随着智能化煤矿建设的逐步推进，构建适用于煤矿安全生产的语音问答交互系统具有重要意义。目前，普通话语音识别技术已经十分成熟，但具有地域性质的方言语音识别研究仍处于发展阶段。陕北地区受黄土高原区域影响，当地大多采用陕北方言作为主要的沟通交流方式。同时陕北地区作为我国煤炭资源最丰富的区域之一，在煤矿生产、会议召开、调度指挥等工作中也多使用陕北方言。

本文基于煤矿安全规程构建陕北方言语音语料库，使用端到端模型和自然语言处理方法，展开基于陕北方言提问的煤矿安全规程智能问答算法研究，主要工作如下：

（1）构建基于煤矿安全规程陕北方言语音语料库。首先，深入研究陕北方言的语音特点、语法特点及词汇特点；其次设计构建陕北方言语音语料库，基于煤矿安全规程选取文本语料，进行录制、核对及标注；最后，构建基于煤矿安全规程的陕北方言语音语料库。

（2）为有效提高基于煤矿安全规程的陕北方言语音识别率，提出基于改进的Conformer端到端陕北方言语音识别模型。首先，针对陕北方言特点设计预处理模块，有效提取陕北方言中独特的发音模式，适应陕北方言语调和语序的变换，理解陕北方言中独特的词汇和表达方式；其次，针对Conformer模型文本生成能力不足，使用Transformer作为解码器，提高文本的生成能力，同时引入CTC（Connectionist Temporal Classification，CTC）联合训练，减少模型计算，提高模型性能。实验结果表明，与其他主流模型相比，基于改进的Conformer端到端陕北方言语音识别模型降低了错字率。

（3）为有效提高基于煤矿安全规程问答模型的文本理解力，提出基于改进的RoBERTa-wwm煤矿安全规程问答模型。首先，基于前文建立的陕北方言语音语料库分析整理其中的煤矿安全规程语料文本，人工标注少量问答对，结合RoBERTa-wwm与UniLM，提出基于改进的RoBERTa-wwm问答对自动生成算法，构建数量更大的煤矿安全规程问答对数据集；其次，针对无法回答问题和无关问题，引入问题相似度机制，构建基于问题相似度机制的答案生成模型，使其只关注可回答问题，提升模型的推理能力。实验结果表明，生成的煤矿安全规程问答对，相较于其他方法具有更好的文本生成能力；同时提高了模型识别无法回答问题和无关问题的效率。

综上所述，本文主要研究了在陕北煤矿安全生产场景中执行方言识别与智能问答任务，并基于此对端到端相关算法展开了深入研究，研究成果对于不同场景下执行不同类型任务时的方言识别与智能问答具有一定的理论意义和应用价值。

With the gradual advancement of the construction of intelligent coal mine, it is of great significance to construct a voice question and answer interactive system suitable for coal mine safety production. At present, Putonghua speech recognition technology has been very mature, but the regional dialect speech recognition research is still in the development stage. Northern Shaanxi is affected by the Loess Plateau, most of the local people adopt northern Shaanxi dialect as the main way of communication. At the same time, as one of the regions with the most abundant coal resources in China, northern Shaanxi also uses northern Shaanxi dialect in coal mine production, meeting, dispatching and commanding.

In this paper, a speech corpus of northern Shaanxi dialect is constructed based on coal mine safety regulations. An end-to-end model and natural language processing method are used to carry out the research on intelligent question answering algorithm of coal mine safety regulations based on northern Shaanxi dialect. The main work is as follows:

(1) Construct a speech corpus of Northern Shaanxi dialect based on coal mine safety regulations. Firstly, the paper studies the phonological features, grammatical features and lexical features of northern Shaanxi dialect. Secondly, we designed and constructed the speech corpus of northern Shaanxi dialect, selected the text corpus based on the coal mine safety regulations, recorded, checked and marked it. Finally, the paper constructs a phonetic corpus of northern Shaanxi dialect based on coal mine safety regulations.

(2) In order to effectively improve the speech recognition rate of northern Shaanxi dialect based on coal mine safety regulations, an end-to-end speech recognition model of northern Shaanxi dialect based on improved Conformer was proposed. Firstly, according to the characteristics of northern Shaanxi dialect, a pre-processing module is designed to effectively extract the unique pronunciation patterns in northern Shaanxi dialect, adapt to the changes of intonation and word order, and understand the unique vocabulary and expression modes in northern Shaanxi dialect. Secondly, in view of the insufficient text generation ability of Conformer model, Transformer is used as a decoder to improve the text generation ability. At the same time, Connectionist Temporal Classification (CTC) joint training is introduced to reduce model calculation. Improve model performance. The experimental results show that compared with other mainstream models, the improved Conformer end-to-end speech recognition model reduces the error rate of Shaanxi dialect.

(3) In order to effectively improve the text understanding of the coal mine safety regulation Q&A model, the improved RoBERTa-wwm coal mine safety regulation Q&A model is proposed. Firstly, the coal mine safety regulation corpus text was analyzed and sorted out based on the northern Shaanxi dialect speech corpus established above, and a small number of question and answer pairs were manually labeled. Combining RoBERTa-wwm and UniLM, an improved RoBERTa-wwm question and answer generation algorithm was proposed to build a larger number of coal mine safety regulation question and answer datasets. Secondly, for unanswerable questions and irrelevant questions, a problem similarity mechanism is introduced to build an answer generation model based on problem similarity mechanism, so that it only focuses on answerable questions and improves the reasoning ability of the model. The experimental results show that the generated questions and answers of coal mine safety regulations have better text generation ability than other methods. At the same time, the efficiency of model identification of unanswerable questions and irrelevant questions is improved.

To sum up, this paper mainly studies the implementation of dialect recognition and intelligent question answering tasks in the safety production scene of coal mine in northern Shaanxi Province, and conducts in-depth research on the end-to-end relevant algorithms based on this. The research results have certain theoretical significance and application value for dialect recognition and intelligent question answering in different scenarios and different types of tasks.