CN111625992B - Mechanical fault prediction method based on self-optimal deep learning - Google Patents
Mechanical fault prediction method based on self-optimal deep learning Download PDFInfo
- Publication number
- CN111625992B CN111625992B CN202010437200.9A CN202010437200A CN111625992B CN 111625992 B CN111625992 B CN 111625992B CN 202010437200 A CN202010437200 A CN 202010437200A CN 111625992 B CN111625992 B CN 111625992B
- Authority
- CN
- China
- Prior art keywords
- network
- fault diagnosis
- training
- main
- cnn
- 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]
- G06F30/20—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
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T90/00—Enabling technologies or technologies with a potential or indirect contribution to GHG emissions mitigation
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Software Systems (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Mathematical Physics (AREA)
- Computational Linguistics (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Medical Informatics (AREA)
- Computer Hardware Design (AREA)
- Geometry (AREA)
- Testing And Monitoring For Control Systems (AREA)
Abstract
The invention provides a mechanical fault prediction method based on self-optimization deep learning, which comprises the following steps: constructing a fault diagnosis model based on CNN; constructing a reinforcement learning model by utilizing the fault diagnosis model; training the reinforcement learning model, and utilizing the reinforcement learning model to adaptively adjust the learning rate of the fault diagnosis model; and performing fault diagnosis by using the fault diagnosis model capable of adaptively adjusting the learning rate. The beneficial effects of the invention are as follows: a convolutional neural network reinforcement learning hybrid model is constructed, and the model realizes automatic adjustment of learning rate according to the real-time state of a CNN-based fault diagnosis model, so that the learning efficiency and learning effect of the fault diagnosis model are improved, and the fault diagnosis performance is improved.
Description
Technical Field
The invention relates to the field of fault prediction, in particular to a mechanical fault prediction method based on self-optimizing deep learning.
Background
Fault diagnosis plays an important role in modern equipment manufacturing. The fault diagnosis method can detect whether the equipment is in dangerous working state or not so as to ensure safe working state or save time to take corresponding remedial measures. The accurate and efficient fault diagnosis method is widely focused by academia and engineering. Deep learning is the most commonly used analysis method in data driving fault diagnosis, and has been widely applied to bearings, gear boxes, wind turbines, reciprocating compressors and the like, and a great deal of research results are achieved.
However, the performance of deep learning depends on the adjustment of the superparameter, whereas default superparameters do not guarantee performance across all datasets, so finding optimal superparameter combinations is an effective way to improve the performance of deep learning methods. The traditional parameter adjusting process is mainly trial and error or manual searching. Because the parameter adjusting process is a necessary link of the fault diagnosis method based on deep learning and needs to be carried out manually on each data set, the parameter adjusting result is highly dependent on manual experience and is time-consuming and labor-consuming. The method has the characteristics of low data standardization degree, multiple data sources, wide data distribution range and the like of fault diagnosis, and brings greater challenges to the parameter adjustment process of deep learning.
The learning rate is one of the most important parameters of the deep learning model. Too large learning rate will lead to divergent deep learning training process, and an effective fault diagnosis model cannot be obtained. Too small learning rate can cause the deep learning model to fall into a local optimal solution, so that the data characteristics of faults cannot be effectively extracted. Therefore, a suitable learning rate optimizing strategy is always one of the core contents of the fault diagnosis method based on deep learning.
Disclosure of Invention
Aiming at the problems, the invention provides a super-parameter tuning model based on reinforcement learning, and further realizes a fault prediction method based on self-tuning deep learning. The method can adaptively adjust learning rate parameters, extract a high-efficiency learning rate scheduling model aiming at a fault diagnosis data set, enable the learning rate to be always in a reasonable range, improve the data feature extraction capability of a deep learning model for fault diagnosis, and further improve the effect of final fault diagnosis.
S101: acquiring a fault diagnosis data set;
s102: constructing a fault diagnosis model; the model comprises a main CNN network, a game CNN network, a main Q network and a target Q network; the main CNN outputs a final fault diagnosis result; the game CNN network has the same structure as the main CNN network and is a clone network; the main Q network is a 3-layer artificial neural network structure; the target Q network has the same structure as the main Q network, and is a clone network; initializing learning rate eta of a main CNN network and a game CNN network, wherein training step t=0;
s103: establishing a characterization method of states, behaviors and rewards of a fault diagnosis model; when the fault diagnosis model CNN network performs t-th training, the state s t A method for representing a fault diagnosis model CNN network; the fault diagnosis model CNN network comprises a main CNN network and a game CNN network; the fault diagnosis model Q network predicts and obtains predicted reward values corresponding to all behaviors through the state of the fault diagnosis model CNN network, and selects the largest reward value y t Behavior a corresponding to the same t The method comprises the steps of carrying out a first treatment on the surface of the The fault diagnosis model Q network comprises a main Q network and a target Q network;
the fault diagnosis model Q network uses the behavior function to make the behavior a t The learning rate eta is used for updating, and the fault diagnosis model CNN network is trained by one step according to the learning rate eta, so that the training error f of the step is obtained t The prize value r is obtained by conversion t+1 At the same time, the CNN network reaches the next state s t+1 ;
Finally, state s t Behavior a t Prize value r t+1 Next state s t+1 Storing for standby;
s104: the training method of the reinforcement learning network structure comprises the following steps: by accumulating state s in the previous step t Behavior a t Prize value r t+1 Next state s t+1 Training a main Q network; the training process adopts a double Q-network training method, namely adopts a target Q network to predict the next state s t+1 Is the most significant of (3)The large reward value is used for constructing a training function of the main Q network; after the main Q network training is completed, cloning network parameters of the main Q network to a target Q network for next training;
s105: the fault diagnosis network model training method comprises the following steps: cloning network parameters of the main Q network to a game CNN network; training a game CNN network step; then after training the main Q network, cloning network parameters of the main Q network to a target Q network; training a step of a main CNN network to update the effect of the CNN network on fault diagnosis;
s106: and applying the trained main Q network to fault data diagnosis.
Further, the specific method for acquiring the fault diagnosis data set in step S101 is as follows: randomly intercepting time sequence signal samples from acquired vibration signals, converting signals in the time sequence signal samples from time domains to time domains by adopting S conversion, and adjusting the obtained two-dimensional matrix into a 224 multiplied by 224 dimensional matrix to serve as 1 piece of fault diagnosis data; m pieces of fault diagnosis data are collected together to form a fault diagnosis data set.
Further, in step S102, the fault diagnosis model CNN network is a network structure modified based on a classical LeNet-5 model, and includes six groups of alternating convolution layers and pooling layers, and includes 2 groups of full connection layers and Softmax fault classifiers, specifically: inserting a plurality of convolutional layers before the max pooling layer; comprising the following steps: 1 convolution layer 7 x 64 and 3 convolution layers 5 x 96 inserted before the first max pooling layer; 3 x 128 convolutional layers inserted before the second max pooling layer; 2 3 x 256 convolutional layers inserted before the third max pooling layer; 1 convolutional layer of 3×3×256 inserted before the fourth max pooling layer; wherein 3×3, 5×5, and 7×7 denote convolutional filter sizes of the convolutional layers of 3×3, 5×5, and 7×7, respectively; the 64, 128 and 256 represent the depths of the convolutional layers of 64, 128 and 256, respectively; the convolution layer steps were all 1×1 except for the 7×7×64 convolution layer step size of 2×2.
Further, in step S102, the fault diagnosis model Q network is a four-layer artificial neural network structure, and the network structure thereof is [6,16,16,5]; the input layer is 6-dimensional, corresponds to the dimension of the CNN network state s, and the output layer is 5-dimensional, corresponds to the state dimension of the behavior a; the hidden layer node number of the middle two layers is 16.
Further, in step S103, the state S of the fault diagnosis model CNN network t Represented as a set of 6-dimensional vectors, comprising: current learning rate, current training loss value, gradient of current networkSquare sum of (v), current iteration number, maximum minimum coding and state alignment;
the current learning rate, the current training loss value and the gradient of the current networkThe sum of squares of (2) and the current number of iterations is a known value;
the maximum and minimum codes [ s ] t ]For the current loss value f t Comparing the M minimum loss values F obtained in the training process t-1 As shown in formula (1); the state alignment represents the consistency of the gradient of the current step with the gradient of the previous step, as shown in formula (2);
[s t ] alignment =mean(sign(g t ·g t-1 )) (2)
in the formula (2) [ s ] t ] alignment Representing state alignment; sign (·) is a sign function and mean (·) is an average function.
Further, in step S103, the behavior a t The method comprises 5 steps of greatly increasing, slightly increasing, keeping unchanged, slightly decreasing and greatly decreasing respectively; the behaviour a t Updating the learning rate eta by matching with a behavior function: when the behavior a t For a large increase, the learning rate η=η/α 1 The method comprises the steps of carrying out a first treatment on the surface of the When the behavior is a small increase, the learning rate η=η/α 2 The method comprises the steps of carrying out a first treatment on the surface of the When the behavior a t In order to be kept unchanged, in order to keep the same,learning rate η=η; when the behavior a t For small amplitude decreases, the learning rate η=η×α 2 The method comprises the steps of carrying out a first treatment on the surface of the When the behavior a t To decrease greatly, the learning rate η=ηxα 1 The method comprises the steps of carrying out a first treatment on the surface of the Wherein alpha is 1 、α 2 Is a preset value.
In step S103, the prize value r t Representing CNN network execution behavior a t The evaluation value given later represents the execution behavior a t The quality of (3); the calculation is shown in formula (3):
in step S104, in the double Q-network training method, the target Q-network QNet' predicts the next state S t+1 The maximum prize value of the product multiplied by the discount factor is equal to the prize value r t+1 Together establishing the true prize value y forming the primary Q network QNet t As shown in formula (4); the predicted prize value for the primary Q network QNet isAs shown in formula (5); the training formula of the main Q network is shown in equation (6), i.e. minimizing the square difference of the predicted prize value and the true prize value:
y t =r t+1 +γmax a′ [QNet′(s t+1 ;θ)] a′ (4)
wherein, gamma is a discount factor, which is a preset value, and represents the weight of rewards, and the preset values are all 1; t represents the current time step, T is the preset maximum time step, and θ represents the weight of QNet.
A storage device stores instructions and data for implementing a mechanical failure prediction method based on self-tuning deep learning.
A mechanical failure prediction device based on self-optimizing deep learning, comprising: a processor and the storage device; the processor loads and executes the instructions and data in the storage device to realize a mechanical fault prediction method based on self-optimizing deep learning.
The technical scheme provided by the invention has the beneficial effects that: a convolutional neural network reinforcement learning hybrid model is constructed, and the model realizes automatic adjustment of learning rate according to the real-time state of a CNN-based fault diagnosis model, so that the learning efficiency and learning effect of the fault diagnosis model are improved, and the fault diagnosis performance is improved.
Drawings
The invention will be further described with reference to the accompanying drawings and examples, in which:
FIG. 1 is a flow chart of a method for mechanical failure prediction based on self-optimizing deep learning in an embodiment of the invention;
FIG. 2 is a schematic diagram of the operation of a hardware device in an embodiment of the invention.
Detailed Description
For a clearer understanding of technical features, objects and effects of the present invention, a detailed description of embodiments of the present invention will be made with reference to the accompanying drawings.
The embodiment of the invention provides a mechanical fault prediction method based on self-optimizing deep learning.
Referring to fig. 1, fig. 1 is a flowchart of a mechanical fault prediction method based on self-tuning deep learning in an embodiment of the invention, which specifically includes the following steps:
s101: acquiring a fault diagnosis data set;
s102: constructing a fault diagnosis model; the model comprises a main CNN network, a game CNN network, a main Q network and a target Q network; the main CNN outputs a final fault diagnosis result; the game CNN network has the same structure as the main CNN network and is a clone network; the main Q network is a 3-layer artificial neural network structure; the target Q network has the same structure as the main Q network, and is a clone network; initializing learning rate eta of a main CNN network and a game CNN network, wherein training step t=0;
s103: establishing a characterization method of states, behaviors and rewards of a fault diagnosis model; when the fault diagnosis model CNN network performs t-th training, the state s t A method for representing a fault diagnosis model CNN network; the fault diagnosis model CNN network comprises a main CNN network and a game CNN network; the fault diagnosis model Q network predicts and obtains predicted reward values corresponding to all behaviors through the state of the fault diagnosis model CNN network, and selects the largest reward value y t Behavior a corresponding to the same t The method comprises the steps of carrying out a first treatment on the surface of the The fault diagnosis model Q network comprises a main Q network and a target Q network;
the fault diagnosis model Q network uses the behavior function to make the behavior a t The learning rate eta is used for updating, and the fault diagnosis model CNN network is trained by one step according to the learning rate eta, so that the training error f of the step is obtained t The prize value r is obtained by conversion t+1 At the same time, the CNN network reaches the next state s t+1 ;
Finally, state s t Behavior a t Prize value r t+1 Next state s t+1 Storing for standby;
s104: the training method of the reinforcement learning network structure comprises the following steps: by accumulating state s in the previous step t Behavior a t Prize value r t+1 Next state s t+1 Training a main Q network; the training process adopts a double Q-network training method, namely adopts a target Q network to predict the next state s t+1 The maximum reward value of (2) is used for constructing a training function of the main Q network; after the main Q network training is completed, cloning network parameters of the main Q network to a target Q network for next training;
s105: the fault diagnosis network model training method comprises the following steps: cloning network parameters of the main Q network to a game CNN network; training a game CNN network step; then after training the main Q network, cloning network parameters of the main Q network to a target Q network; training a step of a main CNN network to update the effect of the CNN network on fault diagnosis;
s106: applying the trained main Q network to fault data diagnosis;
the specific method for acquiring the fault diagnosis data set in step S101 is as follows: randomly intercepting time sequence signal samples from acquired vibration signals, converting signals in the time sequence signal samples from time domains to time domains by adopting S conversion, and adjusting the obtained two-dimensional matrix into a 224 multiplied by 224 dimensional matrix to serve as 1 piece of fault diagnosis data; m pieces of fault diagnosis data are collected together to form a fault diagnosis data set.
In step S102, the fault diagnosis model CNN network is an improved network structure based on a classical LeNet-5 model, and includes six groups of alternating convolution layers and pooling layers, and includes 2 groups of full connection layers and Softmax fault classifiers, specifically: inserting a plurality of convolutional layers before the max pooling layer; comprising the following steps: 1 convolution layer 7 x 64 and 3 convolution layers 5 x 96 inserted before the first max pooling layer; 3 x 128 convolutional layers inserted before the second max pooling layer; 2 3 x 256 convolutional layers inserted before the third max pooling layer; 1 convolutional layer of 3×3×256 inserted before the fourth max pooling layer; wherein 3×3, 5×5, and 7×7 denote convolutional filter sizes of the convolutional layers of 3×3, 5×5, and 7×7, respectively; the 64, 128 and 256 represent the depths of the convolutional layers of 64, 128 and 256, respectively; the convolution layer steps were all 1×1 except for the 7×7×64 convolution layer step size of 2×2.
TABLE 1 Main convolutional neural network Structure
The main Q network adopts a 3-layer artificial neural network (QNet); in step S102, the fault diagnosis model Q network is a four-layer artificial neural network structure, and the network structure thereof is [6,16,16,5]; the input layer is 6-dimensional, corresponds to the dimension of the CNN network state s, and the output layer is 5-dimensional, corresponds to the state dimension of the behavior a; the hidden layer node number of the middle two layers is 16.
In step S101, the specific method for acquiring the fault diagnosis data set is as follows:
randomly intercepting time sequence signal samples from acquired vibration signals, converting signals in the time sequence signal samples from time domains to time domains by adopting S conversion, and adjusting the obtained two-dimensional matrix into a 224 multiplied by 224 dimensional matrix to serve as 1 piece of fault diagnosis data; similarly, 10000 pieces of fault diagnosis data are collected in total to form a fault diagnosis data set;
in step S103, the state S of the fault diagnosis model CNN network t Represented as a set of 6-dimensional vectors, comprising: current learning rate, current training loss value, gradient of current networkSquare sum of (v), current iteration number, maximum minimum coding and state alignment;
the current learning rate, the current training loss value and the gradient of the current networkThe sum of squares of (2) and the current number of iterations is a known value;
the maximum and minimum codes [ s ] t ]For the current loss value f t Comparing the M minimum loss values F obtained in the training process t-1 As shown in formula (1); the state alignment represents the consistency of the gradient of the current step with the gradient of the previous step, as shown in formula (2);
[s t ] alignment =mean(sign(g t ·g t-1 )) (2)
in the formula (2) [ s ] t ] alignment Representing state alignment; sign (·) is a sign function and mean (·) is an average function.
In step S103, the behavior a t The method comprises 5 steps of greatly increasing, slightly increasing, keeping unchanged, slightly decreasing and greatly decreasing respectively; the behaviour a t Updating the learning rate eta by matching with a behavior function: when the behavior a t For a large increase, the learning rate η=η/α 1 The method comprises the steps of carrying out a first treatment on the surface of the When the behavior is a small increase, the learning rate η=η/α 2 The method comprises the steps of carrying out a first treatment on the surface of the When the behavior a t To remain unchanged, the learning rate η=η; when the behavior a t For small amplitude decreases, the learning rate η=η×α 2 The method comprises the steps of carrying out a first treatment on the surface of the When the behavior a t To decrease greatly, the learning rate η=ηxα 1 The method comprises the steps of carrying out a first treatment on the surface of the Wherein alpha is 1 、α 2 Is a preset value; alpha 1 =0.99,α 2 =0.995。
In step S103, the prize value r t Representing CNN network execution behavior a t The evaluation value given later represents the execution behavior a t The quality of (3); the calculation is shown in formula (3):
in step S104, in the double Q-network training method, the target Q-network QNet' predicts the next state S t+1 The maximum prize value of the product multiplied by the discount factor is equal to the prize value r t+1 Together establishing the true prize value y forming the primary Q network QNet t As shown in formula (4); the predicted prize value for the primary Q network QNet isAs shown in formula (5); the training formula of the main Q network is shown in equation (6), i.e. minimizing the square difference of the predicted prize value and the true prize value:
y t =r t+1 +γmax a′ [QNet′(s t+1 ;θ)] a′ (4)
wherein, gamma is a discount factor, which is a preset value, and represents the weight of rewards, and the preset values are all 1; t represents the current time step, T is the preset maximum time step, and θ represents the weight of QNet.
The mechanical fault prediction method based on the convolutional neural network is further described in detail below with reference to examples.
Case study:
in the case OF the research, the proposed convolutional network reinforcement learning algorithm tests [5] in a Keste Chu Da (CWR) bearing data table, three error patterns are provided in the data table, namely a Roller Fault (RF), an outer ring fault (OF) and an inner ring fault (IF), and each error pattern has three damage degrees OF 0.18mm,0.36mm and 0.54mm respectively, wherein the nine error patterns and the normal state form ten health conditions OF the bearing. And collecting a vibration signal of the driving end to perform fault diagnosis. The experiment was performed under four load conditions with motor loads set to 0, 1, 2, 3hp, respectively. The initialization learning rate is set to 0.01.
Since there are four base load conditions and each test uses five fold cross validation, there are a total of 20 cases in this case study (4*5 =20). The predictive goal of this case study is to verify the performance of RL-CNN by comparison with other well-known DL and ML algorithms. In this section, the symbol "0-1" indicates that the RL-CNN is trained on 0 load and also tested on 1 load, and there are two configurations in the data. The first is from Chen et al, and the comparison results are shown in table two. The second is from Cinnabaris et al, and the comparison results are shown in Table III.
From Table II, the proposed RL-CNN-Ens is the most top technology of the current generation. The average prediction accuracy of RL-CNN-Ens is 96.76%, which exceeds DCN-SDR, TICNN, WDCNN and NSAE-LCN. In the present case study, the well-known ML algorithm, including ANN and SVM algorithms, were selected as baseline methods for comparison, and the results of the comparison showed that the RL-CNN was significantly improved
In Table three, RL-CNN-Ens is compared with other well-known DL and ML algorithms. The results indicate that RL-CNN-Ens is the best among these algorithms. The average prediction accuracy of RL-CNN-Ens is 97.16%, which is as good as ICN, but better than HCNN, WDCNN (AdaBN), and TICNN is integrated. In this comparison, a comparison was made using the well-known ResNet, alexNet as the baseline, the results of which confirm the performance of RL-CNN-Ens.
TABLE 2 results for CNN single full connection layer under Case1 (%)
TABLE 3 results for CNN single full connection layer under Case1 (%)
Referring to fig. 2, fig. 2 is a schematic working diagram of a hardware device according to an embodiment of the present invention, where the hardware device specifically includes: a mechanical failure prediction device 401, a processor 402 and a storage device 403 based on self-optimizing deep learning.
Mechanical failure prediction device 401 based on self-optimizing deep learning: the mechanical failure prediction device 401 based on self-tuning deep learning implements the mechanical failure prediction method based on self-tuning deep learning.
Processor 402: the processor 402 loads and executes instructions and data in the memory device 403 for implementing the one self-optimizing deep learning based mechanical failure prediction method.
Storage device 403: the storage device 403 stores instructions and data; the storage device 403 is configured to implement the mechanical failure prediction method based on self-optimizing deep learning.
The beneficial effects of the invention are as follows: a convolutional neural network reinforcement learning hybrid model is constructed, and the model realizes automatic adjustment of learning rate according to the real-time state of a CNN-based fault diagnosis model, so that the learning efficiency and learning effect of the fault diagnosis model are improved, and the fault diagnosis performance is improved.
The foregoing description of the preferred embodiments of the invention is not intended to limit the invention to the precise form disclosed, and any such modifications, equivalents, and alternatives falling within the spirit and scope of the invention are intended to be included within the scope of the invention.
Claims (10)
1. A mechanical fault prediction method based on self-adjusting optimal deep learning is characterized in that: the method comprises the following steps:
s101: acquiring a fault diagnosis data set;
s102: constructing a fault diagnosis model; the model comprises a main CNN network, a game CNN network, a main Q network and a target Q network; the main CNN outputs a final fault diagnosis result; the game CNN network has the same structure as the main CNN network and is a clone network; the main Q network is a 3-layer artificial neural network structure; the target Q network has the same structure as the main Q network, and is a clone network; initializing learning rate eta of a main CNN network and a game CNN network, wherein training step t=0;
s103: establishing a characterization method of states, behaviors and rewards of a fault diagnosis model; when the fault diagnosis model CNN network performs t-th training, the state s t A method for representing a fault diagnosis model CNN network; the fault diagnosis model CNN network comprises a main CNN network and a game CNN network; the fault diagnosis model Q network predicts and obtains predicted reward values corresponding to all behaviors through the state of the fault diagnosis model CNN network, and selects the largest reward value y t Behavior a corresponding to the same t The method comprises the steps of carrying out a first treatment on the surface of the The fault diagnosis model Q network comprises a main Q network and a target Q network;
the fault diagnosis model Q network uses the behavior function to make the behavior a t The learning rate eta is used for updating, and the fault diagnosis model CNN network is trained by one step according to the learning rate eta, so that the training error f of the step is obtained t The prize value r is obtained by conversion t+1 At the same time, the CNN network reaches the next state s t+1 ;
Finally, state s t Behavior a t Prize value r t+1 Next state s t+1 Storing for standby;
s104: the training method of the reinforcement learning network structure comprises the following steps: by accumulating state s in the previous step t Behavior a t Prize value r t+1 Next state s t+1 Training a main Q network; the training process adopts a double Q-network training method, namely adopts a target Q network to predict the next state s t+1 The maximum reward value of (2) is used for constructing a training function of the main Q network; after the main Q network training is completed, cloning network parameters of the main Q network to a target Q network for next training;
s105: the fault diagnosis network model training method comprises the following steps: cloning network parameters of the main Q network to a game CNN network; training a game CNN network step; then after training the main Q network, cloning network parameters of the main Q network to a target Q network; training a step of a main CNN network to update the effect of the CNN network on fault diagnosis;
s106: and applying the trained main Q network to fault data diagnosis.
2. The mechanical failure prediction method based on self-optimizing deep learning as claimed in claim 1, wherein: the specific method for acquiring the fault diagnosis data set in step S101 is as follows: randomly intercepting time sequence signal samples from acquired vibration signals, converting signals in the time sequence signal samples from time domains to time domains by adopting S conversion, and adjusting the obtained two-dimensional matrix into a 224 multiplied by 224 dimensional matrix to serve as 1 piece of fault diagnosis data; m pieces of fault diagnosis data are collected together to form a fault diagnosis data set.
3. The mechanical failure prediction method based on self-optimizing deep learning as claimed in claim 1, wherein: in step S102, the fault diagnosis model CNN network is an improved network structure based on a classical LeNet-5 model, and includes six groups of alternating convolution layers and pooling layers, and includes 2 groups of full connection layers and Softmax fault classifiers, specifically: inserting a plurality of convolutional layers before the max pooling layer; comprising the following steps: 1 convolution layer 7 x 64 and 3 convolution layers 5 x 96 inserted before the first max pooling layer; 3 x 128 convolutional layers inserted before the second max pooling layer; 2 3 x 256 convolutional layers inserted before the third max pooling layer; 1 convolutional layer of 3×3×256 inserted before the fourth max pooling layer; wherein 3×3, 5×5, and 7×7 denote convolutional filter sizes of the convolutional layers of 3×3, 5×5, and 7×7, respectively; the 64, 128 and 256 represent the depths of the convolutional layers of 64, 128 and 256, respectively; the convolution layer steps were all 1×1 except for the 7×7×64 convolution layer step size of 2×2.
4. The mechanical failure prediction method based on self-optimizing deep learning as claimed in claim 1, wherein: in step S102, the fault diagnosis model Q network is a four-layer artificial neural network structure, and the network structure thereof is [6,16,16,5]; the input layer is 6-dimensional, corresponds to the dimension of the CNN network state s, and the output layer is 5-dimensional, corresponds to the state dimension of the behavior a; the hidden layer node number of the middle two layers is 16.
5. The mechanical failure prediction method based on self-optimizing deep learning as claimed in claim 1, wherein: in step S103, the state S of the fault diagnosis model CNN network t Represented as a set of 6-dimensional vectors, comprising: current learning rate, current training loss value, gradient of current networkSquare sum of (v), current iteration number, maximum minimum coding and state alignment;
the current learning rate, the current training loss value and the gradient of the current networkThe sum of squares of (2) and the current number of iterations is a known value;
the maximum and minimum codes [ s ] t ]For the current loss value f t Comparing the M minimum loss values F obtained in the training process t-1 As shown in formula (1); the state pairThe alignment represents the consistency of the gradient of the current step with the gradient of the previous step as shown in formula (2);
[s t ] alignment =mean(sign(g t ·g t-1 )) (2)
in the formula (2) [ s ] t ] alignment Representing state alignment; sign (·) is a sign function and mean (·) is an average function.
6. The mechanical failure prediction method based on self-optimizing deep learning as claimed in claim 1, wherein: in step S103, the behavior a t The method comprises 5 steps of greatly increasing, slightly increasing, keeping unchanged, slightly decreasing and greatly decreasing respectively; the behaviour a t Updating the learning rate eta by matching with a behavior function: when the behavior a t For a large increase, the learning rate η=η/α 1 The method comprises the steps of carrying out a first treatment on the surface of the When the behavior is a small increase, the learning rate η=η/α 2 The method comprises the steps of carrying out a first treatment on the surface of the When the behavior a t To remain unchanged, the learning rate η=η; when the behavior a t For small amplitude decreases, the learning rate η=η×α 2 The method comprises the steps of carrying out a first treatment on the surface of the When the behavior a t To decrease greatly, the learning rate η=ηxα 1 The method comprises the steps of carrying out a first treatment on the surface of the Wherein alpha is 1 、α 2 Is a preset value.
7. The mechanical failure prediction method based on self-optimizing deep learning as claimed in claim 5, wherein: in step S103, the prize value r t Representing CNN network execution behavior a t The evaluation value given later represents the execution behavior a t The quality of (3); the calculation is shown in formula (3):
8. the mechanical failure prediction method based on self-optimizing deep learning as claimed in claim 1, wherein: in step S104, in the double Q-network training method, the target Q-network QNet' predicts the next state S t+1 The maximum prize value of the product multiplied by the discount factor is equal to the prize value r t+1 Together establishing the true prize value y forming the primary Q network QNet t As shown in formula (4); the predicted prize value for the primary Q network QNet isAs shown in formula (5); the training formula of the main Q network is shown in equation (6), i.e. minimizing the square difference of the predicted prize value and the true prize value:
y t =r t+1 +γmax a′ [QNet′(s t+1 ;θ)] a′ (4)
wherein, gamma is a discount factor, which is a preset value, and represents the weight of rewards, and the preset values are all 1; t represents the current time step, T is the preset maximum time step, and θ represents the weight of QNet.
9. A memory device, characterized by: the storage device stores instructions and data for implementing a mechanical failure prediction method based on self-optimizing deep learning according to any one of claims 1 to 8.
10. A mechanical failure prediction device based on self-adjusting optimal deep learning is characterized in that: comprising the following steps: a processor and the storage device; the processor loads and executes the instructions and data in the storage device to implement a self-tuning deep learning-based mechanical failure prediction method according to any one of claims 1 to 8.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010437200.9A CN111625992B (en) | 2020-05-21 | 2020-05-21 | Mechanical fault prediction method based on self-optimal deep learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010437200.9A CN111625992B (en) | 2020-05-21 | 2020-05-21 | Mechanical fault prediction method based on self-optimal deep learning |
Publications (2)
Publication Number | Publication Date |
---|---|
CN111625992A CN111625992A (en) | 2020-09-04 |
CN111625992B true CN111625992B (en) | 2023-06-27 |
Family
ID=72272005
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202010437200.9A Active CN111625992B (en) | 2020-05-21 | 2020-05-21 | Mechanical fault prediction method based on self-optimal deep learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN111625992B (en) |
Families Citing this family (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112801272A (en) * | 2021-01-27 | 2021-05-14 | 北京航空航天大学 | Fault diagnosis model self-learning method based on asynchronous parallel reinforcement learning |
CN113095367B (en) * | 2021-03-16 | 2023-04-18 | 华中科技大学 | Compressor fault diagnosis method based on deep reinforcement learning |
CN112924177B (en) * | 2021-04-02 | 2022-07-19 | 哈尔滨理工大学 | Rolling bearing fault diagnosis method for improved deep Q network |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109978134A (en) * | 2019-02-26 | 2019-07-05 | 华中科技大学 | A kind of failure prediction method based on fast integration convolutional neural networks |
CN110839031A (en) * | 2019-11-15 | 2020-02-25 | 中国人民解放军陆军工程大学 | Malicious user behavior intelligent detection method based on reinforcement learning |
CN111079926A (en) * | 2019-11-23 | 2020-04-28 | 同济大学 | Equipment fault diagnosis method with self-adaptive learning rate based on deep learning |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20200119556A1 (en) * | 2018-10-11 | 2020-04-16 | Di Shi | Autonomous Voltage Control for Power System Using Deep Reinforcement Learning Considering N-1 Contingency |
-
2020
- 2020-05-21 CN CN202010437200.9A patent/CN111625992B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109978134A (en) * | 2019-02-26 | 2019-07-05 | 华中科技大学 | A kind of failure prediction method based on fast integration convolutional neural networks |
CN110839031A (en) * | 2019-11-15 | 2020-02-25 | 中国人民解放军陆军工程大学 | Malicious user behavior intelligent detection method based on reinforcement learning |
CN111079926A (en) * | 2019-11-23 | 2020-04-28 | 同济大学 | Equipment fault diagnosis method with self-adaptive learning rate based on deep learning |
Non-Patent Citations (2)
Title |
---|
Long Wen等. A New Two-Level Hierarchical Diagnosis Network Based on Convolutional Neural Network.IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT.2020,第69卷(第02期),第330-338页. * |
许爱华 ; 杜洋 ; 袁涛 ; .基于深度学习的电机轴承故障诊断研究.组合机床与自动化加工技术.2020,(第03期),第45-48+54页. * |
Also Published As
Publication number | Publication date |
---|---|
CN111625992A (en) | 2020-09-04 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN111625992B (en) | Mechanical fault prediction method based on self-optimal deep learning | |
CN107330514B (en) | Air quality prediction method based on integrated extreme learning machine | |
CN115018021B (en) | Machine room abnormity detection method and device based on graph structure and abnormity attention mechanism | |
CN111709448A (en) | Mechanical fault diagnosis method based on migration relation network | |
CN111104522A (en) | Regional industry association effect trend prediction method based on knowledge graph | |
CN111553482B (en) | Machine learning model super-parameter tuning method | |
CN112132102B (en) | Intelligent fault diagnosis method combining deep neural network with artificial bee colony optimization | |
CN111680788A (en) | Equipment fault diagnosis method based on deep learning | |
CN111832663A (en) | Capsule network rolling bearing fault diagnosis system based on gate control circulation unit | |
CN116522795A (en) | Comprehensive energy system simulation method and system based on digital twin model | |
CN106296434A (en) | A kind of Grain Crop Yield Prediction method based on PSO LSSVM algorithm | |
CN115163424A (en) | Wind turbine generator gearbox oil temperature fault detection method and system based on neural network | |
CN112527547B (en) | Mechanical intelligent fault prediction method based on automatic convolution neural network | |
Li et al. | Neural architecture search for fault diagnosis | |
CN111914904B (en) | Image classification method fusing DarkNet and Capsule eNet models | |
CN113469252A (en) | Extra-high voltage converter valve operation state evaluation method considering unbalanced samples | |
CN116740562A (en) | Artificial forest pest and disease damage identification method based on snake group optimization algorithm and CNN algorithm | |
CN117093836A (en) | Last-stage reheater steam temperature early warning method and system based on VMD-EWOA-LSTM | |
CN112149804A (en) | Novel convolutional neural network model and application | |
CN116754213A (en) | Electric gate valve fault diagnosis method, device and equipment based on strong noise background | |
CN116316629A (en) | Data-driven optimal power flow calculation method considering topology feature learning | |
CN114881176A (en) | Non-invasive load identification method based on self-adaptive optimization random forest | |
CN114925620A (en) | Short-term wind power prediction method and system based on ensemble learning algorithm | |
CN114882270A (en) | Aortic dissection CT image classification method based on particle swarm optimization algorithm | |
CN115345192A (en) | Fault diagnosis method for rolling bearing under variable working conditions and complex noise environment |
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 |