A device field collective intelligence representation and hidden horse prediction method

Abstract

The application discloses a device field collective intelligence representation and hidden Markov prediction method, comprising the following steps: step S1, constructing a data acquisition terminal; step S2, constructing a collective intelligence feature evaluation system based on viscosity indexes, uniformity indexes and flexibility indexes; step S3, constructing a candidate index set, screening the candidate indexes based on log-likelihood values, and fusing the final indexes screened; step S4, constructing and screening a group hidden Markov model; step S5, feature migration and state prediction of new group data; and step S6, an intervention mechanism based on hidden state semantic labels. The application can be used for real-time evaluation and prediction of team cooperation and collective intelligence of a group, and further guides external intervention and intervention. The application scheme can also collect feature indexes through a non-contact acquisition model, and has the advantages of strong popularization and no excessive interference on a process.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence application technology, and in particular to one such application. Background Technology

[0002] In the field of modern engineering education, traditional practical teaching is shifting from single-function experiments to a more open and inquiry-based project-based learning model. Against this backdrop, teaching scenarios based on group collaboration to complete project development are being used more widely. As a result, the concept of an equipment field has emerged. An equipment field refers to the work site where learners use equipment to carry out project practice activities. Unlike the traditional concept of equipment storage places such as laboratories, it is a dynamic concept that includes operating equipment, learners using the equipment, and ongoing research projects.

[0003] Psychological research suggests that a "general collective intelligence factor," or collective intelligence, exists within a group during collaborative processes. The equipment environment effectively provides a scenario for the emergence of collective intelligence. In this scenario, members, through balanced interaction and close collaboration, form a team problem-solving ability that surpasses the sum of individual intelligence.

[0004] Therefore, assessing and predicting the teamwork or collective intelligence level of a group is not only beneficial for promoting positive progress but also allows for timely intervention when collaboration problems arise. However, assessing collaboration levels and collective intelligence presents numerous challenges, such as: 1. Lack of dynamic modeling of the "hidden states" of group collaboration. Group collaboration is a complex stochastic process, and its internal "intelligent states" are often not directly observable. Existing technologies mostly employ direct statistical mapping, lacking a mathematical model to establish a probabilistic relationship between observable physical behavior and unobservable collective intelligence states. Due to the lack of modeling of this implicit state transition mechanism, the system cannot understand the inherent evolutionary laws of group collaboration. 2. Lack of real-time or early prediction capabilities. Existing assessment methods often rely on post-project questionnaires, output scores after project completion, or analysis based on static averages over a period of time. This evaluation ignores the dynamic changes of collective intelligence over time. Especially in engineering practice teaching, teachers urgently need a process prediction technique—one that can predict the group's potential collaborative performance from weak behavioral signals in the early or middle stages of a project, enabling timely intervention. 3. Data collection is difficult, deployment costs are high, and it is highly disruptive. For example, existing studies either require participants to wear special sociometric badges or organize specialized psychological tests to assess social perception abilities. This data acquisition method, which relies on special wearable devices or additional testing procedures, is not only expensive in terms of hardware but also highly disruptive in real-world, routine teaching scenarios, making it difficult to deploy seamlessly on a large scale in ordinary online teaching platforms or smart classrooms.

[0005] To address the aforementioned issues, this proposal presents a method for collective intelligent representation and hidden horse prediction of equipment fields. Summary of the Invention

[0006] To address the aforementioned technical problems, embodiments of the present invention provide a method for collective intelligent representation and hidden horse prediction of equipment fields, comprising the following steps: Step S1: Deploy data acquisition terminals in the equipment field to obtain the time-series logs and outputs of each team member; Step S2: Based on the physical space data and digital space data obtained in step S1, calculate the viscosity index V, uniformity index E, and flexibility index F of the group. Step S3: Construct a candidate indicator set, evaluate and screen each indicator in the candidate indicator set, and select the final indicator to form an unlabeled observation set; Step S4: Use the unlabeled observation set from step S3 as training data to train several candidate Hidden Markov Models in the group, construct a model evaluation index, and use the model evaluation index to screen the candidate Hidden Markov Models in the group to obtain and output the optimal Hidden Markov Prediction Model. Step S5: Connect the optimal hidden Markov prediction model obtained in step S4 to the new group. At the same time, use the real-time observation vector obtained and solved by the data acquisition terminal set in step S1 as the input parameter, and use the optimal prediction model to predict and output the prediction status of the group in real time. Step S6: Construct an intervention state set, judge the predicted state output in step S5 based on the intervention state set, and execute the corresponding preset strategy according to the judgment result.

[0007] The data acquisition terminal includes a surveillance camera and a data acquisition system. The surveillance camera can acquire video streams of the team collaborating within the equipment area, and the control system can analyze and generate time-series logs of each member within the equipment area based on the video streams. The data acquisition system can be connected to the group's data platform and obtain the output results of each member of the group.

[0008] The viscosity index V in step S2 includes: multi-person collaborative presence time Vt and multi-person collaborative presence count Vc; The multi-person collaborative presence time Vt refers to the total duration of a scenario in which the number of people in a group are present at the same time in the device field of the video stream is greater than 50% of the total number of people in the group. The number of times Vc of multiple people are present in collaboration refers to the total number of times the number of people in a group at the same time in the device field recorded in the video stream is greater than 50% of the total number of people in the group. The uniformity index E in step S2 includes: the standard deviation of the time spent in the field Et and the standard deviation of the number of times multiple people are present Ec. The formula for calculating the standard deviation of the time spent in the field Et is as follows:

[0009] Where N is the total number of people in the group, and Ti is the individual attendance time of each group member as recorded in the time-series log. The average on-site time of group members is statistically analyzed in the time-series log, where i is the group index; The formula for calculating the standard deviation Ec of the number of times multiple people are present is:

[0010] Where N is the total number of people in the group, and Ci is the number of times each group member was present, as recorded in the time-series log. The average number of times each group member was present, as statistically analyzed in the time-series log. The compliance index F in step S2 is a score of the group's output, which combines the main evaluator's score for the group and the mutual scores between groups. Its calculation formula is as follows:

[0011] Where α is the rating weighting coefficient of the primary rater. The main evaluator scores the output of group i in stage t, where M is the number of groups participating in the peer review. This is the score given by group k to group i for the results produced in stage t.

[0012] The candidate index set in step S3 includes: Viscosity index candidate set SV: {multi-person collaborative presence time Vt, number of multi-person collaborative presences Vc}; The candidate set of uniformity indicators SE is: {standard deviation of time spent in the field Et, standard deviation of the number of times multiple people are present Ec}.

[0013] Step S3 includes the following steps: Step S31: For each candidate indicator in the candidate indicator set, construct a single-variable hidden Markov model. Step S32: Train the univariate hidden Markov model using the data corresponding to the index from the collected historical dataset; Step S33: Calculate the log-likelihood value of the observed data generated by the trained model. The formula for calculating the log-likelihood value is as follows:

[0014] in, Candidate indicators The log-likelihood value, This indicates the candidate metrics currently being evaluated. For the historical dataset used in the evaluation, Indicates based on candidate metrics The dataset of the univariate hidden Markov model trained and converged, where P represents the probability calculation function; Step S34: Select the index with the largest log-likelihood value as the final feature of this dimension, and normalize the selected final feature to form the observation vector Ot of the group at time T:

[0015] Where Vtf is the final feature of the viscosity feature dimension, Etf is the final index of the uniformity feature dimension, and F is the compliance index. Step S35: Collect the observation vector Ot data corresponding to each group in multiple time periods to form an unlabeled observation sequence set O.

[0016] Step S4 includes the following steps: Step S41: Set the shrinkage range K for the number of hidden states. For each K value, several hidden Markov candidate models are obtained by training on an unlabeled observation sequence set O. ; Step S42: Calculate the Akaike Information Criterion (AIC) and reconstruction error L1 for each Hidden Markov candidate model as model evaluation indicators. Find the point where the reconstruction error L1 is significantly reduced and the Akaike Information Criterion (AIC) is at a low inflection point, and determine the corresponding K value as the optimal number of hidden states. Step S43: After determining the optimal number of states and completing the final convergence training, the output includes the following core parameters. The optimal hidden Markov prediction model; Where μ and Σ are Gaussian emission parameters, Let the initial state probability vector be defined as follows: one of them This represents the state S of a group at the very beginning of the task. i The probability, A is the transition matrix, defined as a K×K matrix, with matrix elements... Indicates the group from the current time S i State transition to the next moment S j The probability of a state is defined as:

[0017] In the formula, qt represents the hidden state at time t, and P represents the probability calculation function.

[0018] The Akaike Information Content Criterion (AIC) is used to evaluate the goodness of fit and complex balance of the model, preventing overfitting. Its formula is as follows:

[0019] Where PK is the number of model parameters, L K This is the maximum likelihood estimate of the model on the re-observed data.

[0020] The method for calculating the reconstruction error includes the following steps: Step S421: Calculate the posterior probability For each time point t in the time series, the forward-backward algorithm is used to calculate the posterior probability that the population is in the hidden state K at time t, given the entire set of observations O:

[0021] Step S422: Calculate the expected observations The mean of the Gaussian emission distribution of each hidden state The feature reconstruction vector at time t is generated by weighted summation of the posterior probabilities. The calculation formula is as follows:

[0022] Where μk is the mean vector of viscosity index V, uniformity index E and flexibility index F in the k-th state; Step S423: Calculate the distance between the reconstructed vector and the actual observed vector, and average it over all time steps and all feature dimensions D to obtain the reconstruction error. :

[0023] in, This represents the actual observation vector in the d-th dimension at time t. This represents the corresponding reconstructed vector value, where D represents the number of feature dimensions. Indicates a time step.

[0024] Step S5 includes the following steps: Step S51: Read the global statistics of viscosity index V, uniformity index E, and compliance index F in the training set, and process the observation vectors obtained in real time through the data acquisition terminal described in step S1. Synchronous standardization is performed to obtain the standard observation vector. ;

[0025] in, The mean of the global statistics. The standard deviation of the global statistic; Step S52: Extract the observation sequence of the group at the current time t and the past window length L. The Viterbi algorithm is used to find the most probable hidden state path on the optimal swarm hidden Markov model obtained in step S4. Output the specific state of the group at the current time t; Step S53: Using the state transition matrix A trained in step S52, predict the possible state of the group after K time steps. The calculation method is as follows:

[0026] in, Let be the state probability distribution vector at time t. This represents the transition matrix after K time steps.

[0027] Step S6 includes the following steps: Step S61: Define the set of intervention states Determine whether the future state predicted in step S5 belongs to the set of intervention states; Step S62: If it is determined that the future state belongs to the set of intervention states, a preset intervention signal is triggered and the intervention instruction is executed. If it is determined that the future state does not belong to the set of intervention states, the teamwork level of the group is determined to be normal.

[0028] The embodiments of this invention have the following beneficial effects: 1. For hidden states such as collective intelligence or team collaboration level that are not visible in the system, an evaluation system based on three dimensions—viscosity index, uniformity index, and flexibility index—is established for evaluation and prediction. 2. Based on the above-mentioned evaluation system of viscosity index, uniformity index, and flexibility index, combined with a hidden Markov model, a predictive model capable of real-time, early prediction is constructed, making it possible to predict team collaboration and collective intelligence during the project process. 3. The data acquisition terminal used in this solution is not only cheaper than existing solutions, but also less likely to interfere with the project process, enabling real-time data acquisition and using the real-time acquired data for immediate prediction and evaluation. Attached Figure Description

[0029] Figure 1 This is a schematic block diagram of the present invention; Figure 2 This is a schematic block diagram of step S1 of the present invention; Figure 3 This is a schematic block diagram of step S2 of the present invention; Figure 4 This is a schematic block diagram of step S3 of the present invention; Figure 5 This is a schematic block diagram of step S4 of the present invention; Figure 6 This is a schematic block diagram of step S5 of the present invention; Figure 7 This is a schematic block diagram of step S6 of the present invention. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings.

[0031] like Figure 1 As shown, the equipment field collective intelligence representation and Hidden Horse Prediction method includes the following steps: Step S1: Set up data acquisition terminals in the equipment field to collect time-series logs and outputs of each team member; Step S2: Based on the physical space data and digital space data obtained in step S1, calculate the viscosity index V, uniformity index E, and flexibility index F of the group. Step S3: Construct a candidate indicator set, evaluate and screen each indicator in the candidate indicator set, and select the final indicator to form an observation vector and an unlabeled observation set; Step S4: Use the unlabeled observation set from step S3 as training data to train several hidden Markov candidate models in a group, construct a model evaluation index, and use the evaluation index to screen the several hidden Markov candidate models in the group to obtain and output the optimal hidden Markov prediction model. Step S5: Connect the optimal hidden Markov prediction model obtained in step S4 to the new group. At the same time, use the real-time observation vector obtained and solved by the data acquisition terminal set in step S1 as the input parameter, and use the optimal prediction model to predict and output the prediction status of the group in real time. Step S6: Construct an intervention state set, judge the predicted state output in step S5 based on the intervention state set, and execute the corresponding preset strategy according to the judgment result.

[0032] To better illustrate the specific implementation of the above scheme, this embodiment will describe each step in detail in sequence.

[0033] Step S1: Deploy the data acquisition terminal.

[0034] The data acquisition terminal includes a surveillance camera and a data acquisition system. like Figure 2 As shown, the monitoring camera can acquire video streams of the team collaborating within the equipment area, and the control system can analyze and generate time-series logs of each member within the equipment area based on the video streams.

[0035] Specifically, surveillance cameras, such as behavior-aware cameras and result-aware cameras, are deployed in the equipment area. These cameras identify group members using a face recognition algorithm and generate video streams of each group member in the equipment area. The video streams are analyzed to record the time when each group member enters and leaves the equipment area, generating an accurate time-series log. During this process, noise data that is short-lived, meaning that the time a group member spends in the equipment area is less than ten minutes, is removed.

[0036] It should be noted that the data collection system can obtain the outputs of each team member at each stage of the project through an interface or by crawling. For example, the data collection system can access the team's document sharing platform and obtain the process outputs generated by each team member during the project, such as weekly reports, code, experimental data, project progress, or meeting minutes.

[0037] II. Step S2: Construct a collective intelligence feature evaluation system based on viscosity index V, uniformity index E, and flexibility index F.

[0038] like Figure 3 As shown, the time series logs and output results obtained in step S1 are used as inputs, and viscosity index V, uniformity index E, and flexibility index F are obtained through calculation and scoring.

[0039] Specifically, the viscosity index V includes: the time Vt of multiple people working together and the number of times multiple people work together.

[0040] The multi-person collaborative presence time Vt refers to the total duration of a scenario in which the number of people in a group are present at the same time in the device field of the video stream is greater than 50% of the total number of people in the group. The number of times Vc of multi-person collaboration is recorded in the time series log as the total number of scenarios in which the number of people in a group at the same time in the device field is greater than 50% of the total number of people in the group. Specifically, the uniformity index E includes: the standard deviation of the time spent on site Et and the standard deviation of the number of times multiple people are present Ec. The formula for calculating the standard deviation of the time spent on site Et is as follows:

[0041] Where N is the total number of people in the group, T i The time spent by each group member in the time-series log is statistically analyzed. The average time spent on site by group members, as statistically analyzed in the time-series log. The formula for calculating the standard deviation Ec of the number of times multiple people are present is:

[0042] Where N is the total number of people in the group, C i This refers to the number of times each group member was present, as counted in the time-series log. The average number of times each group member was present, as statistically analyzed in the time-series log. It should be noted that the compliance index F in step S2 is a score of the group's output, which combines the main evaluator's score for the group and the mutual scores between groups. Its calculation formula is as follows:

[0043] Where α is the rating weighting coefficient of the primary rater. The main evaluator scores the output of group i in stage t, where M is the number of groups participating in the peer review. The score given by group k to the output of group i in stage t is used as the flexible index F.

[0044] III. Step S3: Construct a candidate index set, filter the candidate indexes based on the log-likelihood value, and merge the final indexes obtained from the filtering.

[0045] like Figure 4 As shown, in order to construct the observation vector that best represents the pattern of group cooperation, this invention establishes candidate index sets in the dimensions of "viscosity" and "uniformity", and automatically selects the index with the highest fit by training a univariate hidden Markov model and calculating the log likelihood value.

[0046] The candidate index set in step S3 includes: Viscosity index candidate set SV: {multi-person collaborative presence time Vt, multi-person collaborative presence count Vc}.

[0047] The candidate set of uniformity indicators SE is: {standard deviation of time spent in the field Et, standard deviation of the number of times multiple people are present Ec}.

[0048] The log-likelihood-based optimization algorithm performs the following optimization process for each candidate metric in each dimension: Step S31: For each candidate indicator in the candidate indicator set, construct a single-variable hidden Markov model. Step S32: Train the univariate hidden Markov model using the data corresponding to the index from the collected historical dataset; Step S33: Calculate the log-likelihood value of the observed data generated by the trained model. The formula for calculating the log-likelihood value is as follows:

[0049] Where LL(x) is the log-likelihood value of the candidate indicator, x represents the candidate indicator currently being evaluated, and O history For the historical dataset used in the evaluation, denoted as the dataset of a univariate hidden Markov model trained and converged based on candidate metric x, where P represents the probability calculation function. Step S34: Select the indicator with the largest log-likelihood value as the final feature of this dimension:

[0050]

[0051] The selected viscosity features are obtained through multi-dimensional feature fusion. Uniformity characteristics Normalized with the compliance index F, the observation vector Ot of this group at time T is formed:

[0052] Among them, V tf E is the final feature of the viscosity feature dimension. tf Ft is the final index for the uniformity feature dimension, and Ft is the compliance index.

[0053] Step S35: Collect the observation vector Ot data corresponding to each group over multiple time periods to form an unlabeled observation sequence set O:

[0054] Where L represents the number of time periods, and the unlabeled observation sequence set O is the set of all observation vectors Ot from the first stage to the Lth stage. This unlabeled observation sequence set O will be used in step S4 for the training and construction of the swarm hidden Markov model.

[0055] IV. Step S4: Construction and Screening of Population Hidden Markov Models like Figure 5 As shown, step S4 includes the following steps: Step S41: Set the shrinkage range for the number of hidden states. For each K value, a candidate model is obtained by training on the unlabeled observation sequence set O. ; Step S42: Calculate the Akaike Information Criterion (AIC) and reconstruction error L1 for each candidate model. Find the point where the reconstruction error L1 is significantly reduced and the Akaike Information Criterion (AIC) is at a low inflection point, and determine the corresponding K value as the optimal number of hidden states. Step S43: After determining the optimal number of states and completing the final convergence training, the output includes the following core parameters. The optimal hidden Markov prediction model Where π is the initial state probability vector, defined as one of them This indicates that a group is in state t=0 at the beginning of the task. The probability, A is the transition matrix, defined as a K×K matrix with elements... Indicates the group from the current moment State transition to the next moment The probability of a state:

[0056] μ and Σ are Gaussian emission parameters and their mean vectors. ,in, The viscosity is the average value, indicating the average number of members present simultaneously under a given condition. The mean of softness represents the mean of the standard deviation of the member's time of presence. The mean of softness represents the average score F of the members in this state.

[0057] It should be noted that in step S42, this embodiment constructs an evaluation system for the multiple candidate models obtained in step S41, which includes two indicators: Akaike Information Content Criterion (AIC) and Reconstruction Error L1Loss.

[0058] 1. Akaike Information Content Criterion (AIC) The Akaike Information Content Criterion (AIC) is used to evaluate the goodness of fit and complex balance of a model, preventing overfitting. Its formula is as follows:

[0059] Among them, P K L represents the number of model parameters. K This is the maximum likelihood estimate of the model on the re-observed data.

[0060] 2. Reconstruction error L1Loss The method for calculating the reconstruction error includes the following steps: Step S421: Calculate the posterior probability For each time point in the time series Using the forward-backward algorithm, calculate the posterior probability that the population is in hidden state k at time t, given the entire set of observation sequences O:

[0061] in, Let P be the posterior probability, P be the probability calculation function, and qt represent the hidden state of the system at time t. This represents a hidden state belonging to the observation sequence set O. Step S422: Calculate the expected observations The mean of the Gaussian emission distribution of each hidden state The feature reconstruction vector at time t is generated by weighted summation of the posterior probabilities. The calculation formula is as follows:

[0062] in, For posterior probability, Let V be the mean vector of the viscosity index V, uniformity index E, and compliance index F at the k-th state. , The viscosity is the average value, indicating the average number of members present simultaneously under a given condition. The mean of softness represents the mean of the standard deviation of the member's time of presence. Let F be the mean of the softness score, representing the mean score F of the members in this state, and T represent the time.

[0063] Step S423: Calculate the reconstructed vector Compared with the actual observed vector The distance between them, and in all time steps The reconstruction error L1Loss is obtained by averaging across the feature dimensions.

[0064] in, This represents the actual observation vector in the d-th dimension at time t. This represents the corresponding reconstructed vector value, where D represents the number of feature dimensions, namely viscosity, uniformity, and softness. Indicates a time step.

[0065] Finally, based on the two evaluation metrics mentioned above, the optimal population hidden Markov model is selected in this step as the final prediction model for real-time prediction in this embodiment.

[0066] V. Step S5: Feature Transfer and State Prediction of New Population Data Using the optimal population hidden Markov model trained and with fixed parameters in step 4, the newly connected current student population is analyzed.

[0067] like Figure 6 As shown, step S5 includes the following steps: Step S51: Read the global statistics, i.e., the mean, of the viscosity index V, uniformity index E, and compliance index F in the training set. and standard deviation The observation vectors obtained and calculated in real time through the data acquisition terminal described in step S1 are processed. Synchronous standardization is performed to obtain the standard observation vector. :

[0068] Step S52: Extract the observation sequence of the group at the current time t and the past window length L. The Viterbi algorithm is used to find the most probable hidden state path on the optimal swarm hidden Markov model obtained in step S4. Output the specific state of the group at the current time t; Step S53: Utilize the state transition matrix A trained in step S52 K The method for predicting the group's possible state after a future time step is as follows:

[0069] Where, α t Let A be the state probability distribution vector at time t. K State transition matrix VI. Step S6: Intervention mechanism based on hidden state semantic labels.

[0070] like Figure 7 As shown, step S6 includes the following steps: Step S61: Define the set of intervention states Determine whether the future state predicted in step S5 belongs to the set of intervention states; Among them, S 低效态 and S 高效态 A set of states defined based on historical data.

[0071] Step S62: If it is determined that the future state belongs to the set of intervention states, then a preset intervention signal is triggered and the intervention instruction is executed. If it is determined that the future state does not belong to the set of intervention states, then the teamwork level of the group is judged to be normal.

[0072] It should be noted that after the optimal hidden Markov model training phase is completed, each hidden state needs to be analyzed in advance. Gaussian emission parameters The mean values ​​of viscosity index V, uniformity index E, and flexibility index F are assigned specific semantic labels. These labels are used to make attribution judgments on the viscosity index V, uniformity index E, and flexibility index F analyzed by the group in real time. A mechanism is established that can automatically judge when the team's level of collaboration declines. At the same time, combined with pre-set intervention instructions, a complete system of automatic judgment and autonomous intervention is formed.

[0073] To illustrate this step, this embodiment will be used as an example for detailed explanation. In this embodiment, based on the group's teaching objectives, a set of intervention states Q that require system intervention is pre-defined. int Taking the educational example in a school as described in this embodiment, based on teaching experience, the implicit state of a group's teamwork level has three states: Inefficient state: The mean of the observed vector in this state is significantly lower than the global average. From the state transition matrix, this state has an extremely high self-holding probability, meaning that once the group is trapped in this state, it is extremely difficult for it to spontaneously transition to a higher-order state without external stimulus input, and the system tends to a low-energy deadlock equilibrium.

[0074] Transitional state: In this state, observed indicators are at a moderate level, accompanied by significant variance fluctuations. The population lacks sufficient endogenous motivation to break through performance bottlenecks. Without external intervention, this state often exhibits stagnant oscillation characteristics and is unlikely to spontaneously evolve into an efficient cooperative mode.

[0075] High-efficiency state: In this state, all collaboration indicators remain within the optimal range. The team exhibits high robustness and self-organization capabilities. Even when faced with minor external disturbances, it can maintain its current high performance output by utilizing internal adjustment mechanisms with a high self-sustaining probability, forming a positive inertial closed loop.

[0076] Therefore, the intervention state set in this embodiment .

[0077] In conjunction with step S62 of this embodiment, when the system receives the optimal hidden state at the current time output by the Viterbi algorithm... or predicted future state Then, the following judgment logic will be used to determine q. t Does it belong to a set? .

[0078] like t If this happens, an intervention signal will be automatically triggered and the intervention will be executed, such as sending a mandatory collaboration instruction to all members, forcing them to aggregate in physical space.

[0079] like If the system detects a failure, it determines that the group's teamwork level is normal, and the system remains silent.

[0080] It should be noted that some algorithms and technical terms involved in this embodiment, such as the face recognition algorithm used in the surveillance camera in step S1, the feature fusion and normalization processing in step S3, the calculation method of the log-likelihood value in step S3, and the Viterbi algorithm used in step S6, are clearly unique and unambiguous to those skilled in the art, and there are many existing technologies as examples. Those skilled in the art can clearly reproduce the technology, so this solution will not be repeated here.

[0081] Of course, the above embodiments are only for illustrating the technical concept and features of the present invention, and their purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be used to limit the scope of protection of the present invention. All modifications made according to the spirit and essence of the main technical solution of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for collective intelligent representation and Hidden Horse Prediction of equipment fields, characterized in that, Includes the following steps: Step S1: Deploy data acquisition terminals in the equipment field to obtain the time-series logs and outputs of each team member; Step S2: Based on the physical space data and digital space data obtained in step S1, calculate the viscosity index V, uniformity index E, and flexibility index F of the group. Step S3: Construct a candidate indicator set, evaluate and screen each indicator in the candidate indicator set, and select the final indicator to form an unlabeled observation set; Step S4: Use the unlabeled observation set from step S3 as training data to train several candidate Hidden Markov Models in the group, construct a model evaluation index, and use the model evaluation index to screen the candidate Hidden Markov Models in the group to obtain and output the optimal Hidden Markov Prediction Model. Step S5: Connect the optimal hidden Markov prediction model obtained in step S4 to the new group. At the same time, use the real-time observation vector obtained and solved by the data acquisition terminal set in step S1 as the input parameter, and use the optimal prediction model to predict and output the prediction status of the group in real time. Step S6: Construct an intervention state set, judge the predicted state output in step S5 based on the intervention state set, and execute the corresponding preset strategy according to the judgment result.

2. The method for collective intelligent representation and Hidden Horse Prediction of equipment fields according to claim 1, characterized in that, The data acquisition terminal includes a surveillance camera and a data acquisition system. The surveillance camera can acquire video streams of the team collaborating within the equipment area, and the control system can analyze and generate time-series logs of each member within the equipment area based on the video streams. The data acquisition system can be connected to the group's data platform and obtain the output results of each member of the group.

3. The method for collective intelligent representation and hidden horse prediction of equipment fields according to claim 2, characterized in that, The viscosity index V in step S2 includes: multi-person collaborative on-site time Vt and multi-person collaborative on-site number Vc; The multi-person collaborative presence time Vt refers to the total duration of a scenario in which the number of people in a group are present at the same time in the device field of the video stream is greater than 50% of the total number of people in the group. The number of times Vc of multiple people are present in collaboration refers to the total number of times the number of people in a group at the same time in the device field recorded in the video stream is greater than 50% of the total number of people in the group. The uniformity index E in step S2 includes: the standard deviation of the time spent in the field Et and the standard deviation of the number of times multiple people are present Ec. The formula for calculating the standard deviation of the time spent in the field Et is as follows: Where N is the total number of people in the group, and Ti is the individual attendance time of each group member as recorded in the time-series log. The average on-site time of group members is statistically analyzed in the time-series log, where i is the group index; The formula for calculating the standard deviation Ec of the number of times multiple people are present is: Where N is the total number of people in the group, and Ci is the number of times each group member was present, as recorded in the time-series log. The average number of times each group member was present, as statistically analyzed in the time-series log. The compliance index F in step S2 is a score of the group's output, which combines the main evaluator's score for the group and the mutual scores between groups. Its calculation formula is as follows: Where α is the rating weighting coefficient of the primary rater. The main evaluator scores the output of group i in stage t, where M is the number of groups participating in the peer review. This is the score given by group k to group i for the results produced in stage t.

4. The method for collective intelligent representation and hidden horse prediction of equipment fields according to claim 1, characterized in that, The candidate index set in step S3 includes: Viscosity index candidate set SV: {multi-person collaborative presence time Vt, number of multi-person collaborative presences Vc}; The candidate set of uniformity indicators SE is: {standard deviation of time spent in the field Et, standard deviation of the number of times multiple people are present Ec}.

5. The method for collective intelligent representation and Hidden Horse Prediction of equipment fields according to claim 4, characterized in that, Step S3 includes the following steps: Step S31: For each candidate indicator in the candidate indicator set, construct a single-variable hidden Markov model. Step S32: Train the univariate hidden Markov model using the data corresponding to the index from the collected historical dataset; Step S33: Calculate the log-likelihood value of the observed data generated by the trained model. The formula for calculating the log-likelihood value is as follows: in, Candidate indicators The log-likelihood value, This indicates the candidate metrics currently being evaluated. For the historical dataset used in the evaluation, Indicates based on candidate metrics The dataset of the univariate hidden Markov model trained and converged, where P represents the probability calculation function; Step S34: Select the index with the largest log-likelihood value as the final feature of this dimension, and normalize the selected final feature to form the observation vector Ot of the group at time T: Where Vtf is the final feature of the viscosity feature dimension, Etf is the final index of the uniformity feature dimension, and F is the compliance index. Step S35: Collect the observation vector Ot data corresponding to each group in multiple time periods to form an unlabeled observation sequence set O.

6. The method for collective intelligent representation and hidden horse prediction of equipment fields according to claim 5, characterized in that, Step S4 includes the following steps: Step S41: Set the shrinkage range K for the number of hidden states. For each K value, several hidden Markov candidate models are obtained by training on an unlabeled observation sequence set O. ; Step S42: Calculate the Akaike Information Criterion (AIC) and reconstruction error L1 for each Hidden Markov candidate model as model evaluation indicators. Find the point where the reconstruction error L1 is significantly reduced and the Akaike Information Criterion (AIC) is at a low inflection point, and determine the corresponding K value as the optimal number of hidden states. Step S43: After determining the optimal number of states and completing the final convergence training, the output includes the following core parameters. The optimal hidden Markov prediction model; Where μ and Σ are Gaussian emission parameters, Let the initial state probability vector be defined as follows: one of them This represents the state S of a group at the very beginning of the task. i The probability, A is the transition matrix, defined as a K×K matrix, with matrix elements... Indicates the group from the current time S i State transition to the next moment S j The probability of a state is defined as: In the formula, qt represents the hidden state at time t, and P represents the probability calculation function.

7. The method for collective intelligent representation and Hidden Horse Prediction of an equipment field according to claim 6, characterized in that, The Akaike Information Content Criterion (AIC) is used to evaluate the goodness of fit and complex balance of a model, preventing overfitting. Its formula is as follows: Where PK is the number of model parameters, L K This is the maximum likelihood estimate of the model on the re-observed data.

8. The method for collective intelligent representation and hidden horse prediction of equipment fields according to claim 6, characterized in that, The method for calculating the reconstruction error includes the following steps: Step S421: Calculate the posterior probability For each time point t in the time series, the forward-backward algorithm is used to calculate the posterior probability that the population is in the hidden state K at time t, given the entire set of observations O: Step S422: Calculate the expected observations The mean of the Gaussian emission distribution of each hidden state The feature reconstruction vector at time t is generated by weighted summation of the posterior probabilities. The calculation formula is as follows: Where μk is the mean vector of viscosity index V, uniformity index E and flexibility index F in the k-th state; Step S423: Calculate the distance between the reconstructed vector and the actual observed vector, and average it over all time steps and all feature dimensions D to obtain the reconstruction error. : in, This represents the actual observation vector in the d-th dimension at time t. This represents the corresponding reconstructed vector value, where D represents the number of feature dimensions. Indicates a time step.

9. The method for collective intelligent representation and hidden horse prediction of equipment fields according to claim 1, characterized in that, Step S5 includes the following steps: Step S51: Read the global statistics of viscosity index V, uniformity index E, and compliance index F in the training set, and process the observation vectors obtained in real time through the data acquisition terminal described in step S1. Synchronous standardization is performed to obtain the standard observation vector. ; in, The mean of the global statistics. The standard deviation of the global statistic; Step S52: Extract the observation sequence of the group at the current time t and the past window length L. The Viterbi algorithm is used to find the most probable hidden state path on the optimal swarm hidden Markov model obtained in step S4. Output the specific state of the group at the current time t; Step S53: Using the state transition matrix A trained in step S52, predict the possible state of the group after K time steps. The calculation method is as follows: in, Let be the state probability distribution vector at time t. This represents the transition matrix after K time steps.

10. The method for collective intelligent representation and hidden horse prediction of equipment fields according to claim 1, characterized in that, Step S6 includes the following steps: Step S61: Define the set of intervention states Determine whether the future state predicted in step S5 belongs to the set of intervention states; Step S62: If it is determined that the future state belongs to the set of intervention states, a preset intervention signal is triggered and the intervention instruction is executed. If it is determined that the future state does not belong to the set of intervention states, the teamwork level of the group is determined to be normal.

Images