Radar signal processing device and method based on optical convolution calculation

Abstract

A radar signal processing device and method based on optical convolution calculation, the device constructs a convolutional neural network through an optical method, and by adjusting the output delay difference of each path in the optical convolution calculation architecture and the output layer electric sampling clock, a multi-step convolution layer calculation can be constructed, so that feature extraction and dimension reduction of the radar signal are realized.The device has the characteristics of full analog domain processing, can greatly reduce the back-end data sampling and processing pressure, and is truly suitable for the actual needs of large bandwidth radar signal processing.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of optical computing, and in particular to a radar signal processing device and method based on optical convolution calculation. BACKGROUND

[0002] Radar transmits electromagnetic waves through an antenna, receives electromagnetic waves reflected by a target, and processes the received electromagnetic waves through a backend computing device to extract information such as the speed, distance, and angle of the target. Currently, radar is widely used in military reconnaissance, aerospace, weather forecasting, ocean exploration, geological exploration, and other fields. However, as the demand for detection continues to increase, radar systems need to have large bandwidth, multiple channels, and high data throughput for real-time processing. Traditional radar signal processing systems mostly use electrical hardware to complete radar data processing, which cannot meet the real-time processing needs of massive data, and new processing methods are urgently needed.

[0003] Current radar signal processing systems mostly use electrical analog-to-digital conversion combined with backend digital computing architecture processing, such as FPGA, DSP, GPU, CPU, etc. However, as the processing signal bandwidth increases, the number of radar channels needed for processing increases, the signal throughput increases, and the real-time processing requirement becomes higher, requiring high-speed analog-to-digital conversion chips, high-speed processing, and high-speed memory reading. The existing electrical computing architecture cannot meet the real-time processing needs. However, there is a lot of redundant information in the radio frequency analog signal, and if the key information is extracted through a convolutional neural network method, the feature extraction and dimensionality reduction of the signal can be achieved, reducing the backend data sampling and processing pressure. Therefore, how to implement a convolutional neural network for large bandwidth signal input is extremely important.

[0004] Under this background, optical computing has become an effective way to reduce the processing pressure of the backend. Optical computing has developed rapidly in recent years, and compared with electrical computing, optical computing has the advantages of high speed, large bandwidth, low power consumption, etc., and can effectively realize real-time processing for the high-throughput computing required by large bandwidth radar. At the same time, optical computing has the ability to implement convolutional neural networks, and current convolutional computing is achieved through dispersion, delay, etc. With its unique parallelization advantage, it can complete convolution operation at high speed, low delay, and high energy efficiency. However, most of the existing optical convolution computing architectures focus on general-purpose computing, and few works focus on how to apply optical convolution computing to radar systems.

[0005] The prior art document CN113705774A discloses a light signal processing device, comprising: a multiplication sequence acquisition module for acquiring a target convolution multiplication sequence; a light signal conversion module for converting the target convolution multiplication sequence into a light signal; a light convolution processing module for performing convolution calculation processing corresponding to multiple sets of convolution weights on the light signal in parallel by using the aforementioned optical circuit to obtain a convolution-processed light signal; and a photoelectric conversion module for performing photoelectric conversion on the convolution-processed light signal to obtain a convolution calculation result. However, additional operations such as singular value decomposition (implemented in a digital computer) are required. SUMMARY

[0006] The present application aims to provide a radar signal feature extraction and dimension reduction device based on optical convolution calculation. The device constructs a convolutional neural network in an optical manner, and by adjusting the output delay difference of each path and the output layer electric sampling clock in the optical convolution calculation architecture, a multi-step convolution layer calculation can be constructed, thereby realizing feature extraction and dimension reduction of radar signals. The device proposed in the present application has the characteristics of all-analog-domain processing, which can greatly reduce the backend data sampling and processing pressure, and is truly suitable for the actual needs of large-bandwidth radar signal processing.

[0007] To achieve the above-mentioned purpose, the technical solutions of the present application are as follows.

[0008] In one aspect, the present application provides a radar signal processing device based on optical convolution calculation, characterized in that it comprises:

[0009] An input layer module for down-converting and modulating the input radar signal onto an optical carrier;

[0010] An optical convolution layer module for realizing convolution calculation with U-layer step length V, extracting features in the radar signal, and reducing the output signal quantity to 1 / (U*V) of the original;

[0011] The optical convolution layer module constructs a convolutional neural network in an optical manner, and by adjusting the output delay difference of each path and the output layer electric sampling clock in the optical convolution calculation architecture, a multi-step convolution layer calculation is constructed, thereby realizing feature extraction and dimension reduction of radar signals;

[0012] An output layer module for realizing pulse sampling, nonlinear calculation, average pooling layer, and analog-to-digital conversion functions, and outputting a digital signal with a bandwidth of 1 / (2*U*V) of the original signal for processing by a backend electric calculation system.

[0013] The input layer module includes radar signal input device, radar signal down conversion device, adjustable delay line, and functions to down convert radar signal and modulate to optical carrier, the convolution layer module includes wavelength division multiplexer, optical convolution calculation device array, Machenhel modulator array, and functions to realize convolution calculation with U layer step length V, extract features in signal and reduce output signal quantity to original 1 / (U*V), the output layer module includes pulse sampling device, high-speed microwave diode, low-pass filter, analog-digital converter, and functions to realize pulse sampling, nonlinear calculation, average pooling layer, analog-digital conversion function, and output bandwidth is digital signal of original signal 1 / (2*U*V) to be processed by rear-end electric calculation system.

[0014] The specific content of each module of the device is described as follows.

[0015] The input layer module includes M radar receiving antennas, M continuous light sources with different wavelengths, M down conversion modules to down convert radar signal and modulate to M optical carriers with different wavelengths, and M adjustable delay lines to align M signals in time and then input the optical convolution layer module.

[0016] The optical convolution layer module is used to realize optical path calculation. In the present application, the convolution layer of the convolutional neural network used to realize feature extraction of radar signal includes U optical convolution calculation layers. Taking an optical convolution kernel chip as an example, the basic principle is as follows:

[0017] In optical convolution calculation, the output after a convolution layer can be expressed by the following formula:

[0018]

[0019] y v,s represents the output value of the vth convolution channel, which needs to be realized through two-dimensional multiplication and addition operation, one dimension is convolution window, and Q in the formula represents the size of the convolution window. The other dimension is the dimension of the input channel, and C in the formula represents the number of input channels. Therefore, the output of one convolution result needs C*Q times of multiplication and addition calculation.w u,v,q represents the weight value corresponding to the u input channel, the v convolution channel and the q output channel; x u,s+q represents the input value corresponding to the u input channel and the q output channel; corresponding to it, each optical convolution layer includes N optical convolution kernel chips, each of which includes on-chip C wavelength division multiplexer, Q on-chip adjustable optical delay line and Q on-chip micro ring resonator array. The wavelength division multiplexer combines C optical signals with different wavelengths into one, the on-chip adjustable delay line can adjust the convolution layer step length V, and the input signal rate is f s , the delay amount of the first convolution layer should be set to 1 / f s , the delay amount of the second convolution layer should be set to V / fs , and so on. The micro-ring on chip can apply weight to the light of specific wavelength. Through the above configuration, the signal entering the first micro-ring resonator array passes through one delay unit, and the signal is denoted as x u,s+1 ; the signal entering the second micro-ring resonator array passes through two delay units, and the signal is denoted as x u,s+2 ; and so on, the signal entering the last micro-ring resonator array is denoted as x u,s+Q . Therefore, only the corresponding micro-ring modulator needs to be adjusted to the preset weight value, and the input optical signal can be multiplied by CxQ times. The Q photodetectors connected by the Q micro-ring resonator arrays convert the optical signals into electrical signals and realize Q-dimensional addition, and the subsequent electrical power combiner realizes C-dimensional electrical signal addition, realizing CxQ times of addition operation. The N electrical signals output by each convolution layer are loaded on N wavelengths by N parallel Mach-Zehnder modulators to input into the next convolution layer.

[0020] The output layer includes N pulse sampling units to realize the extraction of the reduced dimension signal, N nonlinear calculation units to realize the nonlinear activation function, N low-pass filters to realize the average pooling layer, and N analog-to-digital converters to output digital signals. The nonlinear unit can be realized by a nonlinear device such as a microwave diode or by using the nonlinear output curve of an optoelectronic modulator.

[0021] The radar signal feature extraction dimension reduction principle of the optical convolution calculation is as follows:

[0022] In the optical convolution calculation, the output after a convolution layer can be expressed as:

[0023]

[0024] Suppose the signal rate is f s , the first layer delay is 1 / f s , and if the output y v,s is sampled at a frequency of f s / V, then the convolution layer operation with a step size of V is realized. In the next layer, the delay line delay is set to V / f s , which is equivalent to sampling the output y v,s at a frequency of f s / V, realizing the convolution operation with a step size of V. Therefore, in the convolution module, the first convolution layer delay should be set to 1 / f s , the second convolution layer delay should be set to V / f s , and so on. The final output layer electrical sampling clock should be set to f s / (UxV) to realize signal dimension reduction.

[0025] The specific connection of the device is as follows:

[0026] First signal is output by M-way radar signal input device, and the output port is divided into two ways by 1:1 electric power divider, one of which is connected with a port of double parallel Mach-Zehnder modulator after 90° electric phase shifter, and the other is connected with another port of double parallel Mach-Zehnder modulator, realizing single sideband modulation of carrier suppression, which can down-convert radar signal to baseband and modulate to optical carrier. The output port of M-way double parallel Mach-Zehnder modulator is connected with M-way adjustable delay line to align M-way delay. The output port of M-way adjustable delay line is connected with 1 M-way wavelength division multiplexer. The output port of wavelength division multiplexer is connected with N-way optical computing device to realize convolution layer calculation of first layer with step length V1. The N-way output port of optical computing device array is connected with the input port of N-way Mach-Zehnder modulator. The input port of N-way Mach-Zehnder modulator is connected with 1 N-way wavelength division multiplexer, and the output port of wavelength division multiplexer is connected with optical computing device array to realize convolution layer calculation of second layer with step length V2. Then the above process is repeated to realize U layers of convolution layer. The N-way output of optical computing device array of U layers of convolution layer is connected with N-way pulse sampling module in output layer module. N-way pulse sampling module is connected with N-way nonlinear activation module. N-way nonlinear activation module is connected with N-way low-pass filter to realize average pooling layer function. N-way low-pass filter is connected with N-way analog-to-digital converter to obtain digital signal after dimension reduction.

[0027] In another aspect, the application also provides a radar signal processing method based on optical convolution calculation, characterized in that, comprising:

[0028] Collecting radar signal for down-conversion and modulation to optical carrier;

[0029] Carrying out U layers of convolution calculation with step length V on the radar signal, and extracting features of the signal to reduce the output signal quantity to 1 / (U*V) of the original signal;

[0030] Outputting digital signal with bandwidth of 1 / (2*U*V) of the original signal, and handing it over to the back-end electric computing system for processing.

[0031] Further, the U layers of convolution calculation with step length V, the delay amount of the first convolution layer in the convolution module should be set as 1 / f s , the delay amount of the second convolution layer should be set as V / f s , and so on, and the electric sampling clock of the last output layer should be set as f s / (U*V) to realize signal dimension reduction.

[0032] Further, the C channel input radar signal data is loaded on two wavelengths of light and input to the optical convolution layer module at the same time, the first layer convolution is delayed by 1 / fs through the Q path micro ring resonator array, and the weights in the convolution kernel are loaded into each micro ring resonator, so that the multiplication of the C*Q size signal and the corresponding data of the convolution kernel is realized, and the convolution operation is realized; through 1 / V rate sampling of the data, it is equivalent to the convolution kernel sliding backward V data, realizing the convolution step of V, and finally obtaining the single output channel signal length of 1 / V of the single input channel signal length, realizing one layer convolution layer operation; the sampling process can be equivalent to realize by setting the delay difference to V / fs during the operation of the subsequent convolution layer.

[0033] Compared with the prior art, the technical advantages of the application are:

[0034] (1) The application directly modulates the radio frequency signal into the optical domain and processes it using an optical high-speed convolution kernel chip to extract useful information in the radar signal and realize signal dimension reduction, effectively reducing the pressure on the backend electrical processing.

[0035] (2) The corresponding convolution layer calculation is directly realized by using the propagation process of light, the calculation speed is the speed of light, and the processing of large bandwidth radar signals can be carried out with extremely low calculation delay.

[0036] (3) The delay line is used to realize the rearrangement of data in the optical path, the convolution calculation is all deployed in the optical processor, there is no need to convert into digital signals, and there is no need to use additional digital computer processing, so that multi-layer and multi-step convolution layers can be realized in the system, and then the features in the radar signal are extracted and the output signal quantity is greatly reduced. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 It is an embodiment schematic diagram of the optical device for feature extraction and dimension reduction of the radar signal. A feasible device connection mode when U=2, V=2, M=2, Q=3, C1=2, C2=4, N=4 is shown.

[0038] Figure 2 It is a schematic diagram of the optical convolution calculation device. A feasible device connection mode when N=4, C=4, Q=3 is shown.

[0039] Figure 3 It is a schematic diagram of the convolution layer calculation dimension reduction principle. The convolution layer calculation process when V=2, M=2, N=4, Q=3 is shown.

[0040] Figure 4 It is a schematic diagram of the operation mode of the optical device for feature extraction and dimension reduction of the radar signal. A feasible operation mode when V=2, M=2, N=4, Q=3 is shown. DETAILED DESCRIPTION

[0041] The technical solutions of the application are described in detail below in combination with the drawings and examples, and detailed implementation manners and structures are given, but the protection scope of the application is not limited to the following examples.

[0042] A radar signal processing device based on optical convolution calculation, comprising an input layer module, an optical convolution layer module and an output layer module. The input layer module comprises a radar signal input device, a radar signal down-conversion device and an adjustable delay line, and functions to down-convert and modulate the radar signal onto an optical carrier. The convolution layer module comprises a wavelength division multiplexer, an array of optical convolution calculation devices and an array of Mach-Zehnder modulators, and functions to realize convolution calculation with U-layer step length V, extract features in the signal and reduce the output signal quantity to 1 / (U*V) of the original signal. The output layer module comprises a pulse sampling device, a high-speed microwave diode, a low-pass filter and an analog-to-digital converter, and functions to realize pulse sampling, nonlinear calculation, average pooling layer and analog-to-digital conversion. The output is a digital signal with a bandwidth of 1 / (2*U*V) of the original signal, which is processed by a backend electrical calculation system.

[0043] The specific content of each module of the device is described as follows.

[0044] The input layer module comprises M radar receiving antennas, M continuous light sources with different wavelengths, M down-conversion modules for down-converting and modulating the radar signal onto M optical carriers with different wavelengths, and M adjustable delay lines for aligning the M signals in time and inputting them into the optical convolution layer module.

[0045] First, the signal is output from M radar signal input devices. The output port is split into two paths by a 1:1 electrical power splitter. One path is connected to a 90° electrical phase shifter and then to one port of a dual-parallel Mach-Zehnder modulator. The other path is connected to the other port of the dual-parallel Mach-Zehnder modulator, implementing carrier-suppressed single-sideband modulation. This down-converts the radar signal to baseband and modulates it onto an optical carrier. The output ports of the M dual-parallel Mach-Zehnder modulators are connected to M adjustable delay lines to align the M-path delays. The output ports of the M adjustable delay lines are connected to an M-path wavelength division multiplexer. The output port of the wavelength division multiplexer is connected to N optical computing devices to perform the first-layer convolutional layer calculation with a step size of V1. The N output ports of the optical computing device array are connected to the input ports of the N-path Mach-Zehnder modulators. The input ports of the N-path Mach-Zehnder modulators are connected to an N-path wavelength division multiplexer. The output port of the wavelength division multiplexer is connected to the optical computing device array to perform the second-layer convolutional layer calculation with a step size of V2. The above process is then repeated to implement a total of U convolutional layers. The N outputs of the optical computing device array in the Uth convolutional layer are connected to the N pulse sampling modules in the output layer module. The N pulse sampling modules are connected to the N nonlinear activation modules. The N nonlinear activation modules are connected to the N low-pass filters to implement the average pooling layer function. The N low-pass filters are connected to the N analog-to-digital converters to obtain the reduced-dimensional digital signal.

[0046] The "upper" and "lower" described below are based on Figure 1 The placement shown in the figure does not describe the actual system position. The actual system only needs to ensure that the connection order is the same as the figure. In this embodiment, U = 2, V = 2, M = 2, Q = 4, C = 2, N = 4, as shown in the figure. Figure 1 As shown, an optical device for multi-channel radar signal feature extraction and dimensionality reduction includes an input module 100, a convolutional layer module 200, and an output layer module 300. The device specifically includes a two-channel radar signal input device, two 90° electrical phase shifters, two dual-parallel Mach-Zehnder modulators, two wavelength division multiplexers, two optical computing devices, six adjustable delay lines, four Mach-Zehnder modulators, four pulse sampling modules, four high-speed microwave diodes, four low-pass filters, and four analog-to-digital converters. The specific connection methods and functions of each device are described below:

[0047] Firstly, the signal is output by two radar antennas 101, and the output port is divided into two paths by a 1:1 electrical power divider, one of which is connected to a 90° electrical phase shifter 102 and then connected to one port of a double parallel Mach-Zehnder modulator, and the other is connected to the other port of the double parallel Mach-Zehnder modulator 103 to form a radar signal down-conversion device 104, which realizes carrier-suppressed single sideband modulation and can down-convert the radar signal to the baseband and modulate it to the optical carrier. The two radar signal down-conversion devices 104 are connected to two adjustable delay lines 105 to align the two delays. The output ports of the two adjustable delay lines 104 are connected to a wavelength division multiplexer 201. The output ports of the wavelength division multiplexer 201 are connected to an optical computing device array 202 to realize a convolution layer calculation with a step size of 2 in the first layer. In the embodiment, the optical computing device is represented as a 4×3 on-chip architecture with wavelength division delay characteristics. The four output ports of the optical computing device array 202 are connected to the input ports of four Mach-Zehnder modulators 203. The output ports of the four Mach-Zehnder modulators are connected to four adjustable delay lines 204, and the output ports of the four adjustable delay lines are connected to a wavelength division multiplexer 205. The output ports of the wavelength division multiplexer 205 are connected to an optical computing device array 206 to realize a convolution layer calculation with a step size of 2 in the second layer. The four outputs of the optical computing device array 206 are connected to four pulse sampling modules 301 in the output layer module. The four pulse sampling modules 301 are connected to four high-speed microwave diodes 302. The high-speed microwave diodes have the characteristics of forward conduction and negative cut-off, which can realize the rectified linear unit activation function required by the convolutional neural network. The four high-speed microwave diodes 302 are connected to four low-pass filters 303 with a bandwidth of 1 / 8 of the signal bandwidth. The low-pass filters can add two adjacent data to obtain one data, realizing the average pooling layer function in the convolutional neural network. The four low-pass filters 303 are connected to four analog-to-digital converters 304 to obtain the reduced digital signal.

[0048] Referring to Figure 2 , an embodiment of the optical convolution computing device in the present application is shown. The embodiment shows a 4×3 optical convolution kernel chip 400 architecture, which specifically includes three delay lines 401 to realize a multi-step convolution process, three microring resonator arrays 402, and a total of 4×3=12 microring resonators to realize weight loading functions. Three microring resonator arrays are connected to three photodetectors 403 to realize the addition of different wavelength optical signals, and one electrical power combiner 404 is connected to the three photodetectors to realize the addition of the output power of the three microring resonators. Finally, four such chips form a convolution channel of the optical convolution computing array with a channel of 4, realizing a convolution layer with four convolution channels, four input channels, and three output channels.

[0049] Referring to Figure 3, which shows the principle of convolution layer calculation in the application. In the figure, it is assumed that the length of the single input signal is 9, the size of the convolution kernel is 2x3, and the convolution step is 2. The first 2x3 signal of the two input channels is multiplied by the corresponding data of the convolution kernel and added to realize one convolution operation. The convolution kernel slides back two data, and multiplication and addition are performed again to realize the second convolution operation. In this way, the length of the single output channel signal is finally 4. If the length of the input signal is long enough, the length of the single output channel signal is half the length of the single input channel signal, realizing feature extraction and dimension reduction of the signal.

[0050] Referring to Figure 4 , which shows the operation mode schematic diagram of the optical device for radar signal feature extraction and dimension reduction in the application. In the figure, it is assumed that the length of the single input signal is 9, the size of the convolution kernel is 2x3, and the convolution step is 2. The two-channel input signal data is loaded on the light of two wavelengths and input into the optical convolution calculation device at the same time. By using the delay difference of each micro-ring resonator array and loading the weights in the convolution kernel into each micro-ring resonator, the multiplication and addition of the 2x3 signal and the corresponding data of the convolution kernel can be realized, and the convolution operation can be realized. By sampling the data at half the rate, it is equivalent to the convolution kernel sliding back two data, realizing the convolution with a convolution step of 2. The length of the single output channel signal is finally 4. If the length of the input signal is long enough, the length of the single output channel signal is half the length of the single input channel signal, realizing feature extraction and dimension reduction of the signal.

Claims

1. A radar signal processing device based on optical convolution computation, characterized by The application relates to a radar signal processing system based on optical convolution neural network, which comprises the following parts: An input layer module is used for down-converting input radar signals and modulating the radar signals onto an optical carrier; An optical convolution layer module is used for realizing convolution calculation with U layers and V steps, extracting features in the radar signals and reducing the output signal quantity to 1 / (U*V) of the original signal quantity; The optical convolution layer module is used for constructing a convolution neural network through an optical method, adjusting the output delay difference of each path in the optical convolution calculation architecture and the output layer electric sampling clock, constructing convolution layer calculation with multiple steps, and realizing feature extraction and dimension reduction of the radar signals; An output layer module is used for realizing pulse sampling, nonlinear calculation, average pooling layer and analog-digital conversion functions, and outputting a digital signal with a bandwidth of 1 / (2*U*V) of the original signal to be processed by a rear-end electric calculation system; The convolution layer module comprises U-layer wave division multiplexers (201), an optical convolution calculation device array (202), a Mach-Zehnder modulator array (203) and an adjustable delay line (204); After the radar signals pass through the first-layer wave division multiplexer (201), the radar signals are input into the first-layer optical convolution calculation device array (202) to realize convolution layer calculation with the first-layer step being V, then the radar signals pass through N Mach-Zehnder modulators and the adjustable delay line (204) and the second-layer wave division multiplexer (201) to be input into the second-layer optical convolution calculation device array (202) to realize convolution layer calculation with the second-layer step being V, and the radar signals are input into the U-layer optical convolution calculation device array (202) to realize convolution layer calculation with the U-layer step being V.

2. The radar signal processing apparatus based on optical convolution calculation according to claim 1, characterized in that, The input layer module comprises M radar signal input devices (101), a radar signal down-conversion device (104) and an adjustable delay line (105); After the radar signals are received by the M radar signal input devices (101), the radar signals are down-converted to a baseband by the radar signal down-conversion device (104) and modulated onto an optical carrier, the output ports of the M radar signal down-conversion devices (104) are connected with the M adjustable delay lines (105) to align the delay of the two paths, and the output ports of the M adjustable delay lines (105) are connected with the convolution layer module.

3. The radar signal processing apparatus based on optical convolution calculation according to claim 1, characterized in that, The output layer module comprises a pulse sampling device (301), a high-speed microwave diode (302), a low-pass filter (303) and an analog-digital converter (304), The N pulse sampling modules (301) are connected with the N high-speed microwave diodes (302), the high-speed microwave diode has the characteristics of forward conduction and negative cut-off, can realize the rectified linear unit activation function required by the convolution neural network, the N high-speed microwave diodes (302) are connected with the N low-pass filters (303) with a bandwidth of 1 / 8 of the signal bandwidth, the low-pass filter can add two adjacent data to obtain one data, realizes the average pooling layer function in the convolution neural network, and the N low-pass filters (303) are connected with the N analog-digital converters (304) to obtain the dimension-reduced digital signal.

4. A radar signal processing method applied to the radar signal processing device based on optical convolution calculation according to claim 1, characterized in that, The application relates to a radar signal processing system based on optical convolution neural network, which comprises the following parts: The radar signals are collected, down-converted and modulated onto an optical carrier; The radar signals are subjected to convolution calculation with U layers and V steps, and the features of the signals are extracted, so that the output signal quantity is reduced to 1 / (U*V) of the original signal quantity; The output bandwidth of the digital signal is 1 / (2*U*V) of the original signal, which is processed by a back-end electrical computing system.

5. The radar signal processing method based on optical convolution calculation according to claim 4, characterized in that, The U layer step length is the convolution calculation, the first convolution layer delay amount in the convolution module should be set to 1 / f s , the second convolution layer delay amount should be set to V / f s , and so on, and the last output layer electric sampling clock should be set to f s / (UxV), realizing signal dimension reduction.

6. The radar signal processing method based on optical convolution calculation according to claim 4, characterized in that, The C channel input radar signal data is loaded on two wavelengths of light and input to an optical convolution layer module, a first layer of convolution is delayed by 1 / fs through a Q path micro ring resonator array, and the weights in the convolution kernel are loaded into each micro ring resonator, so that the C*Q size signal is multiplied by the corresponding data of the convolution kernel and added, convolution operation is realized; through 1 / V rate sampling of the data, it is equivalent to the convolution kernel sliding backward V data, realizing convolution step length V, and finally obtaining a single output channel signal length of 1 / V of the single input channel signal length, realizing a layer of convolution layer operation; the sampling process can be equivalent to realizing by setting the delay difference to V / fs during the operation of the subsequent convolution layer.

Images