CN111830408B - Motor fault diagnosis system and method based on edge calculation and deep learning - Google Patents
Motor fault diagnosis system and method based on edge calculation and deep learning Download PDFInfo
- Publication number
- CN111830408B CN111830408B CN202010581758.4A CN202010581758A CN111830408B CN 111830408 B CN111830408 B CN 111830408B CN 202010581758 A CN202010581758 A CN 202010581758A CN 111830408 B CN111830408 B CN 111830408B
- Authority
- CN
- China
- Prior art keywords
- output
- residual error
- module
- layer
- diagnosis
- 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
- G01—MEASURING; TESTING
- G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
- G01R31/00—Arrangements for testing electric properties; Arrangements for locating electric faults; Arrangements for electrical testing characterised by what is being tested not provided for elsewhere
- G01R31/34—Testing dynamo-electric machines
- G01R31/343—Testing dynamo-electric machines in operation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F9/00—Arrangements for program control, e.g. control units
- G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
- G06F9/46—Multiprogramming arrangements
- G06F9/50—Allocation of resources, e.g. of the central processing unit [CPU]
- G06F9/5061—Partitioning or combining of resources
- G06F9/5072—Grid computing
-
- 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
- G06N3/084—Backpropagation, e.g. using gradient descent
Abstract
The invention relates to a motor fault diagnosis system and method based on edge calculation and deep learning, wherein edge side equipment can configure an acquisition device, perform Empirical Mode Decomposition (EMD) after preprocessing vibration or current signals, input n IMF components by a deep residual error network diagnosis model built in a diagnosis module to perform fault diagnosis, and upload a diagnosis result and fault data to a cloud service center; the cloud service center utilizes the historical data and the fault data to regularly train the deep residual error network diagnosis model, sends the deep residual error network diagnosis model to each edge side device after training, and updates the deep residual error network diagnosis model in the edge side device. The motor equipment can be diagnosed in real time by computing the edge side equipment, the motor state can be diagnosed quickly in time, and processed data are transmitted to the cloud center, so that the time delay caused by the limitation of data transmission speed and bandwidth is greatly reduced, and the data processing pressure of the cloud center is remarkably relieved.
Description
Technical Field
The invention relates to the technical field of motor fault diagnosis, in particular to a motor fault diagnosis system and method based on edge calculation and deep learning.
Background
The motor is very common in the modern industry, whether large factory enterprises or small factory enterprises can see the shadow of the motor everywhere, but large motors on some important production lines are more important to the enterprises, once a fault occurs, the whole production line can be stopped, and more serious economic loss and casualties can be brought to the enterprises, so that the normal and stable operation of the motor is very important to the enterprises.
At present, most of the maintenance of the motor is to carry out vibration noise, shell temperature inspection and the like on the motor every day by field maintenance personnel, the state of the motor is evaluated according to field experience for many years, and the other main measure is to carry out regular factory return maintenance on the motor. The experience of field maintenance personnel is likely to fail to find the fault of the motor in time, so that major accidents are caused, in addition, the motor is disassembled and maintained regularly, the work of the whole production line needs to be suspended, and the economic loss caused by excessive maintenance is also caused.
At present, most of the existing technical schemes are to arrange a data acquisition device on site, transmit data acquired in real time on site back to a cloud server, and apply a corresponding algorithm to detect the state of a motor and diagnose a fault. At present, common algorithms also stay in traditional neural network algorithms more frequently, such as BP neural network and probabilistic neural network, but in the face of huge data volume of a motor group, the shallow neural network algorithms have no advantages, so that the training speed is slow, and the precision of a diagnosis model is low;
on the other hand, the cloud processing method cannot meet the real-time requirement, and the detection is performed after a fault occurs, and in addition, the method has high requirements on the storage and calculation capacity of the server, and the safety and stability of transmitting a large amount of original data cannot be guaranteed.
With the development of network technology, the quantity of devices connected to a network is larger and larger, massive data is generated, if the data are directly transmitted back to a cloud server, large bandwidth and delay are generated, and meanwhile, the safety of processing massive data is difficult to guarantee; in addition, the shallow neural network algorithms are difficult to process huge data volume of a motor group, so that the problems of low network training speed, overfitting and the like can be caused, the diagnosis capability of the network is finally reduced, and the actual state of the motor cannot be well reflected.
Disclosure of Invention
The invention provides a motor fault diagnosis system and method based on edge calculation and deep learning, aiming at the problems that diagnosis calculation can be carried out only by transmitting all diagnosis data back to a cloud server and the accuracy rate of shallow neural network diagnosis is low.
In order to achieve the purpose, the invention provides a motor fault diagnosis system based on edge calculation and deep learning, which comprises a plurality of acquisition devices, a plurality of edge side devices and a cloud service center, wherein the acquisition devices are connected with the edge side devices through a network;
the acquisition device is used for acquiring vibration or current signals of the motor and sending the vibration or current signals to the corresponding edge side equipment;
the edge side equipment comprises an acquisition configuration module, a preprocessing module, an empirical mode decomposition module and a diagnosis module; the acquisition configuration module can configure the acquisition device; the preprocessing module is used for preprocessing the vibration or current signal; the empirical mode decomposition module is used for performing Empirical Mode Decomposition (EMD) on the preprocessed vibration or current signals to decompose the signals into n IMF components; the diagnostic module is internally provided with a depth residual error network diagnostic model, n IMF components are input for fault diagnosis, and a diagnostic result and fault data are uploaded to the cloud service center;
the cloud service center utilizes historical data and fault data to perform EMD, trains the depth residual error network diagnosis model periodically, sends the trained depth residual error network diagnosis model to each edge side device, and updates the depth residual error network diagnosis model in the edge side device.
Further, the empirical mode decomposition module processing flow includes:
s1 obtaining a pre-processed vibration or current signal x (t), further i =0;
s2, obtaining an envelope f on x (t) max (t) and lower envelope f min (t), calculating the mean value of the upper envelope line and the lower envelope line as m (t);
s2, making a difference between x (t) and the mean value m (t), and recording the difference as h 1 (t)；
S3, judging whether i is smaller than n, if so, adding 1 to i to obtain the ith IMF component c i (t)＝h 1 (t) and using x (t) -h 1 (t) replacing x (t), and returning to S2; if i is not less than n, n IMF components are output.
Further, the diagnostic module comprises an input convolution unit, 6 residual error units and an output unit; after the n IMF components are subjected to feature extraction through an input convolution unit, sequentially performing treatment through 6 residual error units connected in series; and finally, obtaining a diagnosis result by an output unit.
Further, the residual error unit includes a first batch normalization layer, a first ReLu activation function, a first convolution layer, a second batch normalization layer, a second ReLu activation function, a second convolution layer, a third batch normalization layer, a third ReLu activation function, a third convolution layer, and an identity mapping unit, which are connected in series in sequence, and the identity mapping unit superimposes and outputs an output of the third convolution layer and an input of the first batch normalization layer.
Further, the output unit comprises an output batch standardization layer, an output ReLu activation function, a global mean pooling layer, a Dropout layer and a full-connection output layer which are sequentially connected in series; the method comprises the steps that an output batch standardization layer carries out standardization processing on input features, an output ReLu activation function carries out nonlinear processing, a global mean value pooling layer averages feature maps of all input channels and outputs the averaged feature maps, a Dropout layer discards partial neurons, a full-connection output layer adopts a softmax function to calculate classification results, and the probabilities of various types of faults and no faults are output.
Further, the operation process of the first, second and third batch normalization layers and the output batch normalization layer includes processing n one-dimensional features characterized by n IMF components of each sample, respectively obtaining output features, and calculating a certain one-dimensional feature x of the s-th input sample s Calculating output characteristic y s Comprises the following steps:
wherein: n is a radical of batch Refers to the number of samples contained in each batch, μ being N batch Mean value of samples in a dimension, σ being N batch The standard deviation of each sample in a dimension, γ and β are the scale and shift learnable parameters, respectively, and ε is a constant that is close to zero.
Further, the cloud service center includes: the system comprises a database, a selection module, an IMF component extraction module, a model training module and a packaging output module; the database stores historical vibration or current signals of the motor, vibration or current signals of the motor during the fault uploaded by each edge side device and a manually labeled diagnosis result; the selection module selects a sample from a database and inputs the sample to the IMF component extraction module, and the IMF component extraction module performs Empirical Mode Decomposition (EMD) on the input to be decomposed into n IMF components; the model training module inputs n IMF components into an internal deep residual error network diagnosis model for training, and when the requirements of training times or training precision are met, the deep residual error network diagnosis model is packaged and sent to each edge side device through the packaging output module.
And further, the system also comprises an application layer, and the application layer is used for reading and displaying the data of the cloud service center.
The invention also provides a motor fault diagnosis method based on edge calculation and deep learning, which comprises the following steps:
training a deep residual error network diagnosis model and loading the model to edge side equipment;
each edge side device receives vibration or current signals of the motor collected by the corresponding collecting device and carries out pretreatment; performing Empirical Mode Decomposition (EMD) on the preprocessed vibration or current signals to decompose the signals into n IMF components; inputting n IMF components by the deep residual error network diagnosis model to perform fault diagnosis, and uploading a diagnosis result and fault data to the cloud service center;
the cloud service center conducts EMD by using the diagnosis result and the fault data, trains the depth residual error network diagnosis model regularly, sends the trained depth residual error network diagnosis model to each edge side device, and updates the depth residual error network diagnosis model in the edge side device.
Further, the EMD module performs EMD on the preprocessed vibration or current signals and decomposes the signals into n IMF components;
the processing flow of EMD by the EMD module comprises the following steps:
s1 obtaining a pre-processed vibration or current signal x (t), further i =0;
s2, obtaining an envelope f on x (t) max (t) and lower envelope f min (t), calculating the mean value of the upper envelope line and the lower envelope line as m (t);
s2, making a difference between x (t) and the mean value m (t), and recording the difference as h 1 (t)；
S3, judging whether i is smaller than n, if so, adding 1 to i to obtain the ith IMF component c i (t)＝h 1 (t) and using x (t) -h 1 (t) replacing x (t), and returning to S1; if i is not less than n, n IMF components are output.
Further, the depth residual error network diagnosis model comprises an input convolution unit, 6 residual error units and an output unit; after the n IMF components are subjected to feature extraction through an input convolution unit, the n IMF components are sequentially subjected to treatment through 6 residual error units which are connected in series; and finally, obtaining a diagnosis result by an output unit.
Further, the residual error unit includes a first batch normalization layer, a first ReLu activation function, a first convolution layer, a second batch normalization layer, a second ReLu activation function, a second convolution layer, a third batch normalization layer, a third ReLu activation function, a third convolution layer, and an identity mapping unit, which are connected in series in sequence, and the identity mapping unit superimposes and outputs an output of the third convolution layer and an input of the first batch normalization layer.
Further, the output module comprises an output batch standardization layer, an output ReLu activation function, a global mean pooling layer, a Dropout layer and a full-connection output layer which are sequentially connected in series; the method comprises the steps that an output batch standardization layer carries out standardization processing on input features, an output ReLu activation function carries out nonlinear processing, a global mean value pooling layer averages feature maps of all input channels and outputs the averaged feature maps, a Dropout layer discards partial neurons, a full-connection output layer adopts a softmax function to calculate classification results, and the probabilities of various types of faults and no faults are output.
Further, the cloud service center includes: the system comprises a database, a selection module, an IMF component extraction module, a model training module and a packaging output module; the database stores historical vibration or current signals of the motor, vibration or current signals of the motor during fault uploading by each edge side device and diagnosis results manually marked; the selection module selects a sample from a database and inputs the sample to the IMF component extraction module, and the IMF component extraction module performs Empirical Mode Decomposition (EMD) on the input to be decomposed into n IMF components; the model training module inputs n IMF components into an internal deep residual error network diagnosis model for training, and when the requirements of training times or training precision are met, the packaging output module packages the deep residual error network diagnosis model and sends the deep residual error network diagnosis model to each edge side device.
The technical scheme of the invention has the following beneficial technical effects:
(1) The existing common method adopts a data acquisition device to send locally acquired data to a cloud center for calculation and analysis to obtain a diagnosis result, but due to the huge amount of original data and the numerous field acquisition devices, real-time data transmission and storage are very great challenges for the cloud center, on-site equipment is easy to break down, but fault data are not timely uploaded to the cloud center due to various reasons, so that fault finding is delayed. The method and the device have the advantages that the deep residual error network model is trained in the cloud center, the trained deep residual error network model is issued to the edge side equipment, the motor equipment can be diagnosed in real time through calculation of the edge side equipment, the motor state can be diagnosed rapidly in time, and after the edge side equipment diagnoses a fault, the diagnosis result is transmitted to the cloud server, and fault data are transmitted to the cloud center, so that a fault database can be enriched continuously, and the trained diagnosis model can better accord with the actual condition of the motor.
(2) The existing method directly uses the collected original data as the input of the deep learning network, which also introduces much noise and causes inaccurate diagnosis result; according to the method, the current or vibration signals are collected, the original data are processed through a preprocessing method, and then the intrinsic mode function of the fault signal is extracted through an empirical mode decomposition algorithm, so that useful information of the original signal can be fully obtained, and the accuracy of a depth model is increased;
(3) The invention relates to a fault diagnosis model, in particular to an improved version of a basic convolutional neural network and a deep residual error network model, wherein the network can greatly reduce the training time of the network, improve the accuracy of the model and provide great support for quick update of a diagnosis model by introducing a residual error module.
(4) In addition, the processing of the edge computing is closer to a data source, only the processed data is transmitted to the cloud center, so that the time delay caused by the limitation of data transmission speed and bandwidth is greatly reduced, and the data processing pressure of the cloud center is remarkably relieved.
Drawings
FIG. 1 is a flow chart of a motor fault diagnosis system;
FIG. 2 is a diagram of a ResNet residual unit;
fig. 3 is a schematic diagram of a ResNet network structure.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention will be described in further detail with reference to the accompanying drawings in conjunction with the following detailed description. It should be understood that the description is intended to be exemplary only, and is not intended to limit the scope of the present invention. Moreover, in the following description, descriptions of well-known structures and techniques are omitted so as to not unnecessarily obscure the concepts of the present invention.
The invention aims to provide a motor fault diagnosis system and a motor fault diagnosis method, wherein the diagnosis system comprises a plurality of acquisition devices, a plurality of edge side devices, a cloud service center and an application layer, and is combined with a figure 1.
The acquisition device is used for acquiring vibration or current signals of the motor and sending the vibration or current signals to the corresponding edge side equipment;
the edge side equipment has a data acquisition function, a data preprocessing function, an application deep learning model diagnosis function and functions of data uploading, local data storage and display; the edge side equipment comprises an acquisition configuration module, a preprocessing module, an empirical mode decomposition module and a diagnosis module; the acquisition configuration module can initialize the acquisition device and configure acquisition strategy parameters; the preprocessing module is used for preprocessing the vibration or current signals and adopts preprocessing methods such as filtering processing and the like; the EMD module is used for carrying out EMD on the preprocessed vibration or current signals and decomposing the preprocessed vibration or current signals into n IMF components; the diagnostic module is internally provided with a depth residual error network diagnostic model, n IMF components are input for fault diagnosis, and a diagnostic result and fault data are uploaded to the cloud service center; and applying the trained deep residual error network diagnosis model on the edge side equipment to diagnose the motor in real time, storing and displaying the diagnosis result only without storing normal real-time data when the diagnosis result is normal, and storing and displaying the diagnosis result and fault data when the diagnosis result is faulty. And uploading the diagnosis result and the fault data to a cloud center for displaying and storing, uploading the diagnosis result and the fault data of the corresponding fault to the cloud center when the diagnosis result indicates that a certain part has a fault, and manually confirming and calibrating the fault result. When all the parts of the diagnosis result are normal, the diagnosis result is only sent to the cloud service center, and the real-time data under normal working conditions is not sent to the cloud service center, so that the cost of data transmission can be greatly reduced.
The cloud service center has a function of carrying out initialization configuration on edge side equipment, has functions of receiving data, storing data and displaying, and comprises a fault database module function; the cloud service center utilizes the historical data and the fault data to regularly train the deep residual error network diagnosis model, sends the trained deep residual error network diagnosis model to each edge side device, and updates the deep residual error network diagnosis model in the edge side device.
The application layer comprises portable equipment such as office computers and notebooks, mobile equipment such as mobile phones and the like and functions of various applications for controlling and displaying real-time running states of the motors.
Furthermore, the empirical mode decomposition is a novel adaptive signal time-frequency processing method, and is particularly suitable for analysis processing of nonlinear non-stationary signals, and the method has the advantages that the method processes the signals according to the time scale characteristics of the method, and no basis function needs to be set, which is essentially different from the Fourier decomposition and the wavelet decomposition which are established on the prior harmonic basis function and the wavelet basis function.
The core of the method is empirical mode decomposition, which can decompose a complex signal into a finite number of eigenmode functions (IMFs), each decomposed IMF component comprises local characteristic signals of different time scales of an original signal, and the decomposition is based on the local characteristics of the time scales of a signal sequence, so that the method has self-adaptability. The eigenmode function must have 2 conditions met:
condition 1: in the whole time range of the function, the number of local extreme points and zero-crossing points must be equal, or the difference is at most 1;
condition 2: at any point in time, the local maximum envelope and the local minimum envelope must be 0 on average.
For a given signal, an EMD method is adopted to solve an intrinsic mode function, and the screening process of EMD analysis is as follows: the empirical mode decomposition module processing flow comprises the following steps:
s1 obtaining a pre-processed vibration or current signal x (t), further i =0;
s2 findingAll extreme points on x (t) are obtained, and all extreme points and minimum points are respectively connected by a curve to form an upper envelope f max (t) and lower envelope f min (t), taking the mean value of the upper envelope line and the lower envelope line as m (t);
s3, making a difference between x (t) and the mean value m (t), and recording the difference as h 1 (t); namely, the following equation:
h 1 (t)＝x(t)-m(t)
ideally, h 1 (t) is a fundamental mode component, whereas the actual signal is generally more complex, h 1 (t) repeating the above process as new x (t) until a condition is met, then h 1 (t) is the first fundamental mode component, noted:
c 1 (t)＝h 1 (t)
s4, judging whether i is smaller than n, if so, adding 1 to i to obtain the ith IMF component c i (t)＝h 1 (t) and using x (t) -h 1 (t) replace x (t), i.e., x (t) = x (t) -c 1 (t), return to S2; if i is not less than n, n IMF components are output. Sequentially obtaining a second IMF component c 2 (t), third IMF component c 3 (t) up to the Nth IMF component c n And (t), finally decomposing the original signal into n IMF components after EMD decomposition. The n IMF components characterize n one-dimensional features of the sample.
Further, with reference to fig. 3, the diagnostic module includes an input convolution unit, 6 residual error units, and an output unit; after the n IMF components are subjected to feature extraction through an input convolution unit, the n IMF components are sequentially subjected to input convolution unit and 6 residual error units connected in series; and finally, obtaining a diagnosis result by an output unit. With reference to fig. 2, the residual error unit includes a first batch normalization layer, a first ReLu activation function, a first convolution layer, a second batch normalization layer, a second ReLu activation function, a second convolution layer, a third batch normalization layer, a third ReLu activation function, a third convolution layer, and an identity mapping unit, which are sequentially connected in series, and the identity mapping unit superimposes and outputs an output of the third convolution layer and an input of the first batch normalization layer.
The output unit comprises an output batch standardization layer, an output ReLu activation function, a global mean pooling layer, a Dropout layer and a full-connection output layer which are sequentially connected in series; the method comprises the steps that an output batch standardization layer carries out standardization processing on input features, an output ReLu activation function carries out nonlinear processing, a global mean value pooling layer averages feature maps of all input channels and outputs the averaged feature maps, a Dropout layer discards partial neurons, a full-connection output layer adopts a softmax function to calculate classification results, and the probabilities of various types of faults and no faults are output.
The method comprises the steps of extracting Intrinsic Mode Functions (IMF) on the edge equipment by applying an empirical mode decomposition algorithm to the acquired signals, using a matrix formed by extracted IMF components of various orders as input of a deep learning diagnosis model, diagnosing a field motor in real time through the diagnosis model, uploading a diagnosis result and fault data to a cloud center, storing the fault data, training a deep neural network model regularly by applying historical data and the uploaded fault data to the cloud center, and issuing the trained deep learning model to each edge equipment, so that the deep learning diagnosis model can be updated continuously, the diagnosis response to the motor is faster, the efficiency is higher, and the accuracy of the fault diagnosis result is higher. The system can display the states of the motors through a plurality of platforms, so that a client can clearly master the real-time conditions of the motors.
The residual error network model is an upgraded convolutional neural network (ConvNet) model in nature, and can solve the problem that model training is gradually difficult as the number of layers of the convolutional network model is continuously deepened. The residual network model is typically mapped by input layer, convolutional layer, batch normalization, identity. The global mean pooling and the full-link output layer are combined by important basic modules. For the deep residual network (ResNet) diagnostic model, different basic units have different functions. The ResNet model usually contains a certain number of basic residual units, which are the core components of the ResNet model. The main part of the ResNet model is formed by stacking a plurality of basic residual modules. Common basic residual units are composed of two convolutional layers, two ReLu activation functions, two batch normalizations, and one identity map. The invention applies an improved version of a residual error module, the path of which comprises 'batch standardization-ReLu activation function-convolution layer-batch standardization-ReLu activation function-convolution layer', the residual error unit is more beneficial to the back propagation of errors in the neural network, and the parameters are easier to train. The residual error module used in the present invention is shown in fig. 2.
(1) Convolutional layer
The convolutional layer is one of the core parts of the ConvNet model and the ResNet model. One of its main advantages over the conventional fully-connected layer is that the training weights required for training are significantly reduced. In practice, the convolution kernel may act as a trainable local feature extraction operator. During the convolution operation, the convolution kernel is slid over the input data to extract the basic local features at different locations. The same convolution kernel can be used for extracting the same type local features at different positions, so that the utilization efficiency of the weight in the neural network is improved. As shown in fig. 3, the present invention first uses 3 × 3 input convolution units for feature extraction. And in each residual error unit, three convolution units of 3 x 3 are respectively arranged for feature extraction.
(2) Batch standardization layer (BN)
When the minimum batch stochastic gradient descent method is adopted, a certain number of randomly selected training samples are input into the deep neural network in each iteration of the deep neural network training process, and forward propagation and backward propagation are calculated. However, the feature distribution between these small batches of training samples tends to change continuously during the iterative process, in which case the weights and biases need to be updated continuously to adapt to the distribution after the change, slowing down the speed of deep neural network training. Batch normalization is a normalization method proposed for this internal covariance drift problem. Unlike the general standardization method, the batch standardization is an operation embedded between layers inside the deep neural network, and can be expressed as:
wherein: n is a radical of hydrogen batch Refers to the number of samples contained in each batch, and u is N batch The mean value of the samples in a certain dimension, sigma being N batch The standard deviation of each sample in one dimension, γ and β, are learnable parameters used to scale and translate the normalized value, respectively, and ε is a constant that is close to zero. Y of s Is x s The corresponding output characteristic. Three batch normalization layers (BN) are arranged in each residual unit, and the output unit is also provided with the batch normalization layer (BN) for normalization processing.
During deep neural network training, with continuous updating of parameters, the data distribution input by each middle layer is often greatly different from that before parameter updating, so that the network is required to adapt to new number distribution continuously, and the training of the whole deep network is extremely difficult. By introducing batch standardization, the input layers are not only subjected to standardization treatment, but also the input of each middle layer is subjected to standardization treatment, so that the problem of drift of internal covariates can be avoided, and the problems of difficulty in training and low training speed are solved.
(3) Activating a function
The activation function is usually an indispensable part of a neural network algorithm, and the core role of the activation function is to realize the nonlinear mapping of input data in the neural network, so that the neural network can fit the nonlinear relation between the input data and the label. Because most of the current neural network algorithms need to adopt an error back propagation algorithm to train internal parameters, the output of an activation function needs to be differentiable for input. Currently common activation functions include Sigmoid, tanh, and Relu activation functions, among others. Because the partial derivative number of the Relu activation function is only possibly 0 or 1, compared with the traditional Sigmoid and Tanh activation functions, the Relu activation function can effectively reduce the risks of gradient disappearance and gradient explosion of the deep neural network in the training process and reduce the training difficulty of internal parameters of the deep neural network, so the Relu activation function is adopted by default, and the formula is expressed as follows:
wherein x is ReLu And y ReLu The input characteristic and the output characteristic of the ReLu activation function are respectively.
As shown in fig. 3, each batch normalization layer (BN) is followed by a ReLu activation function, which performs non-linear processing on the output characteristics of the batch normalization layer (BN). The activation function is used for introducing nonlinear factors into the neural network, and the neural network can fit various curves through the activation function, so that the expression capacity of the neural network is stronger.
(4) Global mean pooling GAP
The ResNet model adopted by the invention adopts the operation of global mean pooling at the position close to the output layer. Global mean pooling refers to calculating a corresponding mean value from the feature map of each channel as an output feature, and can be expressed as:
Y GAP (1,1,i ch )＝average(XGAP(i ro ,i co ,i ch ))
wherein: x GAP And Y GAP An input feature map and an output feature map representing global mean pooling, respectively.
The global mean pooling not only can significantly reduce the number of weights to be trained, but also can effectively solve the problem of translational changes. That is, for the vibration signals of the same fault type, the positions of the fault impacts in the sample are different, and the learned writing characteristics in the conventional neural network are also different accordingly. The global mean pooling can realize invariance of the deep learning model to fault impact positions by calculating the global feature of mean, and solves the problem to a certain extent, thereby improving the generalization effect of the deep learning model on test data.
(5) Dropout layer
Dropout is a more common method of mitigating multiple fits. Dropout randomly discards some neurons during each training process, allowing the neural network to forward propagate and reverse update only the remaining neuron parameters. In this way, drop can weaken the "cooperative relationship" in front of the neurons, making the effect of each neuron more independent, thereby achieving the effect of model regularization. During the test, if there is no Dropout operation, i.e. all neurons participate in the calculation, the invention performs Dropout operation on the global mean pooled output, setting the discard ratio to 50%.
(6) Output layer
The cross entropy loss function is a classical loss function and is often used for multi-classification problems of neural networks and deep neural networks. In the ResNet model, the feature dimensions of the output layer are the same as the number of classes participating in training. Before calculating the cross entropy loss function, firstly, the characteristic x of the output layer needs to be processed by adopting a softmax function j ，j＝1,2,…,N class Forced conversion to [0,1 ]]Fraction of interval, wherein N class Indicating the number of categories to which the classification task relates. The Softmax function can be expressed in the form of a formula:
in general, in the multi-classification problem of the shallow and deep neural networks, the labels of each class are in the form of vectors, only one value is 1, and the other values are 0, and in this label form, the cross-entropy loss function can be defined as:
wherein, t j A true label, y, representing the specimen j Representing the output of the softmax function. In general, cross-entropy loss functions can lead to higher training speeds in neural network-based multi-classification problems than more traditional squared error functions.
Step3: first fundamental mode component c 1 After (t) is decomposed, c is subtracted from the original signal x (t) 1 (t) obtaining a residual signal x 1 (t)
x 1 (t)＝x(t)-c 1 (t)
Step4: x is to be 1 (t) as an original signal x (t), repeating the steps 1 to 3 to obtain a second IMF component c 2 (t), third IMF component c 3 (t) up to the Nth IMF component c n And (t), finally decomposing the original signal into n IMF components after EMD decomposition.
The cloud computing transmits data to the cloud end, the application deployed by the cloud end is used for computing, and the computing result is transmitted back to each user after computing is completed, so that the user does not need to pay attention to the problems of environment, hardware conditions and the like. In addition, the processing of the edge computing is closer to a data source, only the processed data is transmitted to the cloud center, so that the time delay caused by the limitation of data transmission speed and bandwidth is greatly reduced, and the data processing pressure of the cloud center is remarkably relieved.
In summary, the present invention relates to a motor fault diagnosis system and method based on edge calculation and deep learning, where the edge device may configure an acquisition device, perform Empirical Mode Decomposition (EMD) after preprocessing a vibration or current signal, input n IMF components to a diagnosis module built-in deep residual error network diagnosis model for fault diagnosis, and upload a diagnosis result and fault data to a cloud service center; the cloud service center utilizes the historical data and the fault data to regularly train the deep residual error network diagnosis model, sends the deep residual error network diagnosis model to each edge side device after training, and updates the deep residual error network diagnosis model in the edge side device. The motor equipment can be diagnosed in real time through calculation on the edge side equipment, and the motor state can be diagnosed in time and rapidly. By sinking the detection calculation to the edge equipment, the collected field data are directly processed and calculated on the edge equipment combined with the deep learning algorithm, and the detection result and the fault data are sent back to the cloud center server, so that the detection efficiency and the diagnosis accuracy are greatly improved, and the data transmission cost can be reduced.
It is to be understood that the above-described embodiments of the present invention are merely illustrative of or explaining the principles of the invention and are not to be construed as limiting the invention. Therefore, any modification, equivalent replacement, improvement and the like made without departing from the spirit and scope of the present invention should be included in the protection scope of the present invention. Further, it is intended that the appended claims cover all such variations and modifications as fall within the scope and boundaries of the appended claims or the equivalents of such scope and boundaries.
Claims (10)
1. A motor fault diagnosis system based on edge calculation and deep learning is characterized by comprising a plurality of acquisition devices, a plurality of edge side devices and a cloud service center;
the acquisition device is used for acquiring vibration or current signals of the motor and sending the vibration or current signals to the corresponding edge side equipment;
the edge side equipment comprises an acquisition configuration module, a preprocessing module, an empirical mode decomposition module and a diagnosis module; the acquisition configuration module can configure the acquisition device; the preprocessing module is used for preprocessing the vibration or current signal; the empirical mode decomposition module is used for performing Empirical Mode Decomposition (EMD) on the preprocessed vibration or current signals to decompose the signals into n IMF components; the diagnostic module is internally provided with a depth residual error network diagnostic model, n IMF components are input for fault diagnosis, and a diagnostic result and fault data are uploaded to the cloud service center;
the cloud service center performs EMD (empirical mode decomposition) by using historical data and fault data, trains the depth residual error network diagnosis model periodically, sends the trained depth residual error network diagnosis model to each edge side device, and updates the depth residual error network diagnosis model in the edge side device;
the diagnosis module comprises an input convolution unit, 6 residual error units and an output unit; after the n IMF components are subjected to feature extraction through an input convolution unit, the n IMF components are sequentially subjected to treatment through 6 residual error units which are connected in series; finally, obtaining a diagnosis result by an output unit;
the residual error unit comprises a first batch of standardized layers, a first ReLu activation function, a first convolution layer, a second batch of standardized layers, a second ReLu activation function, a second convolution layer, a third batch of standardized layers, a third ReLu activation function, a third convolution layer and an identity mapping unit which are sequentially connected in series, and the identity mapping unit outputs the output of the third convolution layer after being overlapped with the input of the first batch of standardized layers.
2. The system according to claim 1, wherein the EMD module process flow comprises:
s1 obtaining a pre-processed vibration or current signal x (t), further i =0;
s2, obtaining an envelope f on x (t) max (t) and lower envelope f min (t), calculating the mean value of the upper envelope line and the lower envelope line as m (t);
s2, making a difference between x (t) and the mean value m (t), and recording the difference as h 1 (t)；
S3, judging whether i is smaller than n, if so, adding 1 to i to obtain the ith IMF component c i (t)＝h 1 (t) and using x (t) -h 1 (t) replacing x (t), and returning to S2; such asIf i is not less than n, n IMF components are output.
3. The system for motor fault diagnosis based on edge calculation and deep learning of claim 2, wherein the output unit comprises an output batch normalization layer, an output ReLu activation function, a global mean pooling layer, a Dropout layer and a fully-connected output layer which are connected in series in sequence; the method comprises the steps that an output batch standardization layer carries out standardization processing on input features, an output ReLu activation function carries out nonlinear processing, a global mean value pooling layer averages feature maps of all input channels and outputs the averaged feature maps, a Dropout layer discards partial neurons, a full-connection output layer adopts a softmax function to calculate classification results, and the probabilities of various types of faults and no faults are output.
4. The system of claim 3, wherein the first, second, third and output batch normalization layers are operated by processing n one-dimensional features of the samples characterized by the n IMF components of each sample to obtain an output feature, and calculating a certain one-dimensional feature x of the s-th input sample s Computing output feature y s Comprises the following steps:
wherein: n is a radical of batch Refers to the number of samples contained in each batch, and μ is N batch Mean value of samples in a dimension, σ being N batch The standard deviation of each sample in one dimension, gamma and beta are the scale and translation, respectivelyThe learnable parameter, ε, is a constant that is close to zero.
5. The edge-computing and deep-learning based motor fault diagnosis system according to claim 1 or 2, wherein the cloud service center comprises: the system comprises a database, a selection module, an IMF component extraction module, a model training module and a packaging output module; the database stores historical vibration or current signals of the motor, vibration or current signals of the motor during fault uploading by each edge side device and diagnosis results manually marked; the selection module selects a sample from a database and inputs the sample to the IMF component extraction module, and the IMF component extraction module performs Empirical Mode Decomposition (EMD) on the input to be decomposed into n IMF components; the model training module inputs n IMF components into an internal deep residual error network diagnosis model for training, and when the requirements of training times or training precision are met, the packaging output module packages the deep residual error network diagnosis model and sends the deep residual error network diagnosis model to each edge side device.
6. The motor fault diagnosis system based on edge computing and deep learning as claimed in claim 1 or 2, further comprising an application layer for reading and displaying data of the cloud service center.
7. A motor fault diagnosis method based on edge calculation and deep learning is characterized by comprising the following steps:
training a deep residual error network diagnosis model and loading the model to edge side equipment;
each edge side device receives vibration or current signals of the motor collected by the corresponding collecting device and carries out pretreatment; performing Empirical Mode Decomposition (EMD) on the preprocessed vibration or current signals to decompose the signals into n IMF components; inputting n IMF components by the deep residual error network diagnosis model to perform fault diagnosis, and uploading a diagnosis result and fault data to a cloud service center;
the cloud service center performs Empirical Mode Decomposition (EMD) by using the diagnosis result and the fault data, trains the deep residual error network diagnosis model at regular intervals, sends the trained deep residual error network diagnosis model to each edge side device, and updates the deep residual error network diagnosis model in the edge side device;
the depth residual error network diagnosis model comprises an input convolution unit, 6 residual error units and an output unit; after the n IMF components are subjected to feature extraction through an input convolution unit, sequentially performing treatment through 6 residual error units connected in series; finally, obtaining a diagnosis result by an output unit;
the residual error unit comprises a first batch of standardized layers, a first ReLu activation function, a first convolution layer, a second batch of standardized layers, a second ReLu activation function, a second convolution layer, a third batch of standardized layers, a third ReLu activation function, a third convolution layer and an identity mapping unit which are sequentially connected in series, and the identity mapping unit outputs the output of the third convolution layer after being overlapped with the input of the first batch of standardized layers.
8. The edge calculation and deep learning based motor fault diagnosis method according to claim 7, wherein the EMD module performs EMD on the preprocessed vibration or current signals and decomposes the EMD into n IMF components;
the processing flow of EMD by the EMD module comprises the following steps:
s1 obtaining a preprocessed vibration or current signal x (t), where i =0;
s2, obtaining an envelope f on x (t) max (t) and the lower envelope f min (t), calculating the mean value of the upper envelope line and the lower envelope line as m (t);
s2, making a difference between x (t) and the mean value m (t), and recording the difference as h 1 (t)；
S3, judging whether i is smaller than n, if so, adding 1 to i to obtain the ith IMF component c i (t)＝h 1 (t) and using x (t) -h 1 (t) replacing x (t), and returning to S1; if i is not less than n, n IMF components are output.
9. The motor fault diagnosis method based on the edge calculation and the deep learning as claimed in claim 8, wherein the output module comprises an output batch standardization layer, an output ReLu activation function, a global mean pooling layer, a Dropout layer and a fully-connected output layer which are connected in series in sequence; the method comprises the steps that an output batch standardization layer carries out standardization processing on input features, an output ReLu activation function carries out nonlinear processing, a global mean value pooling layer averages feature maps of all input channels and outputs the averaged feature maps, a Dropout layer discards partial neurons, a full-connection output layer adopts a softmax function to calculate classification results, and the probabilities of various types of faults and no faults are output.
10. The motor fault diagnosis method based on edge computing and deep learning of claim 8, wherein the cloud service center comprises: the system comprises a database, a selection module, an IMF component extraction module, a model training module and a packaging output module; the database stores historical vibration or current signals of the motor, vibration or current signals of the motor during the fault uploaded by each edge side device and a manually labeled diagnosis result; the selection module selects a sample from a database and inputs the sample to the IMF component extraction module, and the IMF component extraction module performs Empirical Mode Decomposition (EMD) on the input to be decomposed into n IMF components; the model training module inputs n IMF components into an internal deep residual error network diagnosis model for training, and when the requirements of training times or training precision are met, the packaging output module packages the deep residual error network diagnosis model and sends the deep residual error network diagnosis model to each edge side device.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010581758.4A CN111830408B (en) | 2020-06-23 | 2020-06-23 | Motor fault diagnosis system and method based on edge calculation and deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010581758.4A CN111830408B (en) | 2020-06-23 | 2020-06-23 | Motor fault diagnosis system and method based on edge calculation and deep learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN111830408A CN111830408A (en) | 2020-10-27 |
| CN111830408B true CN111830408B (en) | 2023-04-18 |
Family
ID=72898227
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010581758.4A Active CN111830408B (en) | 2020-06-23 | 2020-06-23 | Motor fault diagnosis system and method based on edge calculation and deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111830408B (en) |
Families Citing this family (23)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112307287B (en) * | 2020-11-11 | 2022-08-02 | 国网山东省电力公司威海供电公司 | Cloud edge cooperative architecture based power internet of things data classification processing method and device |
| CN112565409B (en) * | 2020-12-02 | 2022-08-16 | 杭州谐云科技有限公司 | Industrial equipment monitoring method and system based on cloud edge cooperation |
| CN112417772B (en) * | 2020-12-15 | 2022-12-23 | 重庆忽米网络科技有限公司 | Rotary equipment fault diagnosis method based on edge calculation model |
| CN112881518B (en) * | 2021-01-08 | 2022-07-12 | 东冶及策河北能源技术有限公司 | Method for predicting residual life of dynamic filter compensator |
| CN112819646B (en) * | 2021-01-18 | 2022-09-06 | 南京理工大学 | Fault diagnosis system for flexible production line of customized wood furniture |
| CN112947368B (en) * | 2021-02-02 | 2022-06-07 | 安徽理工大学 | Three-phase asynchronous motor fault diagnosis device based on digital twinning and transfer learning |
| CN112908353A (en) * | 2021-02-03 | 2021-06-04 | 天津大学 | Voice enhancement method for hearing aid by combining edge computing and cloud computing |
| CN112862211A (en) * | 2021-03-09 | 2021-05-28 | 国网冀北电力有限公司信息通信分公司 | Method and device for assigning orders of dynamic ring defects of communication management system |
| CN113030717A (en) * | 2021-03-12 | 2021-06-25 | 广东电网有限责任公司 | Circuit breaker fault diagnosis method and circuit based on electric signal deep learning |
| CN112989712B (en) * | 2021-04-27 | 2021-08-03 | 浙大城市学院 | Aeroengine fault diagnosis method based on 5G edge calculation and deep learning |
| CN113473515A (en) * | 2021-05-31 | 2021-10-01 | 杭州电子科技大学 | Industrial wireless network reliability assessment method |
| CN113255584B (en) * | 2021-06-22 | 2021-10-19 | 德明通讯（上海）股份有限公司 | Fault diagnosis and monitoring system based on edge calculation |
| CN113742855B (en) * | 2021-07-27 | 2022-03-18 | 清华大学 | Fault prediction method, system, electronic equipment and readable storage medium |
| CN113610837A (en) * | 2021-08-24 | 2021-11-05 | 华北电力大学（保定） | External insulation equipment discharge fault diagnosis method based on deep learning and edge calculation |
| CN113933658B (en) * | 2021-08-27 | 2023-08-29 | 国网湖南省电力有限公司 | Dry-type transformer discharge detection method and system based on audible sound analysis |
| CN113758709A (en) * | 2021-09-30 | 2021-12-07 | 河南科技大学 | Rolling bearing fault diagnosis method and system combining edge calculation and deep learning |
| CN113934175A (en) * | 2021-10-26 | 2022-01-14 | 华电莱州发电有限公司 | Intelligent wireless steam turbine data acquisition system |
| CN114252725A (en) * | 2021-11-05 | 2022-03-29 | 国网河南省电力公司电力科学研究院 | HHT and ResNet 18-based single-phase earth fault type comprehensive identification method |
| CN114282579A (en) * | 2021-12-30 | 2022-04-05 | 浙大城市学院 | Aviation bearing fault diagnosis method based on variational modal decomposition and residual error network |
| CN114997062A (en) * | 2022-06-15 | 2022-09-02 | 同济大学 | Industrial water pump fault diagnosis light-weight method and system based on edge calculation |
| CN114943259B (en) * | 2022-07-22 | 2022-10-21 | 深圳市世坤科技实业有限公司 | Equipment fault diagnosis system and method for processing electric element by laser |
| CN115267409A (en) * | 2022-09-26 | 2022-11-01 | 北京中电普华信息技术有限公司 | Power equipment fault detection method, device and equipment |
| CN115877198A (en) * | 2022-12-23 | 2023-03-31 | 国网山东省电力公司蒙阴县供电公司 | Primary and secondary fusion switch fault diagnosis early warning system based on edge calculation |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108805195A (en) * | 2018-04-03 | 2018-11-13 | 北京理工大学 | A kind of motor group method for diagnosing faults based on two-value deep-neural-network |
| CN110147714A (en) * | 2019-03-28 | 2019-08-20 | 中国矿业大学 | Coal mine gob crack identification method and detection system based on unmanned plane |
| WO2020081122A1 (en) * | 2018-10-15 | 2020-04-23 | Illumina, Inc. | Deep learning-based techniques for pre-training deep convolutional neural networks |
-
2020
- 2020-06-23 CN CN202010581758.4A patent/CN111830408B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108805195A (en) * | 2018-04-03 | 2018-11-13 | 北京理工大学 | A kind of motor group method for diagnosing faults based on two-value deep-neural-network |
| WO2020081122A1 (en) * | 2018-10-15 | 2020-04-23 | Illumina, Inc. | Deep learning-based techniques for pre-training deep convolutional neural networks |
| CN110147714A (en) * | 2019-03-28 | 2019-08-20 | 中国矿业大学 | Coal mine gob crack identification method and detection system based on unmanned plane |
Non-Patent Citations (2)
| Title |
|---|
| 基于EMD-神经网络的电动机转子故障诊断方法研究;杨存祥等;《微电机》;20110428(第04期);第88-91页 * |
| 深度卷积神经网络LeNet-5和ResNet的对比以及应用分析;许可等;《电子设计工程》;20200120(第02期);第88-91、96页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN111830408A (en) | 2020-10-27 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111830408B (en) | Motor fault diagnosis system and method based on edge calculation and deep learning | |
| CN108197648A (en) | A kind of Fault Diagnosis Method of Hydro-generating Unit and system based on LSTM deep learning models | |
| CN110044554A (en) | A kind of online test method of the metal pressure container leakage based on acoustic emission signal | |
| CN112113755B (en) | Mechanical fault intelligent diagnosis method based on deep convolution-kurtosis neural network | |
| CN116380445B (en) | Equipment state diagnosis method and related device based on vibration waveform | |
| CN112507479B (en) | Oil drilling machine health state assessment method based on manifold learning and softmax | |
| CN112784920B (en) | Yun Bianduan coordinated rotating component reactance domain self-adaptive fault diagnosis method | |
| CN115438726A (en) | Device life and fault type prediction method and system based on digital twin technology | |
| CN115795343A (en) | Method for analyzing state of unbalanced sample ammeter production equipment based on digital twin | |
| CN116451848A (en) | Satellite telemetry data prediction method and device based on space-time attention mechanism | |
| CN113919220A (en) | Intelligent fault diagnosis method for rolling bearing driven by vibration image | |
| CN114492642A (en) | Mechanical fault online diagnosis method for multi-scale element depth residual shrinkage network | |
| CN114897138A (en) | System fault diagnosis method based on attention mechanism and depth residual error network | |
| CN113627091A (en) | Device and method for predicting energy load | |
| CN116714437B (en) | Hydrogen fuel cell automobile safety monitoring system and monitoring method based on big data | |
| CN116910573A (en) | Training method and device for abnormality diagnosis model, electronic equipment and storage medium | |
| CN116910574A (en) | Mechanical equipment diagnosis method and device based on meta-learning and time convolution network | |
| CN114298413A (en) | Hydroelectric generating set runout trend prediction method | |
| CN113935413A (en) | Distribution network wave recording file waveform identification method based on convolutional neural network | |
| Zhi-hong et al. | Sensor Fault Diagnosis Based on Wavelet Analysis and LSTM Neural Network | |
| CN113051809A (en) | Virtual health factor construction method based on improved restricted Boltzmann machine | |
| Zheng et al. | An unsupervised transfer learning method based on SOCNN and FBNN and its application on bearing fault diagnosis | |
| CN112560252A (en) | Prediction method for residual life of aircraft engine | |
| CN115831339B (en) | Medical system risk management and control pre-prediction method and system based on deep learning | |
| CN113537637B (en) | Earthquake destructive power prediction method and device based on gated cycle unit |
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 |