| CPC G06N 3/084 (2013.01) [G06N 3/045 (2023.01)] | 2 Claims |
|
1. A prediction method for stall and surge of an axial compressor based on deep learning, comprising the following steps:
S1. preprocessing data with stall and surge of an aeroengine, comprising:
S1.1. partitioning a test data set and a training data set from experimental data before processing the experimental data;
S1.2. filtering pressure change data measured at all measure points in the training data set by a low-pass filter;
S1.3. down-sampling the filtered pressure change data;
S1.4. sharding time domain data according to the size of a time step and labeling each sharded sample; setting the time step to be 256, setting a time window with a length of 256, sliding the time window over time domain data, sharding the time domain data falling in the time window as a sample; and assigning a label of 1 or 0 to each sample depending on whether a surge occurs or not;
S1.5. partitioning the training data set into a training set and a validation set in a 4:1 ratio;
S2. constructing a logistic regression (LR) branch network module, comprising:
S2.1. extracting six time domain statistical characteristics in total of each sample, including variance, kurtosis, skewness, average value, minimum value and maximum value, and taking same as the input of the LR branch network module;
S2.2. setting up a single-activation-layer neural network with the Rectified Linear Unit (ReLU) activation function, wherein a number of neurons of an input layer is 6 and a number of neurons of the output layer is 1, obtaining the output of the LR branch network module, a dimension thereof being (m,1), where m represents a number of samples being determined in S1.4;
S3. constructing a WaveNet branch network module, comprising:
S3.1. adjusting the dimension of each sample to (n_steps,1), and taking same as the input of the WaveNet branch network module, where n_steps represents time steps;
S3.2. setting up a dilated convolution module based on causal convolution and dilated convolution, and setting two identical dilated convolution modules; introducing gated activations into each layer of convolution to adjust a information transmitted to a next layer, adopting residual connections between one layer and another layer to prevent a gradient from disappearing, adopting skip connections to reserve an output of each convolution layer, and summating output characteristics of all layers to obtain the output characteristics of the dilated convolution module;
S3.3. fully connecting the output characteristics extracted by the dilated convolution module by multiple layers, and activating by means of the ReLU activation function to obtain an output of the WaveNet branch network module, a dimension thereof being (m,1);
S4. constructing an LR-WaveNet prediction model, comprising:
S4.1. fusing the LR branch network module and the WaveNet branch network module by means of a stacking algorithm, splicing the outputs of the LR branch network module and the WaveNet branch network module, obtaining a fusion output of which the dimension is (m,2), and taking a same input of the stacking fusion module;
S4.2. setting up a stacking fusion module, activating an output by means of two layers of neural network plus sigmoid to obtain a probability of surge, which is used as an output of the LR-WaveNet prediction model;
S4.3. handling a problem existing in the training of data with stall and surge by means of a modified focal loss function,
wherein an improved focusing loss function is:
MFL(pt)=−αtβt(1−pt)γ log(pt)
where MFL represents modified focal loss, α1 represents a class weight coefficient, β1 represents a weight coefficient of importance degree, p1 represents a model prediction probability, and γ represents a regulatory factor parameter;
S4.4. based on the modified focal loss function, updating a weight of a model on the training data set, specifically:
the output of an output layer of the WaveNet branch network module is:
α(L)=f(z(L))=f(W(L)x(L)+b(L))
where L represents the output layer of the WaveNet branch network module; W(L) represents connection weight; b(L) represents bias; x(L) represents input of the output layer; z(L) represents a result of x(L) after linear transformation; f( ) represents the ReLU activation function; and a(L) represents the output layer of the WaveNet branch network module;
the output of the output layer of the LR branch network module is:
α(LR)=f(z(LR))=f(W(LR)x(LR)+b(LR))
where LR represents the output layer of the LR branch network module; W(LR) represents connection weight; b(LR) represents bias; x(LR) represents an input characteristic of the branch neural network; z(LR) represents a result of x(LR) after linear transformation; f( ) represents a second ReLU activation function; and a(LR) represents the output of the LR branch network module;
splicing the outputs of the LR branch network module and the WaveNet branch network module:
α(L′)=[α(L),α(LR)]
z(L′)=[z(L),z(LR)]
where L′ represents a new layer after the outputs of the two branch network modules are spliced, a(L′) represents output of a splicing layer, and z(L′) represents an input of an activation function of the splicing layer;
conducting back propagation on an error of the output layer of the LR-WaveNet prediction model according to a back propagation formula, obtaining that an error on the output layer of the WaveNet branch network module is:
 an error on the output layer of the LR branch network module is:
 where δi(l) represents an error of the modified focal loss function on the ith node of the lth layer; Wij(l) represents a weight coefficient between a jth node of a lth layer and an ith node of the previous layer; sl+1 represents a number of nodes of a (l+1)th layer;
conducting back propagation on the branch networks respectively according to errors on the output layers of the two branch network modules, and updating model parameters;
S4.5. saving the trained model and testing on the validation set, adjusting a hyper-parameter of the trained model according to an evaluation index of the validation set, adopting an F2 index as an evaluation index, and saving a model which makes an evaluation index optimal as a final LR-WaveNet prediction model;
S5. conducting real-time prediction on test data:
S5.1. preprocessing test set data according to the steps of S1, and adjusting data dimensions according to input requirements of the LR-WaveNet prediction model;
S5.2. giving surge prediction probabilities of all samples by means of the LR-WaveNet prediction model according to a time sequence;
S5.3. intercepting a segment of data without surge inception from the test set data, adding a segment of noise points to the segment of data, and giving a probability of surge that noise point data changes over time by means of the LR-WaveNet prediction model, to test an anti-interference performance of the model.
|