Application of an Improved BP Network in Fault Diagnosis of Turbogenerator Units

An improved BP network is applied to the fault diagnosis of steam turbine generator set. He Chengbing, Gu Yuyang, and the feedback algorithm of North China Electric Power University, and explains the basis of the connection between the input layer and the output layer and the determination of the connection priority value. Combined with the fault diagnosis example of steam turbine generator set, it is proved that the model has strong fault recognition ability from the perspective of single fault identification and multi-fault identification, and its diagnosis result is more in line with the actual situation of the fault.

The impact of the sign on the output is shown in the figure. A partial connection is established between the output node and the characteristic input node of the corresponding fault, such as an output node representing an unbalanced fault connected to an input node representing a frequency of multiples of the vibration.

For the fault diagnosis of steam turbine generator sets, there is no simple one-to-one correspondence between the symptoms and the faults, but a situation of staggered complexity. A fault may correspond to multiple symptoms, and one symptom may correspond to multiple faults. In a multi-signal failure, some signs play a major role and are the main signs; some signs play a secondary role and are auxiliary signs. The contribution of these symptoms to the diagnosis results is also different. Therefore, different priority values ​​are added between the directly connected input and output neurons, the priority value of the main symptom is high, and the priority value of the auxiliary symptom is low.

The large steam turbine generator set is a complex system. It integrates the machine and electric technology, and has high reliability requirements. Once the system fails, it will cause huge economic losses and even serious social consequences. Therefore, online or offline troubleshooting of such large electromechanical systems must be performed and the diagnosis must be fast, accurate and efficient.

In the long-term practice, many effective fault diagnosis methods have been proposed, such as fuzzy diagnosis, pattern recognition, statistical Bayesian method, grey theory, etc. However, in recent years, neural network theory has been widely used in the field of mechanical fault diagnosis. . We know that fault diagnosis is to classify or identify the fault mode of a machine or system. The neural network has complex nonlinear mapping, associative memory and self-learning ability. It is an excellent pattern recognizer, so it is very suitable for machinery. Troubleshooting. Neural networks have many models, such as feedforward neural networks, self-organizing feature maps, and bidirectional associative memories. Among them, the feedforward BP neural network trained by the error back propagation algorithm has the ability to form an arbitrarily complex decision surface, which is the most widely used model in the neural network. Based on BP neural network, this paper proposes an improved BP network model with priority value and elaborates on it.

1 Improve BP network model 1.1 Improve BP network model structure 1.2 Improve the feedforward and learning algorithm of BP network model The main adjustment design parameter in this network is the connection weight between input layer and hidden layer "(", hidden layer and The connection weight between the output layers is (2), the threshold of each node of the hidden layer and the output layer, 0丨2), and the partial connection weight % between the input layer and the output layer.

The standard BP network has the ability to associate different degrees of single fault conditions, but the resolution is low, and the ability to diagnose multiple faults is poor, and often only one of the faults can be diagnosed. In response to this situation, part of the connection between the input and output neurons can be added on the basis of the BP network to increase the input of He Chengbing (1974-), male, doctoral students. Powertrain, 102206 single fault diagnosis example uses the same fault sample pair BP network and the improved BP network with priority value in this paper. The test samples and results are shown in Table 2. It can be seen from the table: (1) BP network has strong association ability and high resolution for single faults; (2) Two BP networks have similar performance for single fault diagnosis.

Single Fault Diagnosis Example The feedforward calculation formula is still used between the network layers of the model: the connection priority value between the input layer and the output layer.

1.3 Improving the BP network model The determination of the priority value is known from equation (2). The determination of the connection priority value *ji is very important. It actually includes two problems: first, determine which input and output neurons are between Need to connect, which do not need to connect (no doubt, no connection, ji = 0); Second, how to determine the priority value of the input and output neurons that establish the connection. For the improved BP network model in this paper, the output node establishes a connection with the characteristic input node of the corresponding fault, and the priority value can be determined according to the expert experience, or can be obtained from the fault sample by statistical analysis, and the fault sample is not large enough. In this case, the priority value should be determined from the vibration failure mechanism and combined with expert experience. For the main symptom of the fault, the priority value is 1, and other symptoms are based on the contribution of the fault diagnosis result. The smaller the contribution is, the larger the contribution is, the smaller the contribution is, the smaller the contribution is. Taking the misalignment as an example, the fault mechanism analysis shows that the 1st and 2nd frequencies of the vibration frequency are the main symptoms, accompanied by the auxiliary signs of the 3x and above multipliers, then the 4th line.

2 Turbogenerator Fault Diagnosis Example 2.1 Model Structure and Sample Selection In this paper, the calculation example is taken as an example. The rotor test bench simulates three fault modes, taking five vibration symptom inputs and three fault cause outputs. The network structure corresponds to: input layer, 5 neurons; hidden layer, 4 neurons; output layer, 3 neurons. The specific structure is shown in the figure. In the figure, X1X5 corresponds to 0.40.5 octave, 1 octave, 2 octave, 3 octave and greater than 3 octave in the spectrum analysis of vibration signal respectively; F1F3 is three fault forms, respectively corresponding to large Oil film oscillation, large imbalance and big misalignment.

According to the mechanism analysis of the unit vibration failure, combined with the determination of the priority value of the neuron connection between the input layer and the output layer, see Table 1. Table 1 input layer and output layer connection priority value sequence simulation feature indication plus priority The value of the improved network output conclusion AA large, medium and large AA large, medium, large, medium and large AAB Note: A, B, C in the table represent oil film oscillation, imbalance and misalignment. According to the fuzzy mathematics principle, the definition output values ​​>0.75, 0.750.5, 0.5, and <0.5 represent severe, medium, unknown, and non-existent fault levels, respectively. The following tables are the same.

Table 3. Two improved network performances in a single fault case. The improved network output of the BP network output plus the priority value in the simulated characteristic feature is greatly improved.