Pipe network engineering field expert experience guiding method based on voice recognition
By building a lightweight voice recognition model in the field of pipeline engineering, the data collection problem is solved, the efficient collection and sharing of expert experience is achieved, the standardization and systematization of construction knowledge is promoted, and the efficiency and quality of engineering management is improved.
Patent Information
- Application Number
- CN202510589929.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-08
- Publication Date
- 2025-08-22
AI Technical Summary
In the field of pipeline engineering, the application of speech recognition technology faces the difficulty of data collection, and requires more data support to train accurate speech recognition models. The traditional knowledge acquisition method is inefficient, and expert experience is difficult to systematically and share.
Using an expert experience guidance method in the field of pipeline network engineering based on speech recognition, a lightweight speech recognition model is constructed by collecting field speech data, using WeNet framework and Conformer model for training, and combining knowledge distillation technology, a rule base is constructed for knowledge matching analysis.
It has achieved efficient collection and sharing of expert experience, improved the standardization and systematization of construction knowledge, and improved the efficiency and quality of engineering management.
Smart Images

Figure CN120525031A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of infrastructure construction, and in particular to a method for guiding expert experience in the field of pipe network engineering based on speech recognition. Background Art
[0002] The trend toward intelligence and the rise of the knowledge age have driven an increasing emphasis on knowledge resources within businesses, particularly in the construction sector. Construction knowledge is the core foundation for safe and efficient project execution and a vital asset for businesses. However, much of this knowledge resides in the personal experience of experts, inherently implicit and personalized, making it difficult to systematize and share. Traditional knowledge acquisition methods typically involve interviewing experts or completing questionnaires, which are then compiled and stored in knowledge bases by knowledge engineers. These processes are time-consuming, labor-intensive, and inefficient.
[0003] With the continuous development of computer technology, scholars at home and abroad have begun to use computers to design special knowledge acquisition tools. This type of knowledge acquisition method has improved the degree of automation of expert experience knowledge acquisition to a certain extent, but most of them use fixed editing interfaces or question programs. The operation steps are cumbersome and require experts to have a certain level of computer skills. The system is also unable to provide effective feedback on knowledge entry in a timely manner.
[0004] To address this issue, intelligent methods have become an effective way to capture expert construction experience. Speech recognition technology, due to its efficiency and convenience, has become a popular tool. However, its application in pipeline engineering faces certain challenges. Voice data collection in this area is difficult, and more data is needed to train accurate speech recognition models. Therefore, it is necessary to collect extensive domain voice data to train lightweight speech recognition models, improve expert speech recognition, and assist in the experience collection process. Summary of the Invention
[0005] The purpose of the present invention is to provide a method for guiding expert experience in the field of pipeline engineering based on voice recognition. The automated knowledge matching mechanism of the present invention further ensures the compliance of expert experience with construction specifications, and helps to achieve the standardization and systematization of construction experience; through the intelligent generation of guiding questions, the expert experience can be guided more deeply and comprehensively, promoting the efficient sharing and application of construction knowledge, and greatly improving the efficiency and quality of enterprises in engineering management.
[0006] In order to achieve the above technical features, the purpose of the present invention is to provide a method for guiding expert experience in the field of pipe network engineering based on speech recognition, the method comprising: Taking typical construction plans and construction techniques as examples, this paper analyzes the general content structure of these two types of construction experiences. Considering the specific implementation ideas and related technologies of expert construction experience guidance, a voice-based expert construction experience intelligent guidance framework is proposed; Collect domain voice data of domain-specific words. The correct text data corresponding to each audio file is saved in txt format, corresponding one-to-one with the audio data. Data expansion and enhancement of speech datasets; The WeNet framework was used to train the speech recognition model. The model trained on the large-scale Aishell-1 Chinese speech dataset under the WeNet framework was used as the basic model for transfer learning. A speech recognition model was established using transfer learning, with the Conformer model as the encoder and the CTC-Attention mechanism as the joint decoding mechanism. Use knowledge distillation to obtain a lightweight speech recognition model; Construct a rule-based construction knowledge matching method; Based on the constructed rule base and knowledge base, the problem of comparing the key points of expert experience knowledge with the specification provisions is transformed into a construction experience knowledge matching analysis problem.
[0007] Preferably, the analysis of the content structure based on the two types of construction experience, typical construction schemes and construction techniques, specifically includes: Analysis of the knowledge structure of construction experience in construction plan: The key information of construction plan experience knowledge includes: project overview, main construction methods and processes, and construction guarantee measures; Analysis of the knowledge framework of construction experience in construction technology: mainly includes the following contents: scope of application, pre-construction preparation, process flow and construction requirements, safety, quality, and finished product protection issues; among them, the focus includes the applicable conditions of construction technology, preliminary preparation work, specific operating procedures and requirements, and measures to deal with possible safety, quality and finished product protection challenges.
[0008] Preferably, the specific implementation ideas and related technologies of the expert construction experience guidance are considered, and the voice-based expert construction experience intelligent guidance framework is proposed, which specifically includes: Expert construction experience speech recognition: The domain speech recognition model can more accurately recognize expert construction experience speech containing domain-specific terms, providing a more convenient way to enter experience; Expert construction experience knowledge content positioning: Construction experience knowledge has a certain content structure. It is necessary to timely analyze the subject level of expert construction experience knowledge and locate it in specific modules in the guidance template as an aid to judge the integrity of expert experience knowledge, so as to ensure the systematicness and integrity of experience knowledge as much as possible; Matching analysis of key knowledge elements of expert construction experience: Construction experience knowledge involves some fine-grained construction knowledge points, which may differ from the original specifications. To achieve further analysis of the key knowledge points of expert construction experience, it is necessary to match the expert experience with the knowledge in the original construction knowledge base.
[0009] Preferably, the domain voice data of the domain professional words are collected, and the correct text data corresponding to each audio file is saved in txt format, corresponding one to one with the audio data, specifically including: The data was expanded using the speech rate change method commonly used to expand speech training data, and processed using the Audacity tool. A total of 0.9 and 1.1 speech data with different speeds were added, while maintaining the same data format.
[0010] Preferably, the voice data set data expansion and enhancement includes: The voice data of relevant experts from the construction unit who record construction experience are obtained, and the audio data is converted into a unified audio format: a mono WAV format with a sampling frequency of 16kHz and a sampling size of 16 bits.
[0011] Preferably, the WeNet framework is used for speech recognition model training. The model trained on the large-scale Aishell-1 Chinese speech dataset under the WeNet framework is used as the basic model for transfer learning, specifically including: Download and prepare the Aishell-1 Chinese speech dataset, including audio files and corresponding text labels; divide the dataset into training, validation, and test sets to ensure data balance and diversity; Use the WeNet framework to build a basic speech recognition model, choosing an existing structure or customizing it as needed. Use the Aishell-1 dataset as training data and optimize the model parameters through the training process to enable it to accurately convert speech signals into corresponding text. The model trained on the Aishell-1 dataset is used as the basic model for transfer learning and fine-tuned on datasets related to water supply and drainage pipeline construction to adapt to new fields and tasks.
[0012] Preferably, a speech recognition model is established using a transfer learning method, using a Conformer model as an encoder and a CTC-Attention mechanism for joint decoding, specifically including: During the encoding phase, the audio is first processed using a convolutional downsampling layer, and then processed using multiple Conformer modules. The multi-head attention mechanism module replaces the positional encoding in Transformer with the relative sinusoidal positional encoding in Transformer-XL. This allows the attention module to generalize better across different input lengths, and the resulting encoder is more robust to changes in audio sequence length. The convolution module introduces point-by-point convolution, which uses a 1×1 convolution kernel to perform convolution on each point of the feature map and a linear gated unit (GLU) to activate the number of projection channels. The two form a gating mechanism, followed by a one-dimensional depth-separated convolution, which improves model efficiency. Batchnorm is added for normalization to train deeper models. CTC-Attention joint decoding combines the advantages of the connectionist temporal classification model and the Attention model to better process the feature sequence of the speech signal. It uses the CTC model to force the alignment of the input speech signal with the output, then uses the shared network to transfer parameters and uses the Attention model to model and predict the predicted candidate text sequence to obtain an attention vector. Finally, the predicted attention vector is jointly decoded with the input speech signal to generate the final prediction result.
[0013] Preferably, based on the speech recognition model obtained by transfer learning, a knowledge distillation method is used to obtain a lightweight speech recognition model, which maintains high model accuracy while making the model lightweight.
[0014] Preferably, a rule-based construction knowledge matching method is constructed, including: Gain an in-depth understanding of the content of the construction experience text and the referenced normative provisions, identify the key points that require matching and analysis in the construction experience text, and grasp the characteristics of the construction normative text; Using the regular expression rule base and the knowledge graph rule base, we designed a construction experience knowledge matching algorithm to achieve matching analysis of expert construction experience. The matching analysis results of the system are output according to the module, so that experts can respond to corresponding questions in a timely manner based on the system output. Since the construction method content is mostly keywords and corresponding simple numerical values, the rules in the regular expression are used for matching analysis to take the key information extracted from the construction method screening condition text block and the construction method selected in the specific construction process text block as the analysis objects, and then use the set rules to analyze whether the construction method is applicable, or whether there are multiple applicable methods but the expert selected one of them, and then output the system feedback prompts.
[0015] Preferably, for the established rule base and knowledge base, the problem of comparing the key points of expert experience knowledge with the provisions of the specification is transformed into a construction experience knowledge matching analysis problem, including: The relevant rules constructed using the knowledge graph are used to perform matching analysis on each key construction element. The text under each topic after positioning is passed as input to the rule matching analysis module, and then the rules under the module are used for reasoning. If it does not meet the corresponding rule requirements, a prompt will be output.
[0016] The present invention has the following beneficial effects: 1. Through speech recognition technology and intelligent guidance mechanisms, the efficiency of collecting and sharing construction experience has been significantly improved. Compared to traditional manual interviews and questionnaire surveys, this method can automatically transform experts' implicit experience into explicit knowledge and achieve high-precision speech analysis through a lightweight speech recognition model, making it particularly suitable for pipeline network engineering. By leveraging transfer learning and knowledge distillation techniques, the speech recognition model maintains excellent performance while reducing computing resource consumption.
[0017] 2. This invention's automated knowledge matching mechanism further ensures the alignment of expert experience with construction specifications, helping to standardize and systematize construction experience. By intelligently generating guiding questions, expert experience can be guided more deeply and comprehensively, promoting the efficient sharing and application of construction knowledge and significantly improving the efficiency and quality of enterprise project management. BRIEF DESCRIPTION OF THE DRAWINGS
[0018] The present invention will be further described below with reference to the accompanying drawings and examples.
[0019] Figure 1 It is the general content framework of the experience knowledge of the construction plan of the present invention.
[0020] Figure 2 It is the general content framework of the construction process experience knowledge of the present invention.
[0021] Figure 3 This is the intelligent guidance framework of expert construction experience based on voice of the present invention.
[0022] Figure 4 This is the structural diagram of the multi-head attention module of the present invention.
[0023] Figure 5 This is the structural diagram of the convolution module of the present invention.
[0024] Figure 6 This is the knowledge distillation framework for the speech recognition model of the present invention.
[0025] Figure 7 This is the overall framework of the construction experience knowledge matching analysis method of the present invention.
[0026] Figure 8 This is the overall process of the rule matching analysis algorithm of the present invention. DETAILED DESCRIPTION
[0027] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. The components of the embodiments of the present invention generally described and shown in the drawings herein can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.
[0028] To further illustrate the invention process, features and application methods of the present invention, the following examples are given as examples and described as follows: See also Figure 1-8 The present invention provides a method for guiding expert experience in the field of pipe network engineering based on speech recognition, comprising the following steps: Step A: Taking the two typical construction experiences, construction plan and construction technology, as examples, analyze their general content structure.
[0029] Step B: Considering the specific implementation ideas and related technologies of expert construction experience guidance, a voice-based expert construction experience intelligent guidance framework is proposed.
[0030] Step C: Collect domain voice data of domain-specific words. The correct text data corresponding to each audio file is saved in txt format, corresponding one-to-one with the audio data.
[0031] Step D: Select the WeNet framework for speech recognition model training. The model trained on the large-scale Aishell-1 Chinese speech dataset under the WeNet framework is used as the basic model for transfer learning.
[0032] Step E: Use the transfer learning method to build a speech recognition model, use the Conformer model as the encoder, and use the CTC-Attention mechanism for joint decoding.
[0033] Step F: Based on the speech recognition model obtained by transfer learning, a lightweight speech recognition model is obtained by using the knowledge distillation method, which maintains high model accuracy while making the model lightweight.
[0034] Step G: Construct a rule-based construction knowledge matching method. Based on the established rule base and knowledge base, transform the problem of comparing the key points of expert experience knowledge with the specification provisions into a construction experience knowledge matching analysis problem.
[0035] Furthermore, step A specifically includes: Step A1: Analysis of the knowledge structure of construction experience of construction plan. The key information of construction plan knowledge structure usually includes: project overview, main construction methods and processes, construction guarantee measures, etc. The general content structure of construction plan knowledge structure is as follows: Figure 1 shown.
[0036] Step A2: Analysis of the knowledge structure of construction experience in construction technology. The main focus is on the following: scope of application, pre-construction preparation, process flow and construction requirements, safety, quality, and finished product protection. Among them, the focus includes the applicable conditions of the construction technology, preliminary preparation work, specific operation procedures and requirements, and measures to deal with possible challenges such as safety, quality, and finished product protection. The general content structure of construction technology experience knowledge is as follows: Figure 2 shown.
[0037] Furthermore, step B specifically includes: Step B1: Expert construction experience speech recognition. The domain speech recognition model can more accurately recognize expert construction experience speech containing domain-specific terms, providing a more convenient way to enter experience.
[0038] Step B2: Locate the content of expert construction experience knowledge. Construction experience knowledge has a specific content structure. It is necessary to promptly analyze the thematic hierarchy of expert construction experience knowledge and locate it in specific modules within the guidance template. This serves as an aid in determining the completeness of expert experience knowledge and ensuring its systematicity and integrity as much as possible.
[0039] Step B3, matching analysis of key knowledge elements of expert construction experience. Construction experience knowledge involves some fine-grained construction knowledge points, which may be different from the original specifications. To further analyze the key points of expert construction experience knowledge, it is necessary to match the expert experience with the knowledge in the original construction knowledge base. The constructed voice-based expert construction experience intelligent guidance framework is as follows: Figure 3 .
[0040] Furthermore, step C specifically includes: Step C1: obtain the voice data of relevant experts from the construction unit who record construction experience, and convert the audio data into a unified audio format: a mono wav format with a sampling frequency of 16kHz and a sampling size of 16 bits.
[0041] In step C2, in order to improve the model training effect, the speech rate change method commonly used in expanding speech training data is used to expand the data, and the Audacity tool is used to process it. A total of 0.9 and 1.1 speech data with different speeds are added, while maintaining the same data format.
[0042] Furthermore, step D specifically includes: In step D1, download and prepare the Aishell-1 Chinese speech dataset, including audio files and corresponding text labels. Divide the dataset into training, validation, and test sets to ensure data balance and diversity.
[0043] In step D2, a basic speech recognition model is built using the WeNet framework. You can choose an existing model or customize it as needed. The Aishell-1 dataset is used as training data. The model parameters are optimized through training to accurately convert speech signals into corresponding text.
[0044] In step D3, the model trained on the Aishell-1 dataset is used as the base model for transfer learning. It is fine-tuned on datasets related to water supply and drainage pipeline construction to adapt to new domains and tasks.
[0045] Furthermore, step E specifically includes: In step E1, the audio is first processed using a convolutional downsampling layer during the encoding phase, and then processed using multiple Conformer modules.
[0046] In step E2, the multi-head attention mechanism module replaces the positional encoding in Transformer with the relative sinusoidal positional encoding in Transformer-XL, so that the attention module generalizes better across different input lengths, and the generated encoder is more robust to changes in audio sequence length. Figure 4 This is the structural diagram of the multi-head attention mechanism module.
[0047] In step E3, the convolution module introduces point-by-point convolution. This involves convolution of the feature map point by point using a 1×1 convolution kernel and activation of the projection channels using a Gated Linear Unit (GLU). This creates a gating mechanism followed by a one-dimensional depthwise separable convolution, which improves model efficiency. A batchnorm (another network layer inserted between the next hidden layer) is added for normalization to facilitate training deeper models. Figure 5 This is the structure diagram of the convolution module.
[0048] Step E4, CTC-Attention joint decoding, combines the advantages of the connectionist temporal classification model and the attention model to better process the characteristic sequence of the speech signal. Its main principle is: the CTC model is used to force alignment of the input speech signal and the output. Then, through the shared network parameter transfer, the attention model is used to model and predict the predicted candidate text sequence, obtaining an attention vector. Finally, the predicted attention vector is jointly decoded with the input speech signal to generate the final prediction result.
[0049] Furthermore, step F specifically includes: Step F1, build the overall framework of knowledge distillation ( Figure 6 ): Soft labels are the predictions of the teacher model. They contain the high-dimensional representation information of the teacher network on the input samples and are expressed in the form of probability distribution; soft predictions are the output results generated by the student network during the training process. Their form is the same as that of soft labels, both expressed in the form of probability distribution; hard labels refer to the true labels or answers marked for given samples; hard predictions refer to the substantive classification decisions made by the student network on the input samples during the inference process. They are specific category label values. By comparing with the hard labels, the performance and generalization ability of the student network on the test set can be evaluated, thereby guiding the optimization and adjustment of the model.
[0050] Step F2, training loss function.
[0051] ; (1) ; (2) ; (3) Where, It is called student loss, which represents the loss of the student network on the training data, helping the student network itself to have better accuracy and generalization ability. For loss of students based on CTC, is the student loss based on Attention; is the distillation loss, which represents the difference between the output of the student network and the output of the teacher network, helping the student network learn the knowledge of the teacher network. is the distillation loss based on CTC, is the distillation loss based on Attention; It is a combination of distillation loss and student loss; and is an adjustable parameter, where Represents the weight coefficient of the CTC training process, Represents the distillation weight coefficient of CTC. In this embodiment and The value is 0.5.
[0052] Furthermore, step G specifically includes: Step G1: Gain an in-depth understanding of the content of the construction experience text and the content of the referenced specification provisions, clarify the content in the construction experience text that needs to be matched and analyzed, and grasp the characteristics of the construction specification text.
[0053] In step G2, a construction experience knowledge matching algorithm is designed using the regular expression rule base and the knowledge graph rule base to realize the matching analysis of the expert construction experience, and the matching analysis results of the system are output according to the module, so that the expert can respond to the corresponding questions in a timely manner based on the system output.
[0054] In step G3, since the contents of the construction method are mostly keywords and corresponding simple numerical values, the rules in the regular expression are used to perform matching analysis. The key information extracted from the construction method screening condition text block and the construction method selected in the specific construction process text block are used as analysis objects. Then, the set rules are used to analyze whether the construction method is applicable, or whether there are multiple applicable methods but the expert has selected one of them, and then the system feedback prompts are output.
[0055] In step G4, the relevant rules constructed using the knowledge graph are used to perform matching analysis on each key construction element. The text under each topic after positioning is passed as input to the rule matching analysis module. The rules under the module are then used for reasoning. If the corresponding rule requirements are not met, prompts are output. For example, if the constraint attributes of the identified constraint object do not match the specifications, the correct constraint content is output.
[0056] Example 2: To further illustrate the invention process, features, and application methods of the present invention, the following embodiments are given as examples to illustrate a method for guiding expert experience in the field of pipe network engineering based on speech recognition, including the following steps: Step A: Analyze the general content structure of typical construction plans and construction processes.
[0057] Step A1: Analyze the key information of the experience and knowledge of the construction plan, which usually includes: project overview, main construction methods and processes, construction guarantee measures, etc.
[0058] Step A2: Analyze the construction process text, focusing on the scope of application, pre-construction preparation, possible safety issues, quality issues, finished product protection, etc., and take corresponding measures to deal with them.
[0059] Step B: Propose a voice-based expert construction experience intelligent guidance framework.
[0060] Step B1: Propose an expert construction experience speech recognition method. The domain speech recognition model can more accurately recognize expert construction experience speech containing domain professional words, providing a more convenient way to enter experience.
[0061] Step B2: Content positioning of expert construction experience knowledge, analysis of its subject hierarchy, and positioning it to specific modules in the guidance template to ensure the systematicness and integrity of the experience knowledge.
[0062] Step B3: Match and analyze the key knowledge elements of expert construction experience, match them with the knowledge in the original construction knowledge base, and build a voice-based expert construction experience intelligent guidance framework.
[0063] Step C: Collect domain voice data of domain-specific words and save them in a certain format.
[0064] Step C1: collect voice data in the field of water supply and drainage pipeline construction from various audio platforms such as Himalaya and Qingting FM, and segment the long voice data to obtain the original audio content.
[0065] Step C2: manually record relevant audio containing domain terms in a quiet environment using Mandarin voice, and convert the recorded audio data into a unified WAV format.
[0066] In step C3, data expansion is performed by changing the speech speed. The speech data is processed using the Audacity tool, and speech data at 0.9 and 1.1 times the speed are added to improve the model training effect.
[0067] Step D: Use the WeNet framework to train the speech recognition model.
[0068] Step D1: Download and prepare the Aishell-1 Chinese speech dataset, including audio files and corresponding text labels. Divide the dataset into training, validation, and test sets to ensure data balance and diversity.
[0069] In step D2, we build a basic speech recognition model using the WeNet framework. We can choose an existing architecture or customize it as needed. The Aishell-1 dataset is used as training data, and the model parameters are optimized to accurately convert speech signals into text.
[0070] In step D3, the model trained on the Aishell-1 dataset is used as the base model for transfer learning. It is fine-tuned on datasets related to water supply and drainage pipeline construction to adapt to new domains and tasks.
[0071] Step E: Build a speech recognition model using the Conformer model and CTC-Attention mechanism.
[0072] Step E1, construct the Conformer module, including a fully connected module, a multi-head attention mechanism module, a convolution module and a second fully connected module.
[0073] In step E2, the relative sinusoidal position encoding in Transformer-XL is used in the multi-head attention mechanism module to improve the generalization ability of the model.
[0074] In step E3, point-by-point convolution and GLU activation are introduced into the convolution module to improve model efficiency. At the same time, Batchnorm is used for normalization to facilitate the training of deeper models.
[0075] Step F: Use the knowledge distillation method to obtain a high-precision lightweight speech recognition model.
[0076] Step F1: Ensure that a speech dataset related to the field of water supply and drainage pipeline construction has been prepared, including a training set, a validation set, and a test set.
[0077] In step F2, the speech recognition model obtained by transfer learning is used as the teacher model, the training set is inferred, and soft labels are generated as the knowledge transfer target.
[0078] In step F3, a lightweight student model is trained using the soft labels of the teacher model. The loss function of the student model is defined, including the student loss and the distillation loss. The student model is trained on the training set, and the validation set is used to adjust the hyperparameters to optimize the knowledge distillation process.
[0079] In step F4, after knowledge distillation is completed, the lightweight model is evaluated using the test set to calculate its performance indicators on the speech recognition task.
[0080] Step G: Construct a rule-based construction knowledge matching method, compare the key points of expert experience knowledge with the specification provisions, and transform it into a construction experience knowledge matching analysis problem. The overall framework of this method is as follows: Figure 7 shown.
[0081] Step G1: Deeply understand the construction experience text and relevant specification provisions, determine the content that needs to be matched and analyzed, and grasp the characteristics of the construction specification text.
[0082] Step G2 collects construction-related terms and synonyms, understands the text content that needs to be matched under different construction conditions, uses regular expressions to build matching patterns and filtering conditions, and stores rules by category for easy management and call.
[0083] Step G3 uses deep learning methods to specifically extract key construction knowledge elements to assist in matching and analyzing expert construction experience. The text data is preprocessed, and entities and relationships are annotated. Entities and relationships are extracted using deep learning techniques and imported into a knowledge graph to be converted into computer-recognizable rules.
[0084] Step G4, using the rules in the regular expression to perform matching analysis, taking the key information extracted from the construction method screening condition text block and the specific construction process text block as the analysis object, using the set rules to analyze the applicability of the construction method, and outputting the system's feedback prompts.
[0085] Step G5, using the relevant rules constructed by the knowledge graph to perform matching analysis on each key construction element, takes the located text as input and passes it to the rule matching analysis module for reasoning analysis, outputs prompts, such as if the constraint attributes of the identified constraint object do not match the specification, the correct constraint content must be provided. The overall process of the rule matching analysis algorithm is as follows Figure 8 shown.
[0086] Although the preferred embodiments of the present invention have been described above in conjunction with the accompanying drawings, the present invention is not limited to the above-mentioned specific embodiments. The above-mentioned specific embodiments are merely illustrative and not restrictive. Under the guidance of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the present invention and the claims, which all fall within the scope of protection of the present invention.
Claims
1. A method for guiding expert experience in the field of pipe network engineering based on speech recognition, characterized in that: The method comprises: Taking typical construction plans and construction techniques as examples, this paper analyzes the general content structure of these two types of construction experiences. Considering the specific implementation ideas and related technologies of expert construction experience guidance, a voice-based expert construction experience intelligent guidance framework is proposed; Collect domain voice data of domain-specific words. The correct text data corresponding to each audio file is saved in txt format, corresponding one-to-one with the audio data. Data expansion and enhancement of speech datasets; The WeNet framework was used to train the speech recognition model. The model trained on the large-scale Aishell-1 Chinese speech dataset under the WeNet framework was used as the basic model for transfer learning. A speech recognition model was established using transfer learning, with the Conformer model as the encoder and the CTC-Attention mechanism as the joint decoding mechanism. Use knowledge distillation to obtain a lightweight speech recognition model; Construct a rule-based construction knowledge matching method; Based on the constructed rule base and knowledge base, the problem of comparing the key points of expert experience knowledge with the specification provisions is transformed into a construction experience knowledge matching analysis problem.
2. The method for guiding expert experience in the field of pipe network engineering based on speech recognition according to claim 1, characterized in that: Based on the typical construction plan and construction technology, the content structure analysis includes: Analysis of the knowledge structure of construction experience in construction plan: The key information of construction plan experience knowledge includes: project overview, main construction methods and processes, and construction guarantee measures; Analysis of the knowledge framework of construction experience in construction technology: mainly includes the following contents: scope of application, pre-construction preparation, process flow and construction requirements, safety, quality, and finished product protection issues; among them, the focus includes the applicable conditions of construction technology, preliminary preparation work, specific operating procedures and requirements, and measures to deal with possible safety, quality and finished product protection challenges.
3. The method for guiding expert experience in the field of pipe network engineering based on speech recognition according to claim 1 is characterized in that: The specific implementation ideas and related technologies of expert construction experience guidance are considered, and the voice-based expert construction experience intelligent guidance framework is proposed, which specifically includes: Expert construction experience speech recognition: The domain speech recognition model can more accurately recognize expert construction experience speech containing domain-specific terms, providing a more convenient way to enter experience; Expert construction experience knowledge content positioning: Construction experience knowledge has a certain content structure. It is necessary to timely analyze the subject level of expert construction experience knowledge and locate it in specific modules in the guidance template as an aid to judge the integrity of expert experience knowledge, so as to ensure the systematicness and integrity of experience knowledge as much as possible; Matching analysis of key knowledge elements of expert construction experience: Construction experience knowledge involves some fine-grained construction knowledge points, which may differ from the original specifications. To achieve further analysis of the key knowledge points of expert construction experience, it is necessary to match the expert experience with the knowledge in the original construction knowledge base.
4. The method for guiding expert experience in the field of pipe network engineering based on speech recognition according to claim 1 is characterized in that: Collect domain voice data for domain-specific terms. The correct text data corresponding to each audio file is saved in txt format, corresponding one-to-one with the audio data, specifically including: The data was expanded using the speech rate change method commonly used to expand speech training data, and processed using the Audacity tool. A total of 0.9 and 1.1 speech data with different speeds were added, while maintaining the same data format.
5. The method for guiding expert experience in the field of pipe network engineering based on speech recognition according to claim 1, characterized in that: Speech dataset data expansion and enhancement, including: The voice data of relevant experts from the construction unit who record construction experience are obtained, and the audio data is converted into a unified audio format: a mono WAV format with a sampling frequency of 16kHz and a sampling size of 16 bits.
6. The method for guiding expert experience in the field of pipe network engineering based on speech recognition according to claim 1, characterized in that: The WeNet framework is used for speech recognition model training. The model trained on the large-scale Aishell-1 Chinese speech dataset under the WeNet framework is used as the basic model for transfer learning. Specifically, it includes: Download and prepare the Aishell-1 Chinese speech dataset, including audio files and corresponding text labels; divide the dataset into training, validation, and test sets to ensure data balance and diversity; Use the WeNet framework to build a basic speech recognition model, choosing an existing structure or customizing it as needed. Use the Aishell-1 dataset as training data and optimize the model parameters through the training process to enable it to accurately convert speech signals into corresponding text. The model trained on the Aishell-1 dataset is used as the basic model for transfer learning and fine-tuned on datasets related to water supply and drainage pipeline construction to adapt to new fields and tasks.
7. The method for guiding expert experience in the field of pipe network engineering based on speech recognition according to claim 1, characterized in that: A speech recognition model is built using transfer learning, with the Conformer model as the encoder and the CTC-Attention mechanism for joint decoding. Specifically, the following steps are involved: During the encoding phase, the audio is first processed using a convolutional downsampling layer, and then processed using multiple Conformer modules. The multi-head attention mechanism module replaces the positional encoding in Transformer with the relative sinusoidal positional encoding in Transformer-XL. This allows the attention module to generalize better across different input lengths, and the resulting encoder is more robust to changes in audio sequence length. The convolution module introduces point-by-point convolution, which uses a 1×1 convolution kernel to perform convolution on each point of the feature map and a linear gated unit (GLU) to activate the number of projection channels. The two form a gating mechanism, followed by a one-dimensional depth-separated convolution, which improves model efficiency. Batchnorm is added for normalization to train deeper models. CTC-Attention joint decoding combines the advantages of the connectionist temporal classification model and the Attention model to better process the feature sequence of the speech signal. It uses the CTC model to force the alignment of the input speech signal with the output, then uses the shared network to transfer parameters and uses the Attention model to model and predict the predicted candidate text sequence to obtain an attention vector. Finally, the predicted attention vector is jointly decoded with the input speech signal to generate the final prediction result.
8. The method for guiding expert experience in the field of pipe network engineering based on speech recognition according to claim 1, characterized in that: Based on the speech recognition model obtained by transfer learning, a lightweight speech recognition model is obtained using the knowledge distillation method, which maintains high model accuracy while making the model lightweight.
9. The method for guiding expert experience in the field of pipe network engineering based on speech recognition according to claim 1, characterized in that: Construct a rule-based construction knowledge matching method, including: Gain an in-depth understanding of the content of the construction experience text and the referenced normative provisions, identify the key points that require matching and analysis in the construction experience text, and grasp the characteristics of the construction normative text; Using the regular expression rule base and the knowledge graph rule base, we designed a construction experience knowledge matching algorithm to achieve matching analysis of expert construction experience. The matching analysis results of the system are output according to the module, so that experts can respond to corresponding questions in a timely manner based on the system output. Since the construction method content is mostly keywords and corresponding simple numerical values, the rules in the regular expression are used for matching analysis to take the key information extracted from the construction method screening condition text block and the construction method selected in the specific construction process text block as the analysis objects, and then use the set rules to analyze whether the construction method is applicable, or whether there are multiple applicable methods but the expert selected one of them, and then output the system feedback prompts.
10. The method for guiding expert experience in the field of pipe network engineering based on speech recognition according to claim 1, characterized in that: Based on the established rule base and knowledge base, the problem of comparing the key points of expert experience knowledge with the provisions of the regulations is transformed into a construction experience knowledge matching analysis problem, including: The relevant rules constructed using the knowledge graph are used to perform matching analysis on each key construction element. The text under each topic after positioning is passed as input to the rule matching analysis module, and then the rules under the module are used for reasoning. If it does not meet the corresponding rule requirements, a prompt will be output.
