CN111046945A - Fault type and damage degree diagnosis method based on combined convolutional neural network - Google Patents
Fault type and damage degree diagnosis method based on combined convolutional neural network Download PDFInfo
- Publication number
- CN111046945A CN111046945A CN201911258117.9A CN201911258117A CN111046945A CN 111046945 A CN111046945 A CN 111046945A CN 201911258117 A CN201911258117 A CN 201911258117A CN 111046945 A CN111046945 A CN 111046945A
- Authority
- CN
- China
- Prior art keywords
- training
- fault
- model
- layer
- data
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computational Linguistics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Evolutionary Biology (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Testing Of Devices, Machine Parts, Or Other Structures Thereof (AREA)
Abstract
The invention provides a fault type and damage degree diagnosis method based on a combined convolutional neural network, which comprises the following steps of: s1, data acquisition and pretreatment; s2, constructing a one-dimensional convolutional neural network; s3, training a model; s4, adjusting the hyper-parameters and a network framework; s5, preparing a data set for diagnosing fault types and damage degrees; s6, respectively training each model; s7, combining a plurality of convolution networks into a framework; and S8, completing fault type identification and damage degree diagnosis. The method selects a one-dimensional convolution neural network to carry out end-to-end extraction on the characteristics of an original vibration signal; meanwhile, the global maximum pooling layer is used for replacing the full connection layer, so that training parameters are reduced, the training speed is increased, and overfitting is prevented. The one-dimensional original data with different severity degrees are used for respectively training different models, so that the fault type can be identified, the fault damage degree can be classified, and the effect better than that of a single model is achieved.
Description
Technical Field
The invention belongs to the field of deep learning and rotary machine fault diagnosis, and relates to a method for diagnosing the type and the damage degree of a mechanical fault based on a convolutional neural network.
Background
Nowadays, modern industrial equipment is increasingly developed towards large-scale, high-speed, precise and automatic, and is widely applied to industries such as coal mines, petrochemicals, electric power and the like, and the monitoring of the health condition of the equipment becomes very complicated. Major safety accidents can be caused once a large mechanical system fails, and huge economic losses and even casualties are caused. As one of the important parts of a mechanical system, the bearing plays the role of a connecting rod or a gear shaft in the mechanical operation process, and the damaged bearing seriously affects a transmission shaft and a transmission gear, so that the performance, the stability and the service life of mechanical equipment are affected. The fault positions of the bearing generally comprise an inner ring, an outer ring and a rolling body, and the faults can be identified in time by using a fault diagnosis method, so that the safety performance of equipment is improved. However, early weak failure of equipment is often difficult to directly observe, and if the early weak failure is not prevented in time, accidents will be caused, so that failure diagnosis for large mechanical equipment is widely concerned in the current society.
The traditional fault diagnosis method can be divided into three categories, ① analytical model-based methods such as parameter estimation method, equivalent space method and state estimation method, ② signal-based processing methods such as correlation analysis, spectrum analysis and wavelet analysis, ③ knowledge-based methods such as intelligent diagnosis, fuzzy inference and neural network.
With the development of machine learning, researchers train various models of machine learning by using various indexes obtained through signal analysis as training samples (generally, the number of samples is small), and the fault mode identification precision is low. In recent years, with the advent of the big data era and the development of deep learning techniques, intelligent fault diagnosis methods have been widely used. In particular, since 2016, deep learning has revolutionized practice, providing a useful tool for processing and analyzing large-scale data, and data-driven mechanical failure diagnosis and health monitoring technologies have become increasingly popular.
Disclosure of Invention
The invention provides a mechanical fault type and damage degree diagnosis method based on deep learning, aiming at the problem of fault diagnosis. Because the vibration signal is a one-dimensional sequence, a one-dimensional convolution neural network is selected to carry out end-to-end extraction on the characteristics of the original vibration signal. The bearing fault is taken as an example for explanation, the requirement of real-time online monitoring on the working state of the industrial bearing can be met, and the requirement on professional knowledge of technicians and equipment maintenance personnel is low.
Aiming at the problems, the invention adopts a method for diagnosing the type and the damage degree of the mechanical fault based on the convolutional neural network, which comprises the following steps:
s1, data acquisition and preprocessing: and collecting one-dimensional time sequence vibration signals of the mechanical equipment under different operation states by using a sensor. Each collected state signal is divided into trainable samples, and when the samples are insufficient, overlapped sampling can be carried out so as to achieve the purpose of data enhancement; then constructing different data sets according to the requirements of fault diagnosis; the data set is divided into training, verifying and testing samples and input into a one-dimensional convolutional neural network for training.
S2, constructing a one-dimensional convolutional neural network: the first layer of the one-dimensional convolutional neural network uses a convolution kernel with the width of 8, the core of the convolutional neural network framework is a receptive field, and in order to enable the designed one-dimensional convolutional filter to learn characteristics irrelevant to displacement, the receptive field of the neurons of the last pooling layer in the network for an input signal should be larger than the number of sampling points of one rotation of a mechanical system. Furthermore, the present invention replaces the fully-connected layer with a global max-pooling layer after the convolutional layer.
S3, training the model: inputting a data set containing all fault types and damage degrees into a constructed one-dimensional convolutional neural network according to requirements, learning potential complex features in the original vibration data, and establishing a multi-layer mapping relation from the original one-dimensional vibration signal to the fault types or fault damage degrees of the bearing.
S4, adjusting the hyper-parameters and the network architecture: the invention particularly compares the test precision and the running time of the model according to the depth of the neural network, the width of a convolution kernel, a global maximum pooling layer, a Batch Normalization (BN) layer and Padding (Padding), and is beneficial to realizing real-time fault diagnosis of mechanical equipment and achieving higher identification precision.
S5, preparing a data set for diagnosing the fault type and the damage degree, firstly, forming the data containing all the fault types and the damage degree into an integral data set, and then forming the data with different damage degrees under each fault type into a plurality of independent small data sets.
S6, respectively training each model: training a network only identifying fault types by using a data set containing all fault types and serious damage; the data with different damage degrees under each fault type form a plurality of independent data sets to train a plurality of networks for identifying the fault damage degree. The network frameworks are as described above, different mapping relationships are obtained through learning, each framework is more pertinent, the precision of a single network framework is improved, and overall higher precision can be achieved.
S7, combining a plurality of convolution networks into a framework: and combining the trained convolutional neural networks, and aiming at newly collected vibration data, firstly identifying the fault type through a pre-trained combined model, and then judging the fault damage degree.
And S8, completing fault type identification and damage degree diagnosis: and the end-to-end feature automatic extraction, high-precision fault type identification and damage degree diagnosis are completed. Compared with the framework using a small convolution kernel in the first layer and the traditional neural network framework using full connection after the convolution layer, the method provided by the invention has higher accuracy and needs less training parameters. Meanwhile, compared with a single model, the combined model provided by the invention can be used for directly identifying the fault type and diagnosing the damage degree, and the model precision is greatly improved.
Preferably, the data acquisition and preprocessing in S1 includes the following steps:
s1.1, a large amount of one-dimensional time sequence vibration data under different operation conditions of mechanical equipment are collected through a sensor to form a large data set for neural network training.
S1.2, the sample input dimension is the premise of ensuring the diagnosis precision of the model, when the sample input dimension is increased, the diagnosis precision is improved, but the running speed of the model is reduced, so that the sample dimension suitable for mechanical fault diagnosis should be selected on the premise of ensuring the running speed of the model.
S1.3, the data enhancement method used in the invention is overlapped sampling, assuming that the length of a sample is L and the offset is S, if the data set has n data, then (n-L)/S +1 samples can be obtained. The invention uses overlapping sampling to divide the collected one-dimensional time sequence into needed samples, and divides the signals under different running states into single samples to form different data sets.
S1.4, combining the data sets into a data set containing various fault types and damage degrees.
Preferably, the step of constructing the one-dimensional convolutional neural network described in S2 is as follows:
s2.1, according to the mechanical vibration signal, the first layer uses a convolution kernel with the width of 8, and the subsequent convolution kernel uses a convolution kernel with the width of 3.
And S2.2, aiming at the neuron of the last pooling layer, the receptive field of the input signal is required to be larger than the number of sampling points of the mechanical system rotating for one circle. Let the receptive field of the neuron of the last pooling layer in the input signal be R(0)T is the number of points recorded by the accelerometer when the bearing rotates for one circle, L is the length of the input signal, and then the reception field R(0)Should satisfy T ≦ R(0)L is less than or equal to L, and the specific calculation process is as follows:
the receptor field R of the neuron in the last pooling layer in the kth pooling layer(k)And the receptor field R in the k-1 pooling layer(k-1)The relationship between them is:
R(k-1)＝S(k)(P(k)R(k)-1)+W(k)(1)
wherein S(k)Is the step size of the kth convolution kernel, W(k)Is the width of the kth convolution kernel, P(k)Is the number of k-th layer down-sampling points.
When the number of layers k is greater than 1, S(k)＝1，W(k)＝3，P(k)Thus, formula (1) can be arranged as:
R(k-1)＝2R(k)+2 (2)
when k is the last pooling layer n, R (n)1, the reception field of the last pooling layer in the first pooling layer is:
R(1)＝2n-1×3-2 (3)
bringing the above formula into R(k-1)＝S(k)(P(k)R(k)-1)+W(k)The receptive field of the input signal in the last pooling layer is calculated as:
R(0)＝S(k)(P(k)R(k)-1)+W(k)＝2S(1)(2n-1×3-2)+W(1)-S(1)≈S(1)(2n×3-4) (4)
because T is less than or equal to R(0)L is less than or equal to L, T is less than or equal to S(1)(2nX 3-4) is less than or equal to L, and the step length S is required(1)It should be possible to divide the signal length L equally.
And S2.3, Padding is carried out before each convolution, so that the sizes of the feature graphs before and after the convolution are the same, and the purpose is to fully extract the edge features.
And S2.4, adding a BN layer after each convolution layer, wherein the purpose is to enable the average value of data input into the network to be 0 and the variance to be 1, so that gradient propagation is facilitated, and a deeper network is constructed.
And S2.5, using global maximum pooling to realize the dimension reduction of the feature map, reducing the training parameters of the network, accelerating the training speed and preventing overfitting.
S2.6, the optimization method uses Root Mean Square Prop (RMSProp) to solve the problems of convergence speed and local minimum point of small-batch gradient descent.
S2.7, the invention combines model check point (ModelCheckpoint) and early termination (EarlyStopping) callback function, when the monitoring target index no longer changes in the set round, EarlyStopping training is adopted, and meanwhile, the model can be continuously saved in the training process of ModelCheckpoint, so as to obtain the optimal model.
The steps for adjusting the hyper-parameters and the network architecture as described in the preferred S4 are as follows:
s4.1, adjusting the number of filters to avoid the condition of under-fitting or over-fitting of the model; when the filter types are few, the signal features cannot be sufficiently extracted, thereby resulting in model under-fitting; too many filter types will result in an overfitting.
And S4.2, increasing the depth of the network, and measuring the change of the training precision and the running time until a proper network depth is found.
And S4.3, using a convolution kernel with the width of 3 at the first layer of the model, then using a convolution kernel with the width of 8, and simultaneously observing the training precision of the model.
And S4.4, after the last convolutional layer of the network, firstly using a full-link layer, then using a global maximum pooling layer to replace the full-link layer, and simultaneously inspecting the training precision of the model.
And S4.5, adding Padding before convolution, adding a BN layer after convolution, and inspecting whether the training precision of the model reaches the expectation.
Preferably, the combined convolutional neural network described in S7 includes the steps of:
s7.1, dividing a sample data set containing different fault states and different fault damage degrees into training, verifying and testing samples, training a one-dimensional convolution network for fault type identification in S2, and storing the trained model for fault type identification.
And S7.2, dividing the fault signals containing different damage degrees into a plurality of independent data sets, training a plurality of one-dimensional convolution networks in the S2, and storing a plurality of trained models for fault damage degree diagnosis.
And S7.3, combining the trained different models, identifying the fault type, and diagnosing the fault damage degree.
The invention has the advantages that: selecting a one-dimensional convolution neural network to carry out end-to-end extraction on the characteristics of the original vibration signal; meanwhile, the global maximum pooling layer is used for replacing the full connection layer, and the method has the advantages of reducing network training parameters, accelerating training speed and preventing overfitting. In addition, a neural network structure and a self-adaptive optimization algorithm which are advanced and suitable for mechanical vibration signals are used, and high accuracy is achieved. If the failure damage degrees are different, the one-dimensional original data with different severity degrees are used for respectively training different models, and the trained models are combined, so that the purpose of fault type identification and fault damage degree classification can be achieved, a better effect than a single model can be achieved, and a worker can take corresponding measures according to the fault types and the severity degrees, and major accidents are avoided.
Drawings
Fig. 1 is a schematic diagram of a training process of a one-dimensional convolutional neural network 1 for fault type identification according to an embodiment of the present invention.
Fig. 2 is a schematic diagram of a training process of a one-dimensional convolutional neural network 2-1 for identifying the damage degree of an inner ring fault according to an embodiment of the present invention.
Fig. 3 is a schematic diagram of a training process of the one-dimensional convolutional neural network 2-2 for outer ring fault damage degree identification according to an embodiment of the present invention.
Fig. 4 is a schematic diagram of a training process of a one-dimensional convolutional neural network 2-3 for identifying the damage degree of the rolling element fault according to an embodiment of the present invention.
Fig. 5 is a schematic diagram of a process of performing fault type identification and damage degree diagnosis on the one-dimensional convolutional neural network 1, the one-dimensional convolutional neural network 2-2, and the one-dimensional convolutional neural network 2-3 which are trained in a combined manner according to an embodiment of the present invention.
FIG. 6 is a graph of model accuracy versus 5 parallel trainings using a large convolution kernel and a small convolution kernel, respectively, in the first layer, according to an embodiment of the present invention.
FIG. 7 is a graph of model accuracy versus training 5 times using fully-connected and global max-pooling layers, respectively, after convolution according to an embodiment of the present invention.
FIG. 8 is a comparison graph of model accuracy obtained by determining whether Padding and convolution are performed before convolution and then parallel training of BN layer is performed 5 times according to an embodiment of the invention.
Detailed Description
The present invention will be described in detail below with reference to the accompanying drawings.
The effectiveness of the method provided by the invention is verified by taking a bearing data set of Case Western Reserve University (CWRU) as an example, and the specific steps are as follows:
the method comprises the following steps: data collected from 10 different operating conditions at 1 horsepower (hp) were selected as training, validation and test samples, with a sampling frequency of 12 KHz. In the invention, overlapping sampling is adopted, the overlapping rate is 0.8, the rotating speed of the motor is about 1772 Revolutions Per Minute (RPM), and the number of sampling points of one rotation of the bearing is about 400. The input dimension of the sample directly influences the diagnosis precision, specifically, the input dimension is increased, the diagnosis precision is improved, but the model training speed is reduced, wherein the length of the training sample is 1024, namely the number of sampling points is greater than one rotation of the bearing, and the purpose is to ensure the diagnosis precision which is high enough and the running speed which is high enough.
Data including health, inner ring failure, outer ring failure, and rolling element failure, and a failure state including 3 different damage degrees (failure sizes of 0.007, 0.014, and 0.021, respectively) were combined into data set 1. Table 1 lists the data used in this example, which was divided into training, validation and test samples in proportions of 70%, 15% and 15%, respectively, to train a one-dimensional convolutional neural network for fault type identification.
TABLE 1 training data set description
Step two: in this embodiment, the length of the input signal is 1024, the signal period T ≈ 400, and the number of convolution layers is 5. Calculated from S2.2, when the above-mentioned conditions are satisfiedWhen required S(1)It is only 8, the convolution width is not less than 3 steps, and the convolution kernel width is selected to be 24 in this embodiment.
Step three: the data set 1 is first trained by 10 classes, so that the model can directly learn 10 different complex mappings.
Step four: the width of the first convolutional layer convolution kernel in this embodiment is adjusted, and the convolution kernels with the width of 3, the step length of 1, the width of 24 and the step length of 8 are respectively used, the two models are trained in parallel for 5 times, and the diagnosis precision is shown in fig. 6.
After the last convolutional layer, the fully-connected layer or the global maximal pooling layer is respectively used to replace the fully-connected layer, and the two models are trained in parallel for 5 times, and the diagnosis precision of the two models is shown in fig. 7. Therefore, the two models have little influence on the diagnosis precision, but the model parameters using the global maximum pooling layer are less, and the running speed is greatly improved.
Adding padding before the convolutional layer of the convolutional neural network, adding a BN layer after the convolutional layer and comparing with a framework without any processing, and training the two models 5 times respectively, wherein the diagnosis precision is shown in figure 8. Therefore, the former can effectively improve the model diagnosis precision.
The one-dimensional convolutional network architecture constructed by the present invention is detailed in table 2.
TABLE 2 one-dimensional convolutional neural network structural parameters
Step five: the three fault types and the corresponding damage degree data are divided into different sample sets respectively: an inner ring fault data set 2, an outer ring fault data set 3 and a rolling bearing data set 4, which are detailed in a table 3.
TABLE 3 Fault Damage identification required training data set description
Step six: as shown in fig. 1, a data set 1 containing 10 operating states is initially classified by 4, i.e. the type of fault is first identified.
And then training is respectively carried out on the data set inner ring fault data set 2, the data set outer ring fault data set 3 and the data set rolling body fault data set 4, so as to identify the fault damage degree.
Step seven: and aiming at the newly input vibration signal, combining the four pre-trained models in the S6, and then carrying out fault type identification and damage degree diagnosis on the bearing, so that the diagnosis precision of the models can be greatly improved.
Step eight: the combined model proposed by the present invention was verified using the above data and compared with the one-dimensional convolutional neural network directly classified by 10 classes, and the model accuracy obtained by each parallel training 5 times is shown in table 4.
Table 4 model accuracy obtained by training the model 5 times in parallel.
Claims (5)
1. The fault type and damage degree diagnosis method based on the combined convolutional neural network is characterized by comprising the following steps: the method comprises the following steps:
s1, data acquisition and pretreatment: collecting one-dimensional time sequence vibration signals of mechanical equipment in different running states by using a sensor; each collected state signal is divided into trainable samples, and when the samples are insufficient, overlapped sampling can be carried out so as to achieve the purpose of data enhancement; then constructing different data sets according to the requirements of fault diagnosis; dividing a data set into training, verifying and testing samples and inputting the training, verifying and testing samples into a one-dimensional convolutional neural network for training;
s2, constructing a one-dimensional convolutional neural network, and replacing a full-connection layer with a global maximum pooling layer after a convolutional layer;
s3, training a model: inputting a data set containing all fault types and damage degrees into a constructed one-dimensional convolutional neural network according to requirements, wherein the data set is used for learning potential complex features in original vibration data and establishing a multilayer mapping relation from an original one-dimensional vibration signal to a bearing fault type or a fault damage degree;
s4, adjusting the hyper-parameters and a network framework: comparing the test accuracy and runtime of the model for the depth of the neural network, the width of the convolution kernel, the global maximum pooling layer, the Batch Normalization (BN) layer, and the Padding (Padding);
s5, preparing a data set for fault type and damage degree diagnosis, firstly forming data containing all fault types and damage degrees into an integral data set, and then forming data with different damage degrees under each fault type into a plurality of independent small data sets;
s6, respectively training each model: training a network only identifying fault types by using a data set containing all fault types and serious damage; the data with different damage degrees under each fault type form a plurality of independent data sets to train a plurality of networks for identifying the fault damage degrees, and different mapping relationships are obtained through learning;
s7, combining a plurality of convolution networks into a framework: combining the trained convolutional neural networks, and aiming at newly collected vibration data, firstly identifying the fault type through a pre-trained combined model, and then judging the fault damage degree;
s8, completing fault type identification and damage degree diagnosis: and the end-to-end feature automatic extraction, high-precision fault type identification and damage degree diagnosis are completed.
2. The method of claim 1, wherein the method comprises the steps of: the data acquisition and preprocessing in S1 includes the following steps:
s1.1, collecting a large amount of one-dimensional time sequence vibration data under different operation conditions of mechanical equipment through a sensor to form a large data set for neural network training;
s1.2, selecting a sample dimension suitable for mechanical fault diagnosis on the premise of ensuring the running speed of the model;
s1.3, adopting a data enhancement method of overlapped sampling, assuming that the length of a sample is L and the offset is S, and if the data set has n data, obtaining (n-L)/S +1 samples; dividing the acquired one-dimensional time sequence by using overlapped sampling to form a required sample, and dividing signals in different running states into single samples to form different data sets;
s1.4, combining the data sets into a data set containing various fault types and damage degrees.
3. The method of claim 1, wherein the method comprises the steps of: the step of constructing the one-dimensional convolutional neural network described in S2 is as follows:
s2.1, according to the mechanical vibration signal, using a convolution kernel with the width of 8 in the first layer, and using a convolution kernel with the width of 3 in the subsequent convolution kernel;
s2.2, aiming at the neuron of the last pooling layer, the receptive field of the input signal should be larger than the number of sampling points of the mechanical system rotating for one circle; let the receptive field of the neuron of the last pooling layer in the input signal be R(0)T is the number of points recorded by the accelerometer when the bearing rotates for one circle, L is the length of the input signal, and then the reception field R(0)Should satisfy T ≦ R(0)L is less than or equal to L, and the specific calculation process is as follows:
the receptor field R of the neuron in the last pooling layer in the kth pooling layer(k)And the receptor field R in the k-1 pooling layer(k-1)The relationship between them is:
R(k-1)＝S(k)(P(k)R(k)-1)+W(k)(1)
wherein S(k)Is the step size of the kth convolution kernel, W(k)Is the width of the kth convolution kernel, P(k)Is the number of the kth layer down-sampling points;
when the number of layers k is greater than 1, S(k)＝1，W(k)＝3，P(k)Thus, formula (1) can be arranged as:
R(k-1)＝2R(k)+2(2)
when k is the last pooling layer n, R(n)1, the reception field of the last pooling layer in the first pooling layer is:
R(1)＝2n-1×3-2 (3)
bringing the above formula into R(k-1)＝S(k)(P(k)R(k)-1)+W(k)The receptive field of the input signal in the last pooling layer is calculated as:
R(0)＝S(k)(P(k)R(k)-1)+W(k)＝2S(1)(2n-1×3-2)+W(1)-S(1)≈S(1)(2n×3-4) (4)
because T is less than or equal to R(0)L is less than or equal to L, T is less than or equal to S(1)(2nX 3-4) is less than or equal to L, and the step length S is required(1)The signal length L should be divisible;
s2.3, Padding is carried out before each convolution, so that the sizes of the feature graphs before and after the convolution are the same, and the edge features are fully extracted;
s2.4, adding a BN layer after each convolution layer, enabling the average value of data input into the network to be 0 and the variance to be 1, and constructing a deeper network;
s2.5, using global maximum pooling to realize the dimension reduction of the feature map, reducing the training parameters of the network, accelerating the training speed and preventing overfitting;
s2.6, optimizing (RMSProp) by using a Root Mean Square transfer method, and solving the problems of convergence rate and local minimum point of small-batch gradient descent.
And S2.7, combining a model check point (ModelCheckpoint) and an early stopping (Earlystopping) callback function, when the monitored target index does not change any more in the set round, training by adopting the Earlystopping ending model, and simultaneously, the model can be continuously saved in the training process of the ModelCheckpoint so as to obtain the optimal model.
4. The method of claim 1, wherein the method comprises the steps of: the steps of adjusting the hyper-parameters and the network architecture described in S4 are as follows:
s4.1, adjusting the number of filters to avoid the condition of under-fitting or over-fitting of the model;
s4.2, increasing the depth of the network, and measuring the change of training precision and running time until a proper network depth is found;
s4.3, using a convolution kernel with the width of 3 at the first layer of the model, then using a convolution kernel with the width of 8, and simultaneously inspecting the training precision of the model;
s4.4, after the last convolutional layer of the network, firstly using a full-link layer, then using a global maximum pooling layer to replace the full-link layer, and simultaneously inspecting the training precision of the model;
and S4.5, adding Padding before convolution, adding a BN layer after convolution, and inspecting whether the training precision of the model reaches the expectation.
5. The method of claim 1, wherein the method comprises the steps of: the combined convolutional neural network described in S7 includes the steps of:
s7.1, dividing a sample data set containing different fault states and different fault damage degrees into training, verifying and testing samples, and storing the one-dimensional convolution network model for fault type identification trained in S2 for fault type identification;
s7.2, dividing fault signals containing different damage degrees into a plurality of independent data sets, training a plurality of one-dimensional convolution networks in S2, and storing a plurality of trained models for diagnosing the damage degrees of the faults;
and S7.3, combining the trained different models, identifying the fault type, and diagnosing the fault damage degree.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911258117.9A CN111046945B (en) | 2019-12-10 | 2019-12-10 | Fault type and damage degree diagnosis method based on combined convolutional neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911258117.9A CN111046945B (en) | 2019-12-10 | 2019-12-10 | Fault type and damage degree diagnosis method based on combined convolutional neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN111046945A true CN111046945A (en) | 2020-04-21 |
| CN111046945B CN111046945B (en) | 2023-10-24 |
Family
ID=70235379
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201911258117.9A Active CN111046945B (en) | 2019-12-10 | 2019-12-10 | Fault type and damage degree diagnosis method based on combined convolutional neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111046945B (en) |
Cited By (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111582396A (en) * | 2020-05-13 | 2020-08-25 | 江南大学 | Fault diagnosis method based on improved convolutional neural network |
| CN111597182A (en) * | 2020-05-20 | 2020-08-28 | 中国石油化工股份有限公司 | Convolutional neural network-based fault anomaly identification method for oil pumping unit driving motor |
| CN111811819A (en) * | 2020-06-30 | 2020-10-23 | 佛山科学技术学院 | Bearing fault diagnosis method and device based on machine learning |
| CN111897310A (en) * | 2020-07-24 | 2020-11-06 | 华中科技大学 | Industrial process fault classification method and system based on one-dimensional multi-head convolutional network |
| CN111950526A (en) * | 2020-09-01 | 2020-11-17 | 国网河北省电力有限公司检修分公司 | Fault diagnosis method for energy storage mechanism of circuit breaker based on deep learning |
| CN112001417A (en) * | 2020-07-17 | 2020-11-27 | 国网宁夏电力有限公司检修公司 | Monitoring method, medium and system for transformer oil conservator |
| CN112329886A (en) * | 2020-11-26 | 2021-02-05 | 珠海大横琴科技发展有限公司 | Double-license plate recognition method, model training method, device, equipment and storage medium |
| CN112446326A (en) * | 2020-11-26 | 2021-03-05 | 中国核动力研究设计院 | Canned motor pump fault mode identification method and system based on deep rewinding and accumulating network |
| CN112528548A (en) * | 2020-11-27 | 2021-03-19 | 东莞市汇林包装有限公司 | Self-adaptive depth coupling convolution self-coding multi-mode data fusion method |
| CN113033309A (en) * | 2021-02-25 | 2021-06-25 | 北京化工大学 | Fault diagnosis method based on signal downsampling and one-dimensional convolution neural network |
| CN113538385A (en) * | 2021-07-21 | 2021-10-22 | 上海勘察设计研究院(集团)有限公司 | Tunnel apparent disease type and grade discrimination method based on deep learning |
| CN114593919A (en) * | 2022-03-28 | 2022-06-07 | 青岛理工大学 | Novel rolling bearing fault diagnosis method and system |
| CN114638558A (en) * | 2022-05-19 | 2022-06-17 | 天津市普迅电力信息技术有限公司 | Data set classification method for operation accident analysis of comprehensive energy system |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107727395A (en) * | 2017-07-21 | 2018-02-23 | 中国矿业大学 | A kind of Method for Bearing Fault Diagnosis assessed based on full variation and uncompensation distance |
| US20180357542A1 (en) * | 2018-06-08 | 2018-12-13 | University Of Electronic Science And Technology Of China | 1D-CNN-Based Distributed Optical Fiber Sensing Signal Feature Learning and Classification Method |
| CN110033021A (en) * | 2019-03-07 | 2019-07-19 | 华中科技大学 | A kind of Fault Classification based on one-dimensional multichannel convolutional neural networks |
| CN110073301A (en) * | 2017-08-02 | 2019-07-30 | 强力物联网投资组合2016有限公司 | The detection method and system under data collection environment in industrial Internet of Things with large data sets |
-
2019
- 2019-12-10 CN CN201911258117.9A patent/CN111046945B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107727395A (en) * | 2017-07-21 | 2018-02-23 | 中国矿业大学 | A kind of Method for Bearing Fault Diagnosis assessed based on full variation and uncompensation distance |
| CN110073301A (en) * | 2017-08-02 | 2019-07-30 | 强力物联网投资组合2016有限公司 | The detection method and system under data collection environment in industrial Internet of Things with large data sets |
| US20180357542A1 (en) * | 2018-06-08 | 2018-12-13 | University Of Electronic Science And Technology Of China | 1D-CNN-Based Distributed Optical Fiber Sensing Signal Feature Learning and Classification Method |
| CN110033021A (en) * | 2019-03-07 | 2019-07-19 | 华中科技大学 | A kind of Fault Classification based on one-dimensional multichannel convolutional neural networks |
Non-Patent Citations (2)
| Title |
|---|
| LEVENT EREN等: "A Generic Intelligent Bearing Fault Diagnosis System Using Compact Adaptive 1D CNN Classifier", 《JOURNAL OF SIGNAL PROCESSING SYSTEMS》 * |
| 曲建岭;余路;袁涛;田沿平;高峰;: "基于一维卷积神经网络的滚动轴承自适应故障诊断算法", 仪器仪表学报 * |
Cited By (18)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111582396A (en) * | 2020-05-13 | 2020-08-25 | 江南大学 | Fault diagnosis method based on improved convolutional neural network |
| CN111582396B (en) * | 2020-05-13 | 2023-05-02 | 江南大学 | Fault diagnosis method based on improved convolutional neural network |
| CN111597182A (en) * | 2020-05-20 | 2020-08-28 | 中国石油化工股份有限公司 | Convolutional neural network-based fault anomaly identification method for oil pumping unit driving motor |
| CN111811819A (en) * | 2020-06-30 | 2020-10-23 | 佛山科学技术学院 | Bearing fault diagnosis method and device based on machine learning |
| CN112001417B (en) * | 2020-07-17 | 2022-07-05 | 国网宁夏电力有限公司检修公司 | Monitoring method, medium and system for transformer oil conservator |
| CN112001417A (en) * | 2020-07-17 | 2020-11-27 | 国网宁夏电力有限公司检修公司 | Monitoring method, medium and system for transformer oil conservator |
| CN111897310A (en) * | 2020-07-24 | 2020-11-06 | 华中科技大学 | Industrial process fault classification method and system based on one-dimensional multi-head convolutional network |
| CN111950526A (en) * | 2020-09-01 | 2020-11-17 | 国网河北省电力有限公司检修分公司 | Fault diagnosis method for energy storage mechanism of circuit breaker based on deep learning |
| CN112329886A (en) * | 2020-11-26 | 2021-02-05 | 珠海大横琴科技发展有限公司 | Double-license plate recognition method, model training method, device, equipment and storage medium |
| CN112446326A (en) * | 2020-11-26 | 2021-03-05 | 中国核动力研究设计院 | Canned motor pump fault mode identification method and system based on deep rewinding and accumulating network |
| CN112528548A (en) * | 2020-11-27 | 2021-03-19 | 东莞市汇林包装有限公司 | Self-adaptive depth coupling convolution self-coding multi-mode data fusion method |
| CN113033309A (en) * | 2021-02-25 | 2021-06-25 | 北京化工大学 | Fault diagnosis method based on signal downsampling and one-dimensional convolution neural network |
| CN113033309B (en) * | 2021-02-25 | 2023-12-19 | 北京化工大学 | Fault diagnosis method based on signal downsampling and one-dimensional convolutional neural network |
| CN113538385A (en) * | 2021-07-21 | 2021-10-22 | 上海勘察设计研究院(集团)有限公司 | Tunnel apparent disease type and grade discrimination method based on deep learning |
| CN113538385B (en) * | 2021-07-21 | 2022-10-25 | 上海勘察设计研究院(集团)有限公司 | Tunnel apparent disease type and grade discrimination method based on deep learning |
| CN114593919A (en) * | 2022-03-28 | 2022-06-07 | 青岛理工大学 | Novel rolling bearing fault diagnosis method and system |
| CN114593919B (en) * | 2022-03-28 | 2023-11-17 | 青岛理工大学 | Novel rolling bearing fault diagnosis method and system |
| CN114638558A (en) * | 2022-05-19 | 2022-06-17 | 天津市普迅电力信息技术有限公司 | Data set classification method for operation accident analysis of comprehensive energy system |
Also Published As
| Publication number | Publication date |
|---|---|
| CN111046945B (en) | 2023-10-24 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111046945B (en) | Fault type and damage degree diagnosis method based on combined convolutional neural network | |
| CN111709448B (en) | Mechanical fault diagnosis method based on migration relation network | |
| CN109460618B (en) | Rolling bearing residual life online prediction method and system | |
| CN109186973B (en) | Mechanical fault diagnosis method of unsupervised deep learning network | |
| CN111914883B (en) | Spindle bearing state evaluation method and device based on deep fusion network | |
| CN110110768B (en) | Rolling bearing fault diagnosis method based on parallel feature learning and multiple classifiers | |
| CN112254964A (en) | Rolling bearing fault diagnosis method based on rapid multi-scale convolution neural network | |
| CN109655259A (en) | Combined failure diagnostic method and device based on depth decoupling convolutional neural networks | |
| CN109782603A (en) | The detection method and monitoring system of rotating machinery coupling fault | |
| CN112257530B (en) | Rolling bearing fault diagnosis method based on blind signal separation and support vector machine | |
| CN110132554B (en) | Rotary machine fault diagnosis method based on deep Laplace self-coding | |
| CN109186964B (en) | rotary machine fault diagnosis method based on angle resampling and ROC-SVM | |
| CN109946080B (en) | Mechanical equipment health state identification method based on embedded circulation network | |
| CN112257767B (en) | Product key part state classification method aiming at class imbalance data | |
| CN111680788A (en) | Equipment fault diagnosis method based on deep learning | |
| CN113188794B (en) | Gearbox fault diagnosis method and device based on improved PSO-BP neural network | |
| CN112729834B (en) | Bearing fault diagnosis method, device and system | |
| CN112305388B (en) | On-line monitoring and diagnosing method for insulation partial discharge faults of generator stator winding | |
| CN113095179A (en) | Metametric learning driven helicopter planetary gearbox fault diagnosis method | |
| CN114358124A (en) | Rotary machine new fault diagnosis method based on deep-antithetical-convolution neural network | |
| CN114091504A (en) | Rotary machine small sample fault diagnosis method based on generation countermeasure network | |
| Wang et al. | Fault diagnosis of industrial robots based on multi-sensor information fusion and 1D convolutional neural network | |
| CN114429152A (en) | Rolling bearing fault diagnosis method based on dynamic index antagonism self-adaption | |
| CN115587290A (en) | Aero-engine fault diagnosis method based on variational self-coding generation countermeasure network | |
| CN116956215A (en) | Fault diagnosis method and system for transmission system |
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 |