correlations (r 2 =0.87 and r 2 =0.

correlations (r 2 =0.87 and r 2 =0.89, respectively) and higher prediction errors. This may be related to the fact thatthe low NH3 sensor response inadequately reflected the high TVB-N produc- tion observed and the corresponding increasing Pp counts. For comparison with traditional methods a model based on TVB-N, TVC and the time temperature variable gave a slightly inferior model (r2 =0.91 and RMSEP=0.54) (Table 3). However, when using TVB-N instead of the electronic nose sensors with the same combination of the SSO, models with higher correlations were obtained(r2=0.94and RMSEP=0.43to 0.44). Thisindicates that TVB-N is highly relevant to predict the sensory quality in com- bination with the SSO, although, when used as a single quality criterion shorter shelf-life was predicted compared to sensory analysis.
4.2.3. Validation of the PLSR prediction of sensory quality
Prediction of sensory quality (Torry score) was done using PLSR models with different subsets of the data and a test set with one sample group (0°C +abuse) toverifythe abilityofthe models to predict quality of unknown samples. After exploring dif- ferent models five variables that gave the best correlation for the whole data set(three electronic nose sensors (CO; NH3 and H2S), pseudomonads counts and Tacc) were selected for a PLSR model based ona subsetofthe data (N=18)with samples from0°C,7°C and 15 °C storage (2001 experiments) and samples stored at 0 °C in 2003. This model was similar (r2 =0.94; RMSEP=0.49) as the one calculated with the whole data set (r2=0.93; RMSEP=0.49, see Table 3). Models with the other SSOs instead of the pseudomonads had lower correlation and higher errors (H2S- producing bacteria counts r2 =0.90 and RMSEP=0.54; Pp counts r2=0.88 and RMSEP=0.59) suggesting the importance of pseudomonads in the spoilage of haddock fillets. A biplot of the scores and loadings for the model illustrates well that the PC2 is explaining the influence of the temperature as described before and the temperature abused samples are on the upper half while the samples stored at 0°C are located on the lower half,indicating the different spoilage pattern (Fig. 6). The prediction of the Torry score for the test samples (N=5) from (0 °C + abuse) group

Kết quả (Việt) 1:

mối tương quan (r 2 = 0.87 và r 2 = 0.89, tương ứng) và cao hơn dự báo lỗi. Điều này có thể được liên quan đến một thực tế rằng các phản ứng cảm biến NH3 thấp không đầy đủ phản ánh cao TVB-N sản phẩm-tion quan sát và tương ứng tăng Pp đếm. Để so sánh với phương pháp truyền thống, một mô hình dựa trên TVB-N, phim quảng cáo và thay đổi nhiệt độ thời gian cho một mô hình hơi kém (r2 = 0.91 và RMSEP = 0,54) (bảng 3). Tuy nhiên, khi sử dụng TVB-N thay vì các cảm biến điện tử mũi với sự kết hợp cùng của SSO, các mô hình với mối tương quan cao hơn đã được (r2 = 0.94and RMSEP = 0.43to 0,44). Thisindicates TVB-N là có liên quan cao để dự đoán chất lượng cảm giác trong com-bination với SSO, mặc dù, khi được sử dụng như là một tiêu chí duy nhất chất lượng thọ ngắn hơn đã được dự đoán so với cảm giác phân tích.4.2.3. xác nhận dự đoán PLSR cảm giác chất lượngPrediction of sensory quality (Torry score) was done using PLSR models with different subsets of the data and a test set with one sample group (0°C +abuse) toverifythe abilityofthe models to predict quality of unknown samples. After exploring dif- ferent models five variables that gave the best correlation for the whole data set(three electronic nose sensors (CO; NH3 and H2S), pseudomonads counts and Tacc) were selected for a PLSR model based ona subsetofthe data (N=18)with samples from0°C,7°C and 15 °C storage (2001 experiments) and samples stored at 0 °C in 2003. This model was similar (r2 =0.94; RMSEP=0.49) as the one calculated with the whole data set (r2=0.93; RMSEP=0.49, see Table 3). Models with the other SSOs instead of the pseudomonads had lower correlation and higher errors (H2S- producing bacteria counts r2 =0.90 and RMSEP=0.54; Pp counts r2=0.88 and RMSEP=0.59) suggesting the importance of pseudomonads in the spoilage of haddock fillets. A biplot of the scores and loadings for the model illustrates well that the PC2 is explaining the influence of the temperature as described before and the temperature abused samples are on the upper half while the samples stored at 0°C are located on the lower half,indicating the different spoilage pattern (Fig. 6). The prediction of the Torry score for the test samples (N=5) from (0 °C + abuse) group