| Photosynthesis is the basis of plant growth and development and crop yield formation,photosynthetic rate is an important indicator of photosynthesis,and plant photosynthesis monitor is the main instrument to measure the photosynthetic rate of plants.The plant photosynthesis monitors use a traditional windowed keyboard for control,and the user needs to follow the strict requirements and specifications of the control interface to interact with the photosynthesis monitors.Natural language interaction is a popular area of human-computer interaction technology in recent years.This interaction method is more in line with human mind habits and communication styles,making the interaction between humans and instruments more natural and efficient.Based on this,this study aims to improve the intelligent interaction capability of the plant photosynthesis monitor and simplify the operation process of the instrument,and investigate its natural language interaction method based on the self-developed i GH-100 plant photosynthesis monitor.The natural language interface i GH-NLI was constructed to solve the natural language parsing of real-time control and data query commands of the plant photosynthesis monitor,and the interface was divided into two modules of command recognition and structured query language(SQL)generation for its functional and operational characteristics.The main research of this thesis contains the following aspects:1)An instruction recognition model based on intention recognition and slot filling is proposed.Different instruction recognition model implementation methods are compared and the superiority of the intent recognition and slot filling methods for the instruction recognition task is determined.To address the problem of difficult matching of intent and slot in the separate modeling of intent recognition and slot filling tasks,a BERT model is introduced for joint training of the two tasks,and an attention mechanism is incorporated to incorporate the captured intent vector into the slot vector to achieve the display association between the two subtasks.From the experimental results,it can be seen that the joint training model of BERT+Attention has the best results on the slot-filling F1 values,reaching 94.6%,and the intention accuracy has reached 96.2%.2)The SQL generation model of X-SQL + two-layer Bi-LSTM is constructed.A model structure similar to X-SQL is built as the baseline model for the SQL generation module,and a two-layer Bi-LSTM structure is introduced to further extract the instruction text feature information for the shortcomings of the baseline model in the where value extraction task.Experiments show that the SQL generation model with the introduction of a two-layer Bi-LSTM has the highest improvement in logical accuracy and execution accuracy,reaching 86.2% and 89.3%,respectively.3)A joint training model for command recognition and SQL generation was designed.For the characteristics of the similarity of semantic output features of the coding layers of the command recognition and SQL generation modules,the BERT pre-trained model is used as a shared coding layer for both tasks,and the two tasks of command recognition and SQL generation are jointly trained in a multi-task format.The analysis compares the advantages and disadvantages of jointly trained and individually trained models in terms of performance and model size.4)Developed i GH-HMI,a natural language interaction system for the plant photosynthesis instrument i GH-100.Py Qt5 and Flask framework are used to develop the front and back end of the system,My SQL is used for persistent storage of photosynthesis data,and Modbus-TCP protocol is used to transmit the data collected by the collection module.The system integrates the i GH-NLI natural language interface module,which can better enhance the ease of use of the photosynthesis monitor. |
