Broadband network operation and maintenance method and device

Abstract

The invention discloses a broadband network operation and maintenance method and device. The method comprises the following steps: collecting multi-dimensional network data of a user to be evaluated; a pre-constructed broadband quality evaluation model is adopted to evaluate the multi-dimensional network data to obtain an evaluation result, the broadband quality evaluation model comprises a multi-head sparse self-attention sub-layer, the evaluation result is used for representing the broadband quality of the to-be-evaluated user, and the evaluation result is determined based on a score threshold value of the target moment; the score thresholds at different moments are different; under the condition that the evaluation result indicates that the broadband quality of the to-be-evaluated user is abnormal, obtaining a sparse attention weight matrix corresponding to the to-be-evaluated user, and determining a root cause list of the broadband quality abnormality of the to-be-evaluated user based on the sparse attention weight matrix; and determining an operation and maintenance scheme of the to-be-evaluated user according to the root cause list.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of communication, in particular to a broadband network operation and maintenance method and device. BACKGROUND

[0002] With the popularity of high-bandwidth and low-latency services, users' requirements for the quality of service (QoS) and quality of experience (QoE) of broadband networks are increasingly stringent. The traditional broadband network operation and maintenance mode is facing severe challenges. The existing technology mainly relies on: 1. Index monitoring and static threshold alarm based on simple network management protocol (SNMP): discrete performance indicators (such as port traffic, processor / memory utilization, optical power, and bit error rate) are collected from network devices through polling. A static threshold is set, and an alarm is triggered once the indicator exceeds the threshold. 2. Fault diagnosis system based on rule engine: a large number of IF-THEN rule libraries are written based on expert experience, trying to determine the fault type according to a combination of multiple indicators. However, network fault modes are complex and diverse, especially hidden faults and unknown faults, which are difficult to completely cover through limited artificial rules. Moreover, the maintenance cost of the rule library is high, and the update is lagging. 3. Application of traditional machine learning methods: some studies attempt to use traditional machine learning algorithms such as support vector machine (SVM), decision tree, random forest, and naive Bayes for fault classification or anomaly detection. However, most traditional ML models (non-specific time series models) assume that samples are independent and identically distributed, making it difficult to directly and effectively capture long-term dependencies, periodicity, and trendiness in time series. Moreover, the performance of the model is highly dependent on artificially designed features, which require deep domain knowledge, and feature engineering is time-consuming and labor-intensive, making it difficult to discover potential deep patterns in data. 4. Application of basic time series models (ARIMA, LSTM / GRU, etc.): these models can handle time series data and capture time series dependencies, but when dealing with long-distance dependencies (such as event correlation spanning hours or even days), they may still face the problem of gradient vanishing / explosion, and their parallel computing capability is limited by the sequential processing of sequences. For the large amount of concurrent time series data generated by network monitoring, training and reasoning efficiency may become a bottleneck. More importantly, the "black box" nature of these models makes it difficult to directly and reliably locate the root cause. SUMMARY

[0003] The embodiments of the present application provide a broadband network operation and maintenance method and device to at least solve the technical problem of low fault detection accuracy due to the reliance on index static threshold and preset artificial rule alarm in the related art.

[0004] According to an aspect of an embodiment of the present application, a broadband network operation and maintenance method is provided, including: collecting multi-dimensional network data of a user to be evaluated; evaluating the multi-dimensional network data using a pre-constructed broadband quality evaluation model to obtain an evaluation result, wherein the broadband quality evaluation model includes a multi-head sparse self-attention sublayer, the evaluation result is used to represent broadband quality of the user to be evaluated, the evaluation result is determined based on a score threshold value of a target time, and score threshold values of different times are different; in a case where the evaluation result indicates that the broadband quality of the user to be evaluated is abnormal, obtaining a sparse attention weight matrix corresponding to the user to be evaluated, and determining a root cause list of the broadband quality abnormality of the user to be evaluated based on the sparse attention weight matrix; and determining an operation and maintenance scheme of the user to be evaluated according to the root cause list.

[0005] Optionally, the broadband quality evaluation model is constructed by: obtaining broadband network time series data, wherein the broadband network time series data includes user side data, network access layer data, network core layer data and metadata; performing standardization processing on the broadband network time series data to obtain processed data, and performing embedding processing on the processed data to obtain a training data set; dividing the training data set into a plurality of data blocks of a preset length along a time dimension, and constructing an input sequence based on the plurality of data blocks of the preset length; inputting the input sequence into an encoding layer of an initial model to obtain an output result, completing training of the initial model, and obtaining the broadband quality evaluation model, wherein the encoding layer includes the multi-head sparse self-attention sublayer and a feedforward network sublayer, and a loss function value of the initial model is a weighted sum of a mean square error loss value of a regression task and a cross-entropy loss value of a classification task.

[0006] Optionally, the broadband quality evaluation model is constructed by: obtaining broadband network time series data, wherein the broadband network time series data includes user side data, network access layer data, network core layer data and metadata; performing standardization processing on the broadband network time series data to obtain processed data, and performing embedding processing on the processed data to obtain a training data set; dividing the training data set into a plurality of data blocks of a preset length along a time dimension, and constructing an input sequence based on the plurality of data blocks of the preset length; inputting the input sequence into an encoding layer of an initial model to obtain an output result, completing training of the initial model, and obtaining the broadband quality evaluation model, wherein the encoding layer includes the multi-head sparse self-attention sublayer and a feedforward network sublayer, and a loss function value of the initial model is a weighted sum of a mean square error loss value of a regression task and a cross-entropy loss value of a classification task.

[0007] Optionally, the acquiring the score threshold at the target time point comprises: acquiring a broadband network quality score average value of the user to be evaluated in a preset time period, a second score, a third score and a global network health index, wherein the second score represents a broadband network quality score of a user in a preset ranking position in a peer group at the target time point, the third score represents a minimum score in a service level agreement, and the peer group represents a user in a same cell as the user to be evaluated; and determining the score threshold at the target time point based on the broadband network quality score average value of the user to be evaluated in the preset time period, the second score and the third score.

[0008] Optionally, the method further comprises: acquiring a total number of devices connected to a network device corresponding to the user to be evaluated and a number of devices determined to be abnormal; determining that the network device corresponding to the user to be evaluated does not meet a first preset condition when a first ratio of the number of devices determined to be abnormal to the total number of devices is less than a first preset ratio; and acquiring a plurality of root cause clues with the highest attention weight from the root cause list when the network device corresponding to the user to be evaluated does not meet the first preset condition, wherein the target device is a device causing the broadband quality of the user to be evaluated to be abnormal.

[0009] Optionally, the acquiring the target device based on the plurality of root cause clues comprises: acquiring a total attention weight of a plurality of candidate devices corresponding to the plurality of root cause clues; determining a second ratio of an attention weight of each candidate device to the total attention weight; and determining the candidate device with the second ratio greater than a second preset ratio as the target device.

[0010] Optionally, the acquiring the root cause list of the broadband quality abnormality of the user to be evaluated based on the sparse attention weight matrix comprises: extracting a row index representing a final output of the broadband quality evaluation model from the sparse attention weight matrix; acquiring a plurality of data blocks corresponding to a plurality of column indexes with the maximum attention weight from the row index, and determining an abnormal feature input in a time window of the plurality of data blocks; and determining the root cause list based on the abnormal feature and the attention weight of the abnormal feature.

[0011] According to a further aspect of the embodiments of the present application, a broadband network operation and maintenance device is also provided, comprising: a collection module configured to collect multi-dimensional network data of a user to be evaluated; an evaluation module configured to evaluate the multi-dimensional network data by using a pre-constructed broadband quality evaluation model to obtain an evaluation result, wherein the broadband quality evaluation model comprises a multi-head sparse self-attention sublayer, the evaluation result is used to represent the broadband quality of the user to be evaluated, the evaluation result is determined based on a score threshold of a target time point, and the score thresholds of different time points are different; a determination module configured to, in a case where the evaluation result indicates that the broadband quality of the user to be evaluated is abnormal, acquire a sparse attention weight matrix corresponding to the user to be evaluated, and determine a root cause list of the broadband quality abnormality of the user to be evaluated based on the sparse attention weight matrix; and an operation and maintenance module configured to determine an operation and maintenance scheme of the user to be evaluated according to the root cause list.

[0012] According to a further aspect of the embodiments of the present application, a computer device is also provided, comprising a memory and a processor, wherein the memory is configured to store program instructions, and the processor, connected with the memory, is configured to execute the broadband network operation and maintenance method.

[0013] According to a further aspect of the embodiments of the present application, a computer program product is also provided, comprising computer instructions, which, when executed by a processor, implement the broadband network operation and maintenance method.

[0014] In the embodiments of the present application, multi-dimensional network data of a user to be evaluated is collected, the multi-dimensional network data is evaluated by using a pre-constructed broadband quality evaluation model to obtain an evaluation result, wherein the broadband quality evaluation model comprises a multi-head sparse self-attention sublayer, the evaluation result is used to represent the broadband quality of the user to be evaluated, the evaluation result is determined based on a score threshold of a target time point, and the score thresholds of different time points are different, in a case where the evaluation result indicates that the broadband quality of the user to be evaluated is abnormal, a sparse attention weight matrix corresponding to the user to be evaluated is acquired, and a root cause list of the broadband quality abnormality of the user to be evaluated is determined based on the sparse attention weight matrix, and an operation and maintenance scheme of the user to be evaluated is determined according to the root cause list. Through the broadband quality evaluation model and the dynamic threshold of different time points, the broadband quality of the user is accurately evaluated, and the root cause of the network abnormality is located based on the multi-head sparse self-attention sublayer, so that the purpose of improving the accurate evaluation and operation and maintenance of the network broadband quality is achieved, and the technical problem of low fault detection accuracy caused by the dependence on the index static threshold and the preset artificial rule alarm in the related art is solved. BRIEF DESCRIPTION OF DRAWINGS

[0015] The accompanying drawings, which are included to provide a further understanding of the application and are incorporated in and constitute a part of this application, illustrate embodiments of the application and together with the description serve to explain the application. In the drawings:

[0016] Figure 1 is a hardware structure block diagram of a computer terminal for implementing a broadband network operation and maintenance method according to an embodiment of the application;

[0017] Figure 2 is a flowchart of a broadband network operation and maintenance method according to an embodiment of the application;

[0018] Figure 3 is a root cause list determination flowchart according to an embodiment of the application;

[0019] Figure 4 is a flowchart of another broadband network operation and maintenance method according to an embodiment of the application;

[0020] Figure 5 is a structure diagram of a broadband network operation and maintenance device according to an embodiment of the application. DETAILED DESCRIPTION

[0021] In order to enable persons skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by persons skilled in the art without creative work should fall within the scope of protection of the present application.

[0022] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and in the above-described drawings are used to distinguish similar objects, and do not necessarily indicate a specific order or a chronological sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device that includes a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but can include other steps or units that are not clearly listed or inherent to the process, method, product or device.

[0023] The information collected by the embodiments of the present application is information and data authorized by the user or fully authorized by all parties, and the collection, storage, use, processing, transmission, provision, disclosure and application of related data comply with relevant laws, regulations and standards in the relevant region, necessary security measures are taken, public order and good customs are not violated, and appropriate operation portals are provided for the user to choose to authorize or refuse automatic decision results; if the user chooses to refuse, the expert decision process is entered.

[0024] To solve the problems in the related art, the embodiments of the present application provide a broadband network operation and maintenance method, which can be run in Figure 1 The computer terminal is explained and described below.

[0025] The broadband network operation and maintenance method provided by the embodiments of the present application can be executed in a mobile terminal, a computer terminal or a similar computing device. Figure 1 A hardware structure block diagram of a computer terminal for implementing the broadband network operation and maintenance method is shown. As Figure 1 shown, the computer terminal 10 can include one or more processors (the processor can include but is not limited to a microprocessor MCU or a programmable logic device FPGA processing device), a memory 104 for storing data, and a transmission module 106 for communication function through wired and / or wireless network connection. In addition, it can also include a display, a keyboard, a cursor control device, an input / output interface (I / O interface), a universal serial bus (USB) port (which can be included as one of the ports of the I / O interface), a network interface, a BUS bus. Those skilled in the art can understand that Figure 1 The structure shown is only schematic, which does not limit the structure of the above-mentioned electronic device. For example, the computer terminal 10 can include more or fewer components than those shown in Figure 1 or have a different configuration than that shown in Figure 1 .

[0026] It should be noted that the one or more processors and / or other data processing circuits described above can be referred to herein as "data processing circuits" in general. The data processing circuit can be embodied in whole or in part as software, hardware, firmware or any other combination. In addition, the data processing circuit can be a single independent processing module, or all or part of any one of the other elements combined into the computer terminal 10. As referred to in the embodiments of the present application, the data processing circuit serves as a processor to control (for example, selection of a variable resistance terminal path connected to an interface).

[0027] The memory 104 can be used to store software programs of application software and modules, such as program instructions / data storage devices corresponding to the broadband network operation method in the embodiments of the present application. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory 104, that is, implements the broadband network operation method described above. The memory 104 can include a high-speed random access memory, and can also include a non-volatile memory, such as one or more magnetic storage devices, flash memories, or other non-volatile solid-state memories. In some examples, the memory 104 can further include a memory remotely arranged with respect to the processor, which can be connected to the computer terminal 10 through a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and a combination thereof.

[0028] The transmission module 106 is used to receive or send data via a network. Specific examples of the above-mentioned network can include a wireless network provided by a communication provider of the computer terminal 10. In one example, the transmission module 106 includes a network adapter (Network Interface Controller, NIC), which can be connected to other network devices through a base station so as to be able to communicate with the Internet. In one example, the transmission module 106 can be a radio frequency (Radio Frequency, RF) module, which is used to communicate with the Internet in a wireless manner.

[0029] The display can be, for example, a touch screen type liquid crystal display (LCD), which can enable a user to interact with the user interface of the computer terminal 10.

[0030] It should be noted that, in some optional embodiments, the above-mentioned Figure 1 The computer terminal shown can include hardware elements (including circuitry), software elements (including computer code stored on a computer-readable medium), or a combination of both hardware and software elements. It should be noted that, Figure 1 is only one example of a particular implementation and is intended to illustrate the types of components that can be present in the above-described computer terminal.

[0031] Under the above-mentioned operating environment, the embodiments of the present application provide a broadband network operation method embodiment. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a group of computer executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described herein can be executed in a different order.

[0032] Figure 2 is a flowchart of a broadband network operation method according to the embodiments of the present application, such as Figure 2As shown, the method comprises the following steps:

[0033] Step S202, collect multi-dimensional network data of the user to be evaluated;

[0034] In step S202, the multi-dimensional network data includes: wideband network timing data, for example:

[0035] User-side data: user equipment information: ONU (Optical Network Unit) / ONT (Optical Network Terminal) model, firmware version, runtime length, number of restarts. Wireless environment: Wi-Fi signal strength (RSSI), signal-to-noise ratio (SNR), channel utilization, interference strength, number of connected devices. User-side network performance: request round-trip time from local gateway to target server, packet loss rate, DNS (Domain Name System) resolution delay. Application layer experience indicators: video playback stutter frequency / duration, webpage loading time, game latency / jitter. User-initiated actions: trouble ticket records, self-help troubleshooting tool usage records.

[0036] Network access layer data: (mainly collected from OLT (Optical Line Terminal), ONU): ONU status: online / offline status, online / offline reason code, optical power (received P Rx , transmitted P Tx ), bias current BIAS, temperature. PON (Passive Optical Network) port performance: port traffic (uplink / downlink rate, bandwidth utilization), CRC (Cyclic Redundancy Check) error count (multiple types), FEC (Forward Error Correction) statistics (correctable / uncorrectable block count, number of verification failures), burst error count, channel noise ratio (SNR). PON board / line card performance: CPU (Central Processing Unit) utilization, memory occupancy, board temperature, internal queue length / packet loss statistics, ASIC (Application-Specific Integrated Circuit) specific error counter.

[0037] Network convergence / core layer data (collected from OLT uplink port, convergence / core switch, BRAS (Broadband Remote Access Server)): port performance: port traffic, bandwidth utilization, error packets (CRC, Giant, Jabber, etc.), packet loss count, queue congestion state. Device performance: CPU / memory utilization, routing / switching table entries, session number, protocol neighbor status. Quality of service: jitter and packet loss rate of specific service flows (such as IPTV multicast flow). Environmental data: temperature and humidity of the machine room / device deployment point (which may affect device stability).

[0038] Metadata: user service type, broadband rate package, geographic location, network topology (user-ONU-PON port-PON board-OLT-uplink port-switch).

[0039] In step S204, the multi-dimensional network data is evaluated using a pre-constructed broadband quality evaluation model to obtain an evaluation result, wherein the broadband quality evaluation model includes a multi-head sparse self-attention sublayer, the evaluation result is used to represent the broadband quality of the user to be evaluated, the evaluation result is determined based on a score threshold of a target time, and the score thresholds of different times are different.

[0040] In step S204, the broadband quality evaluation model can be a model based on the Transformer architecture.

[0041] In step S206, if the evaluation result indicates that the broadband quality of the user to be evaluated is abnormal, a sparse attention weight matrix corresponding to the user to be evaluated is obtained, and a root cause list of the broadband quality abnormality of the user to be evaluated is determined based on the sparse attention weight matrix.

[0042] In step S208, an operation and maintenance scheme of the user to be evaluated is determined according to the root cause list.

[0043] By the steps S202 to S208, multi-dimensional network data of a user to be evaluated is collected; the multi-dimensional network data is evaluated by using a pre-constructed broadband quality evaluation model to obtain an evaluation result, wherein the broadband quality evaluation model comprises a multi-head sparse self-attention sublayer, the evaluation result is used to represent broadband quality of the user to be evaluated, the evaluation result is determined based on a score threshold value of a target time, and the score threshold values of different times are different; in a case where the evaluation result indicates that the broadband quality of the user to be evaluated is abnormal, a sparse attention weight matrix corresponding to the user to be evaluated is obtained, and a root cause list of the broadband quality abnormality of the user to be evaluated is determined based on the sparse attention weight matrix; and an operation and maintenance scheme of the user to be evaluated is determined according to the root cause list. The broadband quality of the user is accurately evaluated by using the broadband quality evaluation model and the dynamic threshold values of different times, and the root cause of the network abnormality is located based on the multi-head sparse self-attention sublayer, so that the purpose of improving the accurate evaluation and operation and maintenance of the network broadband quality is achieved, and the technical problem of low fault detection accuracy caused by the dependence on the index static threshold value and the preset artificial rule alarm in the related art is solved. The following will be described in detail.

[0044] In some embodiments of the present application, the broadband quality evaluation model is constructed by the following method: obtaining broadband network time series data, wherein the broadband network time series data comprises user side data, network access layer data, network core layer data and metadata; performing standardization processing on the broadband network time series data to obtain processed data, and performing embedding processing on the processed data to obtain a training data set; dividing the training data set into a plurality of data blocks of a preset length along the time dimension, and constructing an input sequence based on the plurality of data blocks of the preset length; inputting the input sequence into an encoding layer of an initial model to obtain an output result, completing training of the initial model, and obtaining the broadband quality evaluation model, wherein the encoding layer comprises the multi-head sparse self-attention sublayer and a feedforward network sublayer, and a loss function value of the initial model is a weighted sum of a mean square error loss value of a regression task and a cross-entropy loss value of a classification task.

[0045] It should be noted that the broadband network time series data and the multi-dimensional network data have the same data type. After obtaining the broadband network time series data, the broadband network time series data needs to be standardized. The specific processing steps are as follows:

[0046] Aligning the broadband network timing data of different sources and different frequencies to a unified timestamp grid, for example, setting the reference frequency to 1 minute. For data with a frequency lower than the reference frequency, forward padding is used to fill in the gaps; for data with a frequency higher than the reference frequency, the average value within a reference time window is calculated as the aggregation result. Then, for missing or invalid data points, the k-Nearest Neighbor (KNN) interpolation method (k=5) based on historical adjacent values is used to fill in the gaps. For continuous large segment missing data (more than 30 consecutive time steps), it is marked and ignored in the loss function during model training through the masking mechanism, and the data quality problem is recorded for subsequent data source optimization. Secondly, the IQR (Interquartile Range) rule is used to identify extreme outliers, specifically, values exceeding the upper and lower quartile boundaries of the data distribution by 1.5 times the IQR are considered abnormal, and their values are adjusted (truncated Winsorization) to the boundary value. Then, the features of different dimensions and different value ranges are uniformly standardized using Z-Score, so that their mean is 0 and standard deviation is 1, to ensure that each feature has comparability in model calculation, and the processed data is then embedded. Specifically, the processed data (where T is the time step of the input sequence, and D is the original feature dimension) is mapped to a higher-dimensional dense vector space. In this scheme, the time step of the input broadband network timing data sequence is 1440 (representing 24 hours of 1-minute granularity data), and if the original features include categories (such as ONU model, firmware version), a learnable embedding layer is used to convert them to fixed-dimensional dense vectors, which are then input into a learnable fully connected layer along with other numerical features. This step is used to capture the initial interaction between the original features and provide a rich representation basis for subsequent self-attention calculations.

[0047] After obtaining the training data set, the process of building the broadband quality evaluation model is as follows:

[0048] The training data set sequence E is divided into P non-overlapping data patches of length L along the time dimension. Each patch is independently mapped back to d model dimension by a linear layer, forming an input sequence E patched with a length of T, where each element in the input sequence represents a patch, as shown in the following formula:

[0049]

[0050] In the formula, represents the real number field, and d model represents the dimension of the feature vector mapped by each patch.

[0051] Next, fixed sinusoidal / cosine positional encodings are added to E patched The positional encoding function is of the form:

[0052]

[0053]

[0054] (pos is the position index of the patch in the sequence, i is the encoding dimension index). The summed result is the final input Input Encoder to the initial model.

[0055]

[0056] where PE denotes the positional encoding of E patched .

[0057] This input Input Encoder is sequentially passed through 12 stacked Transformer Encoder layers (where the initial model contains 12 encoder layers). Inside a single Encoder layer, the data flow follows the following order: first is the Multi-Head Sparse Self-Attention (MHsSA) sublayer, which is mapped to the query matrix Q, the key matrix K, and the value matrix V by three independent linear transformations, respectively. The attention scores AttentionScores are calculated:

[0058]

[0059] where d k is the dimension of the key vectors in the key matrix K.

[0060] The attention scores are then input into the Sparsemax activation function to obtain the sparse attention weight matrix A sparse . The sparsity of the weights is ensured to facilitate root cause localization:

[0061] A sparse = sparsemax(AttentionScores)

[0062] where sparsemax denotes an activation function.

[0063] The output of this attention head is then calculated as the weighted sum of the weights and the value matrix:

[0064] HeadOutput = Asparse V

[0065] h such attention heads are computed in parallel (each head using a different set of parameters). The outputs of the h heads are concatenated and passed through a final linear transformation W o to obtain the final output Output MHsSA of the multi-head sparse self-attention sub-layer. Output MHsSA is then concatenated with the input Input Layer to the sub-layer through a residual connection. Layer Normalization is then applied to the result of the residual connection. The resulting NormOutput MHsSA is used as the input to the next sub-layer:

[0066] NormOutput MHsSA = LayerNorm(Input layer + Output MHsSA )

[0067] where LayerNorm denotes the layer normalization and Input layer denotes the input data to the current sub-layer.

[0068] This expression describes the process in a Transformer Encoder layer where the output of a multi-head sparse self-attention sub-layer is added to the input data through a residual connection and then normalized through layer normalization.

[0069] NormOutput MHsSA is input to a FFN (Feed Forward Neural Network). The FFN consists of two linear transformations with a GELU activation function in between to obtain Output FFN . Similarly, the output Output FFN of the FFN sub-layer is concatenated with the input NormOutput MHsSA to the sub-layer through a residual connection. Layer normalization is applied to the result of the addition. The resulting NormOutput layer is the final output of the Transformer Encoder layer. This Output layer is used as the input Input Layer+1 to the next (l+1)-th Encoder layer. After processing by all 12 Encoder layers, the final sequence output is obtained:

[0070]

[0071] The model output layer is built based on this final sequence output, first an Output Encoder_N An average pooling is performed over the time dimension (i.e. patch sequence dimension) to get a fixed-size context vector z. z is input to a linear layer followed by a scaled Sigmoid function to get a continuous quality score H in the interval [0, 100]:

[0072] Then, the same context vector z is used. z is input to a linear layer to get logits for C classes. A softmax function is applied to get a normalized class probability distribution P state .

[0073] Meanwhile, the computed sparse attention weight matrix A sparse is kept for root cause localization.

[0074] The model is trained using a training dataset with historical quality scores, network state classifications, and user trouble ticket records. The historical quality score labels are generated mainly by combining historical SLA(Service Level Agreement) achievement and part of expert sampled validation data; the network state classification labels are annotated based on explicit business rules (combined with key performance indicator thresholds and duration) and confirmed fault ticket records. The user trouble ticket records are an important event feature, which are encoded and input to the model in the preprocessing stage. The model training is optimized using a combined loss function, which is the weighted sum (weight is 1) of the mean square error (MSE) loss of the regression task and the cross-entropy (CE) loss of the classification task:

[0075]

[0076] L total = L MSE + L CE

[0077] where M denotes the total number of samples in the training dataset, H i denotes the real score of the i-th sample, denotes the predicted score of the i-th sample, L MSE denotes the mean square error loss value, L CE denotes the cross-entropy loss value, C denotes the total number of classes in the classification task, y ic denotes the binary label, if the i-th sample belongs to the c-th class, y ic is 1, otherwise 0, (P state，ic) represents the probability of predicting that the i-th sample belongs to the c-th class.

[0078] The AdamW optimizer is used in the training process, and the cosine annealing scheduling strategy is adopted for the learning rate (the initial learning rate is set to 1e-4, the minimum learning rate is 1e-6, and the scheduling period is the number of iterations of one training epoch). Dropout regularization is applied after the self-attention module and the feedforward network module of the TransformerEncoder, and the Dropout rate is set to 0.1. The training adopts the batch processing mode, and the batch size is set to 64. The labeled data set is strictly divided into a training set (70%), a validation set (15%), and a test set (15%) in chronological order. By monitoring the combined loss and each task evaluation index (mean absolute error MAE of the regression task, and macro average F1-Score of the classification task) on the validation set, the early stopping strategy (stopping training when the validation set loss does not improve for 5 consecutive epochs) is adopted to prevent overfitting. And aiming at the possible class imbalance problem in the classification task, the class weighting method is adopted in the loss calculation, and higher loss weight is given to the classes with less samples. The model training is completed under the condition of reaching the minimum loss on the validation set, and the model weight is kept.

[0079] In some embodiments of the present application, the multi-dimensional network data is evaluated by using a pre-constructed wideband quality evaluation model, and the specific steps of obtaining the evaluation result are as follows: using a pre-constructed wideband quality evaluation model to evaluate the multi-dimensional network data, obtaining the wideband network quality score of the user to be evaluated at the target time and the wideband network state probability distribution of the user to be evaluated at the target time, and determining the wideband network quality score of the user to be evaluated at the target time as the first score, wherein the network state probability distribution includes: the probability that the network state of the user to be evaluated is in different risk levels; obtaining a score threshold and a preset risk threshold at the target time; in the case that the first score is higher than the score threshold or the probability that the network state of the user to be evaluated is in the target risk level is higher than the preset risk threshold, determining that the evaluation result is a target evaluation result, wherein the target evaluation result is used to indicate that the wideband quality of the user to be evaluated at the target time is abnormal.

[0080] The score threshold at the target moment is obtained, comprising: obtaining a broadband network quality score average value of the user to be evaluated in a preset period, a second score, a third score and a global network health index, wherein the second score represents a broadband network quality score of a user in a preset ranking position in a peer group at the target moment, the third score represents a minimum score in a service level agreement, and the peer group represents a user in the same cell as the user to be evaluated; and determining the score threshold at the target moment based on the broadband network quality score average value of the user to be evaluated in the preset period, the second score and the third score.

[0081] After the model is successfully trained and deployed, it enters a real-time running phase. The system continuously receives a multi-dimensional time series data stream collected in real time and preprocessed in step S1, and inputs it into the trained deep model. The model outputs a real-time and dynamic quality evaluation result of each monitored network entity (such as a single user, an active PON port, and a running PON board card) at the current moment t, including a continuous score H at moment t and a discrete state probability P state . The system establishes and implements a dynamic quality difference definition mechanism. This mechanism discards the traditional rule based on a static threshold, and can more sensitively and accurately respond to network state changes. Specifically, when a network object is determined to be “poor quality” at moment t, the judgment standard is that the model predicted quality score H of the object is lower than a threshold τ dynamic (score threshold) dynamically adjusted according to its historical performance and current business priority, or the model predicted probability P state of the object belonging to a certain unhealthy state (such as “risk” or “poor”) exceeds a preset confidence threshold θ. In the system described in the present application, the confidence threshold θ is set to 0.75. The threshold τ dynamic is continuously self-adapted according to the overall network operating condition and operation and maintenance strategy.

[0082] Specifically, τ dynamicThe calculation formula is as follows: First, prepare four necessary input values: one is the average quality score in the past 24 hours (the average value of the broadband network quality score in the preset period); two is the score at the tenth percentile of all members in the peer group to which the object belongs, based on geographical area and subscription rate, after sorting the quality scores of all members at the current time point from low to high (the second score); three is the minimum quality score line guaranteed by the service level agreement set for the object (this is a fixed value, for example, 60 points for a specific package user, the third score); four is a global index reflecting the current overall network macro health status, which is provided by an independent network monitoring module, with a value range of zero to one, and the higher the value, the worse the overall network status (zero represents normal network, one represents serious abnormality). Second, calculate a basic dynamic threshold. This basic value is obtained by multiplying the 24-hour average of the object itself by 0.6, then adding the tenth percentile score of its peer group multiplied by 0.4, and the result of the two is added. Third, adjust the basic threshold obtained in the previous step according to the global network health index. The adjustment method is to multiply the basic threshold by (1 minus the global network health index multiplied by 0.2). The effect of this is that when the global network status deteriorates (the index increases), the threshold will decrease accordingly, with a maximum decrease of 20% of the basic value. Fourth, the final dynamic score threshold is the larger of the two values, the adjusted threshold calculated in the third step and the minimum score line of the service level agreement corresponding to the object. This final threshold ensures that the evaluation standard can dynamically adapt to the current situation, and will never be lower than the bottom line of service commitment.

[0083] Finally, generate and maintain a list of quality difference objects according to this set of standards. The system monitors the quality of all objects in real time. Once a certain object is determined to be of poor quality, the system immediately generates a structured record for it. The record must contain the following information: the unique identifier of the object, the timestamp of the determination, the continuous quality score value at that time, the state probability distribution at that time (five probability values), the specific reason for triggering the determination (whether the score is lower than the dynamic threshold, or the probability of a high-risk state exceeds the fixed threshold), and the reference information pointing to the sparse attention weight data generated when the object is modeled. These records are immediately added to a dedicated, time-ordered quality difference object event queue. This queue serves as the direct input source for subsequent step four (root cause localization), ensuring that all identified problem objects can be analyzed in a timely and in-depth manner. The design of the system ensures that the generation and update of this list are efficient and without omission.

[0084] In some embodiments of the present application, the total number of devices connected to the network device corresponding to the user to be evaluated and the number of devices determined to be abnormal are also obtained; in the case where the first ratio of the number of devices determined to be abnormal to the total number of devices is less than a first preset ratio, it is determined that the network device corresponding to the user to be evaluated does not meet the first preset condition; in the case where the network device corresponding to the user to be evaluated does not meet the first preset condition, a plurality of root cause clues with the highest attention weight are obtained from the root cause list; and the target device is determined according to the plurality of root cause clues, wherein the target device is the device that causes the broadband quality anomaly of the user to be evaluated.

[0085] The specific steps of determining the target device according to the plurality of root cause clues are as follows: obtaining the total attention weight of a plurality of candidate devices corresponding to the plurality of root cause clues; determining the second ratio of the attention weight of each candidate device to the total attention weight, respectively; and determining the candidate device with the second ratio greater than a second preset ratio as the target device.

[0086] Optionally, determining the root cause list of the broadband quality anomaly of the user to be evaluated based on the sparse attention weight matrix comprises: extracting a row index representing the final output of the broadband quality evaluation model from the sparse attention weight matrix; obtaining a plurality of data blocks corresponding to a plurality of column indexes with the maximum attention weight from the row index, and determining abnormal features input in a time window of the plurality of data blocks; and determining the root cause list according to the abnormal features and the attention weight of the abnormal features.

[0087] For the user with broadband network quality anomaly, the sparse attention weight generated inside the broadband quality evaluation model is used for deep analysis to realize accurate root cause positioning of the quality problem, and a network device (PON port, PON board, OLT / switch) quality difference determination model is established based on this. Figure 3 As shown in the figure, first, root cause positioning analysis based on sparse attention weight is performed. When a network object (user to be evaluated) obj is determined to be quality abnormal (quality difference device) at time t, the system obtains the record of the object from the quality abnormal object event queue, which forcibly contains the sparse attention weight matrix generated when the model is inferred (for the input patch sequence that leads to the determination). The specific operation is as follows: 1. Extract the attention weight: the system accurately extracts the sparse attention weight matrix of the multi-head sparse self-attention sublayer from the broadband quality evaluation model last layer encoder corresponding to the last time step of the object. Let the input sequence length of the model be T, and the extracted weight matrix be where N represents the last layer Encoder. To integrate multi-head information, the attention matrices of h heads are element-level average value aggregated to obtain a single aggregated sparse attention matrix 2. Resolve key attention connections: locate to matrix i represents the row index of the final output of the model out (corresponding to the last patch, i.e. i out =T-1). In this row, identify the set of column indices J top with the largest non-zero attention weights. Here K is a pre-set hyperparameter (K=5) to control the number of root causes. This set is defined as:

[0088]

[0089] where, represents the attention weight value of the j-th index in the row i out of the aggregated sparse attention matrix, represents the K indices with the largest element values in the row i out of the aggregated sparse attention matrix.

[0090] These indices j∈J top point to the history input patches that contribute the most to the final quality assessment result. top 3. Map back to original features and time points: map the identified key patch indices j∈J j back to their corresponding specific time windows W j in the original input time series X. 4. Generate quantified Top-K root cause list: the system analyzes which original input features (e.g. CRC error count of a specific PON port, uplink optical power of a certain ONU, CPU utilization of an OLT board card) exhibit abnormal behavior within these key time windows W j . Combining the corresponding attention weight values as quantified confidence, a ranked root cause list is finally generated. The list format is: [(feature name, abnormal behavior description, time window of occurrence, confidence / attention weight), ], ranked in descending order of confidence. This list is the Top-K root cause explanation given by the model about the quality problem.

[0091] Second, we build and implement an intelligent hierarchical equipment quality model. The goal of this model is to go beyond the quality assessment of individual objects (such as user lines) and determine whether there are quality problems in the key equipment (PON port, PON board, OLT / switch) itself in the network. Its judgment logic forcibly integrates information from multiple dimensions: 1. Judgment hierarchy definition: the model covers three key equipment levels: passive optical network port (PON port), passive optical network board (PON board), and optical line terminal / converged switch (OLT / switch). 2. Judgment elements: for each level of equipment Dev, whether it is judged as a "poor quality device" (target device) (IsPoorDevice(Dev, t) = True), the following conditions must be met: a) Model evaluation results or state aggregation of the lower unit: for the PON port level, its own model quality score H t is lower than the calculated dynamic threshold τ dynamic , or its state probability P state，t satisfies the high-risk condition. For the PON board level, it must satisfy: among the PON ports under it, the proportion of "poor quality" ports exceeds the first preset ratio θ port_ratio (θ port_ratio = 30%), or the key performance indicators of the board itself (such as CPU, memory, temperature) exhibit abnormal patterns identified by the S2 model. For the OLT / switch level, it must satisfy: among the PON boards under it, the proportion of "poor quality devices" exceeds the preset threshold θ board_ratio (For example, θ board_ratio = 20%), or the key performance indicators of the device itself (such as main control board performance, backplane bandwidth utilization) exhibit abnormalities. b) Sparse attention root cause attribution verification: if the equipment Dev is judged as poor quality, its root cause must mainly point to the equipment itself or the resources it directly manages. The specific implementation is: analyze the Top-K root cause clue list of all lower units (or the equipment itself) that satisfy condition a). Calculate the proportion of the total attention weight of these clues that point to the local performance indicators of Dev (for example, the chip error count of the PON board, the switch matrix congestion count of the OLT) or its direct interface (for example, the abnormal sending optical power of the PON port). This proportion must exceed the second preset ratio θ attribution (θ attribution= 50%). If most of the root causes point to an upstream device problem (e.g. OLT points to a core router port packet loss) or a downstream terminal problem (e.g. PON port points to internal faults of multiple independent ONUs), even if condition a) is met, the device Dev will not be marked as a poor quality device, but the problem attribution is passed to the corresponding direction. By combining model direct evaluation, down-hanging unit state aggregation, and strict attention root cause attribution verification, this step realizes accurate, intelligent, and hierarchical judgment of the quality state of network devices.

[0092] After determining the target device, the system integrates the following key data: real-time quality evaluation score (Ht) of the user to be evaluated; root cause list, each of which is attached with its initial confidence C attn obtained by directly analyzing the sparse attention weight output by the broadband network quality evaluation model; and device state (whether it is a target device) of the relevant network device (PON port, PON board, OLT, switch). These data constitute a diagnosis sequence. Subsequently, the diagnosis sequence is submitted as input to a processing strategy decision engine. The core of the engine is a pre-defined, structured rule base. That is, it contains standard processing plans for various different network failure situations. The system will carefully compare the current case (including score, root cause list, confidence, device state, and other specific details) with various plans. When the description of a plan completely matches the specific situation of the current case (for example, "applicable to a score lower than 70, the main reasons include 'weak optical signal' (confidence higher than 0.2) and'many port errors' (confidence higher than 0.2), and all relevant device states are normal"), the system determines that this plan should be executed. The matched plan will clearly indicate all the things that need to be done next: it will specify the processing priority of the task (such as "medium"); it will provide all the detailed content required to generate a work order (including user identification, score, reason list ranked by confidence, and recommended inspection and repair sequence); and it will list other auxiliary operations that need to be automatically executed (for example, "automatically start a remote optical path detection", "send a Wi-Fi optimization suggestion to the user's mobile phone App"). Finally, the system will strictly follow the selected plan to execute, and automatically generate a standardized work order and send it to the responsible team or system, while starting all the automatic auxiliary operations required by the plan.

[0093] It can be understood that the broadband network operation and maintenance method provided in the embodiments of the present application adopts the time sequence model technology (Transformer architecture), which can more deeply analyze the complex relationship between the multi-dimensional indicators of the broadband network changing over time, make the network quality evaluation result more accurate, and reflect the network dynamic changes faster, which is significantly improved compared with the traditional method. And by using the attention mechanism, while giving the evaluation result, it can automatically find out a few key factors causing the network problem, providing clear and quantifiable positioning basis for fault troubleshooting, solving the problem of "not understanding" of complex models, greatly improving the efficiency and accuracy of root cause positioning. The intelligent equipment quality difference (abnormality) judgment method based on the evaluation result of the deep model and the root cause direction replaces the simple counting and threshold method in the past, so as to more accurately judge the fault responsibility attribution and the convergence problem range, and improve the accuracy of fault handling. By analyzing the change trend in the data, potential network faults and equipment aging risks can be found in advance, early warning is realized, and the operation and maintenance work is changed from "after repair" to "prevention", which effectively reduces the occurrence of major faults and business interruption losses, and reduces the operation and maintenance cost.

[0094] In order to better illustrate the broadband network operation and maintenance method provided by the present application, another broadband network operation and maintenance method is further provided in the embodiments of the present application, as shown in Figure 4 As shown in the figure, step 1, time sequence data collection and preprocessing implementation, systematic data collection,

[0095] User side data: collect the ONU model (F677) of user A, the running time, the restart record, the home Wi-Fi signal strength (-55dBm), the signal-to-noise ratio (35dB), the number of connected devices (8), the user end Ping RTT (average 15ms), and the video stuttering event (2 times per hour during peak period).

[0096] Network access layer data: ONU of user A, including real-time optical power (reception -23.5dBm, transmission +2.1dBm), bias current (35mA), and temperature (45℃).

[0097] PON port connected by user A: uplink and downlink traffic on the port (peak utilization rate 70%), CRC-8 error count (150 new per hour), uncorrectable FEC block number (5 new per hour), and burst error count. PON board (OLT-SN-01, Slot3): CPU utilization rate (average 40%), memory occupancy rate (55%), and board temperature (50℃).

[0098] Network convergence / core layer data: (collected from OLT uplink port, convergence switch SW-Core-01): OLT uplink port, including traffic (peak utilization rate 60%), error packet count (low), packet loss count (low). Device CPU utilization rate of switch SW-Core-01 (30%).

[0099] Metadata: User A rate package (1000M), geographic location (a certain cell), network topology.

[0100] Data preprocessing: Align different frequency data (such as ONU optical power 5 minutes, PON port CRC error 1 minute, Wi-Fi index 10 seconds) to a unified 1-minute timestamp grid, and use forward filling to handle missing values. Identify and smooth transient CRC error spikes (values more than 3 times the standard deviation of the mean). Z-Score standardization is performed on all numerical features. The standardized multi-dimensional time series data (assuming it contains 50 features) is embedded into the model working dimension d model = 512 by a linear layer.

[0101] Step 2: Build and apply a broadband quality dynamic evaluation model;

[0102] First, patch the time series data of user A after preprocessing and embedding in step 1 with length L = 12 (i.e. 12 minutes of data) and input it into the pre-trained time series Transformer model containing 12 layers of Encoder.

[0103] Get the continuous quality score H t output by the model at each time point t, for example, at a certain time t1 during the evening peak, user A's Get the network state probability distribution P state，t , for example, at time t1,

[0104] It can be understood that, the probability of the network state of the user to be evaluated at time t1 being at different risk levels.

[0105] Get the internal sparse attention weight matrix

[0106] Step 3: Dynamic quality evaluation and quality difference object identification;

[0107] Perform quality difference determination, first calculate the dynamic threshold τ dynamic , input user A's average H t over the past 24 hours (assuming 78), and the current 10th percentile H t(assuming 70), SLA minimum score (60), global network health index (assuming 0.2, indicating slight deterioration). Compute base threshold: 78*0.6 + 70*0.4 = 46.8 + 28 = 74.8. Adjust base threshold: 74.8*(1-0.2*0.2) = 74.8*0.96 = 71.8. Final τ dynamic = max(71.8, 60) = 71.8. Where high-risk probability threshold θ risk = 0.75 (fixed setting).

[0108] User A's state at time t1 is determined as: The score-lower-than-dynamic-threshold condition is satisfied. The high-risk-probability condition is not satisfied.

[0109] Conclusion: Since the score-lower-than-dynamic-threshold condition is satisfied, user A is determined as a poor-quality user at time t1.

[0110] Poor-quality object list: Add user A's poor-quality event (including identifier, timestamp determination reason "score-lower-than-dynamic-threshold", attention weight reference) to the poor-quality object event queue, such as the queue composed of users with abnormal broadband quality.

[0111] Step 4: Root cause positioning and intelligent hierarchical device poor-quality model construction;

[0112] The system then determines the poor quality of the relevant network equipment according to the hierarchical device poor-quality model defined in S4.

[0113] Poor quality (abnormal device) determination of PON port:

[0114] To determine the poor quality of this PON port, conditions a model evaluation / underhanging unit state aggregation condition and condition b attention weight-based root cause attribution verification condition must be met simultaneously.

[0115] Condition a: model evaluation result or underhanging unit state aggregation;

[0116] This PON port (OLT-SN-01 / 3 / 5) has 32 users under it. According to the poor-quality object identification result in step 3, there are currently 8 users (including user A) determined as poor-quality users. Calculate the proportion of poor-quality users: Ratio poor＿users = 8 / 32 = 0.25 (i.e. 25%). First preset proportion θ port＿ratio (30%). Since 25% < θ port＿ratio = 30%, this PON port does not meet condition a triggered by a large number of poor-quality users under it. (Note: If the model directly outputs the PON port's H t and P state，tThen, the dynamic threshold τ is calculated according to S3 dynamic and the high-risk state probability threshold θ, and the result is considered in combination with the aggregation result of the lower hanging unit. In this example, we focus on the aggregation of the lower hanging unit.

[0117] Current conclusion: PON port (OLT-SN-01 / 3 / 5) meets condition a (first preset condition).

[0118] Condition b (second preset condition): sparse attention root cause attribution verification:

[0119] This step verifies the root cause of the poor quality of the lower hanging user (such as user A, as a representative case causing the PON port to be investigated), and to what extent the problem of the PON port (OLT-SN-01 / 3 / 5) itself can be blamed.

[0120] i. Extract the root cause clues of the relevant object: use the Top-3 root cause list and its normalized attention weight (i.e. confidence) obtained by "root cause positioning" for poor quality user A at t1 time:

[0121] Clue 1: (PONDCRC-8 error count, significantly increased, weight w1=0.35)

[0122] Clue 2: (ONU received optical power, low and fluctuating, weight w2=0.28)

[0123] Clue 3: (Home Wi-Fi channel utilization, continuously high, 权重 w3=0.15) The total attention weight of these Top-K clues is W total_attention = w1+w2+w3=0.35+0.28+0.15=0.78. (Note: W total_attention here refers to the sum of the attention of Top-K root cause clues, not the sum of all possible clues, because we are concerned about those considered most important by the model.)

[0124] ii. Classify the root cause clues and calculate the weight of attribution to the PON port:

[0125] Analyze the relationship between each Top-K root cause clue of user A and the PON port (OLT-SN-01 / 3 / 5) under investigation:

[0126] Clue 1 ("PONDCRC-8 error count"): This is an indicator that directly reflects the local performance status of the PON port (OLT-SN-01 / 3 / 5). Therefore, its weight w1=0.35 is attributed to the PON port.

[0127] Lead 2 ("ONU received optical power"): This is a user A's ONU equipment side indicator, the problem point is in the ONU or optical path, which is the directly managed downstream link or terminal problem relative to the PON port, but is not directly equivalent to the PON port itself failure. For the sake of simplicity of attribution, if the problem is clearly directed to the ONU side, it is not counted as the PON port itself. Lead 3 ("Home Wi-Fi channel utilization rate"): This is a user A's home internal wireless environment problem, which is a remote user environment problem relative to the PON port.

[0128] Therefore, in the Top-K root cause leads of user A, the total attention weight of the leads pointing to the local performance indicators (such as CRC error) of the PON port (OLT-SN-01 / 3 / 5) or its direct interface is W attrib＿to＿port = w1= 0.35.iii. Calculate the attribution proportion of the attention weight and compare it with the threshold value:

[0129] Calculate the proportion of the total attention weight of the leads pointing to the local performance indicators of the PON port (OLT-SN-01 / 3 / 5) or its direct interface in these Top-K leads:

[0130] P attribution＿port = W attrib＿to＿port / W total＿attention = 0.35 / 0.78 ≈ 0.4487 (i.e., about 44.87%).

[0131] Compare this proportion with the second preset ratio θ attribution = 50%:

[0132] 44.87% < θ attribution = 50%

[0133] iv. Draw a root cause attribution verification conclusion:

[0134] Since the calculated attention weight proportion attributed to the PON port itself (about 44.87%) fails to exceed the preset attribution threshold (50%), the root cause attribution verification based on the attention weight indicates that the main root cause of the quality problem of user A does not sufficiently point to the PON port (OLT-SN-01 / 3 / 5) itself. That is, the PON port (OLT-SN-01 / 3 / 5) does not meet the condition b) of the equipment quality determination in S4. Final conclusion: since the PON port (OLT-SN-01 / 3 / 5) fails to meet both condition a (the proportion of the underhanging quality problem users does not reach the threshold value) and condition b (the root cause is not mainly attributed to the PON port itself), the system determines that the PON port (OLT-SN-01 / 3 / 5) is not determined as a quality problem PON port.

[0135] Step 5: Process according to model evaluation, root cause positioning, and equipment quality status:

[0136] The system integrates the following information: User A is a poor user (H t = 65.8). Top-3 root causes point to PON port CRC error, ONU optical power problem, and home Wi-Fi interference. Related equipment (PON port, PON board, OLT, switch) is not determined as poor equipment.

[0137] The processing strategy is that the system automatically generates a work order based on the root cause list, with the priority set to “medium”. The work order content: User A (jsyc_userA) has a low broadband quality score (65.8), main root cause clues: 1. High CRC error on OLT-SN-01 / 3 / 5 ports (confidence 0.35), 2. Low and fluctuating ONU received optical power (-23.5 dBm, confidence 0.28), 3. High home Wi-Fi channel utilization (confidence 0.15). The recommended inspection order: user home Wi-Fi environment -> ONU optical path -> OLT port status. The work order is automatically assigned to the outside line maintenance personnel responsible for the area. At the same time, the system automatically triggers a remote optical link diagnosis (OTDR simulation test) on the ONU of user A, and pushes a home Wi-Fi optimization suggestion (such as modifying the channel) to the APP of user A.

[0138] Figure 5 A broadband network operation and maintenance device is shown, which comprises:

[0139] The acquisition module 50 is configured to acquire multi-dimensional network data of a user to be evaluated;

[0140] The evaluation module 52 is configured to evaluate the multi-dimensional network data using a pre-constructed broadband quality evaluation model to obtain an evaluation result, wherein the broadband quality evaluation model includes a multi-head sparse self-attention sublayer, the evaluation result is used to represent the broadband quality of the user to be evaluated, and the evaluation result is determined based on a score threshold value of a target time, and the score threshold values of different times are different.

[0141] The determination module 54 is configured to, in a case where the evaluation result indicates that the broadband quality of the user to be evaluated is abnormal, acquire a sparse attention weight matrix corresponding to the user to be evaluated, and determine a root cause list of the broadband quality abnormality of the user to be evaluated based on the sparse attention weight matrix.

[0142] The operation and maintenance module 56 is configured to determine an operation and maintenance scheme of the user to be evaluated according to the root cause list.

[0143] The broadband network operation and maintenance device adopts multi-dimensional network data of a user to be evaluated for collection, adopts a broadband quality evaluation model constructed in advance to evaluate the multi-dimensional network data, and obtains an evaluation result, wherein the broadband quality evaluation model comprises a multi-head sparse self-attention sublayer, the evaluation result is used to represent broadband quality of the user to be evaluated, the evaluation result is determined based on a score threshold value of a target time, and the score threshold values of different times are different; in a case where the evaluation result indicates that the broadband quality of the user to be evaluated is abnormal, a sparse attention weight matrix corresponding to the user to be evaluated is obtained, and a root cause list of the broadband quality abnormality of the user to be evaluated is determined based on the sparse attention weight matrix; and an operation and maintenance scheme of the user to be evaluated is determined according to the root cause list. The broadband quality evaluation model and the dynamic threshold values of different times are used to accurately evaluate the broadband quality of the user, and the multi-head sparse self-attention sublayer is used to locate the root cause of the network abnormality, so that the purpose of improving the accurate evaluation and operation and maintenance of the broadband quality of the network is achieved, and the technical problem of low fault detection accuracy caused by the dependence on the index static threshold value and the preset artificial rule alarm in the related art is solved.

[0144] The evaluation module 52 comprises a construction sub-module configured to construct the broadband quality evaluation model by: obtaining broadband network time series data, wherein the broadband network time series data comprises user side data, network access layer data, network core layer data and metadata; performing standardization processing on the broadband network time series data to obtain processed data, and performing embedding processing on the processed data to obtain a training data set; dividing the training data set into a plurality of data blocks of a preset length along a time dimension, and constructing an input sequence based on the plurality of data blocks of the preset length; inputting the input sequence into an encoding layer of an initial model to obtain an output result, completing training of the initial model, and obtaining the broadband quality evaluation model, wherein the encoding layer comprises the multi-head sparse self-attention sublayer and a feedforward network sublayer, and a loss function value of the initial model is a weighted sum of a mean square error loss value of a regression task and a cross-entropy loss value of a classification task.

[0145] The constructing submodule comprises an evaluation unit configured to evaluate the multi-dimensional network data by using a pre-constructed broadband quality evaluation model to obtain an evaluation result, comprising: evaluating the multi-dimensional network data by using a pre-constructed broadband quality evaluation model to obtain a broadband network quality score of the user to be evaluated at a target time and a broadband network state probability distribution of the user to be evaluated at the target time, and determining the broadband network quality score of the user to be evaluated at the target time as a first score, wherein the network state probability distribution comprises a probability that the network state of the user to be evaluated is at different risk levels; obtaining a score threshold at the target time and a preset risk threshold; in a case where the first score is higher than the score threshold or the probability that the network state of the user to be evaluated is at a target risk level is higher than the preset risk threshold, determining that the evaluation result is a target evaluation result, wherein the target evaluation result is used to indicate broadband quality anomaly of the user to be evaluated at the target time.

[0146] The evaluation unit comprises an obtaining subunit configured to obtain a score threshold at the target time, comprising: obtaining a broadband network quality score average value of the user to be evaluated within a preset time period, a second score, a third score and a global network health index, wherein the second score represents a broadband network quality score of a user at a preset ranking position in a peer group at the target time, the third score represents a minimum score in a service level agreement, and the peer group represents a user in the same cell as the user to be evaluated; determining the score threshold at the target time based on the broadband network quality score average value of the user to be evaluated within the preset time period, the second score and the third score.

[0147] The broadband network operation and maintenance device further comprises a judgment submodule configured to obtain a total number of devices connected to a network device corresponding to the user to be evaluated and a number of devices determined to be abnormal; in a case where a first ratio of the number of devices determined to be abnormal to the total number of devices is less than a first preset ratio, determining that the network device corresponding to the user to be evaluated does not satisfy a first preset condition; in a case where the network device corresponding to the user to be evaluated does not satisfy the first preset condition, obtaining a plurality of root cause clues with the highest attention weight from the root cause list; determining a target device according to the plurality of root cause clues, wherein the target device is a device causing broadband quality anomaly of the user to be evaluated.

[0148] The judgment submodule comprises a determination unit configured to determine a target device according to the plurality of root cause clues, comprising: obtaining a total attention weight of a plurality of candidate devices corresponding to the plurality of root cause clues; respectively determining a second ratio of the attention weight of each candidate device to the total attention weight, and determining the candidate device with the second ratio greater than a second preset ratio as the target device.

[0149] The broadband network operation and maintenance device further includes a root factor module configured to determine a root factor list of the broadband quality abnormality of the user to be evaluated based on the sparse attention weight matrix, including: extracting a row index representing a final output of the broadband quality evaluation model from the sparse attention weight matrix; obtaining a plurality of data blocks corresponding to a plurality of column index pairs with the largest attention weight from the row index, and determining abnormal features input in a time window of the plurality of data blocks; and determining the root factor list according to the abnormal features and the attention weight of the abnormal features.

[0150] It should be noted that, Figure 5 The broadband network operation and maintenance device is configured to execute the broadband network operation and maintenance method. Figure 2 The broadband network operation and maintenance device is configured to execute the broadband network operation and maintenance method.

[0151] The broadband network operation and maintenance device is configured to execute the broadband network operation and maintenance method.

[0152] The broadband network operation and maintenance device is configured to execute the broadband network operation and maintenance method.

[0153] The above-mentioned serial numbers of the embodiments of the present application are only for description, and do not represent the advantages and disadvantages of the embodiments.

[0154] In the above-described embodiments of the present application, the description of each embodiment has its own emphasis, and the parts not described in detail in a certain embodiment can be referred to the relevant description of other embodiments.

[0155] In the several embodiments provided by the present application, it should be understood that the disclosed technology can be implemented in other ways. Of course, the device embodiment described above is only schematic. For example, the division of the units can be a logical function division, and there can be another division manner in actual implementation, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed mutual coupling or direct coupling or communication connection between each of the units can be indirect coupling or communication connection through some interface, unit or module, and can be electrical or other forms.

[0156] The units described as separate components may or may not be physically separate, and the components displayed as units may or may not be physical units, that is, may be located in one place, or may be distributed to multiple units. Part or all of the units can be selected according to actual needs to achieve the purpose of the embodiment scheme.

[0157] In addition, the functional units in each embodiment of the present application can be integrated in one processing unit, or each unit can be physically present alone, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0158] The integrated unit, if realized in the form of a software functional unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application essentially or the part of the prior art that contributes to the technical solutions or all or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The foregoing storage medium includes: a U disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a mobile hard disk, a magnetic disk or an optical disk, and various program code storage media.

[0159] The above is only the preferred embodiment of the present application, and it should be pointed out that for those skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, and these improvements and refinements should be considered as the protection scope of the present application.

Claims

1. A broadband network operation and maintenance method, characterized in that, include: Collect multi-dimensional network data of the users to be evaluated; The multi-dimensional network data is evaluated using a pre-built broadband quality assessment model to obtain an evaluation result. The broadband quality assessment model includes a multi-head sparse self-attention sublayer. The evaluation result is used to represent the broadband quality of the user to be evaluated. The evaluation result is determined based on a scoring threshold at a target time, and the scoring threshold is different at different times. If the evaluation result indicates that the broadband quality of the user to be evaluated is abnormal, obtain the sparse attention weight matrix corresponding to the user to be evaluated, and determine the root cause list of the broadband quality abnormality of the user to be evaluated based on the sparse attention weight matrix. The operation and maintenance plan for the user to be evaluated is determined based on the root cause list.

2. The method according to claim 1, characterized in that, The broadband quality assessment model is constructed in the following way: Acquire broadband network time-series data, wherein the broadband network time-series data includes: user-side data, network access layer data, network core layer data, and metadata; The broadband network time-series data is standardized to obtain processed data, and the processed data is then embedded to obtain a training dataset. The training dataset is divided into multiple data blocks of preset length along the time dimension, and an input sequence is constructed based on the multiple data blocks of preset length; The input sequence is input into the encoding layer of the initial model to obtain the output result, thus completing the training of the initial model and obtaining the broadband quality assessment model. The encoding layer includes the multi-head sparse self-attention sub-layer and the feedforward network sub-layer. The loss function value of the initial model is a weighted sum of the mean squared error loss value of the regression task and the cross-entropy loss value of the classification task.

3. The method according to claim 2, characterized in that, The multi-dimensional network data is evaluated using a pre-built broadband quality assessment model to obtain evaluation results, including: The multi-dimensional network data is evaluated using a pre-built broadband quality assessment model to obtain the broadband network quality score of the user to be evaluated at the target time and the broadband network status probability distribution of the user to be evaluated at the target time. The broadband network quality score of the user to be evaluated at the target time is determined as the first score. The network status probability distribution includes the probability that the network status of the user to be evaluated is at different risk levels. Obtain the scoring threshold and preset risk threshold at the target time; If the first score is higher than the score threshold or the probability that the network status of the user to be evaluated is at the target risk level is higher than the preset risk threshold, the evaluation result is determined as the target evaluation result, wherein the target evaluation result is used to indicate that the broadband quality of the user to be evaluated is abnormal at the target time.

4. The method according to claim 3, characterized in that, Obtaining the scoring threshold at the target time includes: The average broadband network quality score, second score, third score, and global network health index of the user to be evaluated within a preset time period are obtained. The second score represents the broadband network quality score of the peer group at the preset sorting position at the target time. The third score represents the lowest score in the service level agreement. The peer group represents users in the same cell as the user to be evaluated. The scoring threshold for the target time is determined based on the average broadband network quality score of the user to be evaluated within a preset time period, the second score, and the third score.

5. The method according to claim 1, characterized in that, The method further includes: Obtain the total number of devices connected to the network device corresponding to the user to be evaluated and the number of devices identified as abnormal; If the first ratio of the number of devices identified as abnormal to the total number of devices is less than a first preset ratio, it is determined that the network device corresponding to the user to be evaluated does not meet the first preset condition. If the network device corresponding to the user to be evaluated does not meet the first preset condition, multiple root cause clues with the highest attention weight are obtained from the root cause list. The target device is determined based on the multiple root cause clues, wherein the target device is the device that causes the broadband quality anomaly of the user to be evaluated.

6. The method according to claim 5, characterized in that, Identifying the target device based on the multiple root cause clues includes: Obtain the total attention weight of multiple candidate devices corresponding to the multiple root cause clues; A second ratio of the attention weight of each candidate device to the total attention is determined, and candidate devices whose second ratio is greater than a second preset ratio are determined as the target devices.

7. The method according to claim 1, characterized in that, Based on the sparse attention weight matrix, a list of root causes of broadband quality anomalies for the user to be evaluated is determined, including: Extract the row index representing the final output of the broadband quality assessment model from the sparse attention weight matrix; Obtain multiple data blocks corresponding to the multiple column indices with the highest attention weight from the row index, and determine the abnormal characteristics of the input within the time window of the input of the multiple data blocks; The root cause list is determined based on the anomalous features and their attention weights.

8. A broadband network operation and maintenance determination device, characterized in that, include: The data acquisition module is used to collect multi-dimensional network data of the user to be evaluated. An evaluation module is used to evaluate the multi-dimensional network data using a pre-built broadband quality evaluation model to obtain an evaluation result. The broadband quality evaluation model includes a multi-head sparse self-attention sub-layer. The evaluation result is used to represent the broadband quality of the user to be evaluated. The evaluation result is determined based on a scoring threshold at a target time, and the scoring threshold is different at different times. The determination module is used to obtain the sparse attention weight matrix corresponding to the user to be evaluated when the evaluation result indicates that the broadband quality of the user to be evaluated is abnormal, and to determine the root cause list of the broadband quality abnormality of the user to be evaluated based on the sparse attention weight matrix. The operation and maintenance module is used to determine the operation and maintenance plan for the user to be evaluated based on the root cause list.

9. A computer device, characterized in that, include: A memory and a processor, wherein the memory is used to store program instructions; The processor, connected to the memory, is used to execute the broadband network operation and maintenance method according to any one of claims 1 to 7.

10. A computer program product comprising computer instructions, characterized in that, When the computer instructions are executed by the processor, they implement the broadband network operation and maintenance method according to any one of claims 1 to 7.

Images