An automatic neighbor relationship method based on machine learning

Abstract

The present invention is an automatic neighbor relationship method based on machine learning, comprising: a first step: collecting key performance indicator (KPI) data of a wireless communication system and forming a data set D through preprocessing; a second step: training a neighbor relationship classifier (FDTC) using historically collected KPI data; a third step: inputting KPIs collected in real time into the FDTC to perform neighbor relationship judgment, and collecting KPI data in real time as a test set; a fourth step: updating and training the FDTC using newly collected KPI data; a fifth step: based on the training of the FDTC in the second step, using a serial enhanced boosting strategy for ensemble learning of the FDTC to train the enhanced FDTC (i.e., BFDTC); and a sixth step: based on the training of the FDTC in the second step, using a parallel bagging strategy for ensemble learning of the FDTC to train the random forest FDTC (i.e., RF-FDTC). Compared with traditional automatic neighbor relationship methods, the wireless network neighbor relationship generated by the present method is more accurate and stable.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent wireless networks in wireless communications, and in particular relates to an automatic neighbor relationship method based on machine learning. Background Art

[0002] In wireless communication networks, users can disconnect from their current serving cell and quickly connect to a new cell with better service quality. This phenomenon is called handover (HO). Neighbor relations (NR) serve as a crucial reference for user handovers, impacting communication quality after handover and being the first parameter to be determined when a base station joins the network. The Neighbor Cell List (NCL) records the neighbor relationships between a base station and surrounding cells in a wireless network. As part of the base station database, it requires regular maintenance as the network operates and the network topology changes. Automatic Neighbor Relations (ANR) technology, the foundation of self-configuration technology, requires rapid acquisition of neighbor relationships between cells in the network to accurately reflect the wireless network topology. Neighbor relationships are often assumed to be infrequently changing, so previous ANR technologies focused solely on achieving higher accuracy in determining neighbor relationships, ignoring the time cost of maintaining the neighbor relationship list.

[0003] In future 6G wireless networks, spectrum resources will be more fully utilized, achieving integrated coverage of "air, land, and sea". In particular, satellites, high-altitude platforms, and drones will become important coverage methods. Moreover, since they are mobile infrastructure, their deployment location and density can be dynamically deployed according to user traffic demand. The wireless network topology changes highly dynamically, and the wireless network neighbor relationship becomes unstable.

[0004] To address these issues, the present invention proposes a machine learning-based automatic neighbor relationship method. This method combines the advantages of decision tree models and support vector machines (SVMs) to address the unstable neighbor relationship issues in scenarios where wireless network topologies frequently change, making it suitable for practical communication applications. Summary of the Invention

[0005] Technical Problem: The purpose of this invention is to propose an automatic neighbor relationship method based on machine learning, using a neighbor relationship classifier to solve the problem of determining neighbor relationships in wireless networks. On the one hand, a fuzzy decision tree (FDT) is proposed to segment wireless network data, which can reduce the size of the data set and reduce inter-class overlap. On the other hand, the segmented data set is used to train an SVM to determine neighbor relationships, resulting in a fuzzy decision tree classifier (FDTC) model suitable for scenarios where the future 6G wireless network topology frequently changes. The performance is further improved by ensemble learning of the FDTC. Compared with traditional automatic neighbor relationship methods, the wireless network neighbor relationships generated by this method are more accurate and stable.

[0006] Technical solution: The present invention provides an automatic neighbor relationship method based on machine learning, comprising the following steps:

[0007] Step 1: Collect the key performance indicator (KPI) data of the wireless communication system and form a dataset D through preprocessing;

[0008] Step 2: Utilizing historically collected wireless communication system key performance indicator (KPI) data, the neighbor relationship classifier (FDTC) is trained. First, a fuzzy decision tree (FDT) is generated through training to perform fuzzy classification on the KPI data. This divides the KPI dataset into multiple small sub-datasets with consistent features and stores them at corresponding FDT leaf nodes. Then, a support vector machine (SVM) with a Gaussian kernel function is trained at the FDT leaf nodes using the corresponding sub-datasets for accurate classification. This allows the neighbor relationship classifier (FDTC) to directly determine network neighbor relationships based on the input wireless network KPI data.

[0009] Step 3: Input the real-time collected wireless communication system key performance indicators (KPIs) into the FDTC to perform neighbor relationship judgment. Real-time network KPI data is collected as a test set. After first being identified by the FDT, it is fed into the SVM classifier on the corresponding leaf node. The SVM classifier then directly outputs the neighbor relationship judgment result, determining whether the target cell is a neighbor cell of the source cell.

[0010] Step 4: Update and train the FDTC using the newly collected wireless communication system key performance indicator (KPI) data. The newly collected wireless network KPI data set is divided into sub-datasets using the FDT structure and stored on the corresponding leaf nodes. The classification accuracy of the SVM classifier of each leaf node in the FDTC is calculated. Leaf nodes with classification accuracy below the accuracy threshold are deleted, i.e., pruning. The FDTC is then retrained at the pruned location using the sub-dataset stored on the corresponding leaf node, i.e., update training. The updated FDTC can better judge neighbor relationships based on the real-time collected wireless network KPI data.

[0011] Step 5: Based on the FDTC trained in step 2, a serial boosting strategy is used to integrate and learn the FDTC. This is used to train the enhanced FDTC, or BFDTC. This method improves the accuracy of determining wireless network neighbor relationships by serially boosting the base classifier.

[0012] Step 6: Based on the FDTC training in step 2, a parallel bagging strategy is used to ensemble learn the FDTC to train a random forest FDTC, namely RF-FDTC. The total number of base classifiers in RF-FDTC is set to L, where l represents the base classifier number, l = 1, 2, ..., L. By training multiple base classifiers in parallel, the accuracy of determining wireless network neighbor relationships is improved.

[0013] in:

[0014] The first step is specifically:

[0015] Step 1.1, record the key performance indicators (KPIs) of the wireless communication system;

[0016] Use the wireless network operation and maintenance management system OAM to obtain the KPI data reported by each base station at the tth moment and record it as a data set Where N is the number of elements in the data set X, (bs i ,bt i ,x i ,y i ) is the i-th element in X, i = 1, 2, ..., N; bs i and bt i Respectively represent the identifiers of the source base station and the target base station; x i ∈R K Indicates the K-dimensional key performance indicator KPI data of the target base station collected and reported by the source base station at a certain time, including base station side data {P BS} and user-side data {P UE The base station side data includes the target base station longitude coordinates lat and latitude coordinates lon, the geographical distance dis from the source base station to the target base station, the target base station antenna transmission power tp, and the target base station load rate lr; the user side data includes the target base station average reference signal received power rsrp measured by the source base station serving users, the number of successful user handovers from the source base station to the target base station hos and the handover success rate hor, x i The definition is as follows:

[0017] x i =(lat,lon,dis,tp,lr,rsrp,hos,hor) (1)

[0018] y iis x i The label indicates the neighbor relationship between the target base station and the source base station, y i = +1 means the target base station is a neighbor of the source base station, y i =-1 indicates that the target base station is not a neighbor of the source base station;

[0019] Step 1.2: Preprocessing of KPI data of wireless communication system:

[0020] The KPI data of consecutive T moments are averaged and normalized, where the value of T is determined by the wireless network operation and maintenance management system OAM. The geographical coordinates of the target base station, the geographical distance from the source base station to the target base station, and the radio frequency configuration parameters of the target base station are fixed values ​​that do not change over time. The geographical coordinates are normalized to a number in the range of [-1, 1] according to formula (2), and the geographical distance and radio frequency configuration parameters are normalized to a number in the range of [0, 1] according to formula (3), where x ik Indicates the number x in the i-th element of the dataset X i The kth parameter, k = 1, 2, ..., K, is the KPI parameter mentioned above, x ik ' represents the normalized KPI parameter,

[0021]

[0022] The target base station load rate, user-side data including the target base station average reference signal received power (RSRP) measured by the source base station serving users, the number of successful user handovers from the source base station to the target base station, and the handover success rate are dynamic values ​​that change over time. First, the average value of T consecutive moments is calculated according to formula (4), where represents the average value of the kth KPI parameter,

[0023]

[0024] The range of base station load rate and handover success rate is within [0,1]. The base station reference signal received power RSRP and the number of user handover successes are normalized according to formula (3);

[0025] Key performance indicator KPI data of wireless communication system x i In its parameter x ik According to the above process, perform averaging or normalization respectively and use x i ″ represents that lat, lon, dis and tp are normalized according to formula (3) to obtain lat′, lon′, dis′ and tp′, and lr and hor are averaged according to formula (4) to obtain and First normalize rsrp and hos according to formula (3) and then calculate the average value according to formula (4) to obtain and x i ″Reflects the wireless network status of the past T consecutive moments, and converts the elements (bs i ,bt i ,x i ,y i ) in wireless communication system key performance indicator KPI data x i Replace with x i ″, get the data set

[0026] The second step is as follows:

[0027] Step 2.1: Define the fuzzy decision tree classifier (FDTC). FDTC consists of two parts: fuzzy classification and precise classification. The fuzzy decision tree (FDT) first performs fuzzy classification on dataset D, using a decision tree structure to divide dataset D into sub-datasets with higher purity. At the leaf nodes of the FDT, the sub-datasets are used to train a support vector machine (SVM) to precisely classify wireless network neighbor relationships. The Gini index (GINI) is used to measure dataset purity. For neighbor relationship classification, samples have only two labels: neighbor cell or non-neighbor cell. Assuming the probability of a sample label being a neighbor cell is p, the Gini index (GINI) of the KPI dataset D is calculated as follows:

[0028] GINI(D)=2p(1-p) (5)

[0029] Step 2.2, initialize FDT,

[0030] Define the root node of FDT, store the data set D at the root node, and extract the K-dimensional wireless network KPI data x i The kth KPI parameter x in ″ ik ″ is the kth feature a of the dataset D k , define the feature set A of the dataset D = {a1, a2, ..., a K};

[0031] Step 2.3, FDT generates a decision,

[0032] Check the "purity" and node depth of the dataset at the current node. If the current node depth is greater than the depth threshold Depth_thr or the GINI index of the dataset at the current node is less than the set threshold GINI_thr, terminate the FDT generation step for the current node and jump to step 2.6. The FDT depth threshold Depth_thr limits the maximum height of the FDT tree structure and constrains the number of leaf node classifiers. The GINI index threshold GINI_thr limits the "purity" of the dataset at the current node and affects the training cost of the leaf node classifiers. Both Depth_thr and GINI_thr are determined by the scale of wireless network KPI data.

[0033] Step 2.4, calculate the optimal segmentation feature a of the dataset D * And the best split value x for this feature ** ″,

[0034] Randomly sample the data set D to obtain the data set D k , select a feature a k , respectively according to feature a k The values ​​of are reordered from large to small to obtain the data set D k ′, select the median element of the data as the initial split point x mk ″, according to x mk The dataset D is divided into two parts D left and D right ; According to formula (6), the information gain brought by the current segmentation point is calculated, which is also feature a k Current maximum information gain Gain *k ;

[0035]

[0036] To find the feature a k The best split point x on *k ″,x mk The value of ″ is based on D k ' is sorted towards larger and smaller values; x mk After the value changes, the weights of the divided data sets are no longer equal. Assuming D left The data volume accounts for q, q<0.5, and the optimal information gain is updated according to formula (7). mk ″Exploration until Gain *k Until it stops increasing, record the x at this time mk ″ is the optimal split point x *k ″;

[0037] Gain *k =max(Gain *k ,qGINI(D left)+(1-q)GINI(D right )) (7)

[0038] For each feature a in the feature set A k Repeat this step and compare the k The best information gain extracted from the above is selected, and the x with the largest information gain is selected *k ″ is the best segmentation value x of the dataset D ** ″, its corresponding feature a k is the best segmentation feature a * ;

[0039] Step 2.5, FDT according to a * and x ** ″Split the dataset D into D left and D right , generate a binary decision tree on the node node, and D left and D right Stored in the left and right subtree nodes respectively left and node right Repeat steps 2.3 to 2.5 for the left and right subtree nodes;

[0040] Step 2.6, delete the data set D on the current node node Base station index number (bs i ,bt i ), then the dataset The SVM classifier trained with Gaussian kernel function is stored in the current leaf node of FDT to complete the FDTC training.

[0041] The third step is specifically,

[0042] Step 3.1, each KPI data x i The decision tree structure after FDTC is sent to the corresponding leaf node SVM classifier, and SVM predicts its classification label y i , if y i = +1, then the source base station and the target base station establish a neighbor relationship. If y i =-1, the source base station and the target base station are controlled not to establish a neighbor relationship, and the neighbor relationships between the base stations are summarized into a neighbor relationship table NRT;

[0043] In step 3.2, the FDTC classification results are verified using the actual neighbor relationships automatically corrected during wireless network operation. The overall accuracy of FDTC's neighbor relationship judgment is calculated. Furthermore, the accuracy of the SVM classifier on each leaf node is calculated. An accuracy threshold, Accuracy_thr, is set. Its value is determined by the expected accuracy of the wireless network for neighbor relationships. If the overall FDTC accuracy is higher than Accuracy_thr, the FDTC structure remains unchanged, and NRT is continuously generated and maintained as the wireless network operates. If the overall FDTC accuracy is lower than Accuracy_thr or the wireless network topology changes, the fourth step is to update and train the FDTC.

[0044] The fourth step is as follows:

[0045] Step 4.1: The newly collected wireless communication system key performance indicator KPI dataset D test As a test set, it is divided into sub-datasets through the FDT structure and stored in the corresponding leaf node positions;

[0046] Step 4.2, FDTC pruning: traverse all leaf nodes of FDTC, delete SVM classifiers with accuracy lower than Accuracy_thr, use the test set stored in the corresponding node as the training set, store it in the corresponding leaf node, check the tree structure of the entire FDTC, further delete the tree nodes where both the left and right subtrees are pruned, and store the training set stored in the left and right subtrees at the node position of the previous level;

[0047] Step 4.3, FDTC update, at the pruning position, use the training set stored at the node position to retrain FDTC according to the steps in the second step to complete FDTC.

[0048] The fifth step is as follows:

[0049] Step 5.1: Define the weight W of the wireless communication system key performance indicator KPI dataset D l , W l =(w l1 ,w l2 ,...,w li ,...,w lN ), where w li Indicates the weight of the i-th element in the dataset D during training, l represents the base classifier number in the enhanced fuzzy decision tree classifier BFDTC, l = 1, 2, ..., L, initialization

[0050]

[0051] Where N represents the number of elements in the data set D;

[0052] Step 5.2: When training the first base classifier, follow the second step to train the base classifier FDTC l , use h l FDTC l Network KPI data x i The prediction result of ″ is expressed as h l (x i ″); Calculate the fitting error rate e according to formula (8) l , e l Indicates the current FDTC l The probability of misjudging the network neighbor relationship reflects the learning progress of BFDTC.

[0053]

[0054] Calculate the current FDTC according to formulas (9) and (10): l The weight α l And the sample weight W of the next round of training l+1 ,

[0055]

[0056] Save FDTC l and α l ;

[0057] Step 5.3: When l does not exceed the set total number of base classifiers L, repeat step 5.2 and construct the final neighbor relationship classifier BFDTC according to formula (11).

[0058]

[0059] Step 5.4, neighbor relationship prediction: the newly collected wireless communication system key performance indicator KPI dataset D test The test set is input into BFDTC, which directly outputs the classification label and generates the wireless network NRT.

[0060] The sixth step is as follows:

[0061] Step 6.1: When training the first base classifier, randomly extract some data from the wireless communication system key performance indicator KPI dataset D. The proportion of extracted data is determined by the size of the dataset. Randomly extract some features from the feature set of D to form the dataset D. l , using D l According to the second step, train the base classifier FDTC l , save FDTC l ;

[0062] Step 6.2, when l does not exceed the set total number of base classifiers L, repeat step 6.1; save the L FDTC l Construct a parallel random forest fuzzy decision tree classifier RF-FDTC;

[0063] Step 6.3, neighbor relationship prediction: the newly collected wireless communication system key performance indicator KPI dataset D test As the test set input RF-FDTC, multiple base classifiers FDTC l The output is presented in the form of label probability, and the neighbor relationship is predicted based on the label probability to generate the wireless network NRT.

[0064] Beneficial Effects: The machine learning-based automatic neighbor relationship method proposed in this paper has the following advantages: Compared with traditional automatic neighbor relationship methods, this method reduces training time overhead by utilizing the distributed update strategy of FDTC while maintaining the performance of the neighbor relationship classifier, making it adaptable to future 6G wireless network scenarios with frequently changing network topologies. Furthermore, by using ensemble learning of FDTC, this method further improves the accuracy and stability of neighbor relationship judgment. DETAILED DESCRIPTION

[0065] Step 1: Collect key performance indicator (KPI) data of the wireless communication system.

[0066] Step 1.1: Record wireless network KPI data.

[0067] Use the wireless network operation and maintenance management system (OAM) to obtain the KPI data reported by each base station at the tth moment and record it as a data set x i ∈R K . Where N is the number of elements in the data set X, (bs i ,bt i ,x i ,y i ) is the i-th element in X, i = 1, 2, ..., N. bs i and bt i Respectively represent the identifiers of the source base station and the target base station. i ∈R K , represents the K-dimensional key performance indicator KPI data of the target base station collected and reported by the source base station at a certain moment, including base station side data {P BS} and user-side data {P UEThe base station side data includes the target base station's longitude coordinates lat and latitude coordinates lon, the geographical distance dis from the source base station to the target base station, the target base station's antenna transmit power tp, and the target base station's load rate lr; the user side data includes the target base station's average reference signal received power rsrp measured by the source base station's serving users, the number of successful user handovers from the source base station to the target base station hos, and the handover success rate hor. i The definition is as follows:

[0068] x i =(lat,lon,dis,tp,lr,rsrp,hos,hor) (12)

[0069] y i is x i The label indicates the neighbor relationship between the target base station and the source base station, y i = +1 means the target base station is a neighbor of the source base station, y i =-1 indicates that the target base station is not a neighbor of the source base station.

[0070] Step 1.2: Preprocess wireless network KPI data.

[0071] The KPI data of consecutive T moments are averaged and normalized, where the value of T is determined by the OAM monitoring period, T = 50, and a sampling time is set to 1s. The geographical coordinates of the target base station, the geographical distance from the source base station to the target base station, and the radio frequency configuration parameters of the target base station are fixed values ​​that do not change over time. The geographical coordinates are normalized to a number in the range of [-1, 1] according to formula (13), and the geographical distance and radio frequency configuration parameters are normalized to a number in the range of [0, 1] according to formula (14), where x ik Indicates the number x in the i-th element of the dataset X i The kth parameter, k = 1, 2, ..., K, is the KPI parameter mentioned above, x ik ′ represents the normalized KPI parameter.

[0072]

[0073] The target base station load rate, user-side data including the target base station average RSRP measured by the source base station serving users, the number of successful user handovers from the source base station to the target base station, and the handover success rate are dynamic values ​​that change over time. First, the average value of T consecutive moments is calculated according to formula (15), where Indicates the average value of the kth KPI parameter.

[0074]

[0075] The base station load rate and the handover success rate are within the range of [0, 1]. The base station RSRP and the number of user handover successes are normalized according to formula (14).

[0076] Wireless network KPI datax i In its parameter x ik According to the above process, perform averaging or normalization respectively and use x i ″ represents that lat, lon, dis and tp are normalized according to formula (3) to obtain lat′, lon′, dis′ and tp′, and lr and hor are averaged according to formula (4) to obtain and First normalize rsrp and hos according to formula (3) and then calculate the average value according to formula (4) to obtain and x i "It can reflect the wireless network status of the past T consecutive moments, and the elements (bs i ,bt i ,x i ,y i ) Wireless network KPI data x i Replace with x i ″, get the data set

[0077] Step 2: Use historically collected wireless network KPI data to train the neighbor relationship classifier FDTC.

[0078] This step includes the following processes:

[0079] Step 2.1: Define FDTC. FDTC consists of two parts: fuzzy classification and precise classification. The FDT first performs fuzzy classification on dataset D, using a decision tree structure to segment it into sub-datasets with higher "purity." At the leaf nodes of the FDT, the SVM is trained using the sub-datasets to accurately classify wireless network neighbor relationships. The GINI index is used to measure dataset purity. For neighbor relationship classification, samples have only two labels: neighbor cell or non-neighbor cell. Assuming the probability of a sample label being a neighbor cell is p, the GINI index calculation formula for network KPI data D is as follows:

[0080] GINI(D)=2p(1-p) (16)

[0081] Step 2.2, initialize FDT.

[0082] Define the root node of FDT, store the data set D at the root node, and extract the K-dimensional wireless network KPI data x iThe kth KPI parameter x in ″ ik ″ is the kth feature a of the dataset D k , define the feature set A of the dataset D = {a1, a2, ..., a K}.

[0083] Step 2.3, FDT generates a decision.

[0084] Check the purity and node depth of the dataset at the current node. If the current node depth exceeds the depth threshold Depth_thr or the GINI index of the dataset at the current node is less than the set threshold GINI_thr, terminate the FDT generation process for the current node and jump to step 2.6. The FDT depth threshold Depth_thr limits the maximum height of the FDT tree structure and constrains the number of leaf node classifiers. The GINI index threshold GINI_thr limits the purity of the dataset at the current node and affects the training cost of the leaf node classifiers. Both Depth_thr and GINI_thr are determined by the scale of wireless network KPI data.

[0085] Step 2.4, calculate the optimal segmentation feature a of the dataset D * And the best split value x for this feature ** ″.

[0086] Randomly sample the data set D to obtain the data set D k , select a feature a k , respectively according to feature a k The values ​​of are reordered from large to small to obtain the data set D k ′, select the median element of the data as the initial split point x mk ″, according to x mk The dataset D is divided into two parts D left and D right According to formula (17), the information gain brought by the current segmentation point is calculated, which is also the feature a k Current maximum information gain Gain *k .

[0087]

[0088] To find the feature a k The best split point x on *k ″,x mk The value of ″ is based on D k ' is sorted towards larger and smaller values. mk After the value changes, the weights of the divided data sets are no longer equal. Assuming D left The data volume accounts for q, q<0.5, and the optimal information gain is updated according to formula (18).mk ″Exploration until Gain *k Until it stops increasing, record the x at this time mk ″ is the optimal split point x *k ″.

[0089] Gain *k =max(Gain *k ,qGINI(D left )+(1-q)GINI(D right )) (18)

[0090] Traverse the feature set A, for each feature a k Repeat this step and compare the k The best information gain extracted from the above is selected, and the x with the largest information gain is selected *k ″ is the best segmentation value x of the dataset D ** ″, its corresponding feature a k is the best segmentation feature a * .

[0091] Step 2.5, FDT according to a * and x ** Split the dataset D into D left and D right , generate a binary decision tree on the node node, and D left and D right Stored in the left and right subtree nodes respectively left and node right Repeat steps 2.3 to 2.5 for the left and right subtree nodes.

[0092] Step 2.6, delete the data set D on the current node node Base station index number (bs i ,bt i ), then the dataset The SVM classifier trained with Gaussian kernel function is stored in the current leaf node of FDT to complete the FDTC training.

[0093] Step 3: Input the real-time collected wireless network KPI data into FDTC to determine neighbor relationships.

[0094] Each KPI data x i The decision tree structure after FDTC is sent to the corresponding leaf node SVM classifier, and SVM predicts its classification label y i , if y i = +1, then the source base station and the target base station establish a neighbor relationship. If y i= -1, the source base station and the target base station are controlled not to establish a neighbor relationship, and the neighbor relationships between the base stations are summarized into a neighbor relations table (Neighbor Relations Table, NRT).

[0095] FDTC's classification results are verified using actual neighbor relationships automatically corrected during wireless network operation. The accuracy of FDTC's overall neighbor relationship judgment is calculated, as is the accuracy of the SVM classifier at each leaf node. An accuracy threshold, Accuracy_thr, is set to 0.95, determined by the expected accuracy of the wireless network's neighbor relationship assessment. If the overall FDTC accuracy is higher than Accuracy_thr, the FDTC structure remains unchanged, and NRT is continuously generated and maintained as the wireless network operates. If the overall FDTC accuracy is lower than Accuracy_thr or if the wireless network topology changes, the fourth step, FDTC training, is performed.

[0096] Step 4: Update and train FDTC using the newly collected wireless network KPI data.

[0097] This step includes the following processes:

[0098] Step 4.1: The newly collected wireless network KPI dataset D test As a test set, it is divided into sub-datasets through the FDT structure and stored in the corresponding leaf node positions.

[0099] Step 4.2: Prune the FDTC. Traverse all leaf nodes of the FDTC and remove SVM classifiers with accuracy below the threshold of 0.95. Use the test set stored in the corresponding node as the training set and store it in the corresponding leaf node. Examine the entire FDTC tree structure and further remove nodes where both the left and right subtrees have been pruned. Combine the training sets stored in both subtrees and store them at the previous node.

[0100] Step 4.3: FDTC update. At the pruned location, use the training set stored at the node location to retrain the FDTC according to the steps in step 2 to complete the FDTC.

[0101] Step 5: Based on the FDTC generated by training in the second step, the Boosting strategy is used to integrate the learning of FDTC, train the enhanced FDTC (BFDTC), and set the total number of base classifiers in BFTDC to L = 5.

[0102] This step includes the following processes:

[0103] Step 5.1: Define the weight W of the wireless network KPI dataset D l , W l =(w l1 ,w l2,...,w lN ),initialization

[0104]

[0105] Where N represents the number of elements in the data set D, w li represents the weight of the i-th element in the dataset D during training, l represents the base classifier label in BFDTC, l = 1, 2, ..., L.

[0106] Step 5.2: When training the first base classifier, follow the second step to train the base classifier FDTC l , use h l FDTC l Network KPI data x i The prediction result of ″ is expressed as h l (x i ″). Calculate the fitting error rate e according to formula (19) l , e l Indicates the current FDTC l The probability of misjudging the network neighbor relationship reflects the learning progress of BFDTC.

[0107]

[0108] Calculate the current FDTC according to formulas (20) and (21) l The weight α l And the sample weight W of the next round of training l+1 .

[0109]

[0110] Save FDTC l and α l .

[0111] In step 5.3, when l does not exceed the set total number of base classifiers L, repeat step 5.2 and construct the final neighbor relationship classifier BFDTC according to formula (22).

[0112]

[0113] Step 5.4, neighbor relationship prediction. The newly collected wireless network KPI dataset D test The test set is input into BFDTC, which directly outputs the classification label and generates the wireless network NRT.

[0114] Step 6: Based on the FDTC generated by training in the second step, the bagging strategy is used to ensemble learn the FDTC, train the random forest FDTC (RF-FDTC), and set the total number of base classifiers in RF-FTDC to L = 5.

[0115] This step includes the following processes:

[0116] Step 6.1: When training the first base classifier, randomly extract 70% of the data from the wireless network KPI dataset D and randomly extract 5 features from the feature set of D to form the dataset D. l , using D l According to the second step, train the base classifier FDTC l , save FDTC l .

[0117] Step 6.2: When l does not exceed the set total number of base classifiers L, repeat step 6.1 and save the L FDTCs. l Construct a parallel neighbor relation classifier RF-FDTC.

[0118] Step 6.3, neighbor relationship prediction. The newly collected wireless network KPI dataset D test As the test set input RF-FDTC, multiple base classifiers FDTC l The output is presented in the form of label probability, and the neighbor relationship is predicted based on the label probability to generate the wireless network NRT.

[0119] The specific embodiments described herein are merely illustrative of the spirit of the present invention. Persons skilled in the art may make various modifications, additions, or substitutions to the described specific embodiments without departing from the spirit of the present invention or exceeding the scope of the appended claims.

Claims

1. An automatic neighbor relationship method based on machine learning, characterized in that The method comprises the following steps: Step 1: Collect the key performance indicator (KPI) data of the wireless communication system and form a dataset D through preprocessing; Step 2: Utilizing historically collected wireless communication system key performance indicator (KPI) data, the neighbor relationship classifier (FDTC) is trained. First, a fuzzy decision tree (FDT) is generated through training to perform fuzzy classification on the KPI data. This divides the KPI dataset into multiple small sub-datasets with consistent features and stores them at corresponding FDT leaf nodes. Then, a support vector machine (SVM) with a Gaussian kernel function is trained at the FDT leaf nodes using the corresponding sub-datasets for accurate classification. This allows the neighbor relationship classifier (FDTC) to directly determine network neighbor relationships based on the input wireless network KPI data. Step 3: Input the real-time collected wireless communication system key performance indicators (KPIs) into the FDTC to perform neighbor relationship judgment. Real-time network KPI data is collected as a test set. After first being identified by the FDT, it is fed into the SVM classifier on the corresponding leaf node. The SVM classifier then directly outputs the neighbor relationship judgment result, determining whether the target cell is a neighbor cell of the source cell. Step 4: Update and train the FDTC using the newly collected wireless communication system key performance indicator (KPI) data. The newly collected wireless network KPI data set is divided into sub-datasets using the FDT structure and stored on the corresponding leaf nodes. The classification accuracy of the SVM classifier of each leaf node in the FDTC is calculated. Leaf nodes with classification accuracy below the accuracy threshold are deleted, i.e., pruning. The FDTC is then retrained at the pruned location using the sub-dataset stored on the corresponding leaf node, i.e., update training. The updated FDTC can better judge neighbor relationships based on the real-time collected wireless network KPI data. Step 5: Based on the FDTC trained in step 2, a serial boosting strategy is used to integrate and learn the FDTC. This is used to train the enhanced FDTC, or BFDTC. This method improves the accuracy of determining wireless network neighbor relationships by serially boosting the base classifier. Step 6: Based on the FDTC training in step 2, a parallel bagging strategy is used to ensemble learn the FDTC to train a random forest FDTC, namely RF-FDTC. The total number of base classifiers in RF-FDTC is set to L, where l represents the base classifier number, l = 1, 2, ..., L. By training multiple base classifiers in parallel, the accuracy of determining wireless network neighbor relationships is improved.

2. The automatic neighbor relationship method based on machine learning according to claim 1, characterized in that The first step is specifically: Step 1.1, record the key performance indicators (KPIs) of the wireless communication system; Use the wireless network operation and maintenance management system OAM to obtain the KPI data reported by each base station at the tth moment and record it as a data set Where N is the number of elements in the data set X, (bs i ,bt i ,x i ,y i ) is the i-th element in X, i = 1, 2, ..., N; bs i and bt i Respectively represent the identifiers of the source base station and the target base station; x i ∈R K Indicates the K-dimensional key performance indicator KPI data of the target base station collected and reported by the source base station at a certain time, including base station side data {P BS } and user-side data {P UE The base station side data includes the target base station longitude coordinates lat and latitude coordinates lon, the geographical distance dis from the source base station to the target base station, the target base station antenna transmission power tp, and the target base station load rate lr; the user side data includes the target base station average reference signal received power rsrp measured by the source base station serving users, the number of successful user handovers from the source base station to the target base station hos and the handover success rate hor, x i The definition is as follows: x i =(lat,lon,dis,tp,lr,rsrp,hos,hor) (1) y i is x i The label indicates the neighbor relationship between the target base station and the source base station, y i = +1 means the target base station is a neighbor of the source base station, y i =-1 indicates that the target base station is not a neighbor of the source base station; Step 1.2: Preprocessing of KPI data of wireless communication system: The KPI data of consecutive T moments are averaged and normalized, where the value of T is determined by the wireless network operation and maintenance management system OAM. The geographical coordinates of the target base station, the geographical distance from the source base station to the target base station, and the radio frequency configuration parameters of the target base station are fixed values ​​that do not change over time. The geographical coordinates are normalized to a number in the range of [-1, 1] according to formula (2), and the geographical distance and radio frequency configuration parameters are normalized to a number in the range of [0, 1] according to formula (3), where x ik Indicates the number x in the i-th element of the dataset X i The kth parameter, k = 1, 2, ..., K, is the KPI parameter mentioned above, x ik ' represents the normalized KPI parameter, The target base station load rate, user-side data including the target base station average reference signal received power (RSRP) measured by the source base station serving users, the number of successful user handovers from the source base station to the target base station, and the handover success rate are dynamic values ​​that change over time. First, the average value of T consecutive moments is calculated according to formula (4), where represents the average value of the kth KPI parameter, The range of base station load rate and handover success rate is within [0,1]. The base station reference signal received power RSRP and the number of user handover successes are normalized according to formula (3); Key performance indicator KPI data of wireless communication system x i In the parameter x ik According to the above process, perform averaging or normalization respectively and use x i ″ represents that lat, lon, dis and tp are normalized according to formula (3) to obtain lat′, lon′, dis′ and tp′, and lr and hor are averaged according to formula (4) to obtain and First normalize rsrp and hos according to formula (3) and then calculate the average value according to formula (4) to obtain and x i ″Reflects the wireless network status of the past T consecutive moments, and converts the elements (bs i ,bt i ,x i ,y i ) in wireless communication system key performance indicator KPI data x i Replace with x i ″, get the data set 3. The automatic neighbor relationship method based on machine learning according to claim 1, characterized in that The second step is as follows: Step 2.1: Define the fuzzy decision tree classifier (FDTC). FDTC consists of two parts: fuzzy classification and precise classification. The fuzzy decision tree (FDT) first performs fuzzy classification on dataset D, using a decision tree structure to segment it into sub-datasets with higher purity. At the leaf nodes of the FDT, the sub-datasets are used to train a support vector machine (SVM) to precisely classify wireless network neighbor relationships. The Gini index (GINI) is used to measure dataset purity. For neighbor relationship classification, samples have only two labels: neighbor cell or non-neighbor cell. Assuming the probability of a sample label being a neighbor cell is p, the Gini index (GINI) of KPI dataset D is calculated as follows: GINI(D)=2p(1-p) (5) Step 2.2, initialize FDT, Define the root node of FDT, store the data set D at the root node, and extract the K-dimensional wireless network KPI data x i The kth KPI parameter x in ″ ik ″ is the kth feature a of the dataset D k , define the feature set A of the dataset D = {a1, a2, ..., a K }; Step 2.3, FDT generates a decision, Check the "purity" and node depth of the dataset on the current node. If the current node depth is greater than the depth threshold Depth_thr or the GINI index of the dataset on the current node is less than the set threshold GINI_thr, terminate the FDT generation step for the current node and jump to step 2.

6. The FDT depth threshold Depth_thr limits the maximum height of the FDT tree structure and constrains the number of leaf node classifiers. The GINI index threshold GINI_thr limits the "purity" of the dataset on the current node and affects the training cost of the leaf node classifiers. Both Depth_thr and GINI_thr are determined by the scale of wireless network KPI data. Step 2.4, calculate the optimal segmentation feature a of the dataset D * And the best split value x for this feature ** ″, Randomly sample the data set D to obtain the data set D k , select a feature a k , respectively according to feature a k The values ​​of are reordered from large to small to obtain the data set D k ′, select the median element of the data as the initial split point x mk ″, according to x mk The dataset D is divided into two parts D left and D right ; According to formula (6), the information gain brought by the current segmentation point is calculated, which is also feature a k Current maximum information gain Gain *k ; To find the feature a k The best split point x on *k ″,x mk The value of ″ is based on D k ' is sorted towards larger and smaller values; x mk After the value changes, the weights of the divided data sets are no longer equal. Assuming D left The data volume accounts for q, q<0.5, and the optimal information gain is updated according to formula (7). mk ″Exploration until Gain *k Until it stops increasing, record the x at this time mk ″ is the optimal split point x *k ″; Gain *k =max(Gain *k ,qGINI(D left )+(1-q)GINI(D right )) (7) For each feature a in the feature set A k Repeat this step and compare the k The best information gain extracted from the above is selected, and the x with the largest information gain is selected *k ″ is the optimal segmentation value x of the dataset D ** ″, its corresponding feature a k is the best segmentation feature a * ; Step 2.5, FDT according to a * and x ** ″Split the dataset D into D left and D right , generate a binary decision tree on the node node, and D left and D right Stored in the left and right subtree nodes respectively left and node right Repeat steps 2.3 to 2.5 for the left and right subtree nodes; Step 2.6, delete the data set D on the current node node Base station index number (bs i ,bt i ), then the dataset The SVM classifier trained with Gaussian kernel function is stored in the current leaf node of FDT to complete the FDTC training.

4. The automatic neighbor relationship method based on machine learning according to claim 1, characterized in that The third step is specifically, Step 3.1, each KPI data x i The decision tree structure after FDTC is sent to the corresponding leaf node SVM classifier, and SVM predicts its classification label y i , if y i = +1, then the source base station and the target base station establish a neighbor relationship. If y i =-1, the source base station and the target base station are controlled not to establish a neighbor relationship, and the neighbor relationships between the base stations are summarized into a neighbor relationship table NRT; In step 3.2, the FDTC classification results are verified using the actual neighbor relationships automatically corrected during wireless network operation. The overall accuracy of FDTC's neighbor relationship judgment is calculated. Furthermore, the accuracy of the SVM classifier on each leaf node is calculated. An accuracy threshold, Accuracy_thr, is set. Its value is determined by the expected accuracy of the wireless network for neighbor relationships. If the overall FDTC accuracy is higher than Accuracy_thr, the FDTC structure remains unchanged, and NRT is continuously generated and maintained as the wireless network operates. If the overall FDTC accuracy is lower than Accuracy_thr or the wireless network topology changes, the fourth step is to update and train the FDTC.

5. The automatic neighbor relationship method based on machine learning according to claim 1, characterized in that The fourth step is as follows: Step 4.1: The newly collected wireless communication system key performance indicator KPI dataset D test As a test set, it is divided into sub-datasets through the FDT structure and stored in the corresponding leaf node positions; Step 4.2, FDTC pruning: traverse all leaf nodes of FDTC, delete SVM classifiers with accuracy lower than Accuracy_thr, use the test set stored in the corresponding node as the training set, store it in the corresponding leaf node, check the tree structure of the entire FDTC, further delete the tree nodes where both the left and right subtrees are pruned, and store the training set stored in the left and right subtrees at the node position of the previous level; Step 4.3, FDTC update, at the pruning position, use the training set stored at the node position to retrain FDTC according to the steps in the second step to complete FDTC.

6. The automatic neighbor relationship method based on machine learning according to claim 1, characterized in that The fifth step is as follows: Step 5.1: Define the weight W of the wireless communication system key performance indicator KPI dataset D l , W l =(w l1 ,w l2 ,...,w li ,...,w lN ), where w li Indicates the weight of the i-th element in the dataset D during training, l represents the base classifier number in the enhanced fuzzy decision tree classifier BFDTC, l = 1, 2, ..., L, initialization Where N represents the number of elements in the data set D; Step 5.2: When training the first base classifier, follow the second step to train the base classifier FDTC l , use h l FDTC l Network KPI data x i The prediction result of ″ is expressed as h l (x i ″); Calculate the fitting error rate e according to formula (8) l , e l Indicates the current FDTC l The probability of misjudging the network neighbor relationship reflects the learning progress of BFDTC. Calculate the current FDTC according to formulas (9) and (10): l The weight α l And the sample weight W of the next round of training l+1 , Save FDTC l and α l ; Step 5.3: When l does not exceed the set total number of base classifiers L, repeat step 5.2 and construct the final neighbor relationship classifier BFDTC according to formula (11). Step 5.4, neighbor relationship prediction: the newly collected wireless communication system key performance indicator KPI dataset D test The test set is input into BFDTC, which directly outputs the classification label and generates the wireless network NRT.

7. The automatic neighbor relationship method based on machine learning according to claim 1, characterized in that The sixth step is as follows: Step 6.1: When training the first base classifier, randomly extract some data from the wireless communication system key performance indicator KPI dataset D. The proportion of extracted data is determined by the size of the dataset. Randomly extract some features from the feature set of D to form the dataset D. l , using D l According to the second step, train the base classifier FDTC l , save FDTC l ; Step 6.2, when l does not exceed the set total number of base classifiers L, repeat step 6.1; save the L FDTC l Construct a parallel random forest fuzzy decision tree classifier RF-FDTC; Step 6.3, neighbor relationship prediction: the newly collected wireless communication system key performance indicator KPI dataset D test As the test set input RF-FDTC, multiple base classifiers FDTC l The output is presented in the form of label probability, and the neighbor relationship is predicted based on the label probability to generate the wireless network NRT.

Images