(重慶大學 應(yīng)用物理系，重慶 400044)
摘 要: 根據(jù)7005鋁合金在不同工藝參數(shù)(擠壓溫度、擠壓速度、淬火方式和時效條件)下的力學性能(抗拉強度σb、屈服強度σ0.2和硬度HB)實測數(shù)據(jù)集，應(yīng)用基于粒子群算法(PSO)尋優(yōu)的支持向量回歸(SVR)結(jié)合留一交叉驗證(LOOCV)的方法，對7005鋁合金力學性能進行建模和預(yù)測研究，并與偏最小二乘法(PLS)、反向傳播人工神經(jīng)網(wǎng)絡(luò)(BPNN)和兩者結(jié)合的PLS-BPNN模型的預(yù)測結(jié)果進行比較。結(jié)果表明：基于SVR-LOOCV法的預(yù)測精度最高，對3種力學性能(σb、σ0.2和HB)預(yù)測的均方根誤差(RMSE)分別為4.531 9 MPa、14.550 8 MPa和HB1.414 2，其平均相對誤差(MRE)分別為0.72%、2.61%和0.66%，均比PLS、BPNN和PLS-BPNN方法預(yù)測的RMSE和MRE要小。
關(guān)鍵字: 7005鋁合金；力學性能；支持向量機；粒子群算法；留一交叉驗證法；回歸分析
（Department of Applied Physics, Chongqing University, Chongqing 400044, China）
Abstract:The support vector regression (SVR) approach based on the particle swarm optimization (PSO) for its parameter optimization, combined with leave-one-out cross validation (LOOCV), was proposed to predict the mechanical properties (tensile strength σb, yield strength σ0.2 and hardness HB) of 7005 Al alloys under different processing parameters including extrusion temperature, extrusion velocity, quenching type and aging time. The results strongly support that the prediction precision of SVR-LOOCV method is superior to those of partial least squares (PLS), back-propagation neural networks (BPNN) and their combination PLS-BPNN model by applying the identical dataset. The root mean square errors (RMSE) for σb, σ0.2 and HB achieved by SVR-LOOCV are 4.531 9 MPa, 14.550 8 MPa and HB 1.414 2, respectively, and their mean relative errors (MRE) are 0.72%, 2.61% and 0.66%, respectively, which are less than those predicted by PLS, BPNN or PLS-BPNN approach.
Key words: 7005 Al alloys; mechanical properties; support vector machines; particle swarm optimization; leave-one-out cross validation; regression analysis
