k with the best prediction performance at the moment;
step 4-5), pruning the epsilon output features in the depth feature extraction module on the client k, uploading the pruned model to a central server, and then transmitting the pruned model to all clients by the central server;
step 4-6) repeating the steps 4-3) to 4-5), when the predictive performance of the global model starts to decrease in step 4-4), or the number of communication iteration rounds reaches N round When the model training process ends.
In the method for predicting the residual life of the rolling bearing based on federal learning and model pruning, the method for updating parameters of the received model by each client in the step 4-3) comprises the following steps:
client i (i=1, 2,., M) updates the received model parameters using a random gradient descent algorithm, namely:
in the left side beta i (t) model parameters updated by the client i in the t-th round, delta is the learning rate of each client,as a loss function.
In the rolling bearing residual life prediction method based on federal learning and model pruning, the method for collecting the models by the central server in step 4-3) comprises the following steps:
use of federation by a central serverThe average algorithm aggregates all the collected client model parameters, namely:
wherein beta is g (t) is a global model aggregated from a plurality of client model parameters at round t, M is the number of clients.
In the method for predicting the residual life of the rolling bearing based on federal learning and model pruning, the method for evaluating the global model by each client in the step 4-4) comprises the following steps:
the global model predictive performance is quantitatively evaluated by adopting Root Mean Square Error (RMSE), and the calculation formula is as follows:
in the formula, RUL j Andthe real life and the predicted life of the jth input sequence data are respectively, and Z is the total number of the input sequence data on one client.
In the rolling bearing residual life prediction method based on federal learning and model pruning, the step of pruning in the step 4-5) to remove epsilon output features in the depth feature extraction module comprises the following steps:
step 4-5-1), on a client k, carrying out primary parameter updating on the global model at the moment by utilizing local data, and recording all output characteristics of the depth characteristic extraction module and gradients of network activation functions of each layer on each output characteristic;
step 4-5-2) for the q-th output feature h q The importance of the feature is the gradient of the layer activation function CAnd the output characteristic h q The absolute value of the product, i.e.)>Calculating the importance of all output characteristics;
step 4-5-3) sorting the importance degree of all the output features in the depth feature extraction module, pruning the epsilon output features with the minimum importance degree.
Compared with the prior art, the beneficial effects that this disclosure brought are:
according to the invention, the federal learning technology is used for constructing the prediction model for the distributed data which are not shared with each other, so that the safety of the data of each client is greatly improved, and the problem of data island in life prediction research is solved; the multi-scale feature expansion module constructed by the invention can effectively expand the features of the input signals, provides enough basic features for the extraction of the subsequent deep features and the regression of the features, and improves the prediction accuracy of the model; according to the invention, some unimportant output features in the depth feature extraction module are removed through pruning operation, so that the module becomes light-weighted, and the reasoning speed of the prediction model is improved.
Drawings
In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings that are needed in the embodiments will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments described in the present invention, and other drawings may be obtained according to these drawings for a person having ordinary skill in the art.
In the drawings:
FIG. 1 is a flow chart of a method for predicting the remaining life of a rolling bearing based on federal learning and model pruning provided in one embodiment of the present disclosure;
FIG. 2 is a block diagram of a predictive model provided by one embodiment of the present disclosure;
FIG. 3 is an overall architecture diagram of model training under federal learning provided in one embodiment of the present disclosure;
FIG. 4 is a graph of predicted results on test data provided by one embodiment of the present disclosure.
The invention is further explained below with reference to the drawings and examples.
Detailed Description
Specific embodiments of the present disclosure will be described in more detail below with reference to fig. 1 to 4. While specific embodiments of the disclosure are shown in the drawings, it should be understood that the disclosure may be embodied in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.
It should be noted that certain terms are used throughout the description and claims to refer to particular components. Those of skill in the art will understand that a person may refer to the same component by different names. The description and claims do not identify differences in terms of components, but rather differences in terms of the functionality of the components. As used throughout the specification and claims, the terms "include" and "comprise" are used in an open-ended fashion, and thus should be interpreted to mean "include, but not limited to. The description hereinafter sets forth the preferred embodiments for carrying out the present disclosure, but is not intended to limit the scope of the disclosure in general, as the description proceeds. The scope of the present disclosure is defined by the appended claims.
For the purposes of promoting an understanding of the embodiments of the disclosure, reference will now be made to the embodiments illustrated in the drawings and specific examples, without the intention of being limiting the embodiments of the invention.
For better understanding, as shown in fig. 1 to 4, a rolling bearing remaining life prediction method based on federal learning and model pruning includes the steps of:
1) And (3) data acquisition: collecting vibration signals of the whole life cycle of a plurality of rolling bearings;
2) Data partitioning, storing and preprocessing: dividing the collected multiple samples into a training set and a testing set, storing sample data of the training set to multiple clients by taking a single sample as a basic unit, and preprocessing the original data of each client and a testing part to obtain sequence data which can be input as a model;
3) Model construction: constructing a rolling bearing service life prediction model comprising a multi-scale feature expansion module, a depth feature extraction module and a prediction module;
4) Model training: in the federal learning framework, a central server and a plurality of clients cooperatively train a global model and match pruning strategies, so that the global model becomes a lightweight prediction model on the premise of not losing accuracy.
5) Model prediction: and predicting the residual life of the test set data by using the obtained lightweight global model. The test shows that the lightweight predictive model can well predict the residual life of the rolling bearings of the same type. See the data below for details.
In one embodiment, the multi-scale feature expansion module of the prediction model comprises a full-connection layer FC1, a multi-scale convolution layer MC2 and a multi-scale convolution layer MC3, the depth feature extraction module comprises a convolution layer C1, a convolution layer C2, a convolution layer C3 and a convolution layer C4, and the prediction module comprises a full-connection layer FC2 and a regression layer.
In one embodiment, the sub-steps of building the lightweight predictive model in the step 4) are:
4-1) setting the maximum communication round number N in Federal learning round Characteristic number epsilon of pruning and initial model beta 0 ；
4-2) the central server will initiate a predictive model β 0 Issuing to all clients;
4-3) in each round of communication iteration, each client performs parameter update on the received model, and uploads the updated model to a central server, and the central server aggregates the collected models to form a global model;
4-4) the central server transmits the global model to each client, each client evaluates the global model and marks the client k with the best performance at the moment;
4-5) pruning the epsilon output features in the prediction model depth feature extraction module on the client k, uploading the pruned model to a central server, and then transmitting the pruned model to all clients by the central server;
4-6) repeating steps 4-3) to 4-5), when the predictive performance of the global model begins to drop in step 4-4), or the number of communication iteration rounds reaches N round When the model training process is finished;
further, the method for updating parameters of the received model by each client in the step 4-3) includes:
client i (i=1, 2,., M) uses the local data to update parameters of the received model under gradient descent method, i.e.:
in the left side beta i (t) model parameters updated by the client i in the t-th round, delta is the learning rate of each client,as a loss function.
Further, the method for the central server to aggregate all the model parameters in the step 4-3) includes:
the central server uses the federal averaging algorithm to aggregate all the received client model parameters, namely:
wherein beta is g (t) is a global model aggregated from a plurality of client model parameters at round t, M is the number of clients.
Further, the method for evaluating the global model by each client in the step 4-4) is as follows:
the prediction performance of the model is quantitatively characterized by adopting Root Mean Square Error (RMSE) as an evaluation index, and the calculation formula is as follows:
in the formula, RUL j Andthe real life and the predicted life of the j-th input sequence data are respectively, Z is the total sequence number on one client, and the smaller the RMSE value is, the better the prediction performance of the model is.
Further, the sub-steps of pruning epsilon output features in the predictive model depth feature extraction module in step 4-5) are as follows:
4-5-1) on the client k, updating the prediction model at this time by using local data once, and recording all output features of four convolution layers on the model depth feature extraction module and gradients of network activation functions of each layer on each output feature;
4-5-2) for the q-th output feature h q The importance of the feature is the gradient of the layer activation function CAnd the output characteristic h q The absolute value of the product, i.e.)>Calculating the importance of all output characteristics;
4-5-3) sorting the importance of all the output features in the depth feature extraction module, and pruning epsilon output features with minimum importance.
In one embodiment, a method for predicting the remaining life of a rolling bearing based on federal learning and model pruning mainly comprises the following steps:
1) And (3) data acquisition: and collecting vibration signals of the rolling bearing in the whole life cycle.
In this embodiment, the vibration signal of the rolling bearing during operation can be detected by an acceleration sensor. The direction of the vibration signal comprises a horizontal direction and a vertical direction, the sampling frequency is 25.6kHz, the sampling duration is 1.28 seconds, and the sampling interval is 60 seconds.
2) Data partitioning, storing and preprocessing: dividing the collected multiple samples into a training set and a testing set, storing sample data of the training set to multiple clients by taking a single sample as a basic unit, and preprocessing the original data of each client and a testing part to obtain sequence data which can be input as a model;
in this embodiment, the preprocessing operation includes performing a fast fourier transform on the original signal to obtain a frequency domain amplitude signal thereof, performing normalization on the frequency domain amplitude, and dividing the obtained sequence into a plurality of short sequences as inputs of a prediction model based on 1024 data points.
3) Model construction: and constructing a rolling bearing service life prediction model comprising a multi-scale feature expansion module, a depth feature extraction module and a prediction module.
In this embodiment, as shown in fig. 2, the structure of the constructed prediction model is that the multi-scale feature expansion module includes a full connection layer FC1, a multi-scale convolution layer MC2, and a multi-scale convolution layer MC3, where the "conversion" operation converts the FC1 layer output into a feature with a channel number of 64 and a shape of 32×1, and the "connection by channel" operation stacks the three multi-scale convolved outputs along the channel direction to form a feature with a channel number of 96; the depth feature extraction module comprises four one-dimensional convolution layers, namely a convolution layer C1, a convolution layer C2, a convolution layer C3 and a convolution layer C4, wherein the output of the C4 layer is flattened into a one-dimensional array through 'flattening' operation; the prediction module comprises a full connection layer FC2 and a regression layer, wherein the regression layer outputs a life prediction result, and all layers adopt a tanh activation function, and the full connection layer Dropout is 0.5. The structure and parameters of each module of the prediction model are shown in table 1.
TABLE 1 Structure and parameters of modules of Global model
4) Model training: in the federal learning framework, a central server establishes a global model for local data stored by the central server in combination with a plurality of clients, and the global model becomes a lightweight prediction model under the premise of not losing precision by matching with a pruning strategy.
In this embodiment, the overall architecture of model training under federal learning is as shown in fig. 3, setting the total number of clients m=3.
Further, the sub-steps of building the lightweight predictive model in the step 4) are as follows:
4-1) setting the maximum communication round number N in Federal learning round 10, number of features pruned epsilon=5, initial model beta with randomization parameters 0 ；
4-2) the central server will initiate the model β 0 Issuing to all clients;
4-3) in each round of communication iteration, each client updates the received model, and uploads the updated model to a central server, and the central server aggregates all the models to form a global model;
further, the method for updating the received model by each client in the step 4-3) includes:
client i (i=1, 2,., M) uses a random gradient descent method, with its local pre-processing data to update the received model parameters with Adam optimization algorithm, namely:
in the left side beta i (t) model parameters updated by the client i in the t-th round, learning rate delta=0.001 of each client, and loss functionMean Square Error (MSE), the calculation formula is:
RUL j andthe real life and the predicted life of the jth input sequence data are respectively, and Z is the total number of input sequences on one client.
Further, in the step 4-3), the method for aggregating all models by the central server is as follows:
the central server aggregates all client model parameters using the federal averaging algorithm, i.e
Wherein beta is g (t) is a global model aggregated from a plurality of client model parameters at round t, and m=3 is the number of clients.
4-4) the central server transmits the global model to each client, each client evaluates the global model and marks the client k with the best performance at the moment;
further, the method for evaluating the global model by each client in the step 4-4) is as follows:
the prediction performance of the model is quantitatively characterized by adopting Root Mean Square Error (RMSE) as an evaluation index, and the calculation formula is as follows:
in the formula, RUL j Andthe real life and the predicted life of the j-th input sequence data are respectively, Z is the total sequence number on one client, and the smaller the RMSE value is, the better the prediction performance of the model is.
4-5) on the client k, epsilon=5 output features in the depth feature extraction module of the pruning prediction model, uploading the pruning model to a central server, and then issuing the pruning model to all clients by the central server;
further, the sub-steps of pruning epsilon output features in the predictive model depth feature extraction module in step 4-5) are as follows:
4-5-1) on the client k, updating the prediction model at this time by using local data once, and recording all output features of four convolution layers on the model depth feature extraction module and gradients of network activation functions of each layer on each output feature;
4-5-2) for the q-th output feature h q The importance of the feature is the gradient of the layer activation function CAnd the output characteristic h q The absolute value of the product, i.e.)>Calculating the importance of all output characteristics;
4-5-3) sorting the importance of all the output features in the depth feature extraction module, and pruning epsilon output features with minimum importance.
4-6) repeating steps 4-3) to 4-5), when the predictive performance of the global model begins to drop in step 4-4), or the number of communication iteration rounds reaches N round At=10, the model training process ends.
5) Model prediction: and predicting the residual service life of the test data by utilizing the finally obtained lightweight global model, wherein the result is as follows:
in this embodiment, the method of the present invention is performed on a working condition 2 data set of an XJTU-SY rolling bearing accelerated life test data set, under which 5 rolling bearing full life data are provided, and after the bearing 2_1, the bearing 2_2, the bearing 2_3 and the bearing 2_5 are preprocessed, 720, 2400, 4080 and 4000 input sequences with 1024 data points length are respectively obtained, the first three sets of sequence data are respectively put into 3 clients as training sets to perform model training, the bearing 25 performs model prediction, and two indexes of RMSE and MAE are adopted to evaluate model prediction performance, wherein the calculation formula of MAE is as follows:
in the formula, RUL j Andthe real life and the predicted life of the j-th input sequence data are respectively, and Z is the total sequence number on one client. The smaller the value of the index, the better the predictive performance of the model.
The model training method of the present invention was compared with the conventional FedAvg training method (Federated Averaging, abbreviated as FedAvg) and the results are shown in Table 2. It can be seen that the RMSE of the model obtained using the training method of the present invention was increased by about 22.32% compared to FedAvg and MAE was increased by about 16.87% because the complexity of the model was reduced, the saturation of the model was increased, and the overfitting of the model was prevented. Fig. 4 is a graph of the prediction result of the method provided by the invention on test data, and it can be seen that the prediction model can accurately track the degradation trend of the rolling bearing.
Table 2 comparison of the training method in the present invention with the FedAvg method
| Training method | RMSE | MAE |
| The invention is that | 0.087 | 0.069 |
| FedAvg | 0.112 | 0.083 |
The model structure constructed by the method of the invention is compared and analyzed with the model structure without the multi-scale characteristic expansion module, and the result is shown in table 3. The prediction performance of the model can be obviously improved by using the multi-scale feature expansion module, the RMSE is improved by about 37.41%, the MAE is improved by about 45.67%, the multi-scale feature expansion module is important to the whole model structure, the expanded basic features play a positive role for the subsequent depth feature extraction module and the prediction module, and the accurate prediction of the model is facilitated.
TABLE 3 comparison of results of the proposed model structure with model structure without multi-scale feature expansion module
| Model structure | RMSE | MAE |
| The invention is that | 0.087 | 0.069 |
| Multi-scale-free feature expansion module | 0.139 | 0.127 |
The lightweight global model obtained by the method is compared with the model before pruning, and the result is shown in table 4. The pruning operation cuts off about 57.80% of depth feature extraction module parameters, and the reasoning time of the obtained lightweight model is shortened by 13.53%. After pruning and updating for many times, the evaluation parameters of the prediction model are obviously improved, and the effectiveness of the invention is further proved.
TABLE 4 comparison of results of models before and after pruning in the method of the present invention
| Depth feature extraction module parameter | Model inference time | RMSE | MAE | |
| Before pruning | 173kB | 0.776s | 0.454 | 0.461 |
| After pruning | 73kB | 0.671s | 0.087 | 0.069 |
In summary, by means of the above technical solution of the present invention, data distributed on a plurality of clients is model-trained in a federal learning manner of "data motionless model", and a pruning strategy is matched to obtain a lightweight life prediction model suitable for a rolling bearing. The obtained model has good prediction performance and reasoning speed, the modeling process guarantees the safety of distributed data, and the model can be used for predicting the residual life of the rolling bearings of the same type.
Although embodiments of the present disclosure have been described above with reference to the accompanying drawings, the present disclosure is not limited to the specific embodiments and fields of application described above, which are merely illustrative, instructive, and not restrictive. Those skilled in the art, having the benefit of this disclosure, may effect numerous forms of the invention without departing from the scope of the invention as claimed.
Claims (7)
1. The method for predicting the residual life of the rolling bearing based on federal learning and model pruning is characterized by comprising the following steps of:
step 1, collecting vibration signals of the rolling bearing in the whole life cycle as samples;
step 2, taking part of the collected samples as a training set, taking a single sample as a basic unit, storing the samples of the training set to a plurality of clients, and preprocessing the samples of the training set of each client to obtain sequence data which is input as a model;
step 3, constructing a global model for predicting the residual life of the rolling bearing, wherein the global model comprises a multi-scale feature expansion module, a depth feature extraction module and a prediction module, the multi-scale feature expansion module comprises a full-connection layer FC1, a multi-scale convolution layer MC2 and a multi-scale convolution layer MC3, the depth feature extraction module comprises a convolution layer C1, a convolution layer C2, a convolution layer C3 and a convolution layer C4, and the prediction module comprises a full-connection layer FC2 and a regression layer;
step 4, in the federal learning framework, the central server and a plurality of clients cooperatively train a global model, and training the global model into a lightweight prediction model based on pruning strategies;
and 5, predicting the residual life of the rolling bearings of the same type based on the lightweight global model.
2. The method for predicting the residual life of a rolling bearing based on federal learning and model pruning according to claim 1, wherein the preprocessing preferably comprises performing fast fourier transform on the vibration signal to obtain a frequency domain amplitude signal thereof, normalizing the frequency domain amplitude, and dividing the obtained sequence into a plurality of short sequences as sequence data based on 1024 data points.
3. A method for predicting the remaining life of a rolling bearing based on federal learning and model pruning as set forth in claim 1, wherein said step 4 includes:
step 4-1) setting the maximum communication round number N in federal learning round Characteristic number epsilon of pruning and initial prediction model beta 0 ；
Step 4-2) the center server will initially predict the model β 0 Issuing to all clients;
step 4-3), in each round of communication iteration, each client performs parameter update on the received model, and uploads the updated model to a central server, and the central server aggregates the collected models to form a global model;
step 4-4), the central server transmits the global model to each client, each client evaluates the global model and marks a client k with the best prediction performance at the moment;
step 4-5), pruning the epsilon output features in the depth feature extraction module on the client k, uploading the pruned model to a central server, and then transmitting the pruned model to all clients by the central server;
step 4-6) repeating the steps 4-3) to 4-5), when the predictive performance of the global model starts to decrease in step 4-4), or the number of communication iteration rounds reaches N round When the model training process ends.
4. A method for predicting the remaining life of a rolling bearing based on federal learning and model pruning according to claim 3, wherein the method for updating parameters of the received model by each client in step 4-3) comprises the following steps:
client i (i=1, 2,., M) updates the received model parameters using a random gradient descent algorithm, namely:
in the left side beta i (t) model parameters updated by the client i in the t-th round, delta is the learning rate of each client,as a loss function.
5. A method for predicting the residual life of a rolling bearing based on federal learning and model pruning according to claim 3, wherein the step 4-3) the central server aggregates the collected models is as follows:
the central server uses the federal averaging algorithm to aggregate all the collected client model parameters, namely:
wherein beta is g (t) is a global model aggregated from a plurality of client model parameters at round t, M is the number of clients.
6. A method for predicting the residual life of a rolling bearing based on federal learning and model pruning according to claim 3, wherein the method for evaluating the global model by each client in step 4-4) is as follows:
the global model predictive performance is quantitatively evaluated by adopting Root Mean Square Error (RMSE), and the calculation formula is as follows:
in the formula, RUL j Andthe real life and the predicted life of the jth input sequence data are respectively, and Z is the total number of the input sequence data on one client.
7. A method of predicting the remaining life of a rolling bearing based on federal learning and model pruning as set forth in claim 3, wherein the step of pruning the epsilon output features of the depth feature extraction module in step 4-5) includes:
step 4-5-1), on a client k, carrying out primary parameter updating on the global model at the moment by utilizing local data, and recording all output characteristics of the depth characteristic extraction module and gradients of network activation functions of each layer on each output characteristic;
step 4-5-2) for the q-th output feature h q The importance of the feature is the gradient of the layer activation function CAnd the output characteristic h q The absolute value of the product, i.e.)>Calculating the importance of all output characteristics;
step 4-5-3) sorting the importance degree of all the output features in the depth feature extraction module, pruning the epsilon output features with the minimum importance degree.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310405004.7A CN116502696A (en) | 2023-04-14 | 2023-04-14 | Rolling bearing residual life prediction method based on federal learning and model pruning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310405004.7A CN116502696A (en) | 2023-04-14 | 2023-04-14 | Rolling bearing residual life prediction method based on federal learning and model pruning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116502696A true CN116502696A (en) | 2023-07-28 |
Family
ID=87329559
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310405004.7A Pending CN116502696A (en) | 2023-04-14 | 2023-04-14 | Rolling bearing residual life prediction method based on federal learning and model pruning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116502696A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117172312A (en) * | 2023-08-18 | 2023-12-05 | 南京理工大学 | Equipment fault diagnosis method based on improved federal element learning |
-
2023
- 2023-04-14 CN CN202310405004.7A patent/CN116502696A/en active Pending
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117172312A (en) * | 2023-08-18 | 2023-12-05 | 南京理工大学 | Equipment fault diagnosis method based on improved federal element learning |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108334948B (en) | Mechanical bearing fault diagnosis technology based on wide residual error network learning model | |
| CN110991690B (en) | Multi-time wind speed prediction method based on deep convolutional neural network | |
| CN109000930B (en) | Turbine engine performance degradation evaluation method based on stacking denoising autoencoder | |
| CN112254964A (en) | Rolling bearing fault diagnosis method based on rapid multi-scale convolution neural network | |
| CN107657250B (en) | Bearing fault detection and positioning method and detection and positioning model implementation system and method | |
| CN109829236A (en) | A kind of Compressor Fault Diagnosis method based on XGBoost feature extraction | |
| CN109946080B (en) | Mechanical equipment health state identification method based on embedded circulation network | |
| CN106022954B (en) | Multiple BP neural network load prediction method based on grey correlation degree | |
| CN112651167A (en) | Semi-supervised rolling bearing fault diagnosis method based on graph neural network | |
| CN106920007A (en) | PM based on second order Self-organized Fuzzy Neural Network2.5Intelligent Forecasting | |
| CN110647830A (en) | Bearing fault diagnosis method based on convolutional neural network and Gaussian mixture model | |
| CN108415884B (en) | Real-time tracking method for structural modal parameters | |
| CN116502696A (en) | Rolling bearing residual life prediction method based on federal learning and model pruning | |
| CN113761777B (en) | HP-OVMD-based ultra-short-term photovoltaic power prediction method | |
| CN116244647A (en) | Unmanned aerial vehicle cluster running state estimation method | |
| CN114429152A (en) | Rolling bearing fault diagnosis method based on dynamic index antagonism self-adaption | |
| CN114547795A (en) | Rolling bearing residual life prediction method based on data driving | |
| CN115587290A (en) | Aero-engine fault diagnosis method based on variational self-coding generation countermeasure network | |
| CN112101659A (en) | Complex equipment service life prediction method based on stacking denoising autoencoder | |
| CN116665483A (en) | Novel method for predicting residual parking space | |
| CN116701871A (en) | Rolling bearing composite fault feature selection method based on causal feature weighting network | |
| CN114942140A (en) | Rolling bearing fault diagnosis method based on multi-input parallel graph convolution neural network | |
| CN111353640A (en) | Method for constructing wind speed prediction model by combination method | |
| CN114841412A (en) | Method for predicting pH value of sea cucumber growing water | |
| CN112683532B (en) | Cross-working condition countermeasure diagnostic method for bearing |
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 |