Microwave Photonic Neuron Convolution Acceleration Device for Electromagnetic Signal Processing

Abstract

The present invention discloses a microwave photonic neuron convolution acceleration device for electromagnetic signal processing, which relates to the technical field of signal processing and includes: a microwave photonic neuron convolution accelerator connected to the electromagnetic signal processing main control unit, configured to perform matrix multiplication and addition linear operations on the received electromagnetic signals or processed data and a preset weight vector in the optical analog domain, and output the matrix multiplication and addition linear operation results to the electromagnetic signal processing main control unit; an electromagnetic signal processing main control unit, configured to input the pre-obtained electromagnetic signals in vector format to the microwave photonic neuron convolution accelerator, and configured to process the received matrix multiplication and addition linear operation results to obtain processed data in vector format, and output the processed data to the microwave photonic neuron convolution accelerator when a preset condition is not met. The present invention can achieve fast processing and recognition of broadband electromagnetic signals, reduce the computational energy consumption requirements, and reduce the computing power requirements and energy efficiency ratio requirements for electronic devices.

Description

Technical Field

[0001] The present invention relates to the technical field of signal processing, and particularly to a microwave photonic neuron convolution acceleration device for electromagnetic signal processing. Background Art

[0002] With the development and large-scale popularization of emerging technologies such as the Internet of Things, video communication, satellite communication, and radar, the utilization of electromagnetic spectrum resources has become very important in various fields. Although the International Telecommunication Union (ITU) has divided the uses of frequency bands, electromagnetic interference or illegal use between different services may still damage important electromagnetic systems. In addition, with the development of the above emerging technologies, the requirements for signal bandwidth and real-time processing of large-scale signals are also increasing. Therefore, high-performance electromagnetic signal receiving and processing devices are of great significance for the perception and monitoring of electromagnetic spectrum resources. Traditional electromagnetic signal receiving and processing devices based on microwave / digital signal systems are usually limited by the analog-to-digital conversion (ADC) of the digital signal processing part and the performance of electronic components, and it is difficult to achieve high-speed processing and real-time perception of large-bandwidth signals under the limited cost and power consumption resources.

[0003] The high parallelism and high processing speed of optical computing technology can quickly process a large amount of data to meet the signal processing requirements of systems such as modern communication, radar, remote sensing, and electronic reconnaissance. Secondly, optical computing technology can achieve non-linear signal processing. By constructing single or multiple neuron models and non-linear activation functions, the calculation of neural network models can be accelerated, which is particularly suitable for the intelligent and rapid identification of electromagnetic targets. In addition, combined with the large bandwidth characteristics of microwave photons, it can cover signals in multiple frequency bands, which is beneficial to the classification and identification of electromagnetic signals under multi-scenario transformations. Moreover, optical computing technology has good electromagnetic interference resistance, which is also very beneficial to the normal operation and stable operation of devices in the field of electromagnetic signal processing. Therefore, microwave photon-optical computing technology provides a new idea for electromagnetic signal perception and processing.

[0004] At present, optical neural networks and photonic convolutional calculations based on microwave photonic optoelectronic hybrid processing architectures have been widely studied and applied. For example, the literature titled "11 TOPS photonic convolutional accelerator for optical neural networks" published by Xingyuan Xu, Mengxi Tan, etc. in the journal "Nature" in 2021 proposed an optical vector convolutional accelerator that can operate at a speed of more than 10 TOPS (trillion operations per second) and can be used for large-scale image data processing to achieve fast image recognition. The literature titled "Integrated Photonic Tensor Core Based on Phase-Change Memory Arrays" published by Wolfram Pernice, etc. in the journal "Nature Communications" in 2021 proposed an integrated photonic tensor core based on a phase-change material storage array that can operate at a speed of about 0.5 TOPS (trillion operations per second), achieved parallel matrix-vector multiplication operations, and tested the image recognition effect on a handwritten digit database.

[0005] Currently, the research mainly focuses on the application of microwave photon-optical computing technology in the field of image processing and recognition tasks. The acceleration devices for image processing and recognition tasks cannot be applied to electromagnetic signal fast processing and recognition tasks; moreover, the performance of the current acceleration devices based on microwave photon-optical computing technology highly depends on the performance of optical materials, the device structure is relatively complex, the device construction cost is relatively high, and there are relatively large limitations. Summary of the Invention

[0006] To solve some or all of the above technical problems existing in the prior art, the present invention provides a microwave photon neuron convolutional acceleration device for electromagnetic signal processing.

[0007] The technical solution of the present invention is as follows:

[0008] A microwave photon neuron convolutional acceleration device for electromagnetic signal processing is provided, including:

[0009] A microwave photon neuron convolutional accelerator, connected to the electromagnetic signal processing main control unit, for performing matrix multiplication and addition linear operations on the received electromagnetic signals or processed data and preset weight vectors in the optical analog domain, and outputting the matrix multiplication and addition linear operation results to the electromagnetic signal processing main control unit;

[0010] The main control unit for electromagnetic signal processing is used to input the pre-obtained electromagnetic signal in vector format to the microwave photonic neuron convolution accelerator, and is also used to process the result of matrix multiplication and addition linear operation received, obtain the processed data in vector format, and output the processed data to the microwave photonic neuron convolution accelerator when a preset condition is not met.

[0011] In some alternative embodiments, the microwave photonic neuron convolution accelerator includes:

[0012] A multi-wavelength optical excitation unit, connected to the weight control unit, which is used to generate multiple optical carriers with different wavelengths and equal adjacent wavelength intervals, and synthesize the multiple optical carriers into one and then output it to the weight control unit;

[0013] The weight control unit, connected to the photoelectric modulation unit, has a preset weight vector, which is used to load weights on the received optical carriers according to the weight vector, divide the optical carriers into multiple paths according to different wavelengths, and output the multiple optical carriers in parallel to the photoelectric modulation unit, where the optical carriers of each wavelength are loaded with the same weight;

[0014] An input vector processing unit, connected to the photoelectric modulation unit, which is used to output the electromagnetic signal or processed data sent by the main control unit for electromagnetic signal processing to the photoelectric modulation unit;

[0015] The photoelectric modulation unit, connected to the wavelength delay unit, is used to modulate the received electromagnetic signal or processed data onto the multiple optical carriers input in parallel by the weight control unit to obtain multiple optical signals, and output the multiple optical signals in parallel to the wavelength delay unit, where the optical carriers of each wavelength are loaded with the same modulation signal, and the modulation signal is the electromagnetic signal or processed data received by the photoelectric modulation unit;

[0016] The wavelength delay unit, connected to the feature extraction unit, is used to apply a time delay to the optical signals of different wavelengths, so that the multiple optical signals are sequentially delayed by a set duration according to the wavelength size;

[0017] The feature extraction unit, connected to the photoelectric detection unit, is used to extract the optical signal segments within a set time window from the delayed optical signals and output them to the photoelectric detection unit;

[0018] The photoelectric detection unit is used to perform linear superposition on the received multiple optical signals.

[0019] In some alternative embodiments, the set duration is expressed as:

[0020] ;

[0021] Among them, represents the set duration, represents the modulation period required for the photoelectric modulation unit to modulate the modulation signal onto one optical carrier, represents the number of wavelength types of the optical signal.

[0022] In some alternative embodiments, the multi-wavelength optical excitation unit includes an optical frequency comb.

[0023] In some alternative embodiments, the weight control unit includes a waveform shaper and an optical splitter connected in sequence.

[0024] In some alternative embodiments, the photoelectric modulation unit includes a broadband photoelectric modulator.

[0025] In some alternative embodiments, the wavelength delay unit includes dispersion optical fiber.

[0026] In some alternative embodiments, the photoelectric detection unit includes a broadband flat intensity detector.

[0027] In some alternative embodiments, the microwave photon neuron convolution accelerator includes two input ports and one output port, and the electromagnetic signal processing main control unit includes one input port and two output ports;

[0028] The two input ports of the microwave photon neuron convolution accelerator are respectively connected to the two output ports of the electromagnetic signal processing main control unit, the output port of the microwave photon neuron convolution accelerator is connected to the input port of the electromagnetic signal processing main control unit, the electromagnetic signal processing main control unit inputs electromagnetic signals and processing data to the microwave photon neuron convolution accelerator through the two output ports respectively, and the microwave photon neuron convolution accelerator inputs the matrix multiplication and addition linear operation result to the electromagnetic signal processing main control unit through the output port.

[0029] In some alternative embodiments, a main control host computer is further included;

[0030] The main control host computer is connected to the electromagnetic signal processing main control unit. When the preset conditions are met, the electromagnetic signal processing main control unit outputs the processing data to the main control host computer, and the main control host computer is used to display the received processing data.

[0031] The main advantages of the technical solution of the present invention are as follows:

[0032] The microwave photonic neuron convolution acceleration device of the present invention migrates large-scale matrix multiply-add linear operations to the optical simulation domain by using a microwave photonic neuron convolution accelerator, which can greatly reduce the demand for computing energy consumption of electronic devices, and reduce the computing power demand and energy efficiency ratio demand for electronic devices; by utilizing the advantages of high parallelism, high speed, and large bandwidth of the microwave photonic neuron convolution accelerator, it can achieve fast processing and recognition of broadband electromagnetic signals. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] The drawings described herein are used to provide a further understanding of the embodiments of the present invention, and constitute a part of the present invention. The illustrative embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0034] Figure 1 It is a schematic structural diagram of a microwave photonic neuron convolution acceleration device for electromagnetic signal processing provided by an embodiment of the present invention;

[0035] Figure 2 It is a schematic structural diagram of a microwave photonic neuron convolution accelerator provided by an embodiment of the present invention;

[0036] Figure 3 It is a schematic diagram of the principle of a microwave photonic neuron convolution accelerator provided by an embodiment of the present invention to implement matrix multiply-add linear operations;

[0037] Figure 4 It is a schematic structural diagram of another microwave photonic neuron convolution accelerator provided by an embodiment of the present invention, where the feature extraction unit is not shown;

[0038] Figure 5 It is a schematic diagram of the operation process of 32-way parallel optical signals to implement matrix multiply-add linear operations in the time-wavelength plane provided by an embodiment of the present invention;

[0039] Figure 6 It is a schematic structural diagram of another microwave photonic neuron convolution acceleration device for electromagnetic signal processing provided by an embodiment of the present invention;

[0040] Figure 7 It is a schematic diagram of the recognition accuracy of 8 types of electromagnetic signals provided by Example 1 of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0041] To make the objectives, technical solutions, and advantages of the present invention more clear, the following will clearly and completely describe the technical solutions of the present invention in combination with specific embodiments of the present invention and the corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.

[0042] The following will detail the technical solutions provided by the embodiments of the present invention in combination with the drawings.

[0043] Referring to Figure 1 , the embodiments of the present invention provide a microwave photonic neuron convolution acceleration device for electromagnetic signal processing. The device includes:

[0044] A microwave photonic neuron convolution accelerator 1, connected to an electromagnetic signal processing main control unit 2, for performing matrix multiplication and addition linear operations (MM-MAC) on the received electromagnetic signals or processed data and a preset weight vector in the optical analog domain, and outputting the matrix multiplication and addition linear operation result to the electromagnetic signal processing main control unit 2;

[0045] An electromagnetic signal processing main control unit 2, for inputting the pre-obtained electromagnetic signals in vector format to the microwave photonic neuron convolution accelerator 1, and for processing the received matrix multiplication and addition linear operation result to obtain processed data in vector format, and outputting the processed data to the microwave photonic neuron convolution accelerator 1 when a preset condition is not met.

[0046] Specifically, when the microwave photonic neuron convolution acceleration device provided by the embodiments of the present invention is in use, first, the electromagnetic signal to be recognized and processed is sent to the electromagnetic signal processing main control unit 2. The electromagnetic signal processing main control unit 2 converts the received electromagnetic signal into a vector format and sends it to the microwave photonic neuron convolution accelerator 1. The electromagnetic signal in vector format can be regarded as the input of the first-layer neurons of the neural network. The microwave photonic neuron convolution accelerator 1 performs matrix multiplication and addition linear operations on the received electromagnetic signal and the preset weight vector in the optical simulation domain, and then outputs the result of the matrix multiplication and addition linear operation to the electromagnetic signal processing main control unit 2. The electromagnetic signal processing main control unit 2 performs preset data post-processing on the received matrix multiplication and addition linear operation to obtain processed data in vector format. This processed data can be regarded as the output of the first-layer neurons of the neural network. After obtaining the processed data, the electromagnetic signal processing main control unit 2 determines whether the preset conditions are met. If the preset conditions are not met, the processed data is sent to the microwave photonic neuron convolution accelerator 1. The processed data in vector format can be regarded as the input of the second-layer neurons of the neural network. The microwave photonic neuron convolution accelerator 1 performs matrix multiplication and addition linear operations on the received processed data and the preset weight vector in the optical simulation domain, and then outputs the result of the matrix multiplication and addition linear operation to the electromagnetic signal processing main control unit 2. The electromagnetic signal processing main control unit 2 performs preset data post-processing on the received matrix multiplication and addition linear operation to obtain processed data in vector format. This processed data can be regarded as the output of the second-layer neurons of the neural network. Then, it continues to determine whether the preset conditions are met. If the preset conditions are not met, the above processing process of the processed data is continued until the preset conditions are met, and the finally obtained processed data is used as the electromagnetic signal processing result.

[0047] In the embodiments of the present invention, by using the electromagnetic signal to be recognized and processed as the initial input, and cyclically performing matrix multiplication and addition linear operation processing of the microwave photonic neuron convolution accelerator 1 and data post-processing of the electromagnetic signal processing main control unit 2 for multiple times, the processing and recognition process of the electromagnetic signal by the multi-layer neural network can be realized, thereby completing the rapid processing and recognition of the electromagnetic signal.

[0048] In the embodiments of the present invention, the preset conditions in the electromagnetic signal processing main control unit 2 are specifically set according to the actual situation. For example, the preset condition is that the number of times of data post-processing by the electromagnetic signal processing main control unit 2 reaches the set threshold M. At this time, the matrix multiplication and addition linear operation processing of the microwave photonic neuron convolution accelerator 1 and the data post-processing of the electromagnetic signal processing main control unit 2 can be cyclically performed M times, thereby constituting a neural network with M layers of neurons and realizing the processing and recognition of the electromagnetic signal.

[0049] In the embodiments of the present invention, the post-processing of data performed by the electromagnetic signal processing main control unit 2 is specifically set according to the actual situation. For example, the data post-processing includes: sampling, quantization, biasing, and non-linear activation processing performed in sequence.

[0050] The microwave photonic neuron convolution acceleration device provided by the embodiments of the present invention migrates large-scale matrix multiplication and addition linear operations to the optical analog domain by using the microwave photonic neuron convolution accelerator 1, which can greatly reduce the demand for computing energy consumption of electronic devices, reduce the computing power demand and energy efficiency ratio demand for electronic devices; by utilizing the advantages of high parallelism, high speed, and large bandwidth of the microwave photonic neuron convolution accelerator 1, it can achieve fast processing and recognition of broadband electromagnetic signals.

[0051] Reference Figure 2 , in the embodiments of the present invention, in order to implement the functions of the microwave photonic neuron convolution accelerator 1 defined above, reduce the structural complexity of the microwave photonic neuron convolution accelerator 1, and facilitate the implementation of the microwave photonic neuron convolution accelerator 1, the microwave photonic neuron convolution accelerator 1 includes:

[0052] A multi-wavelength optical excitation unit 11, connected to the weight control unit 12, the multi-wavelength optical excitation unit 11 is used to generate multiple optical carriers with different wavelengths and equal adjacent wavelength intervals, and synthesize the multiple optical carriers into one and output it to the weight control unit 12;

[0053] A weight control unit 12, connected to the optoelectronic modulation unit 14, the weight control unit 12 is preset with a weight vector, and is used to load weights on the received optical carriers according to the weight vector, divide the optical carriers into multiple paths according to different wavelengths, and output the multiple optical carriers in parallel to the optoelectronic modulation unit 14, wherein, the optical carriers of each wavelength are loaded with the same weight;

[0054] An input vector processing unit 13, connected to the optoelectronic modulation unit 14, is used to output the electromagnetic signal or processing data sent by the electromagnetic signal processing main control unit 2 to the optoelectronic modulation unit 14;

[0055] An optoelectronic modulation unit 14, connected to the wavelength delay unit 15, the optoelectronic modulation unit 14 is used to modulate the received electromagnetic signal or processing data onto the multiple optical carriers input in parallel by the weight control unit 12 to obtain multiple optical signals, and output the multiple optical signals in parallel to the wavelength delay unit 15, wherein, the optical carriers of each wavelength are loaded with the same modulation signal, and the modulation signal is the electromagnetic signal or processing data received by the optoelectronic modulation unit 14;

[0056] A wavelength delay unit 15, connected to the feature extraction unit 16, the wavelength delay unit 15 is used to apply a time delay to the optical signals of different wavelengths, so that the multiple optical signals are sequentially delayed by a set duration according to the wavelength size;

[0057] The feature extraction unit 16 is connected to the photoelectric detection unit 17. The feature extraction unit 16 is configured to extract optical signal segments within a set time window from the time-delayed optical signal and output them to the photoelectric detection unit 17;

[0058] The photoelectric detection unit 17 is configured to perform linear superposition on the received multiple optical signals.

[0059] In the embodiment of the present invention, the multi-wavelength optical excitation unit 11 generates multiple parallel optical carriers with equal wavelength intervals to construct a time-wavelength plane. The multi-wavelength optical excitation unit 11 provides multiple parallel optical carriers as calculation carriers. Each wavelength corresponds to an independent signal channel. All wavelengths are output at the initial moment, forming a signal plane where the time dimension and the wavelength dimension are intertwined, providing a physical basis for the subsequent loading of weights and modulation signals. The weight control unit 12 loads the same weight vector onto the optical carriers of each wavelength. The weight control unit 12 maps the weight information onto the optical carriers by means such as optical power adjustment or wavelength offset, so as to store the weight information in the time-wavelength plane and realize the pre-association of weights and modulation signals. The input vector processing unit 13 is used as a modulation signal source to provide modulation signals. The input vector processing unit 13 cooperates with the photoelectric modulation unit 14 to ensure the accurate mapping of the modulation signals in the time-wavelength plane. The photoelectric modulation unit 14 is configured to uniformly load the modulation signals provided by the input vector processing unit 13 onto all wavelengths, so that the optical carriers of each wavelength carry the same modulation signal, and ensure the signal segmentation alignment in the time-wavelength plane. The wavelength delay unit 15 introduces specific delays to the optical signals of different wavelengths to adjust the timing of the optical signals. Specifically, the diagonal elements in the time-wavelength plane are stretched to the same time window, providing timing alignment for subsequent feature extraction and accumulation. The feature extraction unit 16 extracts the minimum feature units from the time-delayed optical signals, such as optical signal segments within a specific time window, completes the segmentation cutting of the signals, and ensures the independence of subsequent accumulation operations. The photoelectric detection unit 17 converts the optical signals extracted by the feature extraction unit 16 into electrical signals, and through the integration effect, completes the linear superposition of all signals within a specific time window, obtaining the matrix multiplication and addition linear operation result of the modulation signal and the weight vector.

[0060] In the embodiments of the present invention, the optoelectronic modulation unit 14, the wavelength delay unit 15, the feature extraction unit 16, and the optoelectronic detection unit 17 each include one or more, and the number of the optoelectronic modulation unit 14, the wavelength delay unit 15, the feature extraction unit 16, and the optoelectronic detection unit 17 is the same. Among them, when the optoelectronic modulation unit 14, the wavelength delay unit 15, the feature extraction unit 16, and the optoelectronic detection unit 17 are all one, the microwave photonic neuron convolution accelerator 1 can form a one-dimensional microwave photonic neuron; when the optoelectronic modulation unit 14, the wavelength delay unit 15, the feature extraction unit 16, and the optoelectronic detection unit 17 are all multiple, the microwave photonic neuron convolution accelerator 1 can form a multi-dimensional microwave photonic neuron. For example, when the optoelectronic modulation unit 14, the wavelength delay unit 15, the feature extraction unit 16, and the optoelectronic detection unit 17 are all N, the microwave photonic neuron convolution accelerator 1 forms an N-dimensional microwave photonic neuron, where N is a preset positive integer.

[0061] It is set that the multi-wavelength optical excitation unit 11 generates n optical carriers with wavelengths of respectively, the weight vector preset in the weight control unit 12 is , and the signal vector output by the input vector processing unit 13 is . Then, the principle of the matrix multiplication and addition linear operation of a single neuron formed by the multi-wavelength optical excitation unit 11, the weight control unit 12, the input vector processing unit 13, one optoelectronic modulation unit 14, one wavelength delay unit 15, one feature extraction unit 16, and one optoelectronic detection unit 17 in the microwave photonic neuron convolution accelerator 1 is as Figure 3 shown. Specifically, the multi-wavelength optical excitation unit 11 generates n optical carriers with wavelengths of respectively, and synthesizes the n optical carriers into one output to construct an initial time-wavelength plane; after the n optical carriers pass through the weight control unit 12, each optical carrier with a wavelength is loaded with the same weight vector ; after the optical carrier loaded with the weight vector is output to the optoelectronic modulation unit 14, the optoelectronic modulation unit 14 uses the signal vector output by the input vector processing unit 13 as a modulation signal and uniformly modulates it onto the n optical carriers, so that each optical carrier with a wavelength is loaded with the same modulation signal ; when the modulation period is , the duration of each element in the signal vector is ; through the processing of the optoelectronic modulation unit 14, the weight vector and the signal vector Perform linear weighting, and store the weight information and signal information in the time-wavelength plane; then use the wavelength delay unit 15 to control the delay of the time-wavelength plane containing the weight information, so as to generate an equal amount of delay between each wavelength on the time-wavelength plane, and the specific duration of the delay is ; After the time-wavelength plane containing the weight information is delayed by the wavelength delay unit 15, it is manifested as the stretching of the time dimension of its plane. By making the delay between different wavelengths , the elements on the diagonal line of the time-wavelength plane containing the weight information can be stretched to the same time window ; The feature extraction unit 16 extracts and divides the minimum feature units of the elements in the same time window, and outputs them to the photoelectric detection unit 17; The photoelectric detection unit 17 has no wavelength selectivity and has the characteristic of optical signal amplitude response, and can extract all the corresponding elements on the time window, complete the accumulation operation of each weighted component, and obtain the final operation result , .

[0062] Reference Figure 4 , in order to facilitate the implementation of the microwave photonic neuron convolution accelerator 1, in the embodiment of the present invention, the multi-wavelength optical excitation unit 11 includes an optical frequency comb, the weight control unit 12 includes a waveform shaper and an optical splitter, the photoelectric modulation unit 14 includes a broadband photoelectric modulator, the wavelength delay unit 15 includes a dispersion fiber, and the photoelectric detection unit 17 includes a broadband flat-type intensity detector.

[0063] Among them, the waveform shaper is connected to the optical splitter. The waveform shaper is used for weight loading, and the optical splitter is used to divide the optical signal loaded with weights into multiple outputs.

[0064] In the embodiment of the present invention, by using the above-mentioned optical devices and electrical devices to construct the microwave photonic neuron convolution accelerator 1, the practicability of the microwave photonic neuron convolution acceleration device can be improved, and the device cost can be reduced.

[0065] Reference Figure 4 , further, considering that the output optical power of the optical frequency comb is low, the multi-wavelength optical excitation unit 11 further includes an optical amplifier, and the optical amplifier is connected to the optical frequency comb. The optical amplifier is used to amplify the optical signal output by the optical frequency comb.

[0066] In the embodiment of the present invention, when the wavelength delay unit 15 includes a dispersion fiber, the delay control of the optical signal is realized by using the dispersion control of the dispersion fiber, and the corresponding delay duration is expressed as:

[0067] ;

[0068] Among them, represents the set duration, represents the dispersion coefficient of the dispersion fiber, with the unit of ps / km·nm, represents the wavelength interval of the optical signal, represents the fiber length of the dispersion fiber.

[0069] According to the above duration expression corresponding to the dispersion fiber, when the wavelength remains unchanged, the adjustment of the delay duration can be achieved by adjusting the dispersion coefficient and the fiber length of the dispersion fiber.

[0070] Furthermore, in the embodiments of the present invention, the broadband optoelectronic modulator can adopt an optoelectronic modulator with a 3dB bandwidth > 20GHz, and the broadband flat-type intensity detector can adopt a flat-type intensity detector with a bandwidth > 10GHz.

[0071] Reference Figure 5 , Figure 5 shows the operation process of 32 parallel optical signals realizing matrix multiplication and addition linear operations in the time-wavelength plane based on the microwave photonic neuron convolution accelerator 1 provided by the embodiments of the present invention. Assuming the input vector is [0,1,0,0], the clock period of the input vector is 32 times the delay amount of adjacent optical signals, and the weight vector is a 32-bit [1,1,…,1]. Using Figure 4 the shown dual-channel parallel microwave photonic neuron convolution accelerator can realize matrix multiplication and addition linear operations.

[0072] Reference Figure 1 , furthermore, in the embodiments of the present invention, the microwave photonic neuron convolution accelerator 1 includes two input ports and one output port, and the electromagnetic signal processing main control unit 2 includes one input port and two output ports;

[0073] The two input ports of the microwave photonic neuron convolution accelerator 1 are respectively connected to the two output ports of the electromagnetic signal processing main control unit 2. The output port of the microwave photonic neuron convolution accelerator 1 is connected to the input port of the electromagnetic signal processing main control unit 2. The electromagnetic signal processing main control unit 2 inputs electromagnetic signals and processing data to the microwave photonic neuron convolution accelerator 1 through the two output ports respectively, and the microwave photonic neuron convolution accelerator 1 inputs the matrix multiplication and addition linear operation result to the electromagnetic signal processing main control unit 2 through the output port.

[0074] In the embodiments of the present invention, by providing two input ports on the microwave photonic neuron convolution accelerator 1 and two output ports on the electromagnetic signal processing main control unit 2, and respectively inputting electromagnetic signals and processing data into the microwave photonic neuron convolution accelerator 1 by using the two output ports, interference during the transmission of electromagnetic signals and processing data can be avoided, ensuring the accuracy of the data processing results.

[0075] Reference Figure 6 , Further, in the embodiments of the present invention, the microwave photonic neuron convolution acceleration device further includes a main control host computer 3;

[0076] The main control host computer 3 is connected to the electromagnetic signal processing main control unit 2. When the electromagnetic signal processing main control unit 2 meets a preset condition, it outputs the processing data to the main control host computer 3, and the main control host computer 3 is used to display the received processing data.

[0077] Among them, the main control host computer 3 can also adjust and control the electromagnetic signal processing main control unit 2. Specifically, it can input the electromagnetic signals to be identified and processed into the electromagnetic signal processing main control unit 2, and set the preset conditions in the electromagnetic signal processing main control unit 2.

[0078] By providing the main control host computer 3, it is convenient to adjust and control the electromagnetic signal processing main control unit 2, and to output and display the finally obtained processing data, which is convenient for subsequent calling.

[0079] Further, in the embodiments of the present invention, the electromagnetic signal processing main control unit 2 is constructed by using a digital-to-analog converter (DAC), an analog-to-digital converter (ADC), a field-programmable gate array (FPGA), and a graphics processing unit (GPU). The main control host computer 3 selects an X86 system or an Ubuntu system and supports 32-bit or 64-bit floating-point calculations.

[0080] The beneficial effects of the microwave photonic neuron convolution acceleration device for electromagnetic signal processing provided by the embodiments of the present invention are described below with specific examples:

[0081] Example 1

[0082] In this Example 1, build as Figure 4The dual-channel parallel microwave photonic neuron convolution accelerator shown performs matrix multiplication and addition linear operations on electromagnetic signals, and an electromagnetic signal processing main control unit 2 is built to perform acquisition, quantization, biasing, and non-linear activation processing on the results of matrix multiplication and addition linear operations. Eight types of electromagnetic signals including AM-DSB signals, AM-SSB signals, BPSK signals, CPFSK signals, GFSK signals, 4-PAM signals, 16-QAM signals, and QPSK signals are selected as the electromagnetic signals to be recognized and processed. The built microwave photonic neuron convolution accelerator 1 and electromagnetic signal processing main control unit 2 are used to recognize and process the electromagnetic signals to be recognized and processed, and the recognition accuracy rate as shown in Figure 7 is obtained. It can be seen that the microwave photonic neuron convolution acceleration device for electromagnetic signal processing provided by the embodiment of the present invention can complete the rapid processing and recognition of electromagnetic signals.

[0083] It should be noted that in this article, relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device. In addition, in this article, "front", "rear", "left", "right", "upper", and "lower" are all referenced based on the placement state shown in the drawings.

[0084] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A microwave photonic neuron convolution acceleration device for electromagnetic signal processing, comprising: A microwave photonic neuron convolution accelerator, connected to the electromagnetic signal processing main control unit, for performing matrix multiplication and addition linear operations on the received electromagnetic signals or processed data and a preset weight vector in the optical analog domain, and outputting the matrix multiplication and addition linear operation results to the electromagnetic signal processing main control unit; The electromagnetic signal processing main control unit is used to input the electromagnetic signals in vector format obtained in advance to the microwave photonic neuron convolution accelerator, and is used to process the received matrix multiplication and addition linear operation results to obtain processed data in vector format, and output the processed data to the microwave photonic neuron convolution accelerator when the preset conditions are not met. It is characterized in that the microwave photonic neuron convolution accelerator includes: A multi-wavelength optical excitation unit, connected to the weight control unit, the multi-wavelength optical excitation unit is used to generate multiple optical carriers with different wavelengths and equal adjacent wavelength intervals, and synthesize the multiple optical carriers into one path and output it to the weight control unit; The weight control unit, connected to the optoelectronic modulation unit, the weight control unit presets a weight vector, and is used to load weights on the received optical carriers according to the weight vector, and divide the optical carriers into multiple paths according to different wavelengths, and output the multiple optical carriers in parallel to the optoelectronic modulation unit, wherein the optical carriers of each wavelength are loaded with the same weight; An input vector processing unit, connected to the optoelectronic modulation unit, for outputting the electromagnetic signals or processed data sent by the electromagnetic signal processing main control unit to the optoelectronic modulation unit; The optoelectronic modulation unit, connected to the wavelength delay unit, the optoelectronic modulation unit is used to modulate the received electromagnetic signals or processed data onto the multiple optical carriers input in parallel by the weight control unit to obtain multiple optical signals, and output the multiple optical signals in parallel to the wavelength delay unit, wherein the optical carriers of each wavelength are loaded with the same modulation signal, and the modulation signal is the electromagnetic signal or processed data received by the optoelectronic modulation unit; The wavelength delay unit, connected to the feature extraction unit, the wavelength delay unit is used to apply a time delay to the optical signals of different wavelengths, so that the multiple optical signals are sequentially delayed by a set duration according to the wavelength size; The feature extraction unit, connected to the optoelectronic detection unit, the feature extraction unit is used to extract the optical signal segments within a set time window from the delayed optical signals and output them to the optoelectronic detection unit; The optoelectronic detection unit is used to perform linear superposition on the received multiple optical signals.

2. The microwave photonic neuron convolution acceleration device for electromagnetic signal processing according to claim 1, wherein The set duration is expressed as: ； Among them, represents the set duration, represents the modulation period required for the photoelectric modulation unit to modulate the modulation signal onto one optical carrier, represents the number of wavelength types of the optical signal.

3. The microwave photonic neuron convolution acceleration device for electromagnetic signal processing according to claim 1, characterized in that The multi-wavelength optical excitation unit includes an optical frequency comb.

4. The microwave photonic neuron convolution acceleration device for electromagnetic signal processing according to claim 1, wherein The weight control unit includes a waveform shaper and an optical splitter connected in sequence.

5. The microwave photonic neuron convolution acceleration device for electromagnetic signal processing according to claim 1, characterized in that, The optoelectronic modulation unit includes a broadband optoelectronic modulator.

6. The microwave photonic neuron convolution acceleration device for electromagnetic signal processing according to claim 1, wherein The wavelength delay unit includes dispersion optical fiber.

7. The microwave photonic neuron convolution acceleration device for electromagnetic signal processing according to claim 1, characterized in that The optoelectronic detection unit includes a broadband flat-type intensity detector.

8. The microwave photonic neuron convolution acceleration device for electromagnetic signal processing according to claim 1, characterized in that The microwave photonic neuron convolution accelerator includes two input ports and one output port, and the electromagnetic signal processing main control unit includes one input port and two output ports; Two input ports of the microwave photonic neuron convolution accelerator are respectively connected to two output ports of the electromagnetic signal processing main control unit, and an output port of the microwave photonic neuron convolution accelerator is connected to an input port of the electromagnetic signal processing main control unit. The electromagnetic signal processing main control unit inputs electromagnetic signals and processing data into the microwave photonic neuron convolution accelerator through two output ports respectively, and the microwave photonic neuron convolution accelerator inputs the matrix multiplication and addition linear operation results into the electromagnetic signal processing main control unit through the output port.

9. The microwave photonic neuron convolution acceleration device for electromagnetic signal processing according to any one of claims 1-8, characterized in that It further includes a main control host computer; The main control host computer is connected to the electromagnetic signal processing main control unit. When the preset conditions are met, the electromagnetic signal processing main control unit outputs the processing data to the main control host computer, and the main control host computer is used to display the received processing data.

Images