Intelligent pre-class meeting system and method based on artificial intelligence

Abstract

The invention discloses an intelligent pre-class meeting system based on artificial intelligence, which relates to the technical field of artificial intelligence and comprises a data acquisition module, a central processing module, an intelligent generation module, a process guidance module, a real-time monitoring module, a summary generation module and an assessment module. The invention further discloses an intelligent pre-class meeting method based on artificial intelligence. According to the intelligent pre-class routine meeting system and method, through the data acquisition and central processing module, deep integration and analysis of headquarter policy documents and branch company operation data are achieved, it is ensured that the meeting content closely meets the actual service requirement, the intelligent generation module combines the branch company operation data, the issue priority is dynamically adjusted, and the meeting efficiency is improved. The personalized conference agenda is generated, the pertinence and effectiveness of the conference are improved, the conference process is monitored and managed in real time through the application of the real-time monitoring module and the process guiding module, and the problem that the conference deviates from the theme is effectively avoided.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an intelligent pre-shift meeting system and method based on artificial intelligence. Background Art

[0002] With the digital transformation of the logistics industry, pre-shift meetings, as an important part of standardized enterprise management, are increasingly in need of intelligent upgrades. Traditional pre-shift meetings rely on manual formulation of meeting content and offline execution, which makes it difficult to meet the requirements of large-scale and refined management of modern enterprises. In recent years, some companies have attempted to realize the onlineization of meeting processes through information technology, but they still face problems such as low content generation efficiency and uncontrollable execution effects. Existing technologies mainly focus on electronic recording and simple data analysis of meeting processes. They lack the ability to intelligently generate meeting content and a dynamic monitoring mechanism for the execution process, making it difficult to adapt to the differentiated management needs across regions and multiple scenarios.

[0003] However, the current technical bottlenecks prevalent in the industry limit the improvement of meeting management efficiency. First, the generation of meeting content relies on manual experience and cannot dynamically adapt to the differentiated needs of different branches, resulting in a disconnect between policy communication and actual business scenarios. Secondly, the meeting execution process lacks real-time supervision methods. Traditional post-inspection methods are difficult to detect execution deviations, and manual verification is costly. Furthermore, the manual compilation of meeting minutes is inefficient, with significant problems of information omissions and inconsistent formats, making it difficult to support subsequent knowledge accumulation and decision-making analysis. In addition, the lack of a quantitative evaluation system makes it difficult to continuously optimize the quality of meetings, and there is a lag in information feedback between headquarters and branches. To this end, we propose an intelligent pre-shift meeting system and method based on artificial intelligence. Summary of the Invention

[0004] In order to solve the above technical problems, an intelligent pre-shift meeting system and method based on artificial intelligence are provided. This technical solution solves the above problems.

[0005] In order to achieve the above objects, the technical solution adopted by the present invention is:

[0006] An intelligent pre-shift meeting system based on artificial intelligence, comprising: a data acquisition module, a central processing module, an intelligent generation module, a process guidance module, a real-time monitoring module, a minutes generation module and an assessment and evaluation module;

[0007] The data collection module is used to obtain headquarters policy documents and branch operation data before the pre-shift meeting;

[0008] The central processing module is used to perform natural language processing using the BERT model, construct a policy text feature vector matrix, and perform in-depth analysis and feature extraction on the obtained headquarters policy documents. The model algorithm expression is:

[0009] V policy =BERT encoder (P text )

[0010] Where V policy The feature vector obtained after the policy text is encoded by the BERT model has a dimension of 512. This vector is used for the subsequent semantic analysis of the policy text and the calculation of the correlation with the branch operation data. encoder represents the encoder part of the BERT model, P text A text sequence representing headquarters policy documents, which is the input to the BERT encoder;

[0011] The intelligent generation module is used to combine the policy text feature vector obtained by the central processing module and the branch operation data to determine the issue weight through the hierarchical analysis method. The calculation formula is:

[0012]

[0013] Where W topic represents the weight of each topic in the pre-shift meeting. The weight can be used to determine the importance of the topic in the meeting and then generate the meeting agenda. n represents the number of branch data indicators, that is, the number of types of branch operation data indicators considered. α i represents the weight of branch data indicators, β i,j It represents the policy correlation coefficient, reflecting the degree of correlation between each branch's data indicators and the headquarters' policies;

[0014] The process guidance module is used to monitor and manage the process of the meeting according to the agenda determined by the intelligent generation module;

[0015] The real-time monitoring module is used to collect video and audio information of the conference site through sound and image acquisition equipment, and monitor the conference situation in real time;

[0016] The minutes generation module is used to combine the information collected by the real-time monitoring module and the speech records during the meeting to summarize and refine the meeting content and generate meeting minutes;

[0017] The assessment and evaluation module is used to assess the quality and effectiveness of pre-shift meetings based on the information provided by the real-time monitoring module and the minutes generation module.

[0018] Preferably, the process guidance module includes a speech recognition subsystem, which uses a CTC-loss algorithm for speech transcription, and its acoustic model training formula is:

[0019]

[0020] Where, L ctc represents the loss function in the CTC-loss algorithm, which is used to train the acoustic model. T represents the time step, that is, the number of time segments into which the speech signal is divided. C represents the character set size, which refers to the number of all characters that may appear in speech recognition. π tc represents the alignment probability of character c at time t, p t,c represents the probability of predicting character c at time t; the anomaly detection module constructs a semantic deviation model through the LSTM network, and the formula is:

[0021] D deviation =sigmoid(W h ·h t +b)

[0022] Where D deviation It represents the semantic deviation, which can be used to judge whether the speech content in the meeting deviates from the meeting theme. When the semantic deviation exceeds the predetermined threshold, a reminder is issued. Sigmoid represents the activation function. W h Represents the weight matrix, which is used to adjust the influence of the hidden layer state in the LSTM network. t It represents the hidden state of the LSTM network at the tth moment, which contains the information of the previous moment and is used to capture the long-term dependency in the sequence data. b represents the bias term.

[0023] Preferably, the real-time monitoring module uses the OpenPose algorithm to perform human posture recognition and calculate the concentration of participants:

[0024]

[0025] Where A concentration represents the concentration of participants, N represents the number of participants, θ i represents the deviation angle of the i-th participant’s sight. The larger the angle, the less focused the participant is. The video analysis module detects the meeting scene using the YOLOv5 model, and its loss function formula is:

[0026] L yolo =λ coord ·L box +λ obj ·L obj +λ cls ·L cls

[0027] Where, L yolo Represents the loss function of the YOLOv5 model, which is used to optimize the detection performance of the model, λ coord ,λ obj ,λ clsRepresents the weight coefficients of different parts of the loss, which are used to adjust the relative importance of bounding box loss, target confidence loss and classification loss in the total loss. box Represents the bounding box loss, which is used to measure the position error between the detection box and the real box, L obj Represents the target confidence loss, which is used to evaluate the accuracy of the model's judgment on the existence of the target. cls Represents the classification loss, which is used to determine the category to which the detected target belongs.

[0028] Preferably, the summary generation module uses the TextRank algorithm to generate summaries, and its iterative formula is:

[0029]

[0030] Where, Represents sentence v in the TextRank algorithm i The score of , d represents the damping coefficient, which is used to control the randomness of the iterative process. Indicates pointing to sentence v i The set of all sentences that are related to sentence v i There are related sentences. Represents sentence v j The set of all sentences pointed to by reflects the sentence v j The associated range, w ji Represents the edge weight, which measures the sentence v j With sentence v i The structured processing module applies regular expressions to extract key information, and the regular pattern is:

[0031] Regex={(Task|Decision):\w+,\s+\d4-\d2-\d2}

[0032] In the formula, Regex represents a regular pattern for extracting key information from meeting minutes, \w+ represents matching one or more word characters, which is used to match the name of a task or decision, \s+ represents matching one or more whitespace characters, and \d4-\d2-\d2 is used to match the date format. \d4 represents matching four digits, corresponding to the year, and \d2 represents matching two digits, corresponding to the month and day, respectively.

[0033] An intelligent pre-shift meeting method based on artificial intelligence includes the following steps:

[0034] S1. Data integration and analysis: Collect headquarters policy documents and branch operation data, extract the core content of the policy through natural language processing technology, and generate a preliminary topic pool based on the branch data characteristics. Specifically, the headquarters policy documents are converted into a TF-IDF feature matrix using the formula:

[0035]

[0036] Where, Indicates word frequency, represents the document frequency, and N represents the total number of documents;

[0037] S2. Personalized agenda generation: Dynamically adjust the priority of topics based on the current operating status of the branch. Algorithms are used to match policy relevance with the actual needs of the branch to generate a detailed meeting agenda that includes discussion time and key content. Topic combinations are optimized using a genetic algorithm, and the fitness function is:

[0038]

[0039] Where, F fitness Represents the fitness function in the genetic algorithm, which is used to evaluate the quality of each topic combination, m represents the number of topics, γ k represents the topic priority coefficient, T k represents the actual discussion time of topic k, represents the planned discussion duration of topic k;

[0040] S3. Meeting process guidance: This system monitors meeting content in real time through voice recognition and semantic analysis, automatically detecting abnormalities such as deviations from the topic and overtime discussions, prompting the host to adjust the agenda through voice or text, and simultaneously tracking the progress of each topic and reminding the time nodes;

[0041] S4. Multi-dimensional real-time monitoring: Using cameras and microphones to collect audio and video data from the meeting, computer vision is used to analyze the posture and attention of participants. Participation indicators are calculated based on speaking frequency, and a real-time data dashboard of the meeting execution is generated simultaneously.

[0042] S5. Intelligent Minutes Generation: Transcribe the entire meeting recording into text, extract key decision-making, task assignment, and other information through natural language processing technology, automatically generate structured meeting minutes, format them according to preset templates, and simultaneously send them to relevant personnel and headquarters for archiving;

[0043] S6. Quality Assessment and Feedback: Establish a scoring model based on indicators such as agenda completion, participation, and time compliance to generate a quantitative assessment report; analyze common issues through historical data mining, provide optimization suggestions to headquarters, and dynamically adjust subsequent meeting generation strategies.

[0044] Preferably, the step of guiding the conference process during the conference includes real-time voice analysis and progress management mechanism;

[0045] Among them, real-time speech analysis uses the VAD algorithm for endpoint detection, and its expression is:

[0046]

[0047] Where, E voice An indicator used to judge speech activity, T represents the time length of the speech signal, s t represents the energy value of the speech signal at the tth moment, μ represents the mean value of the speech signal energy, σ 2 Represents the variance of the speech signal energy, τ is the set threshold, when E voice When it is greater than τ, it is considered that there is speech activity;

[0048] Among them, the progress management mechanism uses a dynamic programming algorithm to optimize the agenda time allocation, and its objective function is:

[0049]

[0050] The constraints are Where, represents the actual discussion time of topic i, is the planned discussion time of topic i, n is the number of topics, T total It is the upper limit of the total time of the meeting. The dynamic programming algorithm decomposes the complex agenda time allocation problem into multiple sub-problems and solves the optimal solution of the sub-problems to obtain the optimal solution of the original problem, thereby effectively managing the meeting progress.

[0051] Preferably, the real-time monitoring step is performed during the conference;

[0052] The participation calculation model is used to evaluate the participation of participants in the meeting. The formula is:

[0053]

[0054] Where, P participation Indicates the participation of participants, N represents the number of participants, that is, the total number of people participating in the pre-shift meeting, S i is the number of speeches made by the i-th participant, δ i is the speaking interval of the i-th participant;

[0055] Outliers are detected using the isolation forest algorithm, and the anomaly score formula is:

[0056] s(xψ)=2 -E(h(x)) / c(ψ)

[0057] Where s(x,ψ) represents the anomaly score of data point x, E(h(x)) represents the average path length of data point x in the isolation forest, and c(ψ) is an expected value related to the number of samples, which is used to normalize the path length.

[0058] Preferably, the meeting minutes generation step is performed after the meeting, which includes two parts: multimodal information fusion and structured output; multimodal information fusion uses an attention mechanism to fuse speech and text features, and the structured output generates a standard format through a template matching algorithm.

[0059] Preferably, the quality of the pre-shift meeting is evaluated through a multi-dimensional scoring system, using the following formula:

[0060] Q score =ω1·C completion +ω2·P participation +ω3·T timeliness

[0061] Where Q score represents the comprehensive score of the meeting, ω1, ω2, and ω3 represent the weight coefficients of the three dimensions of meeting completion, participant participation, and meeting time control, respectively. completion Indicates the degree of completion of the meeting agenda, P participation Indicates the participation of participants, T timeliness It is about the control of meeting time.

[0062] Preferably, a cloud deployment architecture and a data security mechanism are adopted, wherein the cloud deployment architecture adopts a microservice architecture, and the data security mechanism adopts an AES-256 encryption algorithm.

[0063] Compared with the prior art, the present invention has the following beneficial effects:

[0064] The intelligent pre-shift meeting system and method proposed in the present invention realizes the in-depth integration and analysis of headquarters policy documents and branch operation data through data collection and central processing modules, ensuring that the content of the meeting is closely aligned with actual business needs. The intelligent generation module combines branch operation data, dynamically adjusts topic priorities, and generates personalized meeting agendas, thereby improving the pertinence and effectiveness of the meeting. The application of real-time monitoring module and process guidance module enables the meeting process to be monitored and managed in real time, effectively avoiding problems such as the meeting deviating from the topic or lagging behind in progress. At the same time, through the participation calculation model and anomaly detection algorithm, it can accurately evaluate the participation of participants and identify abnormal behavior, ensuring the smooth progress of the meeting. The minutes generation module adopts advanced text summarization algorithm to realize automatic summary and refinement of meeting content, greatly improving the efficiency and quality of meeting minutes generation. The assessment and evaluation module uses a multi-dimensional scoring system to comprehensively evaluate the quality and effectiveness of the meeting, providing data support for the continuous optimization of the meeting. The cloud deployment architecture and data security mechanism are adopted to ensure the security and reliability of data, providing a strong guarantee for the digital transformation of enterprises. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 It is a system module diagram of the present invention;

[0066] Figure 2 Flow chart of the method of the present invention. DETAILED DESCRIPTION

[0067] The following description is intended to disclose the present invention so that those skilled in the art can implement the present invention. The preferred embodiments described below are merely examples, and those skilled in the art may conceive of other obvious variations.

[0068] Reference Figure 1 As shown, an intelligent pre-shift meeting system based on artificial intelligence includes: a data acquisition module, a central processing module, an intelligent generation module, a process guidance module, a real-time monitoring module, a minutes generation module and an assessment and evaluation module.

[0069] The data collection module is used to obtain headquarters policy documents and branch operations data. This module connects with headquarters' file storage systems and branch data management systems to collect relevant information on a scheduled or real-time basis. For headquarters policy documents, the module captures their textual content, including various rules and regulations, strategic plans, and more. For branch operations data, the module collects data on sales performance, production indicators, staff attendance, and other aspects, providing the foundation for subsequent analysis and processing.

[0070] The central processing module uses the BERT model for natural language processing to construct a policy text feature vector matrix. Its primary function is to perform in-depth analysis and feature extraction on headquarters policy documents acquired by the data collection module. First, the policy text is input into the BERT model, where the model's encoder processes the text to obtain the policy text's feature vector. The specific formula is:

[0071] V policy =BERT encoder (P text )

[0072] Among them, P text is the policy text sequence, V policy is a feature vector with dimension 512. This feature vector can effectively represent the semantic information of the policy text and provide a basis for the subsequent determination of the relevance to the branch operation.

[0073] The intelligent generation module determines topic weights using the Analytic Hierarchy Process (AHP). This module combines the policy text feature vectors generated by the central processing module with branch operating data, comprehensively considering various factors to determine the topics and their weights for the pre-shift meeting. When determining weights, the AHP method is used to break down complex issues into multiple levels and factors, and the relative importance of each factor is determined through pairwise comparisons. The calculation formula is:

[0074]

[0075] Where, α i is the weight of branch data indicators, β i,j is the policy relevance coefficient. This formula is used to calculate the weight of each topic, thus providing support for generating personalized pre-shift meeting agendas.

[0076] The process guidance module provides real-time guidance and prompts during pre-shift meetings. It monitors and manages the meeting process based on the agenda determined by the intelligent generation module. For example, if the meeting deviates from the agenda, the module can issue a prompt reminder; if a topic is discussed for an extended period, it can prompt the host to adjust the time limit. The module also features voice interaction, recognizing participants' voice commands and providing appropriate information and guidance based on those commands.

[0077] The real-time monitoring module monitors pre-shift meetings in real time. It uses cameras, microphones, and other devices to capture video and audio from the meeting. For video, computer vision technologies, such as posture recognition and facial expression analysis, are used to assess attendees' attention and engagement. For audio, speech recognition and sentiment analysis are used to understand attendees' speeches and emotional states. This real-time monitoring allows for timely identification of any issues during meetings and provides a basis for subsequent adjustments.

[0078] The minutes generation module automatically generates meeting minutes after the pre-shift meeting. This module combines information collected by the real-time monitoring module with recorded speeches during the meeting to summarize and refine the meeting content. Using natural language processing techniques, such as text summarization algorithms, it extracts key information and decision-making outcomes from the meeting and generates meeting minutes in a specific format. Furthermore, this module structures meeting minutes, categorizing and organizing different types of information for easier review and management.

[0079] The assessment and evaluation module evaluates the quality and effectiveness of pre-shift meetings. Based on information provided by the real-time monitoring module and the minutes generation module, it assesses meetings from multiple perspectives, such as meeting completion, participant engagement, and meeting time management. By establishing a multi-dimensional scoring system, it provides a comprehensive score for each pre-shift meeting and offers improvement recommendations based on the results. This module also analyzes and mines historical meeting data to provide data support for optimizing subsequent pre-shift meetings.

[0080] The process guidance module includes a speech recognition subsystem, which uses the CTC-loss algorithm for speech transcription. The acoustic model training formula is:

[0081]

[0082] Where T is the time step, C is the character set size, π t,c is the alignment probability of character c at time t. The primary function of the speech recognition subsystem is to accurately convert speech information from the meeting site into text. When training the acoustic model, by minimizing the CTC-loss function, the model can better learn the mapping relationship between speech signals and characters. In practice, this subsystem can convert participant speeches into text in real time and display it on the meeting interface, making it convenient for the host and other participants to review and record. Furthermore, the recognized text can be further semantically analyzed to determine whether the speech content is relevant to the meeting agenda, thereby providing more accurate information for meeting guidance and management.

[0083] The anomaly detection module constructs a semantic deviation model through the LSTM network, and the formula is:

[0084] D deviation =sigmoid(W h ·h t +b)

[0085] Among them, h t is the LSTM hidden layer state, W h is the weight matrix, D deviation The threshold is set to 0.7. The main function of the anomaly detection module is to detect semantic deviations in speech content during meetings. During meetings, participants' speeches may deviate from the meeting topic. The LSTM network models and analyzes the speech content and calculates the degree of semantic deviation. When the semantic deviation exceeds the set threshold (0.7), the system issues a reminder, prompting the host to guide the speech, ensuring that the meeting stays on topic and improving efficiency and quality.

[0086] The real-time monitoring module uses the OpenPose algorithm to recognize human postures and calculate the concentration of participants:

[0087]

[0088] Among them, θ iis the angle of gaze deviation of the i-th person, and N is the number of participants. The real-time monitoring module collects video information from the meeting site through a camera and uses the OpenPose algorithm to identify the body posture of participants, particularly their gaze direction. By calculating the gaze deviation angle, the participant's attention concentration is determined. This metric can intuitively reflect the participant's level of focus during the meeting. If the majority of participants' attention is low at a given moment, it may mean that the meeting content is not engaging enough or the meeting pace needs to be adjusted. The host can take timely measures based on this metric, such as adjusting the presentation method or adding interactive elements, to improve participant engagement and meeting effectiveness.

[0089] The video analysis module detects the conference scene using the YOLOv5 model, and its loss function formula is:

[0090] L yolo =λ coord ·L box +λ obj ·L obj +λ cls ·L cls

[0091] The main function of the video analysis module is to detect and analyze the video scene of the conference. Using the YOLOv5 model, the module can identify various objects and targets in the conference scene, such as participants, tables and chairs, equipment, etc. By minimizing the loss function L yolo , continuously optimizing the detection performance of the model. coord ,λ obj ,λ cls are the weight coefficients of the losses in different parts, L box is the bounding box loss, which is used to measure the position error between the detection box and the real box; L obj is the target confidence loss, reflecting the accuracy of the model’s judgment on the existence of the target; L cls Classification loss is used to determine the category of the detected target. Through accurate detection and analysis of the meeting scene, the system can further understand the layout and situation of the meeting site, providing a reference for subsequent meeting management and optimization.

[0092] The summary generation module uses the TextRank algorithm to generate summaries, and its iterative formula is:

[0093]

[0094] Where d is the damping coefficient, which is 0.85, and w jiis the edge weight. After the meeting, the minutes generation module needs to summarize and refine the meeting content to produce concise and clear minutes. The TextRank algorithm is a graph-based ranking algorithm that treats sentences in the meeting text as nodes in the graph and calculates edge weights based on the similarity between sentences. By repeatedly iterating the above formula, the score of each sentence is calculated, and sentences with high scores are designated as key sentences. The system selects these key sentences to form the core content of the meeting minutes, thereby effectively summarizing the meeting content.

[0095] The structured processing module applies regular expressions to extract key information. The regular pattern is:

[0096] Regex={(Task|Decision):\w+,\s+\d4-\d2-\d2}

[0097] The main task of the structured processing module is to further structure the generated meeting minutes. During meetings, various types of information are generated, such as task assignments and decision results. By applying regular expressions, this module accurately extracts this key information from the meeting minutes and organizes it into a specific format. For example, for task assignments, the module extracts the task name and completion date; for decision results, the module extracts the decision content and related time information. This results in better structure and readability of the generated meeting minutes, making them easier to review and manage.

[0098] refer to Figure 2 As shown, an intelligent pre-shift meeting method based on artificial intelligence includes the following steps:

[0099] S1: Integrate data, extract policy cores using NLP, and generate an issue pool;

[0100] S2: Dynamically adjust the priority of topics and generate a detailed agenda;

[0101] S3: guides the meeting process, monitors content and indicates anomalies;

[0102] S4: Monitor audio and video data to assess participant engagement;

[0103] S5: Transcribe the recording, automatically generate and send meeting minutes;

[0104] S6: Evaluate meeting quality and provide optimization suggestions.

[0105] During the meeting, the conference process guidance step is executed, which includes real-time voice analysis and progress management mechanism;

[0106] Real-time speech analysis uses the VAD algorithm for endpoint detection. The core of this algorithm is to determine the start and end points of speech by analyzing the energy of the speech signal. The formula is:

[0107]

[0108] Among them, E voice It is an indicator used to judge speech activity. T is the time length of the speech signal, that is, the number of time segments into which the speech signal is divided; s t is the energy value of the speech signal at the tth moment; μ is the mean value of the speech signal energy, which reflects the average level of the speech signal energy; σ 2 is the variance of the speech signal energy, which describes the discrete degree of the speech signal energy; τ is the set threshold, when E voice When it is greater than τ, it is considered that there is speech activity. In this way, the silent part in the speech signal can be effectively removed, improving the accuracy and efficiency of subsequent speech recognition.

[0109] The progress management mechanism uses a dynamic programming algorithm to optimize the agenda time allocation. The goal is to minimize the deviation between the actual discussion time and the planned discussion time, while ensuring that the total discussion time does not exceed the set upper limit. The objective function is:

[0110]

[0111] The constraints are in, represents the actual discussion time of topic i, is the planned discussion time of topic i, n is the number of topics, T total It is the upper limit of the total time of the meeting. The dynamic programming algorithm decomposes the complex agenda time allocation problem into multiple sub-problems and solves the optimal solution of the sub-problems to obtain the optimal solution of the original problem, thereby effectively managing the meeting progress, ensuring that the meeting can proceed according to the predetermined agenda and time schedule, and improving meeting efficiency.

[0112] During the meeting, a real-time monitoring step is performed, which includes a participation calculation model and an abnormality early warning mechanism;

[0113] The participation calculation model is used to evaluate the degree of participation of participants in the meeting. It takes into account the number of speeches and the interval between speeches. The formula is:

[0114]

[0115] Among them, P participation represents the participation of the participants, N is the number of participants, that is, the total number of people participating in the pre-shift meeting; S iis the number of speeches of the i-th participant. The more speeches the participant makes, the more actively the participant participates in the meeting. i is the speaking interval of the i-th participant. The shorter the interval, the higher the enthusiasm of the participant. By calculating this indicator, the activity level of the meeting and the enthusiasm of the participants can be intuitively reflected, providing a reference for subsequent meeting adjustments and improvements. The abnormal warning mechanism detects outliers through the isolation forest algorithm. The isolation forest algorithm is an anomaly detection algorithm based on a tree structure. It divides data points into different leaf nodes by constructing multiple decision trees. The anomaly score formula is:

[0116] s(xψ)=2 -E(h(x)) / c(ψ)

[0117] Among them, s(x,ψ) is the anomaly score of data point x. The higher the score, the more likely the data point is an outlier. E(h(x)) is the average path length of data point x in the isolation forest. The shorter the path length, the more likely the data point is an outlier. c(ψ) is an expected value related to the number of samples and is used to normalize the path length. When the anomaly score exceeds the set threshold, the system will issue an early warning to remind the meeting organizer or host to pay attention to possible abnormal situations, such as abnormal behavior of individual participants or sudden deviation from the meeting discussion direction, so that timely measures can be taken to deal with them.

[0118] After the meeting, the meeting minutes generation step is executed, which includes two parts: multimodal information fusion and structured output;

[0119] Multimodal information fusion uses the attention mechanism to fuse speech and text features. During a meeting, multiple types of information such as speech and text are generated. In order to make full use of this information, effective fusion is required. The attention mechanism can perform weighted fusion based on the importance of different information. The formula is:

[0120]

[0121] Among them, a i,j is the attention weight, which indicates the importance of the j-th feature to the i-th query during the fusion process; q i is the query vector, which is used to represent the information that needs attention at present; k j The key vector is used to calculate the similarity with the query vector to determine the attention weight. In this way, the key information in the meeting can be extracted more accurately, improving the quality and accuracy of the meeting minutes.

[0122] Structured output is generated into a standard format through a template matching algorithm. To make the generated meeting minutes more standardized and readable, and to facilitate access and management, they need to be organized according to the standard format. The template matching degree calculation formula is:

[0123]

[0124] Among them, M match represents the template matching degree, which is used to measure the similarity between the generated meeting minutes and the standard template; F is the field set, that is, all the fields that need to be matched in the meeting minutes; v f is the value of field f in the actual information; is the value of field f in the standard template; |F| represents the number of elements in the field set F. When the template matching degree reaches a certain threshold, it means that the generated meeting minutes meet the standard format requirements. Through continuous adjustment and optimization, the structured output of meeting minutes can be ensured.

[0125] After the meeting, a meeting quality assessment step is performed, which includes a multi-dimensional scoring system and the generation of improvement suggestions;

[0126] The multi-dimensional scoring system is used to comprehensively evaluate the quality of pre-shift meetings, taking into account multiple dimensions such as meeting completion, participant engagement, and meeting time management. The formula is:

[0127] Q score =ω1·C completion +ω2·P participation +ω3·T timeliness

[0128] Among them, Q score is the comprehensive score of the meeting, ω1, ω2, and ω3 are the weight coefficients of the three dimensions of meeting completion, participant participation, and meeting time control, respectively. Here, ω1 = 0.4, ω2 = 0.3, and ω3 = 0.3; C completion Indicates the degree of completion of the meeting agenda and reflects whether the meeting has completed various tasks according to the predetermined plan; participation The participation of participants reflects the activity and attractiveness of the meeting; T timeliness The control of meeting time measures whether the meeting starts and ends on time. This scoring system can intuitively reflect the quality level of the meeting and provide a basis for subsequent meeting improvements.

[0129] Improvement suggestion generation uses the Apriori algorithm to mine association rules. Association rule mining can discover the potential relationship between different events in the meeting data, thereby providing valuable suggestions for meeting improvements. The support formula is:

[0130]

[0131] The confidence formula is:

[0132]

[0133] in, It is an association rule The support is used to measure the frequency of the rule in the data set. The higher the support, the more common the rule. It is an association rule The confidence level is used to measure the probability of Y occurring when X occurs. The higher the confidence level, the more reliable the rule. T represents the dataset, i.e., the set of all meeting records. X and Y are data item sets, representing different events or conditions respectively. |T| is the number of elements in the dataset T.

[0134] |T∩(X∪Y)| represents the number of records containing both X and Y; sup(X∪Y) is the support for the simultaneous occurrence of X and Y; sup(X) is the support for the occurrence of X. By mining association rules, we can discover, for example, the relationship between prolonged discussion of certain topics and low participant participation, thereby providing targeted improvement suggestions.

[0135] The system adopts cloud deployment architecture and data security mechanism;

[0136] The cloud deployment architecture adopts a microservices architecture, which splits the system into multiple small, autonomous services to facilitate development, deployment, and maintenance. The service call delay calculation formula is:

[0137]

[0138] Among them, D latency represents the delay time of service calls in the cloud deployment architecture, reflecting the efficiency of service calls; n is the number of service calls; T send,i is the sending time of the i-th service call, that is, the time required for the request to be sent from the client to the server; T proc,i is the processing time of the i-th service call, that is, the time required for the server to process the request; T recv,i is the reception time of the i-th service call, that is, the time required for the server to return the processing result to the client;

[0139] The data security mechanism uses the AES-256 encryption algorithm to ensure the security of data during transmission and storage. The key generation formula is:

[0140]

[0141] Among them, K is the key generated by the AES-256 encryption algorithm; SHA-256 is a hash algorithm used to hash the input and generate a hash value of fixed length; seed is a random seed that provides randomness for generating the key; timestamp is a timestamp that represents the current time information; It is an exclusive OR operator used to perform an exclusive OR operation on the random seed and timestamp to increase the randomness and security of the key. At the same time, the data transmission process adopts a secure transmission protocol, such as the TLS protocol, to prevent data from being stolen or tampered with during transmission, ensuring the security of data throughout its life cycle and protecting meeting-related data from illegal acquisition and use.

[0142] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions merely illustrate the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims

1. An intelligent pre-shift meeting system based on artificial intelligence, characterized by: include: Data acquisition module, central processing module, intelligent generation module, process guidance module, real-time monitoring module, minutes generation module and assessment and evaluation module; The data collection module is used to obtain headquarters policy documents and branch operation data before the pre-shift meeting; The central processing module is used to perform natural language processing using the BERT model, construct a policy text feature vector matrix, and perform in-depth analysis and feature extraction on the obtained headquarters policy documents. The model algorithm expression is: In policy =BERT encoder (P text ) Where V policy The feature vector obtained after the policy text is encoded by the BERT model has a dimension of 512. This vector is used for the subsequent semantic analysis of the policy text and the calculation of the correlation with the branch operation data. encoder represents the encoder part of the BERT model, P text A text sequence representing headquarters policy documents, which is the input to the BERT encoder; The intelligent generation module is used to combine the policy text feature vector obtained by the central processing module and the branch operation data to determine the issue weight through the hierarchical analysis method. The calculation formula is: Where W topic represents the weight of each topic in the pre-shift meeting. The weight can be used to determine the importance of the topic in the meeting and then generate the meeting agenda. n represents the number of branch data indicators, that is, the number of types of branch operation data indicators considered. α i represents the weight of branch data indicators, β i,j It represents the policy correlation coefficient, reflecting the degree of correlation between each branch's data indicators and the headquarters' policies; The process guidance module is used to monitor and manage the process of the meeting according to the agenda determined by the intelligent generation module; The real-time monitoring module is used to collect video and audio information of the conference site through sound and image acquisition equipment, and monitor the conference situation in real time; The minutes generation module is used to combine the information collected by the real-time monitoring module and the speech records during the meeting to summarize and refine the meeting content and generate meeting minutes; The assessment and evaluation module is used to assess the quality and effectiveness of pre-shift meetings based on the information provided by the real-time monitoring module and the minutes generation module.

2. The intelligent pre-shift meeting system based on artificial intelligence according to claim 1 is characterized in that: The process guidance module includes a speech recognition subsystem, which uses the CTC-loss algorithm for speech transcription. The acoustic model training formula is: Where, L ctc represents the loss function in the CTC-loss algorithm, which is used to train the acoustic model. T represents the time step, that is, the number of time segments into which the speech signal is divided. C represents the character set size, which refers to the number of all characters that may appear in speech recognition. π tc represents the alignment probability of character c at time t, p t,c represents the probability of predicting character c at time t; the anomaly detection module constructs a semantic deviation model through the LSTM network, and the formula is: D deviation =sigmoid(W h ·h t +b) Where D deviation It represents the semantic deviation, which can be used to judge whether the speech content in the meeting deviates from the meeting theme. When the semantic deviation exceeds the predetermined threshold, a reminder is issued. Sigmoid represents the activation function. W h Represents the weight matrix, which is used to adjust the influence of the hidden layer state in the LSTM network. t It represents the hidden state of the LSTM network at the tth moment, which contains the information of the previous moment and is used to capture the long-term dependency in the sequence data. b represents the bias term.

3. The intelligent pre-shift meeting system based on artificial intelligence according to claim 1 is characterized in that: The real-time monitoring module uses the OpenPose algorithm to recognize human postures and calculate the concentration of participants: Where A concentration represents the concentration of participants, N represents the number of participants, θ i represents the deviation angle of the i-th participant’s sight. The larger the angle, the less focused the participant is. The video analysis module detects the meeting scene using the YOLOv5 model, and its loss function formula is: L yolo =λ coord ·L box +λ obj ·L obj +λ cls ·L cls Where, L yolo Represents the loss function of the YOLOv5 model, which is used to optimize the detection performance of the model, λ coord ,λ obj ,λ cls Represents the weight coefficients of different parts of the loss, which are used to adjust the relative importance of bounding box loss, target confidence loss and classification loss in the total loss. box Represents the bounding box loss, which is used to measure the position error between the detection box and the real box, L obj Represents the target confidence loss, which is used to evaluate the accuracy of the model's judgment on the existence of the target. cls Represents the classification loss, which is used to determine the category to which the detected target belongs.

4. The intelligent pre-shift meeting system based on artificial intelligence according to claim 1 is characterized in that: The summary generation module uses the TextRank algorithm to generate summaries, and its iterative formula is: In the formula, S(v i ) represents the sentence v in the TextRank algorithm i The score of d is the damping coefficient, which is used to control the randomness of the iteration process. i ) indicates pointing to sentence v i The set of all sentences that are related to sentence v i There are related sentences, Out(v j ) represents the sentence v j The set of all sentences pointed to by reflects the sentence v j The associated range, w ji Represents the edge weight, which measures the sentence v j With sentence v i The structured processing module applies regular expressions to extract key information, and the regular pattern is: Regex={(Task|Decision):\w+,\s+\d4-\d2-\d2} In the formula, Regex represents a regular pattern for extracting key information from meeting minutes, \w+ represents matching one or more word characters, which is used to match the name of a task or decision, \s+ represents matching one or more whitespace characters, and \d4-\d2-\d2 is used to match the date format. \d4 represents matching four digits, corresponding to the year, and \d2 represents matching two digits, corresponding to the month and day, respectively.

5. An intelligent pre-shift meeting method based on artificial intelligence, characterized in that: The following steps are involved: S1. Data integration and analysis: Collect headquarters policy documents and branch operation data, extract the core content of the policy through natural language processing technology, and generate a preliminary topic pool based on the branch data characteristics. Specifically, the headquarters policy documents are converted into a TF-IDF feature matrix using the formula: Where TF i,j Indicates word frequency, DF i represents the document frequency, and N represents the total number of documents; S2. Personalized agenda generation: Dynamically adjust the priority of topics based on the current operating status of the branch. Algorithms are used to match policy relevance with the actual needs of the branch to generate a detailed meeting agenda that includes discussion time and key content. Topic combinations are optimized using a genetic algorithm, and the fitness function is: Where, F fitness Represents the fitness function in the genetic algorithm, which is used to evaluate the quality of each topic combination, m represents the number of topics, γ k represents the topic priority coefficient, T k represents the actual discussion time of topic k, represents the planned discussion duration of topic k; S3. Meeting process guidance: This system monitors meeting content in real time through voice recognition and semantic analysis, automatically detecting abnormalities such as deviations from the topic and overtime discussions, prompting the host to adjust the agenda through voice or text, and simultaneously tracking the progress of each topic and reminding the time nodes; S4. Multi-dimensional real-time monitoring: Using cameras and microphones to collect audio and video data from the meeting, computer vision is used to analyze the posture and attention of participants. Participation indicators are calculated based on speaking frequency, and a real-time data dashboard of the meeting execution is generated simultaneously. S5. Intelligent Minutes Generation: Transcribe the entire meeting recording into text, extract key decision-making, task assignment, and other information through natural language processing technology, automatically generate structured meeting minutes, format them according to preset templates, and simultaneously send them to relevant personnel and headquarters for archiving; S6. Quality Assessment and Feedback: Establish a scoring model based on indicators such as agenda completion, participation, and time compliance to generate a quantitative assessment report; analyze common issues through historical data mining, provide optimization suggestions to headquarters, and dynamically adjust subsequent meeting generation strategies.

6. The method for intelligent pre-shift meeting based on artificial intelligence according to claim 5, characterized in that: Execute meeting process guidance steps during the meeting, including real-time voice analysis and progress management mechanism; Among them, real-time speech analysis uses the VAD algorithm for endpoint detection, and its expression is: Where, E voice An indicator used to judge speech activity, T represents the time length of the speech signal, s t represents the energy value of the speech signal at the tth moment, μ represents the mean value of the speech signal energy, σ 2 Represents the variance of the speech signal energy, τ is the set threshold, when E voice When it is greater than τ, it is considered that there is speech activity; Among them, the progress management mechanism uses a dynamic programming algorithm to optimize the agenda time allocation, and its objective function is: The constraints are Where, represents the actual discussion time of topic i, is the planned discussion time of topic i, n is the number of topics, T total It is the upper limit of the total time of the meeting. The dynamic programming algorithm decomposes the complex agenda time allocation problem into multiple sub-problems and solves the optimal solution of the sub-problems to obtain the optimal solution of the original problem, thereby effectively managing the meeting progress.

7. The method for intelligent pre-shift meeting based on artificial intelligence according to claim 5 is characterized in that: Perform real-time monitoring steps while the meeting is ongoing; The participation calculation model is used to evaluate the participation of participants in the meeting. The formula is: Where, P participation Indicates the participation of participants, N represents the number of participants, that is, the total number of people participating in the pre-shift meeting, S i is the number of speeches made by the i-th participant, δ i is the speaking interval of the i-th participant; Outliers are detected using the isolation forest algorithm, and the anomaly score formula is: s(xψ)=2 -E(h(x)) / c(ψ) Where s(x,ψ) represents the anomaly score of data point x, E(h(x)) represents the average path length of data point x in the isolation forest, and c(ψ) is an expected value related to the number of samples, which is used to normalize the path length.

8. The method for intelligent pre-shift meeting based on artificial intelligence according to claim 5 is characterized in that: After the meeting, the meeting minutes generation step is executed. This step includes two parts: multimodal information fusion and structured output. Multimodal information fusion uses the attention mechanism to fuse speech and text features, and the structured output generates a standard format through the template matching algorithm.

9. The method for intelligent pre-shift meeting based on artificial intelligence according to claim 5, characterized in that: The quality of the pre-shift meeting is evaluated through a multi-dimensional scoring system, the formula is: Q score =ω1·C completion +ω2·P participation +ω3·T timeliness Where Q score represents the comprehensive score of the meeting, ω1, ω2, and ω3 represent the weight coefficients of the three dimensions of meeting completion, participant participation, and meeting time control, respectively. completion Indicates the degree of completion of the meeting agenda, P participation Indicates the participation of participants, T timeliness It is about the control of meeting time.

10. The method for intelligent pre-shift meeting based on artificial intelligence according to claim 5, characterized in that: A cloud deployment architecture and data security mechanism are adopted. The cloud deployment architecture adopts a microservice architecture, and the data security mechanism adopts an AES-256 encryption algorithm.

Images