A Configurable BP Neural Network Accelerator for Laser Welding Parameter Calculation

Boyu Fan; Zaifeng Shi; Zhe Wang; Shaoxiong Li; Tao Luo

doi:10.3788/LOP202259.0214001

Journals >Laser & Optoelectronics Progress >Volume 59 >Issue 2 >Page 0214001 > Article

Laser & Optoelectronics Progress
Vol. 59, Issue 2, 0214001 (2022)

A Configurable BP Neural Network Accelerator for Laser Welding Parameter Calculation

Boyu Fan^1、*, Zaifeng Shi^1、3、**, Zhe Wang¹, Shaoxiong Li¹, and Tao Luo²

Author Affiliations

¹School of Microelectronics, Tianjin University, Tianjin 300072, China

²College of Intelligence and Computing, Tianjin University, Tianjin 300072, China

³Tianjin Key Laboratory of Microelectronic Technology for Imaging and Sensing, Tianjin 300072, China

show less

DOI: 10.3788/LOP202259.0214001 Cite this Article Set citation alerts

Boyu Fan, Zaifeng Shi, Zhe Wang, Shaoxiong Li, Tao Luo. A Configurable BP Neural Network Accelerator for Laser Welding Parameter Calculation[J]. Laser & Optoelectronics Progress, 2022, 59(2): 0214001 Copy Citation Text

show less

Fig. 1. Parallel acceleration technology of BP neural network

Download full size

Fig. 2. Accelerator architecture and multi-level storage of data

Download full size

Fig. 3. Data transmission module of operation matrix

Download full size

Fig. 4. Structure of our accelerator specific verification platform

Download full size

Fig. 5. Running time comparison between accelerator and embedded processor platform

Download full size

Fig. 6. Running time comparison of accelerator and embedded processor platform under typical number neurons and maximum number of neurons. (a) Typical number of neurons; (b) maximum number of neurons

Download full size

Literature	Application direction	Network model used	Result
Literature［9］	Laser induced breakdown spectroscopy	Radial basis function neural network	The accuracy of some elements is improved
Literature［10］	Optimization of laser cutting parameters	Convolution neural network	About 92% accuracy
Literature［11］	Laser additive manufacturing control	Alex net	Effectively used in the process of image segmentation process
Literature［12］	Analysis of laser ranging data	Deep neural network	It is impossible for network to find deep information from satellite laser ranging data
Literature［13］	Spectral analysis	BP neural network	The modeling effect is improved
Literature［14］	Color laser marking	BP neural network	The feasibility is demonstrated
Literature［15］	Intensity calibration of laser scanner	BP neural network	The system response time is effectively shortened
Literature［16］	Laser induced breakdown spectroscopy	BP neural network	The results are satisfactory
Literature［17］	Target detection of optical genetic laser projection system	Convolution neural	Highly accurate detection effect is realized
Literature［2］	Optimization of laser welding parameters	BP neural network	The design goals of high precision，high quality，and high stability are realized
Literature［18］	Optimization of laser welding parameters	BP neural network	The relative error is small and the effect is good
Literature［19］	Laser welding control	Cellular neural network	The algorithm complexity are reduced and the control rate is high
Literature［20］	Optimization of laser cutting parameters	BP neural network	It has achieved obvious success

Table 1. Typical application of artificial neural network in laser technology in recent years

Scanning speed /（m⋅min^-1）	Power /W	Defocus /mm	Weld width /μm	Penetration /μm	Ratio of penetration to weld width
1.3	1000	10	1037.3130	1985.92	1.914484
1.5	800	0	733.7967	1859.60	2.534217
1.5	1000	3	1482.8600	1989.30	1.341529
1.8	1000	0	819.4533	1976.73	2.412255
2.5	1500	3	1166.8770	1958.60	1.678498
3.0	1500	0	776.9867	1858.92	2.392473
5.3	2000	3	617.0433	1900.31	3.079703
5.5	2000	0	597.3000	1855.23	3.106027
7.0	2500	0	676.2867	1856.02	2.744428

Table 2. Typical mapping relationship of parameters in data set

Number of neurons			Processing time of bp1 t_bp1 /ms		Processing time of bp2 t_bp2 /ms
Input layer	Hidden layer	Output layer	Accelerator	Embedded processor platform	Accelerator	Embedded processor platform
3	4	3	0.651360	3.217697	0.557184	2.809763
3	6	4	0.645399	4.746675	0.546455	5.123854
3	8	3	0.649691	6.266117	0.546694	4.949093
3	9	3	0.665188	7.077932	0.545502	5.591869
3	9	4	0.668287	6.937027	0.582457	7.375479
3	10	2	0.664234	7.764816	0.548601	4.140139
3	10	3	0.663519	7.673979	0.619888	6.249666
3	10	4	0.653505	7.800817	0.561714	8.289099
4	4	3	0.648975	4.116058	0.545502	2.835274
4	6	4	0.647545	5.889177	0.549316	5.116224
4	8	3	0.658512	7.790327	0.544310	5.586386
4	9	3	0.666857	8.561611	0.621557	5.743742
4	9	4	0.665903	8.777618	0.571012	7.631302
4	10	2	0.678539	9.542465	0.555515	4.715443
4	10	3	0.666857	9.572506	0.617504	6.770372

Table 3. Comparison of accelerator processing speed

Number of neurons			Processing time of bp1 t_{bp1_a} /ms		Processing time of bp2 t_{bp2_a} /ms		Proportion
Input layer	Hidden layer	Output layer	General accelerator	Proposed accelerator	General accelerator	Proposed accelerator	Proportion
16	32	16	1.061678	0.909328	0.926495	0.851393	0.8877
16	32	32	0.994444	0.916243	1.302719	1.132965	0.8955
32	64	32	2.145052	1.786947	2.032518	1.715899	0.8386
32	64	64	2.478838	1.778364	3.809452	2.868414	0.7352
64	128	64	6.696224	5.238771	6.526232	5.165577	0.7869
64	128	128	6.678343	5.263567	12.558460	9.703398	0.7804

Table 4. Comparison of processing speed between accelerators

Boyu Fan, Zaifeng Shi, Zhe Wang, Shaoxiong Li, Tao Luo. A Configurable BP Neural Network Accelerator for Laser Welding Parameter Calculation[J]. Laser & Optoelectronics Progress, 2022, 59(2): 0214001

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information