Data relationship identification method and device, equipment and storage medium

Abstract

A data relationship identification method, device and equipment and a storage medium are disclosed. The method comprises: obtaining to-be-tested data; the to-be-tested data comprises first sequence data of multiple dimensions within a first time period and second sequence data of a time dimension within a second time period; inputting the to-be-tested data into a pre-trained relationship representation model to output a relationship representation vector of the to-be-tested data through the relationship representation model; the relationship representation vector is used to represent the relationship between the first sequence data of different dimensions; the relationship representation model is obtained by training a prediction model based on training data; the training data comprises third sequence data of multiple dimensions within a third time period and fourth sequence data of a time dimension within a fourth time period.

Description

Technical Field

[0001] This invention relates to the field of wireless communication technology, and in particular to a data relationship identification method, apparatus, device, and storage medium. Background Technology

[0002] Currently, in intelligent network operation and maintenance scenarios, different scenarios require different algorithm models. However, for multiple tasks in the existing network, such as time series prediction, root cause localization, and network health analysis, independent model development is carried out. In some model development, it is necessary to obtain various common data, such as key performance indicators (KPIs), measurement reports (MRs), and alarm data. In each model, the underlying principle is to obtain various data relationship expressions for each network element in different ways. Because each independent model needs to be developed, the efficiency of representing the relationships between various data is not high. Summary of the Invention

[0003] In view of this, embodiments of the present invention aim to provide a data relationship identification method, apparatus, device, and storage medium.

[0004] The technical solution of this invention is implemented as follows:

[0005] This invention provides a data relationship identification method, the method comprising:

[0006] Acquire the data to be tested; the data to be tested includes first sequence data of multiple dimensions within a first time period and second sequence data of time dimension within a second time period;

[0007] The test data is input into a pre-trained relation representation model to output a relation representation vector of the test data; the relation representation vector is used to represent the relationship between first sequence data in different dimensions.

[0008] The relationship representation model is obtained by training the prediction model based on training data; the training data includes third sequence data of multiple dimensions in the third time period and fourth sequence data of time dimension in the fourth time period.

[0009] Furthermore, according to at least one embodiment of the present invention, the relation representation model includes an encoder and a decoder;

[0010] The step of inputting the test data into a pre-trained relation representation model to output a relation representation vector of the test data through the relation representation model includes:

[0011] The first sequence data of multiple dimensions within a first time period is used as the first input data; the encoder encodes the first input data to obtain a first relation representation vector; the first relation representation vector represents the first relationship between the first sequence data of different dimensions within the first time period.

[0012] Using the first relationship between first sequence data of different dimensions within a first time period and the second sequence data of time dimension within a second time period, second input data is obtained; the second input data is decoded by the decoder to obtain a second relationship representation vector; the second relationship representation vector represents the first relationship between first sequence data of different dimensions within the first time period and the second relationship between second sequence data of time dimension within the second time period.

[0013] Furthermore, according to at least one embodiment of the present invention, the method further includes:

[0014] Obtain the training data;

[0015] Based on the training data, the prediction model is trained to obtain the relationship representation model.

[0016] Furthermore, according to at least one embodiment of the present invention, training the prediction model based on the training data to obtain the relation representation model includes:

[0017] The third sequence data of multiple dimensions within the third time period is used as the third input data; the third input data is encoded by the encoder in the initial prediction model to obtain the third relation representation vector; the third relation representation vector represents the third relation between the third sequence data of different dimensions within the third time period.

[0018] Using the third relationship between the third sequence data of different dimensions in the third time period and the fourth sequence data of the time dimension in the fourth time period, fourth input data is obtained; the fourth input data is decoded by the decoder in the initial prediction model to obtain the fourth relationship representation vector; the fourth relationship representation vector represents the third relationship between the third sequence data of different dimensions in the third time period and the fourth relationship between the fourth sequence data of the time dimension in the fourth time period.

[0019] Based on the third sequence data of multiple dimensions within the third time period, a first loss value is determined; based on the fourth input data, a second loss value is determined.

[0020] The first loss value and the second loss value are summed to obtain the third loss value;

[0021] The parameters of the prediction model are updated using the third loss value to obtain the relationship representation model.

[0022] Furthermore, according to at least one embodiment of the present invention, determining the first loss value based on the third sequence data of multiple dimensions within the third time period includes:

[0023] Based on the third sequence data of multiple dimensions within the third time period, a first sequence, a second sequence, a third sequence, and a fourth sequence are obtained; the first sequence represents the first indicator data; the second sequence represents the second indicator data, alarm data, and resource data; the third sequence represents parameter data; and the fourth sequence represents the influencing factors.

[0024] Randomly mask a portion of the data in at least one of the first sequence, the second sequence, the third sequence, and the fourth sequence;

[0025] The randomly obscured data is restored to obtain the restored data.

[0026] The first root mean square error is calculated using the recovered data and the randomly masked data, and the first root mean square error is used as the first loss value.

[0027] Furthermore, according to at least one embodiment of the present invention, determining the second loss value based on the fourth input data includes:

[0028] The fourth input data is used to predict the first indicator data within the fourth time period;

[0029] The second root mean square error is calculated using the predicted first indicator data for the fourth time period and the actual first indicator data.

[0030] The second root mean square error is used as the second loss value.

[0031] At least one embodiment of the present invention provides a data relationship identification device, comprising:

[0032] An acquisition unit is used to acquire data to be tested; the data to be tested includes first sequence data of multiple dimensions within a first time period and second sequence data of time dimension within a second time period.

[0033] The processing unit is configured to input the test data into a pre-trained relational representation model, and output a relational representation vector of the test data through the relational representation model; the relational representation vector is used to represent the relationship between first sequence data in different dimensions;

[0034] The relationship representation model is obtained by training the prediction model based on training data; the training data includes third sequence data of multiple dimensions in the third time period and fourth sequence data of time dimension in the fourth time period.

[0035] At least one embodiment of the present invention provides a data relationship identification device, comprising:

[0036] A communication interface is used to acquire the data to be tested; the data to be tested includes first sequence data with multiple dimensions within a first time period and second sequence data with a time dimension within a second time period.

[0037] A processor is configured to input the test data into a pre-trained relational representation model, and output a relational representation vector of the test data through the relational representation model; the relational representation vector is used to represent the relationship between first sequence data in different dimensions;

[0038] The relationship representation model is obtained by training the prediction model based on training data; the training data includes third sequence data of multiple dimensions in the third time period and fourth sequence data of time dimension in the fourth time period.

[0039] At least one embodiment of the present invention provides a network device, including a processor and a memory for storing a computer program capable of running on the processor.

[0040] When the processor runs the computer program, it executes the steps of any of the methods described above on the network device side.

[0041] At least one embodiment of the present invention provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the above methods.

[0042] This invention provides a data relationship identification method, apparatus, device, and storage medium to acquire test data. The test data includes first sequence data of multiple dimensions within a first time period and second sequence data of a time dimension within a second time period. The test data is input into a pre-trained relationship representation model to output a relationship representation vector of the test data. The relationship representation vector is used to represent the relationship between first sequence data of different dimensions. The relationship representation model is obtained by training a prediction model based on training data. The training data includes third sequence data of multiple dimensions within a third time period and fourth sequence data of a time dimension within a fourth time period. Using the technical solution provided by this invention, based on test data of multiple dimensions in the existing network, such as KPI data, alarm data, MR data, engineering parameter data, neighbor cell relationship data, etc., combined with a pre-trained relationship representation model, a comprehensive representation relationship between individual data of different dimensions is learned. Since a general relationship representation model can be used to express the data relationship between various data, the efficiency of representing relationships between various data is improved. Attached Figure Description

[0043] Figure 1 This is a schematic diagram illustrating the implementation process of the data relationship identification method according to an embodiment of the present invention;

[0044] Figure 2 A schematic diagram illustrating the specific implementation process of the data relationship identification method according to an embodiment of the present invention;

[0045] Figure 3 This is a schematic diagram of the Encoder network according to an embodiment of the present invention;

[0046] Figure 4 This is a schematic diagram of the Decoder network according to an embodiment of the present invention;

[0047] Figure 5 This is a schematic diagram of the prediction model in an embodiment of the present invention;

[0048] Figure 6 This is a schematic diagram of the composition structure of the data relationship recognition device according to an embodiment of the present invention;

[0049] Figure 7 This is a schematic diagram of the composition structure of a network device according to an embodiment of the present invention. Detailed Implementation

[0050] Before introducing the technical solutions of the embodiments of the present invention, the relevant technologies will be explained first.

[0051] The inventors discovered that in intelligent network operation and maintenance scenarios, different scenarios require different algorithm models. However, considering the multiple tasks in the existing network, such as time series prediction, root cause localization, and network health analysis, each model development requires the acquisition of various common data, such as KPI data, MR data, and alarm data. Furthermore, in each model, the underlying data relationships of each network element are obtained in different ways. Therefore, they considered using a universal model to aggregate multiple data types, complete the representation of network element data relationships, and then support downstream tasks.

[0052] However, related technologies do not base their approach on a holistic representation of network elements; instead, each task has its own separate model. Let's take a few typical tasks as examples:

[0053] First, there's the issue of identifying wireless network disturbance work orders. This typically involves acquiring the KPI and MR data of the network elements involved in the work order, extracting N-dimensional mathematical features from the time-series data, establishing a classification model, and learning the differences between disturbance and non-disturbance work orders.

[0054] Second, address the root cause of wireless network issues. Obtain KPI data, MR data, alarm data, and neighbor cell data for the network elements involved in the work order, learn the characteristics of multiple dimensions of network element data, thereby identifying the true cause of the anomaly and then classifying the problem.

[0055] Third, the time-series prediction problem. This involves acquiring multi-dimensional metrics of the wireless network, including KPI data and MR data, fitting the characteristics of historical data, and then inferring data for subsequent time periods.

[0056] The aforementioned issues are all addressed using independent models in current network implementations. Once the requirements are defined, a complete solution design needs to be completed independently, including requirements analysis, data collection, data modeling, and finally deployment as an independent capability or application.

[0057] In other words, the main problems existing in the current data relation representation include:

[0058] First, model training is isolated and the development cycle is long.

[0059] Specifically, currently, different application scenarios are developed by different teams, requiring a significant investment of manpower. Each project involves data collection, data understanding, and data processing, resulting in a long development cycle. While the essence of these projects is to learn the relationships between data, they are currently treated as completely separate projects, designing different isolated models.

[0060] Second, there is no focus on the processing and classification of data.

[0061] Specifically, in the current scenario, the model treats all types of data as having the same impact on the model, failing to highlight the degree of influence of different types of data on the relationships between each network element in the real scenario. Therefore, it may overlook key indicators and features that need to be focused on, resulting in poor model performance.

[0062] Third, the current model has limited representation of network elements, resulting in limited model performance.

[0063] Specifically, current models for multiple tasks mainly rely on machine learning algorithms to extract features from multidimensional data. However, due to factors such as the selected monitoring indicators, the subjective awareness of algorithm engineers, and operational efficiency, it is difficult to extract all features.

[0064] For example, when inputting time-series data into a model, the time-series features must first be transformed into 1×N-dimensional feature data. At this point, features such as mean and variance need to be extracted. These features are only one type of complex relationship and are insufficient to represent the complex network element features, resulting in poor model performance.

[0065] Based on this, in this embodiment of the invention, test data is acquired; the test data includes first sequence data of multiple dimensions within a first time period and second sequence data of time dimension within a second time period; the test data is input into a pre-trained relation representation model to output a relation representation vector of the test data; the relation representation vector is used to represent the relationship between first sequence data of different dimensions; wherein, the relation representation model is obtained by training a prediction model based on training data; the training data includes third sequence data of multiple dimensions within a third time period and fourth sequence data of time dimension within a fourth time period.

[0066] Figure 1 This is a schematic diagram illustrating the implementation process of the data relationship identification method according to an embodiment of the present invention, as shown below. Figure 1 As shown, the method includes steps 101 to 102:

[0067] Step 101: Obtain the data to be tested; the data to be tested includes first sequence data of multiple dimensions within a first time period and second sequence data of the time dimension within a second time period;

[0068] Step 102: Input the test data into a pre-trained relation representation model to output a relation representation vector of the test data; the relation representation vector is used to represent the relationship between the first sequence data in different dimensions;

[0069] The relationship representation model is obtained by training the prediction model based on training data; the training data includes third sequence data of multiple dimensions in the third time period and fourth sequence data of time dimension in the fourth time period.

[0070] In one embodiment, the relation representation model includes an encoder and a decoder;

[0071] The step of inputting the test data into a pre-trained relation representation model to output a relation representation vector of the test data through the relation representation model includes:

[0072] The first sequence data of multiple dimensions within a first time period is used as the first input data; the encoder encodes the first input data to obtain a first relation representation vector; the first relation representation vector represents the first relationship between the first sequence data of different dimensions within the first time period.

[0073] Using the first relationship between first sequence data of different dimensions within a first time period and the second sequence data of time dimension within a second time period, second input data is obtained; the second input data is decoded by the decoder to obtain a second relationship representation vector; the second relationship representation vector represents the first relationship between first sequence data of different dimensions within the first time period and the second relationship between second sequence data of time dimension within the second time period.

[0074] Here, the first relationship can also be called a static relationship, and the second relationship can also be called a dynamic relationship.

[0075] In one embodiment, the method further includes:

[0076] Obtain the training data;

[0077] Based on the training data, the prediction model is trained to obtain the relationship representation model.

[0078] In one embodiment, training the prediction model based on the training data to obtain the relation representation model includes:

[0079] The third sequence data of multiple dimensions within the third time period is used as the third input data; the third input data is encoded by the encoder in the initial prediction model to obtain the third relation representation vector; the third relation representation vector represents the third relation between the third sequence data of different dimensions within the third time period.

[0080] Using the third relationship between the third sequence data of different dimensions in the third time period and the fourth sequence data of the time dimension in the fourth time period, fourth input data is obtained; the fourth input data is decoded by the decoder in the initial prediction model to obtain the fourth relationship representation vector; the fourth relationship representation vector represents the third relationship between the third sequence data of different dimensions in the third time period and the fourth relationship between the fourth sequence data of the time dimension in the fourth time period.

[0081] Based on the third sequence data of multiple dimensions within the third time period, a first loss value is determined; based on the fourth input data, a second loss value is determined.

[0082] The first loss value and the second loss value are summed to obtain the third loss value;

[0083] The parameters of the prediction model are updated using the third loss value to obtain the relationship representation model.

[0084] In one embodiment, determining the first loss value based on the third sequence data of multiple dimensions within the third time period includes:

[0085] Based on the third sequence data of multiple dimensions within the third time period, a first sequence, a second sequence, a third sequence, and a fourth sequence are obtained; the first sequence represents the first indicator data; the second sequence represents the second indicator data, alarm data, and resource data; the third sequence represents parameter data; and the fourth sequence represents the influencing factors.

[0086] Randomly mask a portion of the data in at least one of the first sequence, the second sequence, the third sequence, and the fourth sequence;

[0087] The randomly obscured data is restored to obtain the restored data.

[0088] The first root mean square error is calculated using the recovered data and the randomly masked data, and the first root mean square error is used as the first loss value.

[0089] Here, during the data access process, various types of primary data from the wireless network operation and maintenance process can be collected. During the data classification process, the collected primary data of various types can first be classified and converted into corresponding sequence data, resulting in multi-dimensional third sequence data within the third time period, which is subsequently used for relation learning. During the data representation process, the multi-dimensional third sequence data within the third time period is represented to obtain a first sequence, a second sequence, a third sequence, and a fourth sequence.

[0090] The various types of initial data collected here include at least:

[0091] Wireless cell KPI data;

[0092] Alarm;

[0093] Wireless cell MR;

[0094] Parameter data;

[0095] Resource data.

[0096] Here, the first sequence represents the first indicator data, denoted by Y. Specifically, it can include three main indicators: wireless dropout rate, handover success rate, and wireless connection success rate, as well as other user-level data. These other user-level data can refer to RRC connection count, uplink and downlink traffic, etc. Y reflects the stability and usage of the current network, and is the most important indicator when problems occur in the current network. A total of N indicators are represented. y dimension.

[0097] Here, the second sequence represents the second indicator data, alarm data, and resource data, denoted by X. Specifically, it may include wireless cell KPI data, MR data, alarm data, resource data, etc. These data have a causal relationship or a strong correlation with Y, totaling N. x dimension.

[0098] Here, the third sequence characterizing parameter data, denoted by T, refers to parameters set by existing network equipment. These parameters are typically those added to network elements or adjusted during faults or network optimization. Changes in these parameters directly affect Y, totaling N. t dimension.

[0099] Here, the fourth sequence represents the influence factor, denoted by C, and can specifically include many unobserved influence factors C. Influence factors C affect the relationships between data; therefore, N is added. c The confounding factor, initialized with a Gaussian distribution, learns the implicit relationships between variables during the encoding phase and assists in data reconstruction during the decoding phase.

[0100] It should be noted that, referencing the different roles of different types of data in relation learning within causal inference theory, and considering the actual meaning of data fields in wireless networks, the current network data is divided into three categories: X, Y, and T. Furthermore, based on causal learning theory, a confounding factor C is added. In other words, considering the characteristics of wireless network data, the data is divided into three parts, and a confounding factor is added, resulting in a clearer understanding of data relationships and facilitating accurate model learning.

[0101] In one embodiment, determining the second loss value based on the fourth input data includes:

[0102] The fourth input data is used to predict the first indicator data within the fourth time period;

[0103] The second root mean square error is calculated using the predicted first indicator data for the fourth time period and the actual first indicator data.

[0104] The second root mean square error is used as the second loss value.

[0105] The embodiments of the present invention have the following advantages:

[0106] (1) Based on the various dimensions of test data in the current network, such as KPI data, alarm data, MR data, engineering parameter data, neighbor cell relationship data, etc., combined with the pre-trained relationship representation model, we learn the comprehensive representation relationship between individual data in different dimensions. Since a general relationship representation model is used, the data relationship expression between the data can be completed, thus improving the efficiency of relationship representation of various data.

[0107] (2) Using the relationship representation model and the acquired data from multiple dimensions, the overall representation of data relationships is learned, capturing the static and dynamic relationships between data, and providing a more comprehensive representation of the data.

[0108] Figure 2 This is a schematic diagram illustrating the specific implementation process of the data relationship identification method according to an embodiment of the present invention, as follows: Figure 2 As shown, the method includes steps 201 to 203:

[0109] Step 201: Obtain the training data; the training data includes third sequence data of multiple dimensions within the third time period and fourth sequence data of the time dimension within the fourth time period.

[0110] Here, during the data access process, various types of primary data from the wireless network operation and maintenance process can be collected. During the data classification process, the collected primary data of various types can first be classified and converted into corresponding sequence data, resulting in multi-dimensional third sequence data within the third time period, which is subsequently used for relation learning. During the data representation process, the multi-dimensional third sequence data within the third time period is represented to obtain a first sequence, a second sequence, a third sequence, and a fourth sequence.

[0111] The various types of initial data collected here include at least:

[0112] Wireless cell KPI data;

[0113] Alarm;

[0114] Wireless cell MR;

[0115] Parameter data;

[0116] Resource data.

[0117] Here, the first sequence represents the first indicator data, denoted by Y. Specifically, it can include three main indicators: wireless dropout rate, handover success rate, and wireless connection success rate, as well as other user-level data. These other user-level data can refer to RRC connection count, uplink and downlink traffic, etc. Y reflects the stability and usage of the current network, and is the most important indicator when problems occur in the current network. A total of N indicators are represented. y dimension.

[0118] Here, the second sequence represents the second indicator data, alarm data, and resource data, denoted by X. Specifically, it may include wireless cell KPI data, MR data, alarm data, resource data, etc. These data have a causal relationship or a strong correlation with Y, totaling N. x dimension.

[0119] Here, the third sequence characterizing parameter data, denoted by T, refers to parameters set by existing network equipment. These parameters are typically those added to network elements or adjusted during faults or network optimization. Changes in these parameters directly affect Y, totaling N. t dimension.

[0120] Here, the fourth sequence represents the influence factor, denoted by C, and can specifically include many unobserved influence factors C. Influence factors C affect the relationships between data; therefore, N is added. c The confounding factor, initialized with a Gaussian distribution, learns the implicit relationships between variables during the encoding phase and assists in data reconstruction during the decoding phase.

[0121] It should be noted that, referencing the different roles of different types of data in relation learning within causal inference theory, and considering the actual meaning of data fields in wireless networks, the current network data is divided into three categories: X, Y, and T. Furthermore, based on causal learning theory, a confounding factor C is added. In other words, by combining the characteristics of wireless network data and dividing it into three parts, while simultaneously adding a confounding factor, the data relationships are more clearly defined, facilitating accurate model learning.

[0122] Here, the collected first data of various types is converted into corresponding sequence data, specifically including the following cases:

[0123] In the first case, for time series data, it can be used directly.

[0124] In the second scenario, for non-time-series data, one-hot encoding is used to encode the data and convert it into a numerical representation. Simultaneously, replication and expansion are performed along the time dimension to transform the data into a time series.

[0125] In the third scenario, data of categorized types is encoded before use. Ultimately, the input data for each network element is a time series.

[0126] For example, regarding alarm data: First, all alarm types are summarized. Wireless network alarms are divided into N categories, which, after one-hot encoding, result in an N-column vector. The alarm occurrence time and alarm clearance time are then queried. For this time period, all alarm characteristic columns are encoded using the one-hot encoding method described above. For other times, it is assumed that no alarms occurred, so all columns are set to 0.

[0127] Table 1 illustrates the conversion of alarm data into corresponding sequence data. As shown in Table 1, assume that alarm C1 occurred in network element A at time T1, and the corresponding one-hot encoding is [1, 0, ..., 0]. The alarm ends at time T4.

[0128] Table 1

[0129]

[0130]

[0131] Here, data representation (embedding) is analogous to the encoding of word vectors in NLP. By using the time dimension to encode the current time-series data, it is possible to achieve both feature relationships and time-series relationships between data blocks.

[0132] For example, suppose the current network collects data at a frequency of one point every 15 minutes, resulting in 96 points in a 24-hour day. Each dimension of data is represented using these 96 points. Since the current network data is mostly collected on a daily basis, this approach can encompass all data features for the entire period.

[0133] Table 2 is the data representation of X, Table 3 is the data representation of Y, Table 4 is the data representation of T, and Table 5 is the data representation of C.

[0134] Table 2

[0135] <![CDATA[x2 t1 ]]> <![CDATA[x2 t2 ]]> … <![CDATA[x2 t95 ]]> <![CDATA[x2 t96 ]]> … … … … … <![CDATA[xn t1 ]]> <![CDATA[xn t2 ]]> … <![CDATA[xn t95 ]]> <![CDATA[xn t96 ]]>

[0136] Table 3

[0137] <![CDATA[y2 t1 ]]> <![CDATA[y2 t2 ]]> … <![CDATA[y2 t95 ]]> <![CDATA[y2 t96 ]]> … … … … … <![CDATA[yn t1 ]]> <![CDATA[yn t2 ]]> … <![CDATA[yn t95 ]]> <![CDATA[yn t96 ]]>

[0138] Table 4

[0139] <![CDATA[t2 t1 ]]> <![CDATA[t2 t2 ]]> … <![CDATA[t2 t95 ]]> <![CDATA[t2 t96 ]]> … … … … … <![CDATA[tn t1 ]]> <![CDATA[tn t2 ]]> … <![CDATA[tn t95 ]]> <![CDATA[tn t96 ]]>

[0140] Table 5

[0141] <![CDATA[c2 t1 ]]> <![CDATA[c2 t2 ]]> … <![CDATA[c2 t95 ]]> <![CDATA[c2 t96 ]]> … … … … … <![CDATA[Cn t1 ]]> <![CDATA[Cn t2 ]]> … <![CDATA[cn t95 ]]> <![CDATA[cn t96 ]]>

[0142] Step 202: Based on the training data, train the prediction model to obtain the relationship representation model.

[0143] Here, based on the above data representation methods, we propose a basic model for general relation representation based on transformers. The specific principles and architecture are as follows.

[0144] The design principles of the prediction model mentioned here include the following two aspects:

[0145] First, a multi-head attention mechanism forms the model skeleton: To learn the data relationships between different dimensions after the above data representation, an attention mechanism is introduced as the skeleton structure of the model, setting up multiple heads. Each head can learn a certain relationship between features of all dimensions in the entire time series under the current focus. When the multi-head aggregation is completed, various complex relationships of the entire time series globally can be obtained.

[0146] Second, the Encoder and Decoder architecture: To learn the overall representation between network elements, two types of relationships between data are proposed. One is the static relationship between various data (X, Y, C, T) of the network element. Simultaneously, it is also necessary to learn the influence of T on other variables. This allows for the capture of dynamic relationships between various data (X, Y, C, T) when the network data changes. The superposition of these two relationships constitutes the relationship representation between all data of the network element. Therefore, the model references the Encoder and Decoder architecture of the transformer, and completes the capture of static and dynamic relationships based on a self-designed data classification and input method.

[0147] Here, the structural design of the prediction model includes the following two parts:

[0148] First, the Encoder network: The encoder is designed to capture the static relationships between network data.

[0149] Second, the Decoder network: The design of the Decoder part mainly involves learning the dynamic relationships of network data.

[0150] Figure 3 This is a schematic diagram of the Encoder network according to an embodiment of the present invention. Figure 3As shown, the third sequence data with multiple dimensions within the third time period will be used as the third input data. For example, the time series consisting of four parts—Y, X, T, and C—within the third time period from time 0 to t1 will be used as the third input data, totaling 96 points. The embedding method is shown in the data section. Therefore, the data block input into the encoder of the prediction model is [N]. x +N y +N c +N t ×96. The overall network structure is based on a multi-layer multi-head-attention structure. During computation, some rows of data in at least one of the first, second, third, and fourth sequences are randomly masked. After the stacking module of the attention layer is completed, the randomly masked rows are restored to obtain the restored data. Using the restored data and the randomly masked data, the first root mean square error (RMSE) is calculated, and the first RMSE is used as the first loss value, denoted by the first Loss. Based on this structure, when the first Loss is sufficiently small, the static relationships between all dimensions of the third sequence data within the third time interval from time 0 to time t1 can be learned, including the relationship between the confusion factor and the variables between network elements.

[0151] Figure 4 This is a schematic diagram of the Decoder network according to an embodiment of the present invention. Figure 4As shown, the fourth input data is obtained by utilizing the first relationship between the third sequence data of different dimensions within the third time period and the fourth sequence data of the time dimension within the fourth time period. For example, the first part of the data is the fourth sequence data T of the time dimension within the fourth time period from t1 to t2. In order to learn dynamic relationships in the network, the second part of the data is the output of the encoder, which is equivalent to the static relationship between X, Y, C, and T data within the third time period from 0 to t1. The model structure is also based on a multi-head-attention structure. First, a dynamic relationship learning network is constructed. For the fourth sequence data T of the time dimension within the fourth time period from t1 to 2×t1 and the result of the encoder output, a mutual attention relationship learning network is designed. The resulting weight matrix represents the relationship between the next time T and the upper layer data result. If T changes at this time, the dynamic relationship can be captured naturally. If T does not change, a stable relationship is learned. Then, self-attention relationship learning is performed, mainly by introducing weights based on the pairwise relationships to strengthen the connection between the mutual attention results, thereby improving accuracy. After learning, a fully connected layer is used to adjust the data dimensions and reconstruct the first indicator data Y for the fourth time period from time t1 to 2×t1. Using the real Y value and the reconstructed Y value for the fourth time period, the second RMSE is calculated and represented by the second Loss. Based on this structure, when the second Loss is small enough, the static relationship between all dimensions of the third sequence data in the third time period from 0 to t1 and the dynamic relationship between the time dimension of the fourth sequence data in the fourth time period can be learned, emphasizing the impact of the changing T on the Y value.

[0152] Figure 5 This is a schematic diagram of the prediction model according to an embodiment of the present invention. Figure 5 As shown, the final loss of the entire network is the sum of Loss1 in the Encoder part and Loss2 in the Decoder part, which are used to adjust the model through backpropagation. Here, model training relies on a large amount of live network data to obtain a pre-trained base model, which serves as a public model. When integrating new data, several rounds of fine-tuning can be performed on this model to quickly obtain a model adapted to the new data, thus greatly improving the efficiency of model adaptation.

[0153] Here, using downstream task applications as examples, based on the aforementioned general basic model, the model can be fine-tuned to design a lightweight network structure, which can then be merged with the larger model to quickly support downstream tasks. Take the following two tasks as examples: In the perturbation work order recognition project, to obtain the reconstruction relationship between data, only the Encoder part of the model is used. Data from one day, three days, and seven days prior to the dispatch time are processed through the Encoder part and then connected to a lightweight classification network for rapid training of the perturbation binary classification model. In the root cause localization project, to obtain the root causes corresponding to different Y values, a lightweight root cause classification network can be added after the Decoder network for rapid training of root cause classification.

[0154] Step 203: Obtain the data to be tested; the data to be tested includes first sequence data of multiple dimensions within a first time period and second sequence data of time dimension within a second time period; input the data to be tested into the relation representation model, so as to output the relation representation vector of the data to be tested through the relation representation model; the relation representation vector is used to represent the relationship between the first sequence data of different dimensions.

[0155] This example has the following advantages:

[0156] (1) To address the problem of isolated model training in related technologies, a general model based on the transformer architecture is proposed. Based on various types of data from the existing network, such as KPI data, alarm data, MR data, engineering parameter data, neighbor cell relationship data, etc., and combined with a general basic model, the comprehensive representational relationship between individual network data is learned.

[0157] For downstream tasks, this representational relationship can be used to design lightweight algorithms that can quickly support individual tasks. In other words, because it can learn the comprehensive and essential relationships between wireless network indicator data, it can quickly connect to existing or new downstream tasks.

[0158] (2) In view of the problem that the current model in related technologies has limited representation of network elements, a data processing method and a two-stage model for learning time-series data relationships were designed.

[0159] Step 1: In data processing, following the existing network's cycle, the data is divided into blocks and encoded on a daily basis. Step 2: In model design, there are two stages: Encoder and Decoder. In the Encoder stage, the static relationships between all data in a third time period (e.g., time 0 to t) are learned. In the Decoder stage, the impact of parameter adjustments on data relationships is emphasized, and the dynamic relationships between data in a fourth time period (e.g., time t to 2t) are learned. The entire network structure leverages the strong ability of multi-head attention to capture data relationships, learning the correlation and temporal sequence of data.

[0160] (3) In view of the lack of focus in data processing and classification in related technologies, the data is divided into three types by referring to the different roles of different types of data in relation learning in causal models and combining the data classification in wireless networks. A confusion factor is also added to fit unobserved variables.

[0161] Specifically, the data is divided into three types: 1) characteristics of static indicators from the current network (most KPI data, MR data, alarm data, etc.); 2) three major manifestation indicators: wireless drop rate, handover success rate, wireless connection success rate, and other user-level data; 3) most parameter data, which can affect the data changes in 1) and 2). An influencing factor, also known as a confounding factor, is added. Here, these indicators play different roles at different stages of the model.

[0162] To implement the data relationship identification method of the present invention, the present invention also provides a data relationship identification device. Figure 6 This is a schematic diagram of the composition structure of the data relationship recognition device according to an embodiment of the present invention, as shown below. Figure 6 As shown, the device includes:

[0163] Acquisition unit 61 is used to acquire data to be tested; the data to be tested includes first sequence data of multiple dimensions within a first time period and second sequence data of time dimension within a second time period.

[0164] Processing unit 62 is used to input the test data into a pre-trained relation representation model, so as to output a relation representation vector of the test data through the relation representation model; the relation representation vector is used to represent the relationship between first sequence data in different dimensions;

[0165] The relationship representation model is obtained by training the prediction model based on training data; the training data includes third sequence data of multiple dimensions in the third time period and fourth sequence data of time dimension in the fourth time period.

[0166] In one embodiment, the relation representation model includes an encoder and a decoder; the processing unit 62 is configured to:

[0167] The first sequence data of multiple dimensions within a first time period is used as the first input data; the encoder encodes the first input data to obtain a first relation representation vector; the first relation representation vector represents the first relationship between the first sequence data of different dimensions within the first time period.

[0168] Using the first relationship between first sequence data of different dimensions within a first time period and the second sequence data of time dimension within a second time period, second input data is obtained; the second input data is decoded by the decoder to obtain a second relationship representation vector; the second relationship representation vector represents the first relationship between first sequence data of different dimensions within the first time period and the second relationship between second sequence data of time dimension within the second time period.

[0169] In one embodiment, the acquisition unit 61 is further configured to: acquire the training data;

[0170] Accordingly, the processing unit 62 is further configured to: train the prediction model based on the training data to obtain the relation representation model.

[0171] In one embodiment, the processing unit 62 is configured to:

[0172] The third sequence data of multiple dimensions within the third time period is used as the third input data; the third input data is encoded by the encoder in the initial prediction model to obtain the third relation representation vector; the third relation representation vector represents the third relation between the third sequence data of different dimensions within the third time period.

[0173] Using the third relationship between the third sequence data of different dimensions in the third time period and the fourth sequence data of the time dimension in the fourth time period, fourth input data is obtained; the fourth input data is decoded by the decoder in the initial prediction model to obtain the fourth relationship representation vector; the fourth relationship representation vector represents the third relationship between the third sequence data of different dimensions in the third time period and the fourth relationship between the fourth sequence data of the time dimension in the fourth time period.

[0174] Based on the third sequence data of multiple dimensions within the third time period, a first loss value is determined; based on the fourth input data, a second loss value is determined.

[0175] The first loss value and the second loss value are summed to obtain the third loss value;

[0176] The parameters of the prediction model are updated using the third loss value to obtain the relationship representation model.

[0177] In one embodiment, the processing unit 62 is configured to:

[0178] Based on the third sequence data of multiple dimensions within the third time period, a first sequence, a second sequence, a third sequence, and a fourth sequence are obtained; the first sequence represents the first indicator data; the second sequence represents the second indicator data, alarm data, and resource data; the third sequence represents parameter data; and the fourth sequence represents the influencing factors.

[0179] Randomly mask a portion of the data in at least one of the first sequence, the second sequence, the third sequence, and the fourth sequence;

[0180] The randomly obscured data is restored to obtain the restored data.

[0181] The first root mean square error is calculated using the recovered data and the randomly masked data, and the first root mean square error is used as the first loss value.

[0182] In one embodiment, the processing unit 62 is configured to:

[0183] The fourth input data is used to predict the first indicator data within the fourth time period;

[0184] The second root mean square error is calculated using the predicted first indicator data for the fourth time period and the actual first indicator data.

[0185] The second root mean square error is used as the second loss value.

[0186] In practical applications, the acquisition unit 61 can be implemented by the communication interface in the data relationship identification device; the processing unit 62 can be implemented by the processor in the data relationship identification device.

[0187] It should be noted that the data relationship recognition device provided in the above embodiments is only illustrated by the division of the above program modules when performing data relationship recognition. In actual applications, the above processing can be assigned to different program modules as needed, that is, the internal structure of the device can be divided into different program modules to complete all or part of the processing described above. In addition, the data relationship recognition device and the data relationship recognition method embodiments provided in the above embodiments belong to the same concept, and their specific implementation process can be found in the method embodiments, which will not be repeated here.

[0188] This invention also provides a network device, such as... Figure 7 As shown, it includes:

[0189] The communication interface 71 enables information exchange with other devices;

[0190] The processor 72, connected to the communication interface 71, is used to execute the methods provided by one or more technical solutions on the network device side when running a computer program. The computer program is stored in the memory 73.

[0191] It should be noted that the specific processing procedures of the processor 72 and the communication interface 71 are detailed in the method embodiment and will not be repeated here.

[0192] Of course, in practical applications, the various components in network device 70 are coupled together through bus system 74. It can be understood that bus system 74 is used to implement communication between these components. In addition to a data bus, bus system 74 also includes a power bus, a control bus, and a status signal bus. However, for clarity, in... Figure 7 The general labeled all buses as Bus System 74.

[0193] The memory 73 in this embodiment is used to store various types of data to support the operation of the network device 70. Examples of such data include any computer program used to operate on the network device 70.

[0194] The methods disclosed in the embodiments of this application can be applied to the processor 72, or implemented by the processor 72. The processor 72 may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method can be completed by the integrated logic circuit of the hardware in the processor 72 or by instructions in the form of software. The processor 72 may be a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The processor 72 can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor may be a microprocessor or any conventional processor, etc. The steps of the methods disclosed in the embodiments of this application can be directly reflected as being executed by a hardware decoding processor, or being executed by a combination of hardware and software modules in the decoding processor. The software modules may be located in a storage medium, which is located in the memory 73. The processor 72 reads the information in the memory 73 and combines its hardware to complete the steps of the aforementioned method.

[0195] In an exemplary embodiment, the network device 70 may be implemented by one or more application-specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), general-purpose processors, controllers, microcontrollers (MCUs), microprocessors, or other electronic components to perform the aforementioned method.

[0196] It is understood that the memory (memory 73) in the embodiments of this application can be volatile memory or non-volatile memory, or both. Non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), ferromagnetic random access memory (FRAM), flash memory, magnetic surface memory, optical disc, or compact disc read-only memory (CD-ROM); magnetic surface memory can be disk storage or magnetic tape storage. Volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as Static Random Access Memory (SRAM), Synchronous Static Random Access Memory (SSRAM), Dynamic Random Access Memory (DRAM), Synchronous Dynamic Random Access Memory (SDRAM), Double Data Rate Synchronous Dynamic Random Access Memory (DDRSDRAM), Enhanced Synchronous Dynamic Random Access Memory (ESDRAM), SyncLink Dynamic Random Access Memory (SLDRAM), and Direct Rambus Random Access Memory (DRRAM).The memories described in the embodiments of this application are intended to include, but are not limited to, these and any other suitable types of memories.

[0197] In an exemplary embodiment, the present invention also provides a storage medium, namely a computer storage medium, specifically a computer-readable storage medium, such as a memory that stores a computer program, which can be executed by the processor 72 of the network device 70 to complete the steps described in the aforementioned network device-side method. The computer-readable storage medium may be a memory such as FRAM, ROM, PROM, EPROM, EEPROM, Flash Memory, magnetic surface memory, optical disc, or CD-ROM.

[0198] It should be noted that terms such as "first" and "second" are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence.

[0199] Furthermore, the technical solutions described in the embodiments of the present invention can be combined arbitrarily without conflict.

[0200] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention.

Claims

1. A data relationship identification method, characterized in that, The data relationship refers to the data relationship of network elements in a network intelligent operation and maintenance scenario, and the method includes: Acquire the data to be tested; the data to be tested includes first sequence data of multiple dimensions within a first time period and second sequence data of time dimension within a second time period; The test data is input into a pre-trained relation representation model to output a relation representation vector of the test data; the relation representation vector is used to represent the relationship between first sequence data in different dimensions. The relationship representation model is obtained by training the prediction model based on training data; the training data includes third sequence data of multiple dimensions in the third time period and fourth sequence data of time dimension in the fourth time period. The relation representation model includes an encoder and a decoder; The step of inputting the test data into a pre-trained relation representation model to output a relation representation vector of the test data through the relation representation model includes: The first sequence data of multiple dimensions within a first time period is used as the first input data; the encoder encodes the first input data to obtain a first relation representation vector; the first relation representation vector represents the first relationship between the first sequence data of different dimensions within the first time period. Using the first relationship between first sequence data of different dimensions within a first time period and the second sequence data of time dimension within a second time period, second input data is obtained; the second input data is decoded by the decoder to obtain a second relationship representation vector; the second relationship representation vector represents the first relationship between first sequence data of different dimensions within the first time period and the second relationship between second sequence data of time dimension within the second time period.

2. The method according to claim 1, characterized in that, The method further includes: Obtain the training data; Based on the training data, the prediction model is trained to obtain the relationship representation model.

3. The method according to claim 2, characterized in that, The step of training the prediction model based on the training data to obtain the relation representation model includes: The third sequence data of multiple dimensions within the third time period is used as the third input data; the third input data is encoded by the encoder in the initial prediction model to obtain the third relation representation vector; the third relation representation vector represents the third relation between the third sequence data of different dimensions within the third time period. Using the third relationship between the third sequence data of different dimensions in the third time period and the fourth sequence data of the time dimension in the fourth time period, fourth input data is obtained; the fourth input data is decoded by the decoder in the initial prediction model to obtain the fourth relationship representation vector; the fourth relationship representation vector represents the third relationship between the third sequence data of different dimensions in the third time period and the fourth relationship between the fourth sequence data of the time dimension in the fourth time period. Based on the third sequence data of multiple dimensions within the third time period, a first loss value is determined; based on the fourth input data, a second loss value is determined. The first loss value and the second loss value are summed to obtain the third loss value; The parameters of the prediction model are updated using the third loss value to obtain the relationship representation model.

4. The method according to claim 3, characterized in that, The determination of the first loss value based on the third sequence data of multiple dimensions within the third time period includes: Based on the third sequence data of multiple dimensions within the third time period, a first sequence, a second sequence, a third sequence, and a fourth sequence are obtained; the first sequence represents the first indicator data; the second sequence represents the second indicator data, alarm data, and resource data; the third sequence represents parameter data; and the fourth sequence represents the influencing factors. Randomly mask a portion of the data in at least one of the first sequence, the second sequence, the third sequence, and the fourth sequence; The randomly obscured data is restored to obtain the restored data. The first root mean square error is calculated using the recovered data and the randomly masked data, and the first root mean square error is used as the first loss value.

5. The method according to claim 3, characterized in that, Determining the second loss value based on the fourth input data includes: The fourth input data is used to predict the first indicator data within the fourth time period; The second root mean square error is calculated using the predicted first indicator data for the fourth time period and the actual first indicator data. The second root mean square error is used as the second loss value.

6. A data relationship identification device, characterized in that, The data relationships described refer to the data relationships of network elements in intelligent network operation and maintenance scenarios, including: An acquisition unit is used to acquire data to be tested; the data to be tested includes first sequence data of multiple dimensions within a first time period and second sequence data of time dimension within a second time period. The processing unit is configured to input the test data into a pre-trained relational representation model, and output a relational representation vector of the test data through the relational representation model; the relational representation vector is used to represent the relationship between first sequence data in different dimensions; The relationship representation model is obtained by training the prediction model based on training data; the training data includes third sequence data of multiple dimensions in the third time period and fourth sequence data of time dimension in the fourth time period. The relation representation model includes an encoder and a decoder; The processing unit is further configured to take the first sequence data of multiple dimensions within the first time period as the first input data; and to encode the first input data through the encoder to obtain a first relation representation vector; the first relation representation vector represents the first relation between the first sequence data of different dimensions within the first time period. Using the first relationship between first sequence data of different dimensions within a first time period and the second sequence data of time dimension within a second time period, second input data is obtained; the second input data is decoded by the decoder to obtain a second relationship representation vector; the second relationship representation vector represents the first relationship between first sequence data of different dimensions within the first time period and the second relationship between second sequence data of time dimension within the second time period.

7. A data relationship identification device, wherein the data relationship is the data relationship of network elements in a network intelligent operation and maintenance scenario, characterized in that, include: A communication interface is used to acquire the data to be tested; the data to be tested includes first sequence data with multiple dimensions within a first time period and second sequence data with a time dimension within a second time period. A processor is configured to input the test data into a pre-trained relational representation model, and output a relational representation vector of the test data through the relational representation model. The relation representation vector is used to represent the relationship between the first sequence data in different dimensions; The relationship representation model is obtained by training the prediction model based on training data; the training data includes third sequence data of multiple dimensions in the third time period and fourth sequence data of time dimension in the fourth time period. The relation representation model includes an encoder and a decoder; The processor is further configured to take the first sequence data of multiple dimensions within the first time period as the first input data; and to encode the first input data by the encoder to obtain a first relation representation vector; the first relation representation vector represents the first relation between the first sequence data of different dimensions within the first time period. Using the first relationship between first sequence data of different dimensions within a first time period and the second sequence data of time dimension within a second time period, second input data is obtained; the second input data is decoded by the decoder to obtain a second relationship representation vector; the second relationship representation vector represents the first relationship between first sequence data of different dimensions within the first time period and the second relationship between second sequence data of time dimension within the second time period.

8. A network device, characterized in that, This includes a processor and memory for storing computer programs that can run on the processor. When the processor is used to run the computer program, it performs the steps of the method according to any one of claims 1 to 5.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 5.

Images