Abnormality detection method and system for B5G base station

Abstract

The invention discloses an anomaly detection method and system for a B5G base station. The method comprises the following steps: S1, acquiring an SS-RSRP data set based on a BBU network management system; step S2, a ContrastPatchAD model is constructed, and the ContrastPatchAD model is constructed; s3, the SS-RSRP data set is used for training the ContrastPatchAD model, and the trained ContrastPatchAD model is obtained, and the SS-RSRP data set is used for training the ContrastPatchAD model; and S4, inputting to-be-detected B5G base station data into the trained ContrastPatchAD model to obtain a detection result, and completing anomaly detection of the B5G base station. According to the ContrastPatchAD model disclosed by the invention, the accuracy of B5G base station anomaly detection is improved through a feature extraction method, a fragment embedding encoder method and a comparative learning method.

Description

Technical Field

[0001] The present invention relates to the technical field of base station anomaly detection, and particularly to an anomaly detection method and system for B5G base stations. Background Art

[0002] With the continuous breakthrough of 5G key technologies, the Beyond Fifth Generation (B5G) mobile communication system realizes ultra-high reliability and ultra-high-precision real-time communication through technologies such as ultra-massive antennas, terahertz communication, and network intelligence, and makes up for the deficiencies of 5G NR in intelligent network management, making B5G gradually become the main direction of the development of the new generation of mobile communication technologies.

[0003] With the development of the Internet and communication technologies, the B5G system has evolved into a complex network system, including various wired (such as metropolitan area network / wide area network) and wireless networks (such as 5G access network). Due to the introduction of a large number of cutting-edge new technologies, the complexity of base station equipment is high, and the failure frequency of B5G base stations has increased significantly, especially in complex network scenarios (such as high-speed movement, network congestion, and adjacent cell interference, etc.). And operation and maintenance is the key to ensuring the efficient and stable operation of complex, hybrid, and large B5G base stations. For the accurate and efficient operation and maintenance requirements of B5G base stations, the anomaly detection algorithm can quickly locate the abnormal data in the network and provide real-time fault warnings, thereby significantly shortening the fault response and repair time, improving the stability of the network and the user experience, promoting the process of operation and maintenance intelligence, and also providing strong technical support for the efficient operation and market response of B5G base stations.

[0004] The existing base station data anomaly detection algorithms are designed for 5G base stations. Facing the characteristics of lack of anomaly labels in B5G base station data and limited ability to express normal features, it is difficult to achieve high-precision anomaly detection for B5G base stations.

[0005] Previous research on anomaly detection algorithms mainly focused on using reconstruction-based unsupervised methods to solve this problem. These methods assume that the model has been perfectly trained on normal data and assign higher anomaly scores to abnormal data during the test phase. However, simply relying on reconstruction is not enough to improve the detection ability of the model and is difficult to meet the actual needs of high-precision anomaly detection for B5G base stations. In addition, the existing models have limited ability to express normal features and cannot model normal or abnormal data in a targeted manner. Therefore, it is necessary to establish an anomaly detection model according to the characteristics of lack of anomaly labels in B5G base station data to achieve high-precision anomaly detection for B5G base stations. Summary of the Invention

[0006] In order to solve the above technical problems, the present invention provides an anomaly detection method and system for B5G base stations, and the method includes:

[0007] Step S1: Collect the SS-RSRP dataset based on the BBU network management system;

[0008] Step S2: Construct the ContrastPatchAD model;

[0009] Step S3: Use the SS-RSRP dataset to train the ContrastPatchAD model to obtain a trained ContrastPatchAD model;

[0010] Step S4: Input the data of the to-be-detected B5G base station into the trained ContrastPatchAD model to obtain a detection result, and complete the anomaly detection of the B5G base station.

[0011] Optionally, in step S1, part of the BBU network management system adopts the TR069 protocol, docks with the base station equipment through the southbound interface, and the other part docks with the integrated network management system.

[0012] Optionally, in step S2, constructing the ContrastPatchAD model specifically includes:

[0013] Input the time series data into the feature extractor for slicing and then perform feature mapping to obtain a sliced mapping result;

[0014] Embed the sliced mapping result into the sliced embedding encoder to obtain a slice with an embedded query, define a query matrix based on the slice with the embedded query, aggregate the head features based on the query matrix, and perform a fully connected operation after normalizing the head features;

[0015] Adopt the contrast fusion method of contrastive learning, and based on the fully connected result, amplify the distribution difference between normal data and abnormal data.

[0016] Optionally, the process of adopting the method of contrastive learning and amplifying the distribution difference between normal data and abnormal data based on the fully connected result specifically includes:

[0017] Generate negative samples using Gaussian noise, denoted as X n = X + α·J, where J is sampled from the standard Gaussian distribution, and α is used to control the noise level. Integrate the denoising loss into the objective function:

[0018]

[0019] where, is the denoised negative sample;

[0020] Input the positive sample X into the feature extractor and the sliced embedding encoder to obtain the hidden feature representation N of the positive sample X, and use the projection head Obtain characterization:

[0021]

[0022] The present invention also discloses an anomaly detection system for a B5G base station, and the system includes:

[0023] A data acquisition module, configured to collect an SS-RSRP data set based on a BBU network management system;

[0024] A detection model construction module, configured to construct a ContrastPatchAD model;

[0025] A model training module, configured to use the SS-RSRP data set to train the ContrastPatchAD model to obtain a trained ContrastPatchAD model;

[0026] An anomaly detection module, configured to input the data of the to-be-detected B5G base station into the trained ContrastPatchAD model to obtain a detection result, and complete the anomaly detection of the B5G base station.

[0027] Optionally, in the data acquisition module, a part of the BBU network management system adopts the TR069 protocol, docks with base station devices through a southbound interface, and another part docks with an integrated network management system.

[0028] Optionally, in the detection model construction module, constructing the ContrastPatchAD model specifically includes:

[0029] Inputting time series data into a feature extractor for cutting and then performing feature mapping to obtain a segment mapping result;

[0030] Embedding the segment mapping result into a segment embedding encoder to obtain a segment with an embedded query, defining a query matrix based on the segment with the embedded query, aggregating head features based on the query matrix, and performing a fully connected operation after normalizing the head features;

[0031] Adopting a contrast fusion method of contrastive learning, and based on the fully connected result, amplifying the distribution difference between normal data and abnormal data.

[0032] Optionally, the process of adopting the method of contrastive learning and amplifying the distribution difference between normal data and abnormal data based on the fully connected result specifically includes:

[0033] Generating negative samples using Gaussian noise, denoted as X n = X + α·J, where J is sampled from a standard Gaussian distribution, and α is used to control the noise level, and integrating the denoising loss into the objective function:

[0034]

[0035] Among them, is the negative sample after denoising;

[0036] Input the positive sample X into the feature extractor and the segment embedding encoder to obtain the hidden feature representation N of the positive sample X, and use the projection head to obtain the representation:

[0037]

[0038] Compared with the prior art, the beneficial effects of the present invention are:

[0039] The advantage of the B5G base station anomaly detection method based on ContrastPatchAD of the present invention lies in high accuracy: different from the anomaly detection model that relies on data reconstruction in the traditional anomaly detection method, the B5G base station anomaly detection device based on ContrastPatchAD of the present invention learns data patterns through feature extraction, segment embedding encoding and contrast fusion, improves the model's ability to identify abnormal data, amplifies the distribution difference between normal data and abnormal data, reduces the influence of abnormal data labels on the anomaly detection accuracy of B5G base station data, and realizes the high-precision anomaly detection ability of B5G base stations. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the technical solutions of the present invention, the accompanying drawings required for use in the embodiments will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other accompanying drawings can be obtained based on these drawings without creative efforts.

[0041] Figure 1 It is the method step diagram of the anomaly detection method for B5G base stations according to the embodiment of the present invention;

[0042] Figure 2 It is the schematic diagram of the ContrastPatchAD model according to the embodiment of the present invention;

[0043] Figure 3 It is the anomaly detection result diagram of the ContrastPatchAD model for the SS-RSRP dataset according to the embodiment of the present invention. Among them, Figure 3 (a) is the prediction result of the ContrastPatchAD model, Figure 3 (b) is the anomaly score of the ContrastPatchAD model. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0044] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0045] To make the above objects, features, and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0046] The following explains the proprietary terms used in the present invention:

[0047] SS-RSRP (Secondary Synchronization Signal Received Signal Power) data is an index commonly used in the industry to test the quality and coverage of wireless signals. It is a measure of the received signal power of the reference signal transmitted by the base station, which reflects the signal strength received by the user equipment.

[0048] Embodiment 1

[0049] An anomaly detection method for a B5G base station, as Figure 1 shown, the method includes:

[0050] Step S1: Collect the SS-RSRP data set based on the BBU network management system.

[0051] Create an SS-RSRP data set. The SS-RSRP data set is a real index data set collected based on the BBU network management system of a certain operator. This network management system consists of two parts: one part is docked with the device through the southbound interface and uses the TR069 protocol; the other part is docked with the comprehensive network management systems of the provincial company and the group for reporting resources, alarms, and performance data.

[0052] Step S2: Construct a ContrastPatchAD model.

[0053] As Figure 2 shown, S21. First, input the original time series data X into the feature extractor and perform positional encoding so that the ContrastPatchAD model can actively learn the dependencies between data. Divide X' into N segments with a length of and rearrange them into VE is a simple linear transformation that maps all segments to a unified space The processing performed by the feature extractor can be described as follows:

[0054] X′ = PE(IN(X))

[0055] N = VE(Patching(X))

[0056] The feature extractor enables the ContrastPatchAD model to process data within a longer time window.

[0057] S22. To enhance the memory ability of ContrastPatchAD for normal samples and improve its generalization ability, this section incorporates the segments with embedded queries into the ContrastPatchAD model. Here, V represents the number of segments embedded in ContrastPatchAD, and D represents the dimension. After segment embedding, a linear projection is adopted to obtain where N is much smaller than V to reduce the dimension and enable the ContrastPatchAD model to effectively learn relevant information from normal samples rather than relying solely on prior knowledge.

[0058] In addition, the query matrix is redefined as the key matrix as and the value matrix as where k = {1, 2,..., U}, d is defined as

[0059] The attention values between different segments in each layer are:

[0060]

[0061] At the end of each layer, a linear layer is used to aggregate the features from different heads to obtain and then the head features are normalized to generate the input N(i + 1) for the next layer:

[0062] N (i)′ = LayerNorm(N (i) + Z0 (i) )

[0063] N (i+1) = LayerNorm(N (i)′ + FFN(N (i)′ ))

[0064] In the formula, FFN(·) represents a multi-layer feed-forward network. Finally, a simple fully connected layer is used to restore Z (L) from the previous layer to It is the output result of the model, that is, the prediction result of the original input time series data X. During the training phase, the Mean Square Error (MSE) is used to measure the difference between the original time series data X and and the segment-based similarity loss is used to ensure the continuity between segments. During the testing phase, the ContrastPatchAD model uses to calculate the anomaly score and detect anomalies.

[0065]

[0066] S23. Although the segment embedding encoder enables the ContrastPatchAD model to memorize normal samples, it does not significantly enhance the difference between normal samples and abnormal samples. To solve this problem, this embodiment develops a contrast fusion module using contrastive learning to increase the difference between normal and abnormal data features, thereby improving the detection accuracy.

[0067] To generate negative samples for contrastive learning without introducing too much prior knowledge, a method of generating negative samples using Gaussian noise is adopted, that is, X n = X + α·J, where J is sampled from the standard Gaussian distribution, α is used to control the noise level, and the denoising loss is integrated into the final objective function.

[0068]

[0069] To efficiently and quickly learn invariant features, the positive sample X and the negative sample X n are simultaneously input into the feature extractor and the segment embedding encoder to obtain the hidden feature representations N and N n . Then, the projection head is used to obtain the representation:

[0070]

[0071] Then, this embodiment respectively derives the low-dimensional hidden representations H and H n of N and N n . In addition, to design an asymmetric feature contrast loss, gradient stopping is adopted.

[0072]

[0073] ContrastPatchAD seeks to optimize the reconstruction of positive samples, denoise noisy samples, and maximize the feature difference between positive and negative samples. In the initial stage of training, the model focuses mainly on reconstruction and denoising, and the contrast loss introduces additional training complexity. Therefore, in the initial Nwarm-up iteration, the weight of the contrast loss is gradually increased with each iteration until it reaches β max The maximum value of:

[0074]

[0075] The final objective function combines reconstruction loss, denoising loss and contrast loss. It can be expressed as:

[0076]

[0077] like Figure 3 As shown, step S3, using the SS-RSRP data set to train the ContrastPatchAD model to obtain a trained ContrastPatchAD model.

[0078] Figure 3 (a) is the prediction result of the ContrastPatchAD model. The black dotted line is the true value of SS-RSRP, and the red solid line is the predicted value of the SS-RSRP data set. It can be seen from the figure that the model can well predict the data trend of the SS-RSRP data set. Figure 3 (b) is the anomaly score of the ContrastPatchAD model. It can be observed that the anomaly score is higher where the prediction results are more different, and the anomaly occurrence point can be accurately detected.

[0079] Step S4: Input the data of the B5G base station to be detected into the trained ContrastPatchAD model to obtain the detection results, thereby completing the anomaly detection of the B5G base station.

[0080] In order to accurately observe the accuracy of the anomaly detection model proposed in this embodiment, the ContrastPatchAD model is used to compare and verify the accuracy with multiple baseline models, as shown in Tables 1 and 2.

[0081] Table 1

[0082]

[0083] Table 2

[0084]

[0085] In Table 1, ContrastPatchAD demonstrates good anomaly detection capabilities on 4 public datasets. In terms of precision performance, the precision of each dataset is above 0.9, and the highest precision of 0.99 is achieved on the UCR dataset. The recall rate of the model on six datasets exceeds 0.90, significantly outperforming the baseline model. The AUC value of each dataset is around 0.9, especially the recall rate on the SS-RSRP dataset is as high as 0.93. For the F1 value, the F1 score of the ContrastPatchAD model on each dataset is above 0.9. Therefore, the ContrastPatchAD model is not only applicable to the anomaly detection of B5G base stations but can also be applied to the anomaly detection in other fields.

[0086] In Table 2, the anomaly detection effects of ContrastPatchAD and comparison models such as Random, USAD, TCN-ED, Anom Trans, DCdetector, NPSR, and DualAttenAD on the B5G base station dataset are given. In the SS-RSRP dataset, the F1 score of the ContrastPatchAD model is 0.98, which is better than all comparison models. Compared with the state-of-the-art baseline model, the F1 score of the ContrastPatchAD model has increased by 6.53%. In the SS-SINR dataset, the average F1 score of the ContrastPatchAD model is 0.96, which is better than all comparison models. Compared with the state-of-the-art baseline model, the F1 score of the ContrastPatchAD model has increased by 6.67%.

[0087] Use accuracy, recall, F1 score, and area under the ROC curve (AUC) to evaluate the anomaly detection accuracy.

[0088]

[0089] TPR (True Positive Rate): True positive rate, representing the proportion of correctly identified anomaly samples.

[0090] FPR (False Positive Rate): False positive rate, representing the proportion of normal samples misjudged as anomalies.

[0091] Example 2

[0092] The data acquisition module is used to collect the SS-RSRP dataset based on the BBU network management system;

[0093] Create an SS-RSRP dataset. The SS-RSRP dataset is a real metric dataset collected based on the BBU network management system of a certain operator. The network management system consists of two parts: one part is docked with the device through the southbound interface and uses the TR069 protocol; the other part is docked with the comprehensive network management systems of the provincial company and the group for reporting resource, alarm, and performance data.

[0094] The detection model construction module is used to construct the ContrastPatchAD model.

[0095] First, input the original time series data X into the feature extractor and perform positional encoding so that the ContrastPatchAD model can actively learn the dependencies between data. Divide X' into N segments of length and rearrange them into VE is a simple linear transformation that maps all segments to a unified space The processing performed by the feature extractor can be described as follows:

[0096] X′ = PE(IN(X))

[0097] N = VE(Patching(x))

[0098] The feature extractor enables the ContrastPatchAD model to process data within a longer time window.

[0099] To enhance the memory ability of ContrastPatchAD for normal samples and improve its generalization ability, in this section, the segments with embedded queries are merged into the ContrastPatchAD model. Among them, V represents the number of embedded segments in ContrastPatchAD. After segment embedding, linear projection is adopted to obtain where N is much smaller than V to reduce the dimension and enable the ContrastPatchAD model to effectively learn relevant information from normal samples rather than relying solely on prior knowledge.

[0100] In addition, redefine the query matrix as the key matrix as the value matrix as where k = {1, 2,..., U}, d is defined as

[0101] The attention values between different segments in each layer are:

[0102]

[0103] At the end of each layer, a linear layer is used to aggregate features from different heads to obtain and then normalize the head features to generate the input N(i + 1) for the next layer:

[0104] N (i)′ = LayerNorm(N (i) + Z0 (i) )

[0105] N (i+1) = LayerNorm(N (i)′ + FFNN((N (i) ))

[0106] where FFN(·) represents a multi - layer feed - forward network. Finally, a simple fully - connected layer is used to restore Z (L) from the previous layer to In the training stage, the Mean Square Error (MSE) is used to measure the difference between X and and a fragment - based similarity loss is used to ensure the continuity between fragments. In the testing stage, the ContrastPatchAD model uses to calculate the anomaly score and detect anomalies.

[0107]

[0108] Although the fragment embedding encoder enables the ContrastPatchAD model to memorize normal samples, it does not significantly enhance the difference between normal and abnormal samples. To solve this problem, this embodiment develops a contrast fusion module using contrastive learning to increase the difference between normal and abnormal data features, thereby improving the detection accuracy.

[0109] To generate negative samples for contrastive learning without introducing too much prior knowledge, a method of generating negative samples using Gaussian noise is adopted, that is, X n = X + α·J, where J is sampled from a standard Gaussian distribution, α is used to control the noise level, and the denoising loss is integrated into the final objective function.

[0110]

[0111] To learn invariant features efficiently and quickly, the positive sample X and the negative sample X n are simultaneously input into the feature extractor and the fragment embedding encoder to obtain the hidden feature representations N and N n . Then the projection head is used to obtain the characterization:

[0112]

[0113] Then, in this embodiment, the low-dimensional hidden representations H and Hn of N and Nn are respectively derived. Additionally, to design an asymmetric feature contrast loss, gradient stopping is used.

[0114]

[0115] ContrastPatchAD aims to optimize the reconstruction of positive samples, denoise noisy samples, and maximize the feature difference between positive and negative samples. In the initial stage of training, the model mainly focuses on reconstruction and denoising, and the contrast loss introduces additional training complexity. Therefore, in the initial N warm-up iterations, the weight of the contrast loss gradually increases with each iteration until it reaches the maximum value of β max :

[0116]

[0117] The final objective function combines the reconstruction loss, denoising loss, and contrast loss. It can be expressed as:

[0118]

[0119] A model training module for training the ContrastPatchAD model using the SS-RSRP dataset to obtain a trained ContrastPatchAD model;

[0120] An anomaly detection module for inputting the B5G base station data to be detected into the trained ContrastPatchAD model to obtain a detection result and complete the anomaly detection of the B5G base station.

[0121] To accurately observe the accuracy of the anomaly detection model proposed in this embodiment, the ContrastPatchAD model is compared with multiple baseline models to verify the accuracy.

[0122] The accuracy of anomaly detection is evaluated using accuracy, recall, F1-score, and the area under the ROC curve (AUC).

[0123]

[0124] TPR (True Positive Rate): The true positive rate, representing the proportion of correctly identified abnormal samples.

[0125] FPR (False Positive Rate): The false positive rate, representing the proportion of normal samples misjudged as abnormal.

[0126] The embodiments described above are only descriptions of the preferred embodiments of the present invention, and do not limit the scope of the present invention. Without departing from the design spirit of the present invention, various deformations and improvements made by those of ordinary skill in the art to the technical solutions of the present invention shall fall within the protection scope determined by the claims of the present invention.

Claims

1. An anomaly detection method for a B5G base station, characterized in that, The method includes: Step S1: Collect the SS-RSRP data set based on the BBU network management system; Step S2: Construct the ContrastPatchAD model; Step S3: Use the SS-RSRP data set to train the ContrastPatchAD model to obtain a trained ContrastPatchAD model; Step S4: Input the data of the to-be-detected B5G base station into the trained ContrastPatchAD model to obtain a detection result, and complete the anomaly detection of the B5G base station.

2. The anomaly detection method for a B5G base station according to claim 1, characterized in that, In the step S1, a part of the BBU network management system adopts the TR069 protocol and docks with the base station device through the southbound interface, and the other part docks with the integrated network management system.

3. The anomaly detection method for a B5G base station according to claim 1, wherein, In the step S2, constructing the ContrastPatchAD model specifically includes: Input the time series data into the feature extractor for slicing and then perform feature mapping to obtain a sliced mapping result; Embed the sliced mapping result into the sliced embedding encoder to obtain a slice with an embedded query, define a query matrix based on the slice with the embedded query, aggregate the head features based on the query matrix, and perform a fully connected operation after normalizing the head features; Adopt the contrast fusion method of contrastive learning, and based on the fully connected result, amplify the distribution difference between normal data and abnormal data.

4. The anomaly detection method for a B5G base station according to claim 3, wherein The process of adopting the method of contrastive learning and amplifying the distribution difference between normal data and abnormal data based on the fully connected result specifically includes: Generate negative samples using Gaussian noise, denoted as X n = X + α·J, where J is sampled from the standard Gaussian distribution and α is used to control the noise level. Integrate the denoising loss into the objective function: Among them, is the negative sample after denoising; Input the positive sample X into the feature extractor and the segment embedding encoder to obtain the hidden feature representation N of the positive sample X, and use the projection head to obtain the characterization:

5. An anomaly detection system for a B5G base station, the system being used to implement the anomaly detection method described in any one of claims 1-4, characterized in that, The system includes: A data acquisition module for collecting the SS-RSRP data set based on the BBU network management system; A detection model construction module for constructing the ContrastPatchAD model; A model training module for using the SS-RSRP data set to train the ContrastPatchAD model to obtain a trained ContrastPatchAD model; An anomaly detection module for inputting the data of the to-be-detected B5G base station into the trained ContrastPatchAD model to obtain a detection result, and complete the anomaly detection of the B5G base station.

6. The anomaly detection system for a B5G base station according to claim 5, wherein, In the data acquisition module, a part of the BBU network management system adopts the TR069 protocol and docks with the base station device through the southbound interface, and the other part docks with the integrated network management system.

7. The anomaly detection system for a B5G base station according to claim 5, characterized in that In the detection model construction module, constructing the ContrastPatchAD model specifically includes: Input the time series data into the feature extractor for slicing and then perform feature mapping to obtain a sliced mapping result; Embed the sliced mapping result into the sliced embedding encoder to obtain a slice with an embedded query, define a query matrix based on the slice with the embedded query, aggregate the head features based on the query matrix, and perform a fully connected operation after normalizing the head features; Adopt the contrast fusion method of contrastive learning, and based on the fully connected result, amplify the distribution difference between normal data and abnormal data.

8. The anomaly detection system for a B5G base station according to claim 7, wherein, The process of adopting the method of contrastive learning and amplifying the distribution difference between normal data and abnormal data based on the fully connected result specifically includes: Negative samples are generated using Gaussian noise, denoted as X n = X + α·J, where J is sampled from the standard Gaussian distribution and α is used to control the noise level. The denoising loss is integrated into the objective function: Among them, is the negative sample after denoising; Input the positive sample X into the feature extractor and the segment embedding encoder to obtain the hidden feature representation N of the positive sample X, and use the projection head to obtain the characterization:

Images