Network quality poor root cause identification model training method and device, equipment and medium

Abstract

The embodiment of the application provides a network quality difference root cause identification model training method, device and equipment and medium, and belongs to the technical field of communication network operation and maintenance. The method comprises the following steps: extracting features of first network performance time series data through an initial encoder to obtain first communication quality difference features, and updating the initial encoder according to the first communication quality difference features and first quality difference root cause category labels to obtain a reference encoder; extracting features of second network performance time series data through the reference encoder to obtain second communication quality difference features, and determining a predicted quality difference root cause category of the second network performance time series data based on the second communication quality difference features through an initial classifier; calculating a cross-entropy loss according to the second quality difference root cause category label and the predicted quality difference root cause category, and updating the reference encoder and the classifier according to the cross-entropy loss to obtain a network quality difference root cause identification model. The embodiment of the application can improve the accuracy of network quality difference root cause identification.

Description

Technical Field

[0001] This application relates to the field of communication network operation and maintenance technology, and in particular to a training method, apparatus, equipment and medium for a network quality poor root cause identification model. Background Technology

[0002] In scenarios such as high-definition video, online gaming, connected vehicles, and industrial IoT, end users have higher requirements for low latency, high throughput, and high stability network services. Even slight fluctuations in network performance can lead to a significant decline in the perceived experience for end users. To ensure network service quality, it is necessary to identify the root causes of network experience degradation and optimize them quickly. Related technologies use fault diagnosis rules to identify the root causes of network quality degradation. However, these fault diagnosis rules rely on manual setting, making them unsuitable for new or complex scenarios. Furthermore, their diagnostic granularity is coarse, making it difficult to finely distinguish the types of root causes, resulting in low accuracy in root cause identification. Summary of the Invention

[0003] The main objective of this application is to propose a training method, apparatus, device, and medium for identifying the root causes of poor network quality, aiming to improve the accuracy of identifying the root causes of poor network quality.

[0004] To achieve the above objectives, a first aspect of this application proposes a training method for a network poor quality root cause identification model, the method comprising:

[0005] Obtain first network performance time series data; wherein, the first network performance time series data has a first poor quality root cause category label, and the first poor quality root cause category label is obtained by marking the first network performance time series data according to a preset evaluation rule;

[0006] The initial encoder extracts features from the first network performance time series data to obtain the first communication quality poor feature, and updates the initial encoder based on the first communication quality poor feature and the first quality poor root cause category label to obtain the reference encoder.

[0007] Obtain second network performance time series data; wherein, the second network performance time series data has a second poor quality root cause category label, and the second poor quality root cause category label is obtained by an expert model labeling the second network performance time series data;

[0008] The reference encoder extracts features from the second network performance time series data to obtain the second communication quality poor feature, and the initial classifier determines the predicted quality poor root cause category of the second network performance time series data based on the second communication quality poor feature.

[0009] The cross-entropy loss is calculated based on the second poor quality root cause category label and the predicted poor quality root cause category, and the reference encoder and the classifier are updated based on the cross-entropy loss to obtain the network poor quality root cause identification model; wherein, the network poor quality root cause identification model is used to identify the poor quality root causes of the target network performance time series data of the target communication network.

[0010] In some embodiments, the step of extracting features from the first network performance time-series data using an initial encoder to obtain a first communication quality defect feature includes:

[0011] The first network performance time-series data is augmented to obtain first augmented time-series data and second augmented time-series data; wherein the augmentation strength of the first augmented time-series data is less than the augmentation strength of the second augmented time-series data.

[0012] The first sub-quality difference feature of the first enhanced time series data is extracted through the initial encoder, and the second sub-quality difference feature of the second enhanced time series data is extracted through the initial encoder.

[0013] The first communication quality defect feature is determined based on the first sub-quality defect feature and the second sub-quality defect feature.

[0014] In some embodiments, the step of data augmentation of the first network performance time-series data to obtain first augmented time-series data and second augmented time-series data includes:

[0015] Calculate the time standard deviation based on the first network performance time series data;

[0016] Gaussian noise is obtained by sampling a first normal distribution; wherein the first normal distribution has a first standard deviation;

[0017] The Gaussian noise is scaled according to the time standard deviation to obtain the reference noise;

[0018] The reference noise is added to the first network performance time series data to obtain the first enhanced time series data;

[0019] Sampling is performed on a second normal distribution to obtain random multiplicative noise; wherein the second normal distribution has a second standard deviation, which is smaller than the first standard deviation;

[0020] The random multiplicative noise is added to the first network performance time series data to obtain the second enhanced time series data.

[0021] In some embodiments, the initial encoder includes a first convolutional block and a second convolutional block, both of which are one-dimensional convolutional blocks. Extracting the first sub-quality difference feature of the first enhanced temporal data through the initial encoder includes:

[0022] The first convolutional feature is obtained by performing convolution processing on the first enhanced temporal data using the first convolutional block.

[0023] The second convolutional feature is obtained by performing convolution processing on the first convolutional feature using the second convolutional block;

[0024] The second convolutional feature is flattened to obtain the convolutionally flattened feature;

[0025] The convolution flattening features are normalized to obtain the first sub-quality difference feature.

[0026] In some embodiments, the first convolutional block includes a temporal convolutional layer, a batch normalization layer, an activation layer, and a max pooling layer. The step of performing convolutional processing on the first augmented temporal data through the first convolutional block to obtain the first convolutional feature includes:

[0027] The first enhanced temporal data is temporally convolved by the temporal convolutional layer to obtain initial convolutional features;

[0028] The initial convolutional features are batch normalized through the batch normalization layer to obtain normalized features;

[0029] The normalized features are activated by the activation layer to obtain activated features;

[0030] The activation features are max-pooled through the max-pooling layer to obtain the first convolutional feature.

[0031] In some embodiments, updating the initial encoder based on the first communication quality defect feature and the first quality defect root cause category label to obtain a reference encoder includes:

[0032] For each first communication quality poor feature, other first communication quality poor features that have the same first quality poor root cause category label as the current first communication quality poor feature are obtained to obtain a positive sample set;

[0033] Obtain other first communication quality poor features besides the current first communication quality poor feature to obtain a negative sample set;

[0034] Comparative learning is performed based on each first communication quality poor feature, the corresponding positive sample set, and the corresponding negative sample set to obtain a comparative loss;

[0035] The initial encoder is updated based on the contrast loss to obtain the reference encoder.

[0036] To achieve the above objectives, a second aspect of this application proposes a method for identifying the root causes of poor network quality, the method comprising:

[0037] Acquire target network performance time-series data at the network protocol layer of the target communication network;

[0038] The target network performance time series data is used to identify the root causes of poor quality using a network poor quality root cause identification model to obtain the target poor quality root cause category; wherein, the network poor quality root cause identification model is trained according to the training method of the network poor quality root cause identification model described in the first aspect.

[0039] To achieve the above objectives, a third aspect of this application provides a network quality poor root cause identification device, the device comprising:

[0040] The acquisition module is used to acquire the target network performance time-series data at the network protocol layer of the target communication network;

[0041] The identification module is used to identify the root causes of poor network performance in the time series data of the target network using a network poor quality root cause identification model, and to obtain the target poor quality root cause category; wherein the network poor quality root cause identification model is trained according to the training method of the network poor quality root cause identification model described in the first aspect.

[0042] To achieve the above objectives, a fourth aspect of the present application provides an electronic device, the electronic device including a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method described in the first or second aspect.

[0043] To achieve the above objectives, a fifth aspect of the present application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described in the first or second aspect.

[0044] The training method, device, electronic device, and computer-readable storage medium for identifying the root causes of poor network quality proposed in this application acquire first network performance time-series data and first poor quality root cause category labels, and train the model based on the first network performance time-series data and the first poor quality root cause category labels. An initial encoder extracts features from the first network performance time-series data to obtain key features affecting the identification of poor quality root causes, and constructs a root cause discrimination space with a preliminary semantic structure to obtain first communication poor quality features. To ensure the accuracy of feature extraction, the initial encoder is updated based on the first communication poor quality features and the first poor quality root cause category labels to obtain a reference encoder. Because the first poor quality root cause category labels contain inherent noise, the feature space extracted based on the reference encoder contains pseudo-features, resulting in an inability to accurately distinguish different root cause categories and an inaccurate decision boundary. To solve this problem, second network performance time-series data and second poor quality root cause category labels are acquired for model updating, improving the semantic resolution of root cause analysis. Feature extraction is performed on the second network performance time-series data using a reference encoder to obtain a general representation of network quality degradation, resulting in the second communication quality deterioration feature. To accurately distinguish different types of root causes, a classifier is introduced. An initial classifier determines the predicted root cause category of the second network performance time-series data based on the second communication quality deterioration feature. Cross-entropy loss is calculated based on the second root cause category label and the predicted root cause category, and the reference encoder and classifier are updated according to the cross-entropy loss to obtain the network quality deterioration root cause identification model. Adjusting the reference encoder effectively eliminates residual semantic ambiguity, and adjusting the classifier enables automated root cause identification, solving the problem of insufficient identification accuracy caused by manually set rules and improving the accuracy of root cause identification. Attached Figure Description

[0045] Figure 1 This is a flowchart of the training method for the network poor quality root cause identification model provided in the embodiments of this application;

[0046] Figure 2 yes Figure 1 The flowchart of step S120 in the middle;

[0047] Figure 3 yes Figure 2 The flowchart of step S210 in the middle;

[0048] Figure 4 yes Figure 2 The flowchart of step S220 in the text;

[0049] Figure 5 yes Figure 4 The flowchart of step S410 in the middle;

[0050] Figure 6 yes Figure 1 Another flowchart of step S120 in the process;

[0051] Figure 7 This is a flowchart of the network quality poor root cause identification method provided in the embodiments of this application;

[0052] Figure 8 This is a diagram illustrating the effect of the training method for the network quality poor root cause identification model provided in the embodiments of this application;

[0053] Figure 9 This is another effect diagram of the training method of the network poor quality root cause identification model provided in the embodiments of this application;

[0054] Figure 10 This is a schematic diagram of the network quality poor root cause identification device provided in the embodiments of this application;

[0055] Figure 11 This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0057] It should be noted that although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., in the specification, claims, and the aforementioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0058] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0059] With the rapid development of mobile communication networks, end users are placing higher demands on network services with low latency, high throughput, and high stability. Especially in typical scenarios such as high-definition video, online gaming, connected vehicles, and the Industrial Internet of Things (IIoT), even slight fluctuations in network performance can lead to a significant decline in user experience. Therefore, to ensure network service quality, the ability to detect network degradation in real time, identify its root causes, and quickly optimize it is essential. Against this backdrop, root cause analysis has become a crucial component of intelligent operation and maintenance systems for communication networks.

[0060] Existing root cause analysis methods can be broadly categorized into two types: rule-based heuristic methods and machine learning-based supervised modeling methods. Rule-based root cause analysis typically requires pre-defined fault diagnosis rules, such as setting fixed thresholds for signal quality metrics like Signal-to-Interference plus Noise Ratio (SINR) and Reference Signal Received Power (RSRP), and then determining the fault category based on rule matching results. These methods offer the advantage of low deployment costs, but the rules rely on manual setting, making them unsuitable for novel or complex scenarios, and their diagnostic granularity is relatively coarse, making it difficult to finely distinguish root cause types. Machine learning-based supervised modeling methods identify root causes using deep neural network models based on supervised learning. These models heavily rely on large-scale, high-quality expert-annotated data. However, the cost of data annotation in communication networks is extremely high, and expert resources are limited, severely restricting the number of training samples and resulting in lower accuracy in root cause identification.

[0061] Based on this, embodiments of this application provide a training method for a network poor quality root cause identification model, a network poor quality root cause identification method, a network poor quality root cause identification device, an electronic device, and a computer-readable storage medium, aiming to improve the accuracy of network poor quality root cause identification.

[0062] The training method, identification device, electronic device, and computer-readable storage medium for identifying poor network quality root causes provided in this application are specifically described through the following embodiments. First, the training method for identifying poor network quality root causes in this application is described.

[0063] The training method for identifying the root causes of poor network quality provided in this application relates to the field of communication network operation and maintenance. This method can be applied to a terminal, a server, or software running on either a terminal or a server. In some embodiments, the terminal can be a smartphone, tablet, laptop, desktop computer, etc.; the server can be configured as an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms; the software can be an application implementing the training method for identifying the root causes of poor network quality, but is not limited to the above forms.

[0064] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0065] Figure 1 This is an optional flowchart of the training method for the network poor quality root cause identification model provided in the embodiments of this application. Figure 1 The method may include, but is not limited to, steps S110 to S150.

[0066] Step S110: Obtain first network performance time series data; wherein, the first network performance time series data has a first poor quality root cause category label, and the first poor quality root cause category label is obtained by marking the first network performance time series data according to a preset evaluation rule;

[0067] Step S120: Extract features from the first network performance time series data using the initial encoder to obtain the first communication quality poor feature, and update the initial encoder based on the first communication quality poor feature and the first quality poor root cause category label to obtain the reference encoder;

[0068] Step S130: Obtain second network performance time series data; wherein, the second network performance time series data has a second poor quality root cause category label, and the second poor quality root cause category label is obtained by the expert model labeling the second network performance time series data;

[0069] Step S140: The second network performance time series data is feature extracted by the reference encoder to obtain the second communication quality poor feature, and the predicted quality poor root cause category of the second network performance time series data is determined by the initial classifier based on the second communication quality poor feature.

[0070] Step S150: Calculate the cross-entropy loss based on the second poor quality root cause category label and the predicted poor quality root cause category, and update the reference encoder and classifier based on the cross-entropy loss to obtain the network poor quality root cause identification model; wherein, the network poor quality root cause identification model is used to identify the poor quality root causes of the target network performance time series data of the target communication network.

[0071] In step S110 of some embodiments, at fixed time intervals, such as 5 seconds, data acquisition modules deployed on the base station or core network side acquire various performance indicators (KPIs) of the communication network at different times at the network protocol layer, forming multi-dimensional time series data. This multi-dimensional time series data is then segmented by end user to form unit time series samples. These communication data samples exhibit characteristics such as cross-layer, multivariate, and time-series correlation. The network protocol layer includes the Physical Layer (PHY), Media Access Control Layer (MAC), Radio Link Control Layer (RLC), and Packet Data Convergence Layer (PDCP). The performance indicators of each network protocol layer are shown in Table 1.

[0072] Table 1

[0073]

[0074] The communication data samples, i.e., the KPI sequences, are divided in a 7:3 ratio to obtain the training set and the validation set. The KPI sequences are then extracted from the training set to obtain the first network performance time series data.

[0075] To ensure the data quality of the KPI sequences, preprocessing is required. Specifically, missing values ​​in the KPI sequences are filled using a bidirectional linear interpolation algorithm. If missing fields still exist after interpolation, the mode of all KPI sequences is used for filling. Outlier detection is performed on the KPI sequences; if missing fields still exist or interpolation fails after linear interpolation and global mode filling, the KPI sequence is discarded. To facilitate subsequent model training, the KPI sequences are aligned in length by extracting fixed-length sequence segments from each KPI sequence to unify the temporal input dimension. For example, the fixed length can be 40.

[0076] Tensor transformation is performed on all processed KPI sequences to improve computational efficiency and enhance model training performance by optimizing data representation. Specifically, all processed KPI sequences are transformed into shapes... The tensor is denoted by R, where R represents the real number space, N is the number of samples, m is the number of KPI channels (i.e., the dimension of the KPI field), and L is the uniform time window dimension (i.e., fixed length).

[0077] KPI sequences are labeled with root cause category tags for poor quality. These tags indicate the reasons for network service quality degradation, such as uplink interference, weak uplink coverage, downlink interference, and weak downlink coverage. The root cause category tags are obtained by labeling the KPI sequences according to preset evaluation rules. These preset evaluation rules are root cause judgment rules, consisting of a finite number of logical conditions, and the judgment logic covers multiple layers of KPI cross-constraints. The preset evaluation rules are existing rule templates or domain-specific heuristics in the communication system. Logical judgments are performed on the KPI sequences one by one according to the preset evaluation rules, automatically generating weak tags to obtain the root cause category tags for poor quality. By reusing existing system rules to generate weak tags, a large amount of precisely labeled data is not required, which can adapt to the reality of scarce tags in communication networks and improve the semantic guidance capability of KPI sequences. The first network performance time series data has a first root cause category tag for poor quality, which is the root cause category tag for poor quality in the first network performance time series data. The first root cause category tag for poor quality is obtained by labeling the first network performance time series data according to preset evaluation rules.

[0078] This application's embodiments consider six different types of root causes, namely... This includes uplink interference, weak uplink coverage, downlink interference, weak downlink coverage, traffic channel load, and control channel load. A dataset with rule labels can be obtained based on preset evaluation rules, denoted as […]. The preset evaluation rules are shown in the table below:

[0079] Table 2

[0080]

[0081] Please see Figure 2 In some embodiments, step S120 may include, but is not limited to, steps S210 to S230:

[0082] Step S210: Perform data augmentation on the first network performance time series data to obtain first augmented time series data and second augmented time series data; wherein, the augmentation strength of the first augmented time series data is less than the augmentation strength of the second augmented time series data.

[0083] Step S220: Extract the first sub-quality difference feature of the first enhanced time series data through the initial encoder, and extract the second sub-quality difference feature of the second enhanced time series data through the initial encoder;

[0084] Step S230: Determine the first communication quality difference feature based on the first sub-quality difference feature and the second sub-quality difference feature.

[0085] In step S210 of some embodiments, after constructing the KPI time-series samples into a two-dimensional matrix, two transformations are performed on each KPI time-series sample, i.e., the first network performance time-series data, to obtain semantically consistent but differently distributed enhanced views, resulting in first enhanced time-series data and second enhanced time-series data. The enhancement strength of the first enhanced time-series data is less than that of the second enhanced time-series data; the enhancement strength of the first enhanced time-series data is weak, while the enhancement strength of the second enhanced time-series data is strong. Data augmentation methods can include data jitter, field reorganization, perturbation enhancement, etc.

[0086] The first network performance time series data is represented as follows: Where R is the real number space and m is the number of KPI channels. Given the sequence length, the first and second augmented time-series data can be represented as follows: .

[0087] In step S220 of some embodiments, the deep neural network model structure is relatively large, with high computational complexity, and the model interpretability depends on fuzzy rule design, increasing the difficulty of system engineering implementation. To improve model training efficiency and reduce the difficulty of system engineering implementation, this embodiment uses a lightweight convolutional neural network as the initial encoder, which is used to extract communication quality defects features in time series form. The enhanced KPI sequence is used as input, and the shared initial encoder encodes the first and second enhanced time series data respectively to obtain low-dimensional embedding representations, resulting in the first sub-quality defects feature corresponding to the first enhanced time series data and the second sub-quality defects feature corresponding to the second enhanced time series data. The feature extraction process of the initial encoder for the first and second enhanced time series data is the same.

[0088] In step S230 of some embodiments, the first sub-quality difference feature and the second sub-quality difference feature are used as the first communication quality difference feature.

[0089] Steps S210 to S230 above, by performing data augmentation on the KPI sequence to different degrees, enable the model to learn core quality difference features with invariance and robustness, thereby constructing an accurate semantic feature space and improving the accuracy of root cause identification.

[0090] Please see Figure 3 In some embodiments, step S210 may include, but is not limited to, steps S310 to S360:

[0091] Step S310: Calculate the time standard deviation based on the first network performance time series data;

[0092] Step S320: Sample the first normal distribution to obtain Gaussian noise; wherein the first normal distribution has a first standard deviation;

[0093] Step S330: Scaling the Gaussian noise according to the time standard deviation to obtain the reference noise;

[0094] Step S340: Add reference noise to the first network performance time series data to obtain the first enhanced time series data;

[0095] Step S350: Sample the second normal distribution to obtain random multiplicative noise; wherein the second normal distribution has a second standard deviation, which is less than the first standard deviation;

[0096] Step S360: Add random multiplicative noise to the first network performance time series data to obtain the second enhanced time series data.

[0097] In step S310 of some embodiments, the time standard deviation is the standard deviation of the first network performance time series data over the time channels. Specifically, for each KPI channel of the first network performance time series data, the mean of the index values ​​of the first network performance time series data over each time channel is calculated to obtain the time-series average value of the KPI channel. The standard deviation is calculated based on the index values ​​and time-series average value of the first network performance time series data over each time channel to obtain the time standard deviation of the KPI channel. The formula for calculating the time standard deviation is expressed as:

[0098] ,

[0099] ,

[0100] Where j represents the j-th KPI channel; This represents the t-th time channel; This represents the sequence length, i.e., the number of time channels. This represents the time-series average value of the j-th KPI channel; This represents the time standard deviation of the j-th KPI channel; This represents the index value of the KPI sequence in the j-th KPI channel and the t-th time channel.

[0101] In step S320 of some embodiments, the weak enhancement is achieved by injecting Gaussian noise scaled by the time standard deviation into each KPI channel. Specifically, for each KPI channel of the first network performance time series data, a first normal distribution is sampled to obtain the Gaussian noise corresponding to the KPI channel. The first normal distribution is a standard normal distribution, denoted as... ,in Let be the Gaussian noise corresponding to the j-th KPI channel. The first normal distribution has a first standard deviation of 1.

[0102] In step S330 of some embodiments, the time standard deviation of the KPI channel is multiplied by the Gaussian noise to scale the Gaussian noise by the time standard deviation, so as to obtain the reference noise corresponding to the KPI channel.

[0103] In step S340 of some embodiments, for each KPI channel of the first network performance time series data, corresponding reference noise is added to the KPI channel to obtain the first enhanced time series data. The weak enhancement process can be represented as:

[0104] ,

[0105] in, This represents the index value of the first enhanced time series data in the j-th KPI channel. , Number of KPI channels; This represents the index value of the KPI series in the j-th KPI channel; and These represent the Gaussian noise and time standard deviation corresponding to the j-th KPI channel, respectively; + indicates addition. It represents the multiplication operation.

[0106] In step S350 of some embodiments, strong enhancement simulates intensity fluctuations by applying random multiplicative scaling to each KPI channel. For each KPI channel of the first network performance time series data, a second normal distribution is sampled to obtain the random multiplicative noise corresponding to the KPI channel. The second normal distribution has a second standard deviation. To ensure that the enhancement intensity of strong enhancement is greater than that of weak enhancement, the second standard deviation is set to be less than the first standard deviation. The second standard deviation is a fixed hyperparameter used to control the intensity of the scaling noise, and its value range is typically 0.01-0.2. The random multiplicative noise can be expressed as:

[0107] ,

[0108] in, This represents the random multiplicative noise of the j-th KPI channel; Indicates the second normal distribution; This represents the second standard deviation.

[0109] In step S360 of some embodiments, random multiplicative noise is added to each KPI channel of the first network performance time series data to obtain the second enhanced time series data. The process of strong enhancement can be represented as follows:

[0110] ,

[0111] in, This represents the index value of the second enhanced time series data in the j-th KPI channel; This represents the index value of the KPI series in the j-th KPI channel; This represents the random multiplicative noise of the j-th KPI channel; It represents the multiplication operation.

[0112] Steps S310 to S360 above, through two different forms of data perturbation, can maintain the root cause semantics in communication degradation scenarios, such as the decrease in SINR caused by interference, while enhancing robustness to noise.

[0113] Please see Figure 4 In some embodiments, the initial encoder includes a first convolutional block and a second convolutional block, which are cascaded. Considering that most poor-quality patterns exhibit local temporal variations, one-dimensional convolutional blocks are better at extracting trends within short time windows and modeling local temporal dependent structures compared to fully connected or recurrent structures. Therefore, both the first and second convolutional blocks are one-dimensional convolutional blocks. Step S220 may include, but is not limited to, steps S410 to S440:

[0114] Step S410: Perform convolution processing on the first enhanced temporal data using the first convolution block to obtain the first convolution feature;

[0115] Step S420: Perform convolution processing on the first convolution feature using the second convolution block to obtain the second convolution feature;

[0116] Step S430: Flatten the second convolutional feature to obtain the flattened convolutional feature;

[0117] Step S440: Normalize the convolution flattened features to obtain the first sub-prime difference feature.

[0118] In step S410 of some embodiments, the first enhanced temporal data is input into the first convolutional block for convolution processing to obtain the first convolutional feature.

[0119] In step S420 of some embodiments, the first convolutional feature is input into the second convolutional block for convolution processing to obtain the second convolutional feature. The network structure of the second convolutional block is the same as that of the first convolutional block. The second convolutional feature is represented as:

[0120] ,

[0121] in, Indicates the second convolution feature. R represents the real number space. Indicates the number of feature channels. Indicates the sequence length; Indicates the initial encoder; This represents the first enhanced time series data; Represents the first convolutional block; This represents the second convolutional block.

[0122] In step S430 of some embodiments, to facilitate subsequent root cause identification, the second convolutional feature output by the convolutional network is flattened into a vector, resulting in the convolutionally flattened feature. The convolutionally flattened feature is represented as... R represents the real number space. Represents the vector dimension.

[0123] In step S440 of some embodiments, to accelerate the convergence speed of model training, the convolution flattening features are normalized to obtain the first sub-quality difference feature. The normalization calculation formula is expressed as:

[0124] ,

[0125] in, Indicates the convolution flattening feature; Indicates the first sub-prime difference characteristic; This represents the L2 norm.

[0126] Steps S410 to S440 above utilize the ability of one-dimensional convolutional blocks to model local time-dependent structures, enabling accurate extraction of poor communication quality features from KPI time series sequences.

[0127] Please see Figure 5 In some embodiments, the first convolutional block includes a temporal convolutional layer, a batch normalization layer, an activation layer, and a max pooling layer. Step S410 may include, but is not limited to, steps S510 to S540:

[0128] Step S510: Perform temporal convolution on the first enhanced temporal data through a temporal convolutional layer to obtain initial convolutional features;

[0129] Step S520: The initial convolutional features are batch normalized through a batch normalization layer to obtain normalized features;

[0130] Step S530: The normalized features are activated by the activation layer to obtain the activated features;

[0131] Step S540: Max pool the activation features through a max pooling layer to obtain the first convolutional features.

[0132] In step S510 of some embodiments, the temporal convolutional layer is provided with a temporal convolutional kernel. The temporal convolutional kernel is slid along the temporal dimension of the first enhanced temporal data to perform convolution operation on the first enhanced temporal data in order to extract local temporal patterns and obtain initial convolutional features.

[0133] In step S520 of some embodiments, batch normalization is a commonly used technique in deep learning, designed to accelerate the training of neural networks and improve convergence speed. By batch normalizing the initial convolutional features through a batch normalization layer, normalized features are obtained, thereby making the data distribution more stable and avoiding gradient vanishing or gradient exploding problems.

[0134] In step S530 of some embodiments, the activation layer may use the ReLU activation function to activate the normalized features and obtain the activated features.

[0135] In step S540 of some embodiments, max pooling reduces the spatial dimension of the feature map by selecting the maximum value of a local feature region, thereby reducing computational cost while retaining the most salient features. Max pooling is then applied to the activation features through a max pooling layer to obtain the first convolutional features.

[0136] Through the above steps S510 to S540, local time-series patterns hidden in the KPI time-series sequence can be extracted.

[0137] Please see Figure 6 In some embodiments, step S120 may include, but is not limited to, steps S610 to S640:

[0138] Step S610: For each first communication quality poor feature, obtain other first communication quality poor features that have the same first quality poor root cause category label as the current first communication quality poor feature, and obtain a positive sample set;

[0139] Step S620: Obtain other first communication quality poor features besides the current first communication quality poor feature to obtain a negative sample set;

[0140] Step S630: Perform comparative learning based on each first communication quality poor feature, the corresponding positive sample set, and the corresponding negative sample set to obtain the comparative loss;

[0141] Step S640: Update the initial encoder based on the contrast loss to obtain the reference encoder.

[0142] In step S610 of some embodiments, in order to cluster samples with similar quality defects in the latent space, different view samples sharing the same rule label are used as a positive sample set. The first communication quality defect feature includes sub-quality defect features, and both the first sub-quality defect feature and the second sub-quality defect feature can be regarded as independent sub-quality defect features. Since the first sub-quality defect feature and the second sub-quality defect feature are obtained by feature extraction based on the first network performance time series data, the first sub-quality defect feature and the second sub-quality defect feature also share the first quality defect root cause category label of the first network performance time series data. For each sub-quality defect feature, other sub-quality defect features with the same first quality defect root cause category label as the current sub-quality defect feature are obtained, and the positive sample set corresponding to the current sub-quality defect feature is obtained.

[0143] In step S620 of some embodiments, other sub-prime features besides the current sub-prime feature are obtained to obtain a negative sample set of the current sub-prime feature.

[0144] In step S630 of some embodiments, contrastive learning is performed based on each sub-poor quality feature, its corresponding positive sample set, and its corresponding negative sample set to obtain a contrastive loss. Given a batch containing B samples, the number of sub-poor quality features is 2B, and the contrastive loss is defined as:

[0145] ,

[0146] in, Indicates comparative loss; Indicates the first Individual quality difference characteristics; T represents the transpose operation; Indicates the first A positive sample set with poor individual quality characteristics; Indicates the number of samples in the positive sample set; Denotes the first positive sample in the positive sample set. One sample; Indicates that, except for the first in the batch All other sub-prime features besides the individual sub-prime features, i.e., the first Negative sample set of individual quality poor characteristics; Represents the negative sample set as the first One sample; This refers to temperature hyperparameters.

[0147] In step S640 of some embodiments, the contrast loss is minimized, the model parameters of the initial encoder are updated, and a reference encoder is obtained.

[0148] Steps S610 to S640 above involve constructing a supervised contrastive learning loss function based on rule labels to conduct supervised contrastive training on the enhanced sample pairs. This ensures that the embeddings of similar root cause samples are close to each other, while the embeddings of dissimilar root cause samples are far apart. The optimized embedding space exhibits semantic clustering characteristics, thereby constructing a root cause discrimination space with a preliminary semantic structure.

[0149] In step S130 of some embodiments, the feature space obtained based on contrastive learning may contain spurious features due to the inherent noise carried by the rule labels, making the decision boundary insufficiently accurate when identifying root causes. This embodiment of the application fine-tunes the model based on a small number of expert-annotated samples to improve the semantic resolution of root cause analysis. Specifically, a small number of KPI sequences are randomly extracted from multiple first network performance time-series data to obtain second network performance time-series data. Considering constraints such as high annotation costs and limited expert resources, the number of samples in the second network performance time-series data is much smaller than that in the first network performance time-series data. The second network performance time-series data is presented to an expert model, and the business knowledge possessed by the expert model is used to label the second network performance time-series data with root cause types to generate expert annotation labels, resulting in second poor-quality root cause category labels. The second poor-quality root cause category labels are the poor-quality root cause category labels of the second network performance time-series data. The accuracy of the second poor-quality root cause category labels is higher than that of the first poor-quality root cause category labels, and they can be used for model fine-tuning and decision boundary correction.

[0150] A dataset with expert labels can be represented as , Indicates the first One input sample, Indicates the first Expert annotation labels for each input sample. The number of expert-annotated samples is determined. Through comparative learning, the reference encoder has learned a general representation of network quality degradation patterns. Therefore, at this stage, only a small amount of highly reliable expert-annotated data is needed for model fine-tuning to effectively eliminate residual semantic ambiguity. Especially in scenarios with similar root cause types (such as uplink interference and weak uplink coverage), the two may exhibit similar KPI dynamics under rule labels. However, through calibration with expert-annotated data, different types of root causes can be further distinguished more accurately.

[0151] In step S140 of some embodiments, feature extraction is performed on the second network performance time-series data using a reference encoder to extract key features from the KPI sequence, thereby obtaining the second communication quality poor feature. The feature extraction process can be referred to steps S410 to S440, and will not be repeated here. A lightweight classifier, implemented as a fully connected layer, is added to the pre-trained encoder. The initial classifier, i.e., the lightweight classifier, classifies the second communication quality poor feature, predicts the root cause category of the second network performance time-series data, and obtains the predicted root cause category of the poor quality.

[0152] In step S150 of some embodiments, the pre-trained embedding space is fine-tuned using expert-annotated samples. The fine-tuning stage uses a cross-entropy loss function to further refine the discrimination boundaries between different categories of root causes. Specifically, the cross-entropy loss is calculated based on the second poor-quality root cause category label and the predicted poor-quality root cause category. The cross-entropy loss is minimized, and the model parameters of the reference encoder and the classifier are updated synchronously to obtain the network poor-quality root cause identification model. This model is used to identify poor-quality root causes in the time-series data of the target communication network's performance. The formula for calculating the cross-entropy loss is expressed as:

[0153] ,

[0154] in, Represents cross-entropy loss; C represents the number of expert-annotated samples; C represents the number of root cause categories. The true label of the i-th sample indicates whether the i-th sample belongs to root cause category j; This represents the probability that the classifier predicts the i-th sample belongs to root cause category j.

[0155] Figure 7 This is an optional flowchart of the network quality poor root cause identification method provided in the embodiments of this application. Figure 7 The method may include, but is not limited to, steps S710 to S720.

[0156] Step S710: Obtain the target network performance time-series data at the network protocol layer of the target communication network;

[0157] Step S720: The network performance time series data is used to identify the root causes of poor quality through the network poor quality root cause identification model to obtain the target poor quality root cause category; wherein, the network poor quality root cause identification model is trained according to the above-mentioned training method of the network poor quality root cause identification model.

[0158] In step S710 of some embodiments, the target communication network is the mobile communication network to be analyzed for root causes. Following a fixed sampling period, such as 5 seconds, data acquisition modules deployed at the base station or core network side acquire various performance indicators of the target communication network at different times within the network protocol layer. The received continuous KPI segments are then concatenated into a fixed-dimensional matrix to obtain the target network performance time-series data. To ensure the accuracy of root cause identification, preprocessing and tensor transformation operations can be performed on the target network performance time-series data, as described in step S110.

[0159] In step S720 of some embodiments, after model training is completed, the network quality deterioration root cause identification model can be deployed on an edge server or network management system for real-time detection and diagnosis of communication quality degradation. The deployed model receives new network performance time-series data through a standardized interface. The network quality deterioration root cause identification model includes a fine-tuned reference encoder and a classifier. The target network performance time-series data is input to the reference encoder, and its embedding representation vector in the latent space is automatically extracted to obtain communication quality deterioration features. The classifier predicts the probability that the target network performance time-series data belongs to each quality deterioration root cause category based on the communication quality deterioration features, and the quality deterioration root cause category with the highest probability is taken as the target quality deterioration root cause category. The target quality deterioration root cause category is used to indicate the root cause of communication degradation and can be any one of uplink interference, weak uplink coverage, downlink interference, weak downlink coverage, traffic channel overload, and control channel overload.

[0160] Through the above steps S710 to S720, the root cause of network service quality degradation can be accurately identified.

[0161] This application proposes a semi-supervised root cause analysis method based on rule-label-guided contrastive learning and expert fine-tuning. This method can be used for intelligent root cause identification of poor mobile communication network quality. By introducing rule labels as explicit semantic supervision signals for contrastive learning, an embedding space with root cause semantic structure is constructed from the training stage. This effectively distinguishes highly similar but semantically different KPI patterns, aiming to construct a discriminative feature space with inter-class semantic structure under weak labeling conditions. It does not require a large amount of precisely labeled data and can generate weak labels based on existing rules within the system. Furthermore, a small number of expert samples are used for fine-tuning the decision boundary, significantly improving the model's usability, deployability, and generalization under limited labeling. Simultaneously, a lightweight convolutional network structure and a two-stage training mechanism are employed, balancing model performance and system complexity. This method is suitable for deployment on edge nodes or lightweight operation and maintenance systems, improving system feasibility and deployment efficiency. This method not only enhances the ability to identify root cause patterns in multi-source KPI sequences of communication networks but also possesses stronger robustness and engineering implementation potential.

[0162] Please see Figure 8 and Figure 9The performance of the embodiments of this application in the root cause analysis of poor quality data in real communication network data was tested. The experimental results show that the method can achieve high accuracy in top1 and top2 root cause analysis.

[0163] Please see Figure 10 This application also provides a network quality poor root cause identification device, which can implement the above-mentioned network quality poor root cause identification method. The network quality poor root cause identification device includes:

[0164] The acquisition module 1010 is used to acquire the target network performance time-series data of the target communication network at the network protocol layer;

[0165] The identification module 1020 is used to identify the root causes of poor network performance in the time series data of the target network through the network poor quality root cause identification model, and to obtain the target poor quality root cause category; wherein, the network poor quality root cause identification model is trained according to the above-mentioned network poor quality root cause identification model training method.

[0166] The specific implementation of this network poor quality root cause identification device is basically the same as the specific implementation of the network poor quality root cause identification method described above, and will not be repeated here.

[0167] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the training method or the network quality poor root cause identification method described above. This electronic device can be any smart terminal, including tablet computers, in-vehicle computers, etc.

[0168] Please see Figure 11 , Figure 11 The hardware structure of an electronic device according to another embodiment is illustrated. The electronic device includes:

[0169] The processor 1110 can be implemented using a general-purpose central processing unit (CPU), microprocessor, application specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application.

[0170] The memory 1120 can be implemented as a read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory 1120 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 1120, and the processor 1110 calls and executes the training method or the network quality poor root cause identification method of the embodiments of this application.

[0171] The input / output interface 1130 is used to implement information input and output;

[0172] The communication interface 1140 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, network cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0173] Bus 1150 transmits information between various components of the device (e.g., processor 1110, memory 1120, input / output interface 1130, and communication interface 1140);

[0174] The processor 1110, memory 1120, input / output interface 1130 and communication interface 1140 are connected to each other within the device via bus 1150.

[0175] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described training method or network poor quality root cause identification method for the network poor quality root cause identification model.

[0176] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0177] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0178] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0179] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; 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.

[0180] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0181] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0182] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0183] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0184] The units described above 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 units can be selected to achieve the purpose of this embodiment according to actual needs.

[0185] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0186] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or 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 multiple 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 of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0187] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A training method for a network quality poor root cause identification model, characterized in that, The method includes: Obtain first network performance time series data; wherein, the first network performance time series data has a first poor quality root cause category label, and the first poor quality root cause category label is obtained by marking the first network performance time series data according to a preset evaluation rule; The initial encoder extracts features from the first network performance time series data to obtain the first communication quality poor feature, and updates the initial encoder based on the first communication quality poor feature and the first quality poor root cause category label to obtain the reference encoder. Obtain second network performance time series data; wherein, the second network performance time series data has a second poor quality root cause category label, and the second poor quality root cause category label is obtained by an expert model labeling the second network performance time series data; The reference encoder extracts features from the second network performance time series data to obtain the second communication quality poor feature, and the initial classifier determines the predicted quality poor root cause category of the second network performance time series data based on the second communication quality poor feature. The cross-entropy loss is calculated based on the second poor quality root cause category label and the predicted poor quality root cause category, and the reference encoder and the classifier are updated based on the cross-entropy loss to obtain the network poor quality root cause identification model; wherein, the network poor quality root cause identification model is used to identify the poor quality root causes of the target network performance time series data of the target communication network.

2. The method according to claim 1, characterized in that, The step of extracting features from the first network performance time-series data using an initial encoder to obtain the first communication quality poor feature includes: The first network performance time-series data is augmented to obtain first augmented time-series data and second augmented time-series data; wherein the augmentation strength of the first augmented time-series data is less than the augmentation strength of the second augmented time-series data. The first sub-quality difference feature of the first enhanced time series data is extracted through the initial encoder, and the second sub-quality difference feature of the second enhanced time series data is extracted through the initial encoder. The first communication quality defect feature is determined based on the first sub-quality defect feature and the second sub-quality defect feature.

3. The method according to claim 2, characterized in that, The step of augmenting the first network performance time-series data to obtain first augmented time-series data and second augmented time-series data includes: Calculate the time standard deviation based on the first network performance time series data; Gaussian noise is obtained by sampling a first normal distribution; wherein the first normal distribution has a first standard deviation; The Gaussian noise is scaled according to the time standard deviation to obtain the reference noise; The reference noise is added to the first network performance time series data to obtain the first enhanced time series data; Sampling is performed on a second normal distribution to obtain random multiplicative noise; wherein the second normal distribution has a second standard deviation, which is smaller than the first standard deviation; The random multiplicative noise is added to the first network performance time series data to obtain the second enhanced time series data.

4. The method according to claim 2, characterized in that, The initial encoder includes a first convolutional block and a second convolutional block, both of which are one-dimensional convolutional blocks. Extracting the first sub-quality difference feature from the first enhanced temporal data using the initial encoder includes: The first convolutional feature is obtained by performing convolution processing on the first enhanced temporal data using the first convolutional block. The second convolutional feature is obtained by performing convolution processing on the first convolutional feature using the second convolutional block; The second convolutional feature is flattened to obtain the convolutionally flattened feature; The convolution flattening features are normalized to obtain the first sub-quality difference feature.

5. The method according to claim 4, characterized in that, The first convolutional block includes a temporal convolutional layer, a batch normalization layer, an activation layer, and a max pooling layer. The first convolutional feature is obtained by performing convolutional processing on the first enhanced temporal data through the first convolutional block. The first enhanced temporal data is temporally convolved by the temporal convolutional layer to obtain initial convolutional features; The initial convolutional features are batch normalized through the batch normalization layer to obtain normalized features; The normalized features are activated by the activation layer to obtain activated features; The activation features are max-pooled through the max-pooling layer to obtain the first convolutional feature.

6. The method according to any one of claims 1 to 5, characterized in that, The step of updating the initial encoder based on the first communication quality defect feature and the first quality defect root cause category label to obtain a reference encoder includes: For each first communication quality poor feature, other first communication quality poor features that have the same first quality poor root cause category label as the current first communication quality poor feature are obtained to obtain a positive sample set; Obtain other first communication quality poor features besides the current first communication quality poor feature to obtain a negative sample set; Comparative learning is performed based on each first communication quality poor feature, the corresponding positive sample set, and the corresponding negative sample set to obtain a comparative loss; The initial encoder is updated based on the contrast loss to obtain the reference encoder.

7. A method for identifying the root causes of poor network quality, characterized in that, The method includes: Acquire target network performance time-series data at the network protocol layer of the target communication network; The target network performance time series data is used to identify the root causes of poor quality using a network poor quality root cause identification model to obtain the target poor quality root cause category; wherein, the network poor quality root cause identification model is trained using the training method of the network poor quality root cause identification model according to any one of claims 1 to 6.

8. A device for identifying the root causes of poor network quality, characterized in that, The device includes: The acquisition module is used to acquire the target network performance time-series data at the network protocol layer of the target communication network; The identification module is used to identify the root causes of poor network performance in the time series data of the target network using a network poor quality root cause identification model, and to obtain the target poor quality root cause category; wherein the network poor quality root cause identification model is trained by the training method of the network poor quality root cause identification model according to any one of claims 1 to 6.

9. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method of any one of claims 1 to 6 or the method of claim 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 6 or the method of claim 7.

Images