CN113850570A - AI-based professional scheme aided decision-making expert system construction method - Google Patents

Abstract

The invention provides a professional scheme assistant decision-making expert system construction method based on AI (artificial intelligence) and aims to solve the technical problems of low efficiency, high cost, long expert culture period and the like in the current professional scheme compiling mode. The method comprises the steps of constructing a database, establishing a data structure, constructing a knowledge map, matching an intelligent template, conforming the scheme, retrieving a butt joint knowledge base and the like, wherein an operator can automatically compile a professional scheme which is highly targeted and meets the requirements of normative standard files by inputting keywords, and meanwhile, the professional scheme can be intelligently checked and evaluated, so that the advantages of big data and artificial intelligence are fully utilized, and the intelligent compilation, check and evaluation of the professional scheme are realized; meanwhile, the problems of long professional scheme compiling and auditing period, high cost and the like are solved, the evaluation standard is unified, the labor cost is obviously reduced, the working efficiency is improved, a professional knowledge base formed in the system can lay a foundation for knowledge economy, and support is provided for establishing enterprise competitiveness.

Description

AI-based professional scheme aided decision-making expert system construction method

Technical Field

The invention relates to the technical field of compiling and auditing professional schemes of construction projects, in particular to a construction method of an AI-based professional scheme auxiliary decision expert system.

Background

The professional scheme is strong in compiling and auditing speciality, wide in knowledge range and multiple in types, manual experience is mainly relied on in the aspect of compiling and auditing the professional scheme, and compiling quality and auditing depth of the professional scheme are mainly dependent on professional ability and professional moral of people. Currently, companies are facing to the problems of personnel shortage, partial youngness and the like in the rapid development process, and the cultivation of backbones and experts depends on long-term course training and the continuous accumulation of engineering experience, so that expert systems in the companies are lost, and the personnel structure distribution is unbalanced. At present, experts in related fields are mainly organized to develop the compilation and review of various professional schemes, and the problems of large regional differentiation, long period, high cost and the like exist, so that the management benefit of a company is low. With the high-speed development of the artificial intelligence technology, the artificial intelligence technology is widely applied to various professional fields, and the artificial intelligence technology is fused in the process of compiling and reviewing the professional schemes, so that the realization of intelligent compiling and intelligent reviewing of the professional schemes is an inevitable direction for future development. The traditional professional scheme auditing mode adopts auditing, so that the cost is high, the auditing mistakes are easy to miss, and meanwhile, due to the problem of inconsistent auditing standards, everyone is depended on manual auditing to have inconsistent grasp on the compiling scheme and the standard scale of the auditing scheme. And the culture of the backbone and the experts has the problems of long period, high cost and the like, and meanwhile, the professional schemes have a plurality of categories, so that the experts are required to be subjected to a plurality of items of sharpening and learning to reach the expert level.

Disclosure of Invention

Aiming at the defects in the background art, the invention provides a professional scheme assistant decision-making expert system construction method based on AI, which realizes the functions of professional scheme intelligent compilation, intelligent approval, intelligent recommendation and the like, solves the problems of low efficiency, high cost, long expert culture period and the like in the current professional scheme compilation mode, fully utilizes the advantages of big data and artificial intelligence, and realizes the intelligent compilation, audit and evaluation of professional schemes.

In order to solve the technical problems, the invention adopts the following technical scheme: a construction method of a professional scheme assistant decision-making expert system based on AI comprises the following steps:

step S1: establishing a database for storing scheme documents, and presetting a normative standard file; manually collecting internal professional schemes of the construction project, importing the internal professional schemes into a database, and importing external finished product scheme data into the database;

step S2: preprocessing the data in the database in the step S1, establishing structured data according to the preprocessed data, and storing the structured data in a local database;

step S3: constructing a professional knowledge graph based on industry basic information to form a multifunctional knowledge graph management platform;

step S4: the method comprises the steps of performing template processing on building scheme contents of building type fixed homogenization by applying big data and deep learning technology to form label data, analyzing word frequency sentence topics and label data through machine learning, and finally independently compiling a professional scheme by using an NLG system;

step S5: judging whether the scheme belongs to a specific special technical scheme type or not according to the characteristics of the compiled special scheme, and in the special technical scheme type, inputting whether the text of the special scheme meets the requirements of an audit system model or not;

step S6: storing the sensitive data of the professional scheme in the local data center for encrypted storage by adopting a local data center and cloud end architecture; the cloud collects industry big data information of the internet and the database through the intelligent crawler to provide a scheme retrieval platform.

The step S1 is to screen for internal professional solutions and to clean up professional solution data in which there is an expired reference specification and there is an error, while keeping valid professional solution data.

In step S2, the effective professional plan data, the manual feedback data, the digital twin system feedback data, and the external finished product plan data are preprocessed, and the preprocessing operations include data extraction, data loading, and data conversion.

The machine learning in step S4 includes a BERT pre-training model, where the BERT pre-training model includes an MLM task and an NSP task, and when the BERT pre-training model is trained, in order to minimize a combined loss function of the two tasks, the MLM task and the NSP task are jointly trained, and the pre-training tasks of the BERT model are dynamically adjusted in different stages, and a pre-training coefficient λ is introduced, so that a loss function is obtained as:

L′(θ,θ₁,θ₂)＝L₁(θ,θ₁)+λL₂(θ,θ₂),λ＝0or1(1)

in the formula: θ is a parameter of the encoder portion of the BERT model; theta₁Is a parameter in the output layer connected to the encoder in the MLM task; theta₂Is the classifier parameters on the encoder side in the NSP task; l' (theta )₁,θ₂) Representing BERT model trainingA loss function of the whole process is trained; l is₁(θ,θ₁) Representing a loss function of training of the MLM task; l is₂(θ,θ₂) Representing the loss function of the NSP task for training.

The step of using the NLG system to compile a professional plan in step S4 is as follows:

step S4.1: determining the scheme content under construction contained in the NLG system;

step S4.2: after determining the scheme content, the NLG system determines a scheme structure;

step S4.3: determining statement content of a scheme, combining a plurality of information to form a statement with a complete structure for expression by adopting a Transformer framework of a natural language generation model GPT-2 by an NLG system according to semantics and adding connecting words;

step S4.4: and finally, automatically generating a complete and high-quality professional scheme.

According to the invention, by inputting professional scheme data, deep learning is carried out on massive professional scheme materials by using a computer technology, semantic parsing and combination are carried out from the level of words, sentences and even paragraphs, and an operator can automatically compile a high-quality professional scheme which is highly targeted and meets the requirements of normative standard files by inputting keywords, and can intelligently audit and evaluate the professional scheme, so that the problems of low efficiency, high cost, long expert culture period and the like in the current professional scheme compiling mode are solved, and the advantages of big data and artificial intelligence are fully utilized to realize intelligent compiling, auditing and evaluating of the professional scheme; meanwhile, the problems of long compiling and auditing period, high cost and the like of a professional scheme are solved, the evaluation standard is unified, the labor cost is obviously reduced, and the working efficiency is improved; the knowledge base/professional knowledge base formed in the system can lay a foundation for knowledge economy and provide support for constructing enterprise competitiveness.

Drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below, it is obvious that the drawings in the following description are only some embodiments of the present invention, and for those skilled in the art, other drawings can be obtained according to the drawings without creative efforts.

FIG. 1 is a flow chart of the project data processing of the present invention;

FIG. 2 is an AI specific project assistance expert system engineering architecture diagram of the present invention;

FIG. 3 is a knowledge graph construction flow diagram of the present invention;

FIG. 4 is a transform framework of the natural language generation model GPT-2 of the present invention;

FIG. 5 is a schematic diagram of a pre-training phase of the BERT model of the present invention;

FIG. 6 is a flow chart of the transfer learning process for dynamically adjusting the pre-training task according to the present invention.

Detailed Description

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be obtained by a person skilled in the art without inventive effort based on the embodiments of the present invention, are within the scope of the present invention.

The invention provides a professional scheme assistant decision-making expert system construction method based on AI, comprising the following steps:

step S1: and establishing a database for storing the scheme documents, and presetting a normative standard file, wherein the normative standard file comprises corresponding national policies, laws and regulations, company systems and the like. Manually collecting internal professional schemes of the building project and importing the internal professional schemes into a database, then screening the internal professional schemes, cleaning professional scheme data with overdue reference specifications and errors, and keeping effective professional scheme data; meanwhile, the external finished product scheme is supported to be imported, and the external finished product scheme data can be imported into the database. Step S2: the method comprises the steps of preprocessing effective professional scheme data and external finished product scheme data imported in the step S1, manual feedback data fed back by manual modification, data fed back by a digital twin system and the like, performing operations such as data extraction, data loading and data conversion on the data by using a database, establishing structured data according to the preprocessed data, storing the structured data into the database in the step S1, namely, performing structured processing on the data after data acquisition is completed to enable the data and the structure thereof to be clear, specifically, as shown in FIG. 1, preprocessing the service data in the step S1 to obtain a data warehouse, and finally, realizing data application conversion by applying technologies such as big data and machine deep learning and the like, namely, realizing functions such as data visualization, intelligent writing, intelligent review, intelligent recommendation and the like of the system.

As shown in fig. 2, the engineering architecture of the present invention includes a data acquisition layer, a data storage layer, an information and knowledge layer, a calculation engine layer, an algorithm training layer, an AI service layer, and a visualization layer, wherein the data acquisition layer further includes network data crawling, professional solution library docking, manual feedback solution data, and digital twin feedback data; the data collected by the data collection layer are stored in a data storage layer, and the data storage layer comprises distributed storage, column storage, index storage and structured storage; and the data stored in the data storage layer is processed to obtain the content contained in the information and knowledge layer, system behavior data, industry search data, a constructed knowledge map, professional field resource data and the like. The computing engine layer comprises an offline and real-time computing engine and an algorithm training engine; the algorithm training layer mainly comprises machine learning and deep learning, the main contents of the machine learning comprise data classification, clustering, regression and association rules, the main contents of the deep learning comprise RNN and CNN neural network models and XLNET and RERT pre-training models, and technical support is provided for the next AI service layer. The AI service layer comprises intelligent writing (compiling), intelligent reviewing, intelligent recommending and a knowledge graph; the visualization layer comprises data analysis, data visualization and model evaluation, namely, the data analysis, the model evaluation and the like are carried out on the scheme automatically compiled by the system. Through the constructed data acquisition layer, the data storage layer, the information and knowledge layer, the calculation engine layer, the algorithm training layer, the AI service layer, the visualization layer and the like, data support is provided for an AI professional scheme auxiliary decision-making expert system and a future accessed application system.

Step S3: based on industry basic information, constructing a professional knowledge graph to form a multifunctional knowledge graph management platform, as shown in fig. 3, the process is as follows: firstly, map production; then creating a database framework, importing data and constructing a knowledge graph; performing strategy training including knowledge extraction, map construction and the like; and finally, knowledge management, including map management and knowledge application. The industry basic information comprises industry characteristics, different professional scheme categories, enterprise knowledge, industry policy knowledge and the like, a complex professional knowledge map is constructed based on the basic information, the professional knowledge map mainly comprises a plurality of building information such as compilation basis, policy and regulation, a company system, an audit book, a building drawing and the like, so that a multifunctional knowledge map management platform is formed, and a user can conveniently screen and perfectly use a professional scheme with high matching similarity according to keyword matching when searching the professional scheme.

Step S4: the method comprises the steps of applying big data and deep learning technology, conducting template processing on building scheme contents with fixed and homogeneous buildings to form tag data, analyzing word frequency sentence topics and tag data through machine learning, and finally independently compiling a professional scheme by using an NLG system, so that the professional scheme can be automatically recommended and automatically compiled by inputting relevant keywords for compiling the professional scheme. The specific operation is as follows: 1. a key vector library is constructed against the schema database. For example, all texts have 1 ten thousand rows of data, but after word segmentation processing, there are probably 6000 different word vectors, the whole text can be represented by a 10000 × 6000 matrix, each row has 6000 elements, and a word vector appears several times in the original corresponding row number, and is assigned as several values for the word vector, if 1 appears, the word vector is assigned as 1, and if the word vector does not appear, the word vector is assigned as 0. 2. A TF-IDF (term frequency-inverse document frequency) method is used to evaluate how important a word is to a solution in the document repository. The importance of a word increases in direct proportion to the number of occurrences in the schema, while decreasing in inverse proportion to the frequency with which it occurs in other files in the document library. That is, a word appears more times in a certain scheme, and other schemes do not appear, which indicates that the word is important for the scheme classification, however, if other schemes also appear more and indicate that the word is not very distinctive, the IDF is used to reduce the weight of the word. 3. And intelligent matching can be started when the previous word frequency matrix is completely converted into a matrix containing the TF-IDF method. Specifically, the keyword input by the user is subjected to jieba word segmentation, then the matrixes with the same width (length) as the word frequency matrix are converted in the same way, then the matrixes are subjected to multiplication operation one by one to calculate the value of each row, the maximum value is the value with the highest matching degree, finally the matrixes in the TF-IDF method are transposed, so that the keywords can be conveniently multiplied by the input matrixes, and then the matrixes are sorted, so that the text with the highest matching degree can be obtained.

In this embodiment, in step S4, machine learning is used to implement automatic recommendation and autonomous compilation functions of a professional scenario. One model applied in the machine learning is a BERT pre-training model, which is a pre-training language model aiming at various Natural Language Processing (NLP) tasks and has the capabilities of bidirectional coding and feature extraction. Text error correction is carried out by using a BERT model, based on the pre-trained BERT model, after unsupervised training is carried out on a label-free related field data set, part of parameters are transferred to a target error correction model, and the model is trained again through a training sample of a building document corpus data set to obtain a better error correction effect.

The BERT pre-training model includes two unsupervised prediction tasks, i.e., MLM task and NSP task, in the task part of the pre-training process, as shown in fig. 5, in the pre-training process of the BERT model, the model receives paired sentences (e.g., a and B) as input, and then predicts whether B is the next sentence of a. There is a 50% probability B of being the next sentence of a and a 50% probability B of being a random sentence in the corpus. When training the BERT pre-training model, in order to minimize the combined loss function of the two tasks, the MLM task and the NSP task are jointly trained, and the loss function is defined as

L(θ,θ₁,θ₂)＝L₁(θ,θ₁)+L₂(θ,θ₂) (2)

In the formula: θ is a parameter of the encoder portion of the BERT model; theta₁Is a parameter in the output layer connected to the encoder in the MLM task; theta₂Is the classifier parameters on the encoder side in the NSP task; l is₁(θ,θ₁) Representing a loss function of training of the MLM task; l is₂(θ,θ₂) Representing the loss function of the NSP task for training.

For the loss function of the first part of the MLM task, if the masked word set is M, since it is a multi-classification problem on a dictionary size | V |, the loss function used is a negative log-likelihood function. And because it needs to be minimized, the loss function of the first part is equivalent to the maximum log-likelihood function, i.e. the loss function L in MLM task training₁(θ,θ₁) The mathematical expression of (a) is:

in the formula: i represents the element order of traversal; m represents a predicted word; m is_iIndicating the correct word in order i; p represents the probability of a prediction result; m represents a word set (number-randomly selected number of covering words) in which the input sentence in the MLM task is randomly covered; | V | represents a dictionary size.

For the loss function of the second part of NSP task, since NSP task is a classification problem, the loss function L in training NSP task₂(θ,θ₂) The mathematical expression of (a) is:

in the formula: j represents the element order of traversal; n represents the word order prediction result; n is_jRepresents the correct word in order j; n represents the number of sentences; IsNext means correct in language order; NotNext refers to an error in language order.

Further, pre-training to dynamically adjust the BERT model for different phasesTask, the pre-training coefficient λ is introduced in this embodiment, and a loss function L' (θ, θ) is obtained₁,θ₂) The mathematical expression of (a):

L′(θ,θ₁,θ₂)＝L₁(θ,θ₁)+λL₂(θ,θ₂),λ＝0or1 (1)

after the pre-training is finished, a BERT pre-training model is obtained. Aiming at different Natural Language Processing (NLP) tasks, fine adjustment is carried out on the basis of a BERT pre-training model, the input and output of a specific task are inserted into the BERT, and downstream tasks are simulated by utilizing a strong attention mechanism of a Transformer. As in the first stage, when performing unsupervised training using the unlabeled related domain data set, for formula (1), let λ be 0, remove the NSP task, and train only the MLM, so as to enhance the learning of the model on the proper nouns in different domains, and at the same time, ensure that the model is not affected by the strong related statements in the same domain. In the second phase, let λ be 1 for equation (1), the training of NSP is unlocked. After the fine tuning training of the target data set, the word sequence false alarm rate of the model can be effectively reduced. The flow of the transfer learning for dynamically adjusting the pre-training task is shown in fig. 6.

In this embodiment, the step of programming the professional scheme by using the NLG system in step S4 is as follows: step S4.1: the contents of the scheme being built contained in the NLG system are determined, i.e., the NLG system needs to determine which information should be contained in the scheme being built. Step S4.2: after determining the scheme content, the NLG system determines the scheme structure, that is, after selecting the scheme content, the NLG system needs to determine the arrangement order of the scheme content. Step S4.3: determining statement content of a scheme, combining a plurality of information to form a statement with a complete structure for expression by adopting a Transformer framework of a natural language generation model GPT-2 and adding connecting words according to semantics. The Transformer framework of the natural language generation model GPT-2 is shown in FIG. 4. Step S4.4: after the steps are completed, a complete and high-quality professional scheme is automatically generated finally.

Step S5: automatically judging whether the scheme belongs to a specific special scheme type or not according to the characteristics (content or attributes) of the compiled special scheme, and inputting whether the text of the special scheme meets the requirements of an auditing system model or not in the special scheme type; whether the auditing scheme meets the requirements of national policy, laws and regulations, standard specifications, company system, expert opinions and the like, and the format, content, conformity and the like of the auditing professional scheme. Step S6: the method comprises the steps that a local data center plus cloud end architecture is adopted, sensitive data of a professional scheme are stored in the local data center for encryption and storage, and the safety of a building scheme is guaranteed; meanwhile, the cloud collects industry big data information of the Internet and the database through the intelligent crawler to provide a scheme retrieval platform, namely the system is connected with large retrieval platforms such as encyclopedia, Wikipedia and the like, and a user can perform scheme retrieval on a current page of scheme compilation in real time to provide reference.

In the embodiment, the Similarity between the professional scheme automatically compiled by the system and the evaluation standard is calculated by using Embedding Similarity, so that the system can automatically check according to the evaluation index AI, and then the system is manually turned to after the checking is passed. For a plurality of different scheme texts, the similarity between the texts is calculated, mainly words in the texts are mapped to a vector space to form a mapping relation between the words in the texts and vector data, and the similarity of the texts is calculated by calculating the difference of several or a plurality of different vectors. Similarity measurement is carried out by adopting an Embedding mode, and a figure for measuring similarity is returned. Meanwhile, an evaluation index system is further optimized by combining with the traditional expert offline evaluation standard.

The above description is only for the purpose of illustrating the preferred embodiments of the present invention and is not to be construed as limiting the invention, and any modifications, equivalents, improvements and the like that fall within the spirit and principle of the present invention are intended to be included therein.

Claims (5)

1. A professional scheme assistant decision-making expert system construction method based on AI is characterized in that: the method comprises the following steps:

step S1: establishing a database for storing scheme documents, and presetting a normative standard file; manually collecting internal professional schemes of the construction project, importing the internal professional schemes into a database, and importing external finished product scheme data into the database;

step S2: preprocessing the data in the database in the step S1, establishing structured data according to the preprocessed data, and storing the structured data in a local database;

step S3: constructing a professional knowledge graph based on industry basic information to form a multifunctional knowledge graph management platform;

step S4: the method comprises the steps of performing template processing on building scheme contents of building type fixed homogenization by applying big data and deep learning technology to form label data, analyzing word frequency sentence topics and label data through machine learning, and finally independently compiling a professional scheme by using an NLG system;

step S5: judging whether the scheme belongs to a specific special technical scheme type or not according to the characteristics of the compiled special scheme, and in the special technical scheme type, inputting whether the text of the special scheme meets the requirements of an audit system model or not;

step S6: storing the sensitive data of the professional scheme in the local data center for encrypted storage by adopting a local data center and cloud end architecture; the cloud collects industry big data information of the internet and the database through the intelligent crawler to provide a scheme retrieval platform.

2. The AI-based expert project aid decision expert system construction method of claim 1, wherein: the step S1 is to screen for internal professional solutions and to clean up professional solution data in which there is an expired reference specification and there is an error, while keeping valid professional solution data.

3. The AI-based expert project aid decision expert system construction method of claim 2, wherein: in step S2, the effective professional plan data, the manual feedback data, the digital twin system feedback data, and the external finished product plan data are preprocessed, and the preprocessing operations include data extraction, data loading, and data conversion.

4. The AI-based expert proposal aided decision making expert system construction method according to any one of claims 1 to 3, wherein: the machine learning in step S4 includes a BERT pre-training model, where the BERT pre-training model includes an MLM task and an NSP task, and when the BERT pre-training model is trained, in order to minimize a combined loss function of the two tasks, the MLM task and the NSP task are jointly trained, and the pre-training tasks of the BERT model are dynamically adjusted in different stages, and a pre-training coefficient λ is introduced, so that a loss function is obtained as:

L′(θ,θ₁,θ₂)＝L₁(θ,θ₁)+λL₂(θ,θ₂),λ＝0 or 1 (1)

in the formula: θ is a parameter of the encoder portion of the BERT model; theta₁Is a parameter in the output layer connected to the encoder in the MLM task; theta₂Is the classifier parameters on the encoder side in the NSP task; l' (theta )₁,θ₂) A loss function representing the whole process of BERT model training; l is₁(θ,θ₁) Representing a loss function of training of the MLM task; l is₂(θ,θ₂) Representing the loss function of the NSP task for training.

5. The AI-based expert proposal aid decision expert system construction method of claim 4, wherein: the step of using the NLG system to compile a professional plan in step S4 is as follows:

step S4.1: determining the scheme content under construction contained in the NLG system;

step S4.2: after determining the scheme content, the NLG system determines a scheme structure;

step S4.3: determining statement content of a scheme, combining a plurality of information to form a statement with a complete structure for expression by adopting a Transformer framework of a natural language generation model GPT-2 by an NLG system according to semantics and adding connecting words;

step S4.4: and finally, automatically generating a complete and high-quality professional scheme.

Images