CN113268927A - High-power laser device output energy prediction method based on full-connection neural network - Google Patents
High-power laser device output energy prediction method based on full-connection neural network Download PDFInfo
- Publication number
- CN113268927A CN113268927A CN202110559599.2A CN202110559599A CN113268927A CN 113268927 A CN113268927 A CN 113268927A CN 202110559599 A CN202110559599 A CN 202110559599A CN 113268927 A CN113268927 A CN 113268927A
- Authority
- CN
- China
- Prior art keywords
- neural network
- data set
- data
- output
- output energy
- 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.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 62
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 29
- 238000012549 training Methods 0.000 claims abstract description 51
- 238000003062 neural network model Methods 0.000 claims abstract description 47
- 238000005259 measurement Methods 0.000 claims abstract description 38
- 239000013598 vector Substances 0.000 claims abstract description 28
- 238000012360 testing method Methods 0.000 claims abstract description 16
- 230000006870 function Effects 0.000 claims description 42
- 230000003287 optical effect Effects 0.000 claims description 22
- 230000002159 abnormal effect Effects 0.000 claims description 17
- 238000002474 experimental method Methods 0.000 claims description 14
- 238000011156 evaluation Methods 0.000 claims description 13
- 230000003321 amplification Effects 0.000 claims description 9
- 238000003199 nucleic acid amplification method Methods 0.000 claims description 9
- 238000004364 calculation method Methods 0.000 claims description 8
- 230000010354 integration Effects 0.000 claims description 8
- 238000012544 monitoring process Methods 0.000 claims description 8
- 230000008569 process Effects 0.000 claims description 6
- 238000012795 verification Methods 0.000 claims description 5
- 238000012545 processing Methods 0.000 claims description 4
- 238000010606 normalization Methods 0.000 claims description 3
- 238000005516 engineering process Methods 0.000 abstract description 4
- 238000013473 artificial intelligence Methods 0.000 abstract description 2
- 238000004088 simulation Methods 0.000 abstract description 2
- 239000000284 extract Substances 0.000 abstract 1
- 230000000694 effects Effects 0.000 description 8
- 230000005540 biological transmission Effects 0.000 description 4
- 238000010586 diagram Methods 0.000 description 4
- 238000005457 optimization Methods 0.000 description 3
- 238000012546 transfer Methods 0.000 description 3
- 230000004913 activation Effects 0.000 description 2
- 238000002790 cross-validation Methods 0.000 description 2
- 230000008030 elimination Effects 0.000 description 2
- 238000003379 elimination reaction Methods 0.000 description 2
- 239000011159 matrix material Substances 0.000 description 2
- 230000005856 abnormality Effects 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 230000008451 emotion Effects 0.000 description 1
- 238000012886 linear function Methods 0.000 description 1
- 230000007774 longterm Effects 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 230000036632 reaction speed Effects 0.000 description 1
- 238000013468 resource allocation Methods 0.000 description 1
- 238000010200 validation analysis Methods 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- 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
-
- 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
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Software Systems (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Mathematical Physics (AREA)
- Computational Linguistics (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Medical Informatics (AREA)
- Computer Hardware Design (AREA)
- Geometry (AREA)
- Lasers (AREA)
Abstract
A high-power laser device output energy prediction method based on a fully-connected neural network relates to an application technology of an artificial intelligence technology in a high-power laser device, and aims to solve the problem that the existing physical model simulation cannot accurately predict the light path output energy. The invention extracts the measurement data and the configuration data thereof; combining input energy and corresponding configuration data into a vector, taking the vector as an input data set, taking output energy as an output data set, and establishing an input and output data set; removing the data set, and dividing the data set into a training data set and a testing data set according to a proportion; training the training data set by using a fully-connected neural network model to obtain a trained neural network model; and integrating the test data set, and inputting the test data set into the optimal solution neural network model to obtain the predicted output energy. The method has the advantage that the prediction accuracy is improved by 10%.
Description
Technical Field
The invention relates to an application technology of an artificial intelligence technology in a high-power laser device.
Background
The high-power laser device belongs to huge and complex large scientific engineering, comprises dozens of light paths, and requires the output energy and power of each light path to be balanced. To meet this requirement, the structure and parameters of each optical path are strictly designed. But because each optical path includes hundreds of optical and electrical components; the components inevitably have small differences in the processes of production, processing and assembly and calibration, and the addition of the differences can cause the output performance of each optical path to be different from the design value and have certain deviation. In addition, the working state of the high-power laser device is only one single pulse with the duration of several picoseconds, the working time is extremely short, the cost is extremely high, and the high-power laser device cannot be measured and adjusted in a real-time monitoring mode. Therefore, the actual output energy of each optical path needs to be accurately predicted, and the requirement of a physical experiment on the balance of the output energy of each optical path is met by means of adjusting the input energy.
For engineers who work in the physical experiment for a long time, the long-term accumulated experience makes them have certain acuity to the fluctuation of the input and output of energy, and the energy output of each optical path of the next experiment can be intuitively judged through the measurement results of the previous experiments. However, the prediction method has the problem of inaccuracy and is difficult to guarantee real-time performance. Human visual prediction has no theoretical basis, the influence of subjective emotion is large, the prediction trend is probably close to the real trend, but the deviation of the prediction value and the real value is large. Meanwhile, dozens of light paths need to predict the output energy and adjust the input energy, and because the calculation and reaction speed of people is limited, the task is difficult to be efficiently completed by engineering personnel, and an intelligent automatic control platform needs to be built to provide decision support for parameter configuration, plan adjustment and resource allocation of the device.
The process of laser transmission amplification can be simulated according to a physical model, but the numerical value calculated by the method belongs to a theoretical value. During the actual operation of the device, there is a deviation between the measured data and the theoretical values, which the theoretical model cannot predict. The existence of deviation can lead the output energy of each optical path not to meet the balance requirement of physical experiments, so that the experimental effect is not ideal, and even the experiment fails. Because the device plays an amplifying role in light beam energy, the energy at the output end of the light path can reach thousands of joules, and a small deviation proportion can also cause a large energy error value, a method capable of accurately predicting the output energy of each light path needs to be designed, and the method is particularly important.
Disclosure of Invention
The invention aims to solve the problem that the existing physical model simulation can not accurately predict the output energy of an optical path, and provides a method for predicting the output energy of a high-power laser device based on a fully-connected neural network.
The invention relates to a high-power laser device output energy prediction method based on a fully-connected neural network, which comprises the following steps:
step one, setting monitoring points at the appointed positions of all light paths of a laser device, measuring the input energy and the output energy of all the monitoring points of all the light paths, extracting the input energy and the output energy of each operation of the appointed light paths as measurement data, and obtaining configuration data;
combining the input energy in the measurement data and the corresponding configuration data into a vector, taking the vector as input data, taking the output energy in the measurement data as output data, and establishing an input-output data set;
removing abnormal data points existing in the input and output data sets, and dividing the input and output data sets into a training data set and a testing data set according to a proportion;
step four, training the training data set by using a plurality of fully-connected neural network models simultaneously, and obtaining the trained neural network models after multiple times of training;
integrating the trained neural network models to obtain an optimal solution neural network model;
and step six, inputting the test data set into the optimal solution neural network model to obtain the predicted output energy.
Further, the configuration data in the first step comprises an optical path number, a time number, amplification piece configuration and a pulse width, and the measurement data in the first step needs to be normalized in the corresponding characteristic dimension;
the specific method of the normalization processing is as follows: the number of optical paths, the time number, the arrangement of amplification pieces, and the pulse width of each operation of the laser device were scaled to values between 0 and 1, and the scaling factor was recorded.
Further, the optical path designation in the step one includes two modes, where the two modes are: specifying a beam mode and a beam group mode;
the specified light beam mode is to establish a network prediction model for dozens of light paths of the device;
the method for specifying the beam group refers to the fact that a network prediction model is established for a plurality of light paths sharing one set of hardware equipment in the laser transmission process of the device together, and the network prediction model is used for predicting the overall trend of the specified beam group and detecting the abnormal condition of a single light path.
Further, the designated position in the step one is the joint of each subsystem in the laser device, and the designated position comprises three measuring points;
each subsystem comprises a pre-discharge system, a main discharge system and a target range system;
the three measuring points constitute two working sections.
Further, the specific method for establishing the input/output data set in the second step is as follows:
when the input energy in the measurement data and the corresponding configuration data are combined into a vector, the input energy measurement data and the configuration data with the same running time are combined into a column vector which is a sample; transversely arranging a plurality of samples, namely column vectors of a plurality of running times to form a two-dimensional vector of the plurality of samples and form input data; and then, the output energy measurement data corresponding to the running time are transversely arranged to form a one-dimensional vector of multiple samples, and output data are formed.
Furthermore, the three steps include two ways of eliminating abnormal data points existing in the input/output data set;
the first method is that a miniaturized network is quickly pre-trained to obtain an approximate relation between input and output energies as a reference, and an input and output data set is removed from data points far away from the reference relation;
and secondly, according to configuration parameters before experiments, comparing expected output energy obtained by theoretical calculation with actual output energy obtained by experimental measurement, calculating deviation percentage, and removing data points with large deviation percentage from an input and output data set.
Further, the training data set in the third step is experimental measurement data and configuration data generated in an early stage; the test data set in step three is the data generated from the most recent experiment.
Furthermore, the structures of a plurality of fully-connected neural network models in the fourth step are the same;
the specific training process of simultaneously training the training data set by the fully-connected neural network models is as follows: meanwhile, carrying out iterative training on a plurality of fully-connected neural network models by using a training data set until the condition of stopping iteration is met, and finishing the training;
the iteration stopping conditions are divided into two types, one type is a loss function obtained by calculation after a training data set is substituted into a neural network model, and the loss function can also be called as an evaluation function, and the numerical value is small enough to meet the preset conditions; and the other is that after the verification data set is substituted into the neural network model, the calculated continuous increase times of the loss function reach the preset threshold condition.
Further, the specific method for integrating the trained neural network models in the step five is as follows: defining a weight value of the loss function calculated by each neural network model, wherein the weight value is the ratio of the reciprocal of the loss function of the sub-network to the reciprocal sum of the loss functions of all the sub-networks; and weighting the prediction results of the sub-network models to obtain a relatively accurate output energy prediction result, namely completing integration.
The invention has the advantages that the deep learning full-connection neural network is adopted, the network model is trained through the historical measurement data sample, the input and output gain relation of the main amplifier part in the high-power laser device is regarded as a transfer function, the network is used for approximating the target transfer function, the transfer function is changed along with the multidimensional independent variable, the output energy is predicted to obtain higher accuracy under different operation time and different operation configuration conditions, and the prediction accuracy of the output energy of the high-power laser device is improved by 10%; meanwhile, the abnormal point elimination and integrated learning method is introduced, so that the influence of failed or non-ideal experimental data on the model training process is eliminated, the unstable single neural network prediction effect with uncertain convergence is stable and accurate, and the convergence is ensured; in addition, the invention adopts a multi-input single-output network model, and considers the influence of various configuration parameters on the gain performance, including the number of amplifier chips, the running time and the like; when the state of the device changes, a plurality of models are required to be established in the conventional prediction method so as to accurately predict the output energy, but by adopting the method disclosed by the invention, the prediction of the output energy can be completed only by establishing one model.
Drawings
Fig. 1 is a flowchart of a method for predicting output energy of a high-power laser device based on a fully-connected neural network according to a first embodiment;
FIG. 2 is a schematic structural diagram of a fully-connected neural network model according to an embodiment;
FIG. 3 is a graph of the predicted effect of the output energy of 10 neural network models before integration in step five of the detailed description;
FIG. 4 is a diagram illustrating the predicted effect of the output energy of the neural network model after integration in step five of the first embodiment;
FIG. 5 is a graph of relative predicted deviations before and after optimization in a first embodiment;
FIG. 6 is a statistical histogram of predicted relative deviations before and after optimization in a first embodiment.
Detailed Description
The first embodiment is as follows: the present embodiment is described with reference to fig. 1 to 6, and the method for predicting the output energy of the high-power laser device based on the fully-connected neural network according to the present embodiment includes the following steps:
step one, setting monitoring points at the appointed positions of all light paths of a laser device, measuring the input energy and the output energy of all the monitoring points of all the light paths, extracting the input energy and the output energy of each operation of the appointed light paths as measurement data, and obtaining configuration data; the measurement data is stored in an excel table, and the relevant measurement data is required to be extracted according to the selected optical path;
combining the input energy in the measurement data and the corresponding configuration data into a vector, taking the vector as input data, taking the output energy in the measurement data as output data, and establishing an input-output data set; combining the measured input energy, the light path number (including two elements of a beam group and a light path), the time number, the number of the intracavity amplification pieces, the number of the boosting amplification pieces and the pulse width into a 7 x n vector, namely 7 characteristic points and n samples; taking the measured output energy as an output vector with the size of 1 x n, wherein n is the number of samples;
removing abnormal data points existing in the input and output data sets, and dividing the input and output data sets into a training data set and a testing data set according to a proportion; abnormal data points exist in the data set, and may include some failed running times and experiments of abnormal components. There are two methods for rejecting outliers, a method for rejecting by using a neural network and a method for rejecting by using prior knowledge; the method of utilizing the neural network needs to establish a small network, and train the small network by using a data set without rejecting abnormal points at present; after training is finished, testing the network effect by using the data set, comparing the network output result with the output data in the data set, and if the point with larger difference is considered as an abnormal point, rejecting the point; the elimination method utilizing priori knowledge is required to extract expected data in annual operation data sets, calculate the relative deviation between the output energy of a main amplifier expected to be obtained by the operation of a specified optical path each time and the output energy of the main amplifier obtained by actual measurement, set a threshold value to be 15%, and if the relative deviation calculated by the data operated at a certain time exceeds the threshold value, the data is considered as abnormal data and needs to be eliminated; arranging the data sets with the abnormal data points removed in a time sequence, and arranging the data sets according to the following steps of 8: 1: the proportion of 1 is divided into a training set, a cross validation set and a test set;
step four, training the training data set by using a plurality of fully-connected neural network models in sequence, and obtaining the trained neural network models after training for a plurality of times; wherein, 10 identical full-connection networks are established, three layers are provided, two hidden layers and one output layer, the number of nodes of the hidden layers is 20 and 40 respectively, the activation function of the hidden layers is a hyperbolic tangent function, the activation function of the output layer is a linear function, as shown in fig. 2, and then the 10 models are trained respectively. The training of the model is to establish a complex function by learning the relation between input energy and output energy from the existing training set data; training the model by using the training data set without the abnormal points, observing the result and carrying out iterative optimization until the result meets the termination condition when the model is verified by using the cross verification set data; the result is characterized by an evaluation function, wherein the evaluation function is a mean square error function and represents the similarity degree of the predicted output and the real output, and the smaller the value is, the predicted value and the true value are representedThe closer together; the termination condition comprises that the iteration times reach a preset upper limit, the evaluation function value is smaller than a preset threshold value, and the continuous increase times of the evaluation function value of the cross validation set validation result reach the preset upper limit; the upper limit of the iteration times is 3000, and the threshold value of the evaluation function is 10-5The upper limit of the number of times of increase of the evaluation function of the verification set is 20;
integrating the trained neural network models to obtain an optimal solution neural network model; the integration of 10 networks is a weighted average method, and the weight value depends on the evaluation function value; the higher the evaluation function value is, the worse the prediction effect is, so that the weight is equal to the reciprocal of the evaluation function value, the contribution of the model with accurate prediction to the final result is larger, the contribution of the inaccurate model to the integrated result is smaller, and the more accurate final prediction result is obtained. Multiplying the output of each network model by the respective weight, and dividing by the sum of the weights to obtain an integration result of weighted average, which is an output energy prediction effect diagram of 10 neural network models before integration as shown in fig. 3; as shown in fig. 4, it is an output energy prediction effect diagram of the integrated 10 neural network models;
and step six, inputting the test data set into the optimal solution neural network model to obtain the predicted output energy. If the output energy of the main amplifier in the optical path at the next operation needs to be predicted, combining the preset output energy and the configuration parameters into a 7 x 1 vector in a data set form, and inputting the vector into a trained optimal solution network model to obtain the predicted output energy of the main amplifier at an output layer;
in the embodiment, the energy measurement data and the corresponding configuration data are obtained from the operation data of the high-power laser device, monitoring points are set at key positions of each optical path, the energy of each key position is measured, and the optical paths are stored according to the time sequence; for the selected light path to be predicted, extracting measurement data and configuration data of each operation as input characteristics of the data, and combining the input characteristics into a matrix form for use; removing abnormal data in the data by using two methods, namely a method using a neural network and a method using priori knowledge; because a single network may have the problems of non-convergence, under-fitting and over-fitting, a plurality of networks are adopted for respective training, and the results of all the networks are integrated by a certain method to obtain a more accurate prediction result; and combining the prediction results of the plurality of networks by using a weighted average mode, wherein the weight is the reciprocal of the evaluation function, and the prediction result is accurate after the weighted average is calculated.
The second embodiment is as follows: in this embodiment, the configuration data in the first step includes a light path number, a time number, an amplification piece configuration and a pulse width, and the measurement data in the first step needs to be normalized in the corresponding characteristic dimension;
the specific method of the normalization processing is as follows: the number of optical paths, the time number, the arrangement of amplification pieces, and the pulse width of each operation of the laser device were scaled to values between 0 and 1, and the scaling factor was recorded.
In the present embodiment, the neural network performs learning by scaling adjustment; the scaling factor is recorded for subsequent restore operations.
The third concrete implementation mode: in this embodiment, the method for predicting the output energy of the high-power laser device based on the fully-connected neural network according to the first embodiment is further limited, and in this embodiment, the specified optical path in the first step includes two modes: specifying a beam mode and a beam group mode;
the specified light beam mode is to establish a network prediction model for dozens of light paths of the device;
the beam group specifying mode refers to that a network prediction model is established for a plurality of light paths sharing one set of hardware equipment in the laser transmission process of the device.
In the embodiment, the prediction result of the specified light beam mode is more accurate; in the mode of specifying the beam group, the beam group is a set of adjacent beams and comprises a plurality of light paths, and a set of hardware equipment is shared in the laser transmission process of the device, so that the trends are similar.
The fourth concrete implementation mode: in this embodiment, the method for predicting the output energy of the high-power laser device based on the fully-connected neural network is further defined as the first embodiment, in this embodiment, the designated position in the first step is a connection point of each subsystem in the laser device, and the designated position includes three measurement points;
each subsystem comprises a pre-discharge system, a main discharge system and a target range system;
the three measuring points constitute two working sections.
In the embodiment, due to different working mechanisms, the input and output energy of a single working section is usually calculated and predicted, so that analysis and decision support can be provided for the abnormality detection of the device; the calculation and prediction of the total input and the total output of the two working segments are not excluded, but the complexity of the neural network model is increased and the prediction accuracy is reduced.
The fifth concrete implementation mode: in this embodiment, the method for predicting the output energy of the high-power laser device based on the fully-connected neural network according to the first embodiment is further defined, and in this embodiment, the specific method for establishing the input/output data set in the second step is as follows:
when the input energy in the measurement data and the corresponding configuration data are combined into a vector, the output energy measurement data and the configuration data at the same running time are combined into a column vector which is a sample; transversely arranging a plurality of samples, namely column vectors of a plurality of running times to form a two-dimensional vector of the plurality of samples and form input data; and then, the output energy measurement data corresponding to the running time are transversely arranged to form a one-dimensional vector of multiple samples, and output data are formed.
In this embodiment, the two-dimensional vector is also called a matrix.
The sixth specific implementation mode: in the present embodiment, the method for predicting the output energy of the high-power laser device based on the fully-connected neural network according to the first embodiment is further limited, and in the present embodiment, two ways of eliminating abnormal data points existing in the input/output data set in the third step are included;
the first method is that a miniaturized network is quickly pre-trained to obtain an approximate relation between input and output energies as a reference, and an input and output data set is removed from data points far away from the relation;
and secondly, according to configuration parameters before experiments, comparing expected output energy obtained by theoretical calculation with actual output energy obtained by experimental measurement, calculating deviation percentage, and removing data points with large deviation percentage from an input and output data set.
The seventh embodiment: in this embodiment, the method for predicting the output energy of the high-power laser device based on the fully-connected neural network is further defined as the first embodiment, and in this embodiment, the training data set in the third step is experimental measurement data and configuration data generated in an early stage; the test data set in step three is the data generated from the most recent experiment.
In the embodiment, the input data set from which the outliers are removed and the corresponding output data set are arranged in time sequence and divided into training data sets according to a certain proportion, and the verification data set and the test data set have a large proportion of the training data set; the training data set is experimental measurement data and configuration data generated in an early stage, and the testing data set is data generated in a latest experiment.
The specific implementation mode is eight: in this embodiment, the method for predicting the output energy of the high-power laser device based on the fully-connected neural network according to the first embodiment is further defined, and in this embodiment, the plurality of fully-connected neural network models in the fourth step have the same structure;
the specific training process of simultaneously training the training data set by the fully-connected neural network models is as follows: meanwhile, carrying out iterative training on a plurality of fully-connected neural network models by using a training data set until the condition of stopping iteration is met, and finishing the training;
the iteration stopping conditions are divided into two types, one type is a loss function obtained by calculation after a training data set is substituted into a neural network model, and the loss function can also be called as an evaluation function, and the numerical value is small enough to meet the preset conditions; and the other is that after the training data set is verified to be substituted into the neural network model, the calculated continuous increasing times of the loss function meet the preset condition.
The specific implementation method nine: in this embodiment, the method for predicting the output energy of the high-power laser device based on the fully-connected neural network according to the first embodiment is further defined, and in the fifth embodiment, the method for integrating the trained neural network models includes: defining a weight value of the loss function calculated by each neural network model, wherein the weight value is the ratio of the derivative of the loss function of the sub-network to the sum of the derivatives of the loss functions of all the sub-networks; and weighting the prediction results of the sub-network models to obtain a relatively accurate output energy prediction result, namely completing integration.
In the embodiment, due to the influence of parameter random initialization and local minimum value problem, the fitting capability of the trained neural network model to the data set is uneven; when the sub-networks are integrated, the results of the sub-networks cannot be weighted averagely directly, and the weight of the sub-networks is defined by using the loss function calculated by each neural network model, wherein the weight is the ratio of the derivative of the loss function of the sub-network to the sum of the derivatives of the loss functions of all the sub-networks; and weighting the prediction results of each sub-network model to obtain a relatively accurate output energy prediction result.
Claims (9)
1. The output energy prediction method of the high-power laser device based on the fully-connected neural network is characterized by comprising the following steps of:
step one, setting monitoring points at the appointed positions of all light paths of a laser device, measuring the input energy and the output energy of all the monitoring points of all the light paths, extracting the input energy and the output energy of each operation of the appointed light paths as measurement data, and obtaining configuration data;
combining the input energy in the measurement data and the corresponding configuration data into a vector, taking the vector as input data, taking the output energy in the measurement data as output data, and establishing an input-output data set;
removing abnormal data points existing in the input and output data sets, and dividing the input and output data sets into a training data set and a testing data set according to a proportion;
step four, training the training data set by using a plurality of fully-connected neural network models in sequence, and obtaining the trained neural network models after training for a plurality of times;
integrating the trained neural network models to obtain an optimal solution neural network model;
and step six, inputting the test data set into the optimal solution neural network model to obtain the predicted output energy.
2. The output energy prediction method of the high-power laser device based on the fully-connected neural network as claimed in claim 1, wherein the configuration data in the first step includes a light path number, a time number, an amplification sheet configuration and a pulse width, and the measurement data in the first step needs to be normalized in the corresponding characteristic dimension;
the specific method of the normalization processing is as follows: the number of optical paths, the time number, the arrangement of amplification pieces, and the pulse width of each operation of the laser device were scaled to values between 0 and 1, and the scaling factor was recorded.
3. The output energy prediction method of the high-power laser device based on the fully-connected neural network as claimed in claim 1, wherein the specified optical path in the first step includes two modes, the two modes are: specifying a beam mode and a beam group mode;
the specified light beam mode is to establish a network prediction model for dozens of light paths of the device;
the method is used for predicting the overall trend of the specified beam group and detecting the abnormal condition of a single optical path.
4. The method for predicting the output energy of the high-power laser device based on the fully-connected neural network as claimed in claim 1, wherein the designated positions in the first step are the joints of the subsystems in the laser device, and the designated positions comprise three measurement points;
each subsystem comprises a pre-discharge system, a main discharge system and a target range system;
the three measuring points constitute two working sections.
5. The output energy prediction method of the high-power laser device based on the fully-connected neural network as claimed in claim 1, wherein the specific method for establishing the input and output data set in the second step is as follows:
when the input energy in the measurement data and the corresponding configuration data are combined into a vector, the input energy measurement data and the configuration data with the same running time are combined into a column vector which is a sample; transversely arranging a plurality of samples, namely column vectors of a plurality of running times to form a two-dimensional vector of the plurality of samples and form input data; and then, the output energy measurement data corresponding to the running time are transversely arranged to form a one-dimensional vector of multiple samples, and output data are formed.
6. The method for predicting the output energy of the high-power laser device based on the fully-connected neural network as claimed in claim 1, wherein the three steps include two ways of eliminating abnormal data points existing in the input and output data sets;
the first method is that a miniaturized network is quickly pre-trained to obtain an approximate relation between input and output energies as a reference, and an input and output data set is removed from data points far away from the reference relation;
and secondly, according to configuration parameters before an experiment, comparing expected output energy obtained by theoretical calculation with actual output energy obtained by experimental measurement, calculating the deviation percentage, and removing data points with large deviation percentage from an input and output data set.
7. The method for predicting the output energy of the high-power laser device based on the fully-connected neural network as claimed in claim 1, wherein the training data set in the third step is experimental measurement data and configuration data generated in an early stage; the test data set in step three is the data generated from the most recent experiment.
8. The method for predicting the output energy of the high-power laser device based on the fully-connected neural network as claimed in claim 1, wherein the plurality of fully-connected neural network models in the fourth step have the same structure;
the specific training process of simultaneously training the training data set by the fully-connected neural network models is as follows: meanwhile, carrying out iterative training on a plurality of fully-connected neural network models by using a training data set until the condition of stopping iteration is met, and finishing the training;
the iteration stopping conditions are divided into two types, one type is a loss function obtained by calculation after a training data set is substituted into a neural network model, and the loss function can also be called as an evaluation function, and the numerical value is small enough to meet the preset conditions; and the other is that after the verification data set is substituted into the neural network model, the calculated continuous increase times of the loss function reach the preset threshold condition.
9. The output energy prediction method of the high-power laser device based on the fully-connected neural network as claimed in claim 1, wherein the concrete method for integrating the trained neural network models in the fifth step is as follows: defining a weight value of the loss function calculated by each neural network model, wherein the weight value is the ratio of the reciprocal of the loss function of the sub-network to the reciprocal sum of the loss functions of all the sub-networks; and weighting the prediction results of the sub-network models to obtain a relatively accurate output energy prediction result, namely completing integration.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110559599.2A CN113268927B (en) | 2021-05-21 | 2021-05-21 | Output energy prediction method of high-power laser device based on fully-connected neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110559599.2A CN113268927B (en) | 2021-05-21 | 2021-05-21 | Output energy prediction method of high-power laser device based on fully-connected neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113268927A true CN113268927A (en) | 2021-08-17 |
| CN113268927B CN113268927B (en) | 2024-04-30 |
Family
ID=77232343
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110559599.2A Active CN113268927B (en) | 2021-05-21 | 2021-05-21 | Output energy prediction method of high-power laser device based on fully-connected neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113268927B (en) |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20170091615A1 (en) * | 2015-09-28 | 2017-03-30 | Siemens Aktiengesellschaft | System and method for predicting power plant operational parameters utilizing artificial neural network deep learning methodologies |
| CN109034388A (en) * | 2018-07-27 | 2018-12-18 | 湖北中烟工业有限责任公司 | A kind of prediction model of cigarette material and mainstream smoke constituents based on Genetic Algorithm Optimized Neural Network |
| CN109242236A (en) * | 2018-07-05 | 2019-01-18 | 西安图迹信息科技有限公司 | Based on the PMU primary frequency modulation load forecasting method for being associated with full Connection Neural Network and LSTM |
-
2021
- 2021-05-21 CN CN202110559599.2A patent/CN113268927B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20170091615A1 (en) * | 2015-09-28 | 2017-03-30 | Siemens Aktiengesellschaft | System and method for predicting power plant operational parameters utilizing artificial neural network deep learning methodologies |
| CN109242236A (en) * | 2018-07-05 | 2019-01-18 | 西安图迹信息科技有限公司 | Based on the PMU primary frequency modulation load forecasting method for being associated with full Connection Neural Network and LSTM |
| CN109034388A (en) * | 2018-07-27 | 2018-12-18 | 湖北中烟工业有限责任公司 | A kind of prediction model of cigarette material and mainstream smoke constituents based on Genetic Algorithm Optimized Neural Network |
Non-Patent Citations (1)
| Title |
|---|
| 梁斌梅;韦琳娜;: "改进的径向基函数神经网络预测模型", 计算机仿真, no. 11, 15 November 2009 (2009-11-15) * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113268927B (en) | 2024-04-30 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108038300B (en) | Optical fiber state evaluation method based on improved membership function combined with neural network | |
| CN111222549B (en) | Unmanned aerial vehicle fault prediction method based on deep neural network | |
| CN109214708B (en) | Electric power system risk assessment method based on cross entropy theory optimization support vector machine | |
| CN106934242B (en) | The health degree appraisal procedure and system of equipment under multi-mode based on Cross-Entropy Method | |
| CN110543932A (en) | air conditioner performance prediction method and device based on neural network | |
| CN109376933A (en) | Lithium ion battery negative material energy density prediction technique neural network based | |
| CN113655768B (en) | Assembly yield control method, equipment and computer readable storage medium | |
| CN115809624B (en) | Automatic analysis design method for integrated circuit microstrip line transmission line | |
| CN111257934A (en) | Seismic oscillation peak acceleration prediction method based on second-order neuron deep neural network | |
| CN113110402B (en) | Knowledge and data driven large-scale industrial system distributed state monitoring method | |
| CN110503288A (en) | Consider the System and method for of the interactive identification yield loss reason of board | |
| Zajačko et al. | Application of artificial intelligence principles in mechanical engineering | |
| CN114266278A (en) | Dual-attention-network-based method for predicting residual service life of equipment | |
| CN106503312B (en) | A kind of blade root stress analysis method based on neural network algorithm | |
| CN113222250B (en) | High-power laser device output waveform prediction method based on convolutional neural network | |
| CN114692507A (en) | Counting data soft measurement modeling method based on stacking Poisson self-encoder network | |
| CN113268927B (en) | Output energy prediction method of high-power laser device based on fully-connected neural network | |
| CN108845546B (en) | Dynamic process monitoring method based on BP neural network autoregressive model | |
| CN117350170B (en) | Nonlinear oscillation analysis method based on KOOPHAN deep neural network | |
| CN110147827A (en) | A kind of failure prediction method based on IAALO-SVM and similarity measurement | |
| CN113919221A (en) | Fan load prediction and analysis method and device based on BP neural network and storage medium | |
| US20210117861A1 (en) | Machine learning variable selection and root cause discovery by cumulative prediction | |
| CN113095482A (en) | Raman amplifier gain value calculation method and system based on BP neural network | |
| CN113204857A (en) | Method for predicting residual life of electronic device based on extreme gradient lifting tree algorithm | |
| CN110244690B (en) | Multivariable industrial process fault identification method and system |
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 |