Multi-dimensional data fusion-based panoramic real-time monitoring and decision-making method for credential terminal

Abstract

The invention discloses a multi-dimensional data fusion-based panoramic real-time monitoring and decision-making method for a credential terminal, and relates to the technical field of credential terminal monitoring, and the method comprises the steps: collecting and carrying out the smoothing of a multi-dimensional index of the terminal; predicting the smoothed data at the next moment by using an LSTM model, and calculating a prediction error; after the prediction error is fuzzified through a membership function, the posterior probability of'immediate scheduling 'is calculated in combination with a Bayesian network, and a decision is triggered; selecting an optimal scheduling action in a current state based on Q-learning; and archiving data and decision records in each step. The panoramic monitoring and decision-making method integrating time sequence anomaly detection, fuzzy Bayesian reasoning and reinforcement learning scheduling is constructed, dynamic perception, intelligent judgment and resource self-adaptive control of the operation state of the credential terminal are achieved, and the method has the advantages that anomaly recognition is accurate, decision logic can be explained, strategy execution can be rolled back and the like; and the management intelligence level and the operation efficiency in a large-scale heterogeneous terminal environment are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of signal creation terminal detection, and in particular to a signal creation terminal panoramic real-time monitoring and decision-making method based on multi-dimensional data fusion. BACKGROUND

[0002] With the comprehensive promotion of signal creation (information technology application innovation) terminals in government and enterprise units at all levels, the number of terminal devices is rapidly increasing, the system heterogeneity is enhanced, and the running load is increasingly complex. In order to ensure that the terminal still has stability, high efficiency and manageability in a large-scale deployment environment, it is a key requirement to build a real-time, accurate and self-decision-making terminal panoramic monitoring and control mechanism.

[0003] The existing terminal monitoring system is mostly based on static index presentation and fixed threshold alarm, lacks the ability of trend modeling and dynamic abnormality recognition of terminal running behavior, and is difficult to timely perceive small abnormal signals in complex environment. At the same time, in the face of different chip architectures, operating systems and business load characteristics, the resource usage behaviors of various terminals are highly different, and the conventional unified scheduling strategy is difficult to realize targeted optimization, which easily leads to over-scheduling or response lag, affecting the overall running efficiency and system stability. In addition, although some current systems introduce a strategy control module, the decision-making process of the module mostly relies on artificial preset rules, lacks deep understanding of the actual running state of the terminal and self-adaptive ability. Once the strategy effect is not good, the system also lacks an effective rollback mechanism, which increases the risk of mis-scheduling and reduces user trust. Therefore, a dynamic monitoring and decision-making method that integrates multi-dimensional monitoring data, has real-time trend perception ability and can be self-adjusted multiple times is needed. SUMMARY

[0004] In view of the problems in the monitoring of signal creation terminals, such as single monitoring method, lack of dynamic abnormality recognition ability, difficulty of unified scheduling strategy to accurately adapt to diversified terminals, and dependence of strategy control on artificial rules, lack of self-adaptive and rollback mechanism, the present application is proposed.

[0005] Therefore, the problem to be solved by the present application is how to provide a signal creation terminal panoramic real-time monitoring and decision-making method based on multi-dimensional data fusion.

[0006] To solve the above technical problems, the present application provides the following technical solutions: In a first aspect, the embodiments of the present application provide a real-time monitoring and decision-making method for a ChinaSoft terminal panoramic view based on multi-dimensional data fusion, comprising: collecting in real time and performing sliding average smoothing processing on multi-dimensional indexes of terminal CPU, memory, network bandwidth and login times; using an LSTM model to predict the next time after smoothing and calculate the prediction error for determining whether an anomaly occurs; after the prediction error is fuzzed by a membership function, combining a Bayesian network to calculate the posterior probability of "immediate scheduling" and triggering decision-making; selecting the optimal scheduling action under the current state based on Q-learning; archiving the data and decision-making records of each step, and periodically updating the conditional probability table of LSTM and Bayesian network online.

[0007] As a preferred scheme of the real-time monitoring and decision-making method for a ChinaSoft terminal panoramic view based on multi-dimensional data fusion, in order to obtain smooth key operation indexes, the system uses a fixed-length sliding window to uniformly collect the multi-dimensional original data of terminal CPU usage, memory usage, network bandwidth occupancy and login times in time sequence, and realizes smoothing processing by performing arithmetic average on all original values in the window in each collection period; the length of the sliding window is dynamically selected in the interval of five to ten minutes according to the actual deployment scene, which is used to balance the smoothing effect and response timeliness, and the smoothed values of each index are updated synchronously according to a unified timestamp.

[0008] As a preferred scheme of the real-time monitoring and decision-making method for a ChinaSoft terminal panoramic view based on multi-dimensional data fusion, the system takes the multi-dimensional smoothed value sequence as input, and uses a trained long short-term memory network model to perform time series prediction on the terminal operation indexes at the next time, the long short-term memory network model contains multiple layers of recurrent neural units and Dropout mechanism to prevent overfitting; the system compares the prediction result with the original smoothed value collected at the same time, and takes the absolute value of the difference between the two as the prediction error, so as to evaluate whether the current operation state of the terminal deviates from the normal trend.

[0009] As a preferred scheme of the real-time monitoring and decision-making method for a ChinaSoft terminal panoramic view based on multi-dimensional data fusion, for a terminal that has not triggered abnormality determination for multiple consecutive detections, the system automatically accumulates the internal counter to a preset value, and then dynamically tightens the abnormal threshold according to the recent operation stability of the terminal, which is used to improve the sensitivity to potential small deviations, and continues to use the new threshold to perform subsequent error determination after the threshold is tightened.

[0010] As a preferred scheme of the signal creation terminal panoramic real-time monitoring and decision method based on multi-dimensional data fusion, wherein: the system maps the multi-dimensional prediction error to a continuous abnormality degree value according to a pre-designed membership function, and inputs the mapped membership value into the Bayesian network as evidence, the Bayesian network performs multi-index causal reasoning according to the set prior probability and conditional probability table, calculates the posterior probability of the terminal "immediate dispatch" event, and if the posterior probability exceeds the set threshold, the resource scheduling decision link is triggered.

[0011] As a preferred scheme of the signal creation terminal panoramic real-time monitoring and decision method based on multi-dimensional data fusion, wherein: the system constructs a state vector containing the smoothed CPU usage rate, memory usage rate, network bandwidth occupation and scheduling urgency probability according to the current multi-dimensional index smoothing value and scheduling urgency probability, and selects the optimal scheduling action in the predefined action set based on the Q-learning algorithm, the action set includes limiting broadband, migrating tasks, cleaning cache and no operation, and the real-time reward is calculated before and after each action execution to guide the Q table update.

[0012] As a preferred scheme of the signal creation terminal panoramic real-time monitoring and decision method based on multi-dimensional data fusion, wherein: for each selected scheduling action, the system sets a fixed monitoring time after the action is executed, and if the CPU or memory usage rate reduction ratio after the scheduling action is determined to be less than 10% within the time , the system state before the action is executed is automatically rolled back, the failure experience is recorded to the rollback log, and the suboptimal execution action is reselected.

[0013] In a second aspect, to further solve the problems in the monitoring of the Xinyuan terminal, the embodiment provides a Xinyuan terminal panoramic real-time monitoring and decision system based on multi-dimensional data fusion, which comprises: a data acquisition module configured to periodically acquire operation data of the Xinyuan terminal and to perform smoothing processing on the acquired data by using a sliding window mechanism; a time sequence prediction module configured to receive the operation data after the smoothing processing, to predict the operation data at a future time based on a long short-term memory network, to generate a predicted value, and to calculate an error between the predicted value and an actually acquired value; a fuzzy reasoning module configured to map the prediction error to a multi-dimensional membership value, to infer whether the terminal is in a scheduling triggering state based on a preset Bayesian network structure and a conditional probability table, and to output a posterior probability of scheduling demand; a scheduling decision module configured to receive the posterior probability and current state information to construct a state vector, to select a scheduling action based on a reinforcement learning strategy, and to feed back a rollback module; the rollback module is configured to determine the scheduling effect after the execution of the scheduling action, to rollback to a state before scheduling if no improvement of resource state is detected within a set monitoring period, and to record the scheduling process to a policy update log; and a model update module configured to update and optimize the reinforcement learning strategy based on the scheduling log provided by the feedback rollback module, and to introduce a state convergence condition and an action success rate threshold into the scheduling strategy iteration process for policy optimization.

[0014] In a third aspect, the embodiment of the present application provides a computer device comprising a memory and a processor, and the memory stores a computer program, wherein the computer program is executed by the processor to implement any step of the Xinyuan terminal panoramic real-time monitoring and decision method based on multi-dimensional data fusion according to the first aspect of the present application.

[0015] In a fourth aspect, the embodiment of the present application provides a computer readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement any step of the Xinyuan terminal panoramic real-time monitoring and decision method based on multi-dimensional data fusion according to the first aspect of the present application.

[0016] The present application has the following beneficial effects: 1. The present application can dynamically model based on terminal historical operation indexes by introducing a time sequence anomaly detection mechanism, timely discovers potential risks and implicit anomalies that cannot be captured by traditional static threshold methods, and effectively improves the sensitivity and foresight of monitoring; 2. The present application uses fuzzy logic to measure the degree of error, and combines Bayesian network for causal reasoning, can comprehensively evaluate the necessity of scheduling according to multiple dimensional factors, avoids false alarms of a single index, and at the same time retains the explainability of the decision logic, which is convenient for managers to understand and intervene; 3、The reinforcement learning module in the application can continuously optimize the decision path based on historical feedback when executing specific scheduling strategies (such as task migration, bandwidth limitation, etc.), realize dynamic adjustment of resource allocation according to the terminal state, and improve the overall operation efficiency of the system; 4、The application designs an effect monitoring and rollback mechanism for scheduling actions, which can automatically undo and reselect other feasible strategies once the effect after execution does not meet expectations, reduce the negative impact of mis-scheduling, and improve the robustness and stability of the system; 5、The application connects data collection, trend judgment, intelligent decision and scheduling execution into a closed-loop process to build a terminal operation guarantee system with self-learning and self-optimization capabilities, significantly improving the intelligentization and automation level of the signal creation terminal management. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the application, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor. Among them: Figure 1 The implementation flowchart of the application in embodiment 1.

[0018] Figure 2 The reinforcement learning scheduling module flowchart of the application in embodiment 1.

[0019] Figure 3 The system module diagram of the application in embodiment 2.

[0020] Figure 4 The probability table of CPU prediction error membership of the application in embodiment 1.

[0021] Figure 5 The probability table of memory prediction error membership of the application in embodiment 1.

[0022] Figure 6 The probability table of network bandwidth error membership of the application in embodiment 1.

[0023] Figure 7 The probability table of login behavior prediction error membership of the application in embodiment 1. DETAILED DESCRIPTION

[0024] In order to make the above-mentioned purposes, features and advantages of the application more obvious and easy to understand, the specific embodiments of the application will be described in detail below with reference to the drawings of the specification.

[0025] In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present application. However, it will be apparent to one skilled in the art that the present application can be practiced without the specific details set forth in this description. In other instances, well-known methods, procedures, components, and circuits have not been described in detail so as not to unnecessarily obscure aspects of the present application.

[0026] Secondly, the "one embodiment" or "an embodiment" as referred to herein means a specific implementation that can include features, structures or characteristics that are not included in other implementations of the application. The various examples of the "one embodiment" or "an embodiment" appearing in the description are not necessarily all referring to the same embodiment, nor are they mutually exclusive, but are provided for explanatory purposes to help understand different implementations of the application.

[0027] Embodiment 1 Referring to Figure 1 , Figure 2 , Figures 4-7 , the first embodiment of the present application provides a multi-dimensional data fusion-based real-time monitoring and decision-making method for a ChinaSoft terminal, comprising the following steps: Step S1: Real-time collection and sliding average smoothing processing of multi-dimensional indexes such as terminal CPU, memory, network bandwidth and login times, comprising the following sub-steps: S1-1: Defining and real-time obtaining of four key original running indexes of the terminal: CPU usage , memory usage , network bandwidth occupancy and login times , and recording according to a unified timestamp.

[0028] S1-2: Selection of a smoothing window length (typically 5-10 minutes) for removing short-term burst fluctuations and retaining trend information.

[0029] S1-3: Smoothing calculation of each index using a sliding average formula, which can be expressed as:

[0030] In the formula, represents the value of the th original running index at time , which is collected in real time by a terminal-side sensor / log; represents the sliding average value of the th original running index at time .

[0031] It should be noted that the step collects multiple key running indicators including CPU usage, memory usage, network bandwidth occupancy and login times on the terminal side, and uses a moving average method to smooth the original data to reduce the interference caused by short-term fluctuations, and to provide a more stable input basis for subsequent anomaly detection.

[0032] Step S2: using the LSTM model to predict the next time after smoothing the data and calculate the prediction error to determine whether an anomaly occurs, which includes the following sub-steps: S2-1: input the past time smoothing vector into the trained LSTM model, and output the prediction . Specifically, the LSTM network architecture design includes: Input layer: input dimension , representing sequence length , feature dimension 4.

[0033] LSTM layer: double-layer LSTM, each layer has H (such as 64) hidden units, and each layer is followed by Dropout (dropout rate 0.2) to prevent overfitting.

[0034] Fully connected layer: map the hidden state of the last time to the output dimension 4, which can be represented by the following formula:

[0035] In the formula, , is the weight matrix of the fully connected layer; , is the bias term of the fully connected layer.

[0036] LSTM model training: Use mean square error (MSE) as the loss function, which can be represented by the following formula:

[0037] Use Adam optimization algorithm, learning rate ; training parameter Batch size , training rounds are 50-100 rounds, and training is stopped when there is no decrease for 10 consecutive rounds on the validation set, and the total samples are divided into training set, validation set and test set according to 8:1:1.

[0038] Online inference and sliding window update: Get the latest moving average vector every minute, construct the input vector , and input it into the above trained LSTM model to output the prediction And the prediction results are de-normalized, which can be embodied by the following formula:

[0039] Window sliding includes updating the sequence cache, discarding the earliest time , adding the latest time , and repeating the execution.

[0040] S2-2: Compare the true value with the predicted value to get the absolute error of each index, which can be embodied by the following formula:

[0041] S2-3: Pre-set threshold , and if any , the terminal enters the "abnormal pending decision" state, and the error vector is passed to the next step S3; Otherwise, let , where represents the counter of each terminal, which reflects how many times the terminal has been in the "normal" state continuously; Further, if exceeds the pre-set window (such as 30 detection periods), the threshold may be appropriately tightened, where is the threshold adjustment step (such as 1%), and is the lower limit of the threshold (such as 5%) to improve the sensitivity to potential minor abnormalities.

[0042] It should be noted that this step uses long short-term memory network (LSTM) to predict the time series of terminal running indicators, and compares the actual acquisition value with the prediction result to evaluate whether the current running state is abnormal; if the error exceeds the pre-set threshold, the "abnormal pending decision" state is triggered, providing input for the next reasoning module.

[0043] Step S3: After the prediction error is fuzzified by the membership function, the posterior probability of "immediate dispatch" is calculated by combining the Bayesian network and triggering the decision, including the following sub-steps: S3-1: For each error , define type membership, which can be embodied by the following formula:

[0044] In the formula, represents the th prediction error "abnormal" membership; represents the reference value (such as 10%); represents the slope (e.g. 0.5), which can be tuned according to historical data.

[0045] S3-2: Set all membership degrees as input evidence of the Bayesian network.

[0046] S3-3: Set "immediate dispatch needed" event , and calculate the posteriori probability of the "immediate dispatch needed" event according to the priori and conditional probability table (obtained by statistical analysis of historical samples), which can be embodied by the following formula:

[0047] wherein, the priori probability can be set as 0.1; represents the priori "no dispatch" probability; represents the probability of the first membership degree under the condition of needing dispatch, which is obtained based on the Bayesian network structure learning; represents the probability of the second membership degree under the condition of not dispatching; represents the posteriori probability of needing dispatch.

[0048] S3-4: Set a probability threshold , if the posteriori probability is greater than the threshold , it is determined that "dispatch is needed", and the next step S4 is entered; otherwise, dispatch is not triggered, and the system enters the "observation" state, and step S2 is re-executed at the next monitoring point (e.g. minutes).

[0049] For example, it is assumed that the current membership degree of a terminal is High; Medium; High; Medium, and the conditional probability table is referred to Figures 4-7 , then: ; If the probability threshold = 0.6, the terminal does not enter the dispatch process, but enters the observation queue.

[0050] It should be noted that, in this step, the abnormal error is first fuzzified to express the continuity of the abnormal degree; then, the abnormal membership degrees of multiple indexes are comprehensively analyzed by means of the Bayesian network, the probability of whether the current terminal needs to immediately perform resource dispatch is calculated, and whether to enter the dispatch link is decided based on the probability and the set threshold.

[0051] ​​Step S4: Select the optimal scheduling action based on Q-learning in the current state, and roll back to the last step to reselect if the effect is not as expected, including the following sub-steps: S4-1: Smooth the current four indicators of the terminal and the scheduling urgency probability State vector: ; S4-2: Define the set of executable actions .

[0052] S4-3: Calculate the reward in real time according to the action effect: Set the CPU or memory usage reduction rate after executing the action to (%) and the immediate reward ; If (5%, for example), it is considered a failure, and a negative value penalty is added to encourage the strategy to avoid invalid operations; Specifically, , where represents the usage rate (CPU or memory) before the action, represents the usage rate (CPU or memory) after the action; For example, if the CPU usage rate of a terminal before the action is 80% and the CPU usage rate after the action is 78%, then: = 2.5 .

[0053] S4-4: Use classic Q-learning iteration:

[0054] In the formula, the learning rate ; the discount factor ; represents the value of action in state at time , which is obtained through the Q table maintained inside the algorithm; represents the immediate reward obtained at time ; represents the optimal action value in state at time , which is obtained by taking the maximum value from the Q table;

[0055] S4-5: Execute action , if within minutes, then roll back to state and record the failure experience, reselect the suboptimal execution action; If , it is considered to meet the standard: calculate the reward function , where is the number of indicators that meet the conditions; and the reward value will be passed into the Q-learning learning formula to guide the selection of the same action in the future similar state; in addition, according to the Q-learning update rule, the corresponding Q value is improved, indicating that this is a "worth learning" behavior, and this successful experience can also be stored in the experience pool for subsequent policy playback or training; and it should be noted that if the terminal is successfully executed and achieves the load reduction goal in similar states for times (such as = 3), the corresponding terminal can be removed from the "abnormal state pool" and restored to the regular monitoring mode, reducing unnecessary frequent intervention.

[0056] It should be noted that when the inference result indicates that scheduling is required, this step constructs a reinforcement learning environment based on the current terminal state, selects the optimal resource scheduling action such as bandwidth limitation, task migration, etc. through the Q-learning strategy, and evaluates the scheduling effect in real time; if the effect is not as expected, an automatic rollback strategy is executed to reselect a more appropriate action.

[0057] Step S5: archive the data and decision records of each step, and periodically update the conditional probability table of LSTM and Bayesian network online, including the following sub-steps: S5-1: save each monitoring value, error, posterior probability, execution action and rollback times for subsequent analysis; S5-2: periodic online update, including daily or weekly update, the update content includes fine-tuning LSTM model parameters to improve prediction accuracy, adjusting the conditional probability table of Bayesian network, and optimizing the scheduling strategy; It should be noted that during the periodic online update process, the system first constructs a training sample set based on the terminal running data collected in the recent period of time, and fine-tunes the LSTM model. Generally, the first few layers of parameters are frozen, and only the output layer is fine-tuned to avoid overfitting. At the same time, the mean square error is used as the loss function, and the prediction accuracy is optimized through small batch training; at the same time, for the conditional probability table in the Bayesian network, the frequency of each membership interval in the scheduling and non-scheduling situations is counted, and the maximum likelihood estimation or Laplace smoothing method is used to update the conditional probability value, or the EM method is used to iteratively correct the posterior probability as a soft label, so as to dynamically adjust the inference model, ensure the response ability and decision accuracy of the system to new abnormal patterns.

[0058] To sum up, the application realizes panoramic real-time monitoring and dynamic decision optimization of the Xinyuan terminal running state by fusing time sequence anomaly detection, fuzzy Bayesian inference and reinforcement learning resource scheduling. Compared with the traditional monitoring method relying on static threshold, the application can identify trend anomalies based on terminal historical behavior, intelligently judge in combination with multi-dimensional indexes, and continuously improve the accuracy and adaptability of resource scheduling through strategy learning. In addition, the system has good explainability, can automatically roll back and reselect an optimized strategy when the decision does not meet expectations, and builds a closed-loop mechanism of data collection, risk identification, strategy execution and effect feedback, thereby improving the accuracy of monitoring and the intelligent level of management, and enhancing the stability and reliability of the terminal system in a complex heterogeneous environment.

[0059] Embodiment 2 Reference Figure 3 For the second embodiment of the application, which is different from the first embodiment, a multi-dimensional data fusion-based Xinyuan terminal panoramic real-time monitoring and decision system is also provided, comprising: a data collection module for periodically collecting running data of the Xinyuan terminal and smoothing the collected data using a sliding window mechanism; a time sequence prediction module for receiving the smoothed running data and predicting the running data at a future time based on a long short-term memory network, generating a predicted value, and calculating the error between the predicted value and the actually collected value; a fuzzy inference module for mapping the prediction error to a multi-dimensional membership value and inferring whether the terminal is in a scheduling trigger state based on a preset Bayesian network structure and a conditional probability table, and outputting the posterior probability of scheduling demand; a scheduling decision module for receiving the posterior probability and current state information to construct a state vector, and selecting a scheduling action based on a reinforcement learning strategy; a feedback rollback module for determining the scheduling effect after the scheduling action is executed, and if no resource state improvement is detected within a set monitoring period, rolling back to the state before scheduling and recording the scheduling process to a strategy update log; a model update module for updating and optimizing the reinforcement learning strategy based on the scheduling log provided by the feedback rollback module, and introducing a state convergence condition and an action success rate threshold in the scheduling strategy iteration process for strategy optimization.

[0060] The embodiment provides a computer device suitable for a multi-dimensional data fusion-based Xinyuan terminal panoramic real-time monitoring and decision method, comprising a memory and a processor; the memory is used to store computer executable instructions, and the processor is used to execute the computer executable instructions to realize a multi-dimensional data fusion-based Xinyuan terminal panoramic real-time monitoring and decision method as proposed in the above embodiment.

[0061] The computer device can be a terminal, which includes a processor, a memory, a communication interface, a display screen and an input device connected by a system bus. The processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for running the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is configured to communicate with external terminals in a wired or wireless manner. The wireless manner can be achieved by WIFI, an operator network, NFC (Near Field Communication) or other technologies. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer overlaid on the display screen, or a key, trackball or touchpad arranged on the shell of the computer device, or an external keyboard, touchpad or mouse, etc.

[0062] The embodiment also provides a storage medium having a computer program stored thereon, the program being executed by a processor to implement the method for realizing a ChinaSoft terminal panoramic real-time monitoring and decision-making based on multi-dimensional data fusion as described in the above embodiment. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as a static random access memory (SRAM), an electrically erasable programmable read-only memory (EEPROM), an erasable programmable read-only memory (EPROM), a programmable read-only memory (PROM), a read-only memory (ROM), a magnetic memory, a flash memory, a magnetic disk or an optical disk.

[0063] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application but not limit the present application. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced by equivalents without departing from the spirit and scope of the present application, which should be covered in the scope of the claims of the present application.

Claims

1. A method for real-time monitoring and decision-making of a Xiong'eng terminal panoramic based on multi-dimensional data fusion, characterized in that, Comprise: Step S1, real-time acquisition and sliding average smoothing processing of multi-dimensional indexes of terminal CPU, memory, network bandwidth and login times are carried out; Step S2, the smoothed data is predicted by using the LSTM model and the prediction error is calculated, which is used to determine whether an anomaly occurs; Step S3, after the prediction error is fuzzy by membership function, the posterior probability of "need to dispatch immediately" is calculated by combining Bayesian network and triggering decision; Step S4, the optimal scheduling action is selected based on Q-learning in the current state; Step S5, the data and decision records of each step are archived, and the conditional probability table of LSTM and Bayesian network is updated online periodically.

2. The method of claim 1, wherein the method is characterized in that: In the step S1, in order to obtain the smoothed key operation index, the system adopts the fixed length sliding window to collect the multi-dimensional original data of terminal CPU usage, memory usage, network bandwidth occupation and login times in a unified time sequence, and realizes the smoothing processing by arithmetic average of all original values in the window in each collection period; the length of the sliding window is dynamically selected in the interval of five minutes to ten minutes according to the actual deployment scene, which is used to balance the smoothing effect and response timeliness, and the smoothed values of each index are updated synchronously according to the unified timestamp.

3. The method of claim 1, wherein the method is characterized by: In the step S2, the system takes the multi-dimensional smoothed value sequence as input, and carries out time sequence prediction of the next time terminal operation index by using the trained long short-term memory network model, which contains multiple recursive neural units and Dropout mechanism to prevent overfitting; the system compares the prediction result with the original smoothed value collected at the same time, and takes the absolute value of the difference between them as the prediction error, so as to evaluate whether the current terminal operation state deviates from the normal trend.

4. The method of claim 1, wherein the method is characterized in that: In the step S2, for the terminal which is not triggered by continuous multiple detections, the internal counter of the system automatically accumulates to the preset value, and then the abnormal threshold is dynamically tightened according to the recent operation stability of the terminal, which is used to improve the sensitivity to potential small deviation, and the new threshold is used to execute subsequent error judgment after the threshold is tightened.

5. The method of claim 1, wherein the method is characterized in that: In the step S3, the system maps the multi-dimensional prediction error to continuous abnormal degree value according to the pre-designed membership function, and inputs the mapped membership value into the Bayesian network as evidence, the Bayesian network carries out reasoning according to the set prior probability and conditional probability table, calculates the posterior probability of "need to dispatch immediately" event of the terminal, and if the posterior probability exceeds the set threshold, the resource scheduling decision link is triggered.

6. The method of claim 1, wherein the method is characterized by: In the step S4, the system constructs a state vector containing smoothed CPU usage, memory usage, network bandwidth occupation and scheduling urgency probability according to the current multi-dimensional index smoothed value and scheduling urgency probability, and selects the optimal scheduling action in the predefined action set based on Q-learning algorithm, the action set includes limiting broadband, migrating task, cleaning cache and no operation, and the real-time reward is calculated before and after each action execution to guide the Q table update.

7. The method of claim 1, wherein the method is characterized by: In the step S4, for each selected scheduling action, the system sets a fixed monitoring time length after the action is executed. If the real-time data analysis determines that the CPU or memory usage rate is reduced by a certain percentage within the time length , the system automatically rolls back to the system state before the action is executed, records the failure experience in the rollback log, and reselects the suboptimal execution action.

8. A real-time monitoring and decision system for a China-inspired terminal panoramic based on multi-dimensional data fusion, based on the method for real-time monitoring and decision of a China-inspired terminal panoramic based on multi-dimensional data fusion according to any one of claims 1-7, characterized in that: Comprise: The data acquisition module is configured to periodically acquire running data of the Xinchuang terminal and to perform smoothing processing on the acquired data by using a sliding window mechanism. The time sequence prediction module is configured to receive the running data after the smoothing processing, to predict the running data at a future time based on a long short-term memory network, to generate a prediction value, and to calculate an error between the prediction value and an actually acquired value. The fuzzy inference module is configured to map the prediction error to a multi-dimensional membership value, to infer whether the terminal is in a dispatch triggering state based on a preset Bayesian network structure and a conditional probability table, and to output a posterior probability of a dispatch demand. The dispatch decision module is configured to receive the posterior probability and current state information to construct a state vector, and to select a dispatch action based on a reinforcement learning strategy. The feedback rollback module is configured to determine a dispatch effect after the dispatch action is executed, to rollback to a state before the dispatch if an improvement in a resource state is not detected within a set monitoring period, and to record the dispatch process to a policy update log. The model update module is configured to update and optimize the reinforcement learning strategy based on the dispatch log provided by the feedback rollback module, and to introduce a state convergence condition and an action success rate threshold into a dispatch strategy iteration process for policy optimization. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is characterized in that: The computer program is executed by the processor to implement the steps of the method for panoramic real-time monitoring and decision-making of the Xinchuang terminal based on multi-dimensional data fusion according to any one of claims 1-7.

10. A computer readable storage medium having stored thereon a computer program, characterized in that: The computer program is executed by the processor to implement the steps of the method for panoramic real-time monitoring and decision-making of the Xinchuang terminal based on multi-dimensional data fusion according to any one of claims 1-7.

Images