CN117763399B - Neural network classification method for self-adaptive variable-length signal input - Google Patents
Neural network classification method for self-adaptive variable-length signal input Download PDFInfo
- Publication number
- CN117763399B CN117763399B CN202410194232.9A CN202410194232A CN117763399B CN 117763399 B CN117763399 B CN 117763399B CN 202410194232 A CN202410194232 A CN 202410194232A CN 117763399 B CN117763399 B CN 117763399B
- Authority
- CN
- China
- Prior art keywords
- neural network
- signal
- size
- pooling
- convolution
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 60
- 238000000034 method Methods 0.000 title claims abstract description 16
- 238000011176 pooling Methods 0.000 claims abstract description 34
- 238000004364 calculation method Methods 0.000 claims abstract description 32
- 238000010586 diagram Methods 0.000 claims abstract description 14
- 238000005457 optimization Methods 0.000 claims abstract description 3
- 230000003044 adaptive effect Effects 0.000 claims description 7
- 238000007781 pre-processing Methods 0.000 claims description 5
- 238000001914 filtration Methods 0.000 claims description 4
- 238000013527 convolutional neural network Methods 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 2
- 238000001514 detection method Methods 0.000 description 2
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000006870 function Effects 0.000 description 1
- 238000003058 natural language processing Methods 0.000 description 1
Landscapes
- Image Analysis (AREA)
Abstract
The invention provides a neural network classification method of self-adaptive variable-length signal input, which comprises the steps of firstly calculating the length of a signal to be processed, and windowing continuous signals according to the length of the signal; on the other hand, calculating the characteristic diagram sizes of different convolution layers and the characteristic diagram sizes of different pooling layers in the neural network classifier according to the length of the signal to be processed; the neural network classifier generates adjustment parameters of the current layer according to the received characteristic diagram sizes of the convolution layers and the characteristic diagram sizes of the pooling layers of different layers, so that the optimization adjustment of the number of registers and operation units used in the convolution operation and pooling operation of the current layer is completed; and finally, finishing classification processing by using the neural network classifier which is optimally adjusted. The invention utilizes the calculation unit which can adapt to the size of the feature map in the neural network to realize the neural network classification based on the variable-length signal input, improves the robustness and the universality of the neural network classifier, and reduces the utilization rate of system resources.
Description
Technical Field
The invention relates to a deep learning technology, in particular to a neural network classification technology for self-adaptive variable-length signal input.
Background
Deep learning is a representation learning method capable of autonomously learning data features, and has high-efficiency and strong classification capability and feature learning capability in an unsupervised or supervised mode. Neural networks, such as convolutional neural network CNN, are widely used in the fields of computer vision, natural language processing, physiological signals, etc., as a common method in the current artificial intelligence field.
The length of an input signal is required to be fixed in training and testing, when the length of the input signal is changed, the original data is required to be truncated or zero-padded, so that the information of part of the original signals is lost or useless information is introduced, and finally, the accuracy of the classification of the neural network is reduced or the redundant calculation amount is increased.
Taking electrocardiographic signal processing as an example, the length of a heart beat of different people at different moments is variable, in the existing method, because the input signal is of a fixed length, in order to ensure that the classification accuracy is not reduced, the length of the input signal is generally set to be the maximum heart beat length which can occur, however, useful information of the heart beat only exists between the beginning and the end of the heart beat, and the rest is redundant input, so that unnecessary calculation is caused.
Disclosure of Invention
The invention aims to solve the technical problems that the length of the acquired real data is uncertain in the applications of voice processing, physiological signal processing and the like, and the size of an input signal cannot be efficiently adapted to the existing fixed-length input neural network classification method, and provides a neural network classification method suitable for variable-length signal input.
The invention discloses a neural network classification method for self-adaptive variable-length signal input, which comprises the following steps:
Pretreatment: receiving an input continuous original signal, and carrying out filtering and denoising pretreatment on the original signal;
Self-adaptive windowing: calculating the length of a signal to be processed according to the preprocessed signal, on one hand, windowing the continuous signal according to the length of the signal, and outputting the windowed signal to a neural network classifier; on the other hand, according to the signal length to be processed, the corresponding sizes of all layers in the neural network classifier, namely the characteristic map sizes of the convolution layers of different layers and the characteristic map sizes of the pooling layers of different layers are calculated and output to the neural network classifier;
And (3) self-adaptive adjustment: generating adjustment parameters of a current layer by the neural network classifier according to the received characteristic diagram sizes of the convolution layers of different layers and the characteristic diagram sizes of the pooling layers of different layers, and then completing optimization adjustment of the number of registers and operation units used in the convolution operation and pooling operation of the current layer of the neural network by utilizing the adjustment parameters of the current layer;
Classification processing: and carrying out convolution operation and pooling operation on the windowed signals layer by layer according to the layer sequence of the neural network to obtain signals with fixed lengths, finally finishing classification processing on the signals with fixed lengths through the full-connection layer, and finally outputting classification results.
The invention has the beneficial effects that the neural network classification based on variable-length signal input is realized by utilizing the calculation unit which can adapt to the feature map size in the neural network, so that the robustness and the universality of the neural network classifier are improved, the utilization rate of system resources is reduced, and finally, the overall cost is reduced, and the invention is suitable for the signal classification processing of voice signals, physiological signals and the like with unfixed length.
Drawings
Fig. 1 is a neural network classification system for adaptive variable length signal input.
Fig. 2 is a flow chart of neural network classification for adaptive variable length signal input.
Detailed Description
The electrocardio or language signal is first pre-processed, such as filtering and noise removing, for example, for electrocardio signal classification, QT wave group detection is required for electrocardio signals, the starting and ending positions of a complete heart beat are located, and then the signals are sent to a neural network for classification processing. Or in voice detection of the language signal, the positions of the beginning and the ending of the voice are determined through methods such as a threshold value, so that the length of the language signal to be processed at the later stage is determined. The embodiment system takes the variable-length signal as the input signal of the neural network and cooperates with the self-adaptive feature map size calculation function in the neural network classifier to realize the method. As shown in fig. 1, the system comprises a preprocessing module, an adaptive windowing module and a variable-length input neural network classifier. The variable-length input neural network classifier comprises a neural network control module, a data buffer, a calculation unit and a full-connection module. The computing unit comprises a convolution computing module and a pooling computing module.
The specific implementation steps are shown in fig. 2:
S0, the preprocessing module receives an input continuous original signal, performs preprocessing such as filtering and denoising on the original signal, and outputs preprocessed information to the self-adaptive windowing module.
S1, an adaptive windowing module receives a preprocessed signal, calculates the length of the signal to be processed, on one hand, windows the continuous signal according to the length of the signal to complete adaptive signal windowing, and then outputs the windowed signal to a data buffer in a neural network classifier; on the other hand, the corresponding sizes of all layers of all modules in the neural network classifier are calculated according to the signal length: the convolutional layer feature map Size conv and the pooled layer feature map Size pooling are output to a neural network control module in the neural network classifier.
Sizeconv=(Sizeinput –Sizekernel +2×Padding)/Stride+1;
Sizepooling=(Sizeinput –Sizekernel)/Stride+1;
Wherein Size kernel is the Size of the convolution kernel in the preset neural network classifier, stride is the convolution step Size, and Padding is the filling Size. Size input is the Size of the input signal for the current convolutional layer and the Size of the output signal of the pooling layer of the current layer for the input signal of the next convolutional layer.
And S2, the neural network control module optimizes and adjusts adjustment parameters of the convolution calculation module and the pooling calculation module aiming at the number of registers and operation units used by convolution operation and pooling operation of each layer according to the received characteristic diagram size of the convolution layer and the characteristic diagram size of the pooling layer of each layer, and outputs the layer-by-layer adjustment parameters to the data buffer.
S3, temporarily storing the input adjustment parameters by the data buffer. The network parameters are pre-stored in the data buffer.
S4, performing layer-by-layer calculation by using the neural network classifier, wherein the calculation unit reads parameters and data of a current layer from the data buffer, the parameters comprise adjustment parameters and network parameters of the current layer, and the data are signals subjected to windowing; after parameter tuning is performed on the convolution calculation module and the pooling calculation module by using the adjustment parameters, performing first-layer convolution operation on the windowed signals through the convolution calculation module according to the layer sequence of the neural network, performing first-layer pooling operation through the pooling calculation module, and storing intermediate calculation results of the convolution calculation module and the pooling calculation module into a data buffer for use by a calculation unit of the next-layer neural network. The network parameters include weights and biases.
S5, the neural network classifier performs next-layer calculation, and the calculation unit reads parameters and data of the current layer from the data buffer, wherein the parameters comprise adjustment parameters and network parameters of the current layer, and the data are intermediate calculation results stored in S4; and outputting an intermediate calculation result to a data buffer through a convolution calculation module and a pooling calculation module according to the layer sequence of the neural network for a calculation unit of the next layer of the neural network.
The layer-by-layer operation of the neural network, for example, the first step is a first layer convolution operation, a first layer pooling operation, obtaining an intermediate result 1, and then using the intermediate result 1 as an input to perform a second layer convolution operation and a second layer pooling operation. The data on the arrow of the data buffer pointing to the calculation unit is the windowed signal, i.e. the input signal or the calculation result of the previous layer.
And S6, after the neural network classifier completes convolution operation and pooling operation of all layers, the pooling operation module converts an output result into a signal with a fixed length and outputs the signal with the fixed length to the full-connection module.
And S7, the full connection layer receives signals with fixed lengths, performs classification processing, and finally outputs classification results.
In the working phase, each time a variable length signal is input, the steps S0, S1, S2, S3, S4, S5, S6 and S7 are circulated once.
Claims (5)
1. The neural network classification method for the self-adaptive variable-length signal input is characterized by comprising the following steps of:
Pretreatment: receiving an input continuous original signal, and carrying out filtering and denoising pretreatment on the original signal;
Self-adaptive windowing: calculating the length of a signal to be processed according to the preprocessed signal, on one hand, windowing the continuous signal according to the length of the signal to complete self-adaptive signal windowing processing, and outputting the windowed signal to a data buffer in a neural network classifier; on the other hand, according to the signal length to be processed, the corresponding sizes of all layers in the neural network classifier, namely the characteristic diagram sizes of the convolution layers of different layers and the characteristic diagram sizes of the pooling layers of different layers, are calculated and output to a neural network control module in the neural network classifier; the neural network classifier comprises a neural network control module, a data buffer, a computing unit and a full-connection module;
And (3) self-adaptive adjustment: generating adjustment parameters of a current layer by the neural network classifier according to the received characteristic diagram sizes of the convolution layers of different layers and the characteristic diagram sizes of the pooling layers of different layers, and then completing optimization adjustment of the number of registers and operation units used in the convolution operation and pooling operation of the current layer of the neural network by utilizing the adjustment parameters of the current layer;
classification processing: carrying out convolution operation and pooling operation on the windowed signals layer by layer according to the layer sequence of the neural network to obtain signals with fixed lengths, finally finishing classification processing on the signals with fixed lengths through a full-connection layer, and finally outputting classification results;
The self-adaptive adjustment step and the classification processing step are implemented by a neural network classifier; the neural network classifier comprises the following specific steps: the neural network control module generates adjustment parameters of the current layer according to the received characteristic diagram sizes of the convolution layers and the characteristic diagram sizes of the pooling layers of different layers and outputs the adjustment parameters to the data buffer; the data buffer stores the input adjustment parameters temporarily; the calculation unit reads the adjustment parameters in the data buffer, after optimizing and adjusting the number of registers and operation units used in the convolution operation and pooling operation of the current layer, reads the windowed signals from the data buffer, carries out convolution operation on the windowed signals, carries out pooling operation on the windowed signals to convert the windowed signals into signals with fixed length, and finally outputs the signals with fixed length to the full-connection module; and the full-connection module receives the signals with fixed length, completes classification processing through the full-connection layer, and finally outputs classification results.
2. The method of claim 1, wherein the convolutional layer feature map Size conv and the pooled layer feature map Size pooling are calculated by:
Sizeconv=(Sizeinput – Sizekernel +2×Padding)/Stride+1;
Sizepooling=(Sizeinput – Sizekernel)/Stride+1;
Wherein Size kernel is the Size of the convolution kernel in the preset neural network classifier, stride is the convolution step Size, padding is the Padding Size, and Size input is the Size of the input signal for the current convolution layer.
3. The method of claim 1, wherein the preprocessing step is performed by a preprocessing module.
4. The method of claim 1, wherein the adaptive windowing step is performed by an adaptive windowing module.
5. The method of claim 1, wherein the convolution operation is performed by a convolution calculation module in the calculation unit; the pooling operation is implemented by a pooling calculation module in the calculation unit.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202410194232.9A CN117763399B (en) | 2024-02-21 | 2024-02-21 | Neural network classification method for self-adaptive variable-length signal input |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202410194232.9A CN117763399B (en) | 2024-02-21 | 2024-02-21 | Neural network classification method for self-adaptive variable-length signal input |
Publications (2)
Publication Number | Publication Date |
---|---|
CN117763399A CN117763399A (en) | 2024-03-26 |
CN117763399B true CN117763399B (en) | 2024-05-14 |
Family
ID=90326016
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202410194232.9A Active CN117763399B (en) | 2024-02-21 | 2024-02-21 | Neural network classification method for self-adaptive variable-length signal input |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN117763399B (en) |
Citations (13)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110069958A (en) * | 2018-01-22 | 2019-07-30 | 北京航空航天大学 | A kind of EEG signals method for quickly identifying of dense depth convolutional neural networks |
CN111053549A (en) * | 2019-12-23 | 2020-04-24 | 威海北洋电气集团股份有限公司 | Intelligent biological signal abnormality detection method and system |
CN111210019A (en) * | 2020-01-16 | 2020-05-29 | 电子科技大学 | Neural network inference method based on software and hardware cooperative acceleration |
CN111445420A (en) * | 2020-04-09 | 2020-07-24 | 北京爱芯科技有限公司 | Image operation method and device of convolutional neural network and electronic equipment |
CN111460932A (en) * | 2020-03-17 | 2020-07-28 | 哈尔滨工程大学 | Underwater sound signal classification and identification method based on self-adaptive convolution |
CN111584029A (en) * | 2020-04-30 | 2020-08-25 | 天津大学 | Electroencephalogram self-adaptive model based on discriminant confrontation network and application of electroencephalogram self-adaptive model in rehabilitation |
CN111738427A (en) * | 2020-08-14 | 2020-10-02 | 电子科技大学 | Operation circuit of neural network |
CN111783876A (en) * | 2020-06-30 | 2020-10-16 | 西安全志科技有限公司 | Self-adaptive intelligent detection circuit and image intelligent detection method |
CN114190889A (en) * | 2021-11-19 | 2022-03-18 | 上海联影智能医疗科技有限公司 | Electrocardiosignal classification method and system, electronic equipment and readable storage medium |
CN114692830A (en) * | 2022-03-25 | 2022-07-01 | 潘振华 | Self-strengthening image and voice deep learning model of promotion network |
CN114742225A (en) * | 2022-04-07 | 2022-07-12 | 中国科学院合肥物质科学研究院 | Neural network reasoning acceleration method based on heterogeneous platform |
CN115221926A (en) * | 2022-07-20 | 2022-10-21 | 吉林大学 | Heart beat signal classification method based on CNN-GRU network model |
CN116027911A (en) * | 2023-03-29 | 2023-04-28 | 北京理工大学 | Non-contact handwriting input recognition method based on audio signal |
Family Cites Families (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
KR102043376B1 (en) * | 2018-08-16 | 2019-11-11 | 한국과학기술연구원 | Method for real time analyzing stress using deep neural network algorithm |
CN109560770B (en) * | 2018-10-29 | 2020-03-13 | 西安交通大学 | Photovoltaic system fault arc detection method based on self-adaptive kernel function and instantaneous frequency estimation |
US11755892B2 (en) * | 2019-11-07 | 2023-09-12 | Alibaba Group Holding Limited | Multi-size convolutional layer |
US20210374518A1 (en) * | 2020-05-27 | 2021-12-02 | Nvidia Corporation | Techniques for modifying and training a neural network |
US11558628B2 (en) * | 2020-12-14 | 2023-01-17 | Nokia Technologies Oy | Caching and clearing mechanism for deep convolutional neural networks |
-
2024
- 2024-02-21 CN CN202410194232.9A patent/CN117763399B/en active Active
Patent Citations (13)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110069958A (en) * | 2018-01-22 | 2019-07-30 | 北京航空航天大学 | A kind of EEG signals method for quickly identifying of dense depth convolutional neural networks |
CN111053549A (en) * | 2019-12-23 | 2020-04-24 | 威海北洋电气集团股份有限公司 | Intelligent biological signal abnormality detection method and system |
CN111210019A (en) * | 2020-01-16 | 2020-05-29 | 电子科技大学 | Neural network inference method based on software and hardware cooperative acceleration |
CN111460932A (en) * | 2020-03-17 | 2020-07-28 | 哈尔滨工程大学 | Underwater sound signal classification and identification method based on self-adaptive convolution |
CN111445420A (en) * | 2020-04-09 | 2020-07-24 | 北京爱芯科技有限公司 | Image operation method and device of convolutional neural network and electronic equipment |
CN111584029A (en) * | 2020-04-30 | 2020-08-25 | 天津大学 | Electroencephalogram self-adaptive model based on discriminant confrontation network and application of electroencephalogram self-adaptive model in rehabilitation |
CN111783876A (en) * | 2020-06-30 | 2020-10-16 | 西安全志科技有限公司 | Self-adaptive intelligent detection circuit and image intelligent detection method |
CN111738427A (en) * | 2020-08-14 | 2020-10-02 | 电子科技大学 | Operation circuit of neural network |
CN114190889A (en) * | 2021-11-19 | 2022-03-18 | 上海联影智能医疗科技有限公司 | Electrocardiosignal classification method and system, electronic equipment and readable storage medium |
CN114692830A (en) * | 2022-03-25 | 2022-07-01 | 潘振华 | Self-strengthening image and voice deep learning model of promotion network |
CN114742225A (en) * | 2022-04-07 | 2022-07-12 | 中国科学院合肥物质科学研究院 | Neural network reasoning acceleration method based on heterogeneous platform |
CN115221926A (en) * | 2022-07-20 | 2022-10-21 | 吉林大学 | Heart beat signal classification method based on CNN-GRU network model |
CN116027911A (en) * | 2023-03-29 | 2023-04-28 | 北京理工大学 | Non-contact handwriting input recognition method based on audio signal |
Non-Patent Citations (2)
Title |
---|
A Lightweight Neural Network Based Respiratory Rate Estimation Approach Using PPG Signal;Xu Wang 等;2023 6th International Conference on Electronics Technology (ICET);20230815;1446-1449 * |
面向智能心电监测系统的自适应可重构心拍检测硬件研究;祝镇;《中国优秀硕士学位论文全文数据库 基础科学辑》;20230115;2023年(第01期);第A006-796页正文第三-四章 * |
Also Published As
Publication number | Publication date |
---|---|
CN117763399A (en) | 2024-03-26 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109410974B (en) | Voice enhancement method, device, equipment and storage medium | |
CN110097172B (en) | Convolutional neural network data processing method and device based on Winograd convolutional operation | |
US20230334632A1 (en) | Image recognition method and device, and computer-readable storage medium | |
CN111814975B (en) | Neural network model construction method and related device based on pruning | |
CN113570508A (en) | Image restoration method and device, storage medium and terminal | |
CN114157539B (en) | Data-aware dual-drive modulation intelligent identification method | |
CN110348564B (en) | SCNN reasoning acceleration device based on systolic array, processor and computer equipment | |
CN109460813B (en) | Acceleration method, device and equipment for convolutional neural network calculation and storage medium | |
CN111950715A (en) | 8-bit integer full-quantization inference method and device based on self-adaptive dynamic shift | |
CN113205048A (en) | Gesture recognition method and system | |
CN115941112B (en) | Portable hidden communication method, computer equipment and storage medium | |
CN113298235A (en) | Neural network architecture of multi-branch depth self-attention transformation network and implementation method | |
CN117763399B (en) | Neural network classification method for self-adaptive variable-length signal input | |
CN115328661B (en) | Computing power balance execution method and chip based on voice and image characteristics | |
CN111667819B (en) | Voice recognition method, system, storage medium and electronic equipment based on CRNN | |
CN117494762A (en) | Training method of student model, material processing method, device and electronic equipment | |
CN113435586B (en) | Convolution operation device and system for convolution neural network and image processing device | |
Lee et al. | MPQ-YOLACT: Mixed-Precision Quantization for Lightweight YOLACT | |
US20200259504A1 (en) | Decompression of model parameters using functions based upon cumulative count distributions | |
CN111354372B (en) | Audio scene classification method and system based on front-end and back-end combined training | |
CN113255446B (en) | Face detection system | |
US20230409869A1 (en) | Process for transforming a trained artificial neuron network | |
CN113673690B (en) | Underwater noise classification convolutional neural network accelerator | |
WO2020051751A1 (en) | Convolution neural network computing acceleration method and apparatus, device, and storage medium | |
CN117236388A (en) | Transformation process of trained artificial neuron network |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |