TW202016805A - System and method of learning-based prediction for anomalies within a base station - Google Patents

Abstract

A system and method of learning-based prediction for anomalies within a base station, the method includes: filtering data to generate filtered data; labeling the filtered data to generate label data; filtering the label data to generate filtered label data; and training, by a deep neural network, according to the filtered label data and training method to update the deep neural network.

Description

System and method for predicting abnormality of base station based on automatic learning

The invention relates to a system and method for predicting abnormality of a base station, and particularly to a system and method for predicting abnormality of a base station based on automatic learning.

The analysis and maintenance diagnosis of mobile network traffic and alarms are often indispensable links for troubleshooting and maintenance of mobile network maintenance and optimization of service quality. In the traditional operation method, base station communication monitoring analysis and maintenance problem diagnosis operations require maintenance personnel relying on the mobile network to enter the computer room, and use the service monitoring analysis to diagnose maintenance obstacles. This method relies heavily on personnel operation and professional experience. However, with the increasing growth of mobile network technology and increasingly complex mobile network architecture, at this time, if the traditional manual experience analysis is used to diagnose maintenance obstacles, it is not only laborious and time-consuming, but also easy to cause human misjudgment, and may not be able to find the problem in real time. The crux of the problem. The mobile network always generates traffic data and alarm information, and analyzing and mining these data has a significant meaning for improving the quality of mobile network maintenance. Not only can the maintenance level be changed from artificial judgment to intelligent decision-making, but also to integrate artificial intelligence technology to transform obstacle detection into active prediction and prevent and eliminate in advance to improve the quality of mobile network services.

The US patent "Network fault prediction and proactive maintenance system (application number: 09/337,209)" mentions a similar obstacle prediction concept. The patent states that multiple feature databases containing valid logs are created, which represent network status The source of anomaly alarms is specifically selected by a network of experts or administrators from a large group of logs. In the future, the network logs will predict future failures based on the analysis of the effective logs and characteristics in the database. This method belongs to offline learning in the field of machine learning, that is, the system parameters of the effective log or key features cannot be dynamically updated with the new log, and the entire batch of data must be entered into the model for parameter or structure adjustment.

The Chinese patent "Mobile network health evaluation method based on neural network and fuzzy comprehensive evaluation (application number: 201510236105.1)" mentions a similar obstacle prediction concept and belongs to online learning. Establishing a health evaluation standard through historical alarm messages in the alarm system combined with expert opinions. The patent states that the lowest number of alarms and the highest value of alarms are selected from the number of each type of alarm, and then combined with historical data and experts according to the size of the number Opinion, the warning indicators are divided into five levels: excellent, healthy, good, medium, and unhealthy. Finally, we train a transitive neural network to predict the health of the new mobile network system. However, this method only considers the health evaluation. The evaluation standard is a quantified index. The sensitivity and the amount of information are determined by the size of the quantization interval. The better way should consider the detection and prediction of various obstacles.

The US patent "Alarm prediction in a telecommunication network (application number: 15/029,834)" also mentions the concept of obstacle prediction that belongs to online learning. It collects key performance indicators in network components. The source of this indicator can be: using memory Volume, power level, dropped call rate and answer rate, etc. measure network quality indicators. The training method adopts the extension of directed graph and undirected graph in graph theory: Markov random field and bass network model. However, the above-mentioned methods are all derived from the field of graphics. In pattern recognition, objects have high similarity and denseness. In contrast, mobile network obstacles are sparse, that is, the number of normal and abnormal networks is significantly different. This will cause extreme deviation of the artificial intelligence training results, that is, the training results show that the mobile network is barrier-free and can achieve a high accuracy rate on average, but the accuracy rate and recall rate are not high. In addition, the alarms of the above methods are limited by the detail level of the alarm system provided by the base station supplier.

It can be seen that there are still many defects in the above-mentioned conventional methods. It is not a good design, but needs to be improved.

In order to improve many defects of the above-mentioned conventional methods, the present invention proposes a system and method for predicting abnormality of base stations based on automatic learning.

The invention provides a system for predicting abnormality of a base station based on automatic learning, including a storage medium and a processor. The storage medium stores multiple modules. The processor is coupled to the storage medium, and accesses and executes the modules of the storage medium. The modules include: a database, storing data; a filtering module, which filters the data to generate filtered data; and barrier stickers Labeling module, labeling the filtered data to generate label data; second screening module, filtering the label data to generate filtered label data; and deep neural network, based on the filtered Label data and training methods are trained to update the deep neural network.

The invention provides a method for predicting abnormality of a base station based on automatic learning, which includes: screening data to generate screened data; labeling the screened data to generate label data; screening the label data to Generate the filtered label data; and train the deep neural network according to the filtered label data and the training method to update the deep neural network.

Based on the above, the present invention aims to use the advantages of big data to reduce the misjudgment introduced by human intervention in the maintenance of a base station, and proposes a method for applying a supervised learning technique to a non-labeled data system.

In order to make the above-mentioned features and advantages of the present invention more obvious and understandable, the embodiments are specifically described below in conjunction with the accompanying drawings for detailed description as follows.

FIG. 1 is a schematic diagram of a system 10 for predicting base station abnormality based on automatic learning according to an embodiment of the present invention. The system 10 may include a storage medium 100 and a processor 200. The storage medium 100 stores multiple modules. The processor 200 is coupled to the storage medium 100 and can access and execute multiple modules of the storage medium 100. The plurality of modules of the storage medium 100 may include a database 110, a screening module 120, an obstacle labeling module 130, a second screening module 140, and a deep neural network 150. The database 110 can store data. The filtering module 120 can filter the data to generate filtered data. The obstacle labeling module 130 can label the filtered data to generate label data. The second filtering module 140 can filter the label data to generate the filtered label data. The deep neural network 150 can be trained according to the filtered label data and the training method to update the deep neural network 150.

The storage medium 100 can be used to store various software, data, and various types of code required by the system 10 during operation. The storage medium 100 may be, for example, any type of fixed or removable random access memory (RAM), read-only memory (ROM), flash memory (flash memory) ), hard disk drive (HDD), solid state drive (SSD) or similar components or a combination of the above components, the invention is not limited thereto.

The processor 200 may be, for example, a central processing unit (CPU), a graphics processing unit (GPU), or other programmable general-purpose or special-purpose microprocessors, digital signal processing Digital signal processor (DSP), programmable controller, application specific integrated circuit (ASIC) or other similar components or combinations of the above components, the invention is not limited thereto.

The system 10 can be implemented in at least one of the following: outdoor large wattage base stations, small wattage micro base stations based on self-organizing network management, etc., but the invention is not limited thereto.

FIG. 2 illustrates a flowchart of a method 20 for predicting base station abnormality based on automatic learning according to an embodiment of the present invention, where the method 20 may be implemented by the system 10 shown in FIG. 1. In step S201, the database 110 can receive data from a mobile network and store the data. The data may include at least one of the following: wireless access terminal data, core network signaling, and/or alarm messages. The system 10 decides to use core network signaling and/or wireless access terminal data as input to the filtering module 120 depending on the type of obstacle. Since the data of the core network is more confidential, if it is impossible to obtain or easily cause difficulty in the operation of the system 10, the data of the wireless access terminal is mainly used, and the data of the core network signaling is supplemented, that is, The wireless access terminal data and the core network signaling are used as the input of the filtering module 120 to generate the label data corresponding to the wireless access terminal data and the core network signaling through the barrier labeling module 130. On the other hand, the deep neural network 150 uses the wireless access terminal data to learn the obstacle labeling module 130 that has been integrated into the core network signaling data to take into account the real-time and convenience of data acquisition while considering the accuracy of prediction. The above data (ie: wireless access terminal data, core network signaling and/or alarm messages) corresponds to the mobile network, and the mobile network provides terminal mobile network services, and uses at least one of the following communication standards One: code division multiple access (CDMA), broadband frequency division multiple access (WCDMA), high-speed packet access (HSPA), evolved high-speed packet access (HSPA+), long-term evolution technology (LTE), global interoperable microwave storage Access (WiMAX) and advanced long-term evolution technology (LTE-A).

In some embodiments, the wireless access terminal data may include at least one of the following: physical resource block utilization rate, reference signal received power, received signal strength indicator, upstream and downstream throughput, number of upstream and downstream users, adjustment Variable and coding indicators, CPU usage, device time since last restart, signaling success rate, and memory usage.

In some embodiments, the alarm message may include at least one of the following: radio frequency alarm baseband processing unit alarm, connection alarm, single-version alarm, software upgrade configuration alarm, maintenance test alarm, idle and restart without warning.

In some embodiments, the core network signaling may include at least one of the following: packet loss rate, transmission delay number, bandwidth utilization rate, channel capacity, mobility management component connection establishment success rate, and user application type .

In step S202, the filtering module 120 receives data from the database 110, and filters the data to generate filtered data. Regardless of the input source of the filtering module 120, statistical key factor identification can be performed. The identification of statistical key factors can screen out key index factors to prevent the input parameter from being too large and affecting the system's immediacy. Specifically, the screening module 120 can generate a key model according to the data and the screening method. The key model can screen out unnecessary indicators in the data to generate the screened data, and the screening method includes at least one of statistical key factor identification and principal component analysis.

In some embodiments, the screening module 120 may decide whether to add a new barrier type to the barrier type according to the number of barrier types in the alarm message, where the added new barrier type may be defined by the personnel based on the use requirements of the personnel. For a detailed implementation of step S202, refer to FIG. 3 and Table 1.

FIG. 3 is a flowchart illustrating a detailed implementation of step S202 according to an embodiment of the present invention. Table 1 describes related parameters used by the method 20 according to an embodiment of the present invention. Table I

Please refer to Figure 3 and Table 1. In step S301, the screening module 120 may decide to use at least one of the wireless access terminal data and the core network signaling according to the type of obstacle to be predicted as the first set. Then in step S302, an example is removed from the first set to generate a second set, where the example makes the model constructed by the first set have the least amount of message reduction. In step S303, the P values of the remaining examples (that is, the examples in the second set) after excluding the above examples can be calculated, that is, the P of each example corresponding to the model constructed by the second set can be calculated value. In step S304, it is judged whether the maximum P value example in the calculated P value has reached the tolerance value (for example, it is judged whether the maximum P value example is less than 1e-6), and if so, step S305 is entered. If not, go to step S306. In step S305, the screening module 120 may use the second set as an input of the obstacle labeling module 130. In step S306, the screening module 120 may delete the example corresponding to the maximum P value and combine the remaining examples (ie, the second set minus the example corresponding to the maximum P value) into the updated first set, and perform step S302 again.

Returning to FIG. 2, in step S203, the obstacle labeling module 130 may label the filtered data to generate label data. Specifically, the obstacle labeling module 130 can perform feature filtering on the filtered data and perform normalization, that is, the data corresponding to each selected feature is converted into a standard score of zero mean and unit variation ( z-score). This action can avoid the inconsistency of the input parameter units or the difference in order of magnitude causing the output variation of the obstacle labeling module 130 to be too large, which can cause the weight variation of the deep neural network 150 that is subsequently trained to be too large.

In step S204, the second screening module 140 may receive the label data from the obstacle labeling module 130, and filter the label data to generate the filtered label data. Specifically, the second filtering module 140 can generate a second key model according to the label data and the second filtering method, wherein the second key model can filter out unnecessary indexes in the label data to generate the filtered label data, And the second screening method may include at least one of statistical key factor identification (or P-value hypothesis test) and principal component analysis (or principal component analysis based on feature vectors). The P-value hypothesis test has good visual intuition, and principal component analysis can convert the input coordinates into a feature coordinate system. This step can avoid the sparseness of the data density caused by the input dimension of the deep neural network 150 being too large, which is called: dimension disaster in the literature.

In step S205, the deep neural network 150 may receive the filtered label data from the second filtering module 140, and train to update the deep neural network 150 according to the filtered label data and the training method, in which the deep layer is updated The neural-like network 150 may, for example, update the weights used by the deep neural-like network 150. The training method may include at least one of the following (but the invention is not limited thereto): adopting the backward transfer training method and increasing the hidden layer of the deep neural network 150 layer by layer; adjusting the training learning rate to improve the training error rate; Use at least one of RMSprop and Momentum as an optimizer; adjust the output threshold of the output layer of the deep neural network 150 according to the ratio of positive examples and negative examples in the filtered label data; and use the deep neural network The output data of the path 150 is fed back to the deep neural network 150, and the weight of the deep neural network 150 is adjusted according to the filtered label data (or filtered data) and the output data of the deep neural network 150.

More specifically, the training method: adjust the output threshold of the output layer of the deep neural network 150 according to the ratio of positive examples and negative examples in the filtered label data, which may include at least one of the following: multiply the output threshold by The reciprocal of the ratio of positive examples and negative examples; and when the number of positive examples and negative examples reaches a certain gap, the larger number of positive examples and negative examples is discarded so that the ratio of positive examples and negative examples approaches one.

In some embodiments, when the output type of the deep neural network 150 to be classified is equal to two, a logarithmic loss function (binary_crossentropy) may be selected as the loss function. On the other hand, if the output type of the deep neural network 150 to be classified is greater than two, a normalized exponential function (softmax) may be selected as the loss function.

The deep neural network 150 may include an input layer and an output layer. The deep neural network 150 can also adjust the input data amount of its input layer according to the front window parameters of the sliding window, and adjust the output data amount of its output layer according to the rear window parameters of the sliding window, as shown in FIG. 4. FIG. 4 is a schematic diagram illustrating a data management method using a sliding window (by a deep neural network 150) according to an embodiment of the present invention. Specifically, the deep neural network 150 may use the filtered label data in the past first period (eg, the filtered label data output from the second filtering module 140) according to the front window parameters of the sliding window as its The input data of the input layer and the obstacles occurring in the second period in the future according to the parameters of the rear window to generate the output data of the output layer. In this embodiment, the front window parameter indicates the use of the past M minutes of data (for example, the filtered label data output from the second filter module 140) as the input of the deep neural network 150, and the rear window parameter indicates the prediction Obstacles will occur in the next N minutes. When N (or the rear window parameter) is set to zero, the system 10 switches from predicting system obstacles to detecting system obstacles.

In some embodiments, the training method of the deep neural network 150 may include feeding back the output data of the deep neural network 150 to the deep neural network 150, and according to the filtered label data (or filtered Data) and the output data to adjust the weight of the deep neural network. Specifically, the deep neural network 150 can compare the output data of the deep neural network 150 with the second period in the future before the data enters the input layer of the deep neural network 150 (for example: the future shown in FIG. 4 N minutes) filtered label data (or filtered data) (for example: filtered label data output from the second filtering module 140 or filtered data output from the filtering module 120), of which The output data of the neural network 150 is used to predict the occurrence of obstacles in the second period in the future. If the comparison result is inconsistent, it means that the filtered label data received by the deep neural network 150 from the second filtering module 140 does not have/has not occurred as predicted by the deep neural network 150, therefore, The deep neural network 150 can adjust the layers of the deep neural network according to the filtered label data (or filtered data) and the output data of the deep neural network 150 (for example: input layer, hidden layer and output layer) The weights of, improve the prediction accuracy of the deep neural network 150.

In some embodiments, the training method of the deep neural network 150 may include adopting the backward transfer training method and adding hidden layers of the deep neural network 150 layer by layer. FIG. 5 illustrates a schematic diagram of a reverse transfer deep neural network 150 according to an embodiment of the present invention. The training method of the deep neural network 150 can adopt the backward transfer training method, and adopt the method of deepening hidden layers layer by layer. When the training error is too low, the action to increase the reservoir layer is stopped. At this time, the deep neural network 150 has enough degrees of freedom to manipulate.

In some embodiments, the training method of the deep neural network 150 may include improving the training error rate by adjusting the training learning rate. Specifically, the way to further improve the test error is to adjust the training learning rate. This method can avoid the occurrence of overfitting, that is, the training error rate decreases while the test error rate shows an increased U-shaped structure.

In some embodiments, the training method of the deep neural network 150 may include using at least one of RMSprop and Momentum as an optimizer. Specifically, the choice of the optimizer falls into the regional optimal solution because the deep network is easily affected by the disappearance of the gradient during the backward transfer, so RMSprop or Momentum is used as the optimizer. There is no significant difference between the two, and it has the advantage of fast convergence and the best solution.

In some embodiments, the training method of the deep neural network 150 may include feeding back the output data of the deep neural network 150 to the deep neural network 150, and according to the filtered label data and the deep neural network 150 output data to adjust the weight of the deep neural network 150. In this way, the deep neural network 150 can use the output of the second filtering module 140 (or the filtering module 120) (that is, the filtered label data or the filtered data) as a benchmark, and then be fed back through the feedback method. The example of error recognition is re-entered into the deep neural network 150 to achieve self-adaptive learning of the deep neural network 150. FIG. 6 shows a flowchart of adaptive learning of a deep neural network according to an embodiment of the present invention.

As shown in FIG. 6, in step S601, a time t is defined. In step S602, data of a second period in the future (for example: N minutes in the future) is obtained. In step S603, the collected data is output through the obstacle labeling module 130 (or the second screening module 140). In step S604, the output of the deep neural network 150 simultaneously with step S603 is obtained. Step S605: Compare the output data (for example, the filtered label data) of the obstacle labeling module 130 (or the second filtering module 140) with the output data of the deep neural network 150, that is, compare two strokes Consistency of information. If the comparison result is consistent, step S606 is entered to re-execute the adaptive learning process at the next time point. If the comparison result is inconsistent, step S607 is entered, and the inconsistent examples are reset and fed back to the deep neural network 150 for parameter correction.

Returning to Figure 2, using the prediction of a lazy obstacle as a simple example, other obstacles or abnormalities can be inferred in the same way. In step S201, the database 110 may receive wireless access terminal data, core network signaling and/or alarm messages from a mobile network. In step S202, the screening module 120 decides to use wireless access terminal data as the input of the obstacle labeling module 130 according to the type of obstacle, and uses core network signaling as a verification tool. In this embodiment, the screening module 120 may determine, for example, statistical key factor identification to determine the physical resource block (PRB) usage and user activation quantity (or UEActive) for one minute. Quantity) to use these parameters as the basis for determining whether to stay idle or not, and use these parameters as the input of the obstacle labeling module 130. In some embodiments, the screening module 120 may determine whether to add a new type of obstacle associated with a dead machine to the type of obstacle according to the number of types of obstacles associated with a dead machine in the warning message.

Next, in step S203, the obstacle labeling module 130 may label the filtered data to generate label data. In this embodiment, the output data of the base station for one minute is used as the input of the obstacle labeling module 130, and then the obstacle labeling module 130 is used as an automatic labeling tool. In step S204, the second screening module 140 may receive the label data from the obstacle labeling module 130, and filter the label data to generate the filtered label data.

In step S205, the deep neural network 150 may receive the filtered label data from the second filtering module 140, and perform training according to the filtered label data and the training method to update the deep neural network 150. The deep neural network 150 can also adjust the input data volume of its input layer according to the front window parameters of the sliding window, and adjust the output data volume of its output layer according to the rear window parameters of the sliding window. The parameters of the sliding window are determined as follows: The back window parameters are used as the basis for the system sensitivity requirements within ten minutes, and the front window parameters are obtained by cross-validation. During the training of the deep neural network 105, the training method adopts the reverse transfer training method, and the training optimization goal is based on the accuracy, precision and recall rate. In addition, in order to avoid falling into the regional optimal solution, the optimizer chooses rmsprop or momentum, while the excitation function uses the ReLU function in the hidden layer, and the final output layer uses the Sigmoid function for the output of the reserved probability form.

Features and effects: The training of deep neural networks is often limited to the need for labeling data, and the alarm system provided by the equipment supplier is often insufficient to fully reflect the service quality of the mobile network. The method for predicting abnormality of a base station based on automatic learning provided by the present invention is to introduce core network signaling data into the obstacle labeling module to increase the accuracy of labeling actions and solve the dilemma of automatic labeling. The neural network input still uses the wireless access terminal data to learn the obstacle labeling module of the learned core network data. This method mainly considers that the core network data is not easy to obtain, but the integration of the obstacle labeling module can increase the obstacle recognition rate. In addition, the sliding window data reading method can effectively use the past traffic data and freely control the prediction interval to determine the system reaction flexibility. The present invention has the following advantages when compared with other conventional technologies:

The method for predicting abnormality of a base station based on automatic learning proposed by the present invention simultaneously considers core network signaling and wireless access terminal data to identify mobile network obstacles or abnormalities. The specific data used depends on the type of obstacle, and is used as an obstacle labeling module, and then the labeling module is used to train a deep neural network for obstacle prediction.

The data reading method of the sliding window proposed by the present invention can use the front window parameters to decide whether to use the data of the past first period (for example: the past M minutes) as the input of the deep neural network, and the rear window parameters to determine the Predict the obstacles that will occur in the base station in the second period of the future (eg: N minutes in the future). This method can effectively use data to avoid the occurrence of dimensional disaster issues, and can adjust the prediction interval according to the system sensitivity requirements.

Subordinate supervised learning methods such as deep learning and machine learning are limited to the need for labeling materials. The alarm system of the base station is a good labeling tool, and this tool is subject to the number of obstacles or abnormal items defined by the system supplier and the level of detail, and may not be able to immediately generate alarms or obtain good actions due to the busy system. Network quality monitoring can define more detailed obstacles. The automatic labeling mechanism is completed through the barrier labeling module and can further provide the labeling examples required for deep neural network training.

Obstacles in the base station often have a significant difference in proportion, that is, the scarcity of training data on the obstacle station will cause the deep learning mechanism to output errors. This feature is that the network output has high accuracy and the accuracy and recall are low, so the adjustment of the threshold is additionally considered, that is, the threshold of the output layer of the deep neural network is multiplied by the reciprocal of the ratio of the positive and negative examples to correct the output judgment. Furthermore, when the amount of data is huge, the above threshold can be replaced by discarding the positive data to find that the ratio of the number of positive and negative examples is close to one.

The invention can compare the trained deep neural network with the output of the obstacle labeling module, and retrain the example of wrong prediction or feed back to the input of the deep neural network with more weights to correct the deep layer Neural network-like parameters achieve the advantages of online learning.

Based on the above, the present invention aims to use the advantages of big data to reduce the misjudgment introduced by human intervention in the maintenance of base stations, and proposes a method for non-labeled data systems to apply supervised learning techniques, so as to detect the level of base station maintenance Turn to proactive prediction. The present invention utilizes the communication quality monitoring instrument installed in the mobile network computer room to automatically monitor all aspects of the mobile network's traffic, and sends the alarm information in the traffic and alarm system to the database for storage . The filtering module extracts database data and creates key models for various obstacles, which can be used as a tool for identifying and detecting various obstacles in the mobile network in the future. The obstacle may also be a maintenance obstacle that is not displayed in the alarm system. If the base station alarm system can not effectively reflect the operation status of the base station, then the artificially defined types of obstacles, traffic data, and core network data are statistically modeled and identified to obtain key models to filter out unnecessary indicators to Accelerate the system's real-time operation and avoid collinearity problems that can cause system instability

The output data of the obstacle labeling module can be used to train a deep neural network to proactively predict whether the network is about to impede. Examples of neural network characteristics include communication data, equipment life and maintenance cycle and other related factors. In order to avoid dimensionality disasters, statistical key factor identification (or significant statistical factor identification and identification) or Principal Components Analysis (PCA) can be performed on the input examples to achieve the purpose of dimensionality reduction.

The data can be used in a sliding window method. The front window parameter indicates that the data in the first period of the past is used as the characteristic dimension of an example, and the rear window parameter indicates that the obstacle or abnormality will be predicted in the second period of the future. The specific window size is obtained by cross-validating the test data set with the minimum test error or depending on the sensitivity and maneuverability of the specific system.

The general situation is that the alarm system cannot cover all network obstacles. The reason may be that the system supplier is unwilling to admit that the defect exists, and the operator must define a more detailed network maintenance obstacle or anomaly to facilitate the maintenance operation or because the system is busy. Output real-time wireless access terminal data, etc. One method is to use the data of the core signaling of the wireless network, such as control signaling, to determine the obstacles for different usage scenarios and application scenarios. The specific method can be used by the network expert for the input of the barrier labeler one by one for the core network signaling abnormal parameters. The advantage of one-by-one inclusion is that the entropy can be monitored serially with the parameters until the entropy reaches the previously defined tolerance value. Another advantage of this approach is that the complexity of the barrier labeler can be controlled to achieve immediate system operability. Another method is to adopt unsupervised learning, and put the control signaling corresponding to the normal and abnormal networks into the clustering algorithm, so that it can usually distinguish the normal or abnormal state significantly. If it does not mean that the mutual information (Mutual Information) is close to zero, the input parameters of the cluster algorithm must be adjusted appropriately. The cluster algorithm can be used to determine whether the Euclidean or Manhattan distance of the input parameters is significantly greater than the normal value. If the input parameter dimension is too large, the Euclidean distance will gradually become unsuitable, and any form of dimensionality reduction and correlation coefficient method, such as the principal component analysis of the feature vector, is necessary for the previous operation.

The training of deep neural networks is carried out layer by layer, which mainly controls the degree of freedom and avoids the gradient disappearing and falling into the regional optimal solution. The parameter optimization method of the deep neural network is the cross-compatibility between the training set and the test set, and the output of the deep neural network of the training is expected to be consistent with the output of the obstacle labeling module. If no agreement can be reached, the output data representing the example of wrong estimation is fed back to the deep neural network for parameter adjustment to achieve the function of online learning and dynamic adjustment. In addition, the threshold adjustment issue is also included in the present invention to avoid the base station Unbalanced labeling of obstacles causes extreme deviations in learning results.

The invention can reduce the judgment error of the operator, improve the operation stability of the base station equipment, and reduce the maintenance cost. The present invention mainly includes: (1) Establishing a key factor analysis system for obstacles, and performing statistical key factor identification or principal component analysis based on alarm messages in the alarm system in combination with expert opinions or based on communication data to obtain key models. Key modules can filter out effective data. If the alarm message in the alarm system cannot effectively reflect the detailed quality index of the mobile network, the data of the mobile network core network signaling can be used as the input parameter of the screening module. (2) Establish a deep neural network, and determine the input and output architecture of the deep neural network based on the alarm messages and communication data. (3) Training reverse transfer deep neural network, according to the training error rate, training accuracy rate or training recall rate of the neural network, the amount of training data, data feature engineering, the number of hidden layers of the deep learning network, the most Fine-tuning the optimizer and training learning ratio. In addition, the output layer of the deep neural network still needs to consider the threshold adjustment issue, so that it can avoid the deviation of the training model due to the extreme disparity between the normal and abnormal examples of the base station training data. (4) Feedback and self-learning, compare the prediction results of the neural network with the output results of the obstacle labeling module, and feed back to the deep neural network for parameter modification or structural adjustment in real time to achieve the purpose of online learning and advantage.

The above detailed description is a specific description of a feasible embodiment of the present invention, but this embodiment is not intended to limit the patent scope of the present invention, and any equivalent implementation or change without departing from the technical spirit of the present invention should be included in The patent scope of the present invention.

10: System for prediction of base station anomalies based on automatic learning 100: storage media 110: database 120: screening module 130: barrier labeling module 140: second screening module 150: deep neural network 20: based on Method 200 for automatically learning base station anomaly prediction: processor I1, I2, In, Ii, H1, H2, Hn, Hj, y1, y2, ym, yk: neurons S201, S202, S203, S204, S205, S301, S302, S303, S304, S305, S306, 601, 602, 603, 604, 605, 606, 607: steps

FIG. 1 is a schematic diagram of a system for predicting base station abnormality based on automatic learning according to an embodiment of the present invention. FIG. 2 is a flowchart illustrating a method for predicting anomaly of a base station based on automatic learning according to an embodiment of the present invention. FIG. 3 is a flowchart illustrating a detailed implementation of step S202 according to an embodiment of the present invention. FIG. 4 is a schematic diagram of a data management method using a sliding window according to an embodiment of the present invention. FIG. 5 shows a schematic diagram of a reverse transfer deep neural network according to an embodiment of the present invention. FIG. 6 shows a flowchart of adaptive learning of a deep neural network according to an embodiment of the present invention.

10: Base station anomaly prediction system based on automatic learning

100: storage media

110: database

120: Screening module

130: Barrier labeling module

140: Second screening module

150: deep neural network

200: processor

Claims (16)

A system for predicting abnormality of a base station based on automatic learning includes: a storage medium storing a plurality of modules; and a processor coupled to the storage medium, the processor accessing and executing the modules of the storage medium, The modules include: a database to store data; a filtering module to filter the data to generate filtered data; an obstacle labeling module to label the filtered data to generate label data; second The filtering module filters the label data to generate filtered label data; and the deep neural network, trains to update the deep neural network according to the filtered label data and the training method. The system as described in item 1 of the patent application scope, wherein the data includes at least one of the following: wireless access terminal data, core network signaling, and alarm messages. The system as described in item 2 of the patent application scope, wherein the screening module determines whether to add a new obstacle type to the obstacle type according to the number of obstacle types in the warning message. The system as described in item 2 of the patent application scope, wherein the screening module generates a key model based on the data and the screening method, wherein the key model screens out unnecessary indicators in the data to generate the screened data, And the screening method includes at least one of statistical key factor identification and principal component analysis. The system as described in item 1 of the patent application scope, wherein the second screening module generates a second key model based on the label data and the second screening method, wherein the second key model screens out unnecessary data in the label data To generate the label data after the screening, and the second screening method includes at least one of statistical key factor identification and principal component analysis. The system as described in item 1 of the patent application scope, wherein the training method includes at least one of the following: adopting backward transfer training and adding hidden layers layer by layer; improving the training error rate by adjusting the training learning rate; using RMSprop and Momentum At least one of them serves as an optimizer; adjusts the output threshold of the output layer of the deep neural network according to the ratio of positive examples and negative examples in the filtered label data; and The output data is fed back to the deep neural network, and the weight of the deep neural network is adjusted according to the filtered label data and the output data. The system as described in item 6 of the patent application scope, wherein the output threshold of the output layer of the deep neural network is adjusted according to the ratio of positive examples and negative examples in the filtered label data, including at least one of the following : Multiply the output threshold by the reciprocal of the ratio of the positive example and the negative example; and when the number of the positive example and the negative example reaches a certain gap, discard the larger number of the positive example and the negative example to make the positive example The ratio of the case and the counterexample is close to one. The system as described in item 6 of the patent application scope, wherein the deep neural network further includes an input layer and an output layer, and the deep neural network further adjusts the input data amount of the input layer according to the front window parameters of the sliding window, And adjusting the output data amount of the output layer according to the rear window parameters of the sliding window, including: using the filtered label data in the past first period as input data of the input layer according to the front window parameters; and The window parameter estimates the obstacles that occur in the second period in the future to generate the output data of the output layer. The system as described in item 8 of the patent application scope, wherein the output data of the deep neural network is fed back to the deep neural network, and the deep neural network is adjusted according to the filtered label data and the output data The weights of the neural network include: comparing the output data with the filtered label data in the second period in the future, and if the comparison result is inconsistent, the deep class is adjusted according to the filtered label data and the output data The weight of the neural network. The system as described in item 1 of the patent application scope, wherein the data corresponds to a mobile network, and the mobile network provides terminal mobile network services, and uses at least one of the following communication standards: code division multiple access ( CDMA), wideband code division multiple access (WCDMA), high-speed packet access (HSPA), evolved high-speed packet access (HSPA+), long-term evolution technology (LTE), global interoperable microwave access (WiMAX) and advanced long-term Evolutionary technology (LTE-A). The system as described in item 1 of the patent application scope, wherein the system is implemented in at least one of the following: outdoor large wattage base stations, small wattage micro base stations based on self-organizing network management and control. The system as described in item 2 of the patent application scope, wherein the alarm message includes at least one of the following: radio frequency alarm baseband processing unit alarm, connection alarm, single version alarm, software upgrade configuration alarm, maintenance test alarm, Stand still and restart without warning. The system as described in item 1 of the scope of patent application, wherein the data of the wireless access terminal is the utilization rate of the physical resource block, the reference signal received power, the received signal strength indicator, the upstream and downstream throughput, the number of upstream and downstream users, the adjustment At least one of the change and coding index, the CPU utilization rate, the time since the device was last restarted, the success rate of signaling establishment, and the memory usage, and the core network signaling is the packet loss rate and the number of transmission delays , At least one of the bandwidth utilization rate, channel capacity, the success rate of link establishment of the mobility management component, and the user application category. The system as described in claim 1 of patent scope, wherein the data is screened to generate the screened data, including: Step 1, including: deciding to use the wireless access terminal data and core network according to the type of obstacle to be predicted At least one of the signalings is used as the first set; Step two includes: removing an example from the first set to generate a second set, where the example minimizes the model constructed by the first set The amount of message reduction; step three, including: calculating the P value of each example corresponding to the model constructed by the second set, and determining whether the maximum P value is less than 1e-6, and if so, using the second set as the barrier sticker If the input of the target module is not, delete the example corresponding to the maximum P value and make the remaining examples combined into the updated first set, and perform step two again. The system as described in item 1 of the patent application scope, where if the output type of the deep neural network to be classified is equal to two, the logarithmic loss function (binary_crossentropy) is selected as the loss function, and if the deep neural network is to be classified If the output type is greater than two, the normalized exponential function (softmax) is selected as the loss function. A method for predicting base station anomalies based on automatic learning, including: filtering data to generate filtered data; labeling the filtered data to generate label data; filtering the tag data to generate filtered data Label data; and the deep neural network is trained according to the filtered label data and the training method to update the deep neural network.

Images