CN114397474B - FCN-MLP-based arc ultrasonic sensing array wind parameter measurement method - Google Patents
FCN-MLP-based arc ultrasonic sensing array wind parameter measurement method Download PDFInfo
- Publication number
- CN114397474B CN114397474B CN202210047540.XA CN202210047540A CN114397474B CN 114397474 B CN114397474 B CN 114397474B CN 202210047540 A CN202210047540 A CN 202210047540A CN 114397474 B CN114397474 B CN 114397474B
- Authority
- CN
- China
- Prior art keywords
- wind
- fcn
- mlp
- array
- wind speed
- 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
- G01P—MEASURING LINEAR OR ANGULAR SPEED, ACCELERATION, DECELERATION, OR SHOCK; INDICATING PRESENCE, ABSENCE, OR DIRECTION, OF MOVEMENT
- G01P5/00—Measuring speed of fluids, e.g. of air stream; Measuring speed of bodies relative to fluids, e.g. of ship, of aircraft
- G01P5/24—Measuring speed of fluids, e.g. of air stream; Measuring speed of bodies relative to fluids, e.g. of ship, of aircraft by measuring the direct influence of the streaming fluid on the properties of a detecting acoustical wave
- G01P5/245—Measuring speed of fluids, e.g. of air stream; Measuring speed of bodies relative to fluids, e.g. of ship, of aircraft by measuring the direct influence of the streaming fluid on the properties of a detecting acoustical wave by measuring transit time of acoustical waves
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01P—MEASURING LINEAR OR ANGULAR SPEED, ACCELERATION, DECELERATION, OR SHOCK; INDICATING PRESENCE, ABSENCE, OR DIRECTION, OF MOVEMENT
- G01P13/00—Indicating or recording presence, absence, or direction, of movement
- G01P13/02—Indicating direction only, e.g. by weather vane
-
- 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
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
- Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02E—REDUCTION OF GREENHOUSE GAS [GHG] EMISSIONS, RELATED TO ENERGY GENERATION, TRANSMISSION OR DISTRIBUTION
- Y02E10/00—Energy generation through renewable energy sources
- Y02E10/70—Wind energy
- Y02E10/72—Wind turbines with rotation axis in wind direction
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y04—INFORMATION OR COMMUNICATION TECHNOLOGIES HAVING AN IMPACT ON OTHER TECHNOLOGY AREAS
- Y04S—SYSTEMS INTEGRATING TECHNOLOGIES RELATED TO POWER NETWORK OPERATION, COMMUNICATION OR INFORMATION TECHNOLOGIES FOR IMPROVING THE ELECTRICAL POWER GENERATION, TRANSMISSION, DISTRIBUTION, MANAGEMENT OR USAGE, i.e. SMART GRIDS
- Y04S10/00—Systems supporting electrical power generation, transmission or distribution
- Y04S10/50—Systems or methods supporting the power network operation or management, involving a certain degree of interaction with the load-side end user applications
Abstract
An FCN-MLP-based arc ultrasonic sensing array wind parameter measurement method belongs to the technical field of wind measurement, and comprises the following steps: constructing an arc array structure; constructing an array flow pattern according to the distribution condition of array elements in the constructed arc array structure; establishing array receiving vectors under different signal-to-noise ratios, performing label processing on wind speed and wind direction, and constructing a wind speed data set and a wind direction data set; randomly disordering the wind speed and wind direction data sets, and dividing the randomly disordering wind speed data sets into a training set, a verification set and a test set; constructing and training two FCN-MLP network models; and respectively inputting the wind speed data set and the wind direction data set into the trained FCN-MLP wind speed measurement neural network model and the trained FCN-MLP wind direction measurement neural network model to obtain measurement wind parameter information. The method of the invention realizes the accurate measurement of the wind parameters when the noise exists in the environment, reduces the complexity of calculation and improves the real-time performance of the measurement.
Description
Technical Field
The invention belongs to the technical field of wind measurement, and particularly relates to an arc ultrasonic sensing array wind parameter measurement method based on FCN-MLP.
Background
Wind energy is a clean energy source, and is successfully applied to the power generation link of a wind motor at present, wind power generation as a clean new energy source has important significance for energy shortage and environmental protection, and the wind power industry as a high and new technology industry has wide prospects. Wind speed and wind direction are two important factors that affect the power generation efficiency of a wind turbine: when the wind speed is low, the power generation capacity of the wind turbine generator is relatively weak, and the working efficiency is low; the working efficiency of the wind turbine generator is improved as the wind speed is increased, but the higher the wind speed is, the more the failure times of the motor are; the blades of the wind driven generator are subjected to wind force as much as possible to drive the wind wheel to rotate, and the wind wheel is converted into kinetic energy for power generation. The monitoring of the research wind speed and the wind direction can provide the capability of changing the direction for the fan, and the variable pitch and yaw system is realized through the output of the sensor and the control of central information, so that the efficiency of the wind driven generator is greatly improved. At present, an anemometer mainly used for measuring wind speed and wind direction includes a mechanical wind speed and wind direction sensor, an ultrasonic wind speed and wind direction sensor, and a thermoelectric wind speed and wind direction sensor. When the mechanical wind speed and direction sensor works, mechanical movement exists, and the friction resistance of the mechanical wind speed and direction sensor influences the wind speed and direction measuring effect and the service life; the ultrasonic wind speed and direction sensor influences the measurement precision of the wind speed and the wind direction on the measurement precision of the propagation time of signals in upwind and downwind; thermoelectric wind speed and direction sensors are expensive and cannot adapt to drastic changes in temperature. Based on other array wind measuring structures in the array signal processing theory, the simulation experiment result shows that when the signal-to-noise ratio is low, the wind speed and direction estimation success rate is low, the root mean square error is large, and the noise has a large influence on the measurement accuracy of the wind signal and the anti-interference capability of the wind speed and direction measurement system. The estimation time of the technology is mostly different from 2 to 5 seconds, and the analysis shows that the conventional wind speed and direction measuring method has great problems in the aspects of wind speed and direction measuring accuracy, measuring real-time performance and anti-interference capability, and the traditional algorithm cannot quickly and fully capture the nonlinear characteristics of wind signals. Therefore, how to improve the measurement accuracy, measurement real-time performance, measurement range and anti-interference performance of the wind speed and direction is an urgent problem to be solved in the field of wind speed and direction measurement.
Disclosure of Invention
Aiming at the technical problems of wind speed and wind direction measurement accuracy, measurement instantaneity, calculation complexity and anti-interference performance in the prior art, the invention aims to provide an arc ultrasonic sensing array wind parameter measurement method based on FCN-MLP, which constructs an arc array structure based on FNC-MLP, so that when noise exists in the environment, accurate measurement of wind parameters is realized, the calculation complexity is reduced, and the measurement instantaneity is improved.
The technical scheme adopted by the invention for realizing the purpose is as follows: the FCN-MLP-based arc ultrasonic sensing array wind parameter measuring method is characterized by comprising the following steps of:
s1, constructing an arc array structure: the arc array structure is composed of transmitting array elements and receiving array elements, the number of the receiving array elements is at least two and is a multiple of 2, and all the receiving array elements are uniformly arranged on an arc which takes the transmitting array elements as the circle center and R as the radius; the transmitting array element and the receiving array element are both ultrasonic sensors;
s2, according to the distribution condition of array elements in the constructed arc array structure, decomposing the wind signal to be tested by a vector decomposition method to obtain components in the connecting line direction of the transmitting array elements and the receiving array elements, selecting a certain receiving array element as a reference array element, obtaining the difference value between the time from the transmitting array element to each receiving array element and the time from the transmitting array element to the reference array element, and constructing an array flow pattern;
s3, establishing array receiving vectors under different signal-to-noise ratios, performing label processing on wind speed and wind direction, and constructing a wind speed data set and a wind direction data set;
s4, randomly scrambling the wind speed data set, dividing the randomly scrambled wind speed data set into a training set, a verification set and a test set, randomly scrambling the wind direction data set, and dividing the randomly scrambled wind direction data set into the training set, the verification set and the test set;
s5, building and training an FCN-MLP network model
Respectively constructing an FCN-MLP wind speed measurement neural network model and an FCN-MLP wind direction measurement neural network model and training, wherein the FCN-MLP wind speed measurement neural network model comprises a first FCN full convolution network, a first MLP full connection neural network and a flat layer connected between the first FCN full convolution network and the first MLP full connection neural network; the FCN-MLP wind direction measurement neural network model comprises a second FCN full convolution network, a second MLP full connection neural network and a flat layer connected between the second FCN full convolution network and the second MLP full connection neural network;
and S6, respectively inputting the wind speed data set and the wind direction data set into the trained FCN-MLP wind speed measurement neural network model and the trained FCN-MLP wind direction measurement neural network model to obtain measurement wind parameter information.
Further, the step S3 specifically includes: obtaining an array receiving vector matrix through an array flow pattern, adding Gaussian white noise with signal-to-noise ratios of 0dB, 5dB, 10dB, 15dB, 20dB and 30dB to obtain array receiving vectors with different signal-to-noise ratios, splicing a real part and an imaginary part of the array receiving vectors in a matrix column, adding wind speed and wind direction labels in the last column of the matrix, and respectively storing the labels into a wind speed data set, a csv file and a wind direction data set, the csv file and constructing the wind speed data set and the wind direction data set.
Further, the wind speed data set and the wind direction data set are divided into 80% of training set, 10% of verification set and 10% of test set.
Further, the first FCN full convolution network is composed of five convolution layers, and a Relu function is adopted as a convolution layer activation function in the first FCN full convolution network; the first MLP full-link neural network is composed of two full-link layers which are respectively named as: the DENSE1, DENSE2 and DENSE1 adopt Relu functions as activation functions, and the DENSE1 outputs 256 neurons; the activation function of DENSE2 adopts a softmax function; the second FCN full convolution network is composed of four convolution layers, and a Relu function is adopted as a convolution layer activation function in the second FCN full convolution network; the second MLP full-link neural network is composed of two full-link layers which are respectively named as: DENSE3, DENSE4 and DENSE3 adopt Relu function, DENSE3 outputs 128 neurons; the activation function of DENSE4 adopts a softmax function.
Further, in the process of training the model, a Cross Entropy loss function is adopted as a loss function, and an SGD (sparse Gate D) random gradient descent method optimizer is selected by the optimizer.
Through the design scheme, the invention can bring the following beneficial effects: compared with the array wind measurement structure and other wind measurement principles provided by the invention, the method for measuring the wind parameters of the arc ultrasonic sensing array based on the FCN-MLP reduces the calculation complexity on the premise of ensuring the high-precision estimation success rate of the wind speed and the wind direction, reduces the estimation time to one tenth of the original time, and realizes the error-free estimation when the signal-to-noise ratio is greater than 10dB.
Drawings
FIG. 1 is a flow chart of a method for measuring wind parameters of an arc ultrasonic sensing array based on FCN-MLP;
FIG. 2 is a schematic view of an arc array structure;
FIG. 3 is a diagram of a FCN-MLP anemometry neural network model architecture;
FIG. 4 is a diagram of a FCN-MLP wind direction measurement neural network model structure;
FIG. 5 is a diagram illustrating the structure of neurons and the calculation process;
FIG. 6 is a flow chart of a simulation experiment of the FCN-MLP-based arc ultrasonic sensing array wind parameter measurement method provided by the invention;
FIG. 7 is a visual diagram of the real part of the receiving vector X (t) of the wind direction 300 array at a wind speed of 10m/s according to the embodiment of the invention;
fig. 8 is a diagram illustrating the effect of estimating success rate in the embodiment of the present invention.
Detailed Description
In order to more clearly illustrate the invention, the invention is further described below with reference to preferred embodiments and the accompanying drawings. It is to be understood by persons skilled in the art that the following detailed description is illustrative and not restrictive, and is not to be taken as limiting the scope of the invention. Unless defined otherwise, technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention belongs.
1. FCN-MLP-based arc ultrasonic sensing array wind parameter measurement method
FIG. 1 shows a flow chart of a FCN-MLP-based method for measuring wind parameters of an arc ultrasonic sensing array, wherein the method comprises the following steps:
building arc array structure
The arc array structure is shown in fig. 2, and is composed of five ultrasonic sensors, one of the five ultrasonic sensors is a transmitting array element (sensor 0 in the figure), the other four ultrasonic sensors are receiving array elements (sensors 1-4 in the figure), the transmitting array element is taken as a center, the four receiving array elements are positioned on the same arc, an included angle between connecting lines of two adjacent receiving array elements and the transmitting array element is alpha, and alpha =200; theta is the wind direction angle to be measured, and the incoming direction of wind deviates from the vertical due north clockwise direction by theta.
According to the geometric structure of the array, the components of the wind speed V to be measured on the connecting lines of the transmitting array element and the four receiving array elements are obtained by using a vector decomposition method, wherein the components are V 1 、V 2 、V 3 And V 4 In which
Obtaining the time of the signal transmitted by the transmitting array element reaching the ith receiving array element as:
r is the arc radius, i.e. the distance between the transmitting array element and the receiving array element, c is the ultrasonic propagation speed, V i For the component of the wind speed V to be measured on the connection line of the transmitting array element and the ith receiving array element, taking the 1 st receiving array element as a reference array element, and obtaining the time delay of the transmitting signal reaching the ith receiving array element relative to the reference array element as follows:
τ 1 ＝t 1 -t 1 ＝0；
thus, the array flow pattern A can be found as:
the array receive vector is: x (t) = As (t) + N (t);
wherein X (t) = [ X ] 1 (t),……,x M (t)] T M is the number of receiving array elements, x i (t) receiving a vector for an array of ith (i =1,2, \8230;, M) array elements,
τ i for the time delay of the transmit signal to the i-th (i =1,2, \ 8230;, M) receive array element relative to the reference array element, f is the ultrasonic signal frequency.
For the transmitted ultrasonic signal, where u (t) is the amplitude of the received signal (which is the slow amplitude modulation function (real envelope)), and φ (t) is the phase of the received signal (which is the slow amplitude modulation function (real envelope)), w 0 And =2 pi f is the frequency of the received signal.
Constructing an array receiving vector X (t) under the condition of different signal-to-noise ratios, extracting a real part real (X (t)) and an imaginary part img (X (t)) of the X (t) for column splicing to obtain X' (t) because the X (t) is a complex matrix,
X′(t)＝[realX(t) img X(t)]
finally, corresponding wind speed and wind direction labels are added to the columns of X' (t), and since the labels can only be integers, the wind speed label v is 0-600, wherein v =0 corresponds to a wind speed of 0m/s, v =1 corresponds to a wind speed of 0.1m/s, and the like, and v =600 corresponds to a wind speed of 60.0m/s.
X' (t) is labeled by wind speed
X′ v (t)＝[real(X(t)) img(X(t)) v]
X 'to which a wind speed tag is added' v (t) storing the wind speed DATA set in a csv file to construct a wind speed DATA set DATA v ，
Wherein, X' v0 (t) is the array received vector, X ', for a wind speed of 0 m/s' v1 (t) is an array received vector, X ', corresponding to a wind speed of 0.1 m/s' v599 (t) is the array received vector, X ', for a wind speed of 59.9 m/s' v600 And (t) is an array receiving vector corresponding to the wind speed of 60.0m/s.
The wind direction label θ is 0 to 359, where θ =0 is 0 ° of the wind direction, θ =1 is 1 ° of the wind direction, and so on, and θ =359 is 359 °. X' (t) plus wind direction label is
X′ θ (t)＝[real(X(t)) img(X(t)) θ]
Adding wind direction label to X' θ (t) storing in a wind direction DATA set, csv file, constructing a wind direction DATA set, DATA θ ，
Wherein X' θ0 (t) is an array reception vector, X ', corresponding to wind direction 0 DEG' θ1 (t) is the array receive vector, X ', for 1 deg. of wind' θ358 (t) is the array receive vector, X ', for 358 deg. into the wind' θ359 (t) receive vectors for the array for 359 ° wind direction.
The csv file (i.e. an array receiving vector X (t) contains wind speed and wind direction information, so that one X (t) corresponds to one wind speed tag and one wind direction tag at the same time, and the difference between the two tag information is the difference between the wind speed and wind direction data sets). And respectively constructing two FNC-MLP network models for the wind speed and the wind direction based on the wind speed and the wind direction and training the two FNC-MLP network models. FIGS. 3 and 4 are FNC-MLP wind speed and wind direction measurement network structures.
The data preprocessing divides the disordered wind speed and wind direction data sets into three parts, wherein 80% of the three parts are training sets, 10% of the three parts are verification sets, and 10% of the three parts are testing sets.
Respectively constructing an FCN-MLP wind speed measurement neural network model and an FCN-MLP wind speed measurement neural network model, wherein the FCN-MLP wind speed measurement neural network model comprises a first FCN full convolution network, a first MLP full connection neural network and a flat layer FLATTEN1 connected between the first FCN full convolution network and the first MLP full connection neural network; the first FCN full convolution network is composed of five convolution layers, namely a convolution layer COV1, a convolution layer COV2, a convolution layer COV3, a convolution layer COV4 and a convolution layer COV5, and the first MLP full-connection neural network is composed of two full-connection layers and named as: the DENSE1, DENSE2 and FCN-MLP wind direction measurement neural network model comprises a second FCN full convolution network, a second MLP full connection neural network and a flat layer FLATTEN2 connected between the second FCN full convolution network and the second MLP full connection neural network; the second FCN full convolution network is composed of four convolution layers which are a convolution layer COV6, a convolution layer COV7, a convolution layer COV8 and a convolution layer COV9 respectively, the second MLP full-connection neural network is composed of two full-connection layers which are named as follows: DENSE3, DENSE4, FCN is a neural network composed only of convolutional layers, and MLP is a neural network composed only of fully-connected layers.
For the FCN-MLP anemometry neural network model, each convolution layer contains 32 neurons, the convolution kernel size of each layer is 1 × 32, and the fully-connected layer DENSE1 contains 256 neurons. The method comprises the steps of inputting an 8 x 320 matrix into a network, converting the matrix into a 4 x 320 x 2 matrix after dimension conversion (data preprocessing), generating a 4 x 289 x 32 matrix after passing through a convolutional layer COV1, generating a 4 x 258 x 32 matrix after passing through a convolutional layer COV2, generating a 4 x 227 x 32 matrix after passing through a convolutional layer COV3, generating a 4 x 196 x 32 matrix after passing through a convolutional layer COV4, generating a 4 x 165 x 32 matrix after passing through a convolutional layer COV5, generating a 1 x 21120 matrix after passing through a flat layer FLATTEN1, sending the generated matrix into a full connection layer DENSE1, and then sending the matrix into DENSE2 to obtain a measurement result.
For the FCN-MLP anemometry neural network model, each convolutional layer contains 64 neurons, the convolutional kernel sizes of COV6 and COV7 are 1 × 32, the convolutional kernel sizes of COV8 and COV9 are 1 × 64, and the fully-connected layer DENSE3 contains 128 neurons. The method comprises the steps of inputting an 8 x 320 matrix by a network, converting the matrix into a 4 x 320 x 2 matrix after dimension conversion (data preprocessing), generating a 4 x 289 x 64 matrix after passing through a convolution layer COV6, generating a 4 x 258 x 64 matrix after passing through a convolution layer COV7, generating a 4 x 195 x 64 matrix after passing through a convolution layer COV8, generating a 4 x 132 x 64 matrix after passing through a convolution layer COV9, generating a 1 x 33792 matrix after passing through a flat layer FLATTEN1, sending the generated matrix into a full connection layer DENSE3, and then sending the generated matrix into DENSE4 to obtain a measurement result.
The basic structure of the neural network is constructed by single neurons, the neurons are formed by data input and output and neural nodes, and the structure and the calculation process of the neurons are shown in fig. 5. The data entry is
X(X 1 ,X 2 ,…,X n )，X 1 Input to the 1 st neuron, X 2 Is the input of the 2 nd neuron, X 3 Input for the 3 rd neuron, X n The weight value is input into the nth neuron, each neuron has a corresponding weight value, and the weight values are repeatedly modified during network training until the output of the network can be close to the expected value, namely the label corresponding to each datum, so that the error of a training set is lower than 10%; the neuron output nodes and bias terms are y and b respectively, and the activation function is f, which is used for establishing a logistic regression relationship between the data input term and the data output term so as to carry out classification.
In the convolutional layers, a convolution kernel is used for extracting the characteristic information of the image, and for a certain convolutional layer, the convolution kernel and the characteristic image of the previous layer are subjected to sliding convolution operation, and then an offset value is added to obtain output; finally, obtaining a convolution result, namely an output characteristic diagram, through an activation function, as shown in the following formula:
in the above-mentioned formula, the compound has the following structure,
feature diagram of j, M, representing the l-th layer j For all the input feature maps,
and
is the convolution kernel of the l layer and the bias, and f (-) is the activation function, usually Relu function or Sigmoid function. The convolutional layer activation function in the present application employs the Relu function.
The flat layer flattens the output of the last convolution layer, outputs a 1 xN matrix, and connects the FCN full convolution network with the MLP full-connection neural network. The activation functions of DENSE1 and DENSE3 adopt Relu functions, in a wind speed measurement network, DENSE1 outputs 256 neurons, and in a wind direction measurement network, DENSE3 outputs 128 neurons; the DENSE2 and DENSE4 classify wind speed and wind direction, and the activation function adopts a softmax function.
The loss function is used for calculating a loss value so as to obtain a gradient Grad, a Cross Entropy loss function is adopted by a network designed by the application, and an SGD (random gradient descent optimizer) is selected by an optimizer.
The loss function is used to measure how close the predicted output value is to the actual value, which is defined in a single training sample, the behavior of which is measured. The cost function is used to train the model, which measures how well the algorithm behaves over all the training samples, by summing the penalty functions of m samples and then dividing by m. The gradient descent method is to minimize a cost function, and includes the steps of firstly, randomly initializing model parameters, and then continuously iterating towards the steepest descent direction until the model parameters reach the global optimal solution or a position close to the global optimal solution, namely a minimum value point of the cost function, so as to obtain the optimal model.
Training of the neural network requires a large number of data sets relevant to tasks to train the model, the model is continuously trained in an iterative mode through errors of the model on the data sets, a model which is reasonably fitted to the data sets is obtained, and finally the trained and adjusted model is applied to a real scene.
Firstly, 10% of a data set is selected for network model training to obtain appropriate network structure parameters, and experimental feasibility is verified. And training the FCN-MPL network model by using the global training set and the verification set, and testing the performance of the trained FCN-MPL network model by using the test set.
Selecting the step length to be 1 degree within the range of the angle between 0 and 359 degrees; and selecting the step length to be 0.1m/s within the range of 0 m/s-60 m/s, and outputting the model as the wind direction angle and the wind speed value to be measured through the trained FCN-MPL network model.
2. Simulation experiment
The simulation experiment process is as follows: selecting an ultrasonic emission signal; setting the frequency of ultrasonic signals to be 40KHz and array parameters: the number M =4 of ultrasonic receiving array elements, the sound velocity c =430M/s, the arc radius R =10cm from an ultrasonic transmitting signal source to the ultrasonic receiving array elements, and the included angle alpha between the signal source and two adjacent receiving array elements =20 degrees; discretizing the transmission signal; inputting a wind speed value in the range of 0-60 m/s and a wind direction value in the range of 0-359 degrees; and estimating the input wind parameter value by applying the trained FCN-MPL network model. The flow chart of the simulation experiment is shown in FIG. 6.
The experimental conditions are as follows: in order to verify the feasibility of the method, the arc radius is 10cm, the frequency of the transmitted ultrasonic signal is 40KHz, and the array element noise is additive white Gaussian noise. The wind speed scanning range is 0 m/s-60 m/s, and the step length is 0.1m/s. The wind direction angle scanning range is 0-359 degrees, and the step size is 1 degree. Fast beat number 320, selected SNR =10dB. The data in the data set is shown in table 1, the graph is an array received vector when the wind speed is 10m/s and the wind direction is 30 °, and fig. 7 is a real part visualization graph of the array received vector.
TABLE 1 wind speed 10m/s, wind direction 30 degree array reception vector X (t)
Wind speed and wind direction data sets constructed by an array receiving vector X (t) are respectively sent into the FCN-MPL wind speed and wind direction measuring network designed by the invention for network training, and the obtained training results are shown in tables 2 and 3, wherein Epoch in the tables is the network iteration times, train loss is the training set error, train accuracy is the training set accuracy, vallos is the verification set error, and Valaccuracyachy is the verification set accuracy.
TABLE 2 FCN-MLP wind direction measurement neural network model training results
TABLE 3 FCN-MLP anemometry neural network model training results
According to simulation experiments, the symmetric array structure provided by the application can realize the homodyne estimation in the range of V =0 m/s-60 m/s and theta = 0-359 degrees by applying the FCN-MLP algorithm. Therefore, the FCN-MLP-based arc-shaped array wind speed and direction measurement method is feasible.
3. Success rate experiment
In order to verify the feasibility of the method, the arc radius is 10cm, the frequency of the transmitted ultrasonic signal is 40KHz, the fast beat number is 320, and the noise of each array element is additive white Gaussian noise. And when the signal-to-noise ratio is SNR =10dB, the wind speed scanning range is 0 m/s-60 m/s, and the step length is 0.1m/s. The wind direction angle scanning range is 0-359 degrees, and the step length is 1 degree. Wind speed and wind direction are randomly selected from wind direction of 0-359 degrees and wind speed of 0-60 m/s, and 100 times of experiments are carried out.
Since the step length of the wind speed scanning is 0.1, when the absolute value of the difference between the given wind speed value and the wind speed value obtained through the experiment is less than 0.2, the experiment is considered to be successful; since the step length of the wind direction scanning is 1, when the absolute value of the difference between the given wind speed value and the wind speed value obtained through the experiment is less than 2, the experiment is considered to be successful, the simulation experiment result is shown in fig. 8, and the total success rate of the experiment is 99.877%.
4. Real time experiment
To verify the real-time performance of the proposed method, it is compared with the MUSIC (multiple signal classification) algorithm in array signal processing. The equipment specifications used are shown in table 4:
TABLE 4 Experimental Equipment Specifications
The radius of the selected arc is 10cm, the frequency of the transmitted ultrasonic signal is 40KHz, the fast beat number is 320, and the noise of each array element is additive white Gaussian noise. And when the signal-to-noise ratio is SNR =10dB, the wind speed scanning range is 0 m/s-60 m/s, and the step length is 0.1m/s. The wind direction angle scanning range is 0-359 degrees, and the step length is 1 degree. And comparing the time of measuring the wind speed and the wind direction of the first time of the two algorithms under the condition of the same signal-to-noise ratio and the same wind speed and wind direction:
TABLE 5 time of measuring primary wind speed and direction by two algorithms
According to simulation experiment results, the arc array wind measurement algorithm based on the FCN-MLP has better measurement real-time performance.
The invention provides an FCN-MLP-based arc ultrasonic sensing array wind parameter measurement method which comprises the following steps:
1. using an array output vector X (t) of the sensor array to construct a wind speed and wind direction data set, and preprocessing the data set;
2. and respectively constructing an FCN-MLP network for measuring the wind speed and the wind direction according to the wind speed and the wind direction. The method adopted by the invention is equivalent to the method that the sensor array detects the wind signal, the detection capability of the wind signal is improved, and then the wind signal is measured through the FCN-MLP network, so that the method is beneficial to reducing the calculation complexity on the premise of ensuring high measurement precision, greatly improves the calculation real-time property, and reduces the estimation time from 2 to 5 seconds to 2 milliseconds. The number of the arc-shaped receiving array elements in the array structure provided by the invention can be correspondingly increased or reduced, but the receiving array elements are ensured to be symmetrical and can form an arc-shaped structure, namely a symmetrical arc-shaped array structure. The FCN-MLP network algorithm provided by the invention can change the network layer number, the network hyper-parameter and the data input dimension, and can also be combined by adding other neural networks.
Claims (4)
1. The FCN-MLP-based arc ultrasonic sensing array wind parameter measurement method is characterized by comprising the following steps of:
s1, constructing an arc array structure: the arc array structure is composed of transmitting array elements and receiving array elements, the number of the receiving array elements is at least two and is a multiple of 2, and all the receiving array elements are uniformly arranged on an arc which takes the transmitting array elements as the circle center and R as the radius; the transmitting array element and the receiving array element are both ultrasonic sensors;
s2, according to the distribution condition of array elements in the constructed arc array structure, decomposing the wind signal to be tested by a vector decomposition method to obtain components in the connecting line direction of the transmitting array elements and the receiving array elements, selecting a certain receiving array element as a reference array element, obtaining the difference value between the time from the transmitting array element to each receiving array element and the time from the transmitting array element to the reference array element, and constructing an array flow pattern;
s3, establishing array receiving vectors under different signal-to-noise ratios, carrying out label processing on wind speed and wind direction, and constructing a wind speed data set and a wind direction data set;
s4, randomly disordering the wind speed data set, dividing the randomly disordering wind speed data set into a training set, a verification set and a test set, randomly disordering the wind direction data set, and dividing the randomly disordering wind direction data set into the training set, the verification set and the test set;
s5, building and training an FCN-MLP network model
Respectively constructing and training an FCN-MLP wind speed measurement neural network model and an FCN-MLP wind direction measurement neural network model, wherein the FCN-MLP wind speed measurement neural network model comprises a first FCN full convolution network, a first MLP full connection neural network and a flat layer connected between the first FCN full convolution network and the first MLP full connection neural network; the FCN-MLP wind direction measurement neural network model comprises a second FCN full convolution network, a second MLP full connection neural network and a flat layer connected between the second FCN full convolution network and the second MLP full connection neural network;
s6, respectively inputting the wind speed data set and the wind direction data set into the trained FCN-MLP wind speed measurement neural network model and the trained FCN-MLP wind direction measurement neural network model to obtain measurement wind parameter information;
the step S3 specifically includes: obtaining an array receiving vector matrix through an array flow pattern, adding Gaussian white noise with signal-to-noise ratios of 0dB, 5dB, 10dB, 15dB, 20dB and 30dB to obtain array receiving vectors with different signal-to-noise ratios, splicing a real part and an imaginary part of the array receiving vectors in a matrix column, adding wind speed and wind direction labels in the last column of the matrix, and respectively storing the wind speed and wind direction labels in a wind speed data set, a csv file and a wind direction data set, and constructing the wind speed data set and the wind direction data set.
2. The method of claim 1 for curved ultrasonic sensing array wind parameter measurement on FCN-MLP, wherein: when the wind speed data set and the wind direction data set are divided, the training set accounts for 80% of the total amount of the data set, the verification set accounts for 10% of the total amount of the data set, and the test set accounts for 10% of the total amount of the data set.
3. The method for measuring the wind parameters of the FCN-MLP arc ultrasonic sensing array according to claim 1, wherein the first FCN full convolution network is composed of five convolution layers, and the activation function of the convolution layers in the first FCN full convolution network adopts Relu function; the first MLP full-link neural network is composed of two full-link layers which are respectively named as: the activation functions of DENSE1, DENSE2 and DENSE1 adopt Relu functions, and DENSE1 outputs 256 neurons; the activation function of DENSE2 adopts a softmax function; the second FCN full convolution network is composed of four convolution layers, and a Relu function is adopted as a convolution layer activation function in the second FCN full convolution network; the second MLP full-connection neural network is composed of two full-connection layers which are named as follows: DENSE3, DENSE4 and DENSE3 adopt Relu function, DENSE3 outputs 128 neurons; the activation function of DENSE4 adopts a softmax function.
4. The method for measuring the wind parameters of the arc-shaped ultrasonic sensor array of the FCN-MLP as defined in claim 1, wherein in the process of training the model, a Cross Entropy loss function is adopted as a loss function, and an SGD random gradient descent method optimizer is selected by the optimizer.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210047540.XA CN114397474B (en) | 2022-01-17 | 2022-01-17 | FCN-MLP-based arc ultrasonic sensing array wind parameter measurement method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210047540.XA CN114397474B (en) | 2022-01-17 | 2022-01-17 | FCN-MLP-based arc ultrasonic sensing array wind parameter measurement method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114397474A CN114397474A (en) | 2022-04-26 |
| CN114397474B true CN114397474B (en) | 2022-11-08 |
Family
ID=81231468
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210047540.XA Active CN114397474B (en) | 2022-01-17 | 2022-01-17 | FCN-MLP-based arc ultrasonic sensing array wind parameter measurement method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114397474B (en) |
Citations (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP2667022A2 (en) * | 2012-05-22 | 2013-11-27 | Siemens Aktiengesellschaft | Method and system for wind velocity field measurements on a wind farm |
| CN104155471A (en) * | 2014-07-10 | 2014-11-19 | 天津大学 | Multiphase flow testing method based on cross-correlation velocity measurement of ultrasonic and electric multiple sensors |
| CN107576819A (en) * | 2017-08-29 | 2018-01-12 | 吉林大学 | A kind of method and system for measuring wind speed and direction |
| CN209624628U (en) * | 2018-10-25 | 2019-11-12 | 深圳慈航无人智能系统技术有限公司 | A kind of unmanned plane being measured wind direction and wind velocity based on neural network |
| CN110987066A (en) * | 2019-11-26 | 2020-04-10 | 青岛科技大学 | Ocean wind speed and direction measuring method and system capable of achieving automatic correction |
| WO2020085312A1 (en) * | 2018-10-22 | 2020-04-30 | 株式会社Ｌｉｌｙ ＭｅｄＴｅｃｈ | Ultrasonic diagnostic system |
| CN112215487A (en) * | 2020-10-10 | 2021-01-12 | 吉林大学 | Vehicle driving risk prediction method based on neural network model |
| CN112255429A (en) * | 2020-10-20 | 2021-01-22 | 吉林大学 | Three-dimensional wind parameter measuring method and system |
| WO2021079640A1 (en) * | 2019-10-23 | 2021-04-29 | ソニー株式会社 | Information processing device, information processing method, and artificial intelligence system |
| CN112785052A (en) * | 2021-01-14 | 2021-05-11 | 吉林大学 | Wind speed and wind direction prediction method based on particle filter algorithm |
| CN113063961A (en) * | 2020-01-02 | 2021-07-02 | 吉林大学 | Ultrasonic sensing array wind measuring device and method thereof |
| CN113537573A (en) * | 2021-06-22 | 2021-10-22 | 中国农业大学 | Wind power operation trend prediction method based on dual space-time feature extraction |
| WO2022005619A2 (en) * | 2020-05-18 | 2022-01-06 | The Regents Of The University Of Colorado, A Body Corporate | Ocean surface wind direction retrieval from reflected radio signals on space-borne platforms |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US11015902B2 (en) * | 2013-05-09 | 2021-05-25 | Shooting Simulator, Llc | System and method for marksmanship training |
| CN113311397B (en) * | 2021-05-25 | 2023-03-10 | 西安电子科技大学 | Large array rapid self-adaptive anti-interference method based on convolutional neural network |
-
2022
- 2022-01-17 CN CN202210047540.XA patent/CN114397474B/en active Active
Patent Citations (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP2667022A2 (en) * | 2012-05-22 | 2013-11-27 | Siemens Aktiengesellschaft | Method and system for wind velocity field measurements on a wind farm |
| CN104155471A (en) * | 2014-07-10 | 2014-11-19 | 天津大学 | Multiphase flow testing method based on cross-correlation velocity measurement of ultrasonic and electric multiple sensors |
| CN107576819A (en) * | 2017-08-29 | 2018-01-12 | 吉林大学 | A kind of method and system for measuring wind speed and direction |
| WO2020085312A1 (en) * | 2018-10-22 | 2020-04-30 | 株式会社Ｌｉｌｙ ＭｅｄＴｅｃｈ | Ultrasonic diagnostic system |
| CN209624628U (en) * | 2018-10-25 | 2019-11-12 | 深圳慈航无人智能系统技术有限公司 | A kind of unmanned plane being measured wind direction and wind velocity based on neural network |
| WO2021079640A1 (en) * | 2019-10-23 | 2021-04-29 | ソニー株式会社 | Information processing device, information processing method, and artificial intelligence system |
| CN110987066A (en) * | 2019-11-26 | 2020-04-10 | 青岛科技大学 | Ocean wind speed and direction measuring method and system capable of achieving automatic correction |
| CN113063961A (en) * | 2020-01-02 | 2021-07-02 | 吉林大学 | Ultrasonic sensing array wind measuring device and method thereof |
| WO2022005619A2 (en) * | 2020-05-18 | 2022-01-06 | The Regents Of The University Of Colorado, A Body Corporate | Ocean surface wind direction retrieval from reflected radio signals on space-borne platforms |
| CN112215487A (en) * | 2020-10-10 | 2021-01-12 | 吉林大学 | Vehicle driving risk prediction method based on neural network model |
| CN112255429A (en) * | 2020-10-20 | 2021-01-22 | 吉林大学 | Three-dimensional wind parameter measuring method and system |
| CN112785052A (en) * | 2021-01-14 | 2021-05-11 | 吉林大学 | Wind speed and wind direction prediction method based on particle filter algorithm |
| CN113537573A (en) * | 2021-06-22 | 2021-10-22 | 中国农业大学 | Wind power operation trend prediction method based on dual space-time feature extraction |
Non-Patent Citations (1)
| Title |
|---|
| MIMO超声波传感器阵列风速风向测量方法;李新波 等;《哈尔滨工程大学学报》;20211231;第42卷(第11期);1694-1700 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114397474A (en) | 2022-04-26 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110365612A (en) | A kind of deep learning Beam Domain channel estimation methods based on approximate Message Passing Algorithm | |
| CN111639747B (en) | GNSS-R sea surface wind speed inversion method and system based on BP neural network | |
| CN104616078B (en) | Photovoltaic system electricity generation power Forecasting Methodology based on Spiking neutral nets | |
| CN111639746B (en) | GNSS-R sea surface wind speed inversion method and system based on CNN neural network | |
| CN110443724B (en) | Electric power system rapid state estimation method based on deep learning | |
| CN110888105B (en) | DOA estimation method based on convolutional neural network and received signal strength | |
| CN114048930B (en) | Ultra-short-term wind power prediction method and device | |
| CN112180369B (en) | Depth learning-based sea surface wind speed inversion method for one-dimensional synthetic aperture radiometer | |
| CN112468230A (en) | Wireless ultraviolet light scattering channel estimation method based on deep learning | |
| CN114492217B (en) | Typhoon and wave height prediction method based on mixed time series perceptron model | |
| CN115204058A (en) | Ground wave radar flow field calculation method and device based on BP neural network | |
| CN114397474B (en) | FCN-MLP-based arc ultrasonic sensing array wind parameter measurement method | |
| Wu et al. | A modular tide level prediction method based on a NARX neural network | |
| CN114880843A (en) | Grid-connected inverter sequence impedance calculation method based on neural network | |
| CN115977874B (en) | Wind turbine generator yaw self-adaptive calibration method and system based on laser wind-finding radar | |
| CN104361399A (en) | Solar irradiation intensity minute-scale predication method | |
| CN116632823A (en) | Short-term photovoltaic power prediction method based on power conversion model and multi-layer perceptron | |
| CN116305967A (en) | Sea surface wind speed inversion method based on convolution neural network and multi-modal feature fusion | |
| CN111859241B (en) | Unsupervised sound source orientation method based on sound transfer function learning | |
| CN114047474A (en) | Generalized regression neural network-based uniform linear array target orientation estimation method | |
| CN113673786A (en) | Effective wind speed estimation system and method of wind driven generator based on RBF and LSTM | |
| Dotche et al. | Evaluating Solar Energy Harvesting using Artificial Neural Networks: A Case study in Togo | |
| CN113063961A (en) | Ultrasonic sensing array wind measuring device and method thereof | |
| CN117872424B (en) | GNSS-R sea surface DDM data generation method and device | |
| Crespo et al. | LSTM for Grid Power Forecasting in Short-Term from Wave Energy Converters |
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 |