S1: Get the original operation of the rolling bearing through field experiments and its working labels, and divided into training and test data sets 3: 1.
S1.1, the field experiments include:
Bearing Type: The support shaft to be detected is a rotating shaft of the nuclear power activation pump motor, the sampling frequency is 12kHz and 48kHz.。
Fault setting: The damage of the bearing is a single point of injury of electric spark processing. The fault diameter is 0.007, 0.014, 0.021, 0.028 and 0.040 mm, respectively. The damage point is provided on the bearing inner ring, the rolling bodies, and the outer ring.
Signal Sampling: The motor wind sector and the drive end are placed over the bearing seat of the drive to collect the vibration acceleration signal and power of the faulty bearing, using a 16-channel data recorder collecting the vibration acceleration signal data of the test stage, data acquisition frequency It is 12kHz, the shaft speed is 1700R / min.
Data acquisition: 8 normal samples, 53 outer ring injury, 23 inner ring injury samples and 11 rolling bodies injury samples.
In S1.2, the data working condition tag includes: fault, no fault, fault diameter, fault location (inner ring, outer ring, rolling body)
In S1.3, the original operational data is cut into training sets and test sets, ratio 3: 1.
The original operation data should be converted to a 24 * 24 two-dimensional feature. Specific methods include:
As the input length of 24 * 24 = 576, the sampling mode of data overlap is 0 is taken, and the original vibration data is divided by 576 points; TF.reshape is applied in Tensorflow (XS, [- 1, 24, 24, 1]) Several one-dimensional data is sequentially sorted by a list of 24 * 24 * 1, as the input of training.
Add a working condition label corresponding to the feature, in this embodiment, in this embodiment, it is: A-inner ring fault diameter is 0.007, B-inner ring fault diameter is 0.014 , The C-inner calendar fault diameter is 0.021, the D-inner ring fault diameter is 0.028, E-normal.
S2: Establish a convolutional neural network bearing fault diagnosis model, use the simulation experiment based on the control variable method to determine the model parameters, use training data sets to train convolutional neural network models, resulting in diagnostic results confidence at 0.95 nuclear power start water pump rolling bearing fault Diagnostic model;
S2.1: Establish a convolutional neural network model, model architecturefigure 1 As shown, the convolutional neural network includes a convolution layer, a cellular layer, a full attachment layer, and a Softmax output layer.
Convolutional layer: The volume of the plurality of convolutionary cores and the input image is used, and after the bias item is added, the activation function can be obtained, and a series of characteristics can be obtained, the mathematical expression of the convolution is:
among them, Section 1 of the first element; mjThe JU Jun area of the L-1 layer characteristics; For the element; For the corresponding weight matrix; For the bias item. f (•) is an activation function.
Convolutional neural network model passed training Weight matrix value and The bias item value implements a classified task.
Pihua layer: The pool layer is to reduce the feature, and ensure the translation of the features; the maximum value pool method is selected in this embodiment. Specifically, the pool layer is taken to take the maximum value by taking the maximum value by the characteristic of the convolutional layer output in each of the non-overlapping magnitudes, and the feature map is reduced in two dimensions N times.
Full connection layer: Expand the characterization into one-dimensional vectors, which is equal to and by activating the function:
yk= F (wkxk-1+ Bk)
Where k is the serial number of the network layer; YkOutput of full connectivity; Xk-1Is a typical vector; WkAs the weight coefficient; BkFor the bias item, the SoftMax activation function is used in this embodiment.
Convolutional neural network fault diagnosis model training method is the reverse propagation algorithm, such asfigure 2 As shown, the algorithm uses a chain guide to calculate the gradient of the loss function to each weight, and the weight update is updated according to the gradient drop algorithm. The cost function used by the convolutional neural network is the cross entropy function, the formula is:
Where the C represents the price, X represents the sample, n represents the total number of samples, and a represents the model output value, y indicates the actual value of the sample.
S2.2: Simulation Experiment: Based on the Tensorflow platform, training set data is used as an input of a general convolutional neural network model. Based on a single control variable method, a comprehensive analysis activation function type, a volume subsidy, and a network layer number of model final accuracy. The influence provides theoretical foundation for subsequent model training.
The simulation is specifically includes:
Evaluation Index: The evaluation index of convolivation neural network model is accurate, as follows:
Among them, TP is the number of sample numbers that the predicted results is true and the TN is actually false and predicted as the false (ie, the correct) sample number, the number of samples of the prediction is true and predicted, and the fn is actual. It is true and predicts the number of samples.
General Convolution Neural Network Model Parameter Settings shown in Table 2:
Table 2 parameter settings for CNN models
The effect of activation function type on the accuracy of the fault diagnosis model of the nuclear power start water pump rolling bearing:
Common activation functions include: Sigmoid function, Tanh function, RELU function, Lrelu function, ELU function; specifically as follows:
R (x) = max (0, x)
LR (X) = max (αx, x)
The training set data is used as an input to a general convolutional neural network model, and the corresponding model classification accuracy is calculated by changing the activation function type.
Such asimage 3 As shown, the RELU function is the highest, the function convergence speed is faster, and the prediction performance is better, and it is suitable for processing the nuclear power start water pump rolling bearing fault data.
Influence of Cemented Nuclear Size on the Fault Diagnosis Model of Nuclear Power Start Water Pump Rolling Bearing Fault Diagnosis Model:
Setting the volume nuclear size is: [1, 1], [2, 2], [3, 3] ... [15, 15], the activation function is RELU, and the other parameters are the same as Table 2, by changing the convolutionary core. Size, test the accuracy of its model.
byFigure 4 The accuracy rate of the model is positively correlated with the volume nuclear growth. After the volume core is increased to [5, 5], the accuracy remains at about 0.97. Considering that a larger volume nucleus causes a long computation time, the memory occupation is large, and the optimal volume nuclear size is [5, 5].
Influence of the number of network layers on the fault diagnosis model of nuclear power start water pump rolling bearing:
The setting layers are 1, 2, 3, 4, 5, and the activation function is RELU, and the other parameters are the same as Table 2, and the accuracy of the model is tested by changing the number of layers.
byFigure 5 The accuracy rate of the model is positively correlated with the number of network layers. However, the number of layers can result in double the amount of data, the amount of calculation, the increase in memory, the reduction of training speed, and determines the number of network layers 2.
S2.3: The final structure and parameters of the final structure and parameters of the fault diagnosis model of the nuclear power starting water pump rolling bearing:
Table 3 Parameters of the CNN model
S3: Real-time monitoring of the nuclear power starting water pump scroll bearing operation, enters the fault diagnosis model in real time to monitor the acquisition of the monitoring data, and troubleshoot the nuclear power start water pump rolling bearing and assess its health status, recording abnormal data and alarm;
S3.1: The rolling bearing vibration acceleration signal, power, and rotational speed is acquired when the acceleration sensor and torque sensor mounted on the start of the pump motor fan end and the drive end bearing housing.
Field monitoring equipment includes acceleration sensors and torque sensors;
The acceleration sensor is located on the bearing housing that activates the pump motor fan end and the drive end for collecting the vibration acceleration signal of the rolling bearing; the torque sensor is connected to the motor load by the elastic column pin, for the measurement of power and speed; use 16 channel data The recorder collects the corresponding data.
S3.2: The acquired data is transmitted to the server through the data line, and the server includes a central processor, a data storage center, and can store and process the data.
S3.3: Bearing real-time operation data Entering the convolutional neural network fault diagnosis model, realizing the diagnosis of the nuclear power start water pump rolling bearing fault monitoring, and the conclusion of bearing fault diagnosis.
S3.4: Using a bearing health assessment method based on fault diameter,Figure 6 ,Figure 7As shown, specifically, as shown in Table 4:
Table 4 Nuclear power start water pump rolling bearing fault assessment standard
S4, comprehensively considering the health status, maintenance costs, maintenance time, etc., formulate the corresponding operation and maintenance scheme;
S4.1: Based on the health, maintenance costs, maintenance time, etc., based on nuclear power expert knowledge, according to different faults, establishing a fortune, establishing the operation and maintenance information library, as shown in Table 5:
Table 5 Nuclear power start water pump rolling bearing transport dimension
S4.2: Find the corresponding operation and maintenance suggested information in the operation and maintenance recommendations, and provide nuclear power maintenance personnel, the test verification results are as follows:
Table 6 Nuclear Power Start Water Pump Rollow Bearing Maintenance Plan Intelligent Decision Test Verification Result
The present invention also provides a nuclear power activation water pump rolling bearing intelligent operation and maintenance system, which includes a bearing online monitoring subsystem and a bearing intelligent transporter subsystem.
The bearing is online monitoring subsystem, including online monitoring modules, fault diagnostic modules, health assessment modules, alarm modules.
The online monitoring module is configured to monitor nuclear power activation pump rolling bearing inner ring, outer ring, rolling body vibration data, acceleration data, and bearing speed and transmission power.
The online monitoring module includes a field detecting device: consisting of an acceleration sensor and a torque sensor, an acceleration sensor is located on a bearing housing that activates a water pump motor fan end and a drive end, for collecting a vibration acceleration signal of the rolling bearing; torque sensor through the elastic column pin coupling The motor load is connected to the power and speed measurement; the corresponding data is collected using a 16-channel data recorder.
Server: The collected data is transmitted to the server via the data line. The server includes a central processor, a data storage center, and can store the collected data for easy data reading.
The fault diagnostic module is configured to perform fault monitoring diagnosis based on real-time monitoring data and convolutional neural network bearing fault diagnosis model, and derive corresponding diagnostic results;
The health assessment module is used to evaluate the health status of the bearing with the bearing wear diameter according to the fault diagnosis.
The alarm module is used to faulty bearing alarm prompt and transmit fault information to a smart transporter subsystem;
The bearing intelligent transporter subsystem, including operation and maintenance recommendations, and intelligent maintenance scheme decision modules.
The transportation recommendation information is used to make a corresponding maintenance plan based on nuclear power expert knowledge based on the health, maintenance costs, maintenance time, etc., based on nuclear power expert knowledge, according to the different faults of nuclear power starting water pump rolling bearings.
The intelligent maintenance scheme decision module is used to find the corresponding operation and maintenance recommendations in the operation and maintenance recommendation library according to the health state of the bearing, and provide nuclear power maintenance staff.
1) Sampling processing: For a sample sequence, several sampled values are sampled once, such a new sequence is the sample of the original sequence, for example, for the image I size of m * n, for it Double sample, that is, the resolution image of (m / s) * (n / s) size, where s should be the number of conventions of M and N
2) Convolutional Neural Networks is a kind of feedforward neural network comprising convolutional calculations and has depth structures, is one of the representative algorithms of deep learning. Convolutional neural networks have the ability to characterize learning, and can translate the input information in its class structure.
3) Model training: Multi-classification of image data sets. The model is constructed in the Tensorflow depth learning framework, using a multi-class combination of the neural network layer such as CNN, the input of the model is a two-dimensional characteristic map, the output of the model is multi-classification probability, and the final output classification category of algorithm, etc. . At training, the model approces the correct trend by the target function such as cross entropy.
4) SoftMAX function: Normalization index function is a promotion of logic functions, as follows:
Where ViIt is the output of the classifier preamplifier output unit. i means that the category index, the total category is C. SiIndicates the ratio of the current element index and all of the element index and the ratio. It enables a KD vector "compress" to another K-Dimensional vector such that each element is between [0, 1], and all elements and 1. That is, through the SoftMAX index, the multi-class output value can be converted to the relative probability.
 The present invention utilizes the characteristics of convolutional neural network "end-to-end", skips the internal mechanism analysis, Fourier transform, etc., directly acts on the original operation data, and avoids the loss of raw information. The super parameters in the traditional convolutional neural network, the selection of the activation function is based on the test error. The present invention adopts the method of experimental research to optimize the convolutional neural network parameters, further improving the model training efficiency. In order to solve the status quo of nuclear power, the real-time monitoring diagnosis and health dynamic evaluation method of the development of bearing faults, and the development of scientific and effective health management strategies based on assessment information, ultimately improve the reliability and stability of nuclear power units, avoid excessive maintenance, improve Enterprise Running Efficiency and Economic Benefits.