Optical gate type pulse width recovery algorithm based on BP neural network

Abstract

The application provides an optical gate type pulse width recovery algorithm based on a BP neural network, and comprises the following steps: obtaining a FROG graph and a time domain signal of a to-be-measured pulse; taking a real shot femtosecond pulse FROG graph and a time domain signal corresponding to the FROG graph as a training set of the BP neural network; creating the BP neural network; inputting the real shot femtosecond pulse FROG graph and the time domain signal corresponding to the FROG graph into the successfully created BP neural network in step 3 to train; putting the real shot femtosecond pulse FROG graph and the time domain signal corresponding to the FROG graph into the trained BP neural network as a test set to test, and obtaining a predicted value; taking the real shot femtosecond pulse FROG graph and the predicted value as a guessed pulse to substitute into a PGCP pulse width recovery algorithm to obtain a pulse width of the real shot femtosecond pulse FROG graph corresponding pulse. The beneficial effects are as follows: the speed and the accuracy of the pulse width recovery algorithm are improved; the algorithm has a fast and efficient effect; the situation of too many iteration times and too large error is solved, and the measurement difficulty is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of optical precision measurement, and in particular to an optical gate-type pulse width restoration algorithm based on BP neural network. Background Art

[0002] Short-pulse lasers, typically lasting on the femtosecond scale, offer advantages such as short pulse width, high peak power, and a wide spectral range, making them an effective means of understanding various microscopic ultrafast processes. High-peak-power ultrashort-pulse lasers have important applications in ultraprecision micro-nanofabrication, ultrafast testing, and high-power pulsed terahertz spectrum generation.

[0003] Pulse width is a key parameter of ultrafast lasers. Its value directly determines the energy and spectral characteristics of the laser. Precision testing is required in the development, modulation, and detection of femtosecond lasers, as well as in the monitoring and evaluation of the interaction between ultrashort pulse lasers and matter. In ultrafine processing, pulse parameters such as the spatial chirp and wavefront tilt of the pulse affect the uniformity of heating during processing, and have a significant impact on the resolution and processing shape of ultrafine processing. During the ultrafast testing process triggered by the interaction between ultrashort pulses and matter, changes in phase and intensity information are closely related to the ultrafast reaction process. In the generation of high-power pulsed terahertz spectra, the temporal chirp of ultrashort pulse lasers causes an uneven terahertz spectrum, reducing the excitation efficiency.

[0004] Traditional frequency-resolved optical gate-type pulse width recovery algorithms (PGCP) based on second harmonic generation require a guessed pulse as the initial input, and then iterate to generate the time-domain signal and pulse width corresponding to the actual FROG image. However, in actual measurements, femtosecond pulses have various forms, such as Gaussian and bimodal pulses. When the specific pulse type is unknown in actual measurements, only a fixed initial guess pulse signal can be input. However, when the guessed pulse type does not match the actual pulse type, the pulse width recovery algorithm will require too many iterations, resulting in large errors and even iteration stagnation.

[0005] Therefore, there is an urgent need for an optical gate-type pulse width recovery algorithm that can automatically identify the pulse type and fit the initial guess pulse signal closest to the actual captured pulse, while effectively improving the iteration speed and accuracy of the PGCP algorithm. Summary of the Invention

[0006] To solve the above problems, the present invention provides an optical gate type pulse width restoration algorithm based on BP neural network to solve the above problems.

[0007] To achieve the above objectives, the present invention adopts the following technical means:

[0008] An optical gate-type pulse width restoration algorithm based on BP neural network includes the following steps:

[0009] Step 1, obtaining the FROG diagram and time domain signal of the pulse to be measured from a femtosecond pulse width instrument based on a second harmonic generation frequency-resolved optical gate structure;

[0010] Step 2: Using a femtosecond pulse width instrument based on a second harmonic generation frequency-resolved optical gate structure to obtain a real-shot femtosecond pulse FROG diagram and guess the corresponding time domain signal, the real-shot femtosecond pulse FROG diagram and the corresponding time domain signal are used as a training set for the BP neural network;

[0011] Step 3: Create a BP neural network. The BP neural network is a three-layer network architecture consisting of an input layer, a hidden layer, and an output layer. Input the training set in step 2 into the BP neural network for training. Determine the weights and thresholds of the initialized neurons based on the number of input layer nodes and output layer nodes. Use the SOA algorithm to find the time domain signal that is closest to the FROG graph of the pulse to be tested in step 1 and its time domain signal to optimize the initialized weights and thresholds. Ensure that a BP neural network that best meets the weights and thresholds used by the data is obtained. Save the successfully created BP neural network.

[0012] Step 4: input the real-shot femtosecond pulse FROG image and its corresponding time domain signal in step 2 into the BP neural network successfully created in step 3 for training, and save the trained BP neural network;

[0013] Step 5: The real-shot femtosecond pulse FROG image and its corresponding time domain signal in step 2 are put into the BP neural network trained in step 4 as a test set for testing to obtain the predicted value of the improved time domain signal;

[0014] Step 6: Substitute the actual femtosecond pulse FROG diagram in step 2 and the predicted value of the improved time domain signal obtained in step 5 as the guessed pulse into the PGCP pulse width restoration algorithm to obtain the pulse width of the pulse corresponding to the actual femtosecond pulse FROG diagram.

[0015] A further preferred embodiment of the present invention is as follows: the real-shot femtosecond pulse FROG image in step 2 is subjected to noise processing, and the processed FROG image is subjected to data dimensionality reduction processing.

[0016] A further preferred embodiment of the present invention is that the noise processing includes a process of cutting out background noise and a process of filtering out high-frequency noise by low-pass filtering.

[0017] A further preferred solution of the present invention is: the actual femtosecond pulse FROG image and the corresponding time domain signal in step 2 are normalized.

[0018] A further preferred embodiment of the present invention is as follows: the real-shot femtosecond pulse FROG image of the training set in step 3 is used as data input into the BP neural network for forward propagation training of the signal and back propagation training of the error. The forward propagation of the signal is that the signal is input from the input layer, passes through the hidden layer, and is output from the output layer. The back propagation of the error is to calculate the output error of each layer of neurons layer by layer starting from the output layer, and then adjust the weights and thresholds of each layer according to the error gradient descent method to optimize the performance of the BP neural network.

[0019] A further preferred solution of the present invention: Step 3 also includes solving the optimal number of neurons in the hidden layer and constructing a BP neural network of the optimal hidden layer.

[0020] Compared with the prior art, the present invention has the following beneficial effects:

[0021] 1. The present invention provides a second harmonic generation frequency-resolving optical gate type pulse width restoration algorithm based on BP neural network. The femtosecond pulse width tester based on second harmonic generation frequency-resolving optical gate type can effectively improve the speed and accuracy of the PGCP femtosecond pulse width restoration algorithm.

[0022] 2. The BP neural network of the present invention can learn the mathematical relationship between input data and output data without giving the mapping relationship between the two in advance. Its basic idea is the gradient descent method, which uses gradient search technology to minimize the mean square error between the actual output value and the expected output value of the network. It can effectively process the relationship between the femtosecond pulse width FROG map and its time domain signal, and has a fast and efficient effect.

[0023] 3. The second harmonic generation frequency-resolved optical gate pulse width recovery algorithm based on the BP neural network of the present invention can effectively solve the problem of excessive iterations and large errors caused by not knowing the type of the pulse to be measured in actual measurement, and can reduce the difficulty of actual measurement of femtosecond pulses due to automatic recognition of the pulse type. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0025] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0026] The present invention is further described below:

[0027] Example

[0028] like Figure 1 As shown, an optical gate type pulse width restoration algorithm based on BP neural network includes the following steps:

[0029] Step 1, obtaining the FROG diagram and time domain signal of the pulse to be measured from a femtosecond pulse width instrument based on a second harmonic generation frequency-resolved optical gate structure;

[0030] Step 2: Using a femtosecond pulse width instrument based on a second harmonic generation frequency-resolved optical gate structure to obtain a real-shot femtosecond pulse FROG diagram and guess the corresponding time domain signal, the real-shot femtosecond pulse FROG diagram and the corresponding time domain signal are used as a training set for the BP neural network;

[0031] The real-life femtosecond pulse FROG images were processed to remove background noise and high-frequency noise through low-pass filtering. The processed FROG images were then subjected to data dimensionality reduction. This setting makes the dataset easier to use and effectively reduces the time and cost of the algorithm.

[0032] The actual femtosecond pulse FROG image and its corresponding time domain signal are normalized;

[0033]

[0034] Where Xnorm is the normalized data, X is the original data, Xmin is the minimum value of a set of data, and Xmax is the maximum value of a set of data. Normalization can convert the data to the range of [0, 1], which can avoid the large difference in order of magnitude between the FROG data and the PGCP pulse width recovery algorithm, which will affect the final result.

[0035] Step 3: Create a BP neural network. The BP neural network is a three-layer network architecture consisting of an input layer, a hidden layer, and an output layer. Input the training set in step 2 into the BP neural network for training. Determine the weights and thresholds of the initialized neurons based on the number of input layer nodes and output layer nodes. Use the SOA algorithm (global search algorithm) to find the time domain signal that is closest to the FROG graph of the pulse to be tested in step 1 and its time domain signal to optimize the initialized weights and thresholds. Ensure that a BP neural network that best meets the weights and thresholds used by the data is obtained, and save the successfully created BP neural network.

[0036] The real-shot femtosecond pulse FROG images of the training set are used as data input into the BP neural network for forward propagation training of signals and back propagation training of errors. The forward propagation of signals is the process of inputting from the input layer, passing through the hidden layer, and then outputting from the output layer. The back propagation of errors is the process of calculating the output errors of neurons in each layer layer by layer starting from the output layer. The weights and thresholds of each layer are then adjusted according to the error gradient descent method to optimize the performance of the BP neural network. This method can minimize the error between the final predicted value of the optimized BP neural network and the expected value.

[0037] Creating a BP neural network also includes solving the optimal number of neurons in the hidden layer. Let the normalized input layer nodes be X n , the weight of the connection between the output layer and the hidden layer is W nm (n is the number of input layer nodes, m is the number of hidden layer neurons), the hidden layer node bias is β m , the connection weight between the hidden layer and the output layer is V i (i is the number of output layer nodes), the bias of the output layer node is λ i , the output data is set to y, and the node activation function is Sigmoid;

[0038]

[0039] When the BP neural network training set data is forward propagated, the result from the input layer data to the second hidden layer data is set as H m =f(X i *W 1m +X2W 2m …-β m ), the result of the hidden layer to output layer data is K=f(H1*V1+H2*V2+H3*V3……-λ m ); The difference error between the output layer data result K and the output layer data is back propagated, and W and V are automatically updated by the trapezoidal descent algorithm according to the error;

[0040]

[0041] The updated value decreases fastest along the negative gradient direction, where Loss is the loss function. and are the partial derivatives of the loss function with respect to the weights W and V, respectively. η represents the learning rate (which is set independently). The BP neural network collects the errors generated by the system during the simulation process, propagates the errors back, and then adjusts the weights. Through continuous iterative updates, the model tends to the overall optimization and constructs the BP neural network with the best hidden layer.

[0042] Step 4: input the real-shot femtosecond pulse FROG image and its corresponding time domain signal in step 2 into the BP neural network successfully created in step 3 for training, and save the trained BP neural network;

[0043] Step 5: The real-shot femtosecond pulse FROG image and its corresponding time domain signal in step 2 are put into the BP neural network trained in step 4 as a test set for testing to obtain the predicted value of the improved time domain signal;

[0044] Step 6: Substitute the actual femtosecond pulse FROG diagram in step 2 and the predicted value of the improved time domain signal obtained in step 5 as the guessed pulse into the PGCP pulse width restoration algorithm to obtain the pulse width of the pulse corresponding to the actual femtosecond pulse FROG diagram.

[0045] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art will be able to modify the technical solutions described in the aforementioned embodiments or substitute equivalents for some of the technical features. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. An optical gate type pulse width restoration algorithm based on BP neural network, characterized in that: The following steps are involved: Step 1, obtaining the FROG diagram and time domain signal of the pulse to be measured from a femtosecond pulse width instrument based on a second harmonic generation frequency-resolved optical gate structure; Step 2: Using a femtosecond pulse width instrument based on a second harmonic generation frequency-resolved optical gate structure to obtain a real-shot femtosecond pulse FROG diagram and guess the corresponding time domain signal, the real-shot femtosecond pulse FROG diagram and the corresponding time domain signal are used as a training set for the BP neural network; Step 3: Create a BP neural network. The BP neural network is a three-layer network architecture consisting of an input layer, a hidden layer, and an output layer. Input the training set in step 2 into the BP neural network for training. Determine the weights and thresholds of the initialized neurons based on the number of input layer nodes and output layer nodes. Use the SOA algorithm to find the time domain signal that is closest to the FROG graph of the pulse to be tested in step 1 and its time domain signal to optimize the initialized weights and thresholds. Ensure that a BP neural network that best meets the weights and thresholds used by the data is obtained. Save the successfully created BP neural network. Step 4: input the real-shot femtosecond pulse FROG image and its corresponding time domain signal in step 2 into the BP neural network successfully created in step 3 for training, and save the trained BP neural network; Step 5: The real-shot femtosecond pulse FROG image and its corresponding time domain signal in step 2 are put into the BP neural network trained in step 4 as a test set for testing to obtain the predicted value of the improved time domain signal; Step 6: Substitute the actual femtosecond pulse FROG diagram in step 2 and the predicted value of the improved time domain signal obtained in step 5 as the guessed pulse into the PGCP pulse width restoration algorithm to obtain the pulse width of the pulse corresponding to the actual femtosecond pulse FROG diagram.

2. The optical gate type pulse width restoration algorithm based on BP neural network according to claim 1 is characterized in that: The actual femtosecond pulse FROG image described in step 2 is subjected to noise processing, and the processed FROG image is subjected to data dimensionality reduction processing.

3. The optical gate type pulse width restoration algorithm based on BP neural network according to claim 2 is characterized in that: The noise processing includes a process of cutting out background noise and a process of filtering out high-frequency noise by low-pass filtering.

4. The optical gate type pulse width restoration algorithm based on BP neural network according to claim 1 is characterized in that: The actual femtosecond pulse FROG image and the corresponding time domain signal in step 2 are normalized.

5. The optical gate type pulse width restoration algorithm based on BP neural network according to claim 1 is characterized in that: The real-shot femtosecond pulse FROG image of the training set in step 3 is used as data input into the BP neural network for forward propagation training of the signal and back propagation training of the error. The forward propagation of the signal is that the signal is input from the input layer, passes through the hidden layer, and then output from the output layer. The back propagation of the error is to calculate the output error of each layer of neurons layer by layer starting from the output layer, and then adjust the weights and thresholds of each layer according to the error gradient descent method to optimize the performance of the BP neural network.

6. The optical gate type pulse width restoration algorithm based on BP neural network according to claim 1 is characterized in that: Step 3 also includes solving the optimal number of neurons in the hidden layer and constructing a BP neural network with the optimal hidden layer.

Images