Neural network architecture based on SAW delay line and signal processing method
Through the neural network architecture based on SAW delay lines, nonlinear operations are used to perform nonlinear operations using parallel and series resonator groups, the problem of low computing efficiency of KAN network is solved, efficient analog signal processing and synaptic transmission are achieved, and energy consumption is reduced.
Patent Information
- Application Number
- CN202510277900.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-10
- Publication Date
- 2025-08-05
AI Technical Summary
The computing efficiency of the KAN network is low, mainly due to the timing control and data conversion caused by relying on digital circuits, which increases the response time.
Using a neural network architecture based on SAW delay lines, the first resonator group in parallel and the second resonator group in series are used to perform nonlinear addition and multiplication operations of the analog signal to avoid timing control and data conversion.
Improves computing efficiency, reduces energy consumption, and operates stably at low voltage and low power, avoiding repeated calculations and redundant operations in digital calculations.
Smart Images
Figure CN120430360A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of neural network technology, and in particular to a neural network architecture and signal processing method based on a SAW delay line. Background Art
[0002] KAN networks, or Kolmogorov-Amnold Networks, are a new type of neural network architecture. Currently, KAN network operations rely on digital circuits, such as operational amplifiers, digital signal processors, and other digital computing hardware units. KAN networks typically involve nonlinear addition and multiplication operations, but operations in digital circuits often involve timing control and data conversion, which increases response time and reduces the operational efficiency of KAN networks. Summary of the Invention
[0003] The main purpose of this application is to provide a neural network architecture and signal processing method based on SAW delay lines, aiming to solve the technical problem of reduced computational efficiency of KAN networks.
[0004] To achieve the above-mentioned object, the present application provides a neural network architecture based on a SAW delay line, comprising an input terminal, a nonlinear calculation module, and an output terminal;
[0005] The nonlinear calculation module includes a plurality of first resonator groups and a plurality of second resonator groups;
[0006] The first resonator group includes at least two resonators, and the resonators in the first resonator group are connected in parallel to each other;
[0007] The second resonator group includes at least one resonator, and the resonators in the second resonator group are connected in series;
[0008] Wherein, the input end is used to receive an analog signal and transmit the analog signal to the nonlinear calculation module;
[0009] The plurality of first resonator groups in the nonlinear calculation module are used to perform addition operation on the analog signal;
[0010] The plurality of second resonator groups in the nonlinear calculation module are used to perform multiplication operation on the analog signal;
[0011] The output end is used to output the analog signal processed by the nonlinear calculation module.
[0012] Optionally, the first resonator group includes N first resonators, where N is a positive integer greater than 1;
[0013] The input end of each first resonator serves as the input end of the first resonator group, and the output end of each first resonator is electrically connected through a target node, and the target node serves as the output end of the first resonator group.
[0014] Optionally, the second resonator group includes M second resonators, where M is a positive integer;
[0015] The input end of the second resonator located at the front of the second resonator group serves as the input end of the second resonator group, and the output end of the second resonator located at the back of the second resonator group serves as the output end of the second resonator group.
[0016] Optionally, the resonator includes a first interdigital electrode, a second interdigital electrode, a magnetic film and a substrate;
[0017] The magnetic film is arranged between the first interdigital electrodes and the second interdigital electrodes, and a portion of the first interdigital electrodes, a portion of the second interdigital electrodes and the magnetic film are wrapped in the substrate.
[0018] Optionally, the magnetic film may be tilted at 15 degrees, 30 degrees, or 45 degrees.
[0019] In addition, to achieve the above-mentioned purpose, the present application also provides a signal processing method, which is applied to the neural network architecture based on the SAW delay line as described above, wherein the neural network architecture includes a nonlinear computing module;
[0020] The signal processing method comprises:
[0021] Receive analog signals;
[0022] The nonlinear calculation module performs nonlinear addition and nonlinear multiplication operations on the analog signal to obtain a processed analog signal.
[0023] Optionally, the nonlinear calculation module includes a plurality of first resonator groups for performing addition operations on the analog signals, the first resonator groups include N first resonators, and N is a positive integer greater than 1;
[0024] Performing a nonlinear addition operation on the analog signal, comprising:
[0025] Calculating a calculation result of each first resonator on the analog signal to obtain N calculation results;
[0026] Calculate the sum of the N calculation results.
[0027] Optionally, calculating the calculation result of each first resonator on the analog signal to obtain N calculation results includes:
[0028] Obtaining the tilt angle corresponding to the magnetic film in each first resonator;
[0029] Determining a first weight value corresponding to each first resonator based on a tilt angle corresponding to the magnetic film in each first resonator to obtain N first weight values;
[0030] Based on the N first weight values, a calculation result of each first resonator on the analog signal is obtained, thereby obtaining N calculation results.
[0031] Optionally, the nonlinear calculation module includes a plurality of second resonator groups for performing multiplication operations on the analog signals, the second resonator groups include M first resonators, and M is a positive integer;
[0032] Performing a nonlinear multiplication operation on the analog signal, comprising:
[0033] Obtaining the activation function corresponding to each second resonator to obtain M activation functions;
[0034] A multiplication operation is performed on the analog signal based on the M activation functions.
[0035] Optionally, performing a multiplication operation on the analog signal based on the M activation functions includes:
[0036] obtaining a tilt angle corresponding to the magnetic film in each second resonator;
[0037] Determining a second weight value corresponding to each second resonator based on a tilt angle corresponding to the magnetic film in each second resonator to obtain M second weight values;
[0038] A multiplication operation is performed on the analog signal based on the M second weight values and the M activation functions.
[0039] Compared with the prior art, the embodiments of the present application have the following beneficial effects:
[0040] The present application provides a neural network architecture and signal processing method based on a SAW delay line. The above-mentioned neural network architecture based on a SAW delay line includes an input end, a nonlinear calculation module and an output end; the nonlinear calculation module includes multiple first resonator groups and multiple second resonator groups; the first resonator group includes at least 2 resonators, and the resonators in the first resonator group are connected in parallel to each other; the second resonator group includes at least 1 resonator, and the resonators in the second resonator group are connected in series to each other; wherein, the input end is used to receive an analog signal and transmit the analog signal to the nonlinear calculation module; the multiple first resonator groups in the nonlinear calculation module are used to perform addition operations on the analog signals; the multiple second resonator groups in the nonlinear calculation module are used to perform multiplication operations on the analog signals; the output end is used to output the analog signal processed by the nonlinear calculation module. The neural network architecture provided in the embodiment of the present application can simulate the nonlinear addition and multiplication operations of the KAN network, and does not involve timing control and data conversion, thereby reducing the delay of the operation and improving the operation efficiency.
[0041] In addition, the neural network architecture provided by the embodiment of the present application does not rely on a large number of switching operations and can operate stably at low voltage and low power, greatly reducing energy consumption.
[0042] In addition, the neural network architecture provided in the embodiment of the present application can directly perform multiplication and addition operations at the hardware level, simulating synaptic transmission between neurons in the neural network, greatly improving computing efficiency, and avoiding the inevitable repeated calculations and redundant operations in digital computing. BRIEF DESCRIPTION OF THE DRAWINGS
[0043] In order to more clearly illustrate the solutions in this application, a brief introduction will be given below to the drawings required for use in the description of the embodiments of this application. Obviously, the drawings described below are some embodiments of this application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.
[0044] Figure 1 It is a structural diagram of the neural network architecture provided in the embodiment of the present application;
[0045] Figure 2-1 to Figure 2-2 is a structural diagram of a first resonator group provided in an embodiment of the present application;
[0046] Figure 3-1 to Figure 3-2 is a schematic structural diagram of a second resonator group provided in an embodiment of the present application;
[0047] Figure 4 is a schematic structural diagram of a resonator provided in an embodiment of the present application;
[0048] Figures 5-1 to 5-12This is one of the structural diagrams of the resonator group provided in the embodiment of the present application;
[0049] Figure 6 is a flowchart of a signal processing method provided in an embodiment of the present application;
[0050] Figure 7 This is the second structural diagram of the resonator group provided in the embodiment of the present application
[0051] Description of reference numerals:
[0052] 10. Input end; 20. Nonlinear computing module; 30. Output end; 21. First resonator group; 22. Second resonator group; 31. First interdigital electrode; 32. Second interdigital electrode; 33. Magnetic film; 34. Substrate. DETAILED DESCRIPTION
[0053] Unless otherwise defined, all technical and scientific terms used herein have the same meanings as commonly understood by those skilled in the art to which this application belongs. The terms used in the specification of the application are for the purpose of describing specific embodiments only and are not intended to limit this application. The terms "including" and "having" and any variations thereof in the specification and claims of this application and the above-mentioned drawings are intended to cover non-exclusive inclusions. The terms "first", "second", etc. in the specification and claims of this application or the above-mentioned drawings are used to distinguish different objects, not to describe a specific order.
[0054] References herein to "embodiments" mean that a particular feature, structure, or characteristic described in connection with the embodiments may be included in at least one embodiment of the present application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor does it constitute an independent or alternative embodiment that is mutually exclusive of other embodiments. It is understood, both explicitly and implicitly, by those skilled in the art that the embodiments described herein may be combined with other embodiments.
[0055] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the accompanying drawings.
[0056] See also Figure 1 , Figure 1 This is a schematic diagram of the structure of the neural network architecture provided in the embodiment of this application. Figure 1 As shown, the neural network architecture includes an input terminal 10, a nonlinear calculation module 20 and an output terminal 30;
[0057] The nonlinear calculation module 20 includes a plurality of first resonator groups 21 and a plurality of second resonator groups 22;
[0058] The first resonator group 21 includes at least two resonators, and the resonators in the first resonator group 21 are connected in parallel with each other;
[0059] The second resonator group 22 includes at least one resonator, and the resonators in the second resonator group 22 are connected in series.
[0060] like Figure 1 As shown, the neural network architecture provided in this embodiment includes an input terminal 10, a nonlinear computing module 20, and an output terminal 30, wherein the nonlinear computing module 20 includes a plurality of first resonator groups 21 and a plurality of second resonator groups 22. It should be noted that the neural network architecture provided in this embodiment is used to simulate a KAN network.
[0061] The working principle of the neural network architecture provided in this embodiment is:
[0062] The input terminal 10 is used to receive an analog signal and transmit the analog signal to the nonlinear calculation module 20;
[0063] The plurality of first resonator groups 21 in the nonlinear calculation module 20 are used to perform addition operations on analog signals;
[0064] The plurality of second resonator groups 22 in the nonlinear calculation module 20 are used to perform multiplication operations on analog signals;
[0065] The output terminal 30 is used to output the analog signal processed by the nonlinear calculation module 20 .
[0066] The neural network architecture provided in the embodiment of the present application can simulate the nonlinear addition and multiplication operations of the KAN network without involving timing control and data conversion, thereby reducing the operation delay and improving the operation efficiency.
[0067] In addition, the neural network architecture provided by the embodiment of the present application does not rely on a large number of switching operations and can operate stably at low voltage and low power, greatly reducing energy consumption.
[0068] In addition, the neural network architecture provided in the embodiment of the present application can directly perform multiplication and addition operations at the hardware level, simulating synaptic transmission between neurons in the neural network, greatly improving computing efficiency, and avoiding the inevitable repeated calculations and redundant operations in digital computing.
[0069] Optionally, the first resonator group includes N first resonators, where N is a positive integer greater than 1;
[0070] The input end of each first resonator serves as the input end of the first resonator group, and the output end of each first resonator is electrically connected through a target node, and the target node serves as the output end of the first resonator group.
[0071] See also Figure 2-1 to Figure 2-2 , Figure 2-1 to Figure 2-2 Schematic diagram of the structure of the first resonator group provided in an embodiment of the present application. Figure 2-1 The first resonator group shown includes 2 first resonators, Figure 2-2 The illustrated first resonator group includes three first resonators.
[0072] The input end of each first resonator serves as the input end of the first resonator group, the output end of each first resonator is electrically connected via a target node, and the target node serves as the output end of the first resonator group. Figure 2-1 to Figure 2-2 The "u" in the figure represents the input of the resonator, and the "y" represents the output of the resonator. In the subsequent drawings of the specification, the "u" represents the input of the resonator and the "y" represents the output of the resonator, and the description will not be repeated.
[0073] It should be understood that in other embodiments, the first resonator group may further include other numbers of first resonators, and the plurality of first resonators are connected in parallel.
[0074] Optionally, the second resonator group includes M second resonators, where M is a positive integer;
[0075] The input end of the second resonator located at the front of the second resonator group serves as the input end of the second resonator group, and the output end of the second resonator located at the back of the second resonator group serves as the output end of the second resonator group.
[0076] See also Figure 3-1 to Figure 3-2 , Figure 3-1 to Figure 3-2 Schematic diagram of the structure of the second resonator group provided in an embodiment of the present application. Figure 3-1 The second resonator group shown includes 2 second resonators, Figure 3-2 The second resonator group shown includes three second resonators.
[0077] The input end of the second resonator located at the front of the second resonator group serves as the input end of the second resonator group, and the output end of the second resonator located at the back of the second resonator group serves as the output end of the second resonator group.
[0078] It should be understood that in other embodiments, the second resonator group may further include other numbers of second resonators, and the plurality of second resonators are connected in series.
[0079] Optionally, the resonator includes a first interdigital electrode, a second interdigital electrode, a magnetic film and a substrate;
[0080] The magnetic film is arranged between the first interdigital electrodes and the second interdigital electrodes, and a portion of the first interdigital electrodes, a portion of the second interdigital electrodes and the magnetic film are wrapped in the substrate.
[0081] See also Figure 4 , Figure 4 Schematic diagram of the structure of the resonator provided in the embodiment of the present application. Figure 4 As shown, the resonator includes a first interdigitated electrode 31, a second interdigitated electrode 32, a magnetic film 33 and a substrate 34; the magnetic film 33 is arranged between the first interdigitated electrode 31 and the second interdigitated electrode 32, and part of the electrodes of the first interdigitated electrode 31, the electrodes of the second interdigitated electrode 32 and the magnetic film 33 are wrapped in the substrate 34.
[0082] The first interdigital electrode 31 serves as the input end of the resonator, and part of the first interdigital electrode 31 is protruded from the surface of the substrate 34 ; the second interdigital electrode 32 serves as the output end of the resonator, and part of the second interdigital electrode 32 is protruded from the surface of the substrate 34 .
[0083] Optionally, the magnetic film may be tilted at 15 degrees, 30 degrees, or 45 degrees.
[0084] See also Figures 5-1 to 5-12 ,like Figure 5-1 As shown, in the second resonator group including two second resonators connected in series, one of the second resonators can be arranged to be inclined at 15 degrees.
[0085] like Figure 5-2 As shown, in the second resonator group including two second resonators connected in series, one of the second resonators can be tilted at 30 degrees.
[0086] like Figure 5-3 As shown, in the second resonator group including two second resonators connected in series, one of the second resonators can be arranged to be inclined at 45 degrees.
[0087] like Figure 5-4 As shown, in the second resonator group including two second resonators connected in series, one of the second resonators can be arranged at an inclination of 15 degrees, and the other second resonator can be arranged at an inclination of 30 degrees.
[0088] like Figure 5-5 As shown, in the second resonator group including two second resonators connected in series, one of the second resonators can be arranged at an inclination of 15 degrees, and the other second resonator can be arranged at an inclination of 45 degrees.
[0089] like Figure 5-6As shown, in the second resonator group including two second resonators connected in series, one of the second resonators can be tilted at 30 degrees, and the other second resonator can be tilted at 45 degrees.
[0090] like Figure 5-7 As shown, in the first resonator group including two first resonators connected in parallel, one of the first resonators can be tilted at 15 degrees.
[0091] like Figure 5-8 As shown, in the first resonator group including two first resonators connected in parallel, one of the first resonators can be tilted at 30 degrees.
[0092] like Figure 5-9 As shown, in the first resonator group including two first resonators connected in parallel, one of the first resonators can be tilted at 45 degrees.
[0093] like Figure 5-10 As shown, in the first resonator group including two first resonators connected in parallel, one of the first resonators can be arranged at an inclination of 15 degrees, and the other can be arranged at an inclination of 30 degrees.
[0094] like Figure 5-11 As shown, in the first resonator group including two first resonators connected in parallel, one of the first resonators can be arranged at an inclination of 15 degrees, and the other can be arranged at an inclination of 45 degrees.
[0095] like Figure 5-12 As shown, in the first resonator group including two first resonators connected in parallel, one of the first resonators can be arranged at an inclination of 15 degrees, and the other can be arranged at an inclination of 45 degrees.
[0096] It should be noted that the first resonators connected in parallel in the first resonator group can be tilted at 15 degrees, 30 degrees, or 45 degrees. The second resonators connected in series in the second resonator group can be tilted at 15 degrees, 30 degrees, or 45 degrees.
[0097] It should be understood that by changing the tilt angle of the magnetic film, the incident angle of the analog signal is adjusted, so that nonlinear addition and multiplication operations are performed on the analog signal using different weight values.
[0098] In other embodiments, the magnetic film may also be tilted at other angles, not limited to 15 degrees, 30 degrees, or 45 degrees, which will not be elaborated in detail here.
[0099] See also Figure 6 , Figure 6 This is a flow chart of the signal processing method provided by the embodiment of the present application. Figure 6 As shown, the signal processing method provided in the embodiment of the present application includes:
[0100] S601, receiving an analog signal.
[0101] The signal processing provided in the embodiments of the present application is applied to a neural network architecture based on a SAW delay line, wherein the neural network architecture includes an input end, a nonlinear computing module, and an output end; the nonlinear computing module includes multiple first resonator groups and multiple second resonator groups; the first resonator group includes at least 2 resonators, and the resonators in the first resonator group are connected in parallel to each other; the second resonator group includes at least 1 resonator, and the resonators in the second resonator group are connected in series to each other.
[0102] In this step, an analog signal is received, wherein the analog signal may be a signal sent by other devices to the neural network architecture.
[0103] S602 , performing nonlinear addition and nonlinear multiplication operations on the analog signal through a nonlinear calculation module to obtain a processed analog signal.
[0104] In this step, after receiving the analog signal, the nonlinear calculation module in the neural network architecture performs nonlinear addition and nonlinear multiplication operations on the analog signal to obtain a processed analog signal.
[0105] The neural network architecture provided in the embodiment of the present application can simulate the nonlinear addition and multiplication operations of the KAN network without involving timing control and data conversion, thereby reducing the operation delay and improving the operation efficiency.
[0106] In addition, the neural network architecture provided by the embodiment of the present application does not rely on a large number of switching operations and can operate stably at low voltage and low power, greatly reducing energy consumption.
[0107] In addition, the neural network architecture provided in the embodiment of the present application can directly perform multiplication and addition operations at the hardware level, simulating synaptic transmission between neurons in the neural network, greatly improving computing efficiency, and avoiding the inevitable repeated calculations and redundant operations in digital computing.
[0108] Optionally, the nonlinear calculation module includes a plurality of first resonator groups for performing addition operations on the analog signals, the first resonator groups include N first resonators, and N is a positive integer greater than 1;
[0109] Performing a nonlinear addition operation on the analog signal, comprising:
[0110] Calculating a calculation result of each first resonator on the analog signal to obtain N calculation results;
[0111] Calculate the sum of the N calculation results.
[0112] In this embodiment, the steps of performing a nonlinear addition operation on the analog signal are:
[0113] The calculation result of each first resonator on the analog signal is calculated to obtain N calculation results, and the sum of the N calculation results is calculated to achieve a nonlinear addition operation on the analog signal.
[0114] Optionally, calculating the calculation result of each first resonator on the analog signal to obtain N calculation results includes:
[0115] Obtaining the tilt angle corresponding to the magnetic film in each first resonator;
[0116] Determining a first weight value corresponding to each first resonator based on a tilt angle corresponding to the magnetic film in each first resonator to obtain N first weight values;
[0117] Based on the N first weight values, a calculation result of each first resonator on the analog signal is obtained, thereby obtaining N calculation results.
[0118] As described above, the magnetic film may be disposed at an inclination.
[0119] In this embodiment, the tilt angle corresponding to the magnetic film in each first resonator is obtained, and the first weight value corresponding to each first resonator is determined based on the tilt angle corresponding to the magnetic film. It should be understood that the mapping relationship between the tilt angle and the first weight value can be preset.
[0120] Furthermore, the product of the calculation result of each first resonator on the analog signal and the corresponding first weight value is used as the final calculation result.
[0121] Optionally, the nonlinear calculation module includes a plurality of second resonator groups for performing multiplication operations on the analog signals, the second resonator groups include M first resonators, and M is a positive integer;
[0122] Performing a nonlinear multiplication operation on the analog signal, comprising:
[0123] Obtaining the activation function corresponding to each second resonator to obtain M activation functions;
[0124] A multiplication operation is performed on the analog signal based on the M activation functions.
[0125] In this embodiment, the steps of performing a nonlinear multiplication operation on the analog signal are:
[0126] An activation function corresponding to each second resonator is obtained to obtain M activation functions; and a multiplication operation is performed on the analog signal based on the M activation functions. Specifically, each second resonator multiplies the input analog signal using the corresponding activation function, thereby achieving nonlinear multiplication of the analog signal.
[0127] Optionally, performing a multiplication operation on the analog signal based on the M activation functions includes:
[0128] obtaining a tilt angle corresponding to the magnetic film in each second resonator;
[0129] Determining a second weight value corresponding to each second resonator based on a tilt angle corresponding to the magnetic film in each second resonator to obtain M second weight values;
[0130] A multiplication operation is performed on the analog signal based on the M second weight values and the M activation functions.
[0131] As described above, the magnetic film may be disposed at an inclination.
[0132] In this embodiment, the tilt angle corresponding to the magnetic film in each second resonator is obtained, and the second weight value corresponding to each second resonator is determined based on the tilt angle corresponding to the magnetic film. It should be understood that the mapping relationship between the tilt angle and the second weight value can be preset.
[0133] Furthermore, each second resonator performs a multiplication operation on the input analog signal through the corresponding activation function and the second weight value.
[0134] To facilitate understanding of the above solution, a specific implementation method is described below. Figure 7 , Figure 7 It includes five resonators, namely S1, S2, S3, S4 and S5, wherein S1 and S2 form a first-stage parallel circuit, S3 and S4 form a second-stage parallel circuit, and S5 forms a third-stage series circuit.
[0135] The output of the first stage circuit is: y1=f1(u)+f2(u)
[0136] The input of the second stage circuit is: u1=f1(u)+f2(u)
[0137] The output of the second stage circuit is: y2=f3(u1)+f4(u1)
[0138] The input of the third stage circuit is: u2=y2=f3(u1)+f4(u1)
[0139] Calculation result: y3=f5(u2)=f5[f3(u1)+f4(u1)]
[0140] Among them, y1 represents the output of the first-level circuit, u represents the input of the first-level circuit, u1 represents the input of the second-level circuit, y2 represents the output of the second-level circuit, u2 represents the input of the third-level circuit, y3 represents the output of the third-level circuit, f1 represents the activation function corresponding to S1, f2 represents the activation function corresponding to S2, f3 represents the activation function corresponding to S3, f4 represents the activation function corresponding to S4, and f5 represents the activation function corresponding to S5.
[0141] Obviously, the embodiments described above are only some of the embodiments of the present application, rather than all of the embodiments. The preferred embodiments of the present application are given in the accompanying drawings, but they do not limit the patent scope of the present application. The present application can be implemented in many different forms. On the contrary, the purpose of providing these embodiments is to make the understanding of the disclosure of the present application more thorough and comprehensive. Although the present application has been described in detail with reference to the aforementioned embodiments, for those skilled in the art, it is still possible to modify the technical solutions described in the aforementioned specific embodiments, or to make equivalent replacements for some of the technical features therein. Any equivalent structure made using the contents of the present application specification and the accompanying drawings, directly or indirectly used in other related technical fields, is also within the scope of patent protection of the present application.
Claims
1. A neural network architecture based on a SAW delay line, characterized in that: It includes an input terminal, a nonlinear calculation module and an output terminal; The nonlinear calculation module includes a plurality of first resonator groups and a plurality of second resonator groups; The first resonator group includes at least two resonators, and the resonators in the first resonator group are connected in parallel to each other; The second resonator group includes at least one resonator, and the resonators in the second resonator group are connected in series; Wherein, the input end is used to receive an analog signal and transmit the analog signal to the nonlinear calculation module; The plurality of first resonator groups in the nonlinear calculation module are used to perform addition operation on the analog signal; The plurality of second resonator groups in the nonlinear calculation module are used to perform multiplication operation on the analog signal; The output end is used to output the analog signal processed by the nonlinear calculation module.
2. The neural network architecture according to claim 1, characterized in that The first resonator group includes N first resonators, where N is a positive integer greater than 1; The input end of each first resonator serves as the input end of the first resonator group, and the output end of each first resonator is electrically connected through a target node, and the target node serves as the output end of the first resonator group.
3. The neural network architecture according to claim 1, characterized in that The second resonator group includes M second resonators, where M is a positive integer; The input end of the second resonator located at the front of the second resonator group serves as the input end of the second resonator group, and the output end of the second resonator located at the back of the second resonator group serves as the output end of the second resonator group.
4. The neural network architecture according to any one of claims 1 to 3, characterized in that The resonator includes a first interdigital electrode, a second interdigital electrode, a magnetic film and a substrate; The magnetic film is arranged between the first interdigital electrodes and the second interdigital electrodes, and a portion of the first interdigital electrodes, a portion of the second interdigital electrodes and the magnetic film are wrapped in the substrate.
5. The neural network architecture according to claim 4, characterized in that The magnetic film can be tilted at 15 degrees, 30 degrees, or 45 degrees.
6. A signal processing method, characterized in that: A neural network architecture based on a SAW delay line as claimed in any one of claims 1 to 5, wherein the neural network architecture comprises a nonlinear computing module; The signal processing method comprises: Receive analog signals; The nonlinear calculation module performs nonlinear addition and nonlinear multiplication operations on the analog signal to obtain a processed analog signal.
7. The method according to claim 6, characterized in that The nonlinear calculation module includes a plurality of first resonator groups for performing addition operations on the analog signals, the first resonator groups including N first resonators, where N is a positive integer greater than 1; Performing a nonlinear addition operation on the analog signal, comprising: Calculating a calculation result of each first resonator on the analog signal to obtain N calculation results; Calculate the sum of the N calculation results.
8. The method according to claim 7, characterized in that The calculating of the calculation result of each first resonator on the analog signal to obtain N calculation results includes: Obtaining the tilt angle corresponding to the magnetic film in each first resonator; Determining a first weight value corresponding to each first resonator based on a tilt angle corresponding to the magnetic film in each first resonator to obtain N first weight values; Based on the N first weight values, a calculation result of each first resonator on the analog signal is obtained, thereby obtaining N calculation results.
9. The method according to claim 6, characterized in that The nonlinear calculation module includes a plurality of second resonator groups for performing multiplication operations on the analog signal, the second resonator group includes M first resonators, where M is a positive integer; Performing a nonlinear multiplication operation on the analog signal, comprising: Obtaining the activation function corresponding to each second resonator to obtain M activation functions; A multiplication operation is performed on the analog signal based on the M activation functions.
10. The method according to claim 9, characterized in that The performing a multiplication operation on the analog signal based on the M activation functions includes: obtaining a tilt angle corresponding to the magnetic film in each second resonator; Determining a second weight value corresponding to each second resonator based on a tilt angle corresponding to the magnetic film in each second resonator to obtain M second weight values; A multiplication operation is performed on the analog signal based on the M second weight values and the M activation functions.