阅读说明：本技术 多尺度时空卷积深度信念网络的风力发电机故障诊断方法 (Wind driven generator fault diagnosis method of multi-scale space-time convolution depth belief network ) 是由 王洪斌 王红 江国乾 王跃灵 郑正 苏博 于 2020-03-19 设计创作，主要内容包括：本发明提出一种多尺度时空卷积深度信念网络的风力发电机故障诊断方法,该方法利用SCADA多变量时间序列固有的时空相关性和交互性特性,设计具有不同卷积核结构的卷积深度信念网络以级联的方式捕获传感器变量间的时空相关性信息,同时以并行的方式在多个滤波器尺度下挖掘变量间交互互补的特征,上述技术手段融合了时空依赖性提取和多尺度特征学习方法,因此能够提取更为丰富的故障诊断信息,与传统的卷积深度信念网络模型及其变体相比,本发明能够增强分类性能,为风力发电机故障诊断领域提供了新的途径。(The invention provides a wind driven generator fault diagnosis method of a multi-scale time-space convolution depth belief network, which utilizes the inherent time-space correlation and interactive characteristics of an SCADA multivariable time sequence to design the convolution depth belief network with different convolution kernel structures to capture the time-space correlation information among sensor variables in a cascading mode and simultaneously excavate the interactive complementary characteristics among the variables in a parallel mode.)

1. A wind driven generator fault diagnosis method of a multi-scale space-time convolution depth belief network is characterized by comprising the following steps:

step S1: collecting multivariate time sequence data collected by a data collection and monitoring control system under different health states of the wind driven generator; respectively preprocessing the multivariate time sequence data aiming at the health state to obtain a two-dimensional multivariate time sequence input matrix with a specified range;

step S2: inputting the two-dimensional multivariable time sequence input matrix into M convolution depth belief networks respectively in a parallel mode for multi-scale spatial feature learning, and extracting multi-scale spatial features under different filter scales, wherein M is a natural number and is more than 1;

step S3: inputting the multi-scale spatial features obtained in the step (2) into other M convolution depth belief networks different from the step (S2) in a parallel mode respectively for multi-scale time feature learning, and extracting effective multi-scale space-time features under different filter scales;

step S4: and (4) inputting the multi-scale space-time characteristics obtained in the step (3) into a softmax classifier for classification to generate a final diagnosis result.

2. The wind turbine generator fault diagnosis method of the multi-scale space-time convolution depth belief network of claim 1, characterized by: the step S1 includes the following specific steps:

step S11, normalizing the acquired original multivariate time series by using a maximum and minimum normalization method, wherein the calculation formula is as follows:wherein y is_ijIs the ith value, x, of the variable j in the normalized multivariate time series_ijIs the ith value, min (x), of variable j in the original multivariate time series_j) And max (x)_j) The minimum and maximum values of variable j, respectively;

and step S12, dividing the standardized multivariable time sequence into a plurality of non-overlapping two-dimensional segments with the length of N by using a sliding window technology to obtain a two-dimensional multivariable time sequence input matrix.

3. The wind driven generator fault diagnosis method of the multi-scale space-time convolution depth belief network as claimed in claim 1, wherein in step S2, the size of the two-dimensional multivariate time sequence input matrix is S × N, wherein S is the number of sensor variables, and N is the number of sampling points, i.e. the length of a sliding window, and the multi-scale space feature learning is performed by using three convolution depth belief networks, specifically comprising the following steps:

step S21, the two-dimensional multivariable time sequence input matrix obtained in the step 1 is respectively input to three convolution depth belief network modules with different filter scales in parallel, and a filter of each convolution depth belief network is designed to slide along a variable axis;

step S22, setting the number of layers of each convolution depth belief network module, wherein each module comprises two hidden layers and a probability maximum pooling layer respectively, and the size of a filter of each probability maximum pooling layer is the same;

and step S23, in order to keep the time dimension unchanged, cascading the local space feature mappings generated by the three convolution depth belief network modules along the direction of the variable axis for further multi-scale time feature extraction.

4. The wind turbine generator fault diagnosis method of the multi-scale space-time convolution depth belief network of claim 1, characterized by: the step S3 includes the following specific steps:

step S31, respectively inputting the multi-scale space characteristics learned in the step 2 into three convolution depth belief network modules with different filter scales in parallel, and designing a filter of each convolution depth belief network to slide along a time axis only;

step S32, setting the number of layers of the three convolution depth belief network modules in the step S31, wherein each module comprises two hidden layers and one probability maximum pooling layer, and the size of the filter of each probability maximum pooling layer is the same;

and step S33, cascading the local time features extracted at different filter scales along a time axis for final fault identification.

5. The wind turbine generator fault diagnosis method of the multi-scale space-time convolution depth belief network of claim 1, characterized by: the step S4 includes the following specific steps:

step S41, defining the classification task of the wind driven generator fault diagnosis as a two-classification problem;

step S42, converting the multi-scale space-time characteristics obtained in the step 3 into a two-dimensional matrix and inputting the two-dimensional matrix into a softmax classifier with a cross entropy loss function to diagnose whether the wind driven generator is in a healthy state; the calculation formula of the cross entropy function is as follows:where p (i) represents the true distribution and q (i) represents the predicted distribution.

Technical Field

The invention belongs to the technical field of wind generating set fault diagnosis, and particularly relates to a wind driven generator fault diagnosis method of a multi-scale space-time convolution depth belief network.

Background

In recent years, wind energy has received much attention from all over the world as a clean renewable energy source which is inexhaustible and rapidly developed. Wind power generators have been widely used on land and at sea because of their important role in wind power generation. However, in practical applications, wind power generators are usually operated in a harsh working environment all the day long, are subject to various complex actions for a long time, are very prone to faults, and even lead to unit shutdown in severe cases. The economic benefit of the wind power plant and the healthy development of the wind power industry are seriously influenced by the faults and the unplanned shutdown, so that the method has important practical significance for timely finding and diagnosing the faults of the wind turbine.

At present, methods based on physical models are widely applied to wind turbine fault diagnosis. However, the wind turbine not only is a complex electromechanical system composed of a plurality of subsystems and components, but also has complex and variable dynamic operation conditions, and it is difficult to effectively establish an accurate mathematical model, which greatly limits the development and application of a physical model-based method in wind turbine fault diagnosis. With the development of advanced sensor technology, a data-driven fault diagnosis method relying only on measurement data becomes a research hotspot. At present, a modern large wind turbine is provided with a Data Acquisition and monitoring Control (SCADA) system, which can collect and record a large amount of operation state information related to the wind turbine and key components thereof, so that the SCADA Data provides possibility for fault diagnosis due to availability and richness of sensor information.

At present, a method for realizing wind turbine fault diagnosis by processing wind turbine SCADA data is available, but the SCADA data is essentially a multivariate time sequence and has typical time-space correlation and interactive characteristics, and the existing wind turbine fault diagnosis method generally lacks the capability of capturing the characteristics, so that the fault classification performance is influenced.

Disclosure of Invention

The invention aims to provide a fault diagnosis method which can effectively identify the fault type of a wind driven generator based on SCADA data and has engineering practical value.

In order to solve the technical problem, the invention provides a wind driven generator fault diagnosis method of a multi-scale space-time convolution depth belief network, which comprises the following steps:

step S1: collecting multivariate time sequence data collected by a data collection and monitoring control system under different health states of the wind driven generator; respectively preprocessing the multivariate time sequence data aiming at the health state to obtain a two-dimensional multivariate time sequence input matrix with a specified range;

step S2: inputting the two-dimensional multivariable time sequence input matrix into M convolution depth belief networks respectively in a parallel mode for multi-scale spatial feature learning, and extracting multi-scale spatial features under different filter scales, wherein M is a natural number and is more than 1;

step S3: inputting the multi-scale spatial features obtained in the step (2) into other M convolution depth belief networks different from the step (S2) in a parallel mode respectively for multi-scale time feature learning, and extracting effective multi-scale space-time features under different filter scales;

step S4: and (4) inputting the multi-scale space-time characteristics obtained in the step (3) into a softmax classifier for classification to generate a final diagnosis result.

Further, the step S1 includes the following specific steps:

step S11, normalizing the acquired original multivariate time series by using a maximum and minimum normalization method, wherein the calculation formula is as follows:wherein y is_ijIs the ith value, x, of the variable j in the normalized multivariate time series_ijIs the ith value, min (x), of variable j in the original multivariate time series_j) And max (x)_j) The minimum and maximum values of variable j, respectively;

and step S12, dividing the standardized multivariable time sequence into a plurality of non-overlapping two-dimensional segments with the length of N by using a sliding window technology to obtain a two-dimensional multivariable time sequence input matrix.

Further, in step S2, the size of the two-dimensional multivariate time series input matrix is S × N, where S is the number of sensor variables, N is the number of sampling points, i.e., the length of the sliding window, and three convolutional deep belief networks are used to perform multi-scale spatial feature learning, specifically including the following steps:

step S21, the two-dimensional multivariable time sequence input matrix obtained in the step 1 is respectively input to three convolution depth belief network modules with different filter scales in parallel, and a filter of each convolution depth belief network is designed to slide along a variable axis;

step S22, setting the number of layers of each convolution depth belief network module, wherein each module comprises two hidden layers and a probability maximum pooling layer respectively, and the size of a filter of each probability maximum pooling layer is the same;

and step S23, in order to keep the time dimension unchanged, cascading the local space feature mappings generated by the three convolution depth belief network modules along the direction of the variable axis for further multi-scale time feature extraction.

Further, the step S3 includes the following specific steps:

step S31, respectively inputting the multi-scale space characteristics learned in the step 2 into three convolution depth belief network modules with different filter scales in parallel, and designing a filter of each convolution depth belief network to slide along a time axis only;

step S32, setting the number of layers of the three convolution depth belief network modules in the step S31, wherein each module comprises two hidden layers and one probability maximum pooling layer, and the size of the filter of each probability maximum pooling layer is the same;

and step S33, cascading the local time features extracted at different filter scales along a time axis for final fault identification.

Further, the step S4 includes the following specific steps:

step S41, defining the classification task of the wind driven generator fault diagnosis as a two-classification problem;

step S42, converting the multi-scale space-time characteristics obtained in the step 3 into a two-dimensional matrix and inputting the two-dimensional matrix into a softmax classifier with a cross entropy loss function to diagnose whether the wind driven generator is in a healthy state; the calculation formula of the cross entropy function is as follows:where p (i) represents the true distribution and q (i) represents the predicted distribution.

Compared with the prior art, the invention has the technical progress that:

the invention provides a wind driven generator fault diagnosis method of a multi-scale time-space convolution depth belief network, which utilizes the inherent time-space correlation and interactive characteristics of an SCADA multivariable time sequence to design the convolution depth belief network with different convolution kernel structures to capture the time-space correlation information among sensor variables in a cascading mode, and simultaneously excavates the interactive complementary characteristics among the variables in a parallel mode under the scale of a plurality of filters.

Drawings

FIG. 1 is a flow chart of one embodiment of the present invention;

FIG. 2 is a flow diagram of multi-scale spatial feature learning according to an embodiment of the present invention;

FIG. 3 is a flow diagram of multi-scale temporal feature learning according to an embodiment of the present invention; and

FIG. 4 is a schematic diagram of performance comparison evaluation of one embodiment of the present invention with other methods.

Detailed Description

Hereinafter, embodiments of the present invention will be described with reference to the drawings.

The embodiment of the invention adopts a universal 5 MW-grade offshore wind power system reference model, the reference model simulates an actual three-blade variable-speed horizontal axis wind power generation system by using a FAST (Fatigue, Aerodynemics, Structures, and Turbulence) simulation platform, the cut-in wind speed, the rated wind speed and the cut-out wind speed are respectively 3m/s, 11.4m/s and 25m/s, and a real wind data sequence with the average wind speed of respectively 11m/s, 14m/s and 17m/s can be generated. In the embodiment of the invention, an experimental data set is generated by using an average wind speed of 17m/s, in addition, the reference model can obtain 15 sensor variables from a real wind power SCADA system, wherein the measured value of each sensor variable is obtained by adding band-limited white Gaussian noise into the real value of the sensor variable to perform modeling, and based on the reference model, the real fault situation of the wind driven generator can be further defined.

The diagnosis process of the wind driven generator fault diagnosis method of the multi-scale space-time convolution depth belief network adopted by the embodiment of the invention is shown in figure 1 and comprises the following steps:

step S1: collecting multivariate time sequence data collected by a data collection and monitoring control system under different health states of the wind driven generator; for each health state, respectively preprocessing the collected data to obtain a two-dimensional multivariable time series input matrix with a specified range, and the specific method is as follows:

step S11, normalizing the acquired original multivariate time series by using a maximum and minimum normalization method, wherein the calculation formula is as follows:wherein y is_ijIs in a multi-variable time series after normalization processingIth value, x, of variable j_ijIs the ith value, min (x), of variable j in the original multivariate time series_j) And max (x)_j) The minimum and maximum values of variable j, respectively;

s12, dividing the standardized multivariate time sequence into a plurality of non-overlapping two-dimensional segments with the length of N by using a sliding window technology to obtain a two-dimensional multivariate time sequence input matrix;

step S2, as shown in FIG. 2, respectively inputting a two-dimensional multivariate time sequence input matrix into M convolution depth belief networks in a parallel mode for multi-scale space feature learning, extracting multi-scale space features under different filter scales, wherein M is a natural number and M is more than 1, the size of the two-dimensional multivariate time sequence input matrix is S × N, wherein S is the number of sensor variables, N is the number of sampling points, namely the length of a sliding window, and the three convolution depth belief networks are adopted for multi-scale space feature learning, and the specific steps are as follows:

step S21, the two-dimensional multivariable time sequence input matrix obtained in the step 1 is respectively input to three convolution depth belief network modules with different filter scales in parallel, and a filter of each convolution depth belief network is designed to slide along a variable axis;

step S22, setting the number of layers of each convolution depth belief network module, wherein each module comprises two hidden layers and a probability maximum pooling layer respectively, and the size of a filter of each probability maximum pooling layer is the same;

step S23, in order to keep the time dimension unchanged, the local space feature mappings generated by the three convolution depth belief network modules are cascaded along the direction of the variable axis for further multi-scale time feature extraction;

step S3: as shown in fig. 3, the multi-scale spatial features obtained in step 2 are respectively input into another three convolution depth belief networks different from that in step S2 in a parallel manner to perform multi-scale temporal feature learning, and multi-scale spatio-temporal features effective under different filter scales are extracted, specifically including the following steps:

step S31, respectively inputting the multi-scale space characteristics learned in the step 2 into three convolution depth belief network modules with different filter scales in parallel, and designing a filter of each convolution depth belief network to slide along a time axis only;

step S32, setting the number of layers of the three convolution depth belief network modules in the step S31, wherein each module comprises two hidden layers and a probability maximum pooling layer, and the size of the filter of each probability maximum pooling layer is the same;

step S33, cascading the local time characteristics extracted under different filter scales along a time axis for final fault identification;

step S4: inputting the multi-scale spatio-temporal features obtained in the step 3 into a softmax classifier for classification to generate a final diagnosis result, and specifically comprising the following steps:

step S41, defining the classification task of the wind driven generator fault diagnosis as a two-classification problem;

step S42, converting the multi-scale space-time characteristics obtained in the step 3 into a two-dimensional matrix and inputting the two-dimensional matrix into a softmax classifier with a cross entropy loss function to diagnose whether the wind driven generator is in a healthy state; the calculation formula of the cross entropy function is as follows:where p (i) represents the true distribution and q (i) represents the predicted distribution.

The experimental results are as follows:

the embodiment of the invention considers six health states of the wind driven generator, and the states comprise different fault types and fault severity degrees. In particular to the normal (state 1) of the wind driven generator, the proportional gross-loss fault of the generator speed sensor (gain factor equal to 1.2, state 2), the proportional gross-loss fault of the generator power sensor (gain factor equal to 1.2, state 3), the viscous fault of the pitch angle sensor (fixed value equal to 1 degree and 5 degrees respectively, state 4 and 5) and the proportional gross-loss fault of the pitch angle sensor (gain factor equal to 1.2, state 6). In the fault classification stage of the wind driven generator, the state 1 is defined as a normal state, and the states 2 to 6 are collectively defined as fault states. In order to effectively diagnose the fault of the wind driven generator, the average result of ten times of repeated operation is used as the final diagnosis result in the experiment. Fig. 4 shows a graph of the average diagnosis result of the present invention and the conventional convolutional deep belief network and its variants, from which it can be seen that the accuracy, precision and F1 score are significantly improved from other variants of the convolutional deep belief network to the present invention, and the present invention obtains enhanced fault diagnosis performance. The wind power generator fault diagnosis method mainly aims at the wind power SCADA multivariable time sequence to extract and classify features so as to realize wind power generator fault diagnosis, and the technical core is that in order to obtain better diagnosis performance, the time-space features hidden in SCADA data are better captured by designing convolution depth belief networks with different convolution kernel structures, and because different filter scales can extract and learn the interactive correlation features beneficial to fault classification, the method simultaneously integrates different filter scales to mine the interactive characteristics among multiple sensor variables. This result further illustrates that the present invention is worth applying to the fault diagnosis of an actual wind turbine.

The above-mentioned embodiments are merely illustrative of the preferred embodiments of the present invention, and do not limit the scope of the present invention, and various modifications and improvements made to the technical solution of the present invention by those skilled in the art without departing from the spirit of the present invention shall fall within the protection scope defined by the claims of the present invention.