Wireless network root cause positioning method and device, electronic equipment, storage medium and product

Abstract

This invention provides a method, apparatus, electronic device, storage medium, and product for root cause localization in wireless networks. The method includes: a multi-scale feature extraction module adaptively extracting and fusing features from time-series data at different time scales to obtain fused features; a temporal dependency modeling module determining target features based on long-distance dependencies in the fused features; and a dynamic classification module generating a classification confidence score and an adaptive classification threshold for each preset root cause category based on the target features, and determining the root cause localization result based on the classification confidence score and the adaptive classification threshold. This invention achieves this by fusing features at different temporal granularities in time-series data through multi-scale feature extraction; by modeling long-distance dependencies in the fused features, it enables a deeper understanding of the dynamic evolution of faults; and by generating an adaptive classification threshold for each sample through the dynamic classification module, it can flexibly handle samples of varying complexity, significantly improving the accuracy and robustness of root cause localization.

Description

Technical Field

[0001] This invention relates to the field of wireless network technology, and in particular to a wireless network root cause localization method, apparatus, electronic device, storage medium, and product. Background Technology

[0002] In the field of wireless networks, existing technologies mainly include human experience-based solutions, solutions based on traditional classification models, and solutions based on general neural networks.

[0003] Human-based solutions rely on expert knowledge, determining root causes by observing the fluctuations of specific indicators within a fixed time window. This process heavily depends on human experience and fixed indicator selection and positioning order, resulting in low positioning efficiency and difficulty in adapting to complex multi-root-cause scenarios. While solutions based on traditional classification models can automatically extract some pre-defined local features for classification, their feature sets are predefined by algorithm developers, thresholds are fixed, and they cannot fully learn the complex deep relationships between indicators. The model's generalization ability is significantly affected by the training samples, and it is also difficult to handle multi-label classification problems.

[0004] However, while solutions using general neural network structures such as convolutional neural networks or attention mechanisms can perform end-to-end learning of time series indicators, they are still insufficient in capturing multi-scale, fine-grained relationships between indicators, and lack mature and efficient multi-label output mechanisms. On the other hand, directly using large-scale time series prediction models with a large number of parameters faces the problems of high model training and inference computation costs and difficulty in deploying and applying them in actual operation and maintenance environments. Summary of the Invention

[0005] This invention provides a method, apparatus, electronic device, storage medium, and product for wireless network root cause localization, which addresses the shortcomings of existing technologies in achieving automation, multi-tag functionality, detailed modeling, and low computational cost in wireless network root cause localization.

[0006] This invention provides a method for root cause localization in wireless networks, comprising the following steps.

[0007] Acquire time-series data of the wireless network to be located; The time series data is input into the root cause localization model to obtain the root cause localization result output by the root cause localization model, wherein the root cause localization result is at least one root cause category of the time series data. The root cause localization model includes a multi-scale feature extraction module, a temporal dependency modeling module, and a dynamic classification module. The multi-scale feature extraction module is used to adaptively extract and fuse features from the time series data at different time scales to obtain fused features. The temporal dependency modeling module is used to determine target features based on long-distance dependencies in the fused features. The dynamic classification module is used to generate a classification confidence score and an adaptive classification threshold for each preset root cause category based on the target features, and to determine the root cause localization result based on the classification confidence score and the adaptive classification threshold.

[0008] According to a wireless network root cause localization method provided by the present invention, the time-series dependency modeling module is specifically used for: The fused features are subjected to dilated convolution to obtain intermediate temporal features; The intermediate time-series features are divided into multiple feature blocks along the time dimension; Self-attention computation is performed within each feature block to obtain locally enhanced features that include local dependency information; The target feature is determined based on multiple local enhancement features.

[0009] According to a wireless network root cause localization method provided by the present invention, the time-series dependency modeling module is specifically used for: Each of the local enhancement features is pooled to obtain a local summary vector. Global attention computation is performed between the local summary vectors to obtain a globally enhanced summary. The target features are obtained based on the global enhanced summary and each of the local summary vectors.

[0010] According to a wireless network root cause localization method provided by the present invention, the time-series dependency modeling module is specifically used for: The fused features are input into the dilated convolution module to obtain the intermediate temporal features; The dilated convolution module includes at least two dilated convolution layers connected in sequence; The dilation rate within the dilated convolutional layer increases sequentially along the feature propagation direction; the output of the dilated convolutional module is the sum of the residuals between the output of the last dilated convolutional layer and the fused features.

[0011] According to a wireless network root cause localization method provided by the present invention, the multi-scale feature extraction module is specifically used for: The short-term features of the first time scale, the medium-term features of the second time scale, and the long-term features of the third time scale are extracted in parallel from the time series data. The short-term features, the medium-term features, and the long-term features are fused to obtain the fused features.

[0012] According to a wireless network root cause localization method provided by the present invention, the multi-scale feature extraction module is specifically used for: The time series data is subjected to a first convolution process to obtain the short-term features; The time series data is subjected to a second convolution process to obtain the intermediate features; The time series data is subjected to mean pooling to obtain pooled features, and the pooled features are subjected to a third convolution to obtain the long-term features.

[0013] According to a wireless network root cause localization method provided by the present invention, the multi-scale feature extraction module is specifically used for: Adjust the short-term, medium-term, and long-term features to the same time length; The short-term, medium-term, and long-term features, adjusted to the same time length, are then spliced ​​along the feature dimension to obtain the fused feature.

[0014] According to a wireless network root cause localization method provided by the present invention, the dynamic classification module is specifically used for: The target features are input into a shared coding network for encoding to obtain the shared features output by the shared coding network; Based on the shared features, a classification confidence score and an adaptive classification threshold are generated for each of the preset root cause categories.

[0015] According to the present invention, a wireless network root cause localization method is provided, wherein the root cause localization model is obtained by iteratively executing the following steps until a preset iteration termination condition is met: Obtain the training dataset; the training dataset includes time series data samples and the true values ​​of the multi-label root cause categories corresponding to the time series data samples; The time series data samples are input into the initial root cause localization model to obtain the predicted score for each root cause category output by the initial root cause localization model. The target loss is determined based on the predicted scores and the true values ​​of the multi-label root cause categories. The model parameters of the initial root cause localization model are updated based on the target loss.

[0016] According to a wireless network root cause localization method provided by the present invention, the step of determining the target loss based on each of the predicted scores and the true values ​​of the multi-label root cause categories includes: Determine the class weight of each root cause category in the preset root cause categories in the training dataset; The target loss is determined based on the predicted scores, the true values ​​of the multi-label root cause categories, and the category weights.

[0017] According to a wireless network root cause localization method provided by the present invention, the category weight is determined based on the number of time series data samples corresponding to each root cause category in the training dataset.

[0018] The present invention also provides a wireless network root cause localization device, comprising the following modules: The acquisition module is used to acquire time-series data of the wireless network to be located. An input module is used to input the time series data into the root cause localization model to obtain the root cause localization result output by the root cause localization model, wherein the root cause localization result is at least one root cause category of the time series data. The root cause localization model includes a multi-scale feature extraction module, a temporal dependency modeling module, and a dynamic classification module. The multi-scale feature extraction module is used to adaptively extract and fuse features from the time series data at different time scales to obtain fused features. The temporal dependency modeling module is used to determine target features based on long-distance dependencies in the fused features. The dynamic classification module is used to generate a classification confidence score and an adaptive classification threshold for each preset root cause category based on the target features, and to determine the root cause localization result based on the classification confidence score and the adaptive classification threshold.

[0019] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the wireless network root cause localization method as described above.

[0020] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the wireless network root cause localization method as described above.

[0021] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the wireless network root cause localization method as described above.

[0022] The present invention provides a wireless network root cause localization method, apparatus, electronic device, storage medium, and product, which acquires time-series data of the wireless network to be located; inputs the time-series data into a root cause localization model to obtain the root cause localization result output by the root cause localization model; wherein the root cause localization model includes a multi-scale feature extraction module, a temporal dependency modeling module, and a dynamic classification module; the multi-scale feature extraction module is used to adaptively extract and fuse features of different time scales from the time-series data to obtain fused features; the temporal dependency modeling module is used to determine target features based on long-distance dependencies in the fused features; the dynamic classification module is used to generate a classification confidence score and an adaptive classification threshold for each preset root cause category based on the target features, and determine the root cause localization result based on the classification confidence score and the adaptive classification threshold. This invention, through multi-scale feature extraction, can fuse features at different time granularities in time series data, avoiding information omissions caused by a single analysis scale. Furthermore, by modeling long-distance dependencies of the fused features, it can gain a deeper understanding of the dynamic evolution of faults. Finally, through a dynamic classification module, it generates adaptive classification thresholds for each sample, which can flexibly handle samples of varying complexity. This solves the problems of traditional methods relying on human experience, low efficiency, and the inability of traditional models to handle complex correlations and multi-root cause scenarios. It achieves accurate and automated location of wireless network faults, significantly improving the accuracy and robustness of root cause localization. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0024] Figure 1 This is a flowchart illustrating the wireless network root cause localization method provided by the present invention.

[0025] Figure 2 This is a flowchart illustrating the process by which the temporal dependency modeling module provided by this invention determines target features.

[0026] Figure 3 This is a schematic diagram of the process by which the multi-scale feature extraction module provided by the present invention determines the fused features.

[0027] Figure 4 This is a schematic diagram of the process by which the dynamic classification module provided by the present invention determines the classification confidence score and the adaptive classification threshold.

[0028] Figure 5 This is a schematic diagram of the structure of the wireless network root cause localization device provided by the present invention.

[0029] Figure 6 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0031] The terms "first," "second," etc., used in this invention are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention can be implemented in orders other than those illustrated or described herein, and that the objects distinguished by "first," "second," etc., are generally of the same class.

[0032] This invention provides a root cause localization method for wireless networks, aiming to address the problems of existing root cause localization methods relying on human experience, being inefficient, and struggling to handle complex scenarios with multiple concurrent root causes. This method constructs a root cause localization model that incorporates multi-scale feature extraction, temporal dependency modeling, and dynamic classification, enabling it to automatically, efficiently, and accurately identify one or more root causes from complex network metrics.

[0033] In one specific embodiment, this method can be applied to intelligent operation and maintenance systems for various wireless communication networks, such as 4G / 5G mobile communication networks and Wi-Fi networks, to automatically process network fault work orders and achieve rapid fault diagnosis and location. Figure 1 This is a flowchart illustrating the wireless network root cause localization method provided by the present invention, as shown below. Figure 1 As shown, the method includes the following: Step 110: Obtain the time series data of the wireless network to be located.

[0034] Specifically, firstly, time-series data of the wireless network to be located can be obtained. The wireless network to be located can refer to any wireless communication network that requires performance monitoring and fault diagnosis.

[0035] In a specific scenario, the wireless network to be located could be a city's 5G commercial network, which contains a large number of base stations and cells, serving a massive number of users, and its network performance status needs to be monitored in real time.

[0036] Here, time-series data refers to a set of one or more monitoring indicators of the operating status of the wireless network to be located over a period of time. Time-series data is typically collected and stored periodically by the network operation and maintenance management system. Time-series data may include key performance indicators (KPIs), measurement reports (MRs), and engineering parameters, etc., which are not specifically limited in this embodiment of the invention.

[0037] Key performance indicators include wireless call success rate, E-RAB call drop rate, handover success rate, uplink / downlink throughput, number of users in the cell, and PRB (Physical Resource Block) utilization. Measurement reports include, for example, Reference Signal Receiving Power (RSRP) and Reference Signal Receiving Quality (RSRQ) reported by the terminal. Engineering parameters include cell configuration parameters such as transmit power and antenna height, which are not specifically limited in this embodiment of the invention.

[0038] In practical applications, to comprehensively characterize the network status, multiple indicators related to common faults are typically selected. For example, 20 key indicators can be chosen, with data collected continuously for 3 days at a 30-minute granularity. Therefore, for each network unit to be analyzed (e.g., a cell), the acquired time-series data can be represented as a matrix of dimension T×F, where T is the time step (e.g., 3 days × 48 points / day = 144 time steps), and F is the number of feature indicators (e.g., 20 indicators). Based on this, the entire training task can be mathematically defined: Let the time-series dataset... ,in Let represent the i-th sample (where T = 144 time steps and F = 20 features). This corresponds to a multi-label vector (where C = 8 categories). The goal of this invention is to learn a mapping function. .

[0039] Step 120: Input the time series data into the root cause localization model to obtain the root cause localization result output by the root cause localization model, wherein the root cause localization result is at least one root cause category of the time series data; The root cause localization model includes a multi-scale feature extraction module, a temporal dependency modeling module, and a dynamic classification module. The multi-scale feature extraction module is used to adaptively extract and fuse features from the time series data at different time scales to obtain fused features. The temporal dependency modeling module is used to determine target features based on long-distance dependencies in the fused features. The dynamic classification module is used to generate a classification confidence score and an adaptive classification threshold for each preset root cause category based on the target features, and to determine the root cause localization result based on the classification confidence score and the adaptive classification threshold.

[0040] Specifically, after obtaining the time series data, the time series data can be input into the root cause localization model to obtain the root cause localization result output by the root cause localization model. The root cause localization result is at least one root cause category of the time series data.

[0041] The root cause localization model is a trained artificial intelligence model capable of end-to-end mapping from input time-series data to output root cause categories; specifically, it can be a deep learning model. The core function of the root cause localization model is to learn and understand the complex patterns of network metrics changing over time, and the deep correlation between these complex patterns and the root causes of specific network failures.

[0042] Here, the root cause localization result specifically refers to one or more root cause categories identified by the root cause localization model based on the input time series data. Root cause categories are predefined classifications of the root causes of network faults. For example, eight common root cause categories can be predefined, including high load, weak coverage, interference, poor VoLTE (Voice over Long-Term Evolution) quality, parameter configuration problems, etc. This embodiment of the invention does not specifically limit these categories.

[0043] In real-world network environments, a single fault may be caused by multiple root causes (e.g., an area may have both weak coverage and external interference). Therefore, the output of a root cause localization model can be multi-labeled, indicating multiple root cause categories simultaneously. For example, for a single sample, the root cause localization result might be {"weak coverage", "interference"}.

[0044] The root cause localization model includes a multi-scale feature extraction module, a temporal dependency modeling module, and a dynamic classification module to achieve accurate root cause localization.

[0045] The multi-scale feature extraction module adaptively extracts and fuses features from different time scales in time series data to obtain fused features. Network faults exhibit diverse manifestations over time. For example, "interference" may manifest as sudden and drastic fluctuations in indicators (short-term pattern), "high load" may manifest as periodic increases in indicators during specific periods (such as evening peak hours) (medium-term pattern), while "weak coverage" may be reflected in indicators remaining at consistently low levels (long-term trend). Traditional single-scale analysis methods tend to overlook key information at certain time scales.

[0046] Therefore, the multi-scale feature extraction module designed in this invention can capture features simultaneously from multiple time granularities. Features at different time scales can specifically include short-term features, medium-term features, and long-term features. Short-term features are used to capture short-term local patterns, such as instantaneous spikes in indicators, noise, and brief, dramatic fluctuations. Medium-term features are used to capture medium-term trend changes and periodic patterns, such as the tidal effect of workdays and periodic performance degradation. Long-term features are used to capture long-term overall trends, such as the slow deterioration or improvement of network performance.

[0047] The multi-scale feature extraction module extracts features at different scales by setting up multiple feature extractors with different receptive fields in parallel, such as convolutional neural networks with different kernel sizes. Then, the multi-scale feature extraction module can adaptively fuse these features. This adaptivity is reflected in the fact that the multi-scale feature extraction module can use learned weights to perform weighted summation or concatenation of features at different scales, allowing the root cause localization model to autonomously determine which time scale information is more important in the current sample.

[0048] Ultimately, the fused features output by the multi-scale feature extraction module are a feature representation with higher information density and stronger representation ability, taking into account both short-term details and long-term trends of time series data.

[0049] Here, the time-series dependency modeling module is used to determine target features based on long-distance dependencies in the fused features. In network operation and maintenance data, there are complex dependencies between metrics at different points in time, and these dependencies can be long-distance. For example, a parameter modification made in the early morning may not have its impact until several hours later during peak business hours. Capturing these long-distance dependencies is crucial for understanding the evolution and root causes of failures.

[0050] The temporal dependency modeling module receives the fused features output from the previous module and performs deep temporal modeling on these features. This module can be implemented using various neural network structures that effectively capture long-range dependencies, such as Recurrent Neural Networks (RNNs), Long Short-Term Memory (LSTM) networks, Gated Recurrent Units (GRUs), or Transformer models based on self-attention mechanisms. Through these structures, the temporal dependency modeling module can learn the correlation between any two time steps in a time series, regardless of their temporal distance.

[0051] Among them, the target feature is a high-level feature representation formed by capturing long-distance dependencies through a temporal dependency modeling module based on the fusion of multi-scale information, which is used to reflect the comprehensive representation of the root cause.

[0052] The dynamic classification module is used to generate a classification confidence score and an adaptive classification threshold for each preset root cause category based on the target features, and to determine the root cause localization result based on the classification confidence score and the adaptive classification threshold.

[0053] Traditional classification methods typically apply a fixed classification threshold, such as 0.5, to all samples. This "one-size-fits-all" approach is ineffective when dealing with real-world problems involving significant differences in sample difficulty and class imbalance. For example, a higher classification threshold might be appropriate for typical fault samples with clear features; however, for complex samples with ambiguous features, a fixed high classification threshold could lead to missed classifications.

[0054] The dynamic classification module proposed in this invention creatively solves this problem. First, the dynamic classification module takes the target feature as input. Second, for each preset root cause category, such as "high load", the dynamic classification module performs two calculations in parallel: generating a classification confidence score and an adaptive classification threshold.

[0055] The classification confidence score represents the probability or confidence that the current sample belongs to a root cause category. Unlike traditional methods, the dynamic classification module can dynamically generate a specific classification threshold for the current sample based on the input target features. Finally, the dynamic classification module compares the classification confidence score with the adaptive classification threshold for each root cause category. If the classification confidence score is greater than the adaptive classification threshold, then that root cause category is determined to be one of the root causes of the current fault. All root cause categories that meet the criteria are combined to obtain the final root cause localization result.

[0056] Here, the final classification decision formula is as follows: ; in, Indicates the root cause category The final predicted label (1 represents the positive class, 0 represents the negative class). Indicates root cause category Predicted classification confidence score This represents the adaptive classification threshold.

[0057] The method provided in this invention acquires time-series data of a wireless network to be located; inputs the time-series data into a root cause localization model to obtain the root cause localization result output by the root cause localization model. The root cause localization model includes a multi-scale feature extraction module, a temporal dependency modeling module, and a dynamic classification module. The multi-scale feature extraction module adaptively extracts and fuses features from different time scales in the time-series data to obtain fused features. The temporal dependency modeling module determines target features based on long-distance dependencies in the fused features. The dynamic classification module generates a classification confidence score and an adaptive classification threshold for each preset root cause category based on the target features, and determines the root cause localization result based on the classification confidence score and the adaptive classification threshold. This invention, through multi-scale feature extraction, can fuse features at different time granularities in time series data, avoiding information omissions caused by a single analysis scale. Furthermore, by modeling long-distance dependencies of the fused features, it can gain a deeper understanding of the dynamic evolution of faults. Finally, through a dynamic classification module, it generates adaptive classification thresholds for each sample, which can flexibly handle samples of varying complexity. This solves the problems of traditional methods relying on human experience, low efficiency, and the inability of traditional models to handle complex correlations and multi-root cause scenarios. It achieves accurate and automated location of wireless network faults, significantly improving the accuracy and robustness of root cause localization.

[0058] Based on the above embodiments, the time-series dependency modeling module is specifically used for: Step 210: Perform dilated convolution processing on the fused features to obtain intermediate temporal features; Step 220: Divide the intermediate time series features into multiple feature blocks in the time dimension; Step 230: Perform self-attention calculation within each feature block to obtain local enhanced features including local dependency information; Step 240: Determine the target feature based on multiple local enhancement features.

[0059] Specifically, Figure 2 This is a flowchart illustrating the process of determining target features using the temporal dependency modeling module provided by this invention, as shown below. Figure 2As shown, firstly, the fused features are subjected to dilated convolution to obtain intermediate time-series features. Dilated convolution is a special type of convolution operation that expands the receptive field by introducing gaps (dilation rates) between elements of the convolution kernel without increasing the number of parameters or computational cost. In this invention, the purpose of using dilated convolution is to cover the receptive field of the entire time-series data with a limited number of neural network layers, thereby efficiently capturing long-distance dependencies. For example, by stacking multiple dilated convolution layers with exponentially increasing dilation rates, such as 1, 2, 4, 8, the root cause localization model can quickly ensure that each time point of the output features perceives a very broad temporal range in the input sequence data.

[0060] In a specific implementation, dilated blocks containing multiple dilated convolutional layers can be designed. For example, two dilated blocks can be included, each containing four dilated convolutional layers. The dilation rate set D of the first dilated block can be set to {1,2,4,8}, and the dilation rate set D of the second dilated block can be set to {1,3,6,12}. Different combinations of dilation rates create multi-granularity time-awareness capabilities; smaller dilation rates are used to capture local mutations and short-term patterns, while larger dilation rates are used to identify long-term trends and periodic patterns.

[0061] The output obtained after dilated convolution is the intermediate temporal feature. Each time point of the intermediate temporal feature incorporates contextual information from a relatively large receptive field within the time series data at that time point.

[0062] Then, the intermediate time-series features are divided into multiple feature blocks along the time dimension. Here, the time dimension refers to the dimension of the time step arrangement in the time series data. Directly performing self-attention calculation on a complete long sequence, such as 144 time points, results in computational complexity proportional to the square of the sequence length, leading to enormous computational overhead. To address this issue, this invention employs a block-based processing strategy.

[0063] Specifically, the intermediate temporal feature H3 of length L is divided into M non-overlapping feature blocks, each of size B = L / M. This process can be achieved through a reshape operation: ; Where Batch is the batch size. This is the feature dimension. Therefore, subsequent computations can be performed efficiently in parallel on these shorter blocks.

[0064] Next, self-attention calculation is performed within each feature block to obtain locally enhanced features that include local dependency information. For each segmented feature block... It independently applies a standard multi-head attention mechanism. Multi-head attention can calculate the relevance weights between any two time steps within a block and aggregate features based on these relevance weights. This process effectively captures local dependency information within a block, which refers to the feature evolution patterns within a short time window.

[0065] After self-attention computation, the training process is typically stabilized by combining residual connections and layer normalization, while preserving the original information. This process can be represented by the following formula: ; in, It is the input feature block. It is a local enhancement feature. This indicates a multi-head attention mechanism. Representation layer normalization. Local enhancement features refer to features obtained by fusing contextual information from all other time points within each feature block at each time point.

[0066] Finally, the target features are determined based on multiple local enhancement features. This involves obtaining the local enhancement features of all feature blocks. Next, further processing is needed to integrate global information across different feature blocks, ultimately forming the target feature.

[0067] The method provided in this invention efficiently expands the receptive field by introducing dilated convolution and finely captures local dependencies by combining a block-based self-attention mechanism. Compared with directly applying standard self-attention or recurrent neural networks to long sequences, it significantly reduces computational complexity while ensuring the ability to model long-distance dependencies, making it more suitable for deployment in real-time operation and maintenance scenarios.

[0068] Based on the above embodiments, the time-series dependency modeling module is specifically used for: Step 2401: Perform pooling processing on each of the local enhancement features to obtain each local summary vector; Step 2402: Perform global attention calculation among the local summary vectors to obtain a globally enhanced summary; Step 2403: Based on the global enhanced summary and each of the local summary vectors, the target features are obtained.

[0069] Specifically, firstly, pooling is performed on each local enhancement feature to obtain a local summary vector. After completing the intra-block attention calculation, each local enhancement feature... This still constitutes a sequence. To enable information exchange between feature blocks, it is necessary to first generate a compact representation for each feature block that can represent the core information of that feature block. This step is accomplished through pooling operations.

[0070] Here, average pooling can be used to enhance each local feature. The processing is performed along the time dimension, compressing it from a feature sequence of size B×d into a vector of size 1×d. This vector is the local summary vector s of that block. i The calculation process can be expressed by the following formula: ; in, Indicates local enhancement features The feature vector at the j-th time step. Through this step, a complex sequence of length L is abstracted into a summary sequence S = [s1, s2, ..., sm] containing only M local summary vectors. m ].

[0071] Then, global attention computation is performed between the local summary vectors to obtain a globally enhanced summary. To enable information exchange between different feature blocks and achieve global context modeling, this invention performs a global self-attention computation on the sequence S composed of all local summary vectors. Since the length M of the summary sequence S is much smaller than the length L of the original sequence, the overhead of this global attention computation is very small.

[0072] This calculation allows each summary vector to incorporate information from all other summary vectors. For example, a summary vector representing the evening peak period can be correlated with a summary vector representing the early morning period, thus establishing causal inferences across long time scales. The calculation process can be represented by the following formula: ; in, To enhance the global summary, This represents the summary sequence. Each local summary vector in Each contains overall contextual information from all other feature blocks.

[0073] Here, the global augmented summary is used to reflect the core feature information of each time block and its correlation with the features of all other time blocks within the global context of the entire time series. Specifically, the original local summary vector only contains local information within its corresponding time block, while the global augmented summary generated after global attention calculation dynamically integrates contextual information from all other time blocks based on this local information through attention weights. In other words, the global augmented summary is a feature representation that embodies a global perspective and is enhanced by contextualization, effectively encoding long-distance dependencies across different time blocks.

[0074] Finally, the target features are obtained based on the global enhanced summary and the local enhanced features. This yields a global enhanced summary containing global contextual information. Then, this high-level global information needs to be "propagated" back to the original, more refined time steps to achieve the fusion of local and global features.

[0075] The specific implementation methods can be varied. For example, a simple and effective approach is to apply the local enhancement feature to each local enhancement feature belonging to the i-th block. All of these will have their corresponding global augmented summaries directly added to them. Other optional fusion methods include concatenating the local augmented features and the global augmented summaries and then projecting them through a linear layer; or designing a gating mechanism to dynamically control the fusion ratio of global and local information; or using a cross-attention mechanism, with local augmented features as queries and global augmented summaries as keys and values ​​for fusion. This embodiment of the invention does not specifically limit these methods.

[0076] Here, the feature vector at each time point in the target feature not only retains its fine local dependencies within a small time window, but also incorporates global contextual information throughout the entire time series, making it a feature with highly condensed information and extremely strong representational ability.

[0077] The method provided in this invention first efficiently captures details locally, then performs global information interaction at a low cost at the summary level, and finally empowers local details with the global view, thereby achieving high-quality global temporal dependency modeling within a controllable computational overhead.

[0078] Based on the above embodiments, the time-series dependency modeling module is specifically used for: Step 2101: Input the fused features into the dilated convolution module to obtain the intermediate temporal features; The dilated convolution module includes at least two dilated convolution layers connected in sequence; The dilation rate within the dilated convolutional layer increases sequentially along the feature propagation direction; the output of the dilated convolutional module is the sum of the residuals between the output of the last dilated convolutional layer and the fused features.

[0079] Specifically, in a particular implementation, to systematically expand the receptive field, the dilated convolution module contains multiple dilated convolutional layers, and the dilation rate of these layers increases sequentially along the feature propagation direction (i.e., the depth direction of the network). For example, a dilated convolution module can contain four dilated convolutional layers, with dilation rates set to 1, 2, 4, and 8 respectively. This design allows the receptive field to grow exponentially, enabling the model to cover a very long time range with fewer layers, thus efficiently capturing long-range dependencies.

[0080] Furthermore, to avoid training difficulties caused by vanishing or exploding gradients in the root cause localization model, this embodiment introduces a residual connection at the output of the dilated convolution module. Specifically, the final output of the dilated convolution module is not directly the output of the last dilated convolution layer, but rather the output of the last dilated convolution layer is added element-wise with the fused features, i.e., the residual sum. This design allows gradients to flow directly during backpropagation, thus ensuring stable training of deep networks and helping the root cause localization model learn identity mappings, avoiding performance degradation. Its calculation process can be expressed by the following formula: ; in, It is the first The input features of the layer, Conv1D represents a one-dimensional convolution operation, and k is the kernel size. For expansion rate, Represents the set of inflation rates, and ReLU is the activation function. This indicates a residual connection.

[0081] The method provided in this invention, by designing a dilated convolution module with an increasing dilation rate and residual connections, can not only systematically expand the receptive field to capture long-term dependencies, but also ensure the trainability of the root cause localization model through residual connections, thereby improving the quality of the final target features and the accuracy of root cause localization.

[0082] Based on the above embodiments, the multi-scale feature extraction module is specifically used for: Step 310: Extract short-term features at the first time scale, medium-term features at the second time scale, and long-term features at the third time scale from the time series data in parallel. Step 320: The short-term feature, the medium-term feature, and the long-term feature are fused to obtain the fused feature.

[0083] Specifically, Figure 3 This is a schematic diagram of the process by which the multi-scale feature extraction module provided by the present invention determines the fused features, as shown below. Figure 3 As shown, firstly, short-term features of the first time scale, medium-term features of the second time scale, and long-term features of the third time scale can be extracted in parallel from time series data.

[0084] The first time scale corresponds to the microscale and aims to capture short-term local patterns in the time series. Short-term features (also known as microscale features) mainly reflect the instantaneous fluctuations, noise, or brief abnormal spikes in the indicators.

[0085] The second time scale corresponds to the mesoscale and aims to capture trends and periodicity within a time series. Medium-term features (also known as mesoscale features) can reflect patterns such as the tidal effect between weekdays and weekends, and the periodic performance degradation during specific time periods (such as evening rush hour).

[0086] The third time scale corresponds to the macroscale and aims to capture the overall trend throughout the entire observation period. The long-term features extracted from it (also known as macroscale features) can reflect long-term states such as the slow deterioration or improvement of network performance.

[0087] Here, parallel extraction means that the extraction processes of these three different scale features are carried out simultaneously and independently, ensuring that the root cause localization model can learn information from its respective scales without interference.

[0088] Furthermore, after obtaining short-term, medium-term, and long-term features, these features can be fused together to obtain fused features.

[0089] The method provided in this invention extracts short-term features at a first time scale, medium-term features at a second time scale, and long-term features at a third time scale from time series data in parallel. Then, the short-term, medium-term, and long-term features are fused to obtain fused features. This ensures that the root cause localization model can comprehensively and without omission capture all time patterns of different granularities related to potential root causes from the time series data, greatly enhancing the feature representation ability and providing a rich and comprehensive information foundation for subsequent accurate root cause localization.

[0090] Based on the above embodiments, the multi-scale feature extraction module is specifically used for: Step 3101: Perform a first convolution process on the time series data to obtain the short-term features; Step 3102: Perform a second convolution process on the time series data to obtain the intermediate features; Step 3103: Perform mean pooling on the time series data to obtain pooled features, and perform a third convolution on the pooled features to obtain the long-term features.

[0091] Specifically, firstly, the time series data can be subjected to a first convolutional process to obtain short-term features, which can be achieved using a set of convolutional kernels with small sizes ( A one-dimensional convolutional layer ({3, 5, 7}) processes the input time series data X. The results of different convolutional kernels are processed by the learned weights. The weighted summation can be calculated using the following formula: ; in, X represents short-term features, X represents time series data, and k represents the convolution kernel size.

[0092] Extracting intermediate-scale features can be done using a set of convolutional kernels with moderate sizes. A one-dimensional convolutional layer ({15, 31, 47}) processes the input time series data X. The results of different convolutional kernels are also processed by the learned weights. The weighted summation can be calculated using the following formula: ; in, X represents the mid-term feature, X represents the time series data, and k represents the convolution kernel size.

[0093] To extract long-term features (macro-scale), the input time series data X is first subjected to mean pooling (Pool(X)) to obtain dimensionality-reduced pooled features. This operation is equivalent to low-pass filtering, preserving macro-trends and reducing computational cost. Then, a set of convolutional kernels with larger sizes ( The process involves processing a one-dimensional convolutional layer (={63,95}) and performing a weighted summation, the calculation of which can be represented by the following formula: ; in, X represents long-term features, k represents time series data, and k represents the kernel size.

[0094] In the above formula, the weights It uses the softmax function to apply the softmax function to a set of learnable parameters. The result is obtained by normalization to ensure that the sum of all weights is 1. The calculation process can be expressed by the following formula: ; Learnable parameters during training , This is an intermediate variable used for mathematical normalization, where the value of j corresponds to each size in the set of convolutional kernel sizes {3, 5, 7}. The root cause localization model can adaptively adjust the importance of each convolutional kernel in feature extraction at its corresponding scale.

[0095] The method provided in this invention offers a specific, adaptive, and computationally efficient multi-scale feature extraction architecture by designing different convolution operations for different time scales and introducing learnable fusion weights. This architecture can automatically learn the optimal feature combination based on data characteristics, making the extracted multi-scale features more targeted and representative.

[0096] The overall algorithm architecture of this invention comprises four main modules: a multi-scale feature extraction module, a temporal dependency modeling module, an efficient temporal attention module, and a dynamic classification module. Its forward propagation process can be summarized as follows: Input time-series data X (dimension B×T×F, where B is the batch size, T is the time step, and F is the number of features); obtain feature H1 through the multi-scale feature extraction module (MultiScaleConv); obtain fused feature H2 through feature fusion; obtain H3 through the temporal dependency modeling module (DilatedBlock); obtain H4 through the efficient temporal attention module (TemporalAttention); obtain H5 through global pooling (GlobalPool); and finally, output the final prediction result Ŷ, classification confidence score S, and adaptive classification threshold through the dynamic classification module (DynamicHead).

[0097] As a further explanation of the aforementioned overall algorithm architecture, the specific processing stages, tensor shape changes, and key operations of the temporal dependency modeling module (i.e., the process of obtaining H3 from H2) can be described in detail in Table 1 below: Table 1

[0098] Based on the above embodiments, the multi-scale feature extraction module is specifically used for: Step 3201: Adjust the short-term feature, the medium-term feature, and the long-term feature to the same time length; Step 3202: The short-term features, medium-term features, and long-term features, after being adjusted to the same time length, are spliced ​​along the feature dimension to obtain the fused features.

[0099] Specifically, firstly, the short-term, medium-term, and long-term features can be adjusted to the same time length.

[0100] Adjusting to the same time length can be achieved through various interpolation or sampling methods. For example, a target length L=48 can be preset, and then the short-term, medium-term, and long-term features at different scales can be adjusted to the same time length using the interpolation function Interp().

[0101] Once the feature sequences at all scales have the same time length, the splicing operation can be performed. The calculation process can be represented by the following formula: ; in, Indicates fusion features, Indicates short-term characteristics, Indicates mid-term characteristics, Indicates long-term characteristics.

[0102] For example, if the adjusted dimensions of the short-term, medium-term, and long-term features are [Batch, 48, 32], [Batch, 48, 32], and [Batch, 48, 32] respectively, then after concatenating along the feature dimensions, the result is... The dimensions are [Batch, 48, 96].

[0103] Furthermore, to further integrate information and adjust feature dimensions to match the input requirements of subsequent modules, the spliced ​​fused features can also be further refined. Perform a projection operation, which involves processing through a linear layer: ; in, This is a projection matrix, for example, with dimensions of 96×96. This is the bias vector. The final result, H2, is the fused feature.

[0104] Based on the above embodiments, the dynamic classification module is specifically used for: Step 410: Input the target features into the shared coding network for encoding to obtain the shared features output by the shared coding network; Step 420: Based on the shared features, generate the classification confidence score and the adaptive classification threshold for each of the preset root cause categories.

[0105] Specifically, Figure 4 This is a schematic diagram illustrating the process by which the dynamic classification module provided by this invention determines the classification confidence score and the adaptive classification threshold, as shown below. Figure 4As shown, after processing through the aforementioned temporal dependency modeling module and attention mechanism, a target feature H4 containing rich local and global temporal information can be obtained. The data dimension of the target feature is usually [B, L, D], where B is the batch size, L is the time length, and D is the feature dimension. For the final classification, a global average pooling layer can first be used to process the target feature in the time dimension, transforming its dimension from [B, L, D] to [B, D], resulting in a feature vector H5 that can represent the entire time series.

[0106] Subsequently, the feature vector H5 is input into the shared encoding network. The shared encoding network is a neural network module used by all predefined root cause categories in classification tasks. Specifically, this shared encoding network... It can be a two-layer perceptron (MLP), whose function is to perform further nonlinear transformation and information extraction on the input feature vector H5. Its calculation process can be represented by the following formula: ; in, , , , These are the weight matrix and bias vector of the two perceptron layers, respectively. For non-linear activation functions, such as ReLU; Indicates shared features.

[0107] The core idea behind the "shared" design is that while different pre-defined root cause categories represent different fault types, they may share potential commonalities and correlations at the feature level. By allowing all classification tasks to share the same shared encoding network, the root cause localization model can learn a more general and robust feature representation. This approach provides a unified feature foundation for subsequent score calculation and threshold generation, avoiding problems caused by mismatched dimensions between scores and thresholds. Furthermore, parameter sharing significantly reduces the total number of parameters in the root cause localization model, saving computational and storage resources, and facilitating knowledge transfer—that is, the knowledge gained by the root cause localization model when learning features of one category can assist in learning features of other categories.

[0108] After obtaining the shared features, the dynamic classification module uses them as a basis to generate a unique classification confidence score for each preset root cause category c (c=1,...,C) in parallel. and adaptive classification threshold .

[0109] Among them, the classification confidence score (Also known as the logit score) represents the probability that a sample belongs to that category, and its calculation process can be expressed by the following formula: ; in, and These are weight vectors and bias terms specific to the preset root cause category c; Indicates shared features.

[0110] The adaptive classification threshold is a specific threshold dynamically generated based on the input features of the current sample. Its calculation process can be expressed by the following formula: ; in, It is a weight vector specific to the preset root cause category c, and tanh is the hyperbolic tangent function. This is a scaling factor used to ensure that the generated threshold matches the logit score. The magnitude matches, based on experience, It can take the value 2.5; This represents the adaptive classification threshold.

[0111] The method provided in this invention inputs target features into a shared coding network for encoding, obtaining shared features output by the shared coding network. Based on these shared features, a classification confidence score and an adaptive classification threshold are generated for each preset root cause category. This not only improves the efficiency and generalization ability of the root cause localization model through parameter sharing, but also provides a unified and high-quality feature foundation for the subsequent generation of scores and thresholds for all root cause categories, thereby enhancing the stability and accuracy of the entire dynamic classification mechanism.

[0112] Based on the above embodiments, the root cause localization model is obtained by iteratively executing the following steps until a preset iteration termination condition is met: Step 510, obtain the training dataset; the training dataset includes time series data samples and the true values ​​of the multi-label root cause categories corresponding to the time series data samples; Step 520: Input the time series data sample into the initial root cause localization model to obtain the prediction score of each root cause category output by the initial root cause localization model; Step 530: Determine the target loss based on each of the predicted scores and the true values ​​of the multi-label root cause categories; Step 540: Update the model parameters of the initial root cause localization model based on the target loss.

[0113] Specifically, firstly, a training dataset can be obtained, which includes time series data samples and the true values ​​of the multi-label root cause categories corresponding to the time series data samples.

[0114] Here, the training dataset is a large collection of labeled data used to train and optimize the root cause localization model. The multi-label root cause category ground truths are standard answers obtained from domain experts or through historical work order analysis. These ground truths explicitly indicate the root cause of the network conditions reflected in the time series data samples. For example, a multi-label root cause category ground truth can be a multi-hot vector, such as [1, 0, 1, 0, 0, 0, 0, 0], indicating that the time series data sample simultaneously contains the first and third root cause categories.

[0115] Then, the time series data samples can be input into the initial root cause localization model to obtain the prediction score for each root cause category output by the initial root cause localization model. The prediction score reflects the confidence level of the root cause localization model in believing that the time series data sample belongs to that category.

[0116] Furthermore, the target loss can be determined based on the predicted scores and the true values ​​of the multi-label root cause categories. Commonly used target loss functions include binary cross-entropy loss, etc. It should be understood that the smaller the target loss value, the closer the prediction of the root cause localization model is to the reality.

[0117] Finally, the model parameters of the initial root cause localization model can be updated based on the target loss. Here, the preset iteration termination condition can be that the number of iterations reaches a preset maximum number of iterations, or that the model performance of the initial root cause localization model no longer improves significantly, etc., and this embodiment of the invention does not specifically limit this.

[0118] Based on the above embodiments, step 530 includes: Step 531: Determine the class weight of each root cause category in the preset root cause categories in the training dataset; Step 532: Determine the target loss based on each of the predicted scores, the true values ​​of the multi-label root cause categories, and the category weights.

[0119] Specifically, in real-world wireless network operation and maintenance, the frequency of occurrence of different root cause categories is often extremely uneven. To address this class imbalance problem, this embodiment introduces class weights when calculating the target loss.

[0120] The category weight is a numerical value assigned to each preset root cause category, used to adjust the contribution of that category in the target loss calculation. The core idea is to assign a higher weight to rarer categories with fewer samples, and a lower weight to common categories with more samples.

[0121] Specifically, this embodiment uses a weighted binary cross-entropy loss function as the optimization objective of the root cause localization model. That is, the target loss is determined based on each predicted score, the true value of the multi-label root cause category, and the category weight.

[0122] For each sample i and each root cause category c, the target loss is... The calculation process can be expressed by the following formula: ; Where N is the total number of samples in the batch; It is the true value of the multi-label root cause category, that is, the true value (0 or 1) of sample i in category c. The root cause localization model outputs the predicted score (logit score) for sample i in category c. It is the sigmoid function, used to convert the logit score into a probability between [0, 1]. It is the category weight specific to category c.

[0123] As can be seen from the above formula, for a positive sample (i.e. = 1), its contribution to the target loss is Weight This amplifies the loss term, causing the initial root cause localization model to be more penalized when it mispredicts rare positive samples, thus forcing the initial root cause localization model to focus more on learning these rare categories. For a negative sample (i.e. = 0), its loss term has a weight of 1, because there are usually a large number of negative samples, so no additional weighting is needed.

[0124] The method provided in this invention effectively solves the class imbalance problem in training data by introducing class weights. This method ensures that the initial root cause localization model gives sufficient attention to all root cause categories (whether common or rare) during training, thereby significantly improving the generalization ability of the root cause localization model in real operation and maintenance scenarios and the detection rate of rare faults, making the root cause localization results more comprehensive and reliable.

[0125] Based on the above embodiments, the category weights are determined based on the number of time series data samples corresponding to each root cause category in the training dataset that belong to the preset root cause categories.

[0126] Specifically, determining class weights is a data-driven, automated process, calculated based on the statistical distribution of each class in the training dataset. A common and effective method is to use a strategy that inversely proportional to the frequency of positive samples in each class. For root cause class c, its class weights are... The calculation process can be expressed by the following formula: ; Where N is the total number of samples in the training dataset; It is the number of samples in the training dataset with positive label c; It is a very small positive number (e.g., 1e-8) used to prevent the denominator from being zero; It is a clipping function that will take the input The value is limited to [ , Within the range of ]. and This is a preset clipping boundary used to limit the range of weights, preventing excessively large weights due to too few samples in one class, or excessively small weights due to too many samples in another class, thus causing training instability. In a specific implementation of this invention, it can be set... =0.02, =50.

[0127] The method provided in this invention determines weights by directly utilizing the statistical characteristics of the training data itself, avoiding the tediousness and subjectivity of manually setting weights. This adaptive weighting strategy based on sample size can accurately quantify and compensate for the imbalance between different categories.

[0128] The wireless network root cause localization device provided by the present invention is described below. The wireless network root cause localization device described below can be referred to in correspondence with the wireless network root cause localization method described above.

[0129] Based on any of the above embodiments, the present invention provides a wireless network root cause localization device. Figure 5 This is a schematic diagram of the structure of the wireless network root cause localization device provided by the present invention, as shown below. Figure 5 As shown, the device includes: The acquisition module 510 is used to acquire time-series data of the wireless network to be located. The input module 520 is used to input the time series data into the root cause localization model to obtain the root cause localization result output by the root cause localization model, wherein the root cause localization result is at least one root cause category of the time series data. The root cause localization model includes a multi-scale feature extraction module, a temporal dependency modeling module, and a dynamic classification module. The multi-scale feature extraction module is used to adaptively extract and fuse features from the time series data at different time scales to obtain fused features. The temporal dependency modeling module is used to determine target features based on long-distance dependencies in the fused features. The dynamic classification module is used to generate a classification confidence score and an adaptive classification threshold for each preset root cause category based on the target features, and to determine the root cause localization result based on the classification confidence score and the adaptive classification threshold.

[0130] The apparatus provided in this embodiment of the invention acquires time-series data of a wireless network to be located; inputs the time-series data into a root cause localization model to obtain the root cause localization result output by the root cause localization model. The root cause localization model includes a multi-scale feature extraction module, a temporal dependency modeling module, and a dynamic classification module. The multi-scale feature extraction module adaptively extracts and fuses features from different time scales in the time-series data to obtain fused features. The temporal dependency modeling module determines target features based on long-distance dependencies in the fused features. The dynamic classification module generates a classification confidence score and an adaptive classification threshold for each preset root cause category based on the target features, and determines the root cause localization result based on the classification confidence score and the adaptive classification threshold. This invention, through multi-scale feature extraction, can fuse features at different time granularities in time series data, avoiding information omissions caused by a single analysis scale. Furthermore, by modeling long-distance dependencies of the fused features, it can gain a deeper understanding of the dynamic evolution of faults. Finally, through a dynamic classification module, it generates adaptive classification thresholds for each sample, which can flexibly handle samples of varying complexity. This solves the problems of traditional methods relying on human experience, low efficiency, and the inability of traditional models to handle complex correlations and multi-root cause scenarios. It achieves accurate and automated location of wireless network faults, significantly improving the accuracy and robustness of root cause localization.

[0131] Based on any of the above embodiments, the time-series dependency modeling module specifically includes: The dilated convolution processing module is used to perform dilated convolution processing on the fused features to obtain intermediate temporal features; The partitioning module is used to divide the intermediate time-series features into multiple feature blocks in the time dimension; A local feature determination module is used to perform self-attention calculation within each feature block to obtain local enhanced features including local dependency information; The target feature determination module is used to determine the target feature based on multiple local enhancement features.

[0132] Based on any of the above embodiments, the target feature determination module is specifically used for: Each of the local enhancement features is pooled to obtain a local summary vector. Global attention computation is performed between the local summary vectors to obtain a globally enhanced summary. The target features are obtained based on the global enhanced summary and each of the local summary vectors.

[0133] Based on any of the above embodiments, the dilated convolution processing module is specifically used for: The fused features are input into the dilated convolution module to obtain the intermediate temporal features; The dilated convolution module includes at least two dilated convolution layers connected in sequence; The dilation rate within the dilated convolutional layer increases sequentially along the feature propagation direction; the output of the dilated convolutional module is the sum of the residuals between the output of the last dilated convolutional layer and the fused features.

[0134] Based on any of the above embodiments, the multi-scale feature extraction module specifically includes: The parallel extraction module is used to extract short-term features at the first time scale, medium-term features at the second time scale, and long-term features at the third time scale from the time series data in parallel. The fusion module is used to fuse the short-term features, the medium-term features, and the long-term features to obtain fused features.

[0135] Based on any of the above embodiments, the parallel extraction module is specifically used for: The time series data is subjected to a first convolution process to obtain the short-term features; The time series data is subjected to a second convolution process to obtain the intermediate features; The time series data is subjected to mean pooling to obtain pooled features, and the pooled features are subjected to a third convolution to obtain the long-term features.

[0136] Based on any of the above embodiments, the fusion module is specifically used for: Adjust the short-term, medium-term, and long-term features to the same time length; The short-term, medium-term, and long-term features, adjusted to the same time length, are then spliced ​​along the feature dimension to obtain the fused feature.

[0137] Based on any of the above embodiments, the dynamic classification module is specifically used for: The target features are input into a shared coding network for encoding to obtain the shared features output by the shared coding network; Based on the shared features, a classification confidence score and an adaptive classification threshold are generated for each of the preset root cause categories.

[0138] Based on any of the above embodiments, a training module is further included, wherein the training module specifically includes: The sample acquisition module is used to acquire the training dataset; the training dataset includes time series data samples and the true values ​​of the multi-label root cause categories corresponding to the time series data samples. The prediction score determination module is used to input the time series data samples into the initial root cause localization model to obtain the prediction score of each root cause category output by the initial root cause localization model; The target loss determination module is used to determine the target loss based on each of the predicted scores and the true values ​​of the multi-label root cause categories; The model parameter update module is used to update the model parameters of the initial root cause localization model based on the target loss.

[0139] Based on any of the above embodiments, the target loss determination module is specifically used for: Determine the class weight of each root cause category in the preset root cause categories in the training dataset; The target loss is determined based on the predicted scores, the true values ​​of the multi-label root cause categories, and the category weights.

[0140] Based on any of the above embodiments, the category weight is determined based on the number of time series data samples corresponding to each root cause category in the training dataset that belongs to the preset root cause categories.

[0141] Figure 6 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 6 As shown, the electronic device may include: a processor 610, a communications interface 620, a memory 630, and a communications bus 640, wherein the processor 610, the communications interface 620, and the memory 630 communicate with each other through the communications bus 640. The processor 610 can call logic instructions in the memory 630 to execute a wireless network root cause localization method. The method includes: acquiring time-series data of the wireless network to be located; inputting the time-series data into a root cause localization model to obtain a root cause localization result output by the root cause localization model, wherein the root cause localization result is at least one root cause category of the time-series data; the root cause localization model includes a multi-scale feature extraction module, a temporal dependency modeling module, and a dynamic classification module; the multi-scale feature extraction module is used to adaptively extract and fuse features from the time-series data at different time scales to obtain fused features; the temporal dependency modeling module is used to determine target features based on long-distance dependencies in the fused features; the dynamic classification module is used to generate a classification confidence score and an adaptive classification threshold for each preset root cause category based on the target features, and determine the root cause localization result based on the classification confidence score and the adaptive classification threshold.

[0142] Furthermore, the logical instructions in the aforementioned memory 630 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0143] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the wireless network root cause localization method provided by the above methods. The method includes: acquiring time-series data of the wireless network to be located; inputting the time-series data into a root cause localization model to obtain a root cause localization result output by the root cause localization model, wherein the root cause localization result is at least one root cause category of the time-series data; the root cause localization model includes a multi-scale feature extraction module, a temporal dependency modeling module, and a dynamic classification module; the multi-scale feature extraction module is used to adaptively extract and fuse features of different time scales from the time-series data to obtain fused features; the temporal dependency modeling module is used to determine target features based on long-distance dependencies in the fused features; the dynamic classification module is used to generate a classification confidence score and an adaptive classification threshold for each preset root cause category based on the target features, and determine the root cause localization result based on the classification confidence score and the adaptive classification threshold.

[0144] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the wireless network root cause localization method provided by the above methods. The method includes: acquiring time-series data of a wireless network to be localized; inputting the time-series data into a root cause localization model to obtain a root cause localization result output by the root cause localization model, wherein the root cause localization result is at least one root cause category of the time-series data; the root cause localization model includes a multi-scale feature extraction module, a temporal dependency modeling module, and a dynamic classification module; the multi-scale feature extraction module is used to adaptively extract and fuse features of different time scales from the time-series data to obtain fused features; the temporal dependency modeling module is used to determine target features based on long-distance dependencies in the fused features; the dynamic classification module is used to generate a classification confidence score and an adaptive classification threshold for each preset root cause category based on the target features, and to determine the root cause localization result based on the classification confidence score and the adaptive classification threshold.

[0145] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0146] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0147] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A wireless network root cause localization method, characterized in that, include: Acquire time-series data of the wireless network to be located; The time series data is input into the root cause localization model to obtain the root cause localization result output by the root cause localization model, wherein the root cause localization result is at least one root cause category of the time series data. The root cause localization model includes a multi-scale feature extraction module, a time-series dependency modeling module, and a dynamic classification module; the multi-scale feature extraction module is used to adaptively extract and fuse features from different time scales in the time series data to obtain fused features; The temporal dependency modeling module is used to determine target features based on the long-distance dependencies in the fused features; The dynamic classification module is used to generate a classification confidence score and an adaptive classification threshold for each preset root cause category based on the target features, and to determine the root cause localization result based on the classification confidence score and the adaptive classification threshold.

2. The wireless network root cause localization method of claim 1, wherein, The time-series dependency modeling module is specifically used for: The fused features are subjected to dilated convolution to obtain intermediate temporal features; The intermediate time-series features are divided into multiple feature blocks along the time dimension; Self-attention computation is performed within each feature block to obtain locally enhanced features that include local dependency information; The target feature is determined based on multiple local enhancement features.

3. The wireless network root cause localization method of claim 2, wherein, The time-series dependency modeling module is specifically used for: Each of the local enhancement features is pooled to obtain a local summary vector. Global attention computation is performed between the local summary vectors to obtain a globally enhanced summary. The target features are obtained based on the global enhanced summary and each of the local summary vectors.

4. The wireless network root cause localization method of claim 2, wherein, The time-series dependency modeling module is specifically used for: The fused features are input into the dilated convolution module to obtain the intermediate temporal features; The dilated convolution module includes at least two dilated convolution layers connected in sequence; The dilation rate within the dilated convolutional layer increases sequentially along the feature propagation direction; the output of the dilated convolutional module is the sum of the residuals between the output of the last dilated convolutional layer and the fused features.

5. The wireless network root cause localization method according to any one of claims 1 to 4, characterized in that, The multi-scale feature extraction module is specifically used for: The short-term features of the first time scale, the medium-term features of the second time scale, and the long-term features of the third time scale are extracted in parallel from the time series data. The short-term features, the medium-term features, and the long-term features are fused to obtain the fused features.

6. The wireless network root cause localization method according to claim 5, characterized in that, The multi-scale feature extraction module is specifically used for: The time series data is subjected to a first convolution process to obtain the short-term features; The time series data is subjected to a second convolution process to obtain the intermediate features; The time series data is subjected to mean pooling to obtain pooled features, and the pooled features are subjected to a third convolution to obtain the long-term features.

7. The wireless network root cause localization method according to claim 5, characterized in that, The multi-scale feature extraction module is specifically used for: Adjust the short-term, medium-term, and long-term features to the same time length; The short-term, medium-term, and long-term features, adjusted to the same time length, are then spliced ​​along the feature dimension to obtain the fused feature.

8. The wireless network root cause localization method according to any one of claims 1 to 4, characterized in that, The dynamic classification module is specifically used for: The target features are input into a shared coding network for encoding to obtain the shared features output by the shared coding network; Based on the shared features, a classification confidence score and an adaptive classification threshold are generated for each of the preset root cause categories.

9. The wireless network root cause localization method according to any one of claims 1 to 4, characterized in that, The root cause localization model is obtained by iteratively executing the following steps until a preset iteration termination condition is met: Obtain the training dataset; the training dataset includes time series data samples and the true values ​​of the multi-label root cause categories corresponding to the time series data samples; The time series data samples are input into the initial root cause localization model to obtain the predicted score for each root cause category output by the initial root cause localization model. The target loss is determined based on the predicted scores and the true values ​​of the multi-label root cause categories. The model parameters of the initial root cause localization model are updated based on the target loss.

10. The wireless network root cause localization method according to claim 9, characterized in that, The determination of the target loss based on each of the predicted scores and the true values ​​of the multi-label root cause categories includes: Determine the class weight of each root cause category in the preset root cause categories in the training dataset; The target loss is determined based on the predicted scores, the true values ​​of the multi-label root cause categories, and the category weights.

11. The wireless network root cause localization method according to claim 10, characterized in that, The category weights are determined based on the number of time-series data samples corresponding to each root cause category in the training dataset.

12. A wireless network root cause localization device, characterized in that, include: The acquisition module is used to acquire time-series data of the wireless network to be located. An input module is used to input the time series data into the root cause localization model to obtain the root cause localization result output by the root cause localization model, wherein the root cause localization result is at least one root cause category of the time series data. The root cause localization model includes a multi-scale feature extraction module, a time-series dependency modeling module, and a dynamic classification module; the multi-scale feature extraction module is used to adaptively extract and fuse features from different time scales in the time series data to obtain fused features; The temporal dependency modeling module is used to determine target features based on the long-distance dependencies in the fused features; The dynamic classification module is used to generate a classification confidence score and an adaptive classification threshold for each preset root cause category based on the target features, and to determine the root cause localization result based on the classification confidence score and the adaptive classification threshold.

13. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the wireless network root cause localization method as described in any one of claims 1 to 11.

14. A non-transitory 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 wireless network root cause localization method as described in any one of claims 1 to 11.

15. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the wireless network root cause localization method as described in any one of claims 1 to 11.

Images