Intelligent office automation service system based on personalized reinforcement learning and task allocation method

Abstract

The invention relates to an intelligent office automation service system based on personalized reinforcement learning and a corresponding task allocation method. The system comprises a background server, an employee computer provided with intelligent office automation service application software, and a wearable electroencephalogram device connected with the employee computer through a collector. According to the method, task data and electroencephalogram data executed on a computer of an employee are collected, and a first data representation model, a second data representation model and a correlation model are constructed by a background server and used for scheduling and distributing tasks. And more personalized, humanized and more scientific task allocation is realized, so that the production efficiency is improved.

Description

Technical Field

[0001] The present invention relates to an intelligent office automation service system and method, and particularly to an intelligent office automation service based on personality reinforcement learning, specifically embodied in a method for task allocation. Background Art

[0002] Existing intelligent office automation services allocate tasks based on the interaction characteristics of the tasks themselves, ignoring whether employees are competent and the best allocation methods.

[0003] However, mechanically allocating based on employees' historical data still takes insufficient factors into account. In the practical process, we found that employees' work capabilities are not constant. With age, stage mood, and occasional emotional problems, they will show stage characteristics. And employees with fewer stage characteristics are actually the ones that should be highly valued. Due to their stable work status, their abilities can be high or low, but some of them have excellent subjective initiative, which is often not discovered and valued by enterprises and institutions in a timely manner. And the accumulated losses caused by the lag of the evaluation model also cause a certain degree of waste of human resources. Therefore, how to monitor employees' work status while providing intelligent office services is an urgent problem to be solved and a new perspective for employee assessment. Summary of the Invention

[0004] To solve the above problems, on the one hand, the present invention provides an intelligent office automation service system based on personality reinforcement learning, including a background server, an employee computer installed with intelligent office automation service application software, and a wearable electroencephalogram device connected to the employee computer. Among them, the application software has a data collection module for recording three types of task data: employee task attributes, task completion progress data, and task completion rating data, and regularly packaging them into a first data set and uploading it to the background server; and the wearable electroencephalogram device monitors electroencephalogram data in real time when the employee is working in front of the computer, and packages it into a second data set and uploads it to the background server during the same period (i.e., the same period as the regular one). The background server receives the package, unpacks it, retrieves the data, and uses it to construct an intelligent model based on personality reinforcement learning. The intelligent model based on personality reinforcement learning includes a first data performance model corresponding to the first data set, which is used to represent the quality of work completion, a second data performance model corresponding to the second data set, which is used to represent the employee's stage work status, and an association model between the two, which is used to reinforce the learning of the correlation between the employee's stage work status and the quality of work completion, for predicting the real-time work completion quality of the employee, and scheduling and allocating tasks according to the first data performance model, the second data performance model, and the association model.

[0005] Optionally, the wearable EEG device has a wearable body and electrodes distributed in different preset areas on the wearable body, the electrodes are connected to a computer host via a data collector, and are regularly packaged via the application software.

[0006] Preferably, the method for presetting different areas includes dividing the middle brain area in the top view of the head into a tic-tac-toe grid, and arranging 2-4 electrodes in each grid area.

[0007] Through our previous EEG distribution analysis, various emotional representations show a nine-square grid distribution state, and there are very few strong and weak signals distributed in a certain area near the grid boundary or intersection. Therefore, the use of the above preset areas is based on the above research and analysis to identify emotional types with a factual basis, rather than the traditional four-division brain area analysis.

[0008] Optionally, the first data representation model construction method is as follows: S1 obtains the employee's historical task data, and calls out the task attributes, task completion progress data, and task completion rating data corresponding to the historical task data; In the actual design process of the system, the task attributes can be selected according to different tasks, such as rated, additional, or usual and challenging, urgent and regular, first-level, second-level and other attribute labels to mark the levels. The progress is divided according to the specific circumstances of the task. For example, the counting is represented by the calculation of numbers, and the staged one is represented by the stage of completion. The completion rating can be represented by the relationship between the progress and the time requirement. For example, half of the time has passed, but the number of counts and the stage are different. This is only for ease of understanding without detailed plan description.

[0009] S2 divides the historical task data into multiple stages according to the period, and in each stage, trains the multivariate support vector machine using the sub-training set divided from the multiple sub-historical task data, and verifies it using the divided sub-verification set, which is used to classify the data in each stage in the three-dimensional space of work task attributes, task completion progress data, and task completion rating data, and further clusters the data under each classification, and clusters according to a preset grade standard to calculate the Euclidean distance between the two types of data; S3 determines the most probable level in each stage according to the Euclidean distance in each stage.

[0010] Optionally, the preset grade standard clustering preset method includes selecting employees with different typical work performances through evaluation, and also completing the data clustering of steps S1-S2 to obtain them.

[0011] Optionally, the period is 3 to 6 months. It is easy to understand that the sub-historical task data is daily or weekly data, and for some specific tasks, it can be hourly data.

[0012] Optionally, the most probable level discrimination method is as follows: according to the data clustering, at least one clustering center is obtained, the Euclidean distances between each clustering center and the clustering centers corresponding to different level criteria are calculated to form a plurality of Euclidean distances, a distance threshold is set, the probabilities of occurrence of each type of Euclidean distance within an integer multiple range of the distance threshold are calculated, and the level to which the clustering corresponding to the level criterion with the largest probability belongs is the most probable level.

[0013] It is easy to understand that by setting the distance threshold and calculating the most probable level, the accidental characteristics of selecting the shortest Euclidean distance as the discrimination result in the traditional sense are excluded, and the best level discrimination result is obtained from a statistical sense.

[0014] Preferably, the distance threshold is selected as the variance of the plurality of Euclidean distances in all stages.

[0015] It should be emphasized that by specifically selecting the variance as the distance threshold, a stability index of the employee's personalization among different levels is given, that is, it characterizes whether an employee stably tends to a certain level in the discrimination of different level affiliations throughout the whole stage. If the variance is smaller, no matter which level the final discrimination result belongs to, it indicates that the affiliation to a certain level is more stable. Therefore, using the variance as the distance threshold can exclude unstable factors, classify the stable type of Euclidean distances, and then calculate the probability of occurrence. If the Euclidean distances of all types are more than twice the distance threshold, the level to which the clustering corresponding to the level criterion with the largest probability within the range of the distance threshold with the smallest multiple belongs is used as the most probable level.

[0016] Optionally, the second data representation model construction method is as follows: Q1 Set the top view of the brain, and according to the spatial distribution of the electrodes, fill in the collected data in pseudo-color to obtain multiple time-dependent brain pseudo-color change diagrams, and divide the pseudo-color diagrams into a training set and a validation set; Q2 Construct a convolutional neural network model, collect the self-descriptions of the employee's working status during the time period corresponding to each sub-historical task data to form status labels, use the training set to train the convolutional neural network model, verify the accuracy of recognizing the status labels with the validation set, and continuously optimize the network parameters until the accuracy stabilizes to complete the training.

[0017] Optionally, the association model construction method is: T1 Construct an adversarial generation network, use the pseudo-color diagram training set in Q1 as the training diagram, and use the three projections of the clustering in S2 in the coordinate plane of the data three-dimensional space as the validation diagram; T2 divides the adversarial generative network into three sub-networks, and uses a training graph and a projected validation graph to train the sub-networks respectively. The training of the sub-networks includes first training the discriminator of the sub-network, and then training the generator of the sub-network. After the training is completed, the discriminator is discarded, and the trained generator is left; T3 re-obtains the sub-historical data of the employee in multiple stages and the pseudo-color map at the corresponding time of the sub-historical data as the associated pseudo-color map, and inputs the associated pseudo-color map into the trained generator in T2 to form three corresponding predicted projection maps. The sub-historical data of the employee re-obtained in multiple stages is also projected in the coordinate plane of the data three-dimensional space, and the real clustering projection is obtained; T4 constructs a three-dimensional loss function, which is composed of sub-functions representing three coordinate planes, calculates the Euclidean distance between the predicted projection map and the real clustering projection in the corresponding coordinate plane as the predicted Euclidean distance, and the three projection distances of the line segment corresponding to the predicted Euclidean distance in the three coordinate planes are used as the function values of the sub-functions. According to the function values, backpropagation is performed to further retrain the generator until the three sub-functions are all stable, and an associated model is obtained.

[0018] It should be understood that through the retraining of reinforcement learning, the generator in the second data representation model is optimized again by using the clustering Euclidean distance in the data three-dimensional space as a metric, realizing the correlation between the EEG data and the task data.

[0019] Optionally, the method of scheduling and allocation includes: P1 obtains the sub-task data of the employee to be tested in the current stage and the EEG data corresponding to the current stage; P2 inputs the sub-task data and the EEG data into the first data representation model and the second data representation model respectively to obtain the current most probable level and the prediction of the current working state; P3 inputs the EEG data into the associated model to obtain the predicted projections on the three coordinate planes (i.e., the prediction of the work completion quality), obtains the projections of the grade standard clustering corresponding to the current most probable level on the three coordinate planes, and calculates the first Euclidean distance between the predicted projections on the three coordinate planes and the projections of the grade standard clustering on the three coordinate planes; P4 obtains the multiple sub-task data of the employee to be tested for multiple stages before the current stage, and repeats steps P1 - P3 to obtain the corresponding multiple Euclidean distances. The difference is that each of the multiple Euclidean distances is calculated based on the multiple second Euclidean distances between the predicted projections on the three coordinate planes in the corresponding stage and the projections of the grade standard clustering corresponding to the current most probable level on the three coordinate planes; If the first Euclidean distance is not greater than any second Euclidean distance, that is, the first case, then assign the task attribute point corresponding to the clustering center of the grade standard clustering corresponding to the current most probable grade, or the previous task attribute point in the task attribute direction, or the same task as the next task attribute point in the direction opposite to the task attribute direction. If the first Euclidean distance is not less than any second Euclidean distance, that is, the second case, then consider the task attribute point corresponding to the clustering center of the grade standard clustering corresponding to the most probable grade in the previous stage, or the previous task attribute point in the task attribute direction, or the same task as the next task attribute point in the direction opposite to the task attribute direction. In other cases, that is, the third case, then consider the task attribute point corresponding to the clustering center of the grade standard clustering corresponding to the most probable grade in the stage corresponding to the second Euclidean distance closest to the first Euclidean distance, or the previous task attribute point in the task attribute direction, or the same task as the next task attribute point in the direction opposite to the task attribute direction.

[0020] Preferably, the current working state includes negative, normal, positive, and extraordinary; if the predicted result of the current working state is negative, then assign the same task as the next task attribute point in the direction opposite to the task attribute direction; if the predicted result of the current working state is normal, then assign the task of the task attribute point corresponding to the clustering center or the same task as the next task attribute point in the direction opposite to the task attribute direction; if the predicted result of the current working state is positive or extraordinary, then assign the same task as the previous task attribute point in the task attribute direction; if the task attribute point corresponding to the clustering center of the grade standard clustering corresponding to the most probable grade is the highest attribute point, then assign the task attribute point corresponding to the clustering center regardless of the working state.

[0021] It is easy to understand that through such an assignment method, the beneficial effects of the present invention can be reflected in the following aspects: on the one hand, from the perspective of time-varying, considering the trend presented by the current most probable grade relative to the previous stage, and the stage based on which the task assignment is selected is different according to different trends. For the first case, it indicates that the current most probable grade is the one that best conforms to the actual work completion quality, so the task is assigned locally based on the most probable grade of this stage, so that the employee is given the most suitable task assignment in the state that best conforms to the current completion quality. For the second case, it indicates that the current most probable grade is the one that least conforms to the actual work completion quality, so consider assigning tasks based on the most probable grade of the previous stage, so that the pressure on the employee from the task assignment has continuity before and after, and the employee's mentality has a continuous adaptation process. The third case indicates that the degree of conformity is between the previous two cases, so select the previous stage closest to this stage and assign tasks based on the most probable grade of that previous stage, so that the employee has a feeling of deja vu and finds the previous feeling, and the change in the task attributes assigned is not too large.

[0022] On the other hand, directly linking EEG emotions with task assignment attributes, obtaining the predicted projection corresponding to the EEG data from the correlation model, thereby calculating the first and second Euclidean distances to characterize the time-dependent changes of the most probable levels, and correlating the EEG data with the most probable levels, realizing intelligent task assignment based on reinforcement learning. It is a task assignment method that comprehensively considers the task completion quality and emotions to construct a correlation model, thereby affecting the task assignment method of the assigned tasks, achieving a more personalized, user-friendly, and scientific task assignment, and thus improving production efficiency.

[0023] On the other hand, the present invention provides a task assignment method based on the above system, including the following steps: The first step is to build an intelligent office automation service system based on personality reinforcement learning; The second step is to use the built intelligent office automation service system based on personality reinforcement learning to construct a first data performance model , a second data performance model, and a correlation model; The third step is to schedule and assign tasks using the first data performance model, the second data performance model, and the correlation model. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 Analysis diagram of the prior art of the electrode spatial distribution of the wearable EEG device, Figure 2 For constructing the first data performance model, the clustering results of multivariate SVM classification, the hierarchical standard clustering distribution, and schematic diagrams of multiple Euclidean distances in the first multi-stage data three-dimensional space, Figure 3 Schematic diagram of the principle for judging the most probable level of the first data performance model in Embodiment 1 of the present invention, Figure 4 Flowchart of the method for constructing the second data performance model, Figure 5 Flowchart of the method for constructing the correlation model, Figure 6 Schematic diagram of the distribution of task attribute points for task assignment, Figure 7 Schematic diagram of the configuration of the intelligent office automation service system based on personality reinforcement learning in Embodiment 1 built in Embodiment 2 of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS Embodiment 1

[0025] This embodiment provides an intelligent office automation service system based on personalized reinforcement learning, including a background server, employee computers installed with intelligent office automation service application software, and wearable electroencephalogram devices connected to the computers. Among them, the employee computer is both a work computer and a unit for electroencephalogram data collection, upload, and task execution. The background server is used to regularly receive the task data of the first data set and the real-time electroencephalogram data of the second data set. The electroencephalogram data comes from the electrodes of the wearable electroencephalogram device with a special well-shaped spatial distribution. As Figure 1 shown, in the existing EGG technology, for three emotions, in different electroencephalogram frequency ranges (4 - 7Hz, 8 - 13Hz, 14 - 29Hz, 30 - 47Hz), the main frequency distributions, including extremely high intensity and extremely low intensity, are mostly distributed in the well-shaped nine-square grid (represented by ○), while there are not many distributions at the boundaries or intersections of the dividing lines of the well-shaped lines (represented by ×). And the distributions of different emotion classifications in different bands show certain specific spatial distribution characteristics. Therefore, arranging the electrodes in these areas helps feature learning and further identifies the working state. In this embodiment, 5 electrodes are set in each grid.

[0026] The intelligent system based on personalized reinforcement learning described in the present invention includes a first data performance model corresponding to the first data set, which is used to characterize the quality of work completion, a second data performance model corresponding to the second data set, which is used to characterize the employee's phased work state, and an association model between the two, which is used to reinforce the learning of the association between the employee's phased work state and the quality of work completion, for predicting the employee's real-time work completion quality, and scheduling and allocating tasks according to the first data performance model, the second data performance model, and the association model.

[0027] Among them, the construction method of the first data performance model is as follows: S1 Obtain the employee's historical task data, and call out the task attributes, task completion progress data, and task completion rating data corresponding to the historical task data; S2 As Figure 2 shown, according to the period, divide the historical task data into N (N > 3) stages, and construct a three-dimensional data space of task attributes, task completion progress data, and task completion rating data in the figure.

[0028] As Figure 3 shown, in each stage, taking the first stage as an example, use the sub-training sets divided by multiple sub-historical task data to train a multi-class support vector machine (multi-class SVM), and use the divided sub-validation sets for verification. The figure shows the arranged data points to represent the data sets in the sub-training sets.

[0029] For each stage of data classification (such as Classification I, Classification II, etc.) in the three-dimensional data space of work task attributes, task completion progress data, and task completion rating data, and further clustering is performed under each classification (including Classification I, Classification II, etc. in the figure), forming corresponding clusters such as Cluster I, Cluster II, etc. for each classification, and clustering according to a preset grade standard, and calculating the Euclidean distance between the two types of data.

[0030] Figure 3 The process flow of the preset grade standard clustering method is given in [reference], which is obtained by evaluating and selecting employees with different typical work performances and performing the same data clustering steps S1 - S2.

[0031] Regarding the calculation of the Euclidean distance between the two types of data, return to Figure 2 In [reference]. Since the task attributes, task completion progress data, and task completion rating data are all discrete (specifically, the data points in the three-dimensional data space are embedded through data numericalization), in order to clearly represent the clustering, Figure 2 the data points in [reference] are densely represented.

[0032] Figure 2 The hyperplane of the multi-class SVM classification is given in [reference]. The clusters of the three grades of excellent, good, and poor in the grade standard clustering are represented by pentagrams, circles, and triangles respectively.

[0033] Figure 2 The clustering of two types of sub-historical task data distinguished by the hyperplane is also given in [reference], and the Euclidean distances between the clustering of the three grade standards of excellent, good, and poor in the first stage are given.

[0034] The specific method for obtaining multiple Euclidean distances continues to refer to Figure 3 , and the ellipsis in the figure indicates that the same process in the first-stage dashed box is repeatedly used in other second and more stages, and the corresponding more Euclidean distances in the cases defined by the dashed box (as Figure 2 shown) are obtained.

[0035] S3 determines the most probable grade ( Figure 3 ) within each stage according to the Euclidean distance within each stage. The specific method is to obtain multiple cluster centers according to the data clusters such as Cluster I, Cluster II, etc. in Figure 3 . In the case of two clusters in Figure 2 , calculate the Euclidean distance between each cluster center and the cluster centers of different grade standards, form multiple Euclidean distances, set the variance of the multiple Euclidean distances of all N stages as the distance threshold, calculate the probability of occurrence of each type of Euclidean distance within the distance threshold, and select the grade corresponding to the grade standard clustering with the highest probability as the most probable grade.

[0036] In other words, in Figure 2Taking the multiple Euclidean distances represented by the first stage as an example, first obtain all such Euclidean distances of N stages, and then calculate their variance as the distance threshold. Then calculate each type of Euclidean distance of each stage (this embodiment takes the first stage as an example) within the distance threshold range, that is, Figure 2 The probability of the two clusters corresponding to the hierarchical standard clusters with an Euclidean distance within the distance threshold range is the largest, such as Figure 2 In the first stage, the Euclidean distances between two of the clusters and the poor grade standard clusters are within the distance threshold range, and the distances with other grade standard clusters are not within the distance threshold range. Therefore, the former has a high probability of appearing, and the most probable grade in this stage is poor.

[0037] If the Euclidean distances between two types of clusters and the standard clusters of each level are all above one times the variance and within the range of several times the variance, the probability of the Euclidean distance within the minimum multiple variance will prevail. The same analysis is applied to the clustering of more amounts within a stage. Even if various types of Euclidean distances, such as those within one times the variance and those within several times the variance, are mixed, if the probability of the Euclidean distance within several times the variance is high, the one with the higher probability will prevail, rather than the minimum Euclidean distance within one times the variance. This shows that the Euclidean distance within one times the variance is most likely to be an accidental data cluster. Therefore, choosing the most probable algorithm is equivalent to eliminating such data clusters that are most likely to be accidental.

[0038] Then, if Figure 4 As shown, the second data representation model construction method will be described as follows: Q1 sets a top view of the brain, and fills in the collected data with pseudo-color according to the spatial distribution of the electrodes, obtaining multiple pseudo-color change maps of the brain over time, and dividing the pseudo-color maps into training sets and validation sets; Q2 builds a convolutional neural network model (CNN), collects employees' self-descriptions of their work status in the time period corresponding to each of the aforementioned sub-historical task data (the first stage, such as the first quarter), forms status labels, uses the training set to train the convolutional neural network model, uses the validation set to verify the accuracy of identifying status labels, and continuously optimizes network parameters until the accuracy stabilizes and the training is completed.

[0039] like Figure 5 As shown, the association model construction method is: T1 constructs a generative adversarial network, uses the pseudo-color image training set in Q1 as the training image, and uses the three projections of the clusters in S2 in the coordinate plane in the three-dimensional space of the data as the verification image; Figure 3Taking the second stage as an example, the projection 1 of data point A on the progress-rating coordinate plane, the projection 2 on the rating-task attribute coordinate plane, and the projection 3 on the task attribute-progress coordinate plane are given. The other data points are projected in the same way. Projections of a data cluster on three coordinate planes are formed, creating three verification graphs. Moreover, because of real-time monitoring and synchronous uploading, the time between the verification graph and the training graph is the same period, that is, they belong to the same stage.

[0040] T2 divides the adversarial generation network into three sub-networks, which are respectively used for three coordinate planes. Figure 5 An example of one of the coordinate planes is given, and the other two sub-networks are constructed in the same way.

[0041] Specifically, the training graph and a projected verification graph are respectively used to train the sub-network. The training of the sub-network includes first training the discriminator of the sub-network, and then training the generator of the sub-network. After training is completed, the discriminator is discarded, leaving the trained generator. Specifically, the training set is input into the generator again to form a generated graph, and the generated graph and the verification graph are input into the discriminator one by one to determine whether it belongs to the verification graph of the same period as the corresponding generated graph.

[0042] T3 re-obtains the sub-historical data of the employee in multiple stages and the pseudo-color graph at the corresponding moment of the sub-historical data as the associated pseudo-color graph. The associated pseudo-color graph is input into the trained generator in T2 to form three corresponding predicted projection graphs. The sub-historical data of the employee re-obtained in multiple stages is also projected in the coordinate planes of the data three-dimensional space, the real clustering projection; T4 constructs a three-dimensional loss function, which is composed of sub-functions representing three coordinate planes. Calculate the Euclidean distance between the predicted projection graph and the real clustering projection in the corresponding coordinate plane as the predicted Euclidean distance. The three projection distances of the line segment corresponding to the predicted Euclidean distance in the three coordinate planes are used as the function values of the sub-functions. According to the function values, backpropagation is performed to further retrain the generator until the three sub-functions are all stable, obtaining an associated model. Therefore, the three sub-functions correspond one-to-one with the sub-networks on the three coordinate planes and are used for backpropagation to optimize the parameters.

[0043] Finally, the method of scheduling and allocation is described as follows: P1 obtains the sub-task data of the current stage of the employee to be tested (still taking the Figure 2 first stage in as an example), and the electroencephalogram data corresponding to the current stage; P3 EEG data is input into the association model to obtain the predicted projection on the three coordinate planes, the projection of the level standard cluster corresponding to the current most probable level on the three coordinate planes is obtained, and the first Euclidean distance between the predicted projection on the three coordinate planes and the projection of the level standard cluster on the three coordinate planes is calculated; P4 obtains multiple subtask data of the employee to be tested for multiple stages before this stage, and repeats steps P1-P3 to obtain multiple corresponding Euclidean distances. The difference is that the calculation of each of the multiple Euclidean distances is based on multiple second Euclidean distances between the predicted projection on the three coordinate planes in the corresponding stage and the projection on the three coordinate planes of the level standard cluster corresponding to the current most probable level; in this embodiment, the corresponding stage should be the stage before the first stage, such as the fourth quarter of last year and the quarter before it. For the aforementioned first stage as an example, if it is changed to other stages, such as the third stage, then the corresponding stage is the stage before the third stage, that is, the second stage, and the first stage before it, etc. At this time, refer to Figure 2 It can be understood that the calculation of each of the multiple Euclidean distances is the calculation of "the predicted projection on the three coordinate planes in the second, first, etc. stages" and "the third stage of this stage ( Figure 2 The Euclidean distance between the projections of the level standard clusters corresponding to the most probable levels (not shown) on the three coordinate planes is called the second Euclidean distance, rather than calculating the Euclidean distance between "the predicted projections on the three coordinate planes in the second and first stages, etc." and "the projections on the three coordinate planes of the level standard clusters corresponding to the most probable levels of the respective stages".

[0044] If the first Euclidean distance is not greater than any second Euclidean distance, that is, the first case, indicating that the current most probable level is the one that best matches the actual work completion quality, then the task attribute point corresponding to the cluster center of the level standard cluster corresponding to the current most probable level, or the previous task attribute point in the direction of the task attribute, or the next task attribute point in the direction opposite to the task attribute direction is assigned; Figure 6 Given that the task attribute point corresponding to the cluster center of the task attribute cluster coordinate point is attribute β, then the previous task attribute point is attribute α, and the next task attribute point is attribute γ.

[0045] If the first Euclidean distance is not less than any of the second Euclidean distances, that is, the second case, indicating that the current most probable level is the least consistent with the actual work completion quality, then the task attribute point corresponding to the cluster center of the level standard cluster corresponding to the most probable level of the previous stage, or the previous task attribute point in the direction of the task attribute, or the task with the same next task attribute point in the direction opposite to the task attribute direction is considered; In other cases, i.e., the third case, indicating that the degree of compliance is between the previous two cases, consider the task attribute point corresponding to the clustering center of the most probable level corresponding to the second Euclidean distance closest to the first Euclidean distance in the corresponding stage of the second Euclidean distance, or the previous task attribute point in the task attribute direction, or the same task as the next task attribute point in the direction opposite to the task attribute direction.

[0046] The current working state includes negative, normal, positive, and extraordinary; if the prediction result of the current working state is negative, assign the same task as the next task attribute point in the direction opposite to the task attribute direction (such as Figure 6 taking attribute β as an example); if the prediction result of the current working state is normal, assign the task of the task attribute point corresponding to the clustering center or the same task as the next task attribute point in the direction opposite to the task attribute direction (such as Figure 6 taking attribute γ as an example), if the prediction result of the current working state is positive or extraordinary, assign the same task as the previous task attribute point in the task attribute direction (such as Figure 6 taking attribute α as an example); if the task attribute point corresponding to the clustering center of the most probable level corresponding to the grade standard clustering is the highest attribute point (i.e., attribute α), then assign the task attribute point α corresponding to the clustering center regardless of the working state. Embodiment 2

[0047] This embodiment provides a task allocation method based on the system given in Embodiment 1, including the following steps: The first step is to build an intelligent office automation service system based on personality reinforcement learning; such as Figure 7 including a background server, an employee computer installed with intelligent office automation service application software, and a wearable electroencephalogram device connected to the employee computer through a data collector.

[0048] The second step is to use the built intelligent office automation service system based on personality reinforcement learning to construct a first data performance model, a second data performance model, and an association model; The third step is to schedule and allocate tasks using the first data performance model, the second data performance model, and the association model.

Claims

1. An intelligent office automation service system based on personality reinforcement learning, characterized in that, It includes a background server, an employee computer installed with intelligent office automation service application software, and a wearable electroencephalogram device connected to the employee computer. Among them, the application software has a data collection module, which is used to record three types of task data: employee task attributes, task completion progress data, and task completion rating data, and regularly packages them into a first data set and uploads it to the background server; and the wearable electroencephalogram device monitors electroencephalogram data in real time when the employee is working in front of the computer, and simultaneously packages it into a second data set using the employee computer and uploads it to the background server. The background server receives the package, unpacks it, retrieves the data, and is used to build an intelligent model based on personality reinforcement learning. The intelligent model based on personality reinforcement learning includes a first data performance model corresponding to the first data set, which is used to represent the quality of work completion, a second data performance model corresponding to the second data set, which is used to represent the employee's phased work status, and an association model between the two, which is used to reinforce the learning of the association between the employee's phased work status and the quality of work completion, and is used for real-time prediction of the employee's work completion quality, and schedules and assigns tasks according to the first data performance model, the second data performance model, and the association model.

2. The system according to claim 1, wherein The wearable electroencephalogram device has a wearable body and electrodes distributed in preset different areas on the wearable body. The electrodes are connected to the computer host through a data collector and are regularly packaged through the application software. The preset method for presetting different areas includes dividing the middle brain area in the top view of the head into a well-shaped nine-square grid, and setting 2 to 4 electrodes in each grid area.

3. The system according to claim 2, wherein The construction method of the first data performance model is as follows: S1 Obtain the employee's historical task data, and call out the task attributes, task completion progress data, and task completion rating data corresponding to the historical task data. S2 According to the period, divide the historical task data into multiple stages. In each stage, use the sub-training sets divided from multiple sub-historical task data to train a multi-class support vector machine, and use the divided sub-validation sets for verification, for data classification in each stage within the three-dimensional space of work task attributes, task completion progress data, and task completion rating data, and further data clustering is performed under each classification, and clustering is performed according to the preset grade standard, and the Euclidean distance between the two types of data is calculated. S3 Determine the most probable grade within each stage according to the Euclidean distance within each stage.

4. The system according to claim 3, wherein The preset method for presetting the grade standard clustering includes selecting employees with different typical work performances through evaluation and obtaining them by performing the same steps S1-S2 data clustering; the period is 3 to 6 months.

5. The system according to claim 3, wherein The method for determining the most probable grade is to obtain at least one clustering center according to the data clustering, calculate the Euclidean distance between each clustering center and the clustering centers of different grade standards, form multiple Euclidean distances, set a distance threshold, calculate the probability of occurrence of each type of Euclidean distance within an integer multiple range of the distance threshold, and select the grade belonging to the grade standard clustering with the highest probability as the most probable grade. The selection method of the distance threshold is the variance of the multiple Euclidean distances in all stages.

6. The system according to claim 1, wherein The method for constructing the second data representation model is as follows: Q1 Set the top view of the brain. According to the spatial distribution of the electrodes, fill in the collected data with pseudocolors to obtain multiple time-dependent brain pseudocolor change maps. Divide the pseudocolor maps into a training set and a validation set; Q2 Construct a convolutional neural network model. Collect the self-descriptions of the working states of employees during the time periods corresponding to each sub-historical task data to form state labels. Use the training set to train the convolutional neural network model, verify the accuracy of identifying state labels with the validation set, and continuously optimize the network parameters until the accuracy stabilizes to complete the training.

7. The system according to claim 1, wherein The method for constructing the correlation model is as follows: T1 Construct an adversarial generative network. Use the pseudocolor map training set in Q1 as the training map, and use the three projections of the clustering in the coordinate plane of the data three-dimensional space in S2 as the validation map; T2 Divide the adversarial generative network into three sub-networks, and use the training map and the validation map of one projection to train the sub-networks respectively. The training of the sub-networks includes first training the discriminator of the sub-network and then training the generator of the sub-network. After the training is completed, discard the discriminator and leave the trained generator; T3 Re-obtain the sub-historical data of employees in multiple stages and the corresponding pseudocolor maps at the corresponding times of the sub-historical data as associated pseudocolor maps. Input the associated pseudocolor maps into the trained generator in T2 to form three corresponding predicted projection maps. Re-obtain the sub-historical data of employees in multiple stages and also perform real clustering projections on the coordinate plane of the data three-dimensional space; T4 Construct a three-dimensional loss function. The three-dimensional loss function is composed of sub-functions representing three coordinate planes. Calculate the Euclidean distance between the predicted projection map and the real clustering projection in the corresponding coordinate plane as the predicted Euclidean distance. The three projection distances of the line segment corresponding to the predicted Euclidean distance in the three coordinate planes are used as the function values of the sub-functions. According to these function values, perform backpropagation to further retrain the generator until the three sub-functions all stabilize to obtain the correlation model.

8. The system according to claim 6 or 7, characterized in that, The method of scheduling and allocation includes: P1 Obtain the sub-task data of the employee to be measured in the current stage and the corresponding electroencephalogram data in the current stage; P2 Input the sub-task data and the electroencephalogram data into the first data representation model and the second data representation model respectively to obtain the current most probable level and the prediction of the current working state; P3 Input the electroencephalogram data into the correlation model to obtain the predicted projections on the three coordinate planes. Obtain the projections of the level standard clustering corresponding to the current most probable level on the three coordinate planes, and calculate the first Euclidean distance between the predicted projections on the three coordinate planes and the projections of the level standard clustering on the three coordinate planes; P4 Obtain the multiple sub-task data of the employee to be measured for multiple stages before the current stage, and repeat steps P1 - P3 to obtain the corresponding multiple Euclidean distances. The difference is that each of the multiple Euclidean distances is calculated based on the multiple second Euclidean distances between the predicted projections on the three coordinate planes in the corresponding stage and the projections of the level standard clustering corresponding to the current most probable level on the three coordinate planes; If the first Euclidean distance is not greater than any of the second Euclidean distances, that is, the first case, then allocate the task attribute point corresponding to the cluster center of the cluster corresponding to the current most probable level, or the previous task attribute point in the task attribute direction, or the same task as the next task attribute point in the direction opposite to the task attribute direction. If the first Euclidean distance is not less than any of the second Euclidean distances, that is, the second case, then consider the task attribute point corresponding to the cluster center of the cluster corresponding to the most probable level in the previous stage, or the previous task attribute point in the task attribute direction, or the same task as the next task attribute point in the direction opposite to the task attribute direction. In other cases, that is, the third case, then consider the task attribute point corresponding to the cluster center of the cluster corresponding to the most probable level in the stage corresponding to the second Euclidean distance closest to the first Euclidean distance, or the previous task attribute point in the task attribute direction, or the same task as the next task attribute point in the direction opposite to the task attribute direction.

9. The system according to claim 8, wherein The current working state includes negative, normal, positive, and extraordinary. If the prediction result of the current working state is negative, then allocate the same task as the next task attribute point in the direction opposite to the task attribute direction. If the prediction result of the current working state is normal, then allocate the task of the task attribute point corresponding to the cluster center or the same task as the next task attribute point in the direction opposite to the task attribute direction. If the prediction result of the current working state is positive or extraordinary, then allocate the same task as the previous task attribute point in the task attribute direction. If the task attribute point corresponding to the cluster center of the cluster corresponding to the most probable level is the highest attribute point, then allocate the task attribute point corresponding to the cluster center regardless of the working state.

10. A task allocation method based on the system according to any one of claims 1-9, characterized in that, It includes the following steps: In the first step, build the system. In the second step, use the built system to construct the first data representation model, the second data representation model, and the association model. In the third step, use the first data representation model, the second data representation model, and the association model to schedule and allocate tasks.

Images